col_1,col_2,col_3,col_4,col_5,col_6,col_7,col_8,abstract,authors,id,references,col_acknowledgments,col_spelling,col_brazilian,col_each,col_error,col_errors,col_last,col_computational,col_deep,col_9,col_squibs,col_evaluation,col_dependent,col_csiro,col_reference,col_book,col_robots,col_judith,col_reviewed,col_adam,col_10,col_automatic,col_type,col_partition,col_ctb,col_previous,col_semeval-2012,col_word,col_edge,col_web,col_when,col_than,col_a,col_compositional,col_linguistic,col_100,col_lexicalization,col_vw-ccg,col_proof,col_appendix,col_language,col_42,col_philipp,col_tu,col_automated,col_claudia,col_obituary,col_jane,col_recognizing,Concatenated Text,Future_Work
"1 introduction :In automatic speech and language processing, many technologies make extensive use of written or read text sets. These linguistic corpora are a necessity to train models or to extract rules, and the quality of the results strongly depends on a corpus’ content. Often, the reference corpus should provide a maximum diversity of content. For example, in Tian, Nurminen, and Kiss (2005), and Tian and Nurminen (2009), it turns out that maximizing the text coverage of the learning corpus improves an automatic syllabification based on a neural network. Similarly, a high quality speech synthesis system based on the selection of speech units requires a rich corpus in terms of diphones, diphones in context, triphones, and prosodic markers. In particular, Bunnell (2010) shows the importance of a good coverage of diphones and triphones for the intelligibility of a voice produced by a unit selection speech synthesis system.
To cover the best attributes needed for a task, several strategies are then possible. A first method—very simple—is to collect text randomly, but it soon becomes expensive because of the natural distribution of linguistic events following the Zipf’s law. Very few events are extremely frequent and many events are very rare. This problem is often made difficult by the fact that many technologies require several variants of the same event (as in a Text-to-Speech [TTS] system using several acoustical versions of the same phonological unit). Usually, a large volume of data needs to be collected. However, and depending on the applications, building such corpora is often achieved under a constraint of parsimony. As an example, for a TTS system, a high-quality synthetic voice generally needs a huge number of speech recordings. But minimizing the duration of a recording is also a critical point to ensure uniform quality of the voice, to reduce the drudgery of the recording, to reduce the financial cost, or to follow a technical constraint on the amount of collected data for embedded systems. Moreover, a reduced set tends to limit the need of human implication for checking the data (transcription and annotation). Similarly, in the natural language processing field (NLP), the adaptation of a generic model to a specific domain often requires new annotated data that illustrate its specificities (as in Candito, Anguiano, and Seddah 2011). However, the creation cost of such data highly depends on the kind of labels used to adapt the model. In particular, the annotation in syntax trees is really more expensive than in Part-of-Speech (POS) tags. Then, it could be more efficient to annotate a compact corpus that reflects the phenomena variability than a corpus with a natural distribution of events, which implies many redundancies (see Neubig and Mori 2010).
In a machine learning framework, the active learning strategy can be used as an alternative that reduces the manual data annotation effort to design the training corpus without diminishing the quality of the model to train (see Settles 2010 or Schein, Sandler, and Ungar 2004). It consists of building the corpus iteratively by choosing an item according to an external source of information (a user or an experimental measure). This approach has been applied in NLP, speech recognition, and spoken language understanding (see for instance Tomanek and Olsson 2009 and Gotab, Béchet, and Damnati 2009).
A second alternative, when no direct quality measure is available, consists of covering a large set of attributes that may impact the final quality (after annotation or recording). This kind of approach might also be preferred when the final corpus is built in one batch (for instance, because of out-sourcing or annotator/performer consistency constraints). A method could be an automatic extraction from a huge text corpus of a minimal sized subset that covers the identified attributes. This
problem is a generalization of the Set-Covering Problem (SCP), which is an NPhard problem, as shown in Karp (1972). It is then necessary to use heuristics or sub-optimal algorithms for a reasonable computation time. Moreover, Raz and Safra (1997) and Alon, Moshkovitz, and Safra (2006) have shown that the SCP cannot be polynomially approximated with ratio c × ln(n) unless P = NP, when c is a constant, and n refers to the size of the universe to cover. That means that one cannot be certain to obtain a result under this ratio with any polynomial algorithm. However, the latter complexity results are given for any kind of distribution in the mono-representation case. One can ask if good multi-represented coverages can be achieved efficiently on data following Zipf’s law, which is usual in the domain of NLP.
Within the field of speech processing, the most frequently used strategy is a greedy method based on an agglomeration policy. This iterative algorithm selects the sentence with the highest score at each iteration. The score reflects the contribution of the sentence to the covering under construction. In Gauvain, Lamel, and Eskénazi (1990), this methodology has been applied to build a database of read speech from a text corpus for the evaluation of speech recognition systems using hierarchically organized covering attributes. Van Santen and Buchsbaum (1997) have tested different variants of greedy selection of texts by varying the units to cover (diphones, duration, etc.) and the “scores” for a sentence depending on the considered applications. In Tian, Nurminen, and Kiss (2005), the learning corpus for an automatic system of syllabification is designed using a greedy approach with the Levenshtein distance as a score function in order to maximize its text diversity. In François and Boëffard (2001), the methodology gives a priority to the rarest categories of allophones. The latter methodology has been implemented for the definition of the multi-speaker corpus Neologos in Krstulović et al. (2006). In the article of Krul et al. (2006), the authors constructed a corpus where the distribution of diphonemes/triphonemes matches a uniform distribution. A greedy algorithm is led by a score function based on the Kullback-Liebler divergence. A similar method is used in Krul et al. (2007) to design a reduced database in accordance with a specific domain distribution. Kawai et al. (2000) propose a pair exchange mechanism that Rojc and Kačič (2000) apply after a first reverse greedy algorithm—also called spitting greedy—deleting the useless sentences. In Cadic, Boidin, and d’Alessandro (2010), the covering of “sandwich” units (defined to be more adapted to corpus-based speech synthesis) is carried out by generating new sentences in a semi-automatic way. Candidates are generated using finite state transducers. The sentences are ordered according to a greedy criterion (their sandwiches richness) and presented to a human evaluator. This collection of artificial and rich sentences enables an effective reduction of the size of the covering but requires expensive human intervention to obtain semantically correct sentences that will be therefore easier to record.
The results of these previously cited studies are difficult to compare because of the different initial corpora and covering constraints (partial or full covering) and evaluation criteria (the number of gathered sentences, the Kullback divergence, etc.). In Zhang and Nakamura (2008), a priority policy for the rare units is added into an agglomerative greedy algorithm in order to get a covering of triphoneme classes from a large text corpus in Chinese language. The results show that this priority policy driven by the score function and the phonetic content of the sentences reduces the covering size compared with a standard agglomerative greedy algorithm.
Similarly, in François and Boëffard (2002), several combinations of greedy algorithms (agglomeration, spitting, pair exchange, or priority to rare units) were applied
to the construction of a corpus for speech synthesis in French containing at least three representatives of the most frequent diphones. Based on this work, the best strategy would be the application of an agglomerative greedy followed by a spitting greedy algorithm. During the agglomeration phase, the score of a sentence corresponds to the number of its unit instances that remain to be covered normalized by its length. During the spitting phase, at each iteration, the longest redundant sentence is removed from the covering. This algorithm is called the Agglomeration and Spitting Algorithm (ASA).
As an alternative to a greedy algorithm, which is sub-optimal, solving the SCP using Lagrangian relaxation principles can provide an exact solution for problems of reasonable size. However, for speech processing, the SCP has several millions of sentences with tens of thousands of covering features. Considering these practical constraints, Chevelu et al. (2007) adapted a Lagrangian relaxation based algorithm proposed by Caprara, Fischetti, and Toth (1999). In the context of Italian railways, Caprara, Fischetti, and Toth proposed heuristics to solve scheduling problems and won a competition, called Faster, organized by the Italian Operational Research Society in 1994, ahead of other Lagrangian relaxation heuristics–based algorithms, like Ceria, Nobili, and Sassano (1998).
In Chevelu et al. (2007, 2008), the algorithm takes into account the constraints of multi-representation. A minimal number of representatives for the same unit may be required. The proposed algorithm, called LamSCP — Lagrangian-based Algorithm for Multi-represented SCP — is applied to extract coverings of diphonemes with a mono- or a 5-representation and coverings of triphonemes with mono-representation constraints. These results are compared with the greedy strategy ASA and are about 5% to 10% better. Besides, the LamSCP provides a lower bound for the cost of the optimal covering and allows for evaluating the quality of the results.
In Barbot, Boëffard, and Delhay (2012), phonological content of diphoneme coverings is studied regarding many parameters. These coverings are obtained by different algorithms (LamSCP, ASA, greedy based on the Kullback divergence) and some of the coverings are randomly completed to reach a given size (from 20,000 to 30,000 phones). It turns out that the coverings obtained using LamSCP and ASA provide a good representation of short units and the representation of long units mainly depends on the length of the corpus.
In this article, we present in more detail the LamSCP algorithm and its score functions and heuristics that take into account multi-representation constraints. We deepen the study about the performance of LamSCP for the construction of a phonologically rich corpus according to the size of the search space. We evaluate LamSCP and ASA algorithms on a corpus of sentences in English for a covering of multi-represented diphones, where the minimal number of required unit representatives varies from one to five. We also compare them in the case of very constrained triphoneme coverings in English and French, which represent about 12 times more units to cover. Additionally, both algorithms are tested to provide multi-represented coverings of POS tags in order to assess their ability to deal with different kinds of linguistic data. A particular effort has been made on methodology to obtain comparable measures, to study the stability of both algorithms, and to establish confidence intervals for each solution.
This article is organized as follows. In Section 2, the SCP framework and the associated notations are introduced. The ASA algorithm is described in Section 3 and the LamSCP is detailed in Section 4. The experimental methodology is presented in Section 5 and results are discussed in Section 6. Before concluding in Section 8, we present experiments in the context of TTS where we evaluate on that task the benefits of a reduction in section 7.","2 the set-covering problem :Before describing the SCP-solving algorithms proposed in this article, we introduce in this section some notations and the Lagrangian properties used by LamSCP.
Let us consider a corpus A composed of n sentences s1, . . . , sn. According to the target applications, these sentences are annotated with respect to phonological, acoustic, prosodic attributes, and so forth. Each sentence is then associated with a family of units of different types. The set of units present in A is denoted U = {u1, . . . , um} and A can be represented by a matrix A = (aij), where aij is the number of instances of unit ui in the sentence sj. Therefore, the j th column of A corresponds to sentence sj in A. To simplify the writing, we define the sets M = {1, . . . , m} and N = {1, . . . , n}.
For a given vector of integers B = (b1, . . . , bm)T, a reduction X of A, also called covering of U , is defined as a subset of A which contains, for every i ∈ M, at least bi instances of ui. It can be described by a vector X = (x1, . . . , xn)T where xj = 1 if sj belongs to X and xj = 0 otherwise. In other words, a covering is a solution X ∈ {0, 1}n of the following system
∀i ∈ M, ∑ j∈N aijxj ≥ bi (1)
that is, AX ≥ B where B is called the constraint vector. Our aim is to optimize a covering according to a cost function minimization criterion. The covering cost is given by summing the costs of the sentences that compose the covering. The optimization problem can be formulated as the following SCP:
X∗ = arg min X∈{0,1}n
AX≥B
CX (2)
where C = (c1, . . . , cn) is the cost vector and cj the cost of the sentence sj. Because of the objective to minimize the total length of the covering, we have chosen to define the cost of a sentence as one of its length features. According to the considered application, the sentence cost can be defined as its number of phones (one of our objectives is to design a phonetically rich script with a minimal speech recording duration), or its number of words, part-of-speech tags, breath groups, and so on.
In Caprara, Fischetti, and Toth (1999), Caprara, Toth, and Fischetti (2000), and Ceria, Nobili, and Sassano (1998), the studied crew scheduling problem is a particular case of Equation (2) where A is a binary matrix and B = 1Rm (i.e., with mono-representation constraints). In order to ensure that Equation (1) admits a solution, we assume that, for each i ∈ M, the minimal number bi of ui instances required in the covering is not greater than the number (A1Rn )i of ui instances in A, that is A1Rn ≥ B. Under this assumption, A is the maximal size solution of Equation (1), represented by X = 1Rn. In the case where bi is greater than the number of ui instances in A, bi is set to (A1Rn )i.
To drive the SCP algorithms during the sentence selection phase, the covering capacity µj of sentence sj is defined as the number of its unit instances required in the covering in view of the constraint vector:
µj = ∑ i∈M min { aij, bi }
(3)
Let us notice that µj does not consider the excess unit instances: For example, if sj contains aij = 10 instances of ui and at least bi = 3 instances of ui are required, the contribution of ui to µj derivation only takes into account three instances of ui.","3 greedy algorithm asa :In this section, the two main steps that compose the algorithm ASA are briefly described. First, an agglomerative greedy procedure is applied to A so as to derive a covering. Next, a spitting greedy procedure reduces this covering in order to approach the optimal solution of Equation (2). The greedy strategy builds a sub-optimal solution to the SCP Equation (2) in an iterative way. At each iteration, the lowest cost sentence is chosen from A. If several sentences correspond to the lowest cost, the one coming first (i.e., the one with the lowest index) is chosen. Initially, the set of selected sentences X is empty, the matrix à associated with the candidate sentences is assigned to A, the current covering capacity of sj is given by µ̃j = µj, and the current constraint vector B̃ = B. The cost of sentence sj is defined by
σj = { cj/µ̃j if µ̃j ̸= 0 ∞ otherwise (4)
Indeed, if µ̃j = 0, it turns out that sj does not cover any unit missing in the solution X under construction and its infinite cost σj avoids its selection.
At each iteration, the selected sentence s is added to X . Taking into account the content of s, B̃ is updated to max { B̃ − Ã∆, 0Rm } where the jth entry of ∆ equals 1 if sj = s and 0 otherwise. Next, the associated column of s in à is set to 0Rm . For each sentence sj with a non-zero µ̃j feature, µ̃j is then updated using à and B̃ in Equation (3).
At last, the agglomerative greedy algorithm is stopped as soon as all the constraints are satisfied, that is, B̃ = 0Rm . The spitting greedy strategy also consists in building iteratively a sub-optimal solution Y to Equation (2) by reducing the size of a covering. The initial covering Y is set to the solution X derived by the agglomerative phase described earlier. At each iteration, the set of the redundant sentences of Y is calculated and the costliest one (according to the cost function C) is removed from Y . An element s of Y is said to be redundant if for each ui ∈ U , its number of instances into Y , denoted mi(Y ), and into Y \ {s}, denoted mi(Y \ {s}), check min {mi(Y ), bi} = min {mi(Y \ {s}), bi}. In other words, s is a redundant element of the covering Y if Y \ {s} is also a covering solution of Equation (1). The spitting greedy algorithm stops when the redundant sentence set is empty.","4 lagrangian relaxation based–algorithm :This section describes the main phases of the algorithm called LamSCP. This algorithm takes advantage of the Lagrangian relaxation properties reviewed herein in order to approach the optimal solution of Equation (2) as close as possible. Strongly inspired by
Caprara, Fischetti, and Toth (1999), but generalized to the multi-representation problem, this algorithm provides a lower bound of the optimal solution cost. Having such information is very useful for assessing the achievements of the SCP algorithms. Let us briefly recall the main principles of Lagrangian relaxation on which LamSCP is based to solve Equation (2) (see Fisher [1981] for more details on Lagrangian relaxation).
First, the Lagrangian function associated with Equation (2) is defined by
L(X,Λ) = CX +ΛT(B − AX) = ΛTB + C(Λ)X
(5)
where Λ ∈ ( R+ )m, X ∈ {0, 1}n, and C(Λ) = C − ΛTA. The coordinates of Λ = (λ1, . . . , λm)T are called Lagrangian multipliers and can be interpreted as a weighting of constraints (1). The jth entry of C(Λ), called Lagrangian cost cj(Λ) of sentence sj, takes into account its cost cj and the adequacy of its composition to address Equation (2). For every covering X and every Λ ∈ ( R+
)m, the Lagrangian function satisfies L(X,Λ) ≤ CX. Thus, the dual Lagrangian function defined by
L(Λ) = min X∈{0,1}n L(X,Λ) (6)
presents the following fundamental property: For every Λ ∈ Rm+ and every covering X, we have L(Λ) ≤ CX. Hence, L(Λ) is a lower bound of the minimal covering cost, CX∗, but does not necessarily correspond to the cost of a covering. In order to compute L(Λ), an acceptable solution for the vector X minimizing L(X,Λ) is X(Λ) = (x1(Λ), . . . , xn(Λ)) T where
xj(Λ) = {
1 if cj(Λ) < 0 0 if cj(Λ) > 0
∈ {0, 1} otherwise (7)
Additionally, the dual Lagrangian function and the Lagrangian costs inform about the potential usefulness of sentences in the optimal covering. More precisely, for a given Λ and a known upper bound UB of minimal covering cost, the gap g(Λ) = UB − L(Λ) measures the relaxation quality. If cj(Λ) is strictly greater than g(Λ), we can check that any covering containing sj has a cost value strictly greater than UB. Hence, sentence sj is not selected and xj can be fixed at zero. Similarly, if cj(Λ) < −g(Λ), any covering with a cost lower than UB contains sj and one can fix xj to 1. Therefore, an optimal covering is made up of sentences with a low Lagrangian cost, as done in Caprara, Toth, and Fischetti (2000) and Ceria, Nobili, and Sassano (1998), and the higher the relaxation quality (i.e., the lower g(Λ)) is, the cheaper the covering will be.
The resolution of the dual problem of Equation (2) consists in finding Λ∗ ∈ Rm+ that maximizes the lower bound L(Λ), that is
Λ∗ = arg max Λ∈Rm+ L(Λ) (8)
Because this real variable function L is concave and piecewise affine, a well-known approach for finding a near-optimal multiplier vector is the subgradient algorithm. The LamSCP is an iterative algorithm, composed of several procedures that aim to either improve the current best solution or reduce the combinatorial issue related to the considered problem. In order to derive a good solution, the algorithm calls on a great number of greedy procedures to solve sub-problems with the help of the Lagrangian costs. As for the combinatorial reduction, the most frequently used heuristic consists of downsizing the problem by mainly considering the sentences with low Lagrangian costs.
The algorithm is organized around a main procedure called 3-phases. This procedure can single-handedly solve a multi-represented SCP. As its name suggests, the 3-phases functioning consists in iterating a sequence of the three following subprocedures as shown in Figure 1:r The subgradient phase calculates an estimation Λ̃ of Λ∗ that maximizes
the dual Lagrangian function. This procedure requires an upper bound UB of the optimal covering cost. UB is initialized by a greedy algorithm (rather than the cost of the whole corpus A). This phase is detailed in Section 4.2.1.r The heuristic phase explores the neighborhood of Λ̃ by generating a great number of Lagrangian vectors Λ̃p. A greedy-like procedure is associated with each Λ̃p so as to compute a covering using the Lagrangian cost vector C(Λ̃p). If, during this exploration, a less costly covering than the best known one (corresponding to the cost UB) is found, the upper bound UB is then updated to the cost of this less costly solution. Similarly, if a better estimation of Λ∗ is obtained, Λ̃ is updated. This phase is described in Section 4.2.2.r The column fixing phase selects a set F of sentences that are most likely to belong to the optimal covering. This phase is detailed in Section 4.2.3.
Following the column fixing phase, the constraint vector is updated and the unselected sentences define a set-covering sub-problem, called a residual problem. This sub-problem is processed similarly, via an additional iteration of the three phases. This iterative process is stopped when the residual problem is empty or when the associated dual Lagrangian function indicates a cost is too high. Indeed, because this function indicates a minimal cost for covering the sub-problem, its addition to the cost CF of the sentences already retained in F gives a lower bound of the total cost of the solution under construction, which should not rise beyond the cost UB of the best known solution so as to be potentially more advantageous.
4.2.1 Subgradient Phase. In order to reach the quality goal, the subgradient phase provides a near-optimal solution Λ̃ of the dual Lagrangian problem (8) using a subgradient
type algorithm. This iterative approach generates a sequence (Λp) using the following updating formula (see Caprara, Fischetti, and Toth 1999)
Λp+1 = max { Λp + µ
g(Λp) ||S(Λp)||2
S(Λp), 0 }
(9)
where S(Λp) = B − AX(Λp) so as to take into account the multi-representation constraints. Parameter µ is adjusted to fit the convergence fastness according to the method proposed by Caprara, Fischetti, and Toth (1999).
At the first call of 3-phases, Λ0 is arbitrarily defined as follows: for each i ∈ M,
λ0i = minj∈N aij ̸=0
cj µj
(10)
As for UB, its initial value is set to the cost of a covering previously calculated. In order to evaluate how much the covering cost derived by ASA can be improved, we have chosen to initialize UB by this value. At the following iterations of 3-phases, Λ0 is given by a random perturbation (less than 10%) of the best known vector Λ̃ (of which the entries of the sentences fixed in the last column fixing phase are removed) and UB corresponds to the cost of the best covering found (after subtraction of the cost of the sentences fixed in the last column fixing phase). In another approach proposed by Ceria, Nobili, and Sassano (1998), UB corresponds to the upper bound of a dual Lagrangian problem, and the subgradient procedure simultaneously estimates the upper bound and the lower bound, generating two sequences of multipliers.
The subgradient phase also calls two procedures: pricing and spitting. Procedure pricing aims to reduce the size of the search space. For each unit ui, the pricing selects the 5bi smallest Lagrangian cost sentences covering ui. If this selection contains less than 5m sentences, where m is the maximal entry of B, it is completed by low Lagrangian cost sentences (less than 0.1) to the limit of 5m sentences. The set of the chosen sentences is denoted P and its design guarantees a sufficient number of instances for each unit to cover and a small variety in its composition. Actually, the subgradient method is applied on P instead of A, and P is updated every 10 subgradient iterations.
Finally, at each iteration, definition (7) of X(Λp) and the large number of Lagrangian costs close to zero allow a considerable number of vectors S(Λp). In order to get around the computational difficulty of finding the steepest accent direction, Caprara, Fischetti, and Toth (1999) propose a heuristic that, according to the experimental results, accelerates the convergence of the subgradient phase. This heuristic is implemented in the spitting procedure. Called at each iteration, this procedure extracts from P the subset S of sentences with a Lagrangian cost lower than 0.001. S is then reduced using a spitting greedy algorithm to remove its redundant elements in decreasing Lagrangian cost order. At last, for every j ∈ M, xj(Λp) = 1 if sj ∈ S and xj(Λp) = 0 otherwise. Thus, S(Λp) does not necessarily correspond to a subgradient vector.
4.2.2 Heuristic Phase. The heuristic phase calculates a large number of coverings before keeping the best one. To that end, a sequence of 150 multiplier vectors is generated by perturbing Λ̃ using the formula Λ̃p+1 = max { Λ̃p + µg(Λ̃p)S(Λ̃p), 0 } where Λ̃0 = Λ̃ and µ is provided by the subgradient phase, so as to allow for a change in a large number of Λ̃p. With each Λ̃p, an agglomerative greedy algorithm followed by a spitting greedy one are associated in order to calculate a covering.
The agglomerative greedy chooses at each iteration the sentence sj with the lowest cost σj(Λ̃p) where
σj(Λ̃p) =  cj(Λ̃p) ∗ µ̃j if cj(Λ̃p) < 0 and µ̃j > 0 cj(Λ̃p)/µ̃j if cj(Λ̃p) ≥ 0 and µ̃j > 0 ∞ if µ̃j = 0
(11)
This cost function provides an advantage to low Lagrangian cost sentences sj containing µ̃j unit instances that could be helpful to the covering under construction. The agglomerative step uses several heuristics so as to reduce the search space. It is run within a limited subset Pl of P , composed of the sentences sj with the lowest costs σj(Λ̃k). At each iteration, a sentence of Pl is selected and the cost of the sentences of P are updated. If the maximum sentence cost in Pl becomes greater than the minimal cost in P \ Pl, the working subset Pl is also updated. The definitions of P and Pl guarantee that the agglomeration step provides a nonpartial solution of the considered SCP. This solution is then reduced during the spitting step by iteratively removing its redundant sentences sj with the highest costs cj.
At the end of the heuristic phase, the best found covering X ∗ and its cost CX∗ (stored in UB) are kept as well as the highest value of L(Λ̃p) (found during the subgradient or heuristic phases).
4.2.3 Column Fixing Phase. The column fixing phase aims to reduce the problem size by choosing “promising” sentences among the ones with very low Lagrangian cost or containing rare unit instances. The unselected sentences are less interesting for resolving the SCP and the residual problem associated with these sentences should be the subject of another call of the 3-phases.
More precisely, the column fixing phase calculates the subset Q composed of sentences sj with a negative Lagrangian cost cj(Λ̃). Q is represented by the binary vector Q = (q1, . . . , qn)T where qj = 1 if sj ∈ Q. For each ui ∈ U , the number of its instances covered by Q is given by (AQ)i. If (AQ)i ≤ bi, then ui is considered as a rare unit and all the elements of Q containing some instances of ui are fixed in a set F . The covering constraints that are not satisfied by F constitute a residual SCP. In order to
complete F with a few sentences of the best known covering, a greedy-type algorithm is run on X ∗ \ F to derive a solution of this residual SCP. From the obtained solution, the max{ ( BT1Rm ) /20, 1} lowest Lagrangian cost sentences are also added in F . The sentences that are “fixed” in F during the column fixing phase stay in F up to the end of the 3-phases. After the column fixing phase, the residual sub-problem is processed by iterating the three phases and the next fixed sentences are added to F . The 3-phases procedure is encapsulated in an outer loop that permits the partial reconsideration of the solution X ∗ provided by this procedure. To that end, the refining procedure, proposed in Caprara, Fischetti, and Toth (1999), is used in order to select elements of X ∗ that contribute at least to the gap g(Λ̃). In the case of the SCP with multirepresentation constraints, the definition of the contribution of sj ∈ X ∗ can be adapted as follows:
δj = max{cj(Λ̃), 0}+ ∑ i∈M
aij>0
λ̃i(AX∗ − B)i aij
(AX∗)i (12)
The second term of Equation (12) consists of sharing the contribution λ̃i(AX∗ − B)i of the excess instances of ui in X ∗ according to the distribution of ui instances in X ∗. Therefore, the refining procedure ranks in an increasing order the elements sj of X ∗ according to their δj value, and fixes the first elements in a set G until its given covering rate τG reaches π. τG represents the rate of covering constraints satisfied by G and is defined by
τG = 1 − ∑
i∈M max{bi − (AG)i, 0} BT1Rm
(13)
where G denotes the binary vector corresponding to G. The LamSCP is made up of the main procedures introduced in the previous sections interlinked by the following steps. First, because of the adaptation of the algorithm to the SCP with multi-representation constraints, the entries of matrix A are clipped to the constraint vector in order to simplify the calculations such as the ones of µ1, . . . ,µn. This threshold application implies that the excess instances of each unit ui are taken into account in each sentence sj beyond the number bi, in the derivation of δj.
After the initialization of Λ0 using Equation (10) and the upper bound UB, procedure 3-phases is called and provides a solution X to the complete SCP. The refining function fixes a sentence subset G such that G covers a given rate π of the covering constraints. Parameter π starts at a minimum value πmin = 0.3. The residual SCP is then processed by 3-phases. The π value grows at a rate of 20 percent whenever 3-phases does not improve the solution to the complete SCP. If π is greater or equal to 1, the refining procedure fixes the whole best solution and the residual problem is then empty. On the other hand, π is set to πmin if a better solution is found in order to challenge G and improve this solution. This iterative sequence composed of the 3-phases and refining procedures is carried out until the residual problem is not empty, the gap g(Λ̃) is positive, and the number of iterations has not reached 20.","5 experiments :We propose a twofold comparison of the ASA and LamSCP algorithms: One part is focused on the covering cost for a large SCP, and the other on the stability of the solutions. Moreover, in order to assess the ability and the behavior of both algorithms to process different linguistic data, a first set of experiments deals with phonological attributes (mainly covering co-occurrences of phonemes) and a second set with grammatical labels (mainly covering co-occurrences of POS labels).
Both attribute types, often involved in automatic linguistic processing, were chosen because their distribution consists of few highly frequent events and numerous rare events. On the one hand, for TTS tasks, the phonological type covering is a useful preliminary step of the text corpus design before the recording step. In order to produce the signal corresponding to a requested sentence, the unit selection engine requires at least one instance of each phone (or 2-phone, depending on the concatenation process). Because the recording and the post-recording annotation process are expensive tasks, the recording length of such a corpus has to be as short as possible. On the other hand, in order to train a domain-specific dependency parser, the covering of POS sequences may be useful for increasing the diversity of syntax patterns. Because the dependency annotation is a highly expensive task, the adaptation corpus to annotate needs to be as small as possible, containing characteristic examples of the specific lexical variation rather than following the natural distribution. One can expect that increasing its diversity of POS sequences may lead to more diversity in the syntax trees.
Experiments on covering co-occurrences of phonemes are carried out on two large phonologically annotated text corpora, and consist of covering at least k instances of each phoneme, diphoneme until n-phoneme (i.e., triphoneme if n = 3, diphoneme if n = 2). The cost cj of the sentence sj is given by its number of phones. From this point, this kind of SCP is called a “k-covering of n-phonemes.”
A first corpus, Gutenberg, is composed of texts in English, mainly extracted from novels and short stories. This corpus is the production of the Gutenberg Project, presented by Hart (2003), and has been used by Kominek and Black (2003) to design the speech corpus Arctic. A second corpus, in French, named Le-Monde, is extracted from articles published in the newspaper Le Monde in 1997. Table 1 summarizes the main features of both corpora.
The phonological annotation of the Gutenberg corpus comes from the Arctic/ Festvox database (see Kominek and Black 2003), and the annotation of the Le-Monde corpus is a by-product of the Neologos project, detailed by Krstulović et al. (2006).
For each corpus, we have collected every phoneme, diphoneme, triphoneme, and their occurrences in each sentence so as to define the set U of units to cover and the matrix A. A is built by collecting one sentence after the other following the ordering inside the corpus, and one unit after the other inside the sentences. After this matrix translation, we obtain two description files and two index files. The first file describes the matrix A and the second one the cost vector C. Because of the low matrix density, we have chosen a sparse representation to save space and computation time: For instance, the 2-phoneme Gutenberg matrix is about 2.2% dense. We only store the cells of A that have a non-zero value so as to get a sparse matrix. The index files are made for the correspondence between the general covering problem and the application domain. The implementation is made in C. In terms of software engineering, our algorithms are working on an SCP that does not depend on the application data. For example, there is no information on what types of units are to be covered. The algorithms only have the matrix of occurrences A, the cost vector C, and the constraint vector B. A set of translation files (from application data to SCP and from SCP to application data) is built before each computation. As a consequence, there is no difficulty in addressing a different set of features to cover on the same or on a different corpus.
To study the achievements of ASA and LamSCP on different types of data, we have also chosen to address the “k-covering of n-POS” on the corpus Le-Monde. The grammatical and syntactical analyses are processed by the Synapse development analyzer presented in Synapse (2011). In order to consider a SCP with a substantial number of required units, a very detailed level of POS tagging has been selected, providing 141 distinct tags after analyzing Le-Monde. For example, this level provides tags like “Determiner male singular Article” or “Noun female singular,” whereas the simplest level gives “Determiner” or “Noun.” This latter level of description would have given only nine different POS tags after analyzing Le-Monde. The main associated statistics are given in Table 2. For these experiments of POS covering, the cost of a sentence is defined as its number of POS occurrences.
We used a PC with 2 CPUs (E5320/1.86GHz/4 cores/64bits) and 32 GB RAM for the phonological coverings and the POS coverings were computed using a PC with 8 CPUs (Intel Xeon X7550/2.00Ghz/8 cores/64bits) and 128 GB RAM. Our implementations do not take advantage of any parallelism.
The following sections detail more precisely the different experiments conducted on French or English. The aim of Experiment 1 is the assessment of the achievements of both algorithms, ASA and LamSCP, and the robustness of the results when the sentence ordering is modified
in the corpus to reduce. Indeed, one of the difficulties of the greedy methodology is that the score function has discrete values and several sentences can yield the same score. In our implementation, among the sentences showing the best current score, the first one encountered is chosen. We would like to measure the influence of this random choice on the stability of the results. LamSCP uses greedy strategies based on Lagrangian costs. Because the Lagrangian costs cj(Λ) take the SCP in its entirety into account and are continuous real-value functions of Λ, they would be more selective than the sentence costs used by ASA. A simple solution for evaluating the stability consists of proceeding with an important amount of experiments on the same SCP by randomly modifying the sentence ranking in A. Experiment 1 measures the impact of these permutations on the solutions computed by both algorithms. The considered SCP is the 1-covering of 2-phonemes on the corpus Le-Monde. Considering the computation time (more than 5 hours for LamSCP), only 47 instances of the SCP are considered, each instance corresponding to a random sentence ordering in Le-Monde. The 95% confidence intervals are derived using the bootstrap method, concerning the covering cost, the number and the length of sentences in coverings, the computation time, and the “distance” between the covering costs and the associated lower bound L(Λ̃).
5.2 Stability of the Algorithms for the k-Covering of 2-Phonemes in English
One of the goals of Experiment 2 is to compare the achievements of both algorithms on corpus Gutenberg, which has different features from the ones of Le-Monde. The sentences in Gutenberg are shorter on average and the associated variation of sentence length is lower. Furthermore, Gutenberg is 10 times smaller than Le-Monde.
In order to compare with the results of the previous experiment done on Le-Monde, we first observed a 1-covering of 2-phonemes on the Gutenberg corpus. The search space seems smaller than in Experiment 1. According to Table 1, Le-Monde is composed of more sentences than Gutenberg. Le-Monde contains 33,165,050 occurrences of 2-phonemes and Gutenberg only 3,025,474. Moreover, the number of attributes to cover is lower: 1,207 2-phonemes in Le-Monde and 2,012 in Gutenberg. We can also notice that five 2-phonemes have only one occurrence in Le-Monde and that the total cost of the five sentences covering these rare units is 751 phones, whereas this is the case for 109 2-phonemes in Gutenberg and the total cost of the 104 concerned sentences is 3,606 phones. Finally, if we consider the density of matrix A, 8.4% of the cells are non-empty for Experiment 1 and 2.2% for Experiment 2. As with Experiment 1, the sentence ordering in Gutenberg has been randomly modified to produce 60 instances of the SCP and similar solution statistics have been computed for both algorithms.
The second objective is to test and compare the ability of two algorithms to deal with the constraints of multi-representation. For this, we apply the same methodology to the k-covering of 2-phonemes in Gutenberg, for k from 2 to 5. We note that for the same original corpus to reduce, the size of the search space decreases when k increases. These different SCPs enable us to compare the performance of two algorithms depending on the size of the search space. In Experiment 3, the aim is to observe the behavior of both algorithms on very constrained problems. For this, we study their ability to treat a covering of 3-phonemes. We try to assess the impact on the solution features and on the stability of such an increase in the number of attributes to cover with many rare events. So as to compute statistics
on the 1-covering of 3-phonemes, an instance of Gutenberg has been proposed to ASA and to LamSCP. This instance counts 29,489 units to cover and the density of matrix A is 0.24%. The computation time is nearly 5 days for LamSCP and we then chose to carry out 35 instances of the 1-covering of 3-phonemes on Gutenberg. Additionally, it is interesting to compare these results with those of Experiment 2 concerning the 1-covering of 2-phonemes on the same corpus, which corresponds to a larger search space. In order to pursue the objective set out in the description of Experiment 3, that is, the ability of algorithms to treat with numerous constraints and a heavy-tail distribution of units, Experiment 4 consisted of testing both algorithms on the 1-covering of 3- phonemes on Le-Monde. The search space seems larger than in the previous experiment. Let us recall that Le-Monde contains 3.18 times more sentences than Gutenberg and it counts 27,650 units to cover. Furthermore, Gutenberg contains 5,000 3-phonemes with only one occurrence, which requires the selection of 4,180 sentences with a total length equal to 137,714 phones, whereas Le-Monde contains 2,274 rare 3-phonemes scattered in 2,107 sentences measuring a total of 283,208 phones. The associated matrix density is 0.69%. Because the computation of a first instance takes more than 8 days, we have limited the number of instances to 30 for this SCP.
5.5 k-Covering of 1-POS and 2-POS in French
The main goal of Experiment 5 is to study the behavior of both algorithms, ASA and LamSCP, dealing with another kind of linguistic attribute, and to compare this with the previous experiments. To achieve this goal, we consider POS attributes and the associated SCP: 1- and 5-coverings of POS, 1- and 5-coverings of 2-POS defined on Le-Monde. Indeed, we can observe in Table 2 that the global statistics of POS tags in Le-Monde are quite different from their phonological counterparts summarized in Table 1. In particular, the density of matrix A is 11.03% for a 1-POS covering and 0.57% for a 2-POS covering. Also, the search space size seems to decrease when considering successively the mono- and the multi-coverings of 1-POS, and the mono- and the multicoverings of 2-POS, permitting us to compare them with the results coming from the experiments on phonological coverings. We evaluate the stability by computing 50 randomly mixed versions of the corpus Le-Monde.","6 results and discussion :In this section, the results of the experiments described in Section 5 are provided and discussed. As a consequence, the organization of this section and Section 5 are similar. Table 3 shows the main results of Experiment 1, concerning the 1-covering of 2- phonemes from the corpus Le-Monde. Symbol ± indicates that the mentioned value corresponds to a 95% confidence interval, calculated using the bootstrap method from
the 47 instances of the SCP. In order to cover each of the 1,207 2-phonemes of Le-Monde, ASA drastically reduces the size of the initial corpus by 99.94% (±0.00). However, on average, LamSCP calculates a 9.00% shorter covering. The lower bound L(Λ̃) for the optimal covering cost is 7,689 ± 5 phones. L(Λ̃) is not a minimum value and may not correspond to the cost of a real covering. Because this lower bound is updated all along the execution of LamSCP, we do not mention a specific calculation time for this result.
For one instance of the SCP, let CX∗ASA be the size of the solution given by ASA. The quantity τASA = 1 − L(Λ̃)/CX∗ASA indicates that the optimum solution to SCP is at most τASA times shorter than the covering calculated by ASA. It can be observed that the optimal solution is at most 10.13% (±0.19) shorter than the one yielded by ASA and at most 1.24% (±0.08) shorter than the solution yielded by LamSCP. The solutions obtained by LamSCP and the optimal solution to the SCP are therefore very close. Considered among the 47 instances of the SCP the best solutions yielded by ASA (8,447 phones) and LamSCP (7,767 phones), LamSCP is 8.75% better than ASA in terms of covering costs, while the best lower bound for the SCP is 7,715 phones, only 0.67% (respectively, 8.66%) shorter than the best covering by LamSCP (respectively, ASA).
The average length of the sentences selected by both algorithms is far below the average length of the sentences in the corpus (96.81 phones). LamSCP tends to choose sentences that are slightly longer than ASA, with an average 28.97 (±0.12) phones compared with 25.48 (±0.10) phones. Moreover, ASA selects on average 335.73 (±1.69) sentences per solution, about 24.91% more than LamSCP, which selects 268.76 (±1.08) sentences on average. This seems to indicate that LamSCP makes fewer local choices than ASA. This hypothesis can also be validated through the analysis of the variability of the results. The relative variation of the covering costs calculated by LamSCP is 13.01/7,786 = 0.16%, and 57.40/8,555 = 0.67% by ASA; that is to say a stability of the costs 4 times greater for the solutions yielded by LamSCP than for ASA. Moreover, the solutions are composed of a very stable number of sentences: The associated relative standard deviation is 5.31/335.73 = 1.58% for the 47 instances solved by ASA, and 3.85/268.76 = 1.43% for the instances solved by LamSCP. It turns out that the results of both algorithms are very stable when the order of the sentences is modified in the original corpus.
Finally, concerning computation time, the resolution of an instance of the SCP lasts on average 5 hr 41 min 18 sec (±8 min 28 sec) for LamSCP versus 51 sec
(±0 sec) for ASA. On average over the 47 instances, LamSCP takes 390 (±9) times as long as ASA.
6.2 Stability of the Algorithms for k-Covering of 2-Phonemes in English
The considered SCP consists of covering at least k times each of the 2,012 2-phonemes of the Gutenberg corpus, with k varying from 1 to 5. The results are summarized in Table 4.
For all instances of these SCPs, it has been observed that LamSCP computes shorter coverings than ASA. However, that advantage diminishes as k grows: The cost advantage offered by LamSCP compared with ASA decreases from 9.73% (±0.13) for k = 1 to 4.50% (±0.04) for k = 5. Also, the solutions obtained from ASA and LamSCP seem to get closer to the optimal solution as k rises. The corresponding figures are presented in Table 5: For instance, the optimal solution is at most 0.75% (±0.02) shorter than that obtained by LamSCP for k = 1, and 0.27% (±0.00) for k = 5.
Because the search area diminishes as k increases, it may be observed that the algorithms tend to be more stable. This is true both for the size of the solutions, as well as for the number of sentences that define them. Table 6 represents the variation of the size of the solutions as a function of k. This variation is calculated as follows: For a given k number and a given algorithm, the standard deviation of the size of the k-covering computed by that algorithm is divided by the average size of these coverings. Thus, it can be noted that LamSCP offers a stability 4 to 8 times superior to ASA concerning the size of the coverings. As for the number of sentences, the relative standard deviation similarly decreases from 0.97% to 0.28% when k increases from 1 to 5 for ASA solutions, and from 0.42% to 0.15% for LamSCP ones.
One can note that the increase of the minimal number k of instances of each unit to cover leads to a selection, by LamSCP and ASA, of longer sentences on average. The average length of the sentences picked for a 1-covering was quite low. As the constraints increase along with k, it only seems natural that the algorithms tend to select longer sentences, as shorter sentences no longer contain enough occurrences of 2-phonemes. Moreover, as described in Section 2, when the minimal number bi of a unit ui demanded in the covering exceeds the number of instances of that unit in the initial corpus, all sentences containing instances of ui in the initial corpus are selected, and bi is set to (A1Rn )i. Thus, as k increases, the algorithm tends to select more and more sentences, and their length tends towards the average value over the whole corpus, which is 28.51 phones for Gutenberg.
As for computation time, although it increases as k grows, because of the increasing number of constraints to update, the ratio between the computation time of LamSCP and ASA tends to diminish, as shown in Table 7. This tendency may find an explanation in the fact that the search space diminishes as k increases, which causes a lesser number of selected sentences to be questioned during the 3-phase iteration of LamSCP. Also, we notice that the average computation time of the two algorithms is greater in Experiment 1, owing to a greater number of sentences in corpus Le-Monde and a higher density of matrix A. Moreover, the ratio between the computation times of
Time LamSCP / Time ASA 333 (± 7) 280 (± 5) 292 (± 6) 218 (± 9) 194 (± 8)
LamSCP and ASA decreases between Experiments 1 and 2, going from 390 (±9) to 333 (±7). Again, this can be explained by the diminishing of the search space.
For k = 1, the advantage offered by LamSCP on the covering costs compared with ASA is slightly higher than that observed in Experiment 1: 9.73% (±0.13) in this case, versus 9.00% (±0.20) in the previous experiment. This seems to contradict the idea that the performance of LamSCP improves as the search area becomes wider. However, the distributions of the units to cover in Gutenberg and Le-Monde are different, and the variation on the length of the sentences in Le-Monde is very high, which may account for this slight difference in terms of gain. Note that the size of the calculated coverings and the lower value L(Λ̃) are closer in the experiment carried out on Gutenberg. It is difficult, however, to perform further comparisons with Experiment 1 regarding the “distance“ between the costs of the solutions computed by these algorithms, and the optimal covering cost, given that the quality of the lower bound cannot be evaluated. The gain in stability offered by LamSCP, both for the costs of the solutions or the number of sentences, is more important than that noticed during the previous experiment. We think that the increase is due to a more restricted search space, and less variability of the length of the sentences in corpus Gutenberg, which may be observed in Table 1. Table 8 sums up the main results of Experiment 3, where 35 instances of 1-covering of 3-phonemes from Gutenberg were processed. According to the L(Λ̃) values, covering all 3-phonemes requires a solution size greater than or equal to 226,635 phones. On average, the solution measures 227,360 ± 12 phones using LamSCP, and 236,828 ± 94
phones using ASA. The optimal covering is at most 0.35% (±0.00) shorter than solutions derived by LamSCP and 4.33% (±0.04) shorter than the ones derived by ASA. We can then observe that both algorithms manage to compute solutions with close sizes when scaling up the required attribute set. The solutions are very stable, even more than in Experiment 2: The relative variation of their size is 0.12% for ASA and 0.01% for LamSCP; the relative variation of their sentence number is 0.17% for ASA and 0.07% for LamSCP. This increase of stability is due to a smaller search space and the increase of the number of rare units required, which also compels the algorithms to select a higher number of inevitable sentences for all the instances of the SCP. Furthermore, the decrease of the ratio between the computation time of LamSCP and ASA from 332 for the 1-covering of 2-phonemes to 130 for the one of 3-phonemes on Gutenberg may confirm this idea, which has also been put forward in Experiment 2.
Concerning the length of the selected sentences by both algorithms, it is greater than the one for the 5-covering of 2-phonemes, observed in Experiment 2, and slightly deviates from the average sentence length for the whole corpus. Consequently, it turns out that covering longer and generally rarer units involves a selection of longer sentences. This is confirmed by the fact that the sentences of Gutenberg covering units with a single occurrence in Gutenberg represent more than half the size of the solutions and are composed of 33 phones on average. In this section, we analyze the results of Experiment 4, the 30 instances of the 1- covering of 3-phonemes from Le-Monde carried out by ASA and LamSCP. The results are given in Table 9. First, although the main features of Le-Monde and Gutenberg are different, notice that the closeness between the size of coverings calculated by both algorithms and the lower bound L(Λ̃) is comparable to the one observed in Experiment 3. Indeed, the optimal covering size is at most 0.48% (±0.03) and 4.35% (±0.05) shorter than the solution size derived by LamSCP and ASA, respectively. Similarly, the size of solutions and the number of selected sentences are as stable as those observed in the previous experiment: The solution length varies from 0.01% for LamSCP to 0.10% for ASA and the number of sentences fluctuates about 0.16% for ASA and 0.10% for LamSCP.
As for the comparison with the results of Experiment 1 (1-covering of 2-phonemes from Le-Monde), the main trends are similar to the ones observed for the transition from the 1-covering of 2-phonemes to the 1-covering of 3-phonemes from Gutenberg.
However, in Experiment 4, the average selected sentence length has markedly increased, approaching the mean value on the whole corpus: 89.02 (±0.03) for ASA and 92.64 (±0.03) for LamSCP, whereas in Experiment 1 these values are, respectively, 25.48 (±0.10) and 28.97 (±0.12). We have already observed in Experiment 3 that covering longer units increases the length of selected sentences but this high amplitude seems to be inherent to the design of corpus Le-Monde. Furthermore, notice that the 1-coverings of 2-phonemes from Le-Monde are almost half as small as the ones from Gutenberg, whereas the 1-coverings of 3-phonemes from Le-Monde are between twice and three times longer than the ones from Gutenberg. This is due to the fact that the 3-phonemes with a single instance in Le-Monde are very scattered in long sentences (their length mean is about 134 phones), and these indispensable sentences represent nearly half the size of the solutions. The other sentences of the solutions are around 70 phones long. Lastly, the ratio between the computation time of both algorithms is about 84, which is smaller than the ratios previously observed, but this SCP is the most time consuming: 2 hr 10 min for ASA and more than 7 days for LamSCP.
6.5 k-Covering of 1-POS and 2-POS in French
Table 10 sums up the main results of Experiment 5, dealing with the 1- and 5-coverings of 1-POS and 2-POS. For all these SCP, LamSCP produces smaller coverings, composed of longer sentences, than the coverings obtained with ASA. When the search space diminishes, the relative “distance” between the size of solutions provided by both algorithms decreases, as well as between the lower bound L(Λ̃) and the size of solutions obtained by ASA. These trends were also observed in the earlier experiments. In particular, as for the 1-covering of 1-POS, not only does LamSCP provide 10.06% (± 0.00) shorter solutions than ASA, but its solutions are optimal for all 50 instances of this SCP. Indeed, the lower bound value varies from 482.51 to 482.87 occurrences of 1-POS while all the solutions given by LamSCP are made of exactly 483 occurrences of POS. For the other k-coverings of n-POS, the optimal solution is at worst 0.39% (±0.02) shorter than the covering given by LamSCP for (k, n) = (5, 1), 0.11% (±0.00) for (k, n) = (1, 2), and 0.22% (±0.00) for (k, n) = (5, 2). The solutions obtained by ASA or by LamSCP are very stable. For example, the relative standard deviation of number of POS in a covering solution varies from 0.00% to 1.23% for ASA, and from 0.00% to 0.04% for LamSCP.
As previously observed for both algorithms, their computation times grow when the number of required covering features increases. However, the ratio between the computation time of LamSCP and ASA does not behave as in Experiment 2 (see Table 7): For the k-covering of 1-POS, this ratio increases from 290 (±11) to 657 (± 43) when k goes from 1 to 5, and for the k-covering of 2-POS, it increases from 75 (±5) to 108 (±6).","7 evaluation on a text-to-speech synthesis system :In the previous sections, different algorithms dealing with corpus reduction were introduced and studied. The proposed experiments mainly evaluate the effects of these algorithms in terms of corpus reduction but not according to a practical task. This section proposes an experiment to assess the impact of the corpus reduction on a unit selection speech synthesis system.
As explained in Section 1, a corpus reduction for a TTS system is a trade-off between minimizing the recording and post-processing time to build the speech corpus and
keeping the highest phonological richness of the corpus to ensure the quality of the synthetic speech. The goal of this experiment is to measure this trade-off by evaluating the quality of the same TTS system fed with different speech corpora uttered by the same speaker. Note that the intrinsic quality of this system is not the purpose here.
Firstly, a brief presentation of a state-of-the-art unit selection–based TTS system is proposed in Section 7.1. The linguistic parameters used by the TTS system are detailed because they are linked to the required features in the reduction stage. In Section 7.2, corpora used in the experiment are introduced. The attributes to cover and the
methodology of evaluation are described in Section 7.3; the results are given and discussed in Section 7.4. For this experiment, a state-of-the-art unit selection–based TTS system is used to produce an acoustic signal from an input text. A linguistic front end processes the text to extract features taken into account by the algorithm that selects segments in a speech corpus (see Boëffard and d’Alessandro 2012).
The input text is converted into a sequence of phonemes using a French phonetizer proposed by Béchet (2001). Non-speech sound labels can be added to this sequence (silences, breaths, para-verbal events, etc.). A vector of features is defined as follows:
1. The phone or non-speech sound label
2. Is the described segment a non-speech sound?
3. Is the phone in the onset of the syllable?
4. Is the phone in the coda of the syllable?
5. Is the phone in the last syllable of its syntagm?
6. Is the current syllable at the end of a word?
7. Is the current syllable at the beginning of a word?
Extraction of features is done using the ROOTS toolkit described in Boëffard et al. (2012). The unit selection process aims to associate a signal segment from the speech corpus to each vector of features computed from the input text. This is performed in two steps. In the first step, for each unit, a set of candidates that match the same features are extracted from the speech corpus. In the second step, given all candidates, the best path is searched using an optimization algorithm so as to produce the sequence of speech units. The algorithm tries to minimize three sub-costs, commonly used in unit selection based TTS systems, which are spectral discrepancies based on MFCC distance, amplitude, and f0 distances. Two corpora are used in this experiment. The first one, Learning corpus, is an annotated acoustic corpus used to provide speech data for the TTS engine. It is an expressive corpus in French, spoken by a male speaker reading Albertine disparue, an excerpt from À la recherche du temps perdu by Marcel Proust. The corpus is composed of 3,138 sentences automatically annotated using a process described in Boëffard et al. (2012). The overall length of the speech corpus is 9 hr 57 min. When creating a voice for a unit selection– based TTS system, long sentences are generally removed or split into syntagm groups in order to help the speaker.
A second corpus, named Test corpus, is a text corpus that is synthesized and used in the listening experiment. It is composed of 30 short sentences randomly extracted from a phonetically balanced corpus in French, proposed by Combescure (1981). The use of a corpus with a different linguistic style minimizes the bias introduced by the learning corpus.
Statistics are given in Table 11. For this experiment, two reduced corpora are evaluated. They are built by reducing the full learning corpus using the two different algorithms presented in the previous sections: ASA and LamSCP.
As described in Section 7.1, the unit selection process of the speech synthesis system is based on a set of phonological attributes. It seems natural to try to cover features that reflect the variability of these attributes. For this experiment, algorithms must cover all the units at least once, where a unit is described by the following:
r Its label, that is, one of the 35 phonemes or a non-speech sound labelr The structure of the syllable that contains the phoneme, if it is a vowelr The position of the associated syllable in the word (start, middle, or end)r A Boolean indicated if the associated syllable is at the end of a syntagm The feature extraction is performed by the same set of tools used by the speech synthesis engine. Given this set of features, the learning corpus contains 1,497 classes of units. The cost function to minimize by the reduction algorithms is the total length of the set of the selected syntagms, in phones.
Two speech synthesis systems are defined, extracting the speech units to concatenate from the coverings provided by LamSCP and ASA. Two other systems are added as baselines. First, a system named Full, built with the whole learning corpus, is used as an upper bound. Second, a system named Random uses a random reduction of the Learning corpus as a pool corpus of speech units. This reduction is done by randomly selecting sentences from the whole learning corpus until the size of the covering obtained by LamSCP is reached. Random is used as a lower bound.
Whereas the optimization efficiency is measured by statistics on the reduced corpora, the quality of the synthesized speech signals is evaluated by a listening test. The protocol is based on a MUSHRA test, presented in ITU-R (2003), where for every sentence of Test corpus, the signals synthesized by the four systems are presented to each tester in a random order. If a system is not able to produce a signal for a requested sentence (because of a missing 2-phone in the pool corpus), an empty signal is presented. Ten native French testers (four naives and six experts) are asked to evaluate the overall quality of the stimuli and to give a mark from 0 to 100 (by steps of 5 points). The Learning corpus composed of 19,587 syntagms is first reduced using ASA and LamSCP. Statistics of the resulting solutions are summarized in Table 12. The covering of the 1,497 constraints divides the input corpus size by almost 10 and reduces the 10 hours of speech to around 1 hr 20 min. As for the previous experiments, even though different kinds of features are mixed (phonemes, syllable structures, position in a word or a syntagm) the ASA algorithm produces a solution close to the optimal one. However, LamSCP is again slightly better in terms of covering size.
For the measure of the acoustic impact of corpus reduction, the listening test results are presented in Table 13 for the average marks and in Table 14 for the average ranks. Note that without a natural speech reference during the test, the marks should not be seen as an absolute score. Even if the LamSCP corpus is slightly smaller than the one from ASA, the acoustic quality of both systems is comparable according to the testers (with a slight advantage for the LamSCP corpus). In comparison with the baseline, which uses the whole learning corpus, the acoustic degradation is significant. This illustrates well the trade-off between corpus size and speech quality: For a 90% corpus size reduction, the acoustic quality drops by 10 points. Further research should be focused on the set of attributes to cover and their number of occurrences in order to improve this compromise. As expected, the baseline built from random sentences is preferred significantly less than the other systems because of the lack of relatively rare acoustic units.","8 conclusion :This article discussed the building of linguistic-rich corpora under an objective of parsimony. This task, a generalization of SCP, turns out to be an NP-hard problem that cannot be polynomially approximated. We studied the behavior of several algorithms in the particular domain of NLP, where the considered events follow a heavy-tailed type distribution.
The proposed algorithms have been compared through three kinds of experiments: The first one is the coverings of 2- and 3-phonemes from two text corpora, one in French, the other in English; the second one consists of the coverings of part-of-speech labels from a corpus in French; the third one evaluates the impact of both algorithms on the acoustic quality of a corpus-based TTS system. The first algorithm, ASA, is composed of an agglomerative greedy strategy followed by a spitting greedy stage. The second one, LamSCP, is based on Lagrangian relaxation principles combined with greedy strategies. LamSCP is our adaptation of an algorithm proposed in Caprara, Fischetti, and Toth (1999) to the multi-representation constraints. The comparison of SCP solutions is mainly about their size, their maximal distance with the optimal covering, and their robustness in case of perturbation of the initial corpus ordering.
Although ASA is much faster than LamSCP, it does not permit us to singlehandedly assess the quality of its solution in terms of size. The main assets of LamSCP are the calculation of a lower bound to the optimal covering size and shorter solutions than the ones obtained by ASA. Indeed, in our experiments of phonological coverings, the optimal solution is at most 1.24% (10.13%, respectively) smaller than the solutions derived by LamSCP (ASA, respectively). As for the coverings of 1-POS, LamSCP provides the optimal solution in a case of a mono-representation constraint, whereas the ASA solution is 10.17% greater than the optimal one. These relative gaps between the lower bounds and solution sizes of both algorithms generally decrease when the size of the search space decreases. Thanks to the lower bound derived by LamSCP, we empirically show that it is possible to get almost optimal solutions in a linguistic framework following Zipf’s law distribution, despite the theoretic complexity of the multi-represented SCP. Concerning the last experiment in the TTS framework, even if LamSCP provides a smaller corpus, the subjective test shows no significant difference between the TTS systems based on LamSCP and ASA corpora. Therefore, we think that ASA remains the most adequate strategy, in terms of performance, ease of development, and computation time to solve SCP in the NLP field. However, it would be interesting
to test a parallelized version of the heuristic phase that calls an important number of greedy sub-procedures.
Our future prospects for this work are in automatic language processing and speech synthesis. First, in the framework of the Phorevox project supported by the French National Research Agency, we are considering the automatic design of exercise contents for language learning by the selection of texts covering some phonological or linguistic difficulties. Secondly, this work is a preliminary step to building a phonetically rich script before its recording in order to produce a high quality speech synthesis. The covering choices, such as the attributes to cover, the number of required occurrences, or the “sentence” length (utterances, syntagms, etc.) need to be validated. Moreover, in this article, we have observed the great impact of the distribution of rare units in the corpus to reduce, and we believe it will be interesting to adapt the “sentence” granularity according to this distribution.","Linguistic corpus design is a critical concern for building rich annotated corpora useful in different domains of applications. For example, speech technologies such as ASR (Automatic Speech Recognition) or TTS (Text-to-Speech) need a huge amount of speech data to train datadriven models or to produce synthetic speech. Collecting data is always related to costs (recording speech, verifying annotations, etc.), and as a rule of thumb, the more data you gather, the more costly your application will be. Within this context, we present in this article solutions to reduce the amount of linguistic text content while maintaining a sufficient level of linguistic richness required by a model or an application. This problem can be formalized as a Set Covering Problem (SCP) and we evaluate two algorithmic heuristics applied to design large text corpora in English and French for covering phonological information or POS labels. The first considered algorithm is a standard greedy solution with an agglomerative/spitting strategy and we propose a second algorithm based on Lagrangian relaxation. The latter approach provides a lower bound to the cost of each covering solution. This lower bound can be used as a metric to evaluate the quality of a reduced corpus whatever the algorithm applied. Experiments show that a suboptimal algorithm like a greedy algorithm achieves good results; the cost of its solutions is not so far from the lower bound (about 4.35% for 3-phoneme coverings). Usually, constraints in SCP are binary; we proposed here a generalization where the constraints on each covering feature can be multi-valued.","[{""affiliations"": [], ""name"": ""Nelly Barbot""}, {""affiliations"": [], ""name"": ""Olivier Bo\u00ebffard""}, {""affiliations"": [], ""name"": ""Jonathan Chevelu""}, {""affiliations"": [], ""name"": ""Arnaud Delhay""}]",SP:066100434f650471b7f8100e81aef651fc0bf5f5,"[{""authors"": [""Alon"", ""Noga"", ""Dana Moshkovitz"", ""Shmuel Safra.""], ""title"": ""Algorithmic construction of sets for k-restrictions"", ""venue"": ""ACM Transactions on Algorithms (TALG), 2(2):153\u2013177."", ""year"": 2006}, {""authors"": [""Barbot"", ""Nelly"", ""Olivier Bo\u00ebffard"", ""Arnaud Delhay.""], ""title"": ""Comparing performance of different set-covering strategies for linguistic content optimization in speech corpora"", ""venue"": ""Proceedings of the International"", ""year"": 2012}, {""authors"": [""B\u00e9chet"", ""Fr\u00e9d\u00e9ric.""], ""title"": ""Liaphon: un systeme complet de phon\u00e9tisation de textes"", ""venue"": ""Traitement automatique des langues, 42(1):47\u201367."", ""year"": 2001}, {""authors"": [""Bo\u00ebffard"", ""Olivier"", ""Laure Charonnat"", ""S\u00e9bastien Le Maguer"", ""Damien Lolive"", ""Ga\u00eblle Vidal.""], ""title"": ""Towards fully automatic annotation of audiobooks for TTS"", ""venue"": ""Proceedings of the International Conference on"", ""year"": 2012}, {""authors"": [""Bunnell"", ""H. Timothy.""], ""title"": ""Crafting small databases for unit selection TTS: Effects on intelligibility"", ""venue"": ""Proceedings of the ISCA Tutorial and Research Workshop on Speech Synthesis (SSW7), pages 40\u201344, Kyoto."", ""year"": 2010}, {""authors"": [""Cadic"", ""Didier"", ""C\u00e9dric Boidin"", ""Christophe d\u2019Alessandro""], ""title"": ""Towards optimal TTS corpora"", ""venue"": ""In Proceedings of the International Conference on Language Resources and Evaluation (LREC),"", ""year"": 2010}, {""authors"": [""Candito"", ""Marie"", ""Enrique Henestroza Anguiano"", ""Djam\u00e9 Seddah.""], ""title"": ""A word clustering approach to domain adaptation: Effective parsing of biomedical texts"", ""venue"": ""Proceedings of the 12th International"", ""year"": 2011}, {""authors"": [""Caprara"", ""Alberto"", ""Matteo Fischetti"", ""Paolo Toth.""], ""title"": ""A heuristic method for the set covering problem"", ""venue"": ""Operations Research, 47(5):730\u2013743."", ""year"": 1999}, {""authors"": [""Caprara"", ""Alberto"", ""Paolo Toth"", ""Matteo Fischetti.""], ""title"": ""Algorithms for the set covering problem"", ""venue"": ""Annals of Operations Research, 98(1-4):353\u2013371."", ""year"": 2000}, {""authors"": [""Ceria"", ""Sebasti\u00e1n"", ""Paolo Nobili"", ""Antonio Sassano.""], ""title"": ""A Lagrangian-based heuristic for large-scale set covering problems"", ""venue"": ""Mathematical Programming, 81(2):215\u2013228."", ""year"": 1998}, {""authors"": [""Chevelu"", ""Jonathan"", ""Nelly Barbot"", ""Olivier Bo\u00ebffard"", ""Arnaud Delhay.""], ""title"": ""Lagrangian relaxation for optimal corpus design"", ""venue"": ""Proceedings of the ISCA Tutorial and Research Workshop on Speech Synthesis"", ""year"": 2007}, {""authors"": [""Chevelu"", ""Jonathan"", ""Nelly Barbot"", ""Olivier Bo\u00ebffard"", ""Arnaud Delhay.""], ""title"": ""Comparing set-covering strategies for optimal corpus design"", ""venue"": ""Proceedings of the International Conference on Language"", ""year"": 2008}, {""authors"": [""Combescure"", ""Pierre.""], ""title"": ""20 listes de 10 phrases phon\u00e9tiquement \u00e9quilibr\u00e9es"", ""venue"": ""Revue d\u2019Acoustique, 56:34\u201338."", ""year"": 1981}, {""authors"": [""Fisher"", ""Marshall L.""], ""title"": ""The Lagrangian relaxation method for solving integer programming problems"", ""venue"": ""Management Science, 27(1):1\u201318."", ""year"": 1981}, {""authors"": [""Fran\u00e7ois"", ""H\u00e9l\u00e8ne"", ""Olivier Bo\u00ebffard.""], ""title"": ""Design of an optimal continuous speech database for text-to-speech synthesis considered as a set covering problem"", ""venue"": ""Proceedings of the European Conference on"", ""year"": 2001}, {""authors"": [""Fran\u00e7ois"", ""H\u00e9l\u00e8ne"", ""Olivier Bo\u00ebffard""], ""title"": ""The greedy algorithm and its application"", ""year"": 2002}, {""authors"": [""Gauvain"", ""Jean-Luc"", ""Lori Lamel"", ""Maxine Esk\u00e9nazi.""], ""title"": ""Design considerations and text selection for Bref, a large French readspeech corpus"", ""venue"": ""Proceedings of the International Conference of Spoken Language"", ""year"": 1990}, {""authors"": [""Gotab"", ""Pierre"", ""Fr\u00e9d\u00e9ric B\u00e9chet"", ""G\u00e9raldine Damnati.""], ""title"": ""Active learning for rule-based and corpus-based spoken language understanding models"", ""venue"": ""Proceedings of the IEEE workshop on"", ""year"": 2009}, {""authors"": [""Hart"", ""Michael.""], ""title"": ""Project gutenberg"", ""venue"": ""http://www.gutenberg.org/ (Last consulted April 2015)."", ""year"": 2003}, {""authors"": [""Karp"", ""Richard M.""], ""title"": ""Reducibility among combinatorial problems"", ""venue"": ""Complexity of Computer Computations, the IBM Research Symposia Series. Springer, pages 85\u2013103."", ""year"": 1972}, {""authors"": [""Kawai"", ""Hisashi"", ""Seiichi Yamamoto"", ""Norio Higuchi"", ""Tohru Shimizu.""], ""title"": ""A design method of speech corpus for text-to-speech synthesis taking account of prosody"", ""venue"": ""Proceedings of the"", ""year"": 2000}, {""authors"": [""Kominek"", ""John"", ""Alan W. Black.""], ""title"": ""The CMU Arctic speech databases for speech synthesis research"", ""venue"": ""Technical Report CMU-LTI-03-177, Carnegie Mellon University Language Technologies"", ""year"": 2003}, {""authors"": [""Krstulovi\u0107"", ""Sacha"", ""Fr\u00e9d\u00e9ric Bimbot"", ""Olivier Bo\u00ebffard"", ""Delphine Charlet"", ""Dominique Fohr"", ""Odile Mella""], ""title"": ""Optimizing the coverage of a speech database through a selection of representative speaker"", ""year"": 2006}, {""authors"": [""Krul"", ""Aleksandra"", ""G\u00e9raldine Damnati"", ""Fran\u00e7ois Yvon"", ""C\u00e9dric Boidin"", ""Thierry Moudenc""], ""title"": ""Adaptive database reduction for domain specific speech"", ""year"": 2007}, {""authors"": [""Krul"", ""Aleksandra"", ""G\u00e9raldine Damnati"", ""Fran\u00e7ois Yvon"", ""Thierry Moudenc""], ""title"": ""Corpus design based on the Kullback-Leibler divergence for Text-To-Speech synthesis application"", ""year"": 2006}, {""authors"": [""Neubig"", ""Graham"", ""Shinsuke Mori.""], ""title"": ""Word-based partial annotation for efficient corpus construction"", ""venue"": ""Proceedings of the International Conference on Language Resources and Evaluation (LREC),"", ""year"": 2010}, {""authors"": [""Raz"", ""Ran"", ""Shmuel Safra.""], ""title"": ""A sub-constant error-probability low-degree test, and a sub-constant error-probability PCP characterization of NP"", ""venue"": ""Proceedings of the twenty-ninth annual ACM symposium"", ""year"": 1997}, {""authors"": [""Rojc"", ""Matej"", ""Zdravko Ka\u010di\u010d.""], ""title"": ""Design of optimal Slovenian speech corpus for use in the concatenative speech synthesis system"", ""venue"": ""Proceedings of the International Conference on Language Resources and"", ""year"": 2000}, {""authors"": [""Schein"", ""Andrew I."", ""Ted S. Sandler"", ""Lyle H. Ungar.""], ""title"": ""Bayesian example selection using BaBiES"", ""venue"": ""Technical Report MS-CIS-04-08, Department of Computer and Information Science, University of"", ""year"": 2004}, {""authors"": [""Settles"", ""Burr.""], ""title"": ""Active learning literature survey"", ""venue"": ""Technical Report 1648, Department of Computer Sciences, University of Wisconsin, Madison."", ""year"": 2010}, {""authors"": [""Synapse.""], ""title"": ""Documentation technique: Composant d\u2019\u00e9tiquetage et lemmatisation"", ""venue"": ""http://www.synapse-fr.com/."", ""year"": 2011}, {""authors"": [""Tian"", ""Jilei"", ""Jani Nurminen.""], ""title"": ""Optimization of text database using hierachical clustering"", ""venue"": ""Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP),"", ""year"": 2009}, {""authors"": [""Tian"", ""Jilei"", ""Jani Nurminen"", ""Imre Kiss.""], ""title"": ""Optimal subset selection from text databases"", ""venue"": ""Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal"", ""year"": 2005}, {""authors"": [""Tomanek"", ""Katrin"", ""Fredrik Olsson.""], ""title"": ""A Web survey on the use of active learning to support annotation of text data"", ""venue"": ""Proceedings of the NAACL HLT 2009 Workshop on Active Learning for Natural"", ""year"": 2009}, {""authors"": [""Van Santen"", ""Jan P.H."", ""Adam L. Buchsbaum.""], ""title"": ""Methods for optimal text selection"", ""venue"": ""Proceedings of the European"", ""year"": 1997}, {""authors"": [""Rhodes. Zhang"", ""Jin-Song"", ""Satoshi Nakamura""], ""title"": ""An improved greedy search"", ""venue"": ""Conference on Speech Communication and Technology (Eurospeech),"", ""year"": 2008}]",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :In automatic speech and language processing, many technologies make extensive use of written or read text sets. These linguistic corpora are a necessity to train models or to extract rules, and the quality of the results strongly depends on a corpus’ content. Often, the reference corpus should provide a maximum diversity of content. For example, in Tian, Nurminen, and Kiss (2005), and Tian and Nurminen (2009), it turns out that maximizing the text coverage of the learning corpus improves an automatic syllabification based on a neural network. Similarly, a high quality speech synthesis system based on the selection of speech units requires a rich corpus in terms of diphones, diphones in context, triphones, and prosodic markers. In particular, Bunnell (2010) shows the importance of a good coverage of diphones and triphones for the intelligibility of a voice produced by a unit selection speech synthesis system.
To cover the best attributes needed for a task, several strategies are then possible. A first method—very simple—is to collect text randomly, but it soon becomes expensive because of the natural distribution of linguistic events following the Zipf’s law. Very few events are extremely frequent and many events are very rare. This problem is often made difficult by the fact that many technologies require several variants of the same event (as in a Text-to-Speech [TTS] system using several acoustical versions of the same phonological unit). Usually, a large volume of data needs to be collected. However, and depending on the applications, building such corpora is often achieved under a constraint of parsimony. As an example, for a TTS system, a high-quality synthetic voice generally needs a huge number of speech recordings. But minimizing the duration of a recording is also a critical point to ensure uniform quality of the voice, to reduce the drudgery of the recording, to reduce the financial cost, or to follow a technical constraint on the amount of collected data for embedded systems. Moreover, a reduced set tends to limit the need of human implication for checking the data (transcription and annotation). Similarly, in the natural language processing field (NLP), the adaptation of a generic model to a specific domain often requires new annotated data that illustrate its specificities (as in Candito, Anguiano, and Seddah 2011). However, the creation cost of such data highly depends on the kind of labels used to adapt the model. In particular, the annotation in syntax trees is really more expensive than in Part-of-Speech (POS) tags. Then, it could be more efficient to annotate a compact corpus that reflects the phenomena variability than a corpus with a natural distribution of events, which implies many redundancies (see Neubig and Mori 2010).
In a machine learning framework, the active learning strategy can be used as an alternative that reduces the manual data annotation effort to design the training corpus without diminishing the quality of the model to train (see Settles 2010 or Schein, Sandler, and Ungar 2004). It consists of building the corpus iteratively by choosing an item according to an external source of information (a user or an experimental measure). This approach has been applied in NLP, speech recognition, and spoken language understanding (see for instance Tomanek and Olsson 2009 and Gotab, Béchet, and Damnati 2009).
A second alternative, when no direct quality measure is available, consists of covering a large set of attributes that may impact the final quality (after annotation or recording). This kind of approach might also be preferred when the final corpus is built in one batch (for instance, because of out-sourcing or annotator/performer consistency constraints). A method could be an automatic extraction from a huge text corpus of a minimal sized subset that covers the identified attributes. This
problem is a generalization of the Set-Covering Problem (SCP), which is an NPhard problem, as shown in Karp (1972). It is then necessary to use heuristics or sub-optimal algorithms for a reasonable computation time. Moreover, Raz and Safra (1997) and Alon, Moshkovitz, and Safra (2006) have shown that the SCP cannot be polynomially approximated with ratio c × ln(n) unless P = NP, when c is a constant, and n refers to the size of the universe to cover. That means that one cannot be certain to obtain a result under this ratio with any polynomial algorithm. However, the latter complexity results are given for any kind of distribution in the mono-representation case. One can ask if good multi-represented coverages can be achieved efficiently on data following Zipf’s law, which is usual in the domain of NLP.
Within the field of speech processing, the most frequently used strategy is a greedy method based on an agglomeration policy. This iterative algorithm selects the sentence with the highest score at each iteration. The score reflects the contribution of the sentence to the covering under construction. In Gauvain, Lamel, and Eskénazi (1990), this methodology has been applied to build a database of read speech from a text corpus for the evaluation of speech recognition systems using hierarchically organized covering attributes. Van Santen and Buchsbaum (1997) have tested different variants of greedy selection of texts by varying the units to cover (diphones, duration, etc.) and the “scores” for a sentence depending on the considered applications. In Tian, Nurminen, and Kiss (2005), the learning corpus for an automatic system of syllabification is designed using a greedy approach with the Levenshtein distance as a score function in order to maximize its text diversity. In François and Boëffard (2001), the methodology gives a priority to the rarest categories of allophones. The latter methodology has been implemented for the definition of the multi-speaker corpus Neologos in Krstulović et al. (2006). In the article of Krul et al. (2006), the authors constructed a corpus where the distribution of diphonemes/triphonemes matches a uniform distribution. A greedy algorithm is led by a score function based on the Kullback-Liebler divergence. A similar method is used in Krul et al. (2007) to design a reduced database in accordance with a specific domain distribution. Kawai et al. (2000) propose a pair exchange mechanism that Rojc and Kačič (2000) apply after a first reverse greedy algorithm—also called spitting greedy—deleting the useless sentences. In Cadic, Boidin, and d’Alessandro (2010), the covering of “sandwich” units (defined to be more adapted to corpus-based speech synthesis) is carried out by generating new sentences in a semi-automatic way. Candidates are generated using finite state transducers. The sentences are ordered according to a greedy criterion (their sandwiches richness) and presented to a human evaluator. This collection of artificial and rich sentences enables an effective reduction of the size of the covering but requires expensive human intervention to obtain semantically correct sentences that will be therefore easier to record.
The results of these previously cited studies are difficult to compare because of the different initial corpora and covering constraints (partial or full covering) and evaluation criteria (the number of gathered sentences, the Kullback divergence, etc.). In Zhang and Nakamura (2008), a priority policy for the rare units is added into an agglomerative greedy algorithm in order to get a covering of triphoneme classes from a large text corpus in Chinese language. The results show that this priority policy driven by the score function and the phonetic content of the sentences reduces the covering size compared with a standard agglomerative greedy algorithm.
Similarly, in François and Boëffard (2002), several combinations of greedy algorithms (agglomeration, spitting, pair exchange, or priority to rare units) were applied
to the construction of a corpus for speech synthesis in French containing at least three representatives of the most frequent diphones. Based on this work, the best strategy would be the application of an agglomerative greedy followed by a spitting greedy algorithm. During the agglomeration phase, the score of a sentence corresponds to the number of its unit instances that remain to be covered normalized by its length. During the spitting phase, at each iteration, the longest redundant sentence is removed from the covering. This algorithm is called the Agglomeration and Spitting Algorithm (ASA).
As an alternative to a greedy algorithm, which is sub-optimal, solving the SCP using Lagrangian relaxation principles can provide an exact solution for problems of reasonable size. However, for speech processing, the SCP has several millions of sentences with tens of thousands of covering features. Considering these practical constraints, Chevelu et al. (2007) adapted a Lagrangian relaxation based algorithm proposed by Caprara, Fischetti, and Toth (1999). In the context of Italian railways, Caprara, Fischetti, and Toth proposed heuristics to solve scheduling problems and won a competition, called Faster, organized by the Italian Operational Research Society in 1994, ahead of other Lagrangian relaxation heuristics–based algorithms, like Ceria, Nobili, and Sassano (1998).
In Chevelu et al. (2007, 2008), the algorithm takes into account the constraints of multi-representation. A minimal number of representatives for the same unit may be required. The proposed algorithm, called LamSCP — Lagrangian-based Algorithm for Multi-represented SCP — is applied to extract coverings of diphonemes with a mono- or a 5-representation and coverings of triphonemes with mono-representation constraints. These results are compared with the greedy strategy ASA and are about 5% to 10% better. Besides, the LamSCP provides a lower bound for the cost of the optimal covering and allows for evaluating the quality of the results.
In Barbot, Boëffard, and Delhay (2012), phonological content of diphoneme coverings is studied regarding many parameters. These coverings are obtained by different algorithms (LamSCP, ASA, greedy based on the Kullback divergence) and some of the coverings are randomly completed to reach a given size (from 20,000 to 30,000 phones). It turns out that the coverings obtained using LamSCP and ASA provide a good representation of short units and the representation of long units mainly depends on the length of the corpus.
In this article, we present in more detail the LamSCP algorithm and its score functions and heuristics that take into account multi-representation constraints. We deepen the study about the performance of LamSCP for the construction of a phonologically rich corpus according to the size of the search space. We evaluate LamSCP and ASA algorithms on a corpus of sentences in English for a covering of multi-represented diphones, where the minimal number of required unit representatives varies from one to five. We also compare them in the case of very constrained triphoneme coverings in English and French, which represent about 12 times more units to cover. Additionally, both algorithms are tested to provide multi-represented coverings of POS tags in order to assess their ability to deal with different kinds of linguistic data. A particular effort has been made on methodology to obtain comparable measures, to study the stability of both algorithms, and to establish confidence intervals for each solution.
This article is organized as follows. In Section 2, the SCP framework and the associated notations are introduced. The ASA algorithm is described in Section 3 and the LamSCP is detailed in Section 4. The experimental methodology is presented in Section 5 and results are discussed in Section 6. Before concluding in Section 8, we present experiments in the context of TTS where we evaluate on that task the benefits of a reduction in section 7. 2 the set-covering problem :Before describing the SCP-solving algorithms proposed in this article, we introduce in this section some notations and the Lagrangian properties used by LamSCP.
Let us consider a corpus A composed of n sentences s1, . . . , sn. According to the target applications, these sentences are annotated with respect to phonological, acoustic, prosodic attributes, and so forth. Each sentence is then associated with a family of units of different types. The set of units present in A is denoted U = {u1, . . . , um} and A can be represented by a matrix A = (aij), where aij is the number of instances of unit ui in the sentence sj. Therefore, the j th column of A corresponds to sentence sj in A. To simplify the writing, we define the sets M = {1, . . . , m} and N = {1, . . . , n}.
For a given vector of integers B = (b1, . . . , bm)T, a reduction X of A, also called covering of U , is defined as a subset of A which contains, for every i ∈ M, at least bi instances of ui. It can be described by a vector X = (x1, . . . , xn)T where xj = 1 if sj belongs to X and xj = 0 otherwise. In other words, a covering is a solution X ∈ {0, 1}n of the following system
∀i ∈ M, ∑ j∈N aijxj ≥ bi (1)
that is, AX ≥ B where B is called the constraint vector. Our aim is to optimize a covering according to a cost function minimization criterion. The covering cost is given by summing the costs of the sentences that compose the covering. The optimization problem can be formulated as the following SCP:
X∗ = arg min X∈{0,1}n
AX≥B
CX (2)
where C = (c1, . . . , cn) is the cost vector and cj the cost of the sentence sj. Because of the objective to minimize the total length of the covering, we have chosen to define the cost of a sentence as one of its length features. According to the considered application, the sentence cost can be defined as its number of phones (one of our objectives is to design a phonetically rich script with a minimal speech recording duration), or its number of words, part-of-speech tags, breath groups, and so on.
In Caprara, Fischetti, and Toth (1999), Caprara, Toth, and Fischetti (2000), and Ceria, Nobili, and Sassano (1998), the studied crew scheduling problem is a particular case of Equation (2) where A is a binary matrix and B = 1Rm (i.e., with mono-representation constraints). In order to ensure that Equation (1) admits a solution, we assume that, for each i ∈ M, the minimal number bi of ui instances required in the covering is not greater than the number (A1Rn )i of ui instances in A, that is A1Rn ≥ B. Under this assumption, A is the maximal size solution of Equation (1), represented by X = 1Rn. In the case where bi is greater than the number of ui instances in A, bi is set to (A1Rn )i.
To drive the SCP algorithms during the sentence selection phase, the covering capacity µj of sentence sj is defined as the number of its unit instances required in the covering in view of the constraint vector:
µj = ∑ i∈M min { aij, bi }
(3)
Let us notice that µj does not consider the excess unit instances: For example, if sj contains aij = 10 instances of ui and at least bi = 3 instances of ui are required, the contribution of ui to µj derivation only takes into account three instances of ui. 3 greedy algorithm asa :In this section, the two main steps that compose the algorithm ASA are briefly described. First, an agglomerative greedy procedure is applied to A so as to derive a covering. Next, a spitting greedy procedure reduces this covering in order to approach the optimal solution of Equation (2). The greedy strategy builds a sub-optimal solution to the SCP Equation (2) in an iterative way. At each iteration, the lowest cost sentence is chosen from A. If several sentences correspond to the lowest cost, the one coming first (i.e., the one with the lowest index) is chosen. Initially, the set of selected sentences X is empty, the matrix à associated with the candidate sentences is assigned to A, the current covering capacity of sj is given by µ̃j = µj, and the current constraint vector B̃ = B. The cost of sentence sj is defined by
σj = { cj/µ̃j if µ̃j ̸= 0 ∞ otherwise (4)
Indeed, if µ̃j = 0, it turns out that sj does not cover any unit missing in the solution X under construction and its infinite cost σj avoids its selection.
At each iteration, the selected sentence s is added to X . Taking into account the content of s, B̃ is updated to max { B̃ − Ã∆, 0Rm } where the jth entry of ∆ equals 1 if sj = s and 0 otherwise. Next, the associated column of s in à is set to 0Rm . For each sentence sj with a non-zero µ̃j feature, µ̃j is then updated using à and B̃ in Equation (3).
At last, the agglomerative greedy algorithm is stopped as soon as all the constraints are satisfied, that is, B̃ = 0Rm . The spitting greedy strategy also consists in building iteratively a sub-optimal solution Y to Equation (2) by reducing the size of a covering. The initial covering Y is set to the solution X derived by the agglomerative phase described earlier. At each iteration, the set of the redundant sentences of Y is calculated and the costliest one (according to the cost function C) is removed from Y . An element s of Y is said to be redundant if for each ui ∈ U , its number of instances into Y , denoted mi(Y ), and into Y \ {s}, denoted mi(Y \ {s}), check min {mi(Y ), bi} = min {mi(Y \ {s}), bi}. In other words, s is a redundant element of the covering Y if Y \ {s} is also a covering solution of Equation (1). The spitting greedy algorithm stops when the redundant sentence set is empty. 4 lagrangian relaxation based–algorithm :This section describes the main phases of the algorithm called LamSCP. This algorithm takes advantage of the Lagrangian relaxation properties reviewed herein in order to approach the optimal solution of Equation (2) as close as possible. Strongly inspired by
Caprara, Fischetti, and Toth (1999), but generalized to the multi-representation problem, this algorithm provides a lower bound of the optimal solution cost. Having such information is very useful for assessing the achievements of the SCP algorithms. Let us briefly recall the main principles of Lagrangian relaxation on which LamSCP is based to solve Equation (2) (see Fisher [1981] for more details on Lagrangian relaxation).
First, the Lagrangian function associated with Equation (2) is defined by
L(X,Λ) = CX +ΛT(B − AX) = ΛTB + C(Λ)X
(5)
where Λ ∈ ( R+ )m, X ∈ {0, 1}n, and C(Λ) = C − ΛTA. The coordinates of Λ = (λ1, . . . , λm)T are called Lagrangian multipliers and can be interpreted as a weighting of constraints (1). The jth entry of C(Λ), called Lagrangian cost cj(Λ) of sentence sj, takes into account its cost cj and the adequacy of its composition to address Equation (2). For every covering X and every Λ ∈ ( R+
)m, the Lagrangian function satisfies L(X,Λ) ≤ CX. Thus, the dual Lagrangian function defined by
L(Λ) = min X∈{0,1}n L(X,Λ) (6)
presents the following fundamental property: For every Λ ∈ Rm+ and every covering X, we have L(Λ) ≤ CX. Hence, L(Λ) is a lower bound of the minimal covering cost, CX∗, but does not necessarily correspond to the cost of a covering. In order to compute L(Λ), an acceptable solution for the vector X minimizing L(X,Λ) is X(Λ) = (x1(Λ), . . . , xn(Λ)) T where
xj(Λ) = {
1 if cj(Λ) < 0 0 if cj(Λ) > 0
∈ {0, 1} otherwise (7)
Additionally, the dual Lagrangian function and the Lagrangian costs inform about the potential usefulness of sentences in the optimal covering. More precisely, for a given Λ and a known upper bound UB of minimal covering cost, the gap g(Λ) = UB − L(Λ) measures the relaxation quality. If cj(Λ) is strictly greater than g(Λ), we can check that any covering containing sj has a cost value strictly greater than UB. Hence, sentence sj is not selected and xj can be fixed at zero. Similarly, if cj(Λ) < −g(Λ), any covering with a cost lower than UB contains sj and one can fix xj to 1. Therefore, an optimal covering is made up of sentences with a low Lagrangian cost, as done in Caprara, Toth, and Fischetti (2000) and Ceria, Nobili, and Sassano (1998), and the higher the relaxation quality (i.e., the lower g(Λ)) is, the cheaper the covering will be.
The resolution of the dual problem of Equation (2) consists in finding Λ∗ ∈ Rm+ that maximizes the lower bound L(Λ), that is
Λ∗ = arg max Λ∈Rm+ L(Λ) (8)
Because this real variable function L is concave and piecewise affine, a well-known approach for finding a near-optimal multiplier vector is the subgradient algorithm. The LamSCP is an iterative algorithm, composed of several procedures that aim to either improve the current best solution or reduce the combinatorial issue related to the considered problem. In order to derive a good solution, the algorithm calls on a great number of greedy procedures to solve sub-problems with the help of the Lagrangian costs. As for the combinatorial reduction, the most frequently used heuristic consists of downsizing the problem by mainly considering the sentences with low Lagrangian costs.
The algorithm is organized around a main procedure called 3-phases. This procedure can single-handedly solve a multi-represented SCP. As its name suggests, the 3-phases functioning consists in iterating a sequence of the three following subprocedures as shown in Figure 1:r The subgradient phase calculates an estimation Λ̃ of Λ∗ that maximizes
the dual Lagrangian function. This procedure requires an upper bound UB of the optimal covering cost. UB is initialized by a greedy algorithm (rather than the cost of the whole corpus A). This phase is detailed in Section 4.2.1.r The heuristic phase explores the neighborhood of Λ̃ by generating a great number of Lagrangian vectors Λ̃p. A greedy-like procedure is associated with each Λ̃p so as to compute a covering using the Lagrangian cost vector C(Λ̃p). If, during this exploration, a less costly covering than the best known one (corresponding to the cost UB) is found, the upper bound UB is then updated to the cost of this less costly solution. Similarly, if a better estimation of Λ∗ is obtained, Λ̃ is updated. This phase is described in Section 4.2.2.r The column fixing phase selects a set F of sentences that are most likely to belong to the optimal covering. This phase is detailed in Section 4.2.3.
Following the column fixing phase, the constraint vector is updated and the unselected sentences define a set-covering sub-problem, called a residual problem. This sub-problem is processed similarly, via an additional iteration of the three phases. This iterative process is stopped when the residual problem is empty or when the associated dual Lagrangian function indicates a cost is too high. Indeed, because this function indicates a minimal cost for covering the sub-problem, its addition to the cost CF of the sentences already retained in F gives a lower bound of the total cost of the solution under construction, which should not rise beyond the cost UB of the best known solution so as to be potentially more advantageous.
4.2.1 Subgradient Phase. In order to reach the quality goal, the subgradient phase provides a near-optimal solution Λ̃ of the dual Lagrangian problem (8) using a subgradient
type algorithm. This iterative approach generates a sequence (Λp) using the following updating formula (see Caprara, Fischetti, and Toth 1999)
Λp+1 = max { Λp + µ
g(Λp) ||S(Λp)||2
S(Λp), 0 }
(9)
where S(Λp) = B − AX(Λp) so as to take into account the multi-representation constraints. Parameter µ is adjusted to fit the convergence fastness according to the method proposed by Caprara, Fischetti, and Toth (1999).
At the first call of 3-phases, Λ0 is arbitrarily defined as follows: for each i ∈ M,
λ0i = minj∈N aij ̸=0
cj µj
(10)
As for UB, its initial value is set to the cost of a covering previously calculated. In order to evaluate how much the covering cost derived by ASA can be improved, we have chosen to initialize UB by this value. At the following iterations of 3-phases, Λ0 is given by a random perturbation (less than 10%) of the best known vector Λ̃ (of which the entries of the sentences fixed in the last column fixing phase are removed) and UB corresponds to the cost of the best covering found (after subtraction of the cost of the sentences fixed in the last column fixing phase). In another approach proposed by Ceria, Nobili, and Sassano (1998), UB corresponds to the upper bound of a dual Lagrangian problem, and the subgradient procedure simultaneously estimates the upper bound and the lower bound, generating two sequences of multipliers.
The subgradient phase also calls two procedures: pricing and spitting. Procedure pricing aims to reduce the size of the search space. For each unit ui, the pricing selects the 5bi smallest Lagrangian cost sentences covering ui. If this selection contains less than 5m sentences, where m is the maximal entry of B, it is completed by low Lagrangian cost sentences (less than 0.1) to the limit of 5m sentences. The set of the chosen sentences is denoted P and its design guarantees a sufficient number of instances for each unit to cover and a small variety in its composition. Actually, the subgradient method is applied on P instead of A, and P is updated every 10 subgradient iterations.
Finally, at each iteration, definition (7) of X(Λp) and the large number of Lagrangian costs close to zero allow a considerable number of vectors S(Λp). In order to get around the computational difficulty of finding the steepest accent direction, Caprara, Fischetti, and Toth (1999) propose a heuristic that, according to the experimental results, accelerates the convergence of the subgradient phase. This heuristic is implemented in the spitting procedure. Called at each iteration, this procedure extracts from P the subset S of sentences with a Lagrangian cost lower than 0.001. S is then reduced using a spitting greedy algorithm to remove its redundant elements in decreasing Lagrangian cost order. At last, for every j ∈ M, xj(Λp) = 1 if sj ∈ S and xj(Λp) = 0 otherwise. Thus, S(Λp) does not necessarily correspond to a subgradient vector.
4.2.2 Heuristic Phase. The heuristic phase calculates a large number of coverings before keeping the best one. To that end, a sequence of 150 multiplier vectors is generated by perturbing Λ̃ using the formula Λ̃p+1 = max { Λ̃p + µg(Λ̃p)S(Λ̃p), 0 } where Λ̃0 = Λ̃ and µ is provided by the subgradient phase, so as to allow for a change in a large number of Λ̃p. With each Λ̃p, an agglomerative greedy algorithm followed by a spitting greedy one are associated in order to calculate a covering.
The agglomerative greedy chooses at each iteration the sentence sj with the lowest cost σj(Λ̃p) where
σj(Λ̃p) =  cj(Λ̃p) ∗ µ̃j if cj(Λ̃p) < 0 and µ̃j > 0 cj(Λ̃p)/µ̃j if cj(Λ̃p) ≥ 0 and µ̃j > 0 ∞ if µ̃j = 0
(11)
This cost function provides an advantage to low Lagrangian cost sentences sj containing µ̃j unit instances that could be helpful to the covering under construction. The agglomerative step uses several heuristics so as to reduce the search space. It is run within a limited subset Pl of P , composed of the sentences sj with the lowest costs σj(Λ̃k). At each iteration, a sentence of Pl is selected and the cost of the sentences of P are updated. If the maximum sentence cost in Pl becomes greater than the minimal cost in P \ Pl, the working subset Pl is also updated. The definitions of P and Pl guarantee that the agglomeration step provides a nonpartial solution of the considered SCP. This solution is then reduced during the spitting step by iteratively removing its redundant sentences sj with the highest costs cj.
At the end of the heuristic phase, the best found covering X ∗ and its cost CX∗ (stored in UB) are kept as well as the highest value of L(Λ̃p) (found during the subgradient or heuristic phases).
4.2.3 Column Fixing Phase. The column fixing phase aims to reduce the problem size by choosing “promising” sentences among the ones with very low Lagrangian cost or containing rare unit instances. The unselected sentences are less interesting for resolving the SCP and the residual problem associated with these sentences should be the subject of another call of the 3-phases.
More precisely, the column fixing phase calculates the subset Q composed of sentences sj with a negative Lagrangian cost cj(Λ̃). Q is represented by the binary vector Q = (q1, . . . , qn)T where qj = 1 if sj ∈ Q. For each ui ∈ U , the number of its instances covered by Q is given by (AQ)i. If (AQ)i ≤ bi, then ui is considered as a rare unit and all the elements of Q containing some instances of ui are fixed in a set F . The covering constraints that are not satisfied by F constitute a residual SCP. In order to
complete F with a few sentences of the best known covering, a greedy-type algorithm is run on X ∗ \ F to derive a solution of this residual SCP. From the obtained solution, the max{ ( BT1Rm ) /20, 1} lowest Lagrangian cost sentences are also added in F . The sentences that are “fixed” in F during the column fixing phase stay in F up to the end of the 3-phases. After the column fixing phase, the residual sub-problem is processed by iterating the three phases and the next fixed sentences are added to F . The 3-phases procedure is encapsulated in an outer loop that permits the partial reconsideration of the solution X ∗ provided by this procedure. To that end, the refining procedure, proposed in Caprara, Fischetti, and Toth (1999), is used in order to select elements of X ∗ that contribute at least to the gap g(Λ̃). In the case of the SCP with multirepresentation constraints, the definition of the contribution of sj ∈ X ∗ can be adapted as follows:
δj = max{cj(Λ̃), 0}+ ∑ i∈M
aij>0
λ̃i(AX∗ − B)i aij
(AX∗)i (12)
The second term of Equation (12) consists of sharing the contribution λ̃i(AX∗ − B)i of the excess instances of ui in X ∗ according to the distribution of ui instances in X ∗. Therefore, the refining procedure ranks in an increasing order the elements sj of X ∗ according to their δj value, and fixes the first elements in a set G until its given covering rate τG reaches π. τG represents the rate of covering constraints satisfied by G and is defined by
τG = 1 − ∑
i∈M max{bi − (AG)i, 0} BT1Rm
(13)
where G denotes the binary vector corresponding to G. The LamSCP is made up of the main procedures introduced in the previous sections interlinked by the following steps. First, because of the adaptation of the algorithm to the SCP with multi-representation constraints, the entries of matrix A are clipped to the constraint vector in order to simplify the calculations such as the ones of µ1, . . . ,µn. This threshold application implies that the excess instances of each unit ui are taken into account in each sentence sj beyond the number bi, in the derivation of δj.
After the initialization of Λ0 using Equation (10) and the upper bound UB, procedure 3-phases is called and provides a solution X to the complete SCP. The refining function fixes a sentence subset G such that G covers a given rate π of the covering constraints. Parameter π starts at a minimum value πmin = 0.3. The residual SCP is then processed by 3-phases. The π value grows at a rate of 20 percent whenever 3-phases does not improve the solution to the complete SCP. If π is greater or equal to 1, the refining procedure fixes the whole best solution and the residual problem is then empty. On the other hand, π is set to πmin if a better solution is found in order to challenge G and improve this solution. This iterative sequence composed of the 3-phases and refining procedures is carried out until the residual problem is not empty, the gap g(Λ̃) is positive, and the number of iterations has not reached 20. 5 experiments :We propose a twofold comparison of the ASA and LamSCP algorithms: One part is focused on the covering cost for a large SCP, and the other on the stability of the solutions. Moreover, in order to assess the ability and the behavior of both algorithms to process different linguistic data, a first set of experiments deals with phonological attributes (mainly covering co-occurrences of phonemes) and a second set with grammatical labels (mainly covering co-occurrences of POS labels).
Both attribute types, often involved in automatic linguistic processing, were chosen because their distribution consists of few highly frequent events and numerous rare events. On the one hand, for TTS tasks, the phonological type covering is a useful preliminary step of the text corpus design before the recording step. In order to produce the signal corresponding to a requested sentence, the unit selection engine requires at least one instance of each phone (or 2-phone, depending on the concatenation process). Because the recording and the post-recording annotation process are expensive tasks, the recording length of such a corpus has to be as short as possible. On the other hand, in order to train a domain-specific dependency parser, the covering of POS sequences may be useful for increasing the diversity of syntax patterns. Because the dependency annotation is a highly expensive task, the adaptation corpus to annotate needs to be as small as possible, containing characteristic examples of the specific lexical variation rather than following the natural distribution. One can expect that increasing its diversity of POS sequences may lead to more diversity in the syntax trees.
Experiments on covering co-occurrences of phonemes are carried out on two large phonologically annotated text corpora, and consist of covering at least k instances of each phoneme, diphoneme until n-phoneme (i.e., triphoneme if n = 3, diphoneme if n = 2). The cost cj of the sentence sj is given by its number of phones. From this point, this kind of SCP is called a “k-covering of n-phonemes.”
A first corpus, Gutenberg, is composed of texts in English, mainly extracted from novels and short stories. This corpus is the production of the Gutenberg Project, presented by Hart (2003), and has been used by Kominek and Black (2003) to design the speech corpus Arctic. A second corpus, in French, named Le-Monde, is extracted from articles published in the newspaper Le Monde in 1997. Table 1 summarizes the main features of both corpora.
The phonological annotation of the Gutenberg corpus comes from the Arctic/ Festvox database (see Kominek and Black 2003), and the annotation of the Le-Monde corpus is a by-product of the Neologos project, detailed by Krstulović et al. (2006).
For each corpus, we have collected every phoneme, diphoneme, triphoneme, and their occurrences in each sentence so as to define the set U of units to cover and the matrix A. A is built by collecting one sentence after the other following the ordering inside the corpus, and one unit after the other inside the sentences. After this matrix translation, we obtain two description files and two index files. The first file describes the matrix A and the second one the cost vector C. Because of the low matrix density, we have chosen a sparse representation to save space and computation time: For instance, the 2-phoneme Gutenberg matrix is about 2.2% dense. We only store the cells of A that have a non-zero value so as to get a sparse matrix. The index files are made for the correspondence between the general covering problem and the application domain. The implementation is made in C. In terms of software engineering, our algorithms are working on an SCP that does not depend on the application data. For example, there is no information on what types of units are to be covered. The algorithms only have the matrix of occurrences A, the cost vector C, and the constraint vector B. A set of translation files (from application data to SCP and from SCP to application data) is built before each computation. As a consequence, there is no difficulty in addressing a different set of features to cover on the same or on a different corpus.
To study the achievements of ASA and LamSCP on different types of data, we have also chosen to address the “k-covering of n-POS” on the corpus Le-Monde. The grammatical and syntactical analyses are processed by the Synapse development analyzer presented in Synapse (2011). In order to consider a SCP with a substantial number of required units, a very detailed level of POS tagging has been selected, providing 141 distinct tags after analyzing Le-Monde. For example, this level provides tags like “Determiner male singular Article” or “Noun female singular,” whereas the simplest level gives “Determiner” or “Noun.” This latter level of description would have given only nine different POS tags after analyzing Le-Monde. The main associated statistics are given in Table 2. For these experiments of POS covering, the cost of a sentence is defined as its number of POS occurrences.
We used a PC with 2 CPUs (E5320/1.86GHz/4 cores/64bits) and 32 GB RAM for the phonological coverings and the POS coverings were computed using a PC with 8 CPUs (Intel Xeon X7550/2.00Ghz/8 cores/64bits) and 128 GB RAM. Our implementations do not take advantage of any parallelism.
The following sections detail more precisely the different experiments conducted on French or English. The aim of Experiment 1 is the assessment of the achievements of both algorithms, ASA and LamSCP, and the robustness of the results when the sentence ordering is modified
in the corpus to reduce. Indeed, one of the difficulties of the greedy methodology is that the score function has discrete values and several sentences can yield the same score. In our implementation, among the sentences showing the best current score, the first one encountered is chosen. We would like to measure the influence of this random choice on the stability of the results. LamSCP uses greedy strategies based on Lagrangian costs. Because the Lagrangian costs cj(Λ) take the SCP in its entirety into account and are continuous real-value functions of Λ, they would be more selective than the sentence costs used by ASA. A simple solution for evaluating the stability consists of proceeding with an important amount of experiments on the same SCP by randomly modifying the sentence ranking in A. Experiment 1 measures the impact of these permutations on the solutions computed by both algorithms. The considered SCP is the 1-covering of 2-phonemes on the corpus Le-Monde. Considering the computation time (more than 5 hours for LamSCP), only 47 instances of the SCP are considered, each instance corresponding to a random sentence ordering in Le-Monde. The 95% confidence intervals are derived using the bootstrap method, concerning the covering cost, the number and the length of sentences in coverings, the computation time, and the “distance” between the covering costs and the associated lower bound L(Λ̃).
5.2 Stability of the Algorithms for the k-Covering of 2-Phonemes in English
One of the goals of Experiment 2 is to compare the achievements of both algorithms on corpus Gutenberg, which has different features from the ones of Le-Monde. The sentences in Gutenberg are shorter on average and the associated variation of sentence length is lower. Furthermore, Gutenberg is 10 times smaller than Le-Monde.
In order to compare with the results of the previous experiment done on Le-Monde, we first observed a 1-covering of 2-phonemes on the Gutenberg corpus. The search space seems smaller than in Experiment 1. According to Table 1, Le-Monde is composed of more sentences than Gutenberg. Le-Monde contains 33,165,050 occurrences of 2-phonemes and Gutenberg only 3,025,474. Moreover, the number of attributes to cover is lower: 1,207 2-phonemes in Le-Monde and 2,012 in Gutenberg. We can also notice that five 2-phonemes have only one occurrence in Le-Monde and that the total cost of the five sentences covering these rare units is 751 phones, whereas this is the case for 109 2-phonemes in Gutenberg and the total cost of the 104 concerned sentences is 3,606 phones. Finally, if we consider the density of matrix A, 8.4% of the cells are non-empty for Experiment 1 and 2.2% for Experiment 2. As with Experiment 1, the sentence ordering in Gutenberg has been randomly modified to produce 60 instances of the SCP and similar solution statistics have been computed for both algorithms.
The second objective is to test and compare the ability of two algorithms to deal with the constraints of multi-representation. For this, we apply the same methodology to the k-covering of 2-phonemes in Gutenberg, for k from 2 to 5. We note that for the same original corpus to reduce, the size of the search space decreases when k increases. These different SCPs enable us to compare the performance of two algorithms depending on the size of the search space. In Experiment 3, the aim is to observe the behavior of both algorithms on very constrained problems. For this, we study their ability to treat a covering of 3-phonemes. We try to assess the impact on the solution features and on the stability of such an increase in the number of attributes to cover with many rare events. So as to compute statistics
on the 1-covering of 3-phonemes, an instance of Gutenberg has been proposed to ASA and to LamSCP. This instance counts 29,489 units to cover and the density of matrix A is 0.24%. The computation time is nearly 5 days for LamSCP and we then chose to carry out 35 instances of the 1-covering of 3-phonemes on Gutenberg. Additionally, it is interesting to compare these results with those of Experiment 2 concerning the 1-covering of 2-phonemes on the same corpus, which corresponds to a larger search space. In order to pursue the objective set out in the description of Experiment 3, that is, the ability of algorithms to treat with numerous constraints and a heavy-tail distribution of units, Experiment 4 consisted of testing both algorithms on the 1-covering of 3- phonemes on Le-Monde. The search space seems larger than in the previous experiment. Let us recall that Le-Monde contains 3.18 times more sentences than Gutenberg and it counts 27,650 units to cover. Furthermore, Gutenberg contains 5,000 3-phonemes with only one occurrence, which requires the selection of 4,180 sentences with a total length equal to 137,714 phones, whereas Le-Monde contains 2,274 rare 3-phonemes scattered in 2,107 sentences measuring a total of 283,208 phones. The associated matrix density is 0.69%. Because the computation of a first instance takes more than 8 days, we have limited the number of instances to 30 for this SCP.
5.5 k-Covering of 1-POS and 2-POS in French
The main goal of Experiment 5 is to study the behavior of both algorithms, ASA and LamSCP, dealing with another kind of linguistic attribute, and to compare this with the previous experiments. To achieve this goal, we consider POS attributes and the associated SCP: 1- and 5-coverings of POS, 1- and 5-coverings of 2-POS defined on Le-Monde. Indeed, we can observe in Table 2 that the global statistics of POS tags in Le-Monde are quite different from their phonological counterparts summarized in Table 1. In particular, the density of matrix A is 11.03% for a 1-POS covering and 0.57% for a 2-POS covering. Also, the search space size seems to decrease when considering successively the mono- and the multi-coverings of 1-POS, and the mono- and the multicoverings of 2-POS, permitting us to compare them with the results coming from the experiments on phonological coverings. We evaluate the stability by computing 50 randomly mixed versions of the corpus Le-Monde. 6 results and discussion :In this section, the results of the experiments described in Section 5 are provided and discussed. As a consequence, the organization of this section and Section 5 are similar. Table 3 shows the main results of Experiment 1, concerning the 1-covering of 2- phonemes from the corpus Le-Monde. Symbol ± indicates that the mentioned value corresponds to a 95% confidence interval, calculated using the bootstrap method from
the 47 instances of the SCP. In order to cover each of the 1,207 2-phonemes of Le-Monde, ASA drastically reduces the size of the initial corpus by 99.94% (±0.00). However, on average, LamSCP calculates a 9.00% shorter covering. The lower bound L(Λ̃) for the optimal covering cost is 7,689 ± 5 phones. L(Λ̃) is not a minimum value and may not correspond to the cost of a real covering. Because this lower bound is updated all along the execution of LamSCP, we do not mention a specific calculation time for this result.
For one instance of the SCP, let CX∗ASA be the size of the solution given by ASA. The quantity τASA = 1 − L(Λ̃)/CX∗ASA indicates that the optimum solution to SCP is at most τASA times shorter than the covering calculated by ASA. It can be observed that the optimal solution is at most 10.13% (±0.19) shorter than the one yielded by ASA and at most 1.24% (±0.08) shorter than the solution yielded by LamSCP. The solutions obtained by LamSCP and the optimal solution to the SCP are therefore very close. Considered among the 47 instances of the SCP the best solutions yielded by ASA (8,447 phones) and LamSCP (7,767 phones), LamSCP is 8.75% better than ASA in terms of covering costs, while the best lower bound for the SCP is 7,715 phones, only 0.67% (respectively, 8.66%) shorter than the best covering by LamSCP (respectively, ASA).
The average length of the sentences selected by both algorithms is far below the average length of the sentences in the corpus (96.81 phones). LamSCP tends to choose sentences that are slightly longer than ASA, with an average 28.97 (±0.12) phones compared with 25.48 (±0.10) phones. Moreover, ASA selects on average 335.73 (±1.69) sentences per solution, about 24.91% more than LamSCP, which selects 268.76 (±1.08) sentences on average. This seems to indicate that LamSCP makes fewer local choices than ASA. This hypothesis can also be validated through the analysis of the variability of the results. The relative variation of the covering costs calculated by LamSCP is 13.01/7,786 = 0.16%, and 57.40/8,555 = 0.67% by ASA; that is to say a stability of the costs 4 times greater for the solutions yielded by LamSCP than for ASA. Moreover, the solutions are composed of a very stable number of sentences: The associated relative standard deviation is 5.31/335.73 = 1.58% for the 47 instances solved by ASA, and 3.85/268.76 = 1.43% for the instances solved by LamSCP. It turns out that the results of both algorithms are very stable when the order of the sentences is modified in the original corpus.
Finally, concerning computation time, the resolution of an instance of the SCP lasts on average 5 hr 41 min 18 sec (±8 min 28 sec) for LamSCP versus 51 sec
(±0 sec) for ASA. On average over the 47 instances, LamSCP takes 390 (±9) times as long as ASA.
6.2 Stability of the Algorithms for k-Covering of 2-Phonemes in English
The considered SCP consists of covering at least k times each of the 2,012 2-phonemes of the Gutenberg corpus, with k varying from 1 to 5. The results are summarized in Table 4.
For all instances of these SCPs, it has been observed that LamSCP computes shorter coverings than ASA. However, that advantage diminishes as k grows: The cost advantage offered by LamSCP compared with ASA decreases from 9.73% (±0.13) for k = 1 to 4.50% (±0.04) for k = 5. Also, the solutions obtained from ASA and LamSCP seem to get closer to the optimal solution as k rises. The corresponding figures are presented in Table 5: For instance, the optimal solution is at most 0.75% (±0.02) shorter than that obtained by LamSCP for k = 1, and 0.27% (±0.00) for k = 5.
Because the search area diminishes as k increases, it may be observed that the algorithms tend to be more stable. This is true both for the size of the solutions, as well as for the number of sentences that define them. Table 6 represents the variation of the size of the solutions as a function of k. This variation is calculated as follows: For a given k number and a given algorithm, the standard deviation of the size of the k-covering computed by that algorithm is divided by the average size of these coverings. Thus, it can be noted that LamSCP offers a stability 4 to 8 times superior to ASA concerning the size of the coverings. As for the number of sentences, the relative standard deviation similarly decreases from 0.97% to 0.28% when k increases from 1 to 5 for ASA solutions, and from 0.42% to 0.15% for LamSCP ones.
One can note that the increase of the minimal number k of instances of each unit to cover leads to a selection, by LamSCP and ASA, of longer sentences on average. The average length of the sentences picked for a 1-covering was quite low. As the constraints increase along with k, it only seems natural that the algorithms tend to select longer sentences, as shorter sentences no longer contain enough occurrences of 2-phonemes. Moreover, as described in Section 2, when the minimal number bi of a unit ui demanded in the covering exceeds the number of instances of that unit in the initial corpus, all sentences containing instances of ui in the initial corpus are selected, and bi is set to (A1Rn )i. Thus, as k increases, the algorithm tends to select more and more sentences, and their length tends towards the average value over the whole corpus, which is 28.51 phones for Gutenberg.
As for computation time, although it increases as k grows, because of the increasing number of constraints to update, the ratio between the computation time of LamSCP and ASA tends to diminish, as shown in Table 7. This tendency may find an explanation in the fact that the search space diminishes as k increases, which causes a lesser number of selected sentences to be questioned during the 3-phase iteration of LamSCP. Also, we notice that the average computation time of the two algorithms is greater in Experiment 1, owing to a greater number of sentences in corpus Le-Monde and a higher density of matrix A. Moreover, the ratio between the computation times of
Time LamSCP / Time ASA 333 (± 7) 280 (± 5) 292 (± 6) 218 (± 9) 194 (± 8)
LamSCP and ASA decreases between Experiments 1 and 2, going from 390 (±9) to 333 (±7). Again, this can be explained by the diminishing of the search space.
For k = 1, the advantage offered by LamSCP on the covering costs compared with ASA is slightly higher than that observed in Experiment 1: 9.73% (±0.13) in this case, versus 9.00% (±0.20) in the previous experiment. This seems to contradict the idea that the performance of LamSCP improves as the search area becomes wider. However, the distributions of the units to cover in Gutenberg and Le-Monde are different, and the variation on the length of the sentences in Le-Monde is very high, which may account for this slight difference in terms of gain. Note that the size of the calculated coverings and the lower value L(Λ̃) are closer in the experiment carried out on Gutenberg. It is difficult, however, to perform further comparisons with Experiment 1 regarding the “distance“ between the costs of the solutions computed by these algorithms, and the optimal covering cost, given that the quality of the lower bound cannot be evaluated. The gain in stability offered by LamSCP, both for the costs of the solutions or the number of sentences, is more important than that noticed during the previous experiment. We think that the increase is due to a more restricted search space, and less variability of the length of the sentences in corpus Gutenberg, which may be observed in Table 1. Table 8 sums up the main results of Experiment 3, where 35 instances of 1-covering of 3-phonemes from Gutenberg were processed. According to the L(Λ̃) values, covering all 3-phonemes requires a solution size greater than or equal to 226,635 phones. On average, the solution measures 227,360 ± 12 phones using LamSCP, and 236,828 ± 94
phones using ASA. The optimal covering is at most 0.35% (±0.00) shorter than solutions derived by LamSCP and 4.33% (±0.04) shorter than the ones derived by ASA. We can then observe that both algorithms manage to compute solutions with close sizes when scaling up the required attribute set. The solutions are very stable, even more than in Experiment 2: The relative variation of their size is 0.12% for ASA and 0.01% for LamSCP; the relative variation of their sentence number is 0.17% for ASA and 0.07% for LamSCP. This increase of stability is due to a smaller search space and the increase of the number of rare units required, which also compels the algorithms to select a higher number of inevitable sentences for all the instances of the SCP. Furthermore, the decrease of the ratio between the computation time of LamSCP and ASA from 332 for the 1-covering of 2-phonemes to 130 for the one of 3-phonemes on Gutenberg may confirm this idea, which has also been put forward in Experiment 2.
Concerning the length of the selected sentences by both algorithms, it is greater than the one for the 5-covering of 2-phonemes, observed in Experiment 2, and slightly deviates from the average sentence length for the whole corpus. Consequently, it turns out that covering longer and generally rarer units involves a selection of longer sentences. This is confirmed by the fact that the sentences of Gutenberg covering units with a single occurrence in Gutenberg represent more than half the size of the solutions and are composed of 33 phones on average. In this section, we analyze the results of Experiment 4, the 30 instances of the 1- covering of 3-phonemes from Le-Monde carried out by ASA and LamSCP. The results are given in Table 9. First, although the main features of Le-Monde and Gutenberg are different, notice that the closeness between the size of coverings calculated by both algorithms and the lower bound L(Λ̃) is comparable to the one observed in Experiment 3. Indeed, the optimal covering size is at most 0.48% (±0.03) and 4.35% (±0.05) shorter than the solution size derived by LamSCP and ASA, respectively. Similarly, the size of solutions and the number of selected sentences are as stable as those observed in the previous experiment: The solution length varies from 0.01% for LamSCP to 0.10% for ASA and the number of sentences fluctuates about 0.16% for ASA and 0.10% for LamSCP.
As for the comparison with the results of Experiment 1 (1-covering of 2-phonemes from Le-Monde), the main trends are similar to the ones observed for the transition from the 1-covering of 2-phonemes to the 1-covering of 3-phonemes from Gutenberg.
However, in Experiment 4, the average selected sentence length has markedly increased, approaching the mean value on the whole corpus: 89.02 (±0.03) for ASA and 92.64 (±0.03) for LamSCP, whereas in Experiment 1 these values are, respectively, 25.48 (±0.10) and 28.97 (±0.12). We have already observed in Experiment 3 that covering longer units increases the length of selected sentences but this high amplitude seems to be inherent to the design of corpus Le-Monde. Furthermore, notice that the 1-coverings of 2-phonemes from Le-Monde are almost half as small as the ones from Gutenberg, whereas the 1-coverings of 3-phonemes from Le-Monde are between twice and three times longer than the ones from Gutenberg. This is due to the fact that the 3-phonemes with a single instance in Le-Monde are very scattered in long sentences (their length mean is about 134 phones), and these indispensable sentences represent nearly half the size of the solutions. The other sentences of the solutions are around 70 phones long. Lastly, the ratio between the computation time of both algorithms is about 84, which is smaller than the ratios previously observed, but this SCP is the most time consuming: 2 hr 10 min for ASA and more than 7 days for LamSCP.
6.5 k-Covering of 1-POS and 2-POS in French
Table 10 sums up the main results of Experiment 5, dealing with the 1- and 5-coverings of 1-POS and 2-POS. For all these SCP, LamSCP produces smaller coverings, composed of longer sentences, than the coverings obtained with ASA. When the search space diminishes, the relative “distance” between the size of solutions provided by both algorithms decreases, as well as between the lower bound L(Λ̃) and the size of solutions obtained by ASA. These trends were also observed in the earlier experiments. In particular, as for the 1-covering of 1-POS, not only does LamSCP provide 10.06% (± 0.00) shorter solutions than ASA, but its solutions are optimal for all 50 instances of this SCP. Indeed, the lower bound value varies from 482.51 to 482.87 occurrences of 1-POS while all the solutions given by LamSCP are made of exactly 483 occurrences of POS. For the other k-coverings of n-POS, the optimal solution is at worst 0.39% (±0.02) shorter than the covering given by LamSCP for (k, n) = (5, 1), 0.11% (±0.00) for (k, n) = (1, 2), and 0.22% (±0.00) for (k, n) = (5, 2). The solutions obtained by ASA or by LamSCP are very stable. For example, the relative standard deviation of number of POS in a covering solution varies from 0.00% to 1.23% for ASA, and from 0.00% to 0.04% for LamSCP.
As previously observed for both algorithms, their computation times grow when the number of required covering features increases. However, the ratio between the computation time of LamSCP and ASA does not behave as in Experiment 2 (see Table 7): For the k-covering of 1-POS, this ratio increases from 290 (±11) to 657 (± 43) when k goes from 1 to 5, and for the k-covering of 2-POS, it increases from 75 (±5) to 108 (±6). 7 evaluation on a text-to-speech synthesis system :In the previous sections, different algorithms dealing with corpus reduction were introduced and studied. The proposed experiments mainly evaluate the effects of these algorithms in terms of corpus reduction but not according to a practical task. This section proposes an experiment to assess the impact of the corpus reduction on a unit selection speech synthesis system.
As explained in Section 1, a corpus reduction for a TTS system is a trade-off between minimizing the recording and post-processing time to build the speech corpus and
keeping the highest phonological richness of the corpus to ensure the quality of the synthetic speech. The goal of this experiment is to measure this trade-off by evaluating the quality of the same TTS system fed with different speech corpora uttered by the same speaker. Note that the intrinsic quality of this system is not the purpose here.
Firstly, a brief presentation of a state-of-the-art unit selection–based TTS system is proposed in Section 7.1. The linguistic parameters used by the TTS system are detailed because they are linked to the required features in the reduction stage. In Section 7.2, corpora used in the experiment are introduced. The attributes to cover and the
methodology of evaluation are described in Section 7.3; the results are given and discussed in Section 7.4. For this experiment, a state-of-the-art unit selection–based TTS system is used to produce an acoustic signal from an input text. A linguistic front end processes the text to extract features taken into account by the algorithm that selects segments in a speech corpus (see Boëffard and d’Alessandro 2012).
The input text is converted into a sequence of phonemes using a French phonetizer proposed by Béchet (2001). Non-speech sound labels can be added to this sequence (silences, breaths, para-verbal events, etc.). A vector of features is defined as follows:
1. The phone or non-speech sound label
2. Is the described segment a non-speech sound?
3. Is the phone in the onset of the syllable?
4. Is the phone in the coda of the syllable?
5. Is the phone in the last syllable of its syntagm?
6. Is the current syllable at the end of a word?
7. Is the current syllable at the beginning of a word?
Extraction of features is done using the ROOTS toolkit described in Boëffard et al. (2012). The unit selection process aims to associate a signal segment from the speech corpus to each vector of features computed from the input text. This is performed in two steps. In the first step, for each unit, a set of candidates that match the same features are extracted from the speech corpus. In the second step, given all candidates, the best path is searched using an optimization algorithm so as to produce the sequence of speech units. The algorithm tries to minimize three sub-costs, commonly used in unit selection based TTS systems, which are spectral discrepancies based on MFCC distance, amplitude, and f0 distances. Two corpora are used in this experiment. The first one, Learning corpus, is an annotated acoustic corpus used to provide speech data for the TTS engine. It is an expressive corpus in French, spoken by a male speaker reading Albertine disparue, an excerpt from À la recherche du temps perdu by Marcel Proust. The corpus is composed of 3,138 sentences automatically annotated using a process described in Boëffard et al. (2012). The overall length of the speech corpus is 9 hr 57 min. When creating a voice for a unit selection– based TTS system, long sentences are generally removed or split into syntagm groups in order to help the speaker.
A second corpus, named Test corpus, is a text corpus that is synthesized and used in the listening experiment. It is composed of 30 short sentences randomly extracted from a phonetically balanced corpus in French, proposed by Combescure (1981). The use of a corpus with a different linguistic style minimizes the bias introduced by the learning corpus.
Statistics are given in Table 11. For this experiment, two reduced corpora are evaluated. They are built by reducing the full learning corpus using the two different algorithms presented in the previous sections: ASA and LamSCP.
As described in Section 7.1, the unit selection process of the speech synthesis system is based on a set of phonological attributes. It seems natural to try to cover features that reflect the variability of these attributes. For this experiment, algorithms must cover all the units at least once, where a unit is described by the following:
r Its label, that is, one of the 35 phonemes or a non-speech sound labelr The structure of the syllable that contains the phoneme, if it is a vowelr The position of the associated syllable in the word (start, middle, or end)r A Boolean indicated if the associated syllable is at the end of a syntagm The feature extraction is performed by the same set of tools used by the speech synthesis engine. Given this set of features, the learning corpus contains 1,497 classes of units. The cost function to minimize by the reduction algorithms is the total length of the set of the selected syntagms, in phones.
Two speech synthesis systems are defined, extracting the speech units to concatenate from the coverings provided by LamSCP and ASA. Two other systems are added as baselines. First, a system named Full, built with the whole learning corpus, is used as an upper bound. Second, a system named Random uses a random reduction of the Learning corpus as a pool corpus of speech units. This reduction is done by randomly selecting sentences from the whole learning corpus until the size of the covering obtained by LamSCP is reached. Random is used as a lower bound.
Whereas the optimization efficiency is measured by statistics on the reduced corpora, the quality of the synthesized speech signals is evaluated by a listening test. The protocol is based on a MUSHRA test, presented in ITU-R (2003), where for every sentence of Test corpus, the signals synthesized by the four systems are presented to each tester in a random order. If a system is not able to produce a signal for a requested sentence (because of a missing 2-phone in the pool corpus), an empty signal is presented. Ten native French testers (four naives and six experts) are asked to evaluate the overall quality of the stimuli and to give a mark from 0 to 100 (by steps of 5 points). The Learning corpus composed of 19,587 syntagms is first reduced using ASA and LamSCP. Statistics of the resulting solutions are summarized in Table 12. The covering of the 1,497 constraints divides the input corpus size by almost 10 and reduces the 10 hours of speech to around 1 hr 20 min. As for the previous experiments, even though different kinds of features are mixed (phonemes, syllable structures, position in a word or a syntagm) the ASA algorithm produces a solution close to the optimal one. However, LamSCP is again slightly better in terms of covering size.
For the measure of the acoustic impact of corpus reduction, the listening test results are presented in Table 13 for the average marks and in Table 14 for the average ranks. Note that without a natural speech reference during the test, the marks should not be seen as an absolute score. Even if the LamSCP corpus is slightly smaller than the one from ASA, the acoustic quality of both systems is comparable according to the testers (with a slight advantage for the LamSCP corpus). In comparison with the baseline, which uses the whole learning corpus, the acoustic degradation is significant. This illustrates well the trade-off between corpus size and speech quality: For a 90% corpus size reduction, the acoustic quality drops by 10 points. Further research should be focused on the set of attributes to cover and their number of occurrences in order to improve this compromise. As expected, the baseline built from random sentences is preferred significantly less than the other systems because of the lack of relatively rare acoustic units. 8 conclusion :This article discussed the building of linguistic-rich corpora under an objective of parsimony. This task, a generalization of SCP, turns out to be an NP-hard problem that cannot be polynomially approximated. We studied the behavior of several algorithms in the particular domain of NLP, where the considered events follow a heavy-tailed type distribution.
The proposed algorithms have been compared through three kinds of experiments: The first one is the coverings of 2- and 3-phonemes from two text corpora, one in French, the other in English; the second one consists of the coverings of part-of-speech labels from a corpus in French; the third one evaluates the impact of both algorithms on the acoustic quality of a corpus-based TTS system. The first algorithm, ASA, is composed of an agglomerative greedy strategy followed by a spitting greedy stage. The second one, LamSCP, is based on Lagrangian relaxation principles combined with greedy strategies. LamSCP is our adaptation of an algorithm proposed in Caprara, Fischetti, and Toth (1999) to the multi-representation constraints. The comparison of SCP solutions is mainly about their size, their maximal distance with the optimal covering, and their robustness in case of perturbation of the initial corpus ordering.
Although ASA is much faster than LamSCP, it does not permit us to singlehandedly assess the quality of its solution in terms of size. The main assets of LamSCP are the calculation of a lower bound to the optimal covering size and shorter solutions than the ones obtained by ASA. Indeed, in our experiments of phonological coverings, the optimal solution is at most 1.24% (10.13%, respectively) smaller than the solutions derived by LamSCP (ASA, respectively). As for the coverings of 1-POS, LamSCP provides the optimal solution in a case of a mono-representation constraint, whereas the ASA solution is 10.17% greater than the optimal one. These relative gaps between the lower bounds and solution sizes of both algorithms generally decrease when the size of the search space decreases. Thanks to the lower bound derived by LamSCP, we empirically show that it is possible to get almost optimal solutions in a linguistic framework following Zipf’s law distribution, despite the theoretic complexity of the multi-represented SCP. Concerning the last experiment in the TTS framework, even if LamSCP provides a smaller corpus, the subjective test shows no significant difference between the TTS systems based on LamSCP and ASA corpora. Therefore, we think that ASA remains the most adequate strategy, in terms of performance, ease of development, and computation time to solve SCP in the NLP field. However, it would be interesting
to test a parallelized version of the heuristic phase that calls an important number of greedy sub-procedures.
Our future prospects for this work are in automatic language processing and speech synthesis. First, in the framework of the Phorevox project supported by the French National Research Agency, we are considering the automatic design of exercise contents for language learning by the selection of texts covering some phonological or linguistic difficulties. Secondly, this work is a preliminary step to building a phonetically rich script before its recording in order to produce a high quality speech synthesis. The covering choices, such as the attributes to cover, the number of required occurrences, or the “sentence” length (utterances, syntagms, etc.) need to be validated. Moreover, in this article, we have observed the great impact of the distribution of rare units in the corpus to reduce, and we believe it will be interesting to adapt the “sentence” granularity according to this distribution. Linguistic corpus design is a critical concern for building rich annotated corpora useful in different domains of applications. For example, speech technologies such as ASR (Automatic Speech Recognition) or TTS (Text-to-Speech) need a huge amount of speech data to train datadriven models or to produce synthetic speech. Collecting data is always related to costs (recording speech, verifying annotations, etc.), and as a rule of thumb, the more data you gather, the more costly your application will be. Within this context, we present in this article solutions to reduce the amount of linguistic text content while maintaining a sufficient level of linguistic richness required by a model or an application. This problem can be formalized as a Set Covering Problem (SCP) and we evaluate two algorithmic heuristics applied to design large text corpora in English and French for covering phonological information or POS labels. The first considered algorithm is a standard greedy solution with an agglomerative/spitting strategy and we propose a second algorithm based on Lagrangian relaxation. The latter approach provides a lower bound to the cost of each covering solution. This lower bound can be used as a metric to evaluate the quality of a reduced corpus whatever the algorithm applied. Experiments show that a suboptimal algorithm like a greedy algorithm achieves good results; the cost of its solutions is not so far from the lower bound (about 4.35% for 3-phoneme coverings). Usually, constraints in SCP are binary; we proposed here a generalization where the constraints on each covering feature can be multi-valued. [{""affiliations"": [], ""name"": ""Nelly Barbot""}, {""affiliations"": [], ""name"": ""Olivier Bo\u00ebffard""}, {""affiliations"": [], ""name"": ""Jonathan Chevelu""}, {""affiliations"": [], ""name"": ""Arnaud Delhay""}] SP:066100434f650471b7f8100e81aef651fc0bf5f5 [{""authors"": [""Alon"", ""Noga"", ""Dana Moshkovitz"", ""Shmuel Safra.""], ""title"": ""Algorithmic construction of sets for k-restrictions"", ""venue"": ""ACM Transactions on Algorithms (TALG), 2(2):153\u2013177."", ""year"": 2006}, {""authors"": [""Barbot"", ""Nelly"", ""Olivier Bo\u00ebffard"", ""Arnaud Delhay.""], ""title"": ""Comparing performance of different set-covering strategies for linguistic content optimization in speech corpora"", ""venue"": ""Proceedings of the International"", ""year"": 2012}, {""authors"": [""B\u00e9chet"", ""Fr\u00e9d\u00e9ric.""], ""title"": ""Liaphon: un systeme complet de phon\u00e9tisation de textes"", ""venue"": ""Traitement automatique des langues, 42(1):47\u201367."", ""year"": 2001}, {""authors"": [""Bo\u00ebffard"", ""Olivier"", ""Laure Charonnat"", ""S\u00e9bastien Le Maguer"", ""Damien Lolive"", ""Ga\u00eblle Vidal.""], ""title"": ""Towards fully automatic annotation of audiobooks for TTS"", ""venue"": ""Proceedings of the International Conference on"", ""year"": 2012}, {""authors"": [""Bunnell"", ""H. Timothy.""], ""title"": ""Crafting small databases for unit selection TTS: Effects on intelligibility"", ""venue"": ""Proceedings of the ISCA Tutorial and Research Workshop on Speech Synthesis (SSW7), pages 40\u201344, Kyoto."", ""year"": 2010}, {""authors"": [""Cadic"", ""Didier"", ""C\u00e9dric Boidin"", ""Christophe d\u2019Alessandro""], ""title"": ""Towards optimal TTS corpora"", ""venue"": ""In Proceedings of the International Conference on Language Resources and Evaluation (LREC),"", ""year"": 2010}, {""authors"": [""Candito"", ""Marie"", ""Enrique Henestroza Anguiano"", ""Djam\u00e9 Seddah.""], ""title"": ""A word clustering approach to domain adaptation: Effective parsing of biomedical texts"", ""venue"": ""Proceedings of the 12th International"", ""year"": 2011}, {""authors"": [""Caprara"", ""Alberto"", ""Matteo Fischetti"", ""Paolo Toth.""], ""title"": ""A heuristic method for the set covering problem"", ""venue"": ""Operations Research, 47(5):730\u2013743."", ""year"": 1999}, {""authors"": [""Caprara"", ""Alberto"", ""Paolo Toth"", ""Matteo Fischetti.""], ""title"": ""Algorithms for the set covering problem"", ""venue"": ""Annals of Operations Research, 98(1-4):353\u2013371."", ""year"": 2000}, {""authors"": [""Ceria"", ""Sebasti\u00e1n"", ""Paolo Nobili"", ""Antonio Sassano.""], ""title"": ""A Lagrangian-based heuristic for large-scale set covering problems"", ""venue"": ""Mathematical Programming, 81(2):215\u2013228."", ""year"": 1998}, {""authors"": [""Chevelu"", ""Jonathan"", ""Nelly Barbot"", ""Olivier Bo\u00ebffard"", ""Arnaud Delhay.""], ""title"": ""Lagrangian relaxation for optimal corpus design"", ""venue"": ""Proceedings of the ISCA Tutorial and Research Workshop on Speech Synthesis"", ""year"": 2007}, {""authors"": [""Chevelu"", ""Jonathan"", ""Nelly Barbot"", ""Olivier Bo\u00ebffard"", ""Arnaud Delhay.""], ""title"": ""Comparing set-covering strategies for optimal corpus design"", ""venue"": ""Proceedings of the International Conference on Language"", ""year"": 2008}, {""authors"": [""Combescure"", ""Pierre.""], ""title"": ""20 listes de 10 phrases phon\u00e9tiquement \u00e9quilibr\u00e9es"", ""venue"": ""Revue d\u2019Acoustique, 56:34\u201338."", ""year"": 1981}, {""authors"": [""Fisher"", ""Marshall L.""], ""title"": ""The Lagrangian relaxation method for solving integer programming problems"", ""venue"": ""Management Science, 27(1):1\u201318."", ""year"": 1981}, {""authors"": [""Fran\u00e7ois"", ""H\u00e9l\u00e8ne"", ""Olivier Bo\u00ebffard.""], ""title"": ""Design of an optimal continuous speech database for text-to-speech synthesis considered as a set covering problem"", ""venue"": ""Proceedings of the European Conference on"", ""year"": 2001}, {""authors"": [""Fran\u00e7ois"", ""H\u00e9l\u00e8ne"", ""Olivier Bo\u00ebffard""], ""title"": ""The greedy algorithm and its application"", ""year"": 2002}, {""authors"": [""Gauvain"", ""Jean-Luc"", ""Lori Lamel"", ""Maxine Esk\u00e9nazi.""], ""title"": ""Design considerations and text selection for Bref, a large French readspeech corpus"", ""venue"": ""Proceedings of the International Conference of Spoken Language"", ""year"": 1990}, {""authors"": [""Gotab"", ""Pierre"", ""Fr\u00e9d\u00e9ric B\u00e9chet"", ""G\u00e9raldine Damnati.""], ""title"": ""Active learning for rule-based and corpus-based spoken language understanding models"", ""venue"": ""Proceedings of the IEEE workshop on"", ""year"": 2009}, {""authors"": [""Hart"", ""Michael.""], ""title"": ""Project gutenberg"", ""venue"": ""http://www.gutenberg.org/ (Last consulted April 2015)."", ""year"": 2003}, {""authors"": [""Karp"", ""Richard M.""], ""title"": ""Reducibility among combinatorial problems"", ""venue"": ""Complexity of Computer Computations, the IBM Research Symposia Series. Springer, pages 85\u2013103."", ""year"": 1972}, {""authors"": [""Kawai"", ""Hisashi"", ""Seiichi Yamamoto"", ""Norio Higuchi"", ""Tohru Shimizu.""], ""title"": ""A design method of speech corpus for text-to-speech synthesis taking account of prosody"", ""venue"": ""Proceedings of the"", ""year"": 2000}, {""authors"": [""Kominek"", ""John"", ""Alan W. Black.""], ""title"": ""The CMU Arctic speech databases for speech synthesis research"", ""venue"": ""Technical Report CMU-LTI-03-177, Carnegie Mellon University Language Technologies"", ""year"": 2003}, {""authors"": [""Krstulovi\u0107"", ""Sacha"", ""Fr\u00e9d\u00e9ric Bimbot"", ""Olivier Bo\u00ebffard"", ""Delphine Charlet"", ""Dominique Fohr"", ""Odile Mella""], ""title"": ""Optimizing the coverage of a speech database through a selection of representative speaker"", ""year"": 2006}, {""authors"": [""Krul"", ""Aleksandra"", ""G\u00e9raldine Damnati"", ""Fran\u00e7ois Yvon"", ""C\u00e9dric Boidin"", ""Thierry Moudenc""], ""title"": ""Adaptive database reduction for domain specific speech"", ""year"": 2007}, {""authors"": [""Krul"", ""Aleksandra"", ""G\u00e9raldine Damnati"", ""Fran\u00e7ois Yvon"", ""Thierry Moudenc""], ""title"": ""Corpus design based on the Kullback-Leibler divergence for Text-To-Speech synthesis application"", ""year"": 2006}, {""authors"": [""Neubig"", ""Graham"", ""Shinsuke Mori.""], ""title"": ""Word-based partial annotation for efficient corpus construction"", ""venue"": ""Proceedings of the International Conference on Language Resources and Evaluation (LREC),"", ""year"": 2010}, {""authors"": [""Raz"", ""Ran"", ""Shmuel Safra.""], ""title"": ""A sub-constant error-probability low-degree test, and a sub-constant error-probability PCP characterization of NP"", ""venue"": ""Proceedings of the twenty-ninth annual ACM symposium"", ""year"": 1997}, {""authors"": [""Rojc"", ""Matej"", ""Zdravko Ka\u010di\u010d.""], ""title"": ""Design of optimal Slovenian speech corpus for use in the concatenative speech synthesis system"", ""venue"": ""Proceedings of the International Conference on Language Resources and"", ""year"": 2000}, {""authors"": [""Schein"", ""Andrew I."", ""Ted S. Sandler"", ""Lyle H. Ungar.""], ""title"": ""Bayesian example selection using BaBiES"", ""venue"": ""Technical Report MS-CIS-04-08, Department of Computer and Information Science, University of"", ""year"": 2004}, {""authors"": [""Settles"", ""Burr.""], ""title"": ""Active learning literature survey"", ""venue"": ""Technical Report 1648, Department of Computer Sciences, University of Wisconsin, Madison."", ""year"": 2010}, {""authors"": [""Synapse.""], ""title"": ""Documentation technique: Composant d\u2019\u00e9tiquetage et lemmatisation"", ""venue"": ""http://www.synapse-fr.com/."", ""year"": 2011}, {""authors"": [""Tian"", ""Jilei"", ""Jani Nurminen.""], ""title"": ""Optimization of text database using hierachical clustering"", ""venue"": ""Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP),"", ""year"": 2009}, {""authors"": [""Tian"", ""Jilei"", ""Jani Nurminen"", ""Imre Kiss.""], ""title"": ""Optimal subset selection from text databases"", ""venue"": ""Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal"", ""year"": 2005}, {""authors"": [""Tomanek"", ""Katrin"", ""Fredrik Olsson.""], ""title"": ""A Web survey on the use of active learning to support annotation of text data"", ""venue"": ""Proceedings of the NAACL HLT 2009 Workshop on Active Learning for Natural"", ""year"": 2009}, {""authors"": [""Van Santen"", ""Jan P.H."", ""Adam L. Buchsbaum.""], ""title"": ""Methods for optimal text selection"", ""venue"": ""Proceedings of the European"", ""year"": 1997}, {""authors"": [""Rhodes. Zhang"", ""Jin-Song"", ""Satoshi Nakamura""], ""title"": ""An improved greedy search"", ""venue"": ""Conference on Speech Communication and Technology (Eurospeech),"", ""year"": 2008}]","8 conclusion :This article discussed the building of linguistic-rich corpora under an objective of parsimony. This task, a generalization of SCP, turns out to be an NP-hard problem that cannot be polynomially approximated. We studied the behavior of several algorithms in the particular domain of NLP, where the considered events follow a heavy-tailed type distribution.
The proposed algorithms have been compared through three kinds of experiments: The first one is the coverings of 2- and 3-phonemes from two text corpora, one in French, the other in English; the second one consists of the coverings of part-of-speech labels from a corpus in French; the third one evaluates the impact of both algorithms on the acoustic quality of a corpus-based TTS system. The first algorithm, ASA, is composed of an agglomerative greedy strategy followed by a spitting greedy stage. The second one, LamSCP, is based on Lagrangian relaxation principles combined with greedy strategies. LamSCP is our adaptation of an algorithm proposed in Caprara, Fischetti, and Toth (1999) to the multi-representation constraints. The comparison of SCP solutions is mainly about their size, their maximal distance with the optimal covering, and their robustness in case of perturbation of the initial corpus ordering.
Although ASA is much faster than LamSCP, it does not permit us to singlehandedly assess the quality of its solution in terms of size. The main assets of LamSCP are the calculation of a lower bound to the optimal covering size and shorter solutions than the ones obtained by ASA. Indeed, in our experiments of phonological coverings, the optimal solution is at most 1.24% (10.13%, respectively) smaller than the solutions derived by LamSCP (ASA, respectively). As for the coverings of 1-POS, LamSCP provides the optimal solution in a case of a mono-representation constraint, whereas the ASA solution is 10.17% greater than the optimal one. These relative gaps between the lower bounds and solution sizes of both algorithms generally decrease when the size of the search space decreases. Thanks to the lower bound derived by LamSCP, we empirically show that it is possible to get almost optimal solutions in a linguistic framework following Zipf’s law distribution, despite the theoretic complexity of the multi-represented SCP. Concerning the last experiment in the TTS framework, even if LamSCP provides a smaller corpus, the subjective test shows no significant difference between the TTS systems based on LamSCP and ASA corpora. Therefore, we think that ASA remains the most adequate strategy, in terms of performance, ease of development, and computation time to solve SCP in the NLP field. However, it would be interesting
to test a parallelized version of the heuristic phase that calls an important number of greedy sub-procedures.
Our future prospects for this work are in automatic language processing and speech synthesis. First, in the framework of the Phorevox project supported by the French National Research Agency, we are considering the automatic design of exercise contents for language learning by the selection of texts covering some phonological or linguistic difficulties. Secondly, this work is a preliminary step to building a phonetically rich script before its recording in order to produce a high quality speech synthesis. The covering choices, such as the attributes to cover, the number of required occurrences, or the “sentence” length (utterances, syntagms, etc.) need to be validated. Moreover, in this article, we have observed the great impact of the distribution of rare units in the corpus to reduce, and we believe it will be interesting to adapt the “sentence” granularity according to this distribution."
"1 introduction :Statistical Machine Translation (SMT) advanced near the beginning of the century from word-based models (Brown et al. 1993) towards more advanced models that take contextual information into account. Phrase-based (Koehn, Och, and Marcu 2003; Och and Ney 2004) and N-gram-based (Casacuberta and Vidal 2004; Mariño et al. 2006) models are two instances of such frameworks. Although the two models have some common properties, they are substantially different. The present work is a step towards combining the benefits and remedying the flaws of these two frameworks.
Phrase-based systems have a simple but effective mechanism that learns larger chunks of translation called bilingual phrases.1 Memorizing larger units enables the phrase-based model to learn local dependencies such as short-distance reorderings, idiomatic collocations, and insertions and deletions that are internal to the phrase pair. The model, however, has the following drawbacks: (i) it makes independence assumptions over phrases, ignoring the contextual information outside of phrases, (ii) the reordering model has difficulties in dealing with long-range reorderings, (iii) problems in both search and modeling require the use of a hard reordering limit, and (iv) it has the spurious phrasal segmentation problem, which allows multiple derivations of a bilingual sentence pair that have the same word alignment but different model scores.
N-gram-based models are Markov models over sequences of tuples that are generated monotonically. Tuples are minimal translation units (MTUs) composed of source and target cepts.2 The N-gram-based model has the following drawbacks: (i) only precalculated orderings are hypothesized during decoding, (ii) it cannot memorize and use lexical reordering triggers, (iii) it cannot perform long distance reorderings, and (iv) using tuples presents a more difficult search problem than in phrase-based SMT.
The Operation Sequence Model. In this article we present a novel model that tightly integrates translation and reordering into a single generative process. Our model explains the translation process as a linear sequence of operations that generates a source and target sentence in parallel, in a target left-to-right order. Possible operations are (i) generation of a sequence of source and target words, (ii) insertion of gaps as explicit target positions for reordering operations, and (iii) forward and backward jump operations that do the actual reordering. The probability of a sequence of operations is defined according to an N-gram model, that is, the probability of an operation depends on the n − 1 preceding operations. Because the translation (lexical generation) and reordering operations are coupled in a single generative story, the reordering decisions may depend on preceding translation decisions and translation decisions may depend
1 A Phrase pair in phrase-based SMT is a pair of sequences of words. The sequences are not necessarily linguistic constituents. Phrase pairs are built by combining minimal translation units and ordering information. As is customary we use the term phrase to refer to phrase pairs if there is no ambiguity. 2 A cept is a group of source (or target) words connected to a group of target (or source) words in a particular alignment (Brown et al. 1993).
on preceding reordering decisions. This provides a natural reordering mechanism that is able to deal with local and long-distance reorderings in a consistent way.
Like the N-gram-based SMT model, the operation sequence model (OSM) is based on minimal translation units and takes both source and target information into account. This mechanism has several useful properties. Firstly, no phrasal independence assumption is made. The model has access to both source and target context outside of phrases. Secondly the model learns a unique derivation of a bilingual sentence given its alignments, thus avoiding the spurious phrasal segmentation problem. The OSM, however, uses operation N-grams (rather than tuple N-grams), which encapsulate both translation and reordering information. This allows the OSM to use lexical triggers for reordering like phrase-based SMT. Our reordering approach is entirely different from the tuple N-gram model. We consider all possible orderings instead of a small set of POS-based pre-calculated orderings, as is used in N-gram-based SMT, which makes their approach dependent on the availability of a source and target POS-tagger. We show that despite using POS tags the reordering patterns learned by N-gram-based SMT are not as general as those learned by our model.
Combining MTU-model with Phrase-Based Decoding. Using minimal translation units makes the search much more difficult because of the poor translation coverage, inaccurate future cost estimates, and pruning of correct hypotheses because of insufficient context. The ability to memorize and produce larger translation units gives an edge to the phrase-based systems during decoding, in terms of better search performance and superior selection of translation units. In this article, we combine N-gram-based modeling with phrase-based decoding to benefit from both approaches. Our model is based on minimal translation units, but we use phrases during decoding. Through an extensive evaluation we found that this combination not only improves the search accuracy but also the BLEU scores. Our in-house phrase-based decoder outperformed state-of-the-art phrase-based (Moses and Phrasal) and N-gram-based (NCode) systems on three translation tasks.
Comparative Experiments. Motivated by these results, we integrated the OSM into the state-of-the-art phrase-based system Moses (Koehn et al. 2007). Our aim was to directly compare the performance of the lexicalized reordering model to the OSM and to see whether we can improve the performance further by using both models together. Our integration of the OSM into Moses gave a statistically significant improvement over a competitive baseline system in most cases.
In order to assess the contribution of improved reordering versus the contribution of better modeling with MTUs in the OSM-augmented Moses system, we removed the reordering operations from the stream of operations. This is equivalent to integrating the conventional N-gram tuple sequence model (Mariño et al. 2006) into a phrasebased decoder, as also tried by Niehues et al. (2011). Small gains were observed in most cases, showing that much of the improvement obtained by the OSM is due to better reordering.
Generalized Operation Sequence Model. The primary strength of the OSM over the lexicalized reordering model is its ability to take advantage of the wider contextual information. In an error analysis we found that the lexically driven OSM often falls back to very small context sizes because of data sparsity. We show that this problem can be addressed by learning operation sequences over generalized representations such as POS tags.
The article is organized into seven sections. Section 2 is devoted to a literature review. We discuss the pros and cons of the phrase-based and N-gram-based SMT frameworks in terms of both model and search. Section 3 presents our model. We
show how our model combines the benefits of both of the frameworks and removes their drawbacks. Section 4 provides an empirical evaluation of our preliminary system, which uses an MTU-based decoder, against state-of-the-art phrase-based (Moses and Phrasal) and N-gram-based (Ncode) systems on three standard tasks of translating German-to-English, Spanish-to-English, and French-to-English. Our results show improvements over the baseline systems, but we noticed that using minimal translation units during decoding makes the search problem difficult, which suggests using larger units in search. Section 5 presents an extension to our system to combine phrasebased decoding with the operation sequence model to address the problems in search. Section 5.1 empirically shows that information available in phrases can be used to improve the search performance and translation quality. Finally, we probe whether integrating our model into the phrase-based SMT framework addresses the mentioned drawbacks and improves translation quality. Section 6 provides an empirical evaluation of our integration on six standard tasks of translating German–English, French–English, and Spanish–English pairs. Our integration gives statistically significant improvements over submission quality baseline systems. Section 7 concludes.","2 previous work : The phrase-based model (Koehn et al. 2003; Och and Ney 2004) segments a bilingual sentence pair into phrases that are continuous sequences of words. These phrases are then reordered through a lexicalized reordering model that takes into account the orientation of a phrase with respect to its previous phrase (Tillmann and Zhang 2005) or block of phrases (Galley and Manning 2008). Phrase-based models memorize local dependencies such as short reorderings, translations of idioms, and the insertion and deletion of words sensitive to local context. Phrase-based systems, however, have the following drawbacks.
Handling of Non-local Dependencies. Phrase-based SMT models dependencies between words and their translations inside of a phrase well. However, dependencies across phrase boundaries are ignored because of the strong phrasal independence assumption. Consider the bilingual sentence pair shown in Figure 1(a).
Reordering of the German word stimmen is internal to the phrase-pair gegen ihre Kampagne stimmen -‘vote against your campaign’ and therefore represented by the translation model. However, the model fails to correctly translate the test sentence shown in Figure 1(b), which is translated as ‘they would for the legalization of abortion in Canada vote’, failing to displace the verb. The language model does not provide enough
evidence to counter the dispreference of the translation model against jumping over the source words für die Legalisieurung der Abtreibung in Kanada and translating stimmen - ‘vote’ at its correct position.
Weak Reordering Model. The lexicalized reordering model is primarily designed to deal with short-distance movement of phrases such as swapping two adjacent phrases and cannot properly handle long-range jumps. The model only learns an orientation of how a phrase was reordered with respect to its previous and next phrase; it makes independence assumptions over previously translated phrases and does not take into account how previous words were translated and reordered. Although such an independence assumption is useful to reduce sparsity, it is overly generalizing and does not help to disambiguate good reorderings from the bad ones.
Moreover, a vast majority of extracted phrases are singletons and the corresponding probability of orientation given phrase-pair estimates are based on a single observation. Due to sparsity, the model falls back to use one-word phrases instead, the orientation of which is ambiguous and can only be judged based on context that is ignored. This drawback has been addressed by Cherry (2013) by using sparse features for reordering models.
Hard Distortion Limit. The lexicalized reordering model fails to filter out bad largescale reorderings effectively (Koehn 2010). A hard distortion limit is therefore required during decoding in order to produce good translations. A distortion limit beyond eight words lets the translation accuracy drop because of search errors (Koehn et al. 2005). The use of a hard limit is undesirable for German–English and similar language pairs with significantly different syntactic structures. Several researchers have tried to address this problem. Moore and Quirk (2007) proposed improved future cost estimation to enable higher distortion limits in phrasal MT. Green, Galley, and Manning (2010) additionally proposed discriminative distortion models to achieve better translation accuracy than the baseline phrase-based system for a distortion limit of 15 words. Bisazza and Federico (2013) recently proposed a novel method to dynamically select which longrange reorderings to consider during the hypothesis extension process in a phrasebased decoder and showed an improvement in a German–English task by increasing the distortion limit to 18.
Spurious Phrasal Segmentation. A problem with the phrase-based model is that there is no unique correct phrasal segmentation of a sentence. Therefore, all possible ways of segmenting a bilingual sentence consistent with the word alignment are learned and used. This leads to two problems: (i) phrase frequencies are obtained by counting all possible occurrences in the training corpus, and (ii) different segmentations producing the same translation are generated during decoding. The former leads to questionable parameter estimates and the latter may lead to search errors because the probability of a translation is fragmented across different segmentations. Furthermore, the diversity in N-best translation lists is reduced. N-gram-based SMT (Mariño et al. 2006) uses an N-gram model that jointly generates the source and target strings as a sequence of bilingual translation units called tuples. Tuples are essentially minimal phrases, atomic units that cannot be decomposed any further. The tuples are generated left to right in target word order. Reordering is not
part of the statistical model. The parameters of the N-gram model are learned from bilingual data where the tuples have been arranged in target word order (see Figure 2).
Decoders for N-gram-based SMT reorder the source words in a preprocessing step so that the translation can be done monotonically. The reordering is performed with POS-based rewrite rules (see Figure 2 for an example) that have been learned from the training data (Crego and Mariño 2006). Word lattices are used to compactly represent a number of alternative reorderings. Using parts of speech instead of words in the rewrite rules makes them more general and helps to avoid data sparsity problems.
The mechanism has several useful properties. Because it is based on minimal units, there is only one derivation for each aligned bilingual sentence pair. The model therefore avoids spurious ambiguity. The model makes no phrasal independence assumption and generates a tuple monotonically by looking at a context of n previous tuples, thus capturing context across phrasal boundaries. On the other hand, N-gram-based systems have the following drawbacks.
Weak Reordering Model. The main drawback of N-gram-based SMT is its poor reordering mechanism. Firstly, by linearizing the source, N-gram-based SMT throws away useful information about how a particular word is reordered with respect to the previous word. This information is instead stored in the form of rewrite rules, which have no influence on the translation score. The model does not learn lexical reordering triggers and reorders through the learned rules only. Secondly, search is performed only on the precalculated word permutations created based on the source-side words. Often, evidence of the correct reordering is available in the translation model and the targetside language model. All potential reorderings that are not supported by the rewrite rules are pruned in the pre-processing step. To demonstrate this, consider the bilingual sentence pair in Figure 2 again. N-gram-based MT will linearize the word sequence gegen ihre Kampagne stimmen to stimmen gegen ihre Kampagne, so that it is in the same order as the English words. At the same time, it learns a POS rule: IN PRP NN VB → VB IN PRP NN. The POS-based rewrite rules serve to precompute the orderings that will be hypothesized during decoding. However, notice that this rule cannot generalize to the test sentence in Figure 1(b), even though the tuple translation model learned the trigram < sie – ‘they’ würden – ‘would’ stimmen – ‘vote’ > and it is likely that the monolingual language model has seen the trigram they would vote.
Hard Reordering Limit. Due to sparsity, only rules with seven or fewer tags are extracted. This subsequently constrains the reordering window to seven or fewer words, preventing the N-gram model from hypothesizing long-range reorderings that require
larger jumps. The need to perform long-distance reordering motivated the idea of using syntax trees (Crego and Mariño 2007) to form rewrite rules. However, the rules are still extracted ignoring the target-side, and search is performed only on the precalculated orderings.
Difficult Search Problem. Using MTUs makes the search problem much more difficult because of poor translation option selection. To illustrate this consider the phrase pair schoss ein Tor – ‘scored a goal’, consisting of units schoss – ‘scored’, ein – ‘a’, and Tor – ‘goal’. It is likely that the N-gram system does not have the tuple schoss – ‘scored’ in its N-best translation options because it is an uncommon translation. Even if schoss – ‘scored’ is hypothesized, it will be ranked quite low in the stack and may be pruned, before ein and Tor are generated in the next steps. A similar problem is also reported in Costa-jussà et al. (2007): When trying to reproduce the sentences in the N-best translation output of the phrase-based system, the N-gram-based system was able to produce only 37.5% of sentences in the Spanish-to-English and English-to-Spanish translation task, despite having been trained on the same word alignment. A phrase-based system, on the other hand, is likely to have access to the phrasal unit schoss ein Tor – ‘scored a goal’ and can generate it in a single step.","3 operation sequence model :Now we present a novel generative model that explains the translation process as a linear sequence of operations that generate a source and target sentence in parallel. Possible operations are (i) generation of a sequence of source and/or target words, (ii) insertion of gaps as explicit target positions for reordering operations, and (iii) forward and backward jump operations that do the actual reordering. The probability of a sequence of operations is defined according to an N-gram model, that is, the probability of an operation depends on the n − 1 preceding operations. Because the translation (generation) and reordering operations are coupled in a single generative story, the reordering decisions may depend on preceding translation decisions, and translation decisions may depend on preceding reordering decisions. This provides a natural reordering mechanism able to deal with local and long-distance reorderings consistently. The generative story of the model is motivated by the complex reordering in the German-to-English translation task. The English words are generated in linear order,3 and the German words are generated in parallel with their English translations. Mostly, the generation is done monotonically. Occasionally the translator inserts a gap on the German side to skip some words to be generated later. Each inserted gap acts as a designated landing site for the translator to jump back to. When the translator needs to cover the skipped words, it jumps back to one of the open gaps. After this is done, the translator jumps forward again and continues the translation. We will now, step by step, present the characteristics of the new model by means of examples.
3 Generating the English words in order is also what the decoder does when translating from German to English.
3.1.1 Basic Operations. The generation of the German–English sentence pair Peter liest – ‘Peter reads’ is straightforward because it is a simple 1-to-1 word-based translation without reordering:
Generate (Peter , Peter) Generate (liest , reads)
3.1.2 Insertions and Deletions. The translation Es ist ja nicht so schlimm – ‘it is not that bad’, requires the insertion of an additional German word ja, which is used as a discourse particle in this construction.
Generate (Es , it) Generate (ist , is) Generate Source Only (ja) Generate (nicht , not) Generate (so , that) Generate (schlimm , bad)
Conversely, the translation Lies mit – ‘Read with me’ requires the deletion of an untranslated English word me.
Generate (Lies , Read) Generate (mit , with) Generate Target Only (me)
3.1.3 Reordering. Let us now turn to an example that requires reordering, and revisit the example in Figure 1(a). The generation of this sentence in our model starts with generating sie – ‘they’, followed by the generation of würden – ‘would’. Then a gap is inserted on the German side, followed by the generation of stimmen – ‘vote’. At this point, the (partial) German and English sentences look as follows:
Operation Sequence Generation
Generate(sie, they) Generate (würden, would) sie würden stimmen ↓ Insert Gap Generate(stimmen, vote)
‘they would vote’
The arrow sign ↓ denotes the position after the previously covered German word. The translation proceeds as follows. We jump back to the open gap on the German side and fill it by generating gegen – ‘against’, Ihre – ‘your’ and Kampagne – ‘campaign’. Let us discuss some useful properties of this mechanism:
1. We have learned a reordering pattern sie würden stimmen – ‘they would vote’, which can be used to generalize the test sentence in Figure 1(b). In this case the translator jumps back and generates the tuples für – ‘for’, die – ‘the’, Legalisierung – ‘legalization’, der – ‘of’, Abtreibung – ‘abortion’, in – ‘in’, Kanada – ‘Canada’.
2. The model handles both local (Figure 1 (a)) and long-range reorderings (Figure 1 (b)) in a unified manner, regardless of how many words separate würden and stimmen.
3. Learning the operation sequence Generate(sie, they) Generate(würden, would) Insert Gap Generate(stimmen, vote) is like learning a phrase pair sie würden X stimmen – ‘they would vote’. The open gap represented by acts as a placeholder for the skipped phrases and serves a similar purpose as the non-terminal category X in a discontinuous phrase-based system.
4. The model couples lexical generation and reordering information. Translation decisions are triggered by reordering decisions and vice
versa. Notice how the reordering decision is triggered by the translation decision in the example. The probability of a gap insertion operation after the generation of the auxiliaries würden – ‘would’ will be high because reordering is necessary in order to move the second part of the German verb complex (stimmen) to its correct position at the end of the clause.
Complex reorderings can be achieved by inserting multiple gaps and/or recursively inserting a gap within a gap. Consider the generation of the example in Figure 3 (borrowed from Chiang [2007]). The generation of this bilingual sentence pair proceeds as follows:
Generate(Aozhou, Australia) Generate(shi, is) Insert Gap Generate(zhiyi, one of )
At this point, the (partial) Chinese and English sentences look like this:
Aozhou shi zhiyi ↓
Australia is one of
The translator now jumps back and recursively inserts a gap inside of the gap before continuing translation:
Jump Back (1) Insert Gap Generate(shaoshu, the few) Generate(guojia, countries)
Aozhou shi shaoshu guojia ↓ zhiyi Australia is one of the few countries
The rest of the sentence pair is generated as follows:
Jump Back (1) Insert Gap Generate(de, that) Jump Back (1) Insert Gap Generate(you, have) Generate(bangjiao, diplomatic relationships) Jump Back (1) Generate(yu, with) Generate(Beihan, North Korea)
Note that the translator jumps back and opens new gaps recursively to exhibit a property similar to the hierarchical model. However, our model uses a deterministic algorithm (see Algorithm 1 later in this article) to convert each bilingual sentence pair given the alignment to a unique derivation, thus avoiding spurious ambiguity unlike hierarchical and phrase-based models.
Multiple gaps can simultaneously exist at any time during generation. The translator decides based on the next English word to be covered which open gap to jump to. Figure 4 shows a German–English subordinate clause pair. The generation of this example is carried out as follows:
Insert Gap Generate(nicht, do not) Insert Gap Generate(wollen, want to)
At this point, the (partial) German and English sentences look as follows:
nicht wollen ↓
do not want to
The inserted gaps act as placeholders for the skipped prepositional phrase über konkrete Zahlen – ‘on specific figures’ and the verb phrase verhandeln – ‘negotiate’. When the translator decides to generate any of the skipped words, it jumps back to one of the open gaps. The Jump Back operation closes the gap that it jumps to. The translator proceeds monotonically from that point until it needs to jump again. The generation proceeds as follows:
Jump Back (1) Generate(verhandeln, negotiate)
nicht verhandeln ↓ wollen do not want to negotiate
The translation ends by jumping back to the open gap and generating the prepositional phrase as follows:
Jump Back (1) Generate(über, on) Generate(konkrete, specific) Generate(Zahlen, figures)
5. Notice that although our model is based on minimal units, we can nevertheless memorize phrases (along with reordering information) through operation subsequences that are memorized by learning an N-gram model over these operation sequences. Some interesting phrases that our model learns are:
Phrases Operation Sub-sequence
nicht X wollen – ‘do not want to’ Generate (nicht , do not) Insert Gap Generate (wollen , want to)
verhandeln wollen – ‘want to negotiate’ Insert Gap Generate (wollen , want to) Jump Back(1) Generate (verhandeln , negotiate)
X represents , the Insert Gap operation on the German side in our notation.
3.1.4 Generation of Discontinuous Source Units. Now we discuss how discontinuous source cepts can be represented in our generative model. The Insert Gap operation discussed in the previous section can also be used to generate discontinuous source cepts. The generation of any such cept is done in several steps. See the example in Figure 5. The gappy cept hat...gelesen – ‘read’ can be generated as shown.
Operation Sequence Generation
Generate(er, he) Generate (hat gelesen, read) er hat gelesen ↓ Insert Gap Continue Source Cept
he read
After the generation of er – ‘he’, the first part of the German complex verb hat is generated as an incomplete translation of ‘read’. The second part gelesen is added to a queue to be generated later. A gap is then inserted for the skipped words ein and Buch. Lastly, the second word (gelesen) of the unfinished German cept hat...gelesen is added to complete the translation of ‘read’ through a Continue Source Cept operation. Discontinuous cepts on the English side cannot be generated analogously because of the fundamental assumption of the model that English (target-side) will be generated from left to right. This is a shortcoming of our approach, which we will discuss later in Section 4.1. Our model uses five translation and three reordering operations, which are repeatedly applied in a sequence. The following is a definition of each of these operations. Generate (X,Y): X and Y are German and English cepts, respectively, each with one or more words. Words in X (German) may be consecutive or discontinuous, but the words in Y (English) must be consecutive. This operation causes the words in Y and the first word in X to be added to the English and German strings, respectively, that were generated so far. Subsequent words in X are added to a queue to be generated later. All the English words in Y are generated immediately because English (target-side) is generated in linear order as per the assumption of the model.4 The generation of the second (and subsequent) German words in a multiword cept can be delayed by gaps, jumps, and other operations defined in the following.
4 Note that when we are translating in the opposite direction (i.e., English-to-German), then German becomes target-side and is generated monotonically and gaps and jumps are performed on English (now source-side).
Continue Source Cept: The German words added to the queue by the Generate (X,Y) operation are generated by the Continue Source Cept operation. Each Continue Source Cept operation removes one German word from the queue and copies it to the German string. If X contains more than one German word, say n many, then it requires n translation operations, an initial Generate (X1...Xn, Y) operation, and n − 1 Continue Source Cept operations. For example kehrten...zurück – ‘returned’ is generated by the operation Generate (kehrten zurück, returned), which adds kehrten and ‘returned’ to the German and English strings and zurück to a queue. A Continue Source Cept operation later removes zurück from the queue and adds it to the German string.
Generate Source Only (X): The words in X are added at the current position in the German string. This operation is used to generate a German word with no coresponding English word. It is performed immediately after its preceding German word is covered. This is because there is no evidence on the English side that indicates when to generate X.5 Generate Source Only (X) helps us learn a source word deletion model. It is used during decoding, where a German word X is either translated to some English word(s) by a Generate (X,Y) operation or deleted with a Generate Source Only (X) operation.
Generate Target Only (Y): The words in Y are added at the current position in the English string. This operation is used to generate an English word with no corresponding German word. We do not utilize this operation in MTU-based decoding where it is hard to predict when to add unaligned target words during decoding. We therefore modified the alignments to remove this, by aligning unaligned target words (see Section 4.1 for details). In phrase-based decoding, however, this is not necessary, as we can easily predict unaligned target words where they are present in a phrase pair.
Generate Identical: The same word is added at the current position in both the German and English strings. The Generate Identical operation is used during decoding for the translation of unknown words. The probability of this operation is estimated from singleton German words that are translated to an identical string. For example, for a tuple QCRI – ‘QCRI’, where German QCRI was observed exactly once during training, we use a Generate Identical operation rather than Generate (QCRI, QCRI). We now discuss the set of reordering operations used by the generative story. Reordering has to be performed whenever the German word to be generated next does not immediately follow the previously generated German word. During the generation process, the translator maintains an index that specifies the position after the previously covered German word (j), an index (Z) that specifies the index after the right-most German word covered so far, and an index of the next German word to be covered (j′). The set of reordering operations used in generation depends upon these indexes. Please refer to Algorithm 1 for details.
5 We want to preserve a 1-to-1 relationship between operation sequences and aligned sentence pairs. If we allowed an unaligned source word to be generated at any time, we would obtain several operation sequences that produce the same aligned sentence pair.
Insert Gap: This operation inserts a gap, which acts as a placeholder for the skipped words. There can be more than one open gap at a time.
Jump Back (W): This operation lets the translator jump back to an open gap. It takes a parameter W specifying which gap to jump to. The Jump Back (1) operation jumps to the closest gap to Z, Jump Back (2) jumps to the second closest gap to Z, and so forth. After the backward jump, the target gap is closed.
Jump Forward: This operation makes the translator jump to Z. It is performed when the next German word to be generated is to the right of the last German word generated and does not follow it immediately. It will be followed by an Insert Gap or Jump Back (W) operation if the next source word is not at position Z. We use Algorithm 1 to convert an aligned bilingual sentence pair to a sequence of operations. Table 1 shows step by step by means of an example (Figure 6) how the conversion is done. The values of the index variables are displayed at each point.
Table 1 Step-wise generation of Example in Figure 6. The arrow indicates position j.
Figure 6 Discontinuous cept translation. Our model is estimated from a sequence of operations obtained through the transformation of a word-aligned bilingual corpus. An operation can be to generate source and target words or to perform reordering by inserting gaps and jumping forward and backward. Let O = o1, . . . , oJ be a sequence of operations as hypothesized by the translator to generate a word-aligned bilingual sentence pair< F, E, A >. The translation model is then defined as:
pT(F, E, A) = p(o1, .., oJ) = J∏
j=1
p(oj|oj−n+1...oj−1)
where n indicates the amount of context used and A defines the word-alignment function between E and F. Our translation model is implemented as an N-gram model of operations using the SRILM toolkit (Stolcke 2002) with Kneser-Ney smoothing (Kneser and Ney 1995). The translate operations in our model (the operations with a name starting with Generate) encapsulate tuples. Tuples are minimal translation units extracted from the word-aligned corpus. The idea is similar to N-gram-based SMT except that the tuples in the N-gram model are generated monotonically. We do not impose the restriction of monotonicity in our model but integrate reordering operations inside the generative model.
As in the tuple N-gram model, there is a 1-to-1 correspondence between aligned sentence pairs and operation sequences, that is, we get exactly one operation sequence per bilingual sentence given its alignments. The corpus conversion algorithm (Algorithm 1) maps each bilingual sentence pair given its alignment into a unique sequence of operations deterministically, thus maintaining a 1-to-1 correspondence. This property of the model is useful because it addresses the spurious phrasal segmentation problem in phrase-based models. A phrase-based model assigns different scores to a derivation based on which phrasal segmentation is chosen. Unlike this, the OSM assigns only one score because the model does not suffer from spurious ambiguity.
3.6.1 Discriminative Model. We use a log-linear approach (Och 2003) to make use of standard features along with several novel features that we introduce to improve endto-end accuracy. We search for a target string E that maximizes a linear combination of feature functions:
Ê = arg max E ⎧⎨ ⎩ J∑ j=1 λjhj(F, E) ⎫⎬ ⎭
where λj is the weight associated with the feature hj(F, E). Apart from the OSM and standard features such as target-side language model, length bonus, distortion limit, and IBM lexical features (Koehn, Och, and Marcu 2003), we used the following new features:
Deletion Penalty. Deleting a source word (Generate Source Only (X)) is a common operation in the generative story. Because there is no corresponding target-side word, the monolingual language model score tends to favor this operation. The deletion penalty counts the number of deleted source words.
Gap and Open Gap Count. These features are introduced to guide the reordering decisions. We observe a large amount of reordering in the automatically word aligned training text. However, given only the source sentence (and little world knowledge), it is not realistic to try to model the reasons for all of this reordering. Therefore we can use a more robust model that reorders less than humans do. The gap count feature sums to the total number of gaps inserted while producing a target sentence.
The open gap count feature is a penalty paid once for each translation operation (Generate(X,Y), Generate Identical, Generate Source Only (X)) performed whose value is the number of currently open gaps. This penalty controls how quickly gaps are closed.
Distance-Based Features. We have two distance-based features to control the reordering decisions. One of the features is the Gap Distance, which calculates the distance between the first word of a source cept X and the start of the leftmost gap. This cost is paid once for each translation operation (Generate, Generate Identical, Generate Source Only (X)). For a source cept covering the positions X1, . . . , Xn, we get the feature value gj = X1 − S, where S is the index of the left-most source word where a gap starts. Another distance-based penalty used in our model is the Source Gap Width. This feature only applies in the case of a discontinuous translation unit and computes the distance between the words of a gappy cept. Let f = f1 . . . , fi, . . . , fn be a gappy source cept where xi is the index of the ith source word in the cept f . The value of the gap-width penalty is calculated as:
wj = n∑
i=2
xi − xi−1 − 1","4 mtu-based search :We explored two decoding strategies in this work. Our first decoder complements the model and only uses minimal translation units in left-to-right stack-based decoding, similar to that used in Pharaoh (Koehn 2004a). The overall process can be roughly divided into the following steps: (i) extraction of translation units, (ii) future cost estimation, (iii) hypothesis extension, and (iv) recombination and pruning. The last two steps are repeated iteratively until all the words in the source sentence have been translated.
Our hypotheses maintain the index of the last source word covered (j), the position of the right-most source word covered so far (Z), the number of open gaps, the number of gaps so far inserted, the previously generated operations, the generated target string, and the accumulated values of all the features discussed in Section 3.6.1. The sequence of operations may include translation operations (generate, continue source cept, etc.) and reordering operations (gap insertions, jumps). Recombination6 is performed on hypotheses having the same coverage vector, monolingual language model context, and OSM context. We do histogram-based pruning, maintaining the 500 best hypotheses for each stack. A large beam size is required to cope with the search errors that result from using minimal translation units during decoding. We address this problem in Section 5.
6 Note that although we are using minimal translation units, recombination is still useful as different derivations can arise through different alignments between source and target fragments. Also, recombination can still take place if hypotheses differ slightly in the output (Koehn 2010). Aligned bilingual training corpora often contain unaligned target words and discontinuous target cepts, both of which pose problems. Unlike discontinuous source cepts, discontinuous target cepts such as hinunterschüttete – ‘poured . . . down’ in constructions like den Drink hinunterschüttete – ‘poured the drink down’ cannot be handled by the operation sequence model because it generates the English words in strict left-to-right order. Therefore they have to be eliminated.
Unaligned target words are only problematic for the MTU-based decoder, which has difficulties predicting where to insert them. Thus, we eliminate unaligned target words in MTU-based decoding.
We use a three-step process (Durrani, Schmid, and Fraser 2011) that modifies the alignments and removes unaligned and discontinuous targets. If a source word is aligned with multiple target words that are not consecutive, first the link to the least frequent target word is identified, and the group (consecutive adjacent words) of links containing this word is retained while the others are deleted. The intuition here is to keep the alignments containing content words (which are less frequent than functional words). For example, the alignment link hinunterschüttete – ‘down’ is deleted and only the link hinunterschüttete – ‘poured’ is retained because ‘down’ occurs more frequently than ‘poured’. Crego and Yvon (2009) used split tokens to deal with this phenomenon.
For MTU-based decoding we also need to deal with unaligned target words. For each unaligned target word, we determine the (left or right) neighbor that it appears more frequently with and align it with the same source word as this neighbor. Crego, de Gispert, and Mariño (2005) and Mariño et al. (2006) instead used lexical probabilities p( f |e) obtained from IBM Model 1 (Brown et al. 1993) to decide whether to attach left or right. A more sophisticated strategy based on part-of-speech entropy was proposed by Gispert and Mariño (2006). We evaluated our systems on German-to-English, French-to-English, and Spanish-toEnglish news translation for the purpose of development and evaluation. We used data from the eighth version of the Europarl Corpus and the News Commentary made available for the translation task of the Eighth Workshop on Statistical Machine Translation.7 The bilingual corpora contained roughly 2M bilingual sentence pairs, which we obtained by concatenating news commentary (≈ 184K sentences) and Europarl for the estimation of the translation model. Word alignments were generated with GIZA++ (Och and Ney 2003), using the grow-diag-final-and heuristic8 (Koehn et al. 2005). All data are lowercased, and we use the Moses tokenizer. We took news-test-2008 as the dev set for optimization and news-test 2009-2012 for testing. The feature weights are tuned with Z-MERT (Zaidan 2009).
4.2.1 Baseline Systems. We compared our system with (i) Moses9 (Koehn et al. 2007), (ii) Phrasal10 (Cer et al. 2010), and (iii) Ncode11 (Crego, Yvon, and Mariño 2011). We used
7 http://www.statmt.org/wmt13/translation-task.html 8 We also tested other symmetrization heuristics such as “Union” and “Intersection” but found the GDFA
heuristic gave best results for all language pairs. 9 http://www.statmt.org/moses/
10 http://nlp.stanford.edu/phrasal/ 11 http://www.limsi.fr/Individu/jmcrego/bincoder/
all these toolkits with their default settings. Phrasal provides two main extensions to Moses: a hierarchical reordering model (Galley and Manning 2008) and discontinuous source and target phrases (Galley and Manning 2010). We used the default stack sizes of 100 for Moses,12 200 for Phrasal, and 25 for Ncode (with 2n stacks). A 5-gram English language model is used. Both phrase-based systems use the 20 best translation options per source phrase; Ncode uses the 25 best tuple translations and a 4-gram tuple sequence model. A hard distortion limit of 6 is used in the default configuration of both phrasebased systems. Among the other defaults, we retained the hard source gap penalty of 15 and a target gap penalty of 7 in Phrasal. We provide Moses and Ncode with the same post-edited alignments13 from which we had removed target-side discontinuities. We feed the original alignments to Phrasal because of its ability to learn discontinuous source and target phrases. All the systems use MERT for the optimization of the weight vector.
4.2.2 Training. Training steps include: (i) post-editing of the alignments (Section 4.1), (ii) generation of the operation sequence (Algorithm 1), and (iii) estimation of the N-gram translation (OSM) and language models using the SRILM toolkit (Stolcke 2002) with Kneser-Ney smoothing. We used 5-gram models.
4.2.3 Summary of Developmental Experiments. During the developent of the MTU-based decoder, we performed a number of experiments to obtain optimal settings for the system. We list here a summary of the results from those experiments:
We found that discontinuous source-side cepts do not improve translation quality in most cases but increase the decoding time by multiple folds. We will therefore only use continuous cepts.
We performed experiments by varying the distortion limit from the conventional window of 6 words to infinity (= no hard limit). We found that the performance of our system is robust when removing the hard reordering constraint and even saw a slight improvement in results in the case of German-to-English systems. Using no distortion limit, however, significantly increases the decoding time. We will therefore use a window of 16 words, which we found to be optimal on the development set.
The performance of the MTU-based decoder is sensitive to the stack size. A high limit of 500 is required for decent search accuracy. We will discuss this further in the next section.
We found using 10 best translation options for each extracted cept during decoding to be optimal.
4.2.4 Comparison with the Baseline Systems. In this section we compare our system (OSMmtu) with the three baseline systems. We used Kevin Gimpel’s tester,14 which uses bootstrap resampling (Koehn 2004b) to test which of our results are significantly better than the baseline results. We mark a baseline result with “*” in order to indicate
12 Using stack sizes from 200–1,000 did not improve results. 13 Using post-processed alignments gave better results than using the original alignments for these baseline
systems. 14 http://www.ark.cs.cmu.edu/MT/
that our model shows a significant improvement over this baseline with a confidence of p < 0.05. We use 1,000 samples during bootstrap resampling.
Our German-to-English results (see Table 2) are significantly better than the baseline systems in most cases. Our French-to-English results show a significant improvement over Moses in three out of four cases, and over Phrasal in half of the cases. The N-gram-based system NCode was better or similar to our system on the French task. Our Spanish-to-English system also showed roughly the same translation quality as the baseline systems, but was significantly worse on the WMT12 task.","5 phrase-based search :The MTU-based decoder is the most straightforward implementation of a decoder for the operation sequence model, but it faces search problems that cause a drop in translation accuracy. Although the OSM captures both source and target contexts and provides a better reordering mechanism, the ability to memorize and produce larger translation units gives an edge to the phrase-based model during decoding in terms of better search performance and superior selection of translation units. In this section, we combine N-gram-based modeling with phrase-based decoding. This combination not only improves search accuracy but also increases translation quality in terms of BLEU.
The operation sequence model, although based on minimal translation units, can learn larger translation chunks by memorizing a sequence of operations. However, it often has difficulties to produce the same translations as the phrase-based system because of the following drawbacks of MTU-based decoding: (i) the MTU-based decoder does not have access to all the translation units that a phrase-based decoder uses as part of a larger phrase, (ii) it requires a larger beam size to prevent early pruning of correct
hypotheses, and (iii) it uses less-powerful future-cost estimates than the phrase-based decoder. To demonstrate these problems, consider the phrase pair
which the model memorizes through the sequence:
Generate(Wie, What is) Insert Gap Generate (Sie, your) Jump Back (1) Generate (heissen, name)
The MTU-based decoder needs three separate tuple translations to generate the same phrasal translation: Wie – ‘What is’, Sie – ‘your’ and heißen – ‘name’. Here we are faced with three challenges.
Translation Coverage: The first problem is that the N-gram model does not have the same coverage of translation options. The English cepts ‘What is’, ‘your’, and ‘name’ are not good candidate translations for the German cepts Wie, Sie, and heißen, which are usually translated to ‘How’, ‘you’, and ‘call’, respectively, in isolation. When extracting tuple translations for these cepts from the Europarl data for our system, the tuple Wie – ‘What is’ is ranked 124th, heißen – ‘name’ is ranked 56th, and Sie – ‘your’ is ranked 9th in the list of n-best translation candidates. Typically, only the 20 best translation options are used, for the sake of efficiency, and such phrasal units with less frequent translations are never hypothesized in the N-gram-based systems. The phrase-based system, on the other hand, can extract the phrase Wie heißen Sie – ‘what is your name’ even if it is observed only once during training.
Larger Beam Size: Even when we allow a huge number of translation options and therefore hypothesize such units, we are faced with another challenge. A larger beam size is required in MTU-based decoding to prevent uncommon translations from getting pruned. The phrase-based system can generate the phrase pair Wie heißen Sie – ‘what is your name’ in a single step, placing it directly into the stack three words to the right. The MTU-based decoder generates this phrase in three stacks with the tuple translations Wie – ‘What is’, Sie – ‘your’, and heißen – ‘name’. A very large stack size is required during decoding to prevent the pruning of Wie – ‘What is’, which is ranked quite low in the stack until the tuple Sie – ‘your’ is hypothesized in the next stack. Although the translation quality achieved by phrase-based SMT remains the same when varying the beam size, the performance of our system varies drastically with different beam sizes (especially for the German–English experiments where the search is more difficult due to a higher number of reorderings). Costa-jussà et al. (2007) also report a significant drop in the performance of N-gram-based SMT when a beam size of 10 is used instead of 50 in their experiments.
Future Cost Estimation: A third problem is caused by inaccurate future cost estimation. Using phrases helps phrase-based SMT to better estimate the future language model cost because of the larger context available, and allows the decoder to capture local (phrase-internal) reorderings in the future cost. In comparison, the future cost for tuples is based on unigram probabilities. The future cost estimate for the phrase pair Wie heißen Sie – ‘What is your name’ is estimated by calculating the cost of each feature.
A bigram language model cost, for example, is estimated in the phrase-based system as follows:
plm = p(What) × p(is|What) × p(your|What is) × p(name|What is your)
The translation model cost is estimated as:
ptm = p(What is your name|Wie heißen Sie)
Phrase-based SMT is aware during the preprocessing step that the words Wie heißen Sie may be translated as a phrase. This is helpful for estimating a more accurate future cost because the context is already available. The same is not true for the MTU-based decoder, to which only minimal units are available. The MTU-based decoder does not have the information that Wie heißen Sie may be translated as a phrase during decoding. The future cost estimate available to the operation sequence model for the span covering Wie heißen Sie will have unigram probabilities for both the translation and language models.
plm = p(What) × p(is|What) × p(your) × p(name)
The translation model cost is estimated as:
ptm = p(Generate(Wie, What is)) × p(Generate(heißen,name)) × p(Generate(Sie, your))
A more accurate future cost estimate for the translation model cost would be:
ptm = p(Generate(Wie,What is)) × p(Insert Gap|C2) × p(Generate(Sie,your)|C3) × p(Jump Back(1)|C4) × p(Generate(heißen,name)|C5)
where Ci is the context for the generation of the ith operation—that is, up to m previous operations. For example C1 = Generate(Wie, What is), C2 = Generate(Wie,What is) Insert Gap, and so on. The future cost estimates computed in this manner are much more accurate because not only do they consider context, but they also take the reordering operations into account (Durrani, Fraser, and Schmid 2013). We extended our in-house OSM decoder to use phrases instead of MTUs during decoding. In order to check whether phrase-based decoding solves the mentioned problems and improves the search accuracy, we evaluated the baseline MTU decoder and the phrase-based decoder with the same model parameters and tuned weights. This allows us to directly compare the model scores. We tuned the feature weights by running MERT with the MTU decoder on the dev set. Table 3 shows results from running both, the MTU-based (OSMmtu) and the phrase-based (OSMphr) decoder, on the WMT09 test set. Improved search accuracy is the percentage of times each decoder was able to produce a better model score than the other. Our phrase-based decoder uses a stack size of 200. Table 3 shows the percentage of times the MTU-based and phrase-based decoder produce better model scores than their counterpart. It shows that the phrase-based decoder produces better model scores for almost 48% of the hypotheses (on average)
across the three language pairs, whereas the MTU-based decoder (using a much higher stack size [500]) produces better hypotheses 8.2% of the time on average.
This improvement in search is also reflected in translation quality. Our phrase-based decoder outperforms the MTU-based decoder in all the cases and gives a significant improvement in 8 out of 12 cases (Table 4). In Section 4.1 we discussed the problem of handling unaligned and discontinuous target words in MTU-based decoding. An advantage of phrase-based decoding is that we can use such units during decoding if they appear within the extracted phrases. We use a Generate Target Only (Y) operation whenever the unaligned target word Y occurs in
a phrase. Similarly, we use the operation Generate (hinunterschüttete, poured down) when the discontinuous tuple hinunterschüttete – ‘poured ... down’ occurs in a phrase. While training the model, we simply ignore the discontinuity and pretend that the word ‘down’ immediately follows ‘poured’. This can be done by linearizing the subsequent parts of discontinuous target cepts to appear after the first word of the cept. During decoding we use phrase-internal alignments to hypothesize such a linearization. This is done only for the estimation of the OSM, and the target for all other purposes is generated in its original order. This heuristic allows us to deal with target discontinuities without extending the operation sequence model in complicated ways. It results in better BLEU accuracy in comparison with the post-editing of the alignments method described in Section 4.1. For details and empirical results refer to Durrani et al. (2013a) (see Table 2 therein, compare Rows 4 and 5).
Note that the OSM, like the discontinuous phrase-based model (Galley and Manning 2010), allows all possible geometries as shown in Figure 7. However, because our decoder only uses continuous phrases, we cannot hypothesize (ii) and (iii) unless they appear inside of a phrase. But our model could be integrated into a discontinuous phrase-based system to overcome this limitation.","6 further comparative experiments :Our model, like the reordering models (Tillmann and Zhang 2005; Galley and Manning 2008) used in phrase-based decoders, is lexicalized. However, our model has richer conditioning as it considers both translation and reordering context across phrasal boundaries. The lexicalized reordering model used in phrase-based SMT only accounts for how a phrase pair was reordered with respect to its previous phrase (or block of phrases). Although such an independence assumption is useful to reduce sparsity, it is overgeneralizing, with only three possible orientations. Moreover, because most of the extracted phrases are observed only once, the corresponding probability of orientation given phrase-pair estimates is very sparse. The model often has to fall back to short oneword phrases. However, most short phrases are observed frequently with all possible orientations during training. This makes it difficult for the decoder to decide which orientation should be picked during decoding. The model therefore overly relies on the language model to break such ties. The OSM may also suffer from data sparsity and the back-off smoothing may fall back to very short contexts. But it might still be able to disambiguate better than the lexicalized reordering models. Also these drawbacks can be addressed by learning an OSM over generalized word representation such as POS tags, as we show in this section.
In an effort to make a comparison of the operation sequence model with the lexicalized reordering model, we incorporate the OSM into the phrase-based Moses decoder. This allows us to exactly compare the two models in identical settings. We integrate the OSM into the hypothesis extension process of the phrase-based decoder. We convert
each phrase pair into a sequence of operations by extracting the MTUs within the phrase pair and using phrase internal alignments. The OSM is used as a feature in the log-linear framework. We also use four supportive features: the Gap, Open Gap, Gap-distance, and Deletion counts, as described earlier (see Section 3.6.1). Our Moses (Koehn et al. 2007) baseline systems are based on the setup described in Durrani et al. (2013b). We trained our systems with the following settings: maximum sentence length 80, grow-diag-final and symmetrization of GIZA++ alignments, an interpolated Kneser-Ney smoothed 5-gram language model with KenLM (Heafield 2011) used at runtime, distortion limit of 6, minimum Bayes-risk decoding (Kumar and Byrne 2004), cube pruning (Huang and Chiang 2007), and the no-reordering-over-punctuation heuristic. We used factored models (Koehn and Hoang 2007), for German–English and English–German. We trained the lexicalized reordering model (Koehn et al. 2005) with msd-bidirectional-fe settings. Table 5 shows that the OSM results in higher gains than the lexicalized reordering model on top of a plain phrase-based baseline (Pb). The average improvement obtained using the lexicalized reordering model (Pblex) over the baseline (Pb) is 0.50. In comparison, the average improvement obtained by using the OSM (Pbosm) over the baseline (Pb) is 0.74. The average improvement obtained by the combination (Pblex+osm) is 0.97. The average improvement obtained by adding the OSM over the baseline (Pblex) is 0.47. We tested for significance and found that in seven out of eight cases adding the OSM on top of Pblex gives a statistically significant improvement with a confidence of p < 0.05. Significant differences are marked with an asterisk. In an additional experiment, we studied how much the translation quality decreases when all reordering operations are removed from the operation sequence model during
training and decoding. The resulting model is similar to the tuple sequence model (TSM) of Mariño et al. (2006), except that we use phrase-internal reordering rather than POS-based rewrite rules to do the source linearization. Table 6 shows an average improvement of just 0.13 on top of the baseline phrase-based system with lexicalized reordering, which is much lower than the 0.46 points obtained with the full operation sequence model.
Bilingual translation models (without reordering) have been integrated into phrase-based systems before, either inside the decoder (Niehues et al. 2011) or to rerank the N-best candidate translations in the output of a phrase-based system (Zhang et al. 2013). Both groups reported improvements of similar magnitude when using a targetorder left-to-right TSM model for German–English and French–English translation with shared task data, but higher gains on other data sets and language pairs. Zhang et al. (2013) showed further gains by combining models with target and source left-to-right and right-to-left orders. The assumption of generating the target in monotonic order is a weakness of our work that can be addressed following Zhang et al. (2013). By generating MTUs in source order and allowing gaps and jumps on the target side, the model will be able to learn other reordering patterns that are ignored by the standard OSM. Because of data sparsity, it is impossible to observe all possible reordering patterns with all possible lexical choices in translation operations. The lexically driven OSM therefore often backs off to very small context sizes. Consider the example shown in Figure 1. The learned pattern sie würden stimmen – ‘they would vote’ cannot be generalized to er würde wählen – ‘he would vote’. We found that the OSM uses only two preceding operations as context on average. This problem can be addressed by replacing words with POS tags (or any other generalized representation such as Morph tags, word clusters) to allow the model to consider a wider syntactic context where this is appropriate, thus improving lexical decisions and the reordering capability of the model. Crego and Yvon (2010) and Niehues et al. (2011) have shown improvements in
translation quality when using a TSM model over POS units. We estimate OSMs over generalized tags and add these as separate features to the loglinear framework.15
Experiments. We enabled factored sequence models (Koehn and Hoang 2007) in German–English language pairs as these have been shown to be useful previously. We used LoPar (Schmid 2000) to obtain morphological analysis and POS annotation of German and MXPOST (Ratnaparkhi 1998), a maximum entropy model for English POS tags. We simply estimate OSMs over POS tags16 by replacing the words by the corresponding tags during training.
Table 7 shows that a system with an additional POS-based OSM (Pblex+osm(s)+osm(p)) gives an average improvement of +0.26 over the baseline (Pblex+osm(s)) system that uses an OSM over surface forms only. The overall gain by using OSMs over the baseline system is +0.70. OSM over surface tags considers 3-gram on average, and OSM over POS tags considers 4.5-grams on average, thus considering wider contextual information when making translation and reordering decisions. Table 8 shows the wall-clock decoding time (in minutes) from running the Moses decoder (on news-test2013) with and without the OSMs. Each decoder is run with 24 threads on a machine with 140GB RAM and 24 processors. Timings vary between experiments because of the fact that machines were somewhat busy in some cases. But generally, the OSM increases decoding time by more than half an hour.17
Table 9 shows the overall sizes of phrase-based translation and reordering models along with the OSMs. It also shows the model sizes when filtered on news-test2013. A similar amount of reduction could be achieved by applying filtering to the OSMs following the language model filtering described by Heafield and Lavie (2010).
15 We also tried to amalgamate lexically driven OSM and generalized OSMs into a single model rather than using these as separate features. However, this attempt was unsuccessful (See Durrani et al. [2014] for details). 16 We also found using morphological tags and automatic word clusters to be useful in our recent IWSLT evaluation campaign (Birch, Durrani, and Koehn 2013; Durrani et al. 2014). 17 The code for the OSM in Moses can be greatly optimized but requires major modifications to source and target phrase classes in Moses.","7 conclusion :In this article we presented a new model for statistical MT that combines the benefits of two state-of-the-art SMT frameworks, namely, N-gram-based and phrase-based SMT. Like the N-gram-based model, it addresses two drawbacks of phrasal MT by better handling dependencies across phrase boundaries, and solving the phrasal segmentation problem. In contrast to N-gram-based MT, our model has a generative story that tightly couples translation and reordering. Furthermore, it is able to consider all possible reorderings, unlike N-gram systems that perform search only on a limited number of pre-calculated orderings. Our model is able to correctly reorder words across large distances, and it memorizes frequent phrasal translations including their reordering as probable operation sequences.
We tested a version of our system that decodes based on minimal translation units (MTUs) against the state-of-the-art phrase-based systems Moses and Phrasal and the N-gram-based system Ncode for German-to-English, French-to-English, and Spanishto-English on three standard test sets. Our system shows statistically significant improvements in 9 out of 12 cases in the German-to-English translation task, and 10 out of 12 cases in the French-to-English translation task. Our Spanish-to-English results are similar to the baseline systems in most of the cases but consistently worse than Ncode.
MTU-based decoding suffers from poor translation coverage, inaccurate future cost estimates, and pruning of correct hypotheses. Phrase-based SMT, on the other hand, avoids these drawbacks by using larger translation chunks during search. We
therefore extended our decoder to use phrases instead of cepts while keeping the statistical model unchanged. We found that combining a model based on minimal units with phrase-based decoding improves both search accuracy and translation quality. Our system extended with phrase-based decoding showed improvements over all the baseline systems, including our MTU-based decoder. In most of the cases, the difference was significant.
Our results show that OSM consistently outperforms the Moses lexicalized reordering model and gives statistically significant gains over a very competitive Moses baseline system. We showed that considering both translation and reordering context is important and ignoring reordering context results in a significant reduction in the performance. We also showed that an OSM based on surface forms suffers from data sparsity and that an OSM based on a generalized representation with part-of-speech tags improves the translation quality by considering a larger context. In the future we would like to study whether the insight of using minimal units for modeling and search based on composed rules would hold for hierarchical SMT. Vaswani et al. (2011) recently showed that a Markov model over the derivation history of minimal rules can obtain the same translation quality as using grammars formed with composed rules, which we believe is quite promising.",,"In this article, we present a novel machine translation model, the Operation Sequence Model (OSM), which combines the benefits of phrase-based and N-gram-based statistical machine translation (SMT) and remedies their drawbacks. The model represents the translation process as a linear sequence of operations. The sequence includes not only translation operations but also reordering operations. As in N-gram-based SMT, the model is: (i) based on minimal translation units, (ii) takes both source and target information into account, (iii) does not make a phrasal independence assumption, and (iv) avoids the spurious phrasal segmentation problem. As in phrase-based SMT, the model (i) has the ability to memorize lexical reordering triggers, (ii) builds the search graph dynamically, and (iii) decodes with large translation units during search. The unique properties of the model are (i) its strong coupling of reordering and translation where translation and reordering decisions are conditioned on n previous translation and reordering decisions, and (ii) the ability to model local and long-range reorderings consistently. Using BLEU as a metric of translation accuracy, we found that our system performs significantly","[{""affiliations"": [], ""name"": ""Nadir Durrani""}, {""affiliations"": [], ""name"": ""QCRI Qatar""}, {""affiliations"": [], ""name"": ""Helmut Schmid""}, {""affiliations"": [], ""name"": ""LMU Munich""}, {""affiliations"": [], ""name"": ""Alexander Fraser""}, {""affiliations"": [], ""name"": ""Philipp Koehn""}, {""affiliations"": [], ""name"": ""Hinrich Sch\u00fctze""}]",SP:314417703be66faa0d2a7d4d072ca98d6ddf9791,"[{""authors"": [""Birch"", ""Alexandra"", ""Nadir Durrani"", ""Philipp Koehn.""], ""title"": ""Edinburgh SLT and MT System Description for the IWSLT 2013 Evaluation"", ""venue"": ""Proceedings of the 10th International Workshop on Spoken Language"", ""year"": 2013}, {""authors"": [""Bisazza"", ""Arianna"", ""Marcello Federico.""], ""title"": ""Efficient Solutions for Word Reordering in German-English Phrase-Based Statistical Machine Translation"", ""venue"": ""Proceedings of the Eighth"", ""year"": 2013}, {""authors"": [""Brown"", ""Peter F"", ""Stephen A. Della Pietra"", ""Vincent J. Della Pietra"", ""R.L. Mercer""], ""title"": ""The Mathematics of Statistical Machine Translation: Parameter"", ""year"": 1993}, {""authors"": [""Casacuberta"", ""Francisco"", ""Enrique Vidal.""], ""title"": ""Machine Translation with Inferred Stochastic Finite-State Transducers"", ""venue"": ""Computational Linguistics, 30:205\u2013225."", ""year"": 2004}, {""authors"": [""Cer"", ""Daniel"", ""Michel Galley"", ""Daniel Jurafsky"", ""Christopher D. Manning.""], ""title"": ""Phrasal: A Statistical Machine Translation Toolkit for Exploring New Model Features"", ""venue"": ""Proceedings of the North American Chapter"", ""year"": 2010}, {""authors"": [""Cherry"", ""Colin.""], ""title"": ""Improved Reordering for Phrase-Based Translation Using Sparse Features"", ""venue"": ""Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics:"", ""year"": 2013}, {""authors"": [""Chiang"", ""David.""], ""title"": ""Hierarchical Phrase-Based Translation"", ""venue"": ""Computational Linguistics, 33(2):201\u2013228."", ""year"": 2007}, {""authors"": [""Costa-juss\u00e0"", ""Marta R"", ""Josep M. Crego"", ""David Vilar"", ""Jos\u00e9 A.R. Fonollosa"", ""Jos\u00e9 B. Mari\u00f1o"", ""Hermann Ney""], ""title"": ""Analysis and System Combination of Phrase- and N-Gram-Based Statistical Machine"", ""year"": 2007}, {""authors"": [""Crego"", ""Josep M."", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Improving Statistical MT by Coupling Reordering and Decoding"", ""venue"": ""Machine Translation, 20(3):199\u2013215."", ""year"": 2006}, {""authors"": [""Crego"", ""Josep M."", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Syntax-Enhanced N-gram-Based SMT"", ""venue"": ""Proceedings of the 11th Machine Translation Summit, pages 111\u2013118, Copenhagen."", ""year"": 2007}, {""authors"": [""Crego"", ""Josep M."", ""Fran\u00e7ois Yvon.""], ""title"": ""Gappy Translation Units under Left-to-Right SMT Decoding"", ""venue"": ""Proceedings of the Meeting of the European Association for Machine Translation,"", ""year"": 2009}, {""authors"": [""Crego"", ""Josep M."", ""Fran\u00e7ois Yvon.""], ""title"": ""Improving Reordering with Linguistically Informed Bilingual N-Grams"", ""venue"": ""COLING 2010: Posters, pages 197\u2013205, Beijing."", ""year"": 2010}, {""authors"": [""Crego"", ""Josep M."", ""Fran\u00e7ois Yvon"", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Ncode: an Open Source Bilingual N-gram SMT Toolkit"", ""venue"": ""The Prague Bulletin of Mathematical Linguistics, 96:49\u201358."", ""year"": 2011}, {""authors"": [""Durrani"", ""Nadir"", ""Alexander Fraser"", ""Helmut Schmid.""], ""title"": ""Model With Minimal Translation Units, But Decode With Phrases"", ""venue"": ""the 2013 Conference of the North American Chapter of the Association for"", ""year"": 2013}, {""authors"": [""Durrani"", ""Nadir"", ""Alexander Fraser"", ""Helmut Schmid"", ""Hieu Hoang"", ""Philipp Koehn""], ""title"": ""Can Markov Models Over Minimal Translation Units Help Phrase-Based SMT"", ""venue"": ""In Proceedings of the 51st Annual Meeting"", ""year"": 2013}, {""authors"": [""Durrani"", ""Nadir"", ""Barry Haddow"", ""Kenneth Heafield"", ""Philipp Koehn.""], ""title"": ""Edinburgh\u2019s Machine Translation Systems for European Language Pairs"", ""venue"": ""Proceedings of the Eighth Workshop on"", ""year"": 2013}, {""authors"": [""Durrani"", ""Nadir"", ""Philipp Koehn"", ""Helmut Schmid"", ""Alexander Fraser.""], ""title"": ""Investigating the Usefulness of Generalized Word Representations in SMT"", ""venue"": ""Proceedings of the 25th Annual"", ""year"": 2014}, {""authors"": [""Durrani"", ""Nadir"", ""Helmut Schmid"", ""Alexander Fraser""], ""title"": ""A Joint Sequence Translation Model with Integrated"", ""year"": 2011}, {""authors"": [""Galley"", ""Michel"", ""Christopher D. Manning.""], ""title"": ""A Simple and Effective Hierarchical Phrase Reordering Model"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 848\u2013856,"", ""year"": 2008}, {""authors"": [""Galley"", ""Michel"", ""Christopher D. Manning.""], ""title"": ""Accurate Non-Hierarchical Phrase-Based Translation"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of"", ""year"": 2010}, {""authors"": [""Gispert"", ""Adri\u00e0"", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Linguistic Tuple Segmentation in N-Gram-Based Statistical Machine Translation"", ""venue"": ""INTERSPEECH, pages 1,149\u20131,152, Pittsburgh, PA."", ""year"": 2006}, {""authors"": [""Green"", ""Spence"", ""Michel Galley"", ""Christopher D. Manning.""], ""title"": ""Improved Models of Distortion Cost for Statistical Machine Translation"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference"", ""year"": 2010}, {""authors"": [""Heafield"", ""Kenneth.""], ""title"": ""KenLM: Faster and Smaller Language Model Queries"", ""venue"": ""Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187\u2013197, Edinburgh."", ""year"": 2011}, {""authors"": [""Heafield"", ""Kenneth"", ""Alon Lavie.""], ""title"": ""Combining Machine Translation Output with Open Source: The Carnegie Mellon Multi-Engine Machine Translation Scheme"", ""venue"": ""The Prague Bulletin of Mathematical"", ""year"": 2010}, {""authors"": [""Huang"", ""Liang"", ""David Chiang.""], ""title"": ""Forest Rescoring: Faster Decoding with Integrated Language Models"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics,"", ""year"": 2007}, {""authors"": [""Kneser"", ""Reinhard"", ""Hermann Ney.""], ""title"": ""Improved Backing-off for M-gram Language Modeling"", ""venue"": ""Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 181\u2013184."", ""year"": 1995}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Pharaoh: A Beam Search Decoder for Phrase-Based Statistical Machine Translation Models"", ""venue"": ""Association for Machine Translation in the Americas, pages 115\u2013124,"", ""year"": 2004}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Statistical Significance Tests for Machine Translation Evaluation"", ""venue"": ""Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388\u2013395, Barcelona."", ""year"": 2004}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Statistical Machine Translation"", ""venue"": ""Cambridge University Press."", ""year"": 2010}, {""authors"": [""Koehn"", ""Philipp"", ""Amittai Axelrod"", ""Alexandra Birch"", ""Chris Callison-Burch"", ""Miles Osborne"", ""David Talbot.""], ""title"": ""Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation"", ""venue"": ""In"", ""year"": 2005}, {""authors"": [""Koehn"", ""Philipp"", ""Hieu Hoang.""], ""title"": ""Factored Translation Models"", ""venue"": ""Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural"", ""year"": 2007}, {""authors"": [""Constantin"", ""Evan Herbst.""], ""title"": ""Moses: Open Source Toolkit for Statistical Machine Translation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics: Demonstrations."", ""year"": 2007}, {""authors"": [""Koehn"", ""Philipp"", ""Franz J. Och"", ""Daniel Marcu.""], ""title"": ""Statistical Phrase-Based Translation"", ""venue"": ""2003 Meeting of the North American Chapter of the Association for Computational Linguistics, pages 127\u2013133,"", ""year"": 2003}, {""authors"": [""Kumar"", ""Shankar"", ""William J. Byrne.""], ""title"": ""Minimum Bayes-Risk Decoding for Statistical Machine Translation"", ""venue"": ""Human Language Technologies: The 2004 Annual Conference of the North American Chapter of"", ""year"": 2004}, {""authors"": [""Mari\u00f1o"", ""Jos\u00e9 B."", ""Rafael E. Banchs"", ""Josep M. Crego"", ""Adri\u00e0 de Gispert"", ""Patrik Lambert"", ""Jos\u00e9 A.R. Fonollosa"", ""Marta R. Costa-juss\u00e0.""], ""title"": ""N-gram-Based Machine Translation"", ""venue"": ""Computational Linguistics,"", ""year"": 2006}, {""authors"": [""Moore"", ""Robert"", ""Chris Quirk.""], ""title"": ""Faster Beam Search Decoding for Phrasal Statistical Machine Translation"", ""venue"": ""Proceedings of the 11th Machine Translation Summit, Copenhagen."", ""year"": 2007}, {""authors"": [""Niehues"", ""Jan"", ""Teresa Herrmann"", ""Stephan Vogel"", ""Alex Waibel.""], ""title"": ""Wider Context by Using Bilingual Language Models in Machine Translation"", ""venue"": ""Proceedings of the Sixth Workshop on"", ""year"": 2011}, {""authors"": [""Och"", ""Franz J.""], ""title"": ""Minimum Error Rate Training in Statistical Machine Translation"", ""venue"": ""Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 160\u2013167, Sapporo."", ""year"": 2003}, {""authors"": [""Och"", ""Franz J."", ""Hermann Ney.""], ""title"": ""A Systematic Comparison of Various Statistical Alignment Models"", ""venue"": ""Computational Linguistics, 29(1):19\u201351."", ""year"": 2003}, {""authors"": [""Och"", ""Franz J."", ""Hermann Ney.""], ""title"": ""The Alignment Template Approach to Statistical Machine Translation"", ""venue"": ""Computational Linguistics, 30(1):417\u2013449."", ""year"": 2004}, {""authors"": [""Ratnaparkhi"", ""Adwait.""], ""title"": ""Maximum Entropy Models for Natural Language Ambiguity Resolution"", ""venue"": ""Ph.D. thesis, University of Pennsylvania, Philadelphia, PA."", ""year"": 1998}, {""authors"": [""Schmid"", ""Helmut.""], ""title"": ""Lopar: Design and implementation"", ""venue"": ""Bericht des sonderforschungsbereiches \u201csprachtheoretische grundlagen f\u00fcr die computerlinguistik.\u201d Technical"", ""year"": 2000}, {""authors"": [""Stolcke"", ""Andreas.""], ""title"": ""SRILM - An Extensible Language Modeling Toolkit"", ""venue"": ""International Conference on Spoken Language Processing, Denver, CO."", ""year"": 2002}, {""authors"": [""Tillmann"", ""Christoph"", ""Tong Zhang.""], ""title"": ""A Localized Prediction Model for Statistical Machine Translation"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2005}, {""authors"": [""Vaswani"", ""Ashish"", ""Haitao Mi"", ""Liang Huang"", ""David Chiang.""], ""title"": ""Rule Markov Models for Fast Tree-to-String Translation"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational"", ""year"": 2011}, {""authors"": [""Zaidan"", ""Omar F.""], ""title"": ""Z-MERT: A Fully Configurable Open Source Tool for Minimum Error Rate Training of Machine Translation Systems"", ""venue"": ""The Prague Bulletin of Mathematical Linguistics,"", ""year"": 2009}, {""authors"": [""Zhang"", ""Hui"", ""Kristina Toutanova"", ""Chris Quirk"", ""Jianfeng Gao.""], ""title"": ""Beyond Left-to-Right: Multiple Decomposition Structures for SMT"", ""venue"": ""the 2013 Conference of the North American Chapter of the"", ""year"": 2013}]",acknowledgments  :We would like to thank the anonymous reviewers and Andreas Maletti and François Yvon for their helpful feedback and suggestions. The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreements 287658 (EU-Bridge) and 287688 (MateCat). Alexander Fraser was funded by Deutsche Forschungsgemeinschaft grant Models of Morphosyntax for Statistical Machine Translation. Helmut Schmid was supported by Deutsche Forschungsgemeinschaft grant SFB 732. This publication only reflects the authors’ views.,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :Statistical Machine Translation (SMT) advanced near the beginning of the century from word-based models (Brown et al. 1993) towards more advanced models that take contextual information into account. Phrase-based (Koehn, Och, and Marcu 2003; Och and Ney 2004) and N-gram-based (Casacuberta and Vidal 2004; Mariño et al. 2006) models are two instances of such frameworks. Although the two models have some common properties, they are substantially different. The present work is a step towards combining the benefits and remedying the flaws of these two frameworks.
Phrase-based systems have a simple but effective mechanism that learns larger chunks of translation called bilingual phrases.1 Memorizing larger units enables the phrase-based model to learn local dependencies such as short-distance reorderings, idiomatic collocations, and insertions and deletions that are internal to the phrase pair. The model, however, has the following drawbacks: (i) it makes independence assumptions over phrases, ignoring the contextual information outside of phrases, (ii) the reordering model has difficulties in dealing with long-range reorderings, (iii) problems in both search and modeling require the use of a hard reordering limit, and (iv) it has the spurious phrasal segmentation problem, which allows multiple derivations of a bilingual sentence pair that have the same word alignment but different model scores.
N-gram-based models are Markov models over sequences of tuples that are generated monotonically. Tuples are minimal translation units (MTUs) composed of source and target cepts.2 The N-gram-based model has the following drawbacks: (i) only precalculated orderings are hypothesized during decoding, (ii) it cannot memorize and use lexical reordering triggers, (iii) it cannot perform long distance reorderings, and (iv) using tuples presents a more difficult search problem than in phrase-based SMT.
The Operation Sequence Model. In this article we present a novel model that tightly integrates translation and reordering into a single generative process. Our model explains the translation process as a linear sequence of operations that generates a source and target sentence in parallel, in a target left-to-right order. Possible operations are (i) generation of a sequence of source and target words, (ii) insertion of gaps as explicit target positions for reordering operations, and (iii) forward and backward jump operations that do the actual reordering. The probability of a sequence of operations is defined according to an N-gram model, that is, the probability of an operation depends on the n − 1 preceding operations. Because the translation (lexical generation) and reordering operations are coupled in a single generative story, the reordering decisions may depend on preceding translation decisions and translation decisions may depend
1 A Phrase pair in phrase-based SMT is a pair of sequences of words. The sequences are not necessarily linguistic constituents. Phrase pairs are built by combining minimal translation units and ordering information. As is customary we use the term phrase to refer to phrase pairs if there is no ambiguity. 2 A cept is a group of source (or target) words connected to a group of target (or source) words in a particular alignment (Brown et al. 1993).
on preceding reordering decisions. This provides a natural reordering mechanism that is able to deal with local and long-distance reorderings in a consistent way.
Like the N-gram-based SMT model, the operation sequence model (OSM) is based on minimal translation units and takes both source and target information into account. This mechanism has several useful properties. Firstly, no phrasal independence assumption is made. The model has access to both source and target context outside of phrases. Secondly the model learns a unique derivation of a bilingual sentence given its alignments, thus avoiding the spurious phrasal segmentation problem. The OSM, however, uses operation N-grams (rather than tuple N-grams), which encapsulate both translation and reordering information. This allows the OSM to use lexical triggers for reordering like phrase-based SMT. Our reordering approach is entirely different from the tuple N-gram model. We consider all possible orderings instead of a small set of POS-based pre-calculated orderings, as is used in N-gram-based SMT, which makes their approach dependent on the availability of a source and target POS-tagger. We show that despite using POS tags the reordering patterns learned by N-gram-based SMT are not as general as those learned by our model.
Combining MTU-model with Phrase-Based Decoding. Using minimal translation units makes the search much more difficult because of the poor translation coverage, inaccurate future cost estimates, and pruning of correct hypotheses because of insufficient context. The ability to memorize and produce larger translation units gives an edge to the phrase-based systems during decoding, in terms of better search performance and superior selection of translation units. In this article, we combine N-gram-based modeling with phrase-based decoding to benefit from both approaches. Our model is based on minimal translation units, but we use phrases during decoding. Through an extensive evaluation we found that this combination not only improves the search accuracy but also the BLEU scores. Our in-house phrase-based decoder outperformed state-of-the-art phrase-based (Moses and Phrasal) and N-gram-based (NCode) systems on three translation tasks.
Comparative Experiments. Motivated by these results, we integrated the OSM into the state-of-the-art phrase-based system Moses (Koehn et al. 2007). Our aim was to directly compare the performance of the lexicalized reordering model to the OSM and to see whether we can improve the performance further by using both models together. Our integration of the OSM into Moses gave a statistically significant improvement over a competitive baseline system in most cases.
In order to assess the contribution of improved reordering versus the contribution of better modeling with MTUs in the OSM-augmented Moses system, we removed the reordering operations from the stream of operations. This is equivalent to integrating the conventional N-gram tuple sequence model (Mariño et al. 2006) into a phrasebased decoder, as also tried by Niehues et al. (2011). Small gains were observed in most cases, showing that much of the improvement obtained by the OSM is due to better reordering.
Generalized Operation Sequence Model. The primary strength of the OSM over the lexicalized reordering model is its ability to take advantage of the wider contextual information. In an error analysis we found that the lexically driven OSM often falls back to very small context sizes because of data sparsity. We show that this problem can be addressed by learning operation sequences over generalized representations such as POS tags.
The article is organized into seven sections. Section 2 is devoted to a literature review. We discuss the pros and cons of the phrase-based and N-gram-based SMT frameworks in terms of both model and search. Section 3 presents our model. We
show how our model combines the benefits of both of the frameworks and removes their drawbacks. Section 4 provides an empirical evaluation of our preliminary system, which uses an MTU-based decoder, against state-of-the-art phrase-based (Moses and Phrasal) and N-gram-based (Ncode) systems on three standard tasks of translating German-to-English, Spanish-to-English, and French-to-English. Our results show improvements over the baseline systems, but we noticed that using minimal translation units during decoding makes the search problem difficult, which suggests using larger units in search. Section 5 presents an extension to our system to combine phrasebased decoding with the operation sequence model to address the problems in search. Section 5.1 empirically shows that information available in phrases can be used to improve the search performance and translation quality. Finally, we probe whether integrating our model into the phrase-based SMT framework addresses the mentioned drawbacks and improves translation quality. Section 6 provides an empirical evaluation of our integration on six standard tasks of translating German–English, French–English, and Spanish–English pairs. Our integration gives statistically significant improvements over submission quality baseline systems. Section 7 concludes. 2 previous work : The phrase-based model (Koehn et al. 2003; Och and Ney 2004) segments a bilingual sentence pair into phrases that are continuous sequences of words. These phrases are then reordered through a lexicalized reordering model that takes into account the orientation of a phrase with respect to its previous phrase (Tillmann and Zhang 2005) or block of phrases (Galley and Manning 2008). Phrase-based models memorize local dependencies such as short reorderings, translations of idioms, and the insertion and deletion of words sensitive to local context. Phrase-based systems, however, have the following drawbacks.
Handling of Non-local Dependencies. Phrase-based SMT models dependencies between words and their translations inside of a phrase well. However, dependencies across phrase boundaries are ignored because of the strong phrasal independence assumption. Consider the bilingual sentence pair shown in Figure 1(a).
Reordering of the German word stimmen is internal to the phrase-pair gegen ihre Kampagne stimmen -‘vote against your campaign’ and therefore represented by the translation model. However, the model fails to correctly translate the test sentence shown in Figure 1(b), which is translated as ‘they would for the legalization of abortion in Canada vote’, failing to displace the verb. The language model does not provide enough
evidence to counter the dispreference of the translation model against jumping over the source words für die Legalisieurung der Abtreibung in Kanada and translating stimmen - ‘vote’ at its correct position.
Weak Reordering Model. The lexicalized reordering model is primarily designed to deal with short-distance movement of phrases such as swapping two adjacent phrases and cannot properly handle long-range jumps. The model only learns an orientation of how a phrase was reordered with respect to its previous and next phrase; it makes independence assumptions over previously translated phrases and does not take into account how previous words were translated and reordered. Although such an independence assumption is useful to reduce sparsity, it is overly generalizing and does not help to disambiguate good reorderings from the bad ones.
Moreover, a vast majority of extracted phrases are singletons and the corresponding probability of orientation given phrase-pair estimates are based on a single observation. Due to sparsity, the model falls back to use one-word phrases instead, the orientation of which is ambiguous and can only be judged based on context that is ignored. This drawback has been addressed by Cherry (2013) by using sparse features for reordering models.
Hard Distortion Limit. The lexicalized reordering model fails to filter out bad largescale reorderings effectively (Koehn 2010). A hard distortion limit is therefore required during decoding in order to produce good translations. A distortion limit beyond eight words lets the translation accuracy drop because of search errors (Koehn et al. 2005). The use of a hard limit is undesirable for German–English and similar language pairs with significantly different syntactic structures. Several researchers have tried to address this problem. Moore and Quirk (2007) proposed improved future cost estimation to enable higher distortion limits in phrasal MT. Green, Galley, and Manning (2010) additionally proposed discriminative distortion models to achieve better translation accuracy than the baseline phrase-based system for a distortion limit of 15 words. Bisazza and Federico (2013) recently proposed a novel method to dynamically select which longrange reorderings to consider during the hypothesis extension process in a phrasebased decoder and showed an improvement in a German–English task by increasing the distortion limit to 18.
Spurious Phrasal Segmentation. A problem with the phrase-based model is that there is no unique correct phrasal segmentation of a sentence. Therefore, all possible ways of segmenting a bilingual sentence consistent with the word alignment are learned and used. This leads to two problems: (i) phrase frequencies are obtained by counting all possible occurrences in the training corpus, and (ii) different segmentations producing the same translation are generated during decoding. The former leads to questionable parameter estimates and the latter may lead to search errors because the probability of a translation is fragmented across different segmentations. Furthermore, the diversity in N-best translation lists is reduced. N-gram-based SMT (Mariño et al. 2006) uses an N-gram model that jointly generates the source and target strings as a sequence of bilingual translation units called tuples. Tuples are essentially minimal phrases, atomic units that cannot be decomposed any further. The tuples are generated left to right in target word order. Reordering is not
part of the statistical model. The parameters of the N-gram model are learned from bilingual data where the tuples have been arranged in target word order (see Figure 2).
Decoders for N-gram-based SMT reorder the source words in a preprocessing step so that the translation can be done monotonically. The reordering is performed with POS-based rewrite rules (see Figure 2 for an example) that have been learned from the training data (Crego and Mariño 2006). Word lattices are used to compactly represent a number of alternative reorderings. Using parts of speech instead of words in the rewrite rules makes them more general and helps to avoid data sparsity problems.
The mechanism has several useful properties. Because it is based on minimal units, there is only one derivation for each aligned bilingual sentence pair. The model therefore avoids spurious ambiguity. The model makes no phrasal independence assumption and generates a tuple monotonically by looking at a context of n previous tuples, thus capturing context across phrasal boundaries. On the other hand, N-gram-based systems have the following drawbacks.
Weak Reordering Model. The main drawback of N-gram-based SMT is its poor reordering mechanism. Firstly, by linearizing the source, N-gram-based SMT throws away useful information about how a particular word is reordered with respect to the previous word. This information is instead stored in the form of rewrite rules, which have no influence on the translation score. The model does not learn lexical reordering triggers and reorders through the learned rules only. Secondly, search is performed only on the precalculated word permutations created based on the source-side words. Often, evidence of the correct reordering is available in the translation model and the targetside language model. All potential reorderings that are not supported by the rewrite rules are pruned in the pre-processing step. To demonstrate this, consider the bilingual sentence pair in Figure 2 again. N-gram-based MT will linearize the word sequence gegen ihre Kampagne stimmen to stimmen gegen ihre Kampagne, so that it is in the same order as the English words. At the same time, it learns a POS rule: IN PRP NN VB → VB IN PRP NN. The POS-based rewrite rules serve to precompute the orderings that will be hypothesized during decoding. However, notice that this rule cannot generalize to the test sentence in Figure 1(b), even though the tuple translation model learned the trigram < sie – ‘they’ würden – ‘would’ stimmen – ‘vote’ > and it is likely that the monolingual language model has seen the trigram they would vote.
Hard Reordering Limit. Due to sparsity, only rules with seven or fewer tags are extracted. This subsequently constrains the reordering window to seven or fewer words, preventing the N-gram model from hypothesizing long-range reorderings that require
larger jumps. The need to perform long-distance reordering motivated the idea of using syntax trees (Crego and Mariño 2007) to form rewrite rules. However, the rules are still extracted ignoring the target-side, and search is performed only on the precalculated orderings.
Difficult Search Problem. Using MTUs makes the search problem much more difficult because of poor translation option selection. To illustrate this consider the phrase pair schoss ein Tor – ‘scored a goal’, consisting of units schoss – ‘scored’, ein – ‘a’, and Tor – ‘goal’. It is likely that the N-gram system does not have the tuple schoss – ‘scored’ in its N-best translation options because it is an uncommon translation. Even if schoss – ‘scored’ is hypothesized, it will be ranked quite low in the stack and may be pruned, before ein and Tor are generated in the next steps. A similar problem is also reported in Costa-jussà et al. (2007): When trying to reproduce the sentences in the N-best translation output of the phrase-based system, the N-gram-based system was able to produce only 37.5% of sentences in the Spanish-to-English and English-to-Spanish translation task, despite having been trained on the same word alignment. A phrase-based system, on the other hand, is likely to have access to the phrasal unit schoss ein Tor – ‘scored a goal’ and can generate it in a single step. 3 operation sequence model :Now we present a novel generative model that explains the translation process as a linear sequence of operations that generate a source and target sentence in parallel. Possible operations are (i) generation of a sequence of source and/or target words, (ii) insertion of gaps as explicit target positions for reordering operations, and (iii) forward and backward jump operations that do the actual reordering. The probability of a sequence of operations is defined according to an N-gram model, that is, the probability of an operation depends on the n − 1 preceding operations. Because the translation (generation) and reordering operations are coupled in a single generative story, the reordering decisions may depend on preceding translation decisions, and translation decisions may depend on preceding reordering decisions. This provides a natural reordering mechanism able to deal with local and long-distance reorderings consistently. The generative story of the model is motivated by the complex reordering in the German-to-English translation task. The English words are generated in linear order,3 and the German words are generated in parallel with their English translations. Mostly, the generation is done monotonically. Occasionally the translator inserts a gap on the German side to skip some words to be generated later. Each inserted gap acts as a designated landing site for the translator to jump back to. When the translator needs to cover the skipped words, it jumps back to one of the open gaps. After this is done, the translator jumps forward again and continues the translation. We will now, step by step, present the characteristics of the new model by means of examples.
3 Generating the English words in order is also what the decoder does when translating from German to English.
3.1.1 Basic Operations. The generation of the German–English sentence pair Peter liest – ‘Peter reads’ is straightforward because it is a simple 1-to-1 word-based translation without reordering:
Generate (Peter , Peter) Generate (liest , reads)
3.1.2 Insertions and Deletions. The translation Es ist ja nicht so schlimm – ‘it is not that bad’, requires the insertion of an additional German word ja, which is used as a discourse particle in this construction.
Generate (Es , it) Generate (ist , is) Generate Source Only (ja) Generate (nicht , not) Generate (so , that) Generate (schlimm , bad)
Conversely, the translation Lies mit – ‘Read with me’ requires the deletion of an untranslated English word me.
Generate (Lies , Read) Generate (mit , with) Generate Target Only (me)
3.1.3 Reordering. Let us now turn to an example that requires reordering, and revisit the example in Figure 1(a). The generation of this sentence in our model starts with generating sie – ‘they’, followed by the generation of würden – ‘would’. Then a gap is inserted on the German side, followed by the generation of stimmen – ‘vote’. At this point, the (partial) German and English sentences look as follows:
Operation Sequence Generation
Generate(sie, they) Generate (würden, would) sie würden stimmen ↓ Insert Gap Generate(stimmen, vote)
‘they would vote’
The arrow sign ↓ denotes the position after the previously covered German word. The translation proceeds as follows. We jump back to the open gap on the German side and fill it by generating gegen – ‘against’, Ihre – ‘your’ and Kampagne – ‘campaign’. Let us discuss some useful properties of this mechanism:
1. We have learned a reordering pattern sie würden stimmen – ‘they would vote’, which can be used to generalize the test sentence in Figure 1(b). In this case the translator jumps back and generates the tuples für – ‘for’, die – ‘the’, Legalisierung – ‘legalization’, der – ‘of’, Abtreibung – ‘abortion’, in – ‘in’, Kanada – ‘Canada’.
2. The model handles both local (Figure 1 (a)) and long-range reorderings (Figure 1 (b)) in a unified manner, regardless of how many words separate würden and stimmen.
3. Learning the operation sequence Generate(sie, they) Generate(würden, would) Insert Gap Generate(stimmen, vote) is like learning a phrase pair sie würden X stimmen – ‘they would vote’. The open gap represented by acts as a placeholder for the skipped phrases and serves a similar purpose as the non-terminal category X in a discontinuous phrase-based system.
4. The model couples lexical generation and reordering information. Translation decisions are triggered by reordering decisions and vice
versa. Notice how the reordering decision is triggered by the translation decision in the example. The probability of a gap insertion operation after the generation of the auxiliaries würden – ‘would’ will be high because reordering is necessary in order to move the second part of the German verb complex (stimmen) to its correct position at the end of the clause.
Complex reorderings can be achieved by inserting multiple gaps and/or recursively inserting a gap within a gap. Consider the generation of the example in Figure 3 (borrowed from Chiang [2007]). The generation of this bilingual sentence pair proceeds as follows:
Generate(Aozhou, Australia) Generate(shi, is) Insert Gap Generate(zhiyi, one of )
At this point, the (partial) Chinese and English sentences look like this:
Aozhou shi zhiyi ↓
Australia is one of
The translator now jumps back and recursively inserts a gap inside of the gap before continuing translation:
Jump Back (1) Insert Gap Generate(shaoshu, the few) Generate(guojia, countries)
Aozhou shi shaoshu guojia ↓ zhiyi Australia is one of the few countries
The rest of the sentence pair is generated as follows:
Jump Back (1) Insert Gap Generate(de, that) Jump Back (1) Insert Gap Generate(you, have) Generate(bangjiao, diplomatic relationships) Jump Back (1) Generate(yu, with) Generate(Beihan, North Korea)
Note that the translator jumps back and opens new gaps recursively to exhibit a property similar to the hierarchical model. However, our model uses a deterministic algorithm (see Algorithm 1 later in this article) to convert each bilingual sentence pair given the alignment to a unique derivation, thus avoiding spurious ambiguity unlike hierarchical and phrase-based models.
Multiple gaps can simultaneously exist at any time during generation. The translator decides based on the next English word to be covered which open gap to jump to. Figure 4 shows a German–English subordinate clause pair. The generation of this example is carried out as follows:
Insert Gap Generate(nicht, do not) Insert Gap Generate(wollen, want to)
At this point, the (partial) German and English sentences look as follows:
nicht wollen ↓
do not want to
The inserted gaps act as placeholders for the skipped prepositional phrase über konkrete Zahlen – ‘on specific figures’ and the verb phrase verhandeln – ‘negotiate’. When the translator decides to generate any of the skipped words, it jumps back to one of the open gaps. The Jump Back operation closes the gap that it jumps to. The translator proceeds monotonically from that point until it needs to jump again. The generation proceeds as follows:
Jump Back (1) Generate(verhandeln, negotiate)
nicht verhandeln ↓ wollen do not want to negotiate
The translation ends by jumping back to the open gap and generating the prepositional phrase as follows:
Jump Back (1) Generate(über, on) Generate(konkrete, specific) Generate(Zahlen, figures)
5. Notice that although our model is based on minimal units, we can nevertheless memorize phrases (along with reordering information) through operation subsequences that are memorized by learning an N-gram model over these operation sequences. Some interesting phrases that our model learns are:
Phrases Operation Sub-sequence
nicht X wollen – ‘do not want to’ Generate (nicht , do not) Insert Gap Generate (wollen , want to)
verhandeln wollen – ‘want to negotiate’ Insert Gap Generate (wollen , want to) Jump Back(1) Generate (verhandeln , negotiate)
X represents , the Insert Gap operation on the German side in our notation.
3.1.4 Generation of Discontinuous Source Units. Now we discuss how discontinuous source cepts can be represented in our generative model. The Insert Gap operation discussed in the previous section can also be used to generate discontinuous source cepts. The generation of any such cept is done in several steps. See the example in Figure 5. The gappy cept hat...gelesen – ‘read’ can be generated as shown.
Operation Sequence Generation
Generate(er, he) Generate (hat gelesen, read) er hat gelesen ↓ Insert Gap Continue Source Cept
he read
After the generation of er – ‘he’, the first part of the German complex verb hat is generated as an incomplete translation of ‘read’. The second part gelesen is added to a queue to be generated later. A gap is then inserted for the skipped words ein and Buch. Lastly, the second word (gelesen) of the unfinished German cept hat...gelesen is added to complete the translation of ‘read’ through a Continue Source Cept operation. Discontinuous cepts on the English side cannot be generated analogously because of the fundamental assumption of the model that English (target-side) will be generated from left to right. This is a shortcoming of our approach, which we will discuss later in Section 4.1. Our model uses five translation and three reordering operations, which are repeatedly applied in a sequence. The following is a definition of each of these operations. Generate (X,Y): X and Y are German and English cepts, respectively, each with one or more words. Words in X (German) may be consecutive or discontinuous, but the words in Y (English) must be consecutive. This operation causes the words in Y and the first word in X to be added to the English and German strings, respectively, that were generated so far. Subsequent words in X are added to a queue to be generated later. All the English words in Y are generated immediately because English (target-side) is generated in linear order as per the assumption of the model.4 The generation of the second (and subsequent) German words in a multiword cept can be delayed by gaps, jumps, and other operations defined in the following.
4 Note that when we are translating in the opposite direction (i.e., English-to-German), then German becomes target-side and is generated monotonically and gaps and jumps are performed on English (now source-side).
Continue Source Cept: The German words added to the queue by the Generate (X,Y) operation are generated by the Continue Source Cept operation. Each Continue Source Cept operation removes one German word from the queue and copies it to the German string. If X contains more than one German word, say n many, then it requires n translation operations, an initial Generate (X1...Xn, Y) operation, and n − 1 Continue Source Cept operations. For example kehrten...zurück – ‘returned’ is generated by the operation Generate (kehrten zurück, returned), which adds kehrten and ‘returned’ to the German and English strings and zurück to a queue. A Continue Source Cept operation later removes zurück from the queue and adds it to the German string.
Generate Source Only (X): The words in X are added at the current position in the German string. This operation is used to generate a German word with no coresponding English word. It is performed immediately after its preceding German word is covered. This is because there is no evidence on the English side that indicates when to generate X.5 Generate Source Only (X) helps us learn a source word deletion model. It is used during decoding, where a German word X is either translated to some English word(s) by a Generate (X,Y) operation or deleted with a Generate Source Only (X) operation.
Generate Target Only (Y): The words in Y are added at the current position in the English string. This operation is used to generate an English word with no corresponding German word. We do not utilize this operation in MTU-based decoding where it is hard to predict when to add unaligned target words during decoding. We therefore modified the alignments to remove this, by aligning unaligned target words (see Section 4.1 for details). In phrase-based decoding, however, this is not necessary, as we can easily predict unaligned target words where they are present in a phrase pair.
Generate Identical: The same word is added at the current position in both the German and English strings. The Generate Identical operation is used during decoding for the translation of unknown words. The probability of this operation is estimated from singleton German words that are translated to an identical string. For example, for a tuple QCRI – ‘QCRI’, where German QCRI was observed exactly once during training, we use a Generate Identical operation rather than Generate (QCRI, QCRI). We now discuss the set of reordering operations used by the generative story. Reordering has to be performed whenever the German word to be generated next does not immediately follow the previously generated German word. During the generation process, the translator maintains an index that specifies the position after the previously covered German word (j), an index (Z) that specifies the index after the right-most German word covered so far, and an index of the next German word to be covered (j′). The set of reordering operations used in generation depends upon these indexes. Please refer to Algorithm 1 for details.
5 We want to preserve a 1-to-1 relationship between operation sequences and aligned sentence pairs. If we allowed an unaligned source word to be generated at any time, we would obtain several operation sequences that produce the same aligned sentence pair.
Insert Gap: This operation inserts a gap, which acts as a placeholder for the skipped words. There can be more than one open gap at a time.
Jump Back (W): This operation lets the translator jump back to an open gap. It takes a parameter W specifying which gap to jump to. The Jump Back (1) operation jumps to the closest gap to Z, Jump Back (2) jumps to the second closest gap to Z, and so forth. After the backward jump, the target gap is closed.
Jump Forward: This operation makes the translator jump to Z. It is performed when the next German word to be generated is to the right of the last German word generated and does not follow it immediately. It will be followed by an Insert Gap or Jump Back (W) operation if the next source word is not at position Z. We use Algorithm 1 to convert an aligned bilingual sentence pair to a sequence of operations. Table 1 shows step by step by means of an example (Figure 6) how the conversion is done. The values of the index variables are displayed at each point.
Table 1 Step-wise generation of Example in Figure 6. The arrow indicates position j.
Figure 6 Discontinuous cept translation. Our model is estimated from a sequence of operations obtained through the transformation of a word-aligned bilingual corpus. An operation can be to generate source and target words or to perform reordering by inserting gaps and jumping forward and backward. Let O = o1, . . . , oJ be a sequence of operations as hypothesized by the translator to generate a word-aligned bilingual sentence pair< F, E, A >. The translation model is then defined as:
pT(F, E, A) = p(o1, .., oJ) = J∏
j=1
p(oj|oj−n+1...oj−1)
where n indicates the amount of context used and A defines the word-alignment function between E and F. Our translation model is implemented as an N-gram model of operations using the SRILM toolkit (Stolcke 2002) with Kneser-Ney smoothing (Kneser and Ney 1995). The translate operations in our model (the operations with a name starting with Generate) encapsulate tuples. Tuples are minimal translation units extracted from the word-aligned corpus. The idea is similar to N-gram-based SMT except that the tuples in the N-gram model are generated monotonically. We do not impose the restriction of monotonicity in our model but integrate reordering operations inside the generative model.
As in the tuple N-gram model, there is a 1-to-1 correspondence between aligned sentence pairs and operation sequences, that is, we get exactly one operation sequence per bilingual sentence given its alignments. The corpus conversion algorithm (Algorithm 1) maps each bilingual sentence pair given its alignment into a unique sequence of operations deterministically, thus maintaining a 1-to-1 correspondence. This property of the model is useful because it addresses the spurious phrasal segmentation problem in phrase-based models. A phrase-based model assigns different scores to a derivation based on which phrasal segmentation is chosen. Unlike this, the OSM assigns only one score because the model does not suffer from spurious ambiguity.
3.6.1 Discriminative Model. We use a log-linear approach (Och 2003) to make use of standard features along with several novel features that we introduce to improve endto-end accuracy. We search for a target string E that maximizes a linear combination of feature functions:
Ê = arg max E ⎧⎨ ⎩ J∑ j=1 λjhj(F, E) ⎫⎬ ⎭
where λj is the weight associated with the feature hj(F, E). Apart from the OSM and standard features such as target-side language model, length bonus, distortion limit, and IBM lexical features (Koehn, Och, and Marcu 2003), we used the following new features:
Deletion Penalty. Deleting a source word (Generate Source Only (X)) is a common operation in the generative story. Because there is no corresponding target-side word, the monolingual language model score tends to favor this operation. The deletion penalty counts the number of deleted source words.
Gap and Open Gap Count. These features are introduced to guide the reordering decisions. We observe a large amount of reordering in the automatically word aligned training text. However, given only the source sentence (and little world knowledge), it is not realistic to try to model the reasons for all of this reordering. Therefore we can use a more robust model that reorders less than humans do. The gap count feature sums to the total number of gaps inserted while producing a target sentence.
The open gap count feature is a penalty paid once for each translation operation (Generate(X,Y), Generate Identical, Generate Source Only (X)) performed whose value is the number of currently open gaps. This penalty controls how quickly gaps are closed.
Distance-Based Features. We have two distance-based features to control the reordering decisions. One of the features is the Gap Distance, which calculates the distance between the first word of a source cept X and the start of the leftmost gap. This cost is paid once for each translation operation (Generate, Generate Identical, Generate Source Only (X)). For a source cept covering the positions X1, . . . , Xn, we get the feature value gj = X1 − S, where S is the index of the left-most source word where a gap starts. Another distance-based penalty used in our model is the Source Gap Width. This feature only applies in the case of a discontinuous translation unit and computes the distance between the words of a gappy cept. Let f = f1 . . . , fi, . . . , fn be a gappy source cept where xi is the index of the ith source word in the cept f . The value of the gap-width penalty is calculated as:
wj = n∑
i=2
xi − xi−1 − 1 4 mtu-based search :We explored two decoding strategies in this work. Our first decoder complements the model and only uses minimal translation units in left-to-right stack-based decoding, similar to that used in Pharaoh (Koehn 2004a). The overall process can be roughly divided into the following steps: (i) extraction of translation units, (ii) future cost estimation, (iii) hypothesis extension, and (iv) recombination and pruning. The last two steps are repeated iteratively until all the words in the source sentence have been translated.
Our hypotheses maintain the index of the last source word covered (j), the position of the right-most source word covered so far (Z), the number of open gaps, the number of gaps so far inserted, the previously generated operations, the generated target string, and the accumulated values of all the features discussed in Section 3.6.1. The sequence of operations may include translation operations (generate, continue source cept, etc.) and reordering operations (gap insertions, jumps). Recombination6 is performed on hypotheses having the same coverage vector, monolingual language model context, and OSM context. We do histogram-based pruning, maintaining the 500 best hypotheses for each stack. A large beam size is required to cope with the search errors that result from using minimal translation units during decoding. We address this problem in Section 5.
6 Note that although we are using minimal translation units, recombination is still useful as different derivations can arise through different alignments between source and target fragments. Also, recombination can still take place if hypotheses differ slightly in the output (Koehn 2010). Aligned bilingual training corpora often contain unaligned target words and discontinuous target cepts, both of which pose problems. Unlike discontinuous source cepts, discontinuous target cepts such as hinunterschüttete – ‘poured . . . down’ in constructions like den Drink hinunterschüttete – ‘poured the drink down’ cannot be handled by the operation sequence model because it generates the English words in strict left-to-right order. Therefore they have to be eliminated.
Unaligned target words are only problematic for the MTU-based decoder, which has difficulties predicting where to insert them. Thus, we eliminate unaligned target words in MTU-based decoding.
We use a three-step process (Durrani, Schmid, and Fraser 2011) that modifies the alignments and removes unaligned and discontinuous targets. If a source word is aligned with multiple target words that are not consecutive, first the link to the least frequent target word is identified, and the group (consecutive adjacent words) of links containing this word is retained while the others are deleted. The intuition here is to keep the alignments containing content words (which are less frequent than functional words). For example, the alignment link hinunterschüttete – ‘down’ is deleted and only the link hinunterschüttete – ‘poured’ is retained because ‘down’ occurs more frequently than ‘poured’. Crego and Yvon (2009) used split tokens to deal with this phenomenon.
For MTU-based decoding we also need to deal with unaligned target words. For each unaligned target word, we determine the (left or right) neighbor that it appears more frequently with and align it with the same source word as this neighbor. Crego, de Gispert, and Mariño (2005) and Mariño et al. (2006) instead used lexical probabilities p( f |e) obtained from IBM Model 1 (Brown et al. 1993) to decide whether to attach left or right. A more sophisticated strategy based on part-of-speech entropy was proposed by Gispert and Mariño (2006). We evaluated our systems on German-to-English, French-to-English, and Spanish-toEnglish news translation for the purpose of development and evaluation. We used data from the eighth version of the Europarl Corpus and the News Commentary made available for the translation task of the Eighth Workshop on Statistical Machine Translation.7 The bilingual corpora contained roughly 2M bilingual sentence pairs, which we obtained by concatenating news commentary (≈ 184K sentences) and Europarl for the estimation of the translation model. Word alignments were generated with GIZA++ (Och and Ney 2003), using the grow-diag-final-and heuristic8 (Koehn et al. 2005). All data are lowercased, and we use the Moses tokenizer. We took news-test-2008 as the dev set for optimization and news-test 2009-2012 for testing. The feature weights are tuned with Z-MERT (Zaidan 2009).
4.2.1 Baseline Systems. We compared our system with (i) Moses9 (Koehn et al. 2007), (ii) Phrasal10 (Cer et al. 2010), and (iii) Ncode11 (Crego, Yvon, and Mariño 2011). We used
7 http://www.statmt.org/wmt13/translation-task.html 8 We also tested other symmetrization heuristics such as “Union” and “Intersection” but found the GDFA
heuristic gave best results for all language pairs. 9 http://www.statmt.org/moses/
10 http://nlp.stanford.edu/phrasal/ 11 http://www.limsi.fr/Individu/jmcrego/bincoder/
all these toolkits with their default settings. Phrasal provides two main extensions to Moses: a hierarchical reordering model (Galley and Manning 2008) and discontinuous source and target phrases (Galley and Manning 2010). We used the default stack sizes of 100 for Moses,12 200 for Phrasal, and 25 for Ncode (with 2n stacks). A 5-gram English language model is used. Both phrase-based systems use the 20 best translation options per source phrase; Ncode uses the 25 best tuple translations and a 4-gram tuple sequence model. A hard distortion limit of 6 is used in the default configuration of both phrasebased systems. Among the other defaults, we retained the hard source gap penalty of 15 and a target gap penalty of 7 in Phrasal. We provide Moses and Ncode with the same post-edited alignments13 from which we had removed target-side discontinuities. We feed the original alignments to Phrasal because of its ability to learn discontinuous source and target phrases. All the systems use MERT for the optimization of the weight vector.
4.2.2 Training. Training steps include: (i) post-editing of the alignments (Section 4.1), (ii) generation of the operation sequence (Algorithm 1), and (iii) estimation of the N-gram translation (OSM) and language models using the SRILM toolkit (Stolcke 2002) with Kneser-Ney smoothing. We used 5-gram models.
4.2.3 Summary of Developmental Experiments. During the developent of the MTU-based decoder, we performed a number of experiments to obtain optimal settings for the system. We list here a summary of the results from those experiments:
We found that discontinuous source-side cepts do not improve translation quality in most cases but increase the decoding time by multiple folds. We will therefore only use continuous cepts.
We performed experiments by varying the distortion limit from the conventional window of 6 words to infinity (= no hard limit). We found that the performance of our system is robust when removing the hard reordering constraint and even saw a slight improvement in results in the case of German-to-English systems. Using no distortion limit, however, significantly increases the decoding time. We will therefore use a window of 16 words, which we found to be optimal on the development set.
The performance of the MTU-based decoder is sensitive to the stack size. A high limit of 500 is required for decent search accuracy. We will discuss this further in the next section.
We found using 10 best translation options for each extracted cept during decoding to be optimal.
4.2.4 Comparison with the Baseline Systems. In this section we compare our system (OSMmtu) with the three baseline systems. We used Kevin Gimpel’s tester,14 which uses bootstrap resampling (Koehn 2004b) to test which of our results are significantly better than the baseline results. We mark a baseline result with “*” in order to indicate
12 Using stack sizes from 200–1,000 did not improve results. 13 Using post-processed alignments gave better results than using the original alignments for these baseline
systems. 14 http://www.ark.cs.cmu.edu/MT/
that our model shows a significant improvement over this baseline with a confidence of p < 0.05. We use 1,000 samples during bootstrap resampling.
Our German-to-English results (see Table 2) are significantly better than the baseline systems in most cases. Our French-to-English results show a significant improvement over Moses in three out of four cases, and over Phrasal in half of the cases. The N-gram-based system NCode was better or similar to our system on the French task. Our Spanish-to-English system also showed roughly the same translation quality as the baseline systems, but was significantly worse on the WMT12 task. 5 phrase-based search :The MTU-based decoder is the most straightforward implementation of a decoder for the operation sequence model, but it faces search problems that cause a drop in translation accuracy. Although the OSM captures both source and target contexts and provides a better reordering mechanism, the ability to memorize and produce larger translation units gives an edge to the phrase-based model during decoding in terms of better search performance and superior selection of translation units. In this section, we combine N-gram-based modeling with phrase-based decoding. This combination not only improves search accuracy but also increases translation quality in terms of BLEU.
The operation sequence model, although based on minimal translation units, can learn larger translation chunks by memorizing a sequence of operations. However, it often has difficulties to produce the same translations as the phrase-based system because of the following drawbacks of MTU-based decoding: (i) the MTU-based decoder does not have access to all the translation units that a phrase-based decoder uses as part of a larger phrase, (ii) it requires a larger beam size to prevent early pruning of correct
hypotheses, and (iii) it uses less-powerful future-cost estimates than the phrase-based decoder. To demonstrate these problems, consider the phrase pair
which the model memorizes through the sequence:
Generate(Wie, What is) Insert Gap Generate (Sie, your) Jump Back (1) Generate (heissen, name)
The MTU-based decoder needs three separate tuple translations to generate the same phrasal translation: Wie – ‘What is’, Sie – ‘your’ and heißen – ‘name’. Here we are faced with three challenges.
Translation Coverage: The first problem is that the N-gram model does not have the same coverage of translation options. The English cepts ‘What is’, ‘your’, and ‘name’ are not good candidate translations for the German cepts Wie, Sie, and heißen, which are usually translated to ‘How’, ‘you’, and ‘call’, respectively, in isolation. When extracting tuple translations for these cepts from the Europarl data for our system, the tuple Wie – ‘What is’ is ranked 124th, heißen – ‘name’ is ranked 56th, and Sie – ‘your’ is ranked 9th in the list of n-best translation candidates. Typically, only the 20 best translation options are used, for the sake of efficiency, and such phrasal units with less frequent translations are never hypothesized in the N-gram-based systems. The phrase-based system, on the other hand, can extract the phrase Wie heißen Sie – ‘what is your name’ even if it is observed only once during training.
Larger Beam Size: Even when we allow a huge number of translation options and therefore hypothesize such units, we are faced with another challenge. A larger beam size is required in MTU-based decoding to prevent uncommon translations from getting pruned. The phrase-based system can generate the phrase pair Wie heißen Sie – ‘what is your name’ in a single step, placing it directly into the stack three words to the right. The MTU-based decoder generates this phrase in three stacks with the tuple translations Wie – ‘What is’, Sie – ‘your’, and heißen – ‘name’. A very large stack size is required during decoding to prevent the pruning of Wie – ‘What is’, which is ranked quite low in the stack until the tuple Sie – ‘your’ is hypothesized in the next stack. Although the translation quality achieved by phrase-based SMT remains the same when varying the beam size, the performance of our system varies drastically with different beam sizes (especially for the German–English experiments where the search is more difficult due to a higher number of reorderings). Costa-jussà et al. (2007) also report a significant drop in the performance of N-gram-based SMT when a beam size of 10 is used instead of 50 in their experiments.
Future Cost Estimation: A third problem is caused by inaccurate future cost estimation. Using phrases helps phrase-based SMT to better estimate the future language model cost because of the larger context available, and allows the decoder to capture local (phrase-internal) reorderings in the future cost. In comparison, the future cost for tuples is based on unigram probabilities. The future cost estimate for the phrase pair Wie heißen Sie – ‘What is your name’ is estimated by calculating the cost of each feature.
A bigram language model cost, for example, is estimated in the phrase-based system as follows:
plm = p(What) × p(is|What) × p(your|What is) × p(name|What is your)
The translation model cost is estimated as:
ptm = p(What is your name|Wie heißen Sie)
Phrase-based SMT is aware during the preprocessing step that the words Wie heißen Sie may be translated as a phrase. This is helpful for estimating a more accurate future cost because the context is already available. The same is not true for the MTU-based decoder, to which only minimal units are available. The MTU-based decoder does not have the information that Wie heißen Sie may be translated as a phrase during decoding. The future cost estimate available to the operation sequence model for the span covering Wie heißen Sie will have unigram probabilities for both the translation and language models.
plm = p(What) × p(is|What) × p(your) × p(name)
The translation model cost is estimated as:
ptm = p(Generate(Wie, What is)) × p(Generate(heißen,name)) × p(Generate(Sie, your))
A more accurate future cost estimate for the translation model cost would be:
ptm = p(Generate(Wie,What is)) × p(Insert Gap|C2) × p(Generate(Sie,your)|C3) × p(Jump Back(1)|C4) × p(Generate(heißen,name)|C5)
where Ci is the context for the generation of the ith operation—that is, up to m previous operations. For example C1 = Generate(Wie, What is), C2 = Generate(Wie,What is) Insert Gap, and so on. The future cost estimates computed in this manner are much more accurate because not only do they consider context, but they also take the reordering operations into account (Durrani, Fraser, and Schmid 2013). We extended our in-house OSM decoder to use phrases instead of MTUs during decoding. In order to check whether phrase-based decoding solves the mentioned problems and improves the search accuracy, we evaluated the baseline MTU decoder and the phrase-based decoder with the same model parameters and tuned weights. This allows us to directly compare the model scores. We tuned the feature weights by running MERT with the MTU decoder on the dev set. Table 3 shows results from running both, the MTU-based (OSMmtu) and the phrase-based (OSMphr) decoder, on the WMT09 test set. Improved search accuracy is the percentage of times each decoder was able to produce a better model score than the other. Our phrase-based decoder uses a stack size of 200. Table 3 shows the percentage of times the MTU-based and phrase-based decoder produce better model scores than their counterpart. It shows that the phrase-based decoder produces better model scores for almost 48% of the hypotheses (on average)
across the three language pairs, whereas the MTU-based decoder (using a much higher stack size [500]) produces better hypotheses 8.2% of the time on average.
This improvement in search is also reflected in translation quality. Our phrase-based decoder outperforms the MTU-based decoder in all the cases and gives a significant improvement in 8 out of 12 cases (Table 4). In Section 4.1 we discussed the problem of handling unaligned and discontinuous target words in MTU-based decoding. An advantage of phrase-based decoding is that we can use such units during decoding if they appear within the extracted phrases. We use a Generate Target Only (Y) operation whenever the unaligned target word Y occurs in
a phrase. Similarly, we use the operation Generate (hinunterschüttete, poured down) when the discontinuous tuple hinunterschüttete – ‘poured ... down’ occurs in a phrase. While training the model, we simply ignore the discontinuity and pretend that the word ‘down’ immediately follows ‘poured’. This can be done by linearizing the subsequent parts of discontinuous target cepts to appear after the first word of the cept. During decoding we use phrase-internal alignments to hypothesize such a linearization. This is done only for the estimation of the OSM, and the target for all other purposes is generated in its original order. This heuristic allows us to deal with target discontinuities without extending the operation sequence model in complicated ways. It results in better BLEU accuracy in comparison with the post-editing of the alignments method described in Section 4.1. For details and empirical results refer to Durrani et al. (2013a) (see Table 2 therein, compare Rows 4 and 5).
Note that the OSM, like the discontinuous phrase-based model (Galley and Manning 2010), allows all possible geometries as shown in Figure 7. However, because our decoder only uses continuous phrases, we cannot hypothesize (ii) and (iii) unless they appear inside of a phrase. But our model could be integrated into a discontinuous phrase-based system to overcome this limitation. 6 further comparative experiments :Our model, like the reordering models (Tillmann and Zhang 2005; Galley and Manning 2008) used in phrase-based decoders, is lexicalized. However, our model has richer conditioning as it considers both translation and reordering context across phrasal boundaries. The lexicalized reordering model used in phrase-based SMT only accounts for how a phrase pair was reordered with respect to its previous phrase (or block of phrases). Although such an independence assumption is useful to reduce sparsity, it is overgeneralizing, with only three possible orientations. Moreover, because most of the extracted phrases are observed only once, the corresponding probability of orientation given phrase-pair estimates is very sparse. The model often has to fall back to short oneword phrases. However, most short phrases are observed frequently with all possible orientations during training. This makes it difficult for the decoder to decide which orientation should be picked during decoding. The model therefore overly relies on the language model to break such ties. The OSM may also suffer from data sparsity and the back-off smoothing may fall back to very short contexts. But it might still be able to disambiguate better than the lexicalized reordering models. Also these drawbacks can be addressed by learning an OSM over generalized word representation such as POS tags, as we show in this section.
In an effort to make a comparison of the operation sequence model with the lexicalized reordering model, we incorporate the OSM into the phrase-based Moses decoder. This allows us to exactly compare the two models in identical settings. We integrate the OSM into the hypothesis extension process of the phrase-based decoder. We convert
each phrase pair into a sequence of operations by extracting the MTUs within the phrase pair and using phrase internal alignments. The OSM is used as a feature in the log-linear framework. We also use four supportive features: the Gap, Open Gap, Gap-distance, and Deletion counts, as described earlier (see Section 3.6.1). Our Moses (Koehn et al. 2007) baseline systems are based on the setup described in Durrani et al. (2013b). We trained our systems with the following settings: maximum sentence length 80, grow-diag-final and symmetrization of GIZA++ alignments, an interpolated Kneser-Ney smoothed 5-gram language model with KenLM (Heafield 2011) used at runtime, distortion limit of 6, minimum Bayes-risk decoding (Kumar and Byrne 2004), cube pruning (Huang and Chiang 2007), and the no-reordering-over-punctuation heuristic. We used factored models (Koehn and Hoang 2007), for German–English and English–German. We trained the lexicalized reordering model (Koehn et al. 2005) with msd-bidirectional-fe settings. Table 5 shows that the OSM results in higher gains than the lexicalized reordering model on top of a plain phrase-based baseline (Pb). The average improvement obtained using the lexicalized reordering model (Pblex) over the baseline (Pb) is 0.50. In comparison, the average improvement obtained by using the OSM (Pbosm) over the baseline (Pb) is 0.74. The average improvement obtained by the combination (Pblex+osm) is 0.97. The average improvement obtained by adding the OSM over the baseline (Pblex) is 0.47. We tested for significance and found that in seven out of eight cases adding the OSM on top of Pblex gives a statistically significant improvement with a confidence of p < 0.05. Significant differences are marked with an asterisk. In an additional experiment, we studied how much the translation quality decreases when all reordering operations are removed from the operation sequence model during
training and decoding. The resulting model is similar to the tuple sequence model (TSM) of Mariño et al. (2006), except that we use phrase-internal reordering rather than POS-based rewrite rules to do the source linearization. Table 6 shows an average improvement of just 0.13 on top of the baseline phrase-based system with lexicalized reordering, which is much lower than the 0.46 points obtained with the full operation sequence model.
Bilingual translation models (without reordering) have been integrated into phrase-based systems before, either inside the decoder (Niehues et al. 2011) or to rerank the N-best candidate translations in the output of a phrase-based system (Zhang et al. 2013). Both groups reported improvements of similar magnitude when using a targetorder left-to-right TSM model for German–English and French–English translation with shared task data, but higher gains on other data sets and language pairs. Zhang et al. (2013) showed further gains by combining models with target and source left-to-right and right-to-left orders. The assumption of generating the target in monotonic order is a weakness of our work that can be addressed following Zhang et al. (2013). By generating MTUs in source order and allowing gaps and jumps on the target side, the model will be able to learn other reordering patterns that are ignored by the standard OSM. Because of data sparsity, it is impossible to observe all possible reordering patterns with all possible lexical choices in translation operations. The lexically driven OSM therefore often backs off to very small context sizes. Consider the example shown in Figure 1. The learned pattern sie würden stimmen – ‘they would vote’ cannot be generalized to er würde wählen – ‘he would vote’. We found that the OSM uses only two preceding operations as context on average. This problem can be addressed by replacing words with POS tags (or any other generalized representation such as Morph tags, word clusters) to allow the model to consider a wider syntactic context where this is appropriate, thus improving lexical decisions and the reordering capability of the model. Crego and Yvon (2010) and Niehues et al. (2011) have shown improvements in
translation quality when using a TSM model over POS units. We estimate OSMs over generalized tags and add these as separate features to the loglinear framework.15
Experiments. We enabled factored sequence models (Koehn and Hoang 2007) in German–English language pairs as these have been shown to be useful previously. We used LoPar (Schmid 2000) to obtain morphological analysis and POS annotation of German and MXPOST (Ratnaparkhi 1998), a maximum entropy model for English POS tags. We simply estimate OSMs over POS tags16 by replacing the words by the corresponding tags during training.
Table 7 shows that a system with an additional POS-based OSM (Pblex+osm(s)+osm(p)) gives an average improvement of +0.26 over the baseline (Pblex+osm(s)) system that uses an OSM over surface forms only. The overall gain by using OSMs over the baseline system is +0.70. OSM over surface tags considers 3-gram on average, and OSM over POS tags considers 4.5-grams on average, thus considering wider contextual information when making translation and reordering decisions. Table 8 shows the wall-clock decoding time (in minutes) from running the Moses decoder (on news-test2013) with and without the OSMs. Each decoder is run with 24 threads on a machine with 140GB RAM and 24 processors. Timings vary between experiments because of the fact that machines were somewhat busy in some cases. But generally, the OSM increases decoding time by more than half an hour.17
Table 9 shows the overall sizes of phrase-based translation and reordering models along with the OSMs. It also shows the model sizes when filtered on news-test2013. A similar amount of reduction could be achieved by applying filtering to the OSMs following the language model filtering described by Heafield and Lavie (2010).
15 We also tried to amalgamate lexically driven OSM and generalized OSMs into a single model rather than using these as separate features. However, this attempt was unsuccessful (See Durrani et al. [2014] for details). 16 We also found using morphological tags and automatic word clusters to be useful in our recent IWSLT evaluation campaign (Birch, Durrani, and Koehn 2013; Durrani et al. 2014). 17 The code for the OSM in Moses can be greatly optimized but requires major modifications to source and target phrase classes in Moses. 7 conclusion :In this article we presented a new model for statistical MT that combines the benefits of two state-of-the-art SMT frameworks, namely, N-gram-based and phrase-based SMT. Like the N-gram-based model, it addresses two drawbacks of phrasal MT by better handling dependencies across phrase boundaries, and solving the phrasal segmentation problem. In contrast to N-gram-based MT, our model has a generative story that tightly couples translation and reordering. Furthermore, it is able to consider all possible reorderings, unlike N-gram systems that perform search only on a limited number of pre-calculated orderings. Our model is able to correctly reorder words across large distances, and it memorizes frequent phrasal translations including their reordering as probable operation sequences.
We tested a version of our system that decodes based on minimal translation units (MTUs) against the state-of-the-art phrase-based systems Moses and Phrasal and the N-gram-based system Ncode for German-to-English, French-to-English, and Spanishto-English on three standard test sets. Our system shows statistically significant improvements in 9 out of 12 cases in the German-to-English translation task, and 10 out of 12 cases in the French-to-English translation task. Our Spanish-to-English results are similar to the baseline systems in most of the cases but consistently worse than Ncode.
MTU-based decoding suffers from poor translation coverage, inaccurate future cost estimates, and pruning of correct hypotheses. Phrase-based SMT, on the other hand, avoids these drawbacks by using larger translation chunks during search. We
therefore extended our decoder to use phrases instead of cepts while keeping the statistical model unchanged. We found that combining a model based on minimal units with phrase-based decoding improves both search accuracy and translation quality. Our system extended with phrase-based decoding showed improvements over all the baseline systems, including our MTU-based decoder. In most of the cases, the difference was significant.
Our results show that OSM consistently outperforms the Moses lexicalized reordering model and gives statistically significant gains over a very competitive Moses baseline system. We showed that considering both translation and reordering context is important and ignoring reordering context results in a significant reduction in the performance. We also showed that an OSM based on surface forms suffers from data sparsity and that an OSM based on a generalized representation with part-of-speech tags improves the translation quality by considering a larger context. In the future we would like to study whether the insight of using minimal units for modeling and search based on composed rules would hold for hierarchical SMT. Vaswani et al. (2011) recently showed that a Markov model over the derivation history of minimal rules can obtain the same translation quality as using grammars formed with composed rules, which we believe is quite promising. In this article, we present a novel machine translation model, the Operation Sequence Model (OSM), which combines the benefits of phrase-based and N-gram-based statistical machine translation (SMT) and remedies their drawbacks. The model represents the translation process as a linear sequence of operations. The sequence includes not only translation operations but also reordering operations. As in N-gram-based SMT, the model is: (i) based on minimal translation units, (ii) takes both source and target information into account, (iii) does not make a phrasal independence assumption, and (iv) avoids the spurious phrasal segmentation problem. As in phrase-based SMT, the model (i) has the ability to memorize lexical reordering triggers, (ii) builds the search graph dynamically, and (iii) decodes with large translation units during search. The unique properties of the model are (i) its strong coupling of reordering and translation where translation and reordering decisions are conditioned on n previous translation and reordering decisions, and (ii) the ability to model local and long-range reorderings consistently. Using BLEU as a metric of translation accuracy, we found that our system performs significantly [{""affiliations"": [], ""name"": ""Nadir Durrani""}, {""affiliations"": [], ""name"": ""QCRI Qatar""}, {""affiliations"": [], ""name"": ""Helmut Schmid""}, {""affiliations"": [], ""name"": ""LMU Munich""}, {""affiliations"": [], ""name"": ""Alexander Fraser""}, {""affiliations"": [], ""name"": ""Philipp Koehn""}, {""affiliations"": [], ""name"": ""Hinrich Sch\u00fctze""}] SP:314417703be66faa0d2a7d4d072ca98d6ddf9791 [{""authors"": [""Birch"", ""Alexandra"", ""Nadir Durrani"", ""Philipp Koehn.""], ""title"": ""Edinburgh SLT and MT System Description for the IWSLT 2013 Evaluation"", ""venue"": ""Proceedings of the 10th International Workshop on Spoken Language"", ""year"": 2013}, {""authors"": [""Bisazza"", ""Arianna"", ""Marcello Federico.""], ""title"": ""Efficient Solutions for Word Reordering in German-English Phrase-Based Statistical Machine Translation"", ""venue"": ""Proceedings of the Eighth"", ""year"": 2013}, {""authors"": [""Brown"", ""Peter F"", ""Stephen A. Della Pietra"", ""Vincent J. Della Pietra"", ""R.L. Mercer""], ""title"": ""The Mathematics of Statistical Machine Translation: Parameter"", ""year"": 1993}, {""authors"": [""Casacuberta"", ""Francisco"", ""Enrique Vidal.""], ""title"": ""Machine Translation with Inferred Stochastic Finite-State Transducers"", ""venue"": ""Computational Linguistics, 30:205\u2013225."", ""year"": 2004}, {""authors"": [""Cer"", ""Daniel"", ""Michel Galley"", ""Daniel Jurafsky"", ""Christopher D. Manning.""], ""title"": ""Phrasal: A Statistical Machine Translation Toolkit for Exploring New Model Features"", ""venue"": ""Proceedings of the North American Chapter"", ""year"": 2010}, {""authors"": [""Cherry"", ""Colin.""], ""title"": ""Improved Reordering for Phrase-Based Translation Using Sparse Features"", ""venue"": ""Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics:"", ""year"": 2013}, {""authors"": [""Chiang"", ""David.""], ""title"": ""Hierarchical Phrase-Based Translation"", ""venue"": ""Computational Linguistics, 33(2):201\u2013228."", ""year"": 2007}, {""authors"": [""Costa-juss\u00e0"", ""Marta R"", ""Josep M. Crego"", ""David Vilar"", ""Jos\u00e9 A.R. Fonollosa"", ""Jos\u00e9 B. Mari\u00f1o"", ""Hermann Ney""], ""title"": ""Analysis and System Combination of Phrase- and N-Gram-Based Statistical Machine"", ""year"": 2007}, {""authors"": [""Crego"", ""Josep M."", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Improving Statistical MT by Coupling Reordering and Decoding"", ""venue"": ""Machine Translation, 20(3):199\u2013215."", ""year"": 2006}, {""authors"": [""Crego"", ""Josep M."", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Syntax-Enhanced N-gram-Based SMT"", ""venue"": ""Proceedings of the 11th Machine Translation Summit, pages 111\u2013118, Copenhagen."", ""year"": 2007}, {""authors"": [""Crego"", ""Josep M."", ""Fran\u00e7ois Yvon.""], ""title"": ""Gappy Translation Units under Left-to-Right SMT Decoding"", ""venue"": ""Proceedings of the Meeting of the European Association for Machine Translation,"", ""year"": 2009}, {""authors"": [""Crego"", ""Josep M."", ""Fran\u00e7ois Yvon.""], ""title"": ""Improving Reordering with Linguistically Informed Bilingual N-Grams"", ""venue"": ""COLING 2010: Posters, pages 197\u2013205, Beijing."", ""year"": 2010}, {""authors"": [""Crego"", ""Josep M."", ""Fran\u00e7ois Yvon"", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Ncode: an Open Source Bilingual N-gram SMT Toolkit"", ""venue"": ""The Prague Bulletin of Mathematical Linguistics, 96:49\u201358."", ""year"": 2011}, {""authors"": [""Durrani"", ""Nadir"", ""Alexander Fraser"", ""Helmut Schmid.""], ""title"": ""Model With Minimal Translation Units, But Decode With Phrases"", ""venue"": ""the 2013 Conference of the North American Chapter of the Association for"", ""year"": 2013}, {""authors"": [""Durrani"", ""Nadir"", ""Alexander Fraser"", ""Helmut Schmid"", ""Hieu Hoang"", ""Philipp Koehn""], ""title"": ""Can Markov Models Over Minimal Translation Units Help Phrase-Based SMT"", ""venue"": ""In Proceedings of the 51st Annual Meeting"", ""year"": 2013}, {""authors"": [""Durrani"", ""Nadir"", ""Barry Haddow"", ""Kenneth Heafield"", ""Philipp Koehn.""], ""title"": ""Edinburgh\u2019s Machine Translation Systems for European Language Pairs"", ""venue"": ""Proceedings of the Eighth Workshop on"", ""year"": 2013}, {""authors"": [""Durrani"", ""Nadir"", ""Philipp Koehn"", ""Helmut Schmid"", ""Alexander Fraser.""], ""title"": ""Investigating the Usefulness of Generalized Word Representations in SMT"", ""venue"": ""Proceedings of the 25th Annual"", ""year"": 2014}, {""authors"": [""Durrani"", ""Nadir"", ""Helmut Schmid"", ""Alexander Fraser""], ""title"": ""A Joint Sequence Translation Model with Integrated"", ""year"": 2011}, {""authors"": [""Galley"", ""Michel"", ""Christopher D. Manning.""], ""title"": ""A Simple and Effective Hierarchical Phrase Reordering Model"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 848\u2013856,"", ""year"": 2008}, {""authors"": [""Galley"", ""Michel"", ""Christopher D. Manning.""], ""title"": ""Accurate Non-Hierarchical Phrase-Based Translation"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of"", ""year"": 2010}, {""authors"": [""Gispert"", ""Adri\u00e0"", ""Jos\u00e9 B. Mari\u00f1o.""], ""title"": ""Linguistic Tuple Segmentation in N-Gram-Based Statistical Machine Translation"", ""venue"": ""INTERSPEECH, pages 1,149\u20131,152, Pittsburgh, PA."", ""year"": 2006}, {""authors"": [""Green"", ""Spence"", ""Michel Galley"", ""Christopher D. Manning.""], ""title"": ""Improved Models of Distortion Cost for Statistical Machine Translation"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference"", ""year"": 2010}, {""authors"": [""Heafield"", ""Kenneth.""], ""title"": ""KenLM: Faster and Smaller Language Model Queries"", ""venue"": ""Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187\u2013197, Edinburgh."", ""year"": 2011}, {""authors"": [""Heafield"", ""Kenneth"", ""Alon Lavie.""], ""title"": ""Combining Machine Translation Output with Open Source: The Carnegie Mellon Multi-Engine Machine Translation Scheme"", ""venue"": ""The Prague Bulletin of Mathematical"", ""year"": 2010}, {""authors"": [""Huang"", ""Liang"", ""David Chiang.""], ""title"": ""Forest Rescoring: Faster Decoding with Integrated Language Models"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics,"", ""year"": 2007}, {""authors"": [""Kneser"", ""Reinhard"", ""Hermann Ney.""], ""title"": ""Improved Backing-off for M-gram Language Modeling"", ""venue"": ""Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing, pages 181\u2013184."", ""year"": 1995}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Pharaoh: A Beam Search Decoder for Phrase-Based Statistical Machine Translation Models"", ""venue"": ""Association for Machine Translation in the Americas, pages 115\u2013124,"", ""year"": 2004}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Statistical Significance Tests for Machine Translation Evaluation"", ""venue"": ""Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388\u2013395, Barcelona."", ""year"": 2004}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Statistical Machine Translation"", ""venue"": ""Cambridge University Press."", ""year"": 2010}, {""authors"": [""Koehn"", ""Philipp"", ""Amittai Axelrod"", ""Alexandra Birch"", ""Chris Callison-Burch"", ""Miles Osborne"", ""David Talbot.""], ""title"": ""Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation"", ""venue"": ""In"", ""year"": 2005}, {""authors"": [""Koehn"", ""Philipp"", ""Hieu Hoang.""], ""title"": ""Factored Translation Models"", ""venue"": ""Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural"", ""year"": 2007}, {""authors"": [""Constantin"", ""Evan Herbst.""], ""title"": ""Moses: Open Source Toolkit for Statistical Machine Translation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics: Demonstrations."", ""year"": 2007}, {""authors"": [""Koehn"", ""Philipp"", ""Franz J. Och"", ""Daniel Marcu.""], ""title"": ""Statistical Phrase-Based Translation"", ""venue"": ""2003 Meeting of the North American Chapter of the Association for Computational Linguistics, pages 127\u2013133,"", ""year"": 2003}, {""authors"": [""Kumar"", ""Shankar"", ""William J. Byrne.""], ""title"": ""Minimum Bayes-Risk Decoding for Statistical Machine Translation"", ""venue"": ""Human Language Technologies: The 2004 Annual Conference of the North American Chapter of"", ""year"": 2004}, {""authors"": [""Mari\u00f1o"", ""Jos\u00e9 B."", ""Rafael E. Banchs"", ""Josep M. Crego"", ""Adri\u00e0 de Gispert"", ""Patrik Lambert"", ""Jos\u00e9 A.R. Fonollosa"", ""Marta R. Costa-juss\u00e0.""], ""title"": ""N-gram-Based Machine Translation"", ""venue"": ""Computational Linguistics,"", ""year"": 2006}, {""authors"": [""Moore"", ""Robert"", ""Chris Quirk.""], ""title"": ""Faster Beam Search Decoding for Phrasal Statistical Machine Translation"", ""venue"": ""Proceedings of the 11th Machine Translation Summit, Copenhagen."", ""year"": 2007}, {""authors"": [""Niehues"", ""Jan"", ""Teresa Herrmann"", ""Stephan Vogel"", ""Alex Waibel.""], ""title"": ""Wider Context by Using Bilingual Language Models in Machine Translation"", ""venue"": ""Proceedings of the Sixth Workshop on"", ""year"": 2011}, {""authors"": [""Och"", ""Franz J.""], ""title"": ""Minimum Error Rate Training in Statistical Machine Translation"", ""venue"": ""Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, pages 160\u2013167, Sapporo."", ""year"": 2003}, {""authors"": [""Och"", ""Franz J."", ""Hermann Ney.""], ""title"": ""A Systematic Comparison of Various Statistical Alignment Models"", ""venue"": ""Computational Linguistics, 29(1):19\u201351."", ""year"": 2003}, {""authors"": [""Och"", ""Franz J."", ""Hermann Ney.""], ""title"": ""The Alignment Template Approach to Statistical Machine Translation"", ""venue"": ""Computational Linguistics, 30(1):417\u2013449."", ""year"": 2004}, {""authors"": [""Ratnaparkhi"", ""Adwait.""], ""title"": ""Maximum Entropy Models for Natural Language Ambiguity Resolution"", ""venue"": ""Ph.D. thesis, University of Pennsylvania, Philadelphia, PA."", ""year"": 1998}, {""authors"": [""Schmid"", ""Helmut.""], ""title"": ""Lopar: Design and implementation"", ""venue"": ""Bericht des sonderforschungsbereiches \u201csprachtheoretische grundlagen f\u00fcr die computerlinguistik.\u201d Technical"", ""year"": 2000}, {""authors"": [""Stolcke"", ""Andreas.""], ""title"": ""SRILM - An Extensible Language Modeling Toolkit"", ""venue"": ""International Conference on Spoken Language Processing, Denver, CO."", ""year"": 2002}, {""authors"": [""Tillmann"", ""Christoph"", ""Tong Zhang.""], ""title"": ""A Localized Prediction Model for Statistical Machine Translation"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2005}, {""authors"": [""Vaswani"", ""Ashish"", ""Haitao Mi"", ""Liang Huang"", ""David Chiang.""], ""title"": ""Rule Markov Models for Fast Tree-to-String Translation"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational"", ""year"": 2011}, {""authors"": [""Zaidan"", ""Omar F.""], ""title"": ""Z-MERT: A Fully Configurable Open Source Tool for Minimum Error Rate Training of Machine Translation Systems"", ""venue"": ""The Prague Bulletin of Mathematical Linguistics,"", ""year"": 2009}, {""authors"": [""Zhang"", ""Hui"", ""Kristina Toutanova"", ""Chris Quirk"", ""Jianfeng Gao.""], ""title"": ""Beyond Left-to-Right: Multiple Decomposition Structures for SMT"", ""venue"": ""the 2013 Conference of the North American Chapter of the"", ""year"": 2013}] acknowledgments  :We would like to thank the anonymous reviewers and Andreas Maletti and François Yvon for their helpful feedback and suggestions. The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreements 287658 (EU-Bridge) and 287688 (MateCat). Alexander Fraser was funded by Deutsche Forschungsgemeinschaft grant Models of Morphosyntax for Statistical Machine Translation. Helmut Schmid was supported by Deutsche Forschungsgemeinschaft grant SFB 732. This publication only reflects the authors’ views.","1 introduction :Statistical Machine Translation (SMT) advanced near the beginning of the century from word-based models (Brown et al. 1993) towards more advanced models that take contextual information into account. Phrase-based (Koehn, Och, and Marcu 2003; Och and Ney 2004) and N-gram-based (Casacuberta and Vidal 2004; Mariño et al. 2006) models are two instances of such frameworks. Although the two models have some common properties, they are substantially different. The present work is a step towards combining the benefits and remedying the flaws of these two frameworks.
Phrase-based systems have a simple but effective mechanism that learns larger chunks of translation called bilingual phrases.1 Memorizing larger units enables the phrase-based model to learn local dependencies such as short-distance reorderings, idiomatic collocations, and insertions and deletions that are internal to the phrase pair. The model, however, has the following drawbacks: (i) it makes independence assumptions over phrases, ignoring the contextual information outside of phrases, (ii) the reordering model has difficulties in dealing with long-range reorderings, (iii) problems in both search and modeling require the use of a hard reordering limit, and (iv) it has the spurious phrasal segmentation problem, which allows multiple derivations of a bilingual sentence pair that have the same word alignment but different model scores.
N-gram-based models are Markov models over sequences of tuples that are generated monotonically. Tuples are minimal translation units (MTUs) composed of source and target cepts.2 The N-gram-based model has the following drawbacks: (i) only precalculated orderings are hypothesized during decoding, (ii) it cannot memorize and use lexical reordering triggers, (iii) it cannot perform long distance reorderings, and (iv) using tuples presents a more difficult search problem than in phrase-based SMT.
The Operation Sequence Model. In this article we present a novel model that tightly integrates translation and reordering into a single generative process. Our model explains the translation process as a linear sequence of operations that generates a source and target sentence in parallel, in a target left-to-right order. Possible operations are (i) generation of a sequence of source and target words, (ii) insertion of gaps as explicit target positions for reordering operations, and (iii) forward and backward jump operations that do the actual reordering. The probability of a sequence of operations is defined according to an N-gram model, that is, the probability of an operation depends on the n − 1 preceding operations. Because the translation (lexical generation) and reordering operations are coupled in a single generative story, the reordering decisions may depend on preceding translation decisions and translation decisions may depend
1 A Phrase pair in phrase-based SMT is a pair of sequences of words. The sequences are not necessarily linguistic constituents. Phrase pairs are built by combining minimal translation units and ordering information. As is customary we use the term phrase to refer to phrase pairs if there is no ambiguity. 2 A cept is a group of source (or target) words connected to a group of target (or source) words in a particular alignment (Brown et al. 1993).
on preceding reordering decisions. This provides a natural reordering mechanism that is able to deal with local and long-distance reorderings in a consistent way.
Like the N-gram-based SMT model, the operation sequence model (OSM) is based on minimal translation units and takes both source and target information into account. This mechanism has several useful properties. Firstly, no phrasal independence assumption is made. The model has access to both source and target context outside of phrases. Secondly the model learns a unique derivation of a bilingual sentence given its alignments, thus avoiding the spurious phrasal segmentation problem. The OSM, however, uses operation N-grams (rather than tuple N-grams), which encapsulate both translation and reordering information. This allows the OSM to use lexical triggers for reordering like phrase-based SMT. Our reordering approach is entirely different from the tuple N-gram model. We consider all possible orderings instead of a small set of POS-based pre-calculated orderings, as is used in N-gram-based SMT, which makes their approach dependent on the availability of a source and target POS-tagger. We show that despite using POS tags the reordering patterns learned by N-gram-based SMT are not as general as those learned by our model.
Combining MTU-model with Phrase-Based Decoding. Using minimal translation units makes the search much more difficult because of the poor translation coverage, inaccurate future cost estimates, and pruning of correct hypotheses because of insufficient context. The ability to memorize and produce larger translation units gives an edge to the phrase-based systems during decoding, in terms of better search performance and superior selection of translation units. In this article, we combine N-gram-based modeling with phrase-based decoding to benefit from both approaches. Our model is based on minimal translation units, but we use phrases during decoding. Through an extensive evaluation we found that this combination not only improves the search accuracy but also the BLEU scores. Our in-house phrase-based decoder outperformed state-of-the-art phrase-based (Moses and Phrasal) and N-gram-based (NCode) systems on three translation tasks.
Comparative Experiments. Motivated by these results, we integrated the OSM into the state-of-the-art phrase-based system Moses (Koehn et al. 2007). Our aim was to directly compare the performance of the lexicalized reordering model to the OSM and to see whether we can improve the performance further by using both models together. Our integration of the OSM into Moses gave a statistically significant improvement over a competitive baseline system in most cases.
In order to assess the contribution of improved reordering versus the contribution of better modeling with MTUs in the OSM-augmented Moses system, we removed the reordering operations from the stream of operations. This is equivalent to integrating the conventional N-gram tuple sequence model (Mariño et al. 2006) into a phrasebased decoder, as also tried by Niehues et al. (2011). Small gains were observed in most cases, showing that much of the improvement obtained by the OSM is due to better reordering.
Generalized Operation Sequence Model. The primary strength of the OSM over the lexicalized reordering model is its ability to take advantage of the wider contextual information. In an error analysis we found that the lexically driven OSM often falls back to very small context sizes because of data sparsity. We show that this problem can be addressed by learning operation sequences over generalized representations such as POS tags.
The article is organized into seven sections. Section 2 is devoted to a literature review. We discuss the pros and cons of the phrase-based and N-gram-based SMT frameworks in terms of both model and search. Section 3 presents our model. We
show how our model combines the benefits of both of the frameworks and removes their drawbacks. Section 4 provides an empirical evaluation of our preliminary system, which uses an MTU-based decoder, against state-of-the-art phrase-based (Moses and Phrasal) and N-gram-based (Ncode) systems on three standard tasks of translating German-to-English, Spanish-to-English, and French-to-English. Our results show improvements over the baseline systems, but we noticed that using minimal translation units during decoding makes the search problem difficult, which suggests using larger units in search. Section 5 presents an extension to our system to combine phrasebased decoding with the operation sequence model to address the problems in search. Section 5.1 empirically shows that information available in phrases can be used to improve the search performance and translation quality. Finally, we probe whether integrating our model into the phrase-based SMT framework addresses the mentioned drawbacks and improves translation quality. Section 6 provides an empirical evaluation of our integration on six standard tasks of translating German–English, French–English, and Spanish–English pairs. Our integration gives statistically significant improvements over submission quality baseline systems. Section 7 concludes. 2 previous work : The phrase-based model (Koehn et al. 2003; Och and Ney 2004) segments a bilingual sentence pair into phrases that are continuous sequences of words. These phrases are then reordered through a lexicalized reordering model that takes into account the orientation of a phrase with respect to its previous phrase (Tillmann and Zhang 2005) or block of phrases (Galley and Manning 2008). Phrase-based models memorize local dependencies such as short reorderings, translations of idioms, and the insertion and deletion of words sensitive to local context. Phrase-based systems, however, have the following drawbacks.
Handling of Non-local Dependencies. Phrase-based SMT models dependencies between words and their translations inside of a phrase well. However, dependencies across phrase boundaries are ignored because of the strong phrasal independence assumption. Consider the bilingual sentence pair shown in Figure 1(a).
Reordering of the German word stimmen is internal to the phrase-pair gegen ihre Kampagne stimmen -‘vote against your campaign’ and therefore represented by the translation model. However, the model fails to correctly translate the test sentence shown in Figure 1(b), which is translated as ‘they would for the legalization of abortion in Canada vote’, failing to displace the verb. The language model does not provide enough
evidence to counter the dispreference of the translation model against jumping over the source words für die Legalisieurung der Abtreibung in Kanada and translating stimmen - ‘vote’ at its correct position.
Weak Reordering Model. The lexicalized reordering model is primarily designed to deal with short-distance movement of phrases such as swapping two adjacent phrases and cannot properly handle long-range jumps. The model only learns an orientation of how a phrase was reordered with respect to its previous and next phrase; it makes independence assumptions over previously translated phrases and does not take into account how previous words were translated and reordered. Although such an independence assumption is useful to reduce sparsity, it is overly generalizing and does not help to disambiguate good reorderings from the bad ones.
Moreover, a vast majority of extracted phrases are singletons and the corresponding probability of orientation given phrase-pair estimates are based on a single observation. Due to sparsity, the model falls back to use one-word phrases instead, the orientation of which is ambiguous and can only be judged based on context that is ignored. This drawback has been addressed by Cherry (2013) by using sparse features for reordering models.
Hard Distortion Limit. The lexicalized reordering model fails to filter out bad largescale reorderings effectively (Koehn 2010). A hard distortion limit is therefore required during decoding in order to produce good translations. A distortion limit beyond eight words lets the translation accuracy drop because of search errors (Koehn et al. 2005). The use of a hard limit is undesirable for German–English and similar language pairs with significantly different syntactic structures. Several researchers have tried to address this problem. Moore and Quirk (2007) proposed improved future cost estimation to enable higher distortion limits in phrasal MT. Green, Galley, and Manning (2010) additionally proposed discriminative distortion models to achieve better translation accuracy than the baseline phrase-based system for a distortion limit of 15 words. Bisazza and Federico (2013) recently proposed a novel method to dynamically select which longrange reorderings to consider during the hypothesis extension process in a phrasebased decoder and showed an improvement in a German–English task by increasing the distortion limit to 18.
Spurious Phrasal Segmentation. A problem with the phrase-based model is that there is no unique correct phrasal segmentation of a sentence. Therefore, all possible ways of segmenting a bilingual sentence consistent with the word alignment are learned and used. This leads to two problems: (i) phrase frequencies are obtained by counting all possible occurrences in the training corpus, and (ii) different segmentations producing the same translation are generated during decoding. The former leads to questionable parameter estimates and the latter may lead to search errors because the probability of a translation is fragmented across different segmentations. Furthermore, the diversity in N-best translation lists is reduced. N-gram-based SMT (Mariño et al. 2006) uses an N-gram model that jointly generates the source and target strings as a sequence of bilingual translation units called tuples. Tuples are essentially minimal phrases, atomic units that cannot be decomposed any further. The tuples are generated left to right in target word order. Reordering is not
part of the statistical model. The parameters of the N-gram model are learned from bilingual data where the tuples have been arranged in target word order (see Figure 2).
Decoders for N-gram-based SMT reorder the source words in a preprocessing step so that the translation can be done monotonically. The reordering is performed with POS-based rewrite rules (see Figure 2 for an example) that have been learned from the training data (Crego and Mariño 2006). Word lattices are used to compactly represent a number of alternative reorderings. Using parts of speech instead of words in the rewrite rules makes them more general and helps to avoid data sparsity problems.
The mechanism has several useful properties. Because it is based on minimal units, there is only one derivation for each aligned bilingual sentence pair. The model therefore avoids spurious ambiguity. The model makes no phrasal independence assumption and generates a tuple monotonically by looking at a context of n previous tuples, thus capturing context across phrasal boundaries. On the other hand, N-gram-based systems have the following drawbacks.
Weak Reordering Model. The main drawback of N-gram-based SMT is its poor reordering mechanism. Firstly, by linearizing the source, N-gram-based SMT throws away useful information about how a particular word is reordered with respect to the previous word. This information is instead stored in the form of rewrite rules, which have no influence on the translation score. The model does not learn lexical reordering triggers and reorders through the learned rules only. Secondly, search is performed only on the precalculated word permutations created based on the source-side words. Often, evidence of the correct reordering is available in the translation model and the targetside language model. All potential reorderings that are not supported by the rewrite rules are pruned in the pre-processing step. To demonstrate this, consider the bilingual sentence pair in Figure 2 again. N-gram-based MT will linearize the word sequence gegen ihre Kampagne stimmen to stimmen gegen ihre Kampagne, so that it is in the same order as the English words. At the same time, it learns a POS rule: IN PRP NN VB → VB IN PRP NN. The POS-based rewrite rules serve to precompute the orderings that will be hypothesized during decoding. However, notice that this rule cannot generalize to the test sentence in Figure 1(b), even though the tuple translation model learned the trigram < sie – ‘they’ würden – ‘would’ stimmen – ‘vote’ > and it is likely that the monolingual language model has seen the trigram they would vote.
Hard Reordering Limit. Due to sparsity, only rules with seven or fewer tags are extracted. This subsequently constrains the reordering window to seven or fewer words, preventing the N-gram model from hypothesizing long-range reorderings that require
larger jumps. The need to perform long-distance reordering motivated the idea of using syntax trees (Crego and Mariño 2007) to form rewrite rules. However, the rules are still extracted ignoring the target-side, and search is performed only on the precalculated orderings.
Difficult Search Problem. Using MTUs makes the search problem much more difficult because of poor translation option selection. To illustrate this consider the phrase pair schoss ein Tor – ‘scored a goal’, consisting of units schoss – ‘scored’, ein – ‘a’, and Tor – ‘goal’. It is likely that the N-gram system does not have the tuple schoss – ‘scored’ in its N-best translation options because it is an uncommon translation. Even if schoss – ‘scored’ is hypothesized, it will be ranked quite low in the stack and may be pruned, before ein and Tor are generated in the next steps. A similar problem is also reported in Costa-jussà et al. (2007): When trying to reproduce the sentences in the N-best translation output of the phrase-based system, the N-gram-based system was able to produce only 37.5% of sentences in the Spanish-to-English and English-to-Spanish translation task, despite having been trained on the same word alignment. A phrase-based system, on the other hand, is likely to have access to the phrasal unit schoss ein Tor – ‘scored a goal’ and can generate it in a single step. 4 mtu-based search :We explored two decoding strategies in this work. Our first decoder complements the model and only uses minimal translation units in left-to-right stack-based decoding, similar to that used in Pharaoh (Koehn 2004a). The overall process can be roughly divided into the following steps: (i) extraction of translation units, (ii) future cost estimation, (iii) hypothesis extension, and (iv) recombination and pruning. The last two steps are repeated iteratively until all the words in the source sentence have been translated.
Our hypotheses maintain the index of the last source word covered (j), the position of the right-most source word covered so far (Z), the number of open gaps, the number of gaps so far inserted, the previously generated operations, the generated target string, and the accumulated values of all the features discussed in Section 3.6.1. The sequence of operations may include translation operations (generate, continue source cept, etc.) and reordering operations (gap insertions, jumps). Recombination6 is performed on hypotheses having the same coverage vector, monolingual language model context, and OSM context. We do histogram-based pruning, maintaining the 500 best hypotheses for each stack. A large beam size is required to cope with the search errors that result from using minimal translation units during decoding. We address this problem in Section 5.
6 Note that although we are using minimal translation units, recombination is still useful as different derivations can arise through different alignments between source and target fragments. Also, recombination can still take place if hypotheses differ slightly in the output (Koehn 2010). Aligned bilingual training corpora often contain unaligned target words and discontinuous target cepts, both of which pose problems. Unlike discontinuous source cepts, discontinuous target cepts such as hinunterschüttete – ‘poured . . . down’ in constructions like den Drink hinunterschüttete – ‘poured the drink down’ cannot be handled by the operation sequence model because it generates the English words in strict left-to-right order. Therefore they have to be eliminated.
Unaligned target words are only problematic for the MTU-based decoder, which has difficulties predicting where to insert them. Thus, we eliminate unaligned target words in MTU-based decoding.
We use a three-step process (Durrani, Schmid, and Fraser 2011) that modifies the alignments and removes unaligned and discontinuous targets. If a source word is aligned with multiple target words that are not consecutive, first the link to the least frequent target word is identified, and the group (consecutive adjacent words) of links containing this word is retained while the others are deleted. The intuition here is to keep the alignments containing content words (which are less frequent than functional words). For example, the alignment link hinunterschüttete – ‘down’ is deleted and only the link hinunterschüttete – ‘poured’ is retained because ‘down’ occurs more frequently than ‘poured’. Crego and Yvon (2009) used split tokens to deal with this phenomenon.
For MTU-based decoding we also need to deal with unaligned target words. For each unaligned target word, we determine the (left or right) neighbor that it appears more frequently with and align it with the same source word as this neighbor. Crego, de Gispert, and Mariño (2005) and Mariño et al. (2006) instead used lexical probabilities p( f |e) obtained from IBM Model 1 (Brown et al. 1993) to decide whether to attach left or right. A more sophisticated strategy based on part-of-speech entropy was proposed by Gispert and Mariño (2006). We evaluated our systems on German-to-English, French-to-English, and Spanish-toEnglish news translation for the purpose of development and evaluation. We used data from the eighth version of the Europarl Corpus and the News Commentary made available for the translation task of the Eighth Workshop on Statistical Machine Translation.7 The bilingual corpora contained roughly 2M bilingual sentence pairs, which we obtained by concatenating news commentary (≈ 184K sentences) and Europarl for the estimation of the translation model. Word alignments were generated with GIZA++ (Och and Ney 2003), using the grow-diag-final-and heuristic8 (Koehn et al. 2005). All data are lowercased, and we use the Moses tokenizer. We took news-test-2008 as the dev set for optimization and news-test 2009-2012 for testing. The feature weights are tuned with Z-MERT (Zaidan 2009).
4.2.1 Baseline Systems. We compared our system with (i) Moses9 (Koehn et al. 2007), (ii) Phrasal10 (Cer et al. 2010), and (iii) Ncode11 (Crego, Yvon, and Mariño 2011). We used
7 http://www.statmt.org/wmt13/translation-task.html 8 We also tested other symmetrization heuristics such as “Union” and “Intersection” but found the GDFA
heuristic gave best results for all language pairs. 9 http://www.statmt.org/moses/
10 http://nlp.stanford.edu/phrasal/ 11 http://www.limsi.fr/Individu/jmcrego/bincoder/
all these toolkits with their default settings. Phrasal provides two main extensions to Moses: a hierarchical reordering model (Galley and Manning 2008) and discontinuous source and target phrases (Galley and Manning 2010). We used the default stack sizes of 100 for Moses,12 200 for Phrasal, and 25 for Ncode (with 2n stacks). A 5-gram English language model is used. Both phrase-based systems use the 20 best translation options per source phrase; Ncode uses the 25 best tuple translations and a 4-gram tuple sequence model. A hard distortion limit of 6 is used in the default configuration of both phrasebased systems. Among the other defaults, we retained the hard source gap penalty of 15 and a target gap penalty of 7 in Phrasal. We provide Moses and Ncode with the same post-edited alignments13 from which we had removed target-side discontinuities. We feed the original alignments to Phrasal because of its ability to learn discontinuous source and target phrases. All the systems use MERT for the optimization of the weight vector.
4.2.2 Training. Training steps include: (i) post-editing of the alignments (Section 4.1), (ii) generation of the operation sequence (Algorithm 1), and (iii) estimation of the N-gram translation (OSM) and language models using the SRILM toolkit (Stolcke 2002) with Kneser-Ney smoothing. We used 5-gram models.
4.2.3 Summary of Developmental Experiments. During the developent of the MTU-based decoder, we performed a number of experiments to obtain optimal settings for the system. We list here a summary of the results from those experiments:
We found that discontinuous source-side cepts do not improve translation quality in most cases but increase the decoding time by multiple folds. We will therefore only use continuous cepts.
We performed experiments by varying the distortion limit from the conventional window of 6 words to infinity (= no hard limit). We found that the performance of our system is robust when removing the hard reordering constraint and even saw a slight improvement in results in the case of German-to-English systems. Using no distortion limit, however, significantly increases the decoding time. We will therefore use a window of 16 words, which we found to be optimal on the development set.
The performance of the MTU-based decoder is sensitive to the stack size. A high limit of 500 is required for decent search accuracy. We will discuss this further in the next section.
We found using 10 best translation options for each extracted cept during decoding to be optimal.
4.2.4 Comparison with the Baseline Systems. In this section we compare our system (OSMmtu) with the three baseline systems. We used Kevin Gimpel’s tester,14 which uses bootstrap resampling (Koehn 2004b) to test which of our results are significantly better than the baseline results. We mark a baseline result with “*” in order to indicate
12 Using stack sizes from 200–1,000 did not improve results. 13 Using post-processed alignments gave better results than using the original alignments for these baseline
systems. 14 http://www.ark.cs.cmu.edu/MT/
that our model shows a significant improvement over this baseline with a confidence of p < 0.05. We use 1,000 samples during bootstrap resampling.
Our German-to-English results (see Table 2) are significantly better than the baseline systems in most cases. Our French-to-English results show a significant improvement over Moses in three out of four cases, and over Phrasal in half of the cases. The N-gram-based system NCode was better or similar to our system on the French task. Our Spanish-to-English system also showed roughly the same translation quality as the baseline systems, but was significantly worse on the WMT12 task. 5 phrase-based search :The MTU-based decoder is the most straightforward implementation of a decoder for the operation sequence model, but it faces search problems that cause a drop in translation accuracy. Although the OSM captures both source and target contexts and provides a better reordering mechanism, the ability to memorize and produce larger translation units gives an edge to the phrase-based model during decoding in terms of better search performance and superior selection of translation units. In this section, we combine N-gram-based modeling with phrase-based decoding. This combination not only improves search accuracy but also increases translation quality in terms of BLEU.
The operation sequence model, although based on minimal translation units, can learn larger translation chunks by memorizing a sequence of operations. However, it often has difficulties to produce the same translations as the phrase-based system because of the following drawbacks of MTU-based decoding: (i) the MTU-based decoder does not have access to all the translation units that a phrase-based decoder uses as part of a larger phrase, (ii) it requires a larger beam size to prevent early pruning of correct
hypotheses, and (iii) it uses less-powerful future-cost estimates than the phrase-based decoder. To demonstrate these problems, consider the phrase pair
which the model memorizes through the sequence:
Generate(Wie, What is) Insert Gap Generate (Sie, your) Jump Back (1) Generate (heissen, name)
The MTU-based decoder needs three separate tuple translations to generate the same phrasal translation: Wie – ‘What is’, Sie – ‘your’ and heißen – ‘name’. Here we are faced with three challenges.
Translation Coverage: The first problem is that the N-gram model does not have the same coverage of translation options. The English cepts ‘What is’, ‘your’, and ‘name’ are not good candidate translations for the German cepts Wie, Sie, and heißen, which are usually translated to ‘How’, ‘you’, and ‘call’, respectively, in isolation. When extracting tuple translations for these cepts from the Europarl data for our system, the tuple Wie – ‘What is’ is ranked 124th, heißen – ‘name’ is ranked 56th, and Sie – ‘your’ is ranked 9th in the list of n-best translation candidates. Typically, only the 20 best translation options are used, for the sake of efficiency, and such phrasal units with less frequent translations are never hypothesized in the N-gram-based systems. The phrase-based system, on the other hand, can extract the phrase Wie heißen Sie – ‘what is your name’ even if it is observed only once during training.
Larger Beam Size: Even when we allow a huge number of translation options and therefore hypothesize such units, we are faced with another challenge. A larger beam size is required in MTU-based decoding to prevent uncommon translations from getting pruned. The phrase-based system can generate the phrase pair Wie heißen Sie – ‘what is your name’ in a single step, placing it directly into the stack three words to the right. The MTU-based decoder generates this phrase in three stacks with the tuple translations Wie – ‘What is’, Sie – ‘your’, and heißen – ‘name’. A very large stack size is required during decoding to prevent the pruning of Wie – ‘What is’, which is ranked quite low in the stack until the tuple Sie – ‘your’ is hypothesized in the next stack. Although the translation quality achieved by phrase-based SMT remains the same when varying the beam size, the performance of our system varies drastically with different beam sizes (especially for the German–English experiments where the search is more difficult due to a higher number of reorderings). Costa-jussà et al. (2007) also report a significant drop in the performance of N-gram-based SMT when a beam size of 10 is used instead of 50 in their experiments.
Future Cost Estimation: A third problem is caused by inaccurate future cost estimation. Using phrases helps phrase-based SMT to better estimate the future language model cost because of the larger context available, and allows the decoder to capture local (phrase-internal) reorderings in the future cost. In comparison, the future cost for tuples is based on unigram probabilities. The future cost estimate for the phrase pair Wie heißen Sie – ‘What is your name’ is estimated by calculating the cost of each feature.
A bigram language model cost, for example, is estimated in the phrase-based system as follows:
plm = p(What) × p(is|What) × p(your|What is) × p(name|What is your)
The translation model cost is estimated as:
ptm = p(What is your name|Wie heißen Sie)
Phrase-based SMT is aware during the preprocessing step that the words Wie heißen Sie may be translated as a phrase. This is helpful for estimating a more accurate future cost because the context is already available. The same is not true for the MTU-based decoder, to which only minimal units are available. The MTU-based decoder does not have the information that Wie heißen Sie may be translated as a phrase during decoding. The future cost estimate available to the operation sequence model for the span covering Wie heißen Sie will have unigram probabilities for both the translation and language models.
plm = p(What) × p(is|What) × p(your) × p(name)
The translation model cost is estimated as:
ptm = p(Generate(Wie, What is)) × p(Generate(heißen,name)) × p(Generate(Sie, your))
A more accurate future cost estimate for the translation model cost would be:
ptm = p(Generate(Wie,What is)) × p(Insert Gap|C2) × p(Generate(Sie,your)|C3) × p(Jump Back(1)|C4) × p(Generate(heißen,name)|C5)
where Ci is the context for the generation of the ith operation—that is, up to m previous operations. For example C1 = Generate(Wie, What is), C2 = Generate(Wie,What is) Insert Gap, and so on. The future cost estimates computed in this manner are much more accurate because not only do they consider context, but they also take the reordering operations into account (Durrani, Fraser, and Schmid 2013). We extended our in-house OSM decoder to use phrases instead of MTUs during decoding. In order to check whether phrase-based decoding solves the mentioned problems and improves the search accuracy, we evaluated the baseline MTU decoder and the phrase-based decoder with the same model parameters and tuned weights. This allows us to directly compare the model scores. We tuned the feature weights by running MERT with the MTU decoder on the dev set. Table 3 shows results from running both, the MTU-based (OSMmtu) and the phrase-based (OSMphr) decoder, on the WMT09 test set. Improved search accuracy is the percentage of times each decoder was able to produce a better model score than the other. Our phrase-based decoder uses a stack size of 200. Table 3 shows the percentage of times the MTU-based and phrase-based decoder produce better model scores than their counterpart. It shows that the phrase-based decoder produces better model scores for almost 48% of the hypotheses (on average)
across the three language pairs, whereas the MTU-based decoder (using a much higher stack size [500]) produces better hypotheses 8.2% of the time on average.
This improvement in search is also reflected in translation quality. Our phrase-based decoder outperforms the MTU-based decoder in all the cases and gives a significant improvement in 8 out of 12 cases (Table 4). In Section 4.1 we discussed the problem of handling unaligned and discontinuous target words in MTU-based decoding. An advantage of phrase-based decoding is that we can use such units during decoding if they appear within the extracted phrases. We use a Generate Target Only (Y) operation whenever the unaligned target word Y occurs in
a phrase. Similarly, we use the operation Generate (hinunterschüttete, poured down) when the discontinuous tuple hinunterschüttete – ‘poured ... down’ occurs in a phrase. While training the model, we simply ignore the discontinuity and pretend that the word ‘down’ immediately follows ‘poured’. This can be done by linearizing the subsequent parts of discontinuous target cepts to appear after the first word of the cept. During decoding we use phrase-internal alignments to hypothesize such a linearization. This is done only for the estimation of the OSM, and the target for all other purposes is generated in its original order. This heuristic allows us to deal with target discontinuities without extending the operation sequence model in complicated ways. It results in better BLEU accuracy in comparison with the post-editing of the alignments method described in Section 4.1. For details and empirical results refer to Durrani et al. (2013a) (see Table 2 therein, compare Rows 4 and 5).
Note that the OSM, like the discontinuous phrase-based model (Galley and Manning 2010), allows all possible geometries as shown in Figure 7. However, because our decoder only uses continuous phrases, we cannot hypothesize (ii) and (iii) unless they appear inside of a phrase. But our model could be integrated into a discontinuous phrase-based system to overcome this limitation. 7 conclusion :In this article we presented a new model for statistical MT that combines the benefits of two state-of-the-art SMT frameworks, namely, N-gram-based and phrase-based SMT. Like the N-gram-based model, it addresses two drawbacks of phrasal MT by better handling dependencies across phrase boundaries, and solving the phrasal segmentation problem. In contrast to N-gram-based MT, our model has a generative story that tightly couples translation and reordering. Furthermore, it is able to consider all possible reorderings, unlike N-gram systems that perform search only on a limited number of pre-calculated orderings. Our model is able to correctly reorder words across large distances, and it memorizes frequent phrasal translations including their reordering as probable operation sequences.
We tested a version of our system that decodes based on minimal translation units (MTUs) against the state-of-the-art phrase-based systems Moses and Phrasal and the N-gram-based system Ncode for German-to-English, French-to-English, and Spanishto-English on three standard test sets. Our system shows statistically significant improvements in 9 out of 12 cases in the German-to-English translation task, and 10 out of 12 cases in the French-to-English translation task. Our Spanish-to-English results are similar to the baseline systems in most of the cases but consistently worse than Ncode.
MTU-based decoding suffers from poor translation coverage, inaccurate future cost estimates, and pruning of correct hypotheses. Phrase-based SMT, on the other hand, avoids these drawbacks by using larger translation chunks during search. We
therefore extended our decoder to use phrases instead of cepts while keeping the statistical model unchanged. We found that combining a model based on minimal units with phrase-based decoding improves both search accuracy and translation quality. Our system extended with phrase-based decoding showed improvements over all the baseline systems, including our MTU-based decoder. In most of the cases, the difference was significant.
Our results show that OSM consistently outperforms the Moses lexicalized reordering model and gives statistically significant gains over a very competitive Moses baseline system. We showed that considering both translation and reordering context is important and ignoring reordering context results in a significant reduction in the performance. We also showed that an OSM based on surface forms suffers from data sparsity and that an OSM based on a generalized representation with part-of-speech tags improves the translation quality by considering a larger context. In the future we would like to study whether the insight of using minimal units for modeling and search based on composed rules would hold for hierarchical SMT. Vaswani et al. (2011) recently showed that a Markov model over the derivation history of minimal rules can obtain the same translation quality as using grammars formed with composed rules, which we believe is quite promising."
"1 introduction :Almost 50 years since Damerau’s groundbreaking work was published (Damerau 1964), the figures he set for the proportion of spelling errors in typed text, along with their classification, still remain in use. According to Damerau, over 80% of all misspelled words present a single error which, in turn, falls into one out of four categories, to wit, Insertion (an extra letter is inserted), Omission (one letter is missing), Substitution (one letter is wrong), and Transposition (two adjacent characters are transposed). In fact, these very same figures lie at the heart of much current mainstream research on related topics, from automatic text spelling correction (e.g., Pirinen and Lindén 2014) to more advanced information retrieval techniques (e.g., Stein, Hoppe, and Gollub 2012) to optical character recognition techniques (e.g., Reynaert 2011). The reason for such popularity may rest in the simplicity of the approach, whereby numbers can be assigned
∗ EACH–USP, Arlindo Béttio, 1000. 03828-000. São Paulo, SP – Brazil. E-mail: norton@usp.br.
Submission received: 30 January 2014; revised submission received: 21 July 2014; accepted for publication: 18 September 2014.
doi:10.1162/COLI a 00216
© 2015 Association for Computational Linguistics
to the probability that a certain type of spelling error might take place, simply because that seems to be the frequency with which people make that kind of mistake.
Surprisingly enough, even though Damerau derived his statistics uniquely from texts in English, his findings are applied, with very little to no adaptation at all, to research in a range of different languages, such as Basque (e.g., Aduriz et al. 1997), Persian (e.g., Miangah 2014), and Arabic (e.g., Alkanhal et al. 2012). The question then becomes how appropriate these figures are when applied to languages other than English. In fact, some researchers have already noticed this potential flaw, and tried to adapt Damerau’s findings to their own language, usually by modifying DamerauLevenstein edit distance (Wagner and Fischer 1974) to match some language-specific data, but still without verifying the appropriateness of Damerau’s statistics in the target language (e.g., Rytting et al. 2011; Richter, Stranák, and Rosen 2012), or by taking into account some feature of that language, such as the presence of diacritics, for example (e.g., Andrade et al. 2012).
In this article, we move a step further by analyzing a set of typed texts from two different corpora in Brazilian Portuguese—a language spoken by over 190 million people1—and deriving statistics tailored to this language. We then compare our statistics with results obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006). As we will show, the behavior demonstrated by native speakers of these languages follow a very similar pattern, straying from Damerau’s original findings while, at the same time, holding them still useful. As a limitation, in this research we only account for non-word errors, that is, errors that, once made, result in some non-existing word, and which may be detected by a regular dictionary look-up (Deorowicz and Ciura 2005).
As an indication of the impact of these results, we added a new module (Gimenes, Roman, and Carvalho 2014)2 to OpenOffice Writer3—a freely available word processor—so as to reorder its suggestions list according to the statistics presented in this article. Within Writer, when typing Pao, instead of Pão (‘bread’ in Portuguese), the correct form comes in fifth place in the suggestion list, with Poca, Pocá, Poça, and Próça coming in the first four positions. With this module, it was observed that, for the first testing corpus, in 27.34% of the cases there was an improvement over Writer’s ranking (i.e., the correct form was ranked higher in the list), whereas in 5.84% the new list was actually worse, and in 66.82% no changes were observed. In the second corpus, an improvement was observed in 19.90% of the cases, in 9.00% figures were worse, with 71.10% remaining unchanged. Some 10–21% increase in accuracy is not something to be neglected, especially at the price of changing weights in an edit distance scheme. ","2 related work :One of the first efforts to derive statistics similar to those by Damerau in a language other than English was made by Medeiros (1995), who analyzed a set of corpora in European Portuguese. In his work, Medeiros found that 80.1% of all spelling errors would fit into one of Damerau’s categories. He also noticed that around 20% of all linguistic errors would be related to the use of diacritics, meaning that they should be taken into account during string edit distance or probability calculations. By
1 According to 2010’s demographic census: http://www.ibge.gov.br/home/estatistica/populacao/ censo2010/caracteristicas_da_populacao/resultados_do_universo.pdf. 2 Available at http://ppgsi.each.usp.br/arquivos/RelTec/PPgSI-001_2014.pdf. 3 https://www.openoffice.org/.
linguistic errors, the author meant linguistically motivated errors, as opposed to those caused by slipping fingers in a keyboard, for example. However illustrative, Medeiros’s findings suffer from a major drawback, to the extent that he relied on pre-existing lists of erroneous words in the related literature, along with handwritten errors made by high school students, with only part of the data coming from e-mail exchanges—that is, from actual typed text. The problem with this approach is that these data come without any frequency analysis, which renders them inappropriate for the application of computer techniques based on statistics, also making it hard to compare to any findings related to error frequency.
Spanish was another language found to confirm Damerau’s findings (Bustamante, Arnaiz, and Ginés 2006). In that case, however, the authors made a much more detailed categorization of the errors they found, beyond the four classes originally proposed by Damerau, in a corpus of written (presumably typed) texts from Mexico and Spain. Still, a careful inspection into these categories (by grouping insertions and repetitions of the same letter, and taking errors related to the misuse of diacritics, capitalization, and cognitive replacements to be substitutions) leads one back to Damerau’s four basic errors. In that case, a total of 86.7% of all errors were found to fit into one of these groups. The main difference, however, between both Damerau and Medeiros, and the work by Bustamante, Arnaiz, and Ginés (2006), is that the majority of spelling errors (54.9% of all errors, corresponding to around 49% of all single errors) were related to the misuse of diacritics, with special emphasis on their omission, which was responsible for 51.5% of all spelling errors (i.e., around 46% of all single errors) found in the corpus.
Finally, one of the latest additions to this list was made by Baba and Suzuki (2012), in which the authors analyzed spelling errors both in English and Japanese, with the last one typed on a Roman keyboard (that is, transliterated). Interestingly, the authors report almost no difference in the distribution of errors among Damerau’s four classes, both for English and Japanese. By looking at the details, however, they found some specific errors in Japanese, which they attribute to the phonological and orthographic characteristics of the language. In fact, a breakdown analysis of the four classes show accentuated differences in the type of substitutions (that is, a vowel by another vowel, a vowel by a consonant, etc.), insertions (inserted vowel vs. consonant), deletions (whether or not the deleted character was a repetition of some of its neighbors), and transpositions (adjacent characters versus syllables) that were made.
As we will show in the following sections, this overall confirmation of Damerau’s findings, but with remarked differences caused by each language’s idiosyncrasies, seems to be the common behavior both in the related work and ours. This, in turn, may shed some light on the reasons why English-based spelling checking software does not seem to perform so badly in other languages, but still not as well as it could, should each language’s own characteristics be taken into account. in C2 for single errors—still a substantial proportion. Interestingly, if one takes errors involving diacritics to be substitutions, then we end up with a total of 89.86% of all single errors falling into one of Damerau’s categories in C1, with 88.91% in C2, thereby confirming Damerau’s statistics. An analysis of the distribution of single errors among Damerau’s four original categories (i.e., disregarding diacritic-related misspellings), and the distribution among the categories related to the use of diacritics also shows no statistically significant difference between both corpora (ks = 0.5, p = 0.6994 for Damerau’s categories and ks = 0.6, p = 0.3291 for the diacritics—including cedilla—set).
Finally, regarding the number of misspelled words, we have once again confirmed Damerau’s results, in that over 85% of all wrong words, be they repetition of existing words or not, have a single spelling error. Table 3 shows these results. In this table,6 we present the figures both when taking word repetitions into account and when ruling them out (that is, when keeping the statistics only for new words). This is also in line with the results obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006), in which it was found that 86.7% of all spelling errors were single errors.
6 Total = 966 × 1 + 70 × 2 + 7 × 3 = 1, 127 errors, with another 12 distributed in the last category, for C1, and 911 × 1 + 95 × 2 + 37 × 3 = 1, 212 errors, with another 48 distributed in the last category, for C2.","3 materials and methods :Because we were interested in source texts written in Brazilian Portuguese, where authors were free to type with little to no intervention by spelling correction facilities, we decided to use the corpus presented in Roman et al. (2013), which comprises 1,808 texts, typed in by 452 different participants in a Web experiment, with a total of 62,858 words in length. As a source, this corpus has the advantages of (1) being produced by a number of different people of different ages, educational attainment, and background;
(2) being produced through a Web browser, without any help from spelling correction modules usually present in text editors; and (3) being freely available for download over the Web. We will refer to this corpus as C1.
In order to verify the statistics gathered on C1, in June 2011 we collected a corpus of blog posts from four different Web sites4 that describe travel diaries along with comments by visitors. With a total of 26,418 words, spread over 192 posts, and being written in Brazilian Portuguese, this corpus presents the same advantages as C1, except for the fact that we cannot verify participants’ details, which means we cannot make any statement on the participants’ distribution according to age, background, and so forth. Still, the main characteristics of availability and free text production remain. We will refer to this corpus as C2. By comparing the statistics from both corpora we intend to reduce the effect of any bias related to writing style and characteristics of the participants.
Given that our goal was to identify non-word errors, we relied on a commercial text editor to highlight misspellings in both corpora. The highlighted words were then manually inspected by one of the researchers, and errors were grouped in categories. Besides Damerau’s four original sets, we identified three extra categories: errors involving the use of diacritics, errors related to the use of the cedilla, and space-related errors. The first of these groups was inspired by the observation that the use of diacritics is a key issue in Portuguese (Andrade et al. 2012), potentially playing an important role in the making of spelling mistakes. Following Damerau, this group can be subdivided into four other subcategories: missing diacritic, addition of diacritic, right diacritic applied to the wrong character, and wrong diacritic applied to the right character.
The second group concerns the misuse of the cedilla—a special diacritical mark that in Portuguese can only be applied to the character c, resulting in ç. The reason for this character deserving a whole category of its own lies in the existence of two distinct keyboard layouts in the Brazilian market: ABNT-2 and US-accents. The main difference between both layouts is that whereas ABNT-2 presents a separate key for ç, that key does not exist on the US-Accents keyboard. Hence, while on ABNT-2 the user hits a single key to get a ç, on US-Accents it is a composite—two keys must be pressed: the single quotation mark and c. Ultimately, errors involving the cedilla may be interpreted both as a regular character mistake, or an error related to the use of diacritics, depending on the keyboard.
Finally, because space-related errors, although not so common (see Section 4), are usually difficult to deal with via spelling correctors, we decided to give them an independent set of categories. As a mistake, spaces have the annoying characteristic of not being handled by edit distance operations (Attia et al. 2012), thereby passing undisturbed by systems that use string distance-based ranking when giving alternative spellings to the user. The reason for this problem is that, by joining two words together, for example, the number of mistakes dramatically rises, if one takes that new string to be a single word. Along with these three categories, there is a “leftover” fourth one— others—comprising linguistic expression errors, first-letter capitalization, and wrong diminutive or augmentative.
4 http://bragatte.wordpress.com/. http://guilhermebragatte.blogspot.com.br/. http://forasteironairlanda.wordpress.com/tag/nomadismo/. http://sussuemdublin.wordpress.com/2011/01/.","4 results and discussion :Results both for C1 and C2 are shown in Tables 1 and 2.5 Even though the total number of mistakes is almost the same in both corpora (1,139 for C1 vs. 1,260 for C2), the proportion of spelling errors in C2 is more than twice as high as that of C1. In this case, C1 had a mean error rate of 1.81% (that is, 0.0181 error per word), while C2 reached 4.77%. This difference may be due to the fact that C1 was generated by students from a university, that is, people with a higher educational level. It may also have something to do with the fact that blogs are more conversation-like, which may lead people to relax the spelling rules they choose to follow, especially when it comes to errors involving diacritics. Still, despite this discrepancy, a two-sample KolmogorovSmirnov test showed no statistically significant differences between these corpora, on the distribution of the proportion of errors among categories, neither for the total number of errors (ks = 0.4286, p = 0.1528), nor for the number of errors per word (ks = 0.4286, p = 0.1528 for single, ks = 0.2143, p = 0.9048 for double, ks = 0.2857, p = 0.6172 for triple, and ks = 0.1429, p = 0.9988 for multiple errors).
As it turns out, errors related to the use of diacritics correspond to approximately half of all spelling errors in both corpora (overall, 47.15% in C1 and 50.08% in C2, corresponding to 49.48% of all single errors in C1 and 58.84% in C2), if we include in this sum cedilla-related errors. By taking the cedilla as a simple substitution, the numbers drop to 41.09% in C1 and 45.63% in C2 (overall), with 44.72% in C1 and 57.08%
5 In these tables, for example, 22 insertions found in words with two errors mean that, from all errors found in such words, 22 were insertions, disregarding whether a word presents two insertions or an insertion along with some other type of error. To save space, rows filled with 0, that is categories with no examples, were removed. Space-related errors were also broken down into Substitution by Space, as in pre qualified instead of pre-qualified; Space Insertion, as in w ord instead of word; Space Transposition, as in m yword instead of my word; and Missing Space, as in myword. For reasons noted in Section 2, we have chosen the work by Bustamante, Arnaiz, and Ginés (2006), on European and Mexican Spanish, as a benchmark against which to compare ours. Table 4 shows the comparison results. In this case, for the sake of comparison, and given that each research has a classification scheme of its own, we had both to reorganize some of the categories and present the figures in terms of number of words, instead of number of errors. Although results are shown considering cedilla-related errors to be cases of a missing diacritic, whenever appropriate the figures for taking such errors as simple substitutions are also presented within parentheses.
By comparing the data distributions in Table 4, one sees no differences7 between our results (both for C1 and C2) and those for Spanish, no matter whether one takes cedillarelated errors to be a missing diacritic or a substitution. This, in turn, might indicate that such errors and their distributions are not related to the language or culture themselves, but instead to the set of characters (and corresponding diacritical marks) allowed within each specific language. Also, the data show the importance of taking diacritical marks into consideration when designing spelling correction systems for these languages. As it turned out, this type of error is responsible for over 40% of all wrong words, both in Brazilian Portuguese and Spanish.","5 conclusion :In this article we presented some statistics collected from two different corpora of typed texts in Brazilian Portuguese. Our first contribution is a description of the error rate distribution among the four original categories defined by Damerau (1964), along with the distribution of errors related to the misuse of diacritical marks. Results show not only that Damerau’s figures still hold, should we put aside diacritics, but also make a
7 With cedilla being a diacritic omission: ks = 0.3, p = 0.7591 both between Spanish and C1, and between Spanish and C2. With cedilla being a character substitution: ks = 0.3, p = 0.7591 both between Spanish and C1, and between Spanish and C2.
point about the importance such marks have on the misspelling of words in Brazilian Portuguese, as indicated by the frequency with which this type of error is made, thereby rendering spelling correction systems that rely solely on Damerau’s results unfit for this language. On this account, a straightforward experiment with a commercially available text editor has shown a 10–21% improvement, depending on the testing corpus, in the ranking of suggestions for misspelled words.
As an additional contribution, we have shown that our results are very much like those obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006), in that we could see no statistically significant difference on error distribution between these two languages. This, in turn, may be taken as an indication that the distribution of errors does not depend on language or culture, but instead on the character set that people are allowed to use. As such, it would not come as a surprise to find out that the same distribution can also be observed in other languages, such as Italian, French, and, to some extent, Turkish, for example, which make an intensive use of a similar set of characters. This is something that is worth investigating, and we leave it for future research.",,,,"Fifty years after Damerau set up his statistics for the distribution of errors in typed texts, his findings are still used in a range of different languages. Because these statistics were derived from texts in English, the question of whether they actually apply to other languages has been raised. We address this issue through the analysis of a set of typed texts in Brazilian Portuguese, deriving statistics tailored to this language. Results show that diacritical marks play a major role, as indicated by the frequency of mistakes involving them, thereby rendering Damerau’s original findings mostly unfit for spelling correction systems, although still holding them useful, should one set aside such marks. Furthermore, a comparison between these results and those published for Spanish show no statistically significant differences between both languages—an indication that the distribution of spelling errors depends on the adopted character set rather than the language itself.","[{""affiliations"": [], ""name"": ""Priscila A. Gimenes""}, {""affiliations"": [], ""name"": ""Norton T. Roman""}, {""affiliations"": [], ""name"": ""M. B. R. Carvalho""}]",SP:1759bf80767626c66b77ddf36a4d2d9c62f4f022,"[{""authors"": [""Aduriz"", ""Itziar"", ""I\u00f1aki Alegria"", ""Xabier Artola"", ""Nerea Ezeiza"", ""Kepa Sarasola"", ""Miriam Urkia.""], ""title"": ""A spelling corrector for Basque based on morphology"", ""venue"": ""Literary and Linguistic Computing,"", ""year"": 1997}, {""authors"": [""Alkanhal"", ""Mohamed I"", ""Mohamed A. Al-Badrashiny"", ""Mansour M. Alghamdi"", ""Abdulaziz O. Al-Qabbany""], ""title"": ""Automatic stochastic Arabic spelling correction with emphasis on space"", ""year"": 2012}, {""authors"": [""Andrade"", ""Guilherme"", ""F. Teixeira"", ""C.R. Xavier"", ""R.S. Oliveira"", ""Leonardo C. da Rocha"", ""A.G. Evsukoff""], ""title"": ""Hasch: High performance automatic spell checker for Portuguese texts"", ""year"": 2012}, {""authors"": [""Attia"", ""Mohammed"", ""Pavel Pecina"", ""Younes Samih"", ""Khaled Shaalan"", ""Josef van Genabith.""], ""title"": ""Improved spelling error detection and correction for Arabic"", ""venue"": ""Proceedings of COLING-2012,"", ""year"": 2012}, {""authors"": [""Baba"", ""Yukino"", ""Hisami Suzuki.""], ""title"": ""How are spelling errors generated and corrected? A study of corrected and uncorrected spelling errors using keystroke logs"", ""venue"": ""Proceedings of ACL-2012,"", ""year"": 2012}, {""authors"": [""Bustamante"", ""Flora Ram\u0131\u0301rez"", ""Alfredo Arnaiz"", ""Mar Gin\u00e9s""], ""title"": ""A spell checker for a world language: The new Microsoft\u2019s Spanish spell checker"", ""venue"": ""In Proceedings of LREC-2006,"", ""year"": 2006}, {""authors"": [""Damerau"", ""Fred J.""], ""title"": ""A technique for computer detection and correction of spelling errors"", ""venue"": ""Communications of the ACM, 7(3):171\u2013176."", ""year"": 1964}, {""authors"": [""Deorowicz"", ""Sebastian"", ""Marcin G. Ciura.""], ""title"": ""Correcting spelling errors by modeling their causes"", ""venue"": ""International Journal of Applied Mathematics and Computer Science, 15(2):275\u2013285."", ""year"": 2005}, {""authors"": [""Gimenes"", ""Priscila Azar"", ""Norton Trevisan Roman"", ""Ariadne Maria Brito Rizzoni Carvalho.""], ""title"": ""An OO writer module for spelling correction in Brazilian Portuguese"", ""venue"": ""Technical Report PPgSI-001/"", ""year"": 2014}, {""authors"": [""Medeiros"", ""Jos\u00e9 Carlos Dinis.""], ""title"": ""Processamento morfol\u00f3gico e correc c\u00e3o ortogr\u00e1fica do portugu\u00eas"", ""venue"": ""Master\u2019s thesis, Instituto Superior T\u00e9cnico \u2013 Universidade T\u00e9cnica de Lisboa, February."", ""year"": 1995}, {""authors"": [""Miangah"", ""Tayebeh Mosavi.""], ""title"": ""Farsispell: A spell-checking system for Persian using a large monolingual corpus"", ""venue"": ""Literary and Linguistic Computing, 29(1):56\u201373."", ""year"": 2014}, {""authors"": [""Pirinen"", ""Tommi A."", ""Krister Lind\u00e9n.""], ""title"": ""State-of-the-art in weighted finite-state spell-checking"", ""venue"": ""Alexander Gelbukh, editor, Computational Linguistics and Intelligent Text Processing,"", ""year"": 2014}, {""authors"": [""Reynaert"", ""Martin W.C.""], ""title"": ""Character confusion versus focus word-based correction of spelling and OCR variants in corpora"", ""venue"": ""International Journal on Document Analysis and Recognition,"", ""year"": 2011}, {""authors"": [""Richter"", ""Michal"", ""Pavel Stran\u00e1k"", ""Alexandr Rosen.""], ""title"": ""Korektor\u2014a system for contextual spell-checking and diacritics completion"", ""venue"": ""Proceedings of COLING-2012, pages 1,019\u20131,028, Mumbai."", ""year"": 2012}, {""authors"": [""Roman"", ""Norton Trevisan"", ""Paul Piwek"", ""Alexandre Rossi Alvares""], ""title"": ""Introducing a corpus of human-authored dialogue summaries in Portuguese"", ""year"": 2013}, {""authors"": [""Christian Hettick"", ""Tim Buckwalter"", ""Charles C. Blake.""], ""title"": ""Spelling correction for dialectal Arabic dictionary lookup"", ""venue"": ""ACM Transactions on Asian Language Information Processing,"", ""year"": 2011}, {""authors"": [""Stein"", ""Benno"", ""Dennis Hoppe"", ""Tim Gollub.""], ""title"": ""The impact of spelling errors on patent search"", ""venue"": ""Proceedings of EACL-2012, pages 570\u2013579, Avignon."", ""year"": 2012}, {""authors"": [""Wagner"", ""Robert A."", ""Michael J. Fischer.""], ""title"": ""The string-to-string correction problem"", ""venue"": ""Journal of the ACM, 21(1): 168\u2013173. 183"", ""year"": 1974}]",,spelling error patterns in :,brazilian portuguese :Priscila A. Gimenes,"each / usp :Norton T. Roman∗ Ariadne M. B. R. Carvalho Institute of Computing / Unicamp
Fifty years after Damerau set up his statistics for the distribution of errors in typed texts, his findings are still used in a range of different languages. Because these statistics were derived from texts in English, the question of whether they actually apply to other languages has been raised. We address this issue through the analysis of a set of typed texts in Brazilian Portuguese, deriving statistics tailored to this language. Results show that diacritical marks play a major role, as indicated by the frequency of mistakes involving them, thereby rendering Damerau’s original findings mostly unfit for spelling correction systems, although still holding them useful, should one set aside such marks. Furthermore, a comparison between these results and those published for Spanish show no statistically significant differences between both languages—an indication that the distribution of spelling errors depends on the adopted character set rather than the language itself.",error category single two three over three total (%) : ,errors words in c1 (%) words in c2 (%) words in c1 (%) words in c2 (%) :,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :Almost 50 years since Damerau’s groundbreaking work was published (Damerau 1964), the figures he set for the proportion of spelling errors in typed text, along with their classification, still remain in use. According to Damerau, over 80% of all misspelled words present a single error which, in turn, falls into one out of four categories, to wit, Insertion (an extra letter is inserted), Omission (one letter is missing), Substitution (one letter is wrong), and Transposition (two adjacent characters are transposed). In fact, these very same figures lie at the heart of much current mainstream research on related topics, from automatic text spelling correction (e.g., Pirinen and Lindén 2014) to more advanced information retrieval techniques (e.g., Stein, Hoppe, and Gollub 2012) to optical character recognition techniques (e.g., Reynaert 2011). The reason for such popularity may rest in the simplicity of the approach, whereby numbers can be assigned
∗ EACH–USP, Arlindo Béttio, 1000. 03828-000. São Paulo, SP – Brazil. E-mail: norton@usp.br.
Submission received: 30 January 2014; revised submission received: 21 July 2014; accepted for publication: 18 September 2014.
doi:10.1162/COLI a 00216
© 2015 Association for Computational Linguistics
to the probability that a certain type of spelling error might take place, simply because that seems to be the frequency with which people make that kind of mistake.
Surprisingly enough, even though Damerau derived his statistics uniquely from texts in English, his findings are applied, with very little to no adaptation at all, to research in a range of different languages, such as Basque (e.g., Aduriz et al. 1997), Persian (e.g., Miangah 2014), and Arabic (e.g., Alkanhal et al. 2012). The question then becomes how appropriate these figures are when applied to languages other than English. In fact, some researchers have already noticed this potential flaw, and tried to adapt Damerau’s findings to their own language, usually by modifying DamerauLevenstein edit distance (Wagner and Fischer 1974) to match some language-specific data, but still without verifying the appropriateness of Damerau’s statistics in the target language (e.g., Rytting et al. 2011; Richter, Stranák, and Rosen 2012), or by taking into account some feature of that language, such as the presence of diacritics, for example (e.g., Andrade et al. 2012).
In this article, we move a step further by analyzing a set of typed texts from two different corpora in Brazilian Portuguese—a language spoken by over 190 million people1—and deriving statistics tailored to this language. We then compare our statistics with results obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006). As we will show, the behavior demonstrated by native speakers of these languages follow a very similar pattern, straying from Damerau’s original findings while, at the same time, holding them still useful. As a limitation, in this research we only account for non-word errors, that is, errors that, once made, result in some non-existing word, and which may be detected by a regular dictionary look-up (Deorowicz and Ciura 2005).
As an indication of the impact of these results, we added a new module (Gimenes, Roman, and Carvalho 2014)2 to OpenOffice Writer3—a freely available word processor—so as to reorder its suggestions list according to the statistics presented in this article. Within Writer, when typing Pao, instead of Pão (‘bread’ in Portuguese), the correct form comes in fifth place in the suggestion list, with Poca, Pocá, Poça, and Próça coming in the first four positions. With this module, it was observed that, for the first testing corpus, in 27.34% of the cases there was an improvement over Writer’s ranking (i.e., the correct form was ranked higher in the list), whereas in 5.84% the new list was actually worse, and in 66.82% no changes were observed. In the second corpus, an improvement was observed in 19.90% of the cases, in 9.00% figures were worse, with 71.10% remaining unchanged. Some 10–21% increase in accuracy is not something to be neglected, especially at the price of changing weights in an edit distance scheme.  2 related work :One of the first efforts to derive statistics similar to those by Damerau in a language other than English was made by Medeiros (1995), who analyzed a set of corpora in European Portuguese. In his work, Medeiros found that 80.1% of all spelling errors would fit into one of Damerau’s categories. He also noticed that around 20% of all linguistic errors would be related to the use of diacritics, meaning that they should be taken into account during string edit distance or probability calculations. By
1 According to 2010’s demographic census: http://www.ibge.gov.br/home/estatistica/populacao/ censo2010/caracteristicas_da_populacao/resultados_do_universo.pdf. 2 Available at http://ppgsi.each.usp.br/arquivos/RelTec/PPgSI-001_2014.pdf. 3 https://www.openoffice.org/.
linguistic errors, the author meant linguistically motivated errors, as opposed to those caused by slipping fingers in a keyboard, for example. However illustrative, Medeiros’s findings suffer from a major drawback, to the extent that he relied on pre-existing lists of erroneous words in the related literature, along with handwritten errors made by high school students, with only part of the data coming from e-mail exchanges—that is, from actual typed text. The problem with this approach is that these data come without any frequency analysis, which renders them inappropriate for the application of computer techniques based on statistics, also making it hard to compare to any findings related to error frequency.
Spanish was another language found to confirm Damerau’s findings (Bustamante, Arnaiz, and Ginés 2006). In that case, however, the authors made a much more detailed categorization of the errors they found, beyond the four classes originally proposed by Damerau, in a corpus of written (presumably typed) texts from Mexico and Spain. Still, a careful inspection into these categories (by grouping insertions and repetitions of the same letter, and taking errors related to the misuse of diacritics, capitalization, and cognitive replacements to be substitutions) leads one back to Damerau’s four basic errors. In that case, a total of 86.7% of all errors were found to fit into one of these groups. The main difference, however, between both Damerau and Medeiros, and the work by Bustamante, Arnaiz, and Ginés (2006), is that the majority of spelling errors (54.9% of all errors, corresponding to around 49% of all single errors) were related to the misuse of diacritics, with special emphasis on their omission, which was responsible for 51.5% of all spelling errors (i.e., around 46% of all single errors) found in the corpus.
Finally, one of the latest additions to this list was made by Baba and Suzuki (2012), in which the authors analyzed spelling errors both in English and Japanese, with the last one typed on a Roman keyboard (that is, transliterated). Interestingly, the authors report almost no difference in the distribution of errors among Damerau’s four classes, both for English and Japanese. By looking at the details, however, they found some specific errors in Japanese, which they attribute to the phonological and orthographic characteristics of the language. In fact, a breakdown analysis of the four classes show accentuated differences in the type of substitutions (that is, a vowel by another vowel, a vowel by a consonant, etc.), insertions (inserted vowel vs. consonant), deletions (whether or not the deleted character was a repetition of some of its neighbors), and transpositions (adjacent characters versus syllables) that were made.
As we will show in the following sections, this overall confirmation of Damerau’s findings, but with remarked differences caused by each language’s idiosyncrasies, seems to be the common behavior both in the related work and ours. This, in turn, may shed some light on the reasons why English-based spelling checking software does not seem to perform so badly in other languages, but still not as well as it could, should each language’s own characteristics be taken into account. in C2 for single errors—still a substantial proportion. Interestingly, if one takes errors involving diacritics to be substitutions, then we end up with a total of 89.86% of all single errors falling into one of Damerau’s categories in C1, with 88.91% in C2, thereby confirming Damerau’s statistics. An analysis of the distribution of single errors among Damerau’s four original categories (i.e., disregarding diacritic-related misspellings), and the distribution among the categories related to the use of diacritics also shows no statistically significant difference between both corpora (ks = 0.5, p = 0.6994 for Damerau’s categories and ks = 0.6, p = 0.3291 for the diacritics—including cedilla—set).
Finally, regarding the number of misspelled words, we have once again confirmed Damerau’s results, in that over 85% of all wrong words, be they repetition of existing words or not, have a single spelling error. Table 3 shows these results. In this table,6 we present the figures both when taking word repetitions into account and when ruling them out (that is, when keeping the statistics only for new words). This is also in line with the results obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006), in which it was found that 86.7% of all spelling errors were single errors.
6 Total = 966 × 1 + 70 × 2 + 7 × 3 = 1, 127 errors, with another 12 distributed in the last category, for C1, and 911 × 1 + 95 × 2 + 37 × 3 = 1, 212 errors, with another 48 distributed in the last category, for C2. 3 materials and methods :Because we were interested in source texts written in Brazilian Portuguese, where authors were free to type with little to no intervention by spelling correction facilities, we decided to use the corpus presented in Roman et al. (2013), which comprises 1,808 texts, typed in by 452 different participants in a Web experiment, with a total of 62,858 words in length. As a source, this corpus has the advantages of (1) being produced by a number of different people of different ages, educational attainment, and background;
(2) being produced through a Web browser, without any help from spelling correction modules usually present in text editors; and (3) being freely available for download over the Web. We will refer to this corpus as C1.
In order to verify the statistics gathered on C1, in June 2011 we collected a corpus of blog posts from four different Web sites4 that describe travel diaries along with comments by visitors. With a total of 26,418 words, spread over 192 posts, and being written in Brazilian Portuguese, this corpus presents the same advantages as C1, except for the fact that we cannot verify participants’ details, which means we cannot make any statement on the participants’ distribution according to age, background, and so forth. Still, the main characteristics of availability and free text production remain. We will refer to this corpus as C2. By comparing the statistics from both corpora we intend to reduce the effect of any bias related to writing style and characteristics of the participants.
Given that our goal was to identify non-word errors, we relied on a commercial text editor to highlight misspellings in both corpora. The highlighted words were then manually inspected by one of the researchers, and errors were grouped in categories. Besides Damerau’s four original sets, we identified three extra categories: errors involving the use of diacritics, errors related to the use of the cedilla, and space-related errors. The first of these groups was inspired by the observation that the use of diacritics is a key issue in Portuguese (Andrade et al. 2012), potentially playing an important role in the making of spelling mistakes. Following Damerau, this group can be subdivided into four other subcategories: missing diacritic, addition of diacritic, right diacritic applied to the wrong character, and wrong diacritic applied to the right character.
The second group concerns the misuse of the cedilla—a special diacritical mark that in Portuguese can only be applied to the character c, resulting in ç. The reason for this character deserving a whole category of its own lies in the existence of two distinct keyboard layouts in the Brazilian market: ABNT-2 and US-accents. The main difference between both layouts is that whereas ABNT-2 presents a separate key for ç, that key does not exist on the US-Accents keyboard. Hence, while on ABNT-2 the user hits a single key to get a ç, on US-Accents it is a composite—two keys must be pressed: the single quotation mark and c. Ultimately, errors involving the cedilla may be interpreted both as a regular character mistake, or an error related to the use of diacritics, depending on the keyboard.
Finally, because space-related errors, although not so common (see Section 4), are usually difficult to deal with via spelling correctors, we decided to give them an independent set of categories. As a mistake, spaces have the annoying characteristic of not being handled by edit distance operations (Attia et al. 2012), thereby passing undisturbed by systems that use string distance-based ranking when giving alternative spellings to the user. The reason for this problem is that, by joining two words together, for example, the number of mistakes dramatically rises, if one takes that new string to be a single word. Along with these three categories, there is a “leftover” fourth one— others—comprising linguistic expression errors, first-letter capitalization, and wrong diminutive or augmentative.
4 http://bragatte.wordpress.com/. http://guilhermebragatte.blogspot.com.br/. http://forasteironairlanda.wordpress.com/tag/nomadismo/. http://sussuemdublin.wordpress.com/2011/01/. 4 results and discussion :Results both for C1 and C2 are shown in Tables 1 and 2.5 Even though the total number of mistakes is almost the same in both corpora (1,139 for C1 vs. 1,260 for C2), the proportion of spelling errors in C2 is more than twice as high as that of C1. In this case, C1 had a mean error rate of 1.81% (that is, 0.0181 error per word), while C2 reached 4.77%. This difference may be due to the fact that C1 was generated by students from a university, that is, people with a higher educational level. It may also have something to do with the fact that blogs are more conversation-like, which may lead people to relax the spelling rules they choose to follow, especially when it comes to errors involving diacritics. Still, despite this discrepancy, a two-sample KolmogorovSmirnov test showed no statistically significant differences between these corpora, on the distribution of the proportion of errors among categories, neither for the total number of errors (ks = 0.4286, p = 0.1528), nor for the number of errors per word (ks = 0.4286, p = 0.1528 for single, ks = 0.2143, p = 0.9048 for double, ks = 0.2857, p = 0.6172 for triple, and ks = 0.1429, p = 0.9988 for multiple errors).
As it turns out, errors related to the use of diacritics correspond to approximately half of all spelling errors in both corpora (overall, 47.15% in C1 and 50.08% in C2, corresponding to 49.48% of all single errors in C1 and 58.84% in C2), if we include in this sum cedilla-related errors. By taking the cedilla as a simple substitution, the numbers drop to 41.09% in C1 and 45.63% in C2 (overall), with 44.72% in C1 and 57.08%
5 In these tables, for example, 22 insertions found in words with two errors mean that, from all errors found in such words, 22 were insertions, disregarding whether a word presents two insertions or an insertion along with some other type of error. To save space, rows filled with 0, that is categories with no examples, were removed. Space-related errors were also broken down into Substitution by Space, as in pre qualified instead of pre-qualified; Space Insertion, as in w ord instead of word; Space Transposition, as in m yword instead of my word; and Missing Space, as in myword. For reasons noted in Section 2, we have chosen the work by Bustamante, Arnaiz, and Ginés (2006), on European and Mexican Spanish, as a benchmark against which to compare ours. Table 4 shows the comparison results. In this case, for the sake of comparison, and given that each research has a classification scheme of its own, we had both to reorganize some of the categories and present the figures in terms of number of words, instead of number of errors. Although results are shown considering cedilla-related errors to be cases of a missing diacritic, whenever appropriate the figures for taking such errors as simple substitutions are also presented within parentheses.
By comparing the data distributions in Table 4, one sees no differences7 between our results (both for C1 and C2) and those for Spanish, no matter whether one takes cedillarelated errors to be a missing diacritic or a substitution. This, in turn, might indicate that such errors and their distributions are not related to the language or culture themselves, but instead to the set of characters (and corresponding diacritical marks) allowed within each specific language. Also, the data show the importance of taking diacritical marks into consideration when designing spelling correction systems for these languages. As it turned out, this type of error is responsible for over 40% of all wrong words, both in Brazilian Portuguese and Spanish. 5 conclusion :In this article we presented some statistics collected from two different corpora of typed texts in Brazilian Portuguese. Our first contribution is a description of the error rate distribution among the four original categories defined by Damerau (1964), along with the distribution of errors related to the misuse of diacritical marks. Results show not only that Damerau’s figures still hold, should we put aside diacritics, but also make a
7 With cedilla being a diacritic omission: ks = 0.3, p = 0.7591 both between Spanish and C1, and between Spanish and C2. With cedilla being a character substitution: ks = 0.3, p = 0.7591 both between Spanish and C1, and between Spanish and C2.
point about the importance such marks have on the misspelling of words in Brazilian Portuguese, as indicated by the frequency with which this type of error is made, thereby rendering spelling correction systems that rely solely on Damerau’s results unfit for this language. On this account, a straightforward experiment with a commercially available text editor has shown a 10–21% improvement, depending on the testing corpus, in the ranking of suggestions for misspelled words.
As an additional contribution, we have shown that our results are very much like those obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006), in that we could see no statistically significant difference on error distribution between these two languages. This, in turn, may be taken as an indication that the distribution of errors does not depend on language or culture, but instead on the character set that people are allowed to use. As such, it would not come as a surprise to find out that the same distribution can also be observed in other languages, such as Italian, French, and, to some extent, Turkish, for example, which make an intensive use of a similar set of characters. This is something that is worth investigating, and we leave it for future research. Fifty years after Damerau set up his statistics for the distribution of errors in typed texts, his findings are still used in a range of different languages. Because these statistics were derived from texts in English, the question of whether they actually apply to other languages has been raised. We address this issue through the analysis of a set of typed texts in Brazilian Portuguese, deriving statistics tailored to this language. Results show that diacritical marks play a major role, as indicated by the frequency of mistakes involving them, thereby rendering Damerau’s original findings mostly unfit for spelling correction systems, although still holding them useful, should one set aside such marks. Furthermore, a comparison between these results and those published for Spanish show no statistically significant differences between both languages—an indication that the distribution of spelling errors depends on the adopted character set rather than the language itself. [{""affiliations"": [], ""name"": ""Priscila A. Gimenes""}, {""affiliations"": [], ""name"": ""Norton T. Roman""}, {""affiliations"": [], ""name"": ""M. B. R. Carvalho""}] SP:1759bf80767626c66b77ddf36a4d2d9c62f4f022 [{""authors"": [""Aduriz"", ""Itziar"", ""I\u00f1aki Alegria"", ""Xabier Artola"", ""Nerea Ezeiza"", ""Kepa Sarasola"", ""Miriam Urkia.""], ""title"": ""A spelling corrector for Basque based on morphology"", ""venue"": ""Literary and Linguistic Computing,"", ""year"": 1997}, {""authors"": [""Alkanhal"", ""Mohamed I"", ""Mohamed A. Al-Badrashiny"", ""Mansour M. Alghamdi"", ""Abdulaziz O. Al-Qabbany""], ""title"": ""Automatic stochastic Arabic spelling correction with emphasis on space"", ""year"": 2012}, {""authors"": [""Andrade"", ""Guilherme"", ""F. Teixeira"", ""C.R. Xavier"", ""R.S. Oliveira"", ""Leonardo C. da Rocha"", ""A.G. Evsukoff""], ""title"": ""Hasch: High performance automatic spell checker for Portuguese texts"", ""year"": 2012}, {""authors"": [""Attia"", ""Mohammed"", ""Pavel Pecina"", ""Younes Samih"", ""Khaled Shaalan"", ""Josef van Genabith.""], ""title"": ""Improved spelling error detection and correction for Arabic"", ""venue"": ""Proceedings of COLING-2012,"", ""year"": 2012}, {""authors"": [""Baba"", ""Yukino"", ""Hisami Suzuki.""], ""title"": ""How are spelling errors generated and corrected? A study of corrected and uncorrected spelling errors using keystroke logs"", ""venue"": ""Proceedings of ACL-2012,"", ""year"": 2012}, {""authors"": [""Bustamante"", ""Flora Ram\u0131\u0301rez"", ""Alfredo Arnaiz"", ""Mar Gin\u00e9s""], ""title"": ""A spell checker for a world language: The new Microsoft\u2019s Spanish spell checker"", ""venue"": ""In Proceedings of LREC-2006,"", ""year"": 2006}, {""authors"": [""Damerau"", ""Fred J.""], ""title"": ""A technique for computer detection and correction of spelling errors"", ""venue"": ""Communications of the ACM, 7(3):171\u2013176."", ""year"": 1964}, {""authors"": [""Deorowicz"", ""Sebastian"", ""Marcin G. Ciura.""], ""title"": ""Correcting spelling errors by modeling their causes"", ""venue"": ""International Journal of Applied Mathematics and Computer Science, 15(2):275\u2013285."", ""year"": 2005}, {""authors"": [""Gimenes"", ""Priscila Azar"", ""Norton Trevisan Roman"", ""Ariadne Maria Brito Rizzoni Carvalho.""], ""title"": ""An OO writer module for spelling correction in Brazilian Portuguese"", ""venue"": ""Technical Report PPgSI-001/"", ""year"": 2014}, {""authors"": [""Medeiros"", ""Jos\u00e9 Carlos Dinis.""], ""title"": ""Processamento morfol\u00f3gico e correc c\u00e3o ortogr\u00e1fica do portugu\u00eas"", ""venue"": ""Master\u2019s thesis, Instituto Superior T\u00e9cnico \u2013 Universidade T\u00e9cnica de Lisboa, February."", ""year"": 1995}, {""authors"": [""Miangah"", ""Tayebeh Mosavi.""], ""title"": ""Farsispell: A spell-checking system for Persian using a large monolingual corpus"", ""venue"": ""Literary and Linguistic Computing, 29(1):56\u201373."", ""year"": 2014}, {""authors"": [""Pirinen"", ""Tommi A."", ""Krister Lind\u00e9n.""], ""title"": ""State-of-the-art in weighted finite-state spell-checking"", ""venue"": ""Alexander Gelbukh, editor, Computational Linguistics and Intelligent Text Processing,"", ""year"": 2014}, {""authors"": [""Reynaert"", ""Martin W.C.""], ""title"": ""Character confusion versus focus word-based correction of spelling and OCR variants in corpora"", ""venue"": ""International Journal on Document Analysis and Recognition,"", ""year"": 2011}, {""authors"": [""Richter"", ""Michal"", ""Pavel Stran\u00e1k"", ""Alexandr Rosen.""], ""title"": ""Korektor\u2014a system for contextual spell-checking and diacritics completion"", ""venue"": ""Proceedings of COLING-2012, pages 1,019\u20131,028, Mumbai."", ""year"": 2012}, {""authors"": [""Roman"", ""Norton Trevisan"", ""Paul Piwek"", ""Alexandre Rossi Alvares""], ""title"": ""Introducing a corpus of human-authored dialogue summaries in Portuguese"", ""year"": 2013}, {""authors"": [""Christian Hettick"", ""Tim Buckwalter"", ""Charles C. Blake.""], ""title"": ""Spelling correction for dialectal Arabic dictionary lookup"", ""venue"": ""ACM Transactions on Asian Language Information Processing,"", ""year"": 2011}, {""authors"": [""Stein"", ""Benno"", ""Dennis Hoppe"", ""Tim Gollub.""], ""title"": ""The impact of spelling errors on patent search"", ""venue"": ""Proceedings of EACL-2012, pages 570\u2013579, Avignon."", ""year"": 2012}, {""authors"": [""Wagner"", ""Robert A."", ""Michael J. Fischer.""], ""title"": ""The string-to-string correction problem"", ""venue"": ""Journal of the ACM, 21(1): 168\u2013173. 183"", ""year"": 1974}] spelling error patterns in : brazilian portuguese :Priscila A. Gimenes each / usp :Norton T. Roman∗ Ariadne M. B. R. Carvalho Institute of Computing / Unicamp
Fifty years after Damerau set up his statistics for the distribution of errors in typed texts, his findings are still used in a range of different languages. Because these statistics were derived from texts in English, the question of whether they actually apply to other languages has been raised. We address this issue through the analysis of a set of typed texts in Brazilian Portuguese, deriving statistics tailored to this language. Results show that diacritical marks play a major role, as indicated by the frequency of mistakes involving them, thereby rendering Damerau’s original findings mostly unfit for spelling correction systems, although still holding them useful, should one set aside such marks. Furthermore, a comparison between these results and those published for Spanish show no statistically significant differences between both languages—an indication that the distribution of spelling errors depends on the adopted character set rather than the language itself. error category single two three over three total (%) :  errors words in c1 (%) words in c2 (%) words in c1 (%) words in c2 (%) :","5 conclusion :In this article we presented some statistics collected from two different corpora of typed texts in Brazilian Portuguese. Our first contribution is a description of the error rate distribution among the four original categories defined by Damerau (1964), along with the distribution of errors related to the misuse of diacritical marks. Results show not only that Damerau’s figures still hold, should we put aside diacritics, but also make a
7 With cedilla being a diacritic omission: ks = 0.3, p = 0.7591 both between Spanish and C1, and between Spanish and C2. With cedilla being a character substitution: ks = 0.3, p = 0.7591 both between Spanish and C1, and between Spanish and C2.
point about the importance such marks have on the misspelling of words in Brazilian Portuguese, as indicated by the frequency with which this type of error is made, thereby rendering spelling correction systems that rely solely on Damerau’s results unfit for this language. On this account, a straightforward experiment with a commercially available text editor has shown a 10–21% improvement, depending on the testing corpus, in the ranking of suggestions for misspelled words.
As an additional contribution, we have shown that our results are very much like those obtained for Spanish (Bustamante, Arnaiz, and Ginés 2006), in that we could see no statistically significant difference on error distribution between these two languages. This, in turn, may be taken as an indication that the distribution of errors does not depend on language or culture, but instead on the character set that people are allowed to use. As such, it would not come as a surprise to find out that the same distribution can also be observed in other languages, such as Italian, French, and, to some extent, Turkish, for example, which make an intensive use of a similar set of characters. This is something that is worth investigating, and we leave it for future research."
"1 the deep learning tsunami :Deep Learning waves have lapped at the shores of computational linguistics for several years now, but 2015 seems like the year when the full force of the tsunami hit the major Natural Language Processing (NLP) conferences. However, some pundits are predicting that the final damage will be even worse. Accompanying ICML 2015 in Lille, France, there was another, almost as big, event: the 2015 Deep Learning Workshop. The workshop ended with a panel discussion, and at it, Neil Lawrence said, “NLP is kind of like a rabbit in the headlights of the Deep Learning machine, waiting to be flattened.” Now that is a remark that the computational linguistics community has to take seriously! Is it the end of the road for us? Where are these predictions of steamrollering coming from?
At the June 2015 opening of the Facebook AI Research Lab in Paris, its director Yann LeCun said: “The next big step for Deep Learning is natural language understanding, which aims to give machines the power to understand not just individual words but entire sentences and paragraphs.”1 In a November 2014 Reddit AMA (Ask Me Anything), Geoff Hinton said, “I think that the most exciting areas over the next five years will be really understanding text and videos. I will be disappointed if in five years’ time we do not have something that can watch a YouTube video and tell a story about what happened. In a few years time we will put [Deep Learning] on a chip that fits into someone’s ear and have an English-decoding chip that’s just like a real Babel fish.”2 And Yoshua Bengio, the third giant of modern Deep Learning, has also increasingly oriented his group’s research toward language, including recent exciting new developments in neural machine translation systems. It’s not just Deep Learning researchers. When leading machine learning researcher Michael Jordan was asked at a September 2014 AMA, “If you got a billion dollars to spend on a huge research project that you get to lead, what would you like to do?”, he answered: “I’d use the billion dollars to build a NASA-size program focusing on natural language processing, in all of its glory (semantics, pragmatics, etc.).” He went on: “Intellectually I think that NLP is fascinating, allowing us to focus on highly structured inference problems, on issues that go to the core of ‘what is thought’ but remain eminently practical, and on a technology
∗ Departments of Computer Science and Linguistics, Stanford University, Stanford CA 94305-9020, U.S.A. E-mail: manning@cs.stanford.edu. 1 http://www.wired.com/2014/12/fb/. 2 https://www.reddit.com/r/MachineLearning/comments/2lmo0l/ama_geoffrey_hinton.
doi:10.1162/COLI a 00239
© 2015 Association for Computational Linguistics
that surely would make the world a better place.” Well, that sounds very nice! So, should computational linguistics researchers be afraid? I’d argue, no. To return to the Hitchhiker’s Guide to the Galaxy theme that Geoff Hinton introduced, we need to turn the book over and look at the back cover, which says in large, friendly letters: “Don’t panic.”","2 the success of deep learning :There is no doubt that Deep Learning has ushered in amazing technological advances in the last few years. I won’t give an extensive rundown of successes, but here is one example. A recent Google blog post told about Neon, the new transcription system for Google Voice.3 After admitting that in the past Google Voice voicemail transcriptions often weren’t fully intelligible, the post explained the development of Neon, an improved voicemail system that delivers more accurate transcriptions, like this: “Using a (deep breath) long short-term memory deep recurrent neural network (whew!), we cut our transcription errors by 49%.” Do we not all dream of developing a new approach to a problem which halves the error rate of the previously state-of-the-art system?","3 why computational linguists need not worry :Michael Jordan, in his AMA, gave two reasons why he wasn’t convinced that Deep Learning would solve NLP: “Although current deep learning research tends to claim to encompass NLP, I’m (1) much less convinced about the strength of the results, compared to the results in, say, vision; (2) much less convinced in the case of NLP than, say, vision, the way to go is to couple huge amounts of data with black-box learning architectures.”4 Jordan is certainly right about his first point: So far, problems in higher-level language processing have not seen the dramatic error rate reductions from deep learning that have been seen in speech recognition and in object recognition in vision. Although there have been gains from deep learning approaches, they have been more modest than sudden 25% or 50% error reductions. It could easily turn out that this remains the case. The really dramatic gains may only have been possible on true signal processing tasks. On the other hand, I’m much less convinced by his second argument. However, I do have my own two reasons why NLP need not worry about deep learning: (1) It just has to be wonderful for our field for the smartest and most influential people in machine learning to be saying that NLP is the problem area to focus on; and (2) Our field is the domain science of language technology; it’s not about the best method of machine learning—the central issue remains the domain problems. The domain problems will not go away. Joseph Reisinger wrote on his blog: “I get pitched regularly by startups doing ‘generic machine learning’ which is, in all honesty, a pretty ridiculous idea. Machine learning is not undifferentiated heavy lifting, it’s not commoditizable like EC2, and closer to design than coding.”5 From this perspective, it is people in linguistics, people in NLP, who are the designers. Recently at ACL conferences, there has been an over-focus on numbers, on beating the state of the art. Call it playing the Kaggle game. More of the field’s effort should go into problems, approaches, and architectures. Recently, one thing that I’ve been devoting a lot of time to—together with many other
3 http://googleblog.blogspot.com/2015/07/neon-prescription-or-rather-new.html. 4 http://www.reddit.com/r/MachineLearning/comments/2fxi6v/ama_michael_i_jordan. 5 http://thedatamines.com/post/13177389506/why-generic-machine-learning-fails.
collaborators—is the development of Universal Dependencies.6 The goal is to develop a common syntactic dependency representation and POS and feature label sets that can be used with reasonable linguistic fidelity and human usability across all human languages. That’s just one example; there are many other design efforts underway in our field. One other current example is the idea of Abstract Meaning Representation.7","4 deep learning of language :Where has Deep Learning helped NLP? The gains so far have not so much been from true Deep Learning (use of a hierarchy of more abstract representations to promote generalization) as from the use of distributed word representations—through the use of real-valued vector representations of words and concepts. Having a dense, multidimensional representation of similarity between all words is incredibly useful in NLP, but not only in NLP. Indeed, the importance of distributed representations evokes the “Parallel Distributed Processing” mantra of the earlier surge of neural network methods, which had a much more cognitive-science directed focus (Rumelhart and McClelland 1986). It can better explain human-like generalization, but also, from an engineering perspective, the use of small dimensionality and dense vectors for words allows us to model large contexts, leading to greatly improved language models. Especially seen from this new perspective, the exponentially greater sparsity that comes from increasing the order of traditional word n-gram models seems conceptually bankrupt.
I do believe that the idea of deep models will also prove useful. The sharing that occurs within deep representations can theoretically give an exponential representational advantage, and, in practice, offers improved learning systems. The general approach to building Deep Learning systems is compelling and powerful: The researcher defines a model architecture and a top-level loss function and then both the parameters and the representations of the model self-organize so as to minimize this loss, in an end-to-end learning framework. We are starting to see the power of such deep systems in recent work in neural machine translation (Sutskever, Vinyals, and Le 2014; Luong et al. 2015).
Finally, I have been an advocate for focusing more on compositionality in models, for language in particular, and for artificial intelligence in general. Intelligence requires being able to understand bigger things from knowing about smaller parts. In particular for language, understanding novel and complex sentences crucially depends on being able to construct their meaning compositionally from smaller parts—words and multiword expressions—of which they are constituted. Recently, there have been many, many papers showing how systems can be improved by using distributed word representations from “deep learning” approaches, such as word2vec (Mikolov et al. 2013) or GloVe (Pennington, Socher, and Manning 2014). However, this is not actually building Deep Learning models, and I hope in the future that more people focus on the strongly linguistic question of whether we can build meaning composition functions in Deep Learning systems.","5 scientific questions that connect computational linguistics and deep learning :I encourage people to not get into the rut of doing no more than using word vectors to make performance go up a couple of percent. Even more strongly, I would like to
6 http://universaldependencies.github.io/docs/. 7 http://amr.isi.edu.
suggest that we might return instead to some of the interesting linguistic and cognitive issues that motivated noncategorical representations and neural network approaches.
One example of noncategorical phenomena in language is the POS of words in the gerund V-ing form, such as driving. This form is classically described as ambiguous between a verbal form and a nominal gerund. In fact, however, the situation is more complex, as V-ing forms can appear in any of the four core categories of Chomsky (1970):
V + −
N + Adjective: an unassuming man Noun: the opening of the store − Verb: she is eating dinner Preposition: concerning your point
What is even more interesting is that there is evidence that there is not just an ambiguity but mixed noun–verb status. For example, a classic linguistic text for being a noun is appearing with a determiner, while a classic linguistic test for being a verb is taking a direct object. However, it is well known that the gerund nominalization can do both of these things at once:
(1) The not observing this rule is that which the world has blamed in our satorist. (Dryden, Essay Dramatick Poesy, 1684, page 310)
(2) The only mental provision she was making for the evening of life, was the collecting and transcribing all the riddles of every sort that she could meet with. (Jane Austen, Emma, 1816)
(3) The difficulty is in the getting the gold into Erewhon. (Sam Butler, Erewhon Revisited, 1902)
This is oftentimes analyzed by some sort of category-change operation within the levels of a phrase-structure tree, but there is good evidence that this is in fact a case of noncategorical behavior in language.
Indeed, this construction was used early on as an example of a “squish” by Ross (1972). Diachronically, the V-ing form shows a history of increasing verbalization, but in many periods it shows a notably non-discrete status. For example, we find clearly graded judgments in this domain:
(4) Tom’s winning the election was a big upset.
(5) ?This teasing John all the time has got to stop.
(6) ?There is no marking exams on Fridays.
(7) *The cessation hostilities was unexpected.
Various combinations of determiner and verb object do not sound so good, but still much better than trying to put a direct object after a nominalization via a derivational morpheme such as -ation. Houston (1985, page 320) shows that assignment of V-ing forms to a discrete part-of-speech classification is less successful (in a predictive sense) than a continuum in explaining the spoken alternation between -ing vs. -in’, suggesting that “grammatical categories exist along a continuum which does not exhibit sharp boundaries between the categories.”
A different, interesting example was explored by one of my graduate school classmates, Whitney Tabor. Tabor (1994) looked at the use of kind of and sort of, an example that I then used in the introductory chapter of my 1999 textbook (Manning and Schütze 1999). The nouns kind or sort can head an NP or be used as a hedging adverbial modifier:
(8) [That kind [of knife]] isn’t used much.
(9) We are [kind of] hungry.
The interesting thing is that there is a path of reanalysis through ambiguous forms, such as the following pair, which suggests how one form emerged from the other.
(10) [a [kind [of dense rock]]]
(11) [a [[kind of] dense] rock]
Tabor (1994) discusses how Old English has kind but few or no uses of kind of. Beginning in Middle English, ambiguous contexts, which provide a breeding ground for the reanalysis, start to appear (the 1570 example in Example (13)), and then, later, examples that are unambiguously the hedging modifier appear (the 1830 example in Example (14)):
(12) A nette sent in to the see, and of alle kind of fishis gedrynge (Wyclif, 1382)
(13) Their finest and best, is a kind of course red cloth (True Report, 1570)
(14) I was kind of provoked at the way you came up (Mass. Spy, 1830)
This is history not synchrony. Presumably kids today learn the softener use of kind/sort of first. Did the reader notice an example of it in the quote in my first paragraph?
(15) NLP is kind of like a rabbit in the headlights of the deep learning machine (Neil Lawrence, DL workshop panel, 2015)
Whitney Tabor modeled this evolution with a small, but already deep, recurrent neural network—one with two hidden layers. He did that in 1994, taking advantage of the opportunity to work with Dave Rumelhart at Stanford.
Just recently, there has started to be some new work harnessing the power of distributed representations for modeling and explaining linguistic variation and change. Sagi, Kaufmann, and Clark (2011)—actually using the more traditional method of Latent Semantic Analysis to generate distributed word representations—show how distributed representations can capture a semantic change: the broadening and narrowing of reference over time. They look at examples such as how in Old English deer was any animal, whereas in Middle and Modern English it applies to one clear animal family. The words dog and hound have swapped: In Middle English, hound was used for any kind of canine, while now it is used for a particular sub-kind, whereas the reverse is true for dog.
Kulkarni et al. (2015) use neural word embeddings to model the shift in meaning of words such as gay over the last century (exploiting the online Google Books Ngrams corpus). At a recent ACL workshop, Kim et al. (2014) use a similar approach—using word2vec—to look at recent changes in the meaning of words. For example, in Figure 1, they show how around 2000, the meaning of the word cell changed rapidly from being
close in meaning to closet and dungeon to being close in meaning to phone and cordless. The meaning of a word in this context is the average over the meanings of all senses of a word, weighted by their frequency of use.
These more scientific uses of distributed representations and Deep Learning for modeling phenomena characterize the previous boom in neural networks. There has been a bit of a kerfuffle online lately about citing and crediting work in Deep Learning, and from that perspective, it seems to me that the two people who scarcely get mentioned any more are Dave Rumelhart and Jay McClelland. Starting from the Parallel Distributed Processing Research Group in San Diego, their research program was aimed at a clearly more scientific and cognitive study of neural networks.
Now, there are indeed some good questions about the adequacy of neural network approaches for rule-governed linguistic behavior. Old timers in our community should remember that arguing against the adequacy of neural networks for rule-governed linguistic behavior was the foundation for the rise to fame of Steve Pinker—and the foundation of the career of about six of his graduate students. It would take too much space to go through the issues here, but in the end, I think it was a productive debate. It led to a vast amount of work by Paul Smolensky on how basically categorical systems can emerge and be represented in a neural substrate (Smolensky and Legendre 2006). Indeed, Paul Smolensky arguably went too far down the rabbit hole, devoting a large part of his career to developing a new categorical model of phonology, Optimality Theory (Prince and Smolensky 2004). There is a rich body of earlier scientific work that has been neglected. It would be good to return some emphasis within NLP to cognitive and scientific investigation of language rather than almost exclusively using an engineering model of research.
Overall, I think we should feel excited and glad to live in a time when Natural Language Processing is seen as so central to both the further development of machine learning and industry application problems. The future is bright. However, I would encourage everyone to think about problems, architectures, cognitive science, and the details of human language, how it is learned, processed, and how it changes, rather than just chasing state-of-the-art numbers on a benchmark task.",,,,"Deep Learning waves have lapped at the shores of computational linguistics for several years now, but 2015 seems like the year when the full force of the tsunami hit the major Natural Language Processing (NLP) conferences. However, some pundits are predicting that the final damage will be even worse. Accompanying ICML 2015 in Lille, France, there was another, almost as big, event: the 2015 Deep Learning Workshop. The workshop ended with a panel discussion, and at it, Neil Lawrence said, “NLP is kind of like a rabbit in the headlights of the Deep Learning machine, waiting to be flattened.” Now that is a remark that the computational linguistics community has to take seriously! Is it the end of the road for us? Where are these predictions of steamrollering coming from? At the June 2015 opening of the Facebook AI Research Lab in Paris, its director Yann LeCun said: “The next big step for Deep Learning is natural language understanding, which aims to give machines the power to understand not just individual words but entire sentences and paragraphs.”1 In a November 2014 Reddit AMA (Ask Me Anything), Geoff Hinton said, “I think that the most exciting areas over the next five years will be really understanding text and videos. I will be disappointed if in five years’ time we do not have something that can watch a YouTube video and tell a story about what happened. In a few years time we will put [Deep Learning] on a chip that fits into someone’s ear and have an English-decoding chip that’s just like a real Babel fish.”2 And Yoshua Bengio, the third giant of modern Deep Learning, has also increasingly oriented his group’s research toward language, including recent exciting new developments in neural machine translation systems. It’s not just Deep Learning researchers. When leading machine learning researcher Michael Jordan was asked at a September 2014 AMA, “If you got a billion dollars to spend on a huge research project that you get to lead, what would you like to do?”, he answered: “I’d use the billion dollars to build a NASA-size program focusing on natural language processing, in all of its glory (semantics, pragmatics, etc.).” He went on: “Intellectually I think that NLP is fascinating, allowing us to focus on highly structured inference problems, on issues that go to the core of ‘what is thought’ but remain eminently practical, and on a technology","[{""affiliations"": [], ""name"": ""Christopher D. Manning""}]",SP:004e03879ab56d9a4ae9ba5b7644f5c0875d2453,"[{""authors"": [""Chomsky"", ""Noam.""], ""title"": ""Remarks on nominalization"", ""venue"": ""R. Jacobs and P. Rosenbaum, editors, Readings in English Transformational Grammar. Ginn, Waltham, MA, pages 184\u2013221."", ""year"": 1970}, {""authors"": [""Houston"", ""Ann Celeste.""], ""title"": ""Continuity and Change in English Morphology: The Variable (ing)"", ""venue"": ""Ph.D. thesis, University of Pennsylvania."", ""year"": 1985}, {""authors"": [""Kim"", ""Yoon"", ""Yi-I Chiu"", ""Kentaro Hanaki"", ""Darshan Hegde"", ""Slav Petrov.""], ""title"": ""Temporal analysis of language through neural language models"", ""venue"": ""Proceedings of the ACL 2014 Workshop on Language"", ""year"": 2014}, {""authors"": [""Kulkarni"", ""Vivek"", ""Rami Al-Rfou"", ""Bryan Perozzi"", ""Steven Skiena.""], ""title"": ""Statistically significant detection of linguistic change"", ""venue"": ""Proceedings of the 24th International World Wide Web Conference"", ""year"": 2015}, {""authors"": [""Luong"", ""Minh-Thang"", ""Ilya Sutskever"", ""Quoc V. Le"", ""Oriol Vinyals"", ""Wojciech Zaremba.""], ""title"": ""Addressing the rare word problem in neural machine translation"", ""venue"": ""Proceedings of the 53rd Annual Meeting of the Association"", ""year"": 2015}, {""authors"": [""Manning"", ""Christopher D."", ""Hinrich Sch\u00fctze.""], ""title"": ""Foundations of Statistical Natural Language Processing"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 1999}, {""authors"": [""Mikolov"", ""Tomas"", ""Ilya Sutskever"", ""Kai Chen"", ""Greg S. Corrado"", ""Jeffrey Dean""], ""title"": ""Distributed representations of words"", ""year"": 2013}, {""authors"": [""Pennington"", ""Jeffrey"", ""Richard Socher"", ""Christopher D. Manning.""], ""title"": ""GloVe: Global vectors for word representation"", ""venue"": ""Proceedings of the 2014 Conference on Empirical Methods in Natural Language"", ""year"": 2014}, {""authors"": [""Prince"", ""Alan"", ""Paul Smolensky.""], ""title"": ""Optimality Theory: Constraint Interaction in Generative Grammar"", ""venue"": ""Blackwell, Oxford."", ""year"": 2004}, {""authors"": [""Ross"", ""John R.""], ""title"": ""The category squish: Endstation Hauptwort"", ""venue"": ""Papers from the Eighth Regional Meeting, pages 316\u2013328, Chicago."", ""year"": 1972}, {""authors"": [""Sagi"", ""Eyal"", ""Stefan Kaufmann"", ""Brady Clark.""], ""title"": ""Tracing semantic change with latent semantic analysis"", ""venue"": ""Kathryn Allen and Justyna Robinson, editors, Current Methods in Historical Semantics. De Gruyter"", ""year"": 2011}, {""authors"": [""Smolensky"", ""Paul"", ""G\u00e9raldine Legendre.""], ""title"": ""The Harmonic Mind: From Neural Computation to Optimality-Theoretic Grammar, volume 1"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 2006}, {""authors"": [""Sutskever"", ""Ilya"", ""Oriol Vinyals"", ""Quoc V. Le.""], ""title"": ""Sequence to sequence learning with neural networks"", ""venue"": ""Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in"", ""year"": 2014}, {""authors"": [""Tabor"", ""Whitney.""], ""title"": ""Syntactic Innovation: A Connectionist Model"", ""venue"": ""Ph.D. thesis, Stanford. 707"", ""year"": 1994}]",acknowledgments  :This Last Words contribution covers part of my 2015 ACL Presidential Address. Thanks to Paola Merlo for suggesting writing it up for publication.,,,,,,last words :,computational linguistics and :,deep learning :Christopher D. Manning∗ Stanford University,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 the deep learning tsunami :Deep Learning waves have lapped at the shores of computational linguistics for several years now, but 2015 seems like the year when the full force of the tsunami hit the major Natural Language Processing (NLP) conferences. However, some pundits are predicting that the final damage will be even worse. Accompanying ICML 2015 in Lille, France, there was another, almost as big, event: the 2015 Deep Learning Workshop. The workshop ended with a panel discussion, and at it, Neil Lawrence said, “NLP is kind of like a rabbit in the headlights of the Deep Learning machine, waiting to be flattened.” Now that is a remark that the computational linguistics community has to take seriously! Is it the end of the road for us? Where are these predictions of steamrollering coming from?
At the June 2015 opening of the Facebook AI Research Lab in Paris, its director Yann LeCun said: “The next big step for Deep Learning is natural language understanding, which aims to give machines the power to understand not just individual words but entire sentences and paragraphs.”1 In a November 2014 Reddit AMA (Ask Me Anything), Geoff Hinton said, “I think that the most exciting areas over the next five years will be really understanding text and videos. I will be disappointed if in five years’ time we do not have something that can watch a YouTube video and tell a story about what happened. In a few years time we will put [Deep Learning] on a chip that fits into someone’s ear and have an English-decoding chip that’s just like a real Babel fish.”2 And Yoshua Bengio, the third giant of modern Deep Learning, has also increasingly oriented his group’s research toward language, including recent exciting new developments in neural machine translation systems. It’s not just Deep Learning researchers. When leading machine learning researcher Michael Jordan was asked at a September 2014 AMA, “If you got a billion dollars to spend on a huge research project that you get to lead, what would you like to do?”, he answered: “I’d use the billion dollars to build a NASA-size program focusing on natural language processing, in all of its glory (semantics, pragmatics, etc.).” He went on: “Intellectually I think that NLP is fascinating, allowing us to focus on highly structured inference problems, on issues that go to the core of ‘what is thought’ but remain eminently practical, and on a technology
∗ Departments of Computer Science and Linguistics, Stanford University, Stanford CA 94305-9020, U.S.A. E-mail: manning@cs.stanford.edu. 1 http://www.wired.com/2014/12/fb/. 2 https://www.reddit.com/r/MachineLearning/comments/2lmo0l/ama_geoffrey_hinton.
doi:10.1162/COLI a 00239
© 2015 Association for Computational Linguistics
that surely would make the world a better place.” Well, that sounds very nice! So, should computational linguistics researchers be afraid? I’d argue, no. To return to the Hitchhiker’s Guide to the Galaxy theme that Geoff Hinton introduced, we need to turn the book over and look at the back cover, which says in large, friendly letters: “Don’t panic.” 2 the success of deep learning :There is no doubt that Deep Learning has ushered in amazing technological advances in the last few years. I won’t give an extensive rundown of successes, but here is one example. A recent Google blog post told about Neon, the new transcription system for Google Voice.3 After admitting that in the past Google Voice voicemail transcriptions often weren’t fully intelligible, the post explained the development of Neon, an improved voicemail system that delivers more accurate transcriptions, like this: “Using a (deep breath) long short-term memory deep recurrent neural network (whew!), we cut our transcription errors by 49%.” Do we not all dream of developing a new approach to a problem which halves the error rate of the previously state-of-the-art system? 3 why computational linguists need not worry :Michael Jordan, in his AMA, gave two reasons why he wasn’t convinced that Deep Learning would solve NLP: “Although current deep learning research tends to claim to encompass NLP, I’m (1) much less convinced about the strength of the results, compared to the results in, say, vision; (2) much less convinced in the case of NLP than, say, vision, the way to go is to couple huge amounts of data with black-box learning architectures.”4 Jordan is certainly right about his first point: So far, problems in higher-level language processing have not seen the dramatic error rate reductions from deep learning that have been seen in speech recognition and in object recognition in vision. Although there have been gains from deep learning approaches, they have been more modest than sudden 25% or 50% error reductions. It could easily turn out that this remains the case. The really dramatic gains may only have been possible on true signal processing tasks. On the other hand, I’m much less convinced by his second argument. However, I do have my own two reasons why NLP need not worry about deep learning: (1) It just has to be wonderful for our field for the smartest and most influential people in machine learning to be saying that NLP is the problem area to focus on; and (2) Our field is the domain science of language technology; it’s not about the best method of machine learning—the central issue remains the domain problems. The domain problems will not go away. Joseph Reisinger wrote on his blog: “I get pitched regularly by startups doing ‘generic machine learning’ which is, in all honesty, a pretty ridiculous idea. Machine learning is not undifferentiated heavy lifting, it’s not commoditizable like EC2, and closer to design than coding.”5 From this perspective, it is people in linguistics, people in NLP, who are the designers. Recently at ACL conferences, there has been an over-focus on numbers, on beating the state of the art. Call it playing the Kaggle game. More of the field’s effort should go into problems, approaches, and architectures. Recently, one thing that I’ve been devoting a lot of time to—together with many other
3 http://googleblog.blogspot.com/2015/07/neon-prescription-or-rather-new.html. 4 http://www.reddit.com/r/MachineLearning/comments/2fxi6v/ama_michael_i_jordan. 5 http://thedatamines.com/post/13177389506/why-generic-machine-learning-fails.
collaborators—is the development of Universal Dependencies.6 The goal is to develop a common syntactic dependency representation and POS and feature label sets that can be used with reasonable linguistic fidelity and human usability across all human languages. That’s just one example; there are many other design efforts underway in our field. One other current example is the idea of Abstract Meaning Representation.7 4 deep learning of language :Where has Deep Learning helped NLP? The gains so far have not so much been from true Deep Learning (use of a hierarchy of more abstract representations to promote generalization) as from the use of distributed word representations—through the use of real-valued vector representations of words and concepts. Having a dense, multidimensional representation of similarity between all words is incredibly useful in NLP, but not only in NLP. Indeed, the importance of distributed representations evokes the “Parallel Distributed Processing” mantra of the earlier surge of neural network methods, which had a much more cognitive-science directed focus (Rumelhart and McClelland 1986). It can better explain human-like generalization, but also, from an engineering perspective, the use of small dimensionality and dense vectors for words allows us to model large contexts, leading to greatly improved language models. Especially seen from this new perspective, the exponentially greater sparsity that comes from increasing the order of traditional word n-gram models seems conceptually bankrupt.
I do believe that the idea of deep models will also prove useful. The sharing that occurs within deep representations can theoretically give an exponential representational advantage, and, in practice, offers improved learning systems. The general approach to building Deep Learning systems is compelling and powerful: The researcher defines a model architecture and a top-level loss function and then both the parameters and the representations of the model self-organize so as to minimize this loss, in an end-to-end learning framework. We are starting to see the power of such deep systems in recent work in neural machine translation (Sutskever, Vinyals, and Le 2014; Luong et al. 2015).
Finally, I have been an advocate for focusing more on compositionality in models, for language in particular, and for artificial intelligence in general. Intelligence requires being able to understand bigger things from knowing about smaller parts. In particular for language, understanding novel and complex sentences crucially depends on being able to construct their meaning compositionally from smaller parts—words and multiword expressions—of which they are constituted. Recently, there have been many, many papers showing how systems can be improved by using distributed word representations from “deep learning” approaches, such as word2vec (Mikolov et al. 2013) or GloVe (Pennington, Socher, and Manning 2014). However, this is not actually building Deep Learning models, and I hope in the future that more people focus on the strongly linguistic question of whether we can build meaning composition functions in Deep Learning systems. 5 scientific questions that connect computational linguistics and deep learning :I encourage people to not get into the rut of doing no more than using word vectors to make performance go up a couple of percent. Even more strongly, I would like to
6 http://universaldependencies.github.io/docs/. 7 http://amr.isi.edu.
suggest that we might return instead to some of the interesting linguistic and cognitive issues that motivated noncategorical representations and neural network approaches.
One example of noncategorical phenomena in language is the POS of words in the gerund V-ing form, such as driving. This form is classically described as ambiguous between a verbal form and a nominal gerund. In fact, however, the situation is more complex, as V-ing forms can appear in any of the four core categories of Chomsky (1970):
V + −
N + Adjective: an unassuming man Noun: the opening of the store − Verb: she is eating dinner Preposition: concerning your point
What is even more interesting is that there is evidence that there is not just an ambiguity but mixed noun–verb status. For example, a classic linguistic text for being a noun is appearing with a determiner, while a classic linguistic test for being a verb is taking a direct object. However, it is well known that the gerund nominalization can do both of these things at once:
(1) The not observing this rule is that which the world has blamed in our satorist. (Dryden, Essay Dramatick Poesy, 1684, page 310)
(2) The only mental provision she was making for the evening of life, was the collecting and transcribing all the riddles of every sort that she could meet with. (Jane Austen, Emma, 1816)
(3) The difficulty is in the getting the gold into Erewhon. (Sam Butler, Erewhon Revisited, 1902)
This is oftentimes analyzed by some sort of category-change operation within the levels of a phrase-structure tree, but there is good evidence that this is in fact a case of noncategorical behavior in language.
Indeed, this construction was used early on as an example of a “squish” by Ross (1972). Diachronically, the V-ing form shows a history of increasing verbalization, but in many periods it shows a notably non-discrete status. For example, we find clearly graded judgments in this domain:
(4) Tom’s winning the election was a big upset.
(5) ?This teasing John all the time has got to stop.
(6) ?There is no marking exams on Fridays.
(7) *The cessation hostilities was unexpected.
Various combinations of determiner and verb object do not sound so good, but still much better than trying to put a direct object after a nominalization via a derivational morpheme such as -ation. Houston (1985, page 320) shows that assignment of V-ing forms to a discrete part-of-speech classification is less successful (in a predictive sense) than a continuum in explaining the spoken alternation between -ing vs. -in’, suggesting that “grammatical categories exist along a continuum which does not exhibit sharp boundaries between the categories.”
A different, interesting example was explored by one of my graduate school classmates, Whitney Tabor. Tabor (1994) looked at the use of kind of and sort of, an example that I then used in the introductory chapter of my 1999 textbook (Manning and Schütze 1999). The nouns kind or sort can head an NP or be used as a hedging adverbial modifier:
(8) [That kind [of knife]] isn’t used much.
(9) We are [kind of] hungry.
The interesting thing is that there is a path of reanalysis through ambiguous forms, such as the following pair, which suggests how one form emerged from the other.
(10) [a [kind [of dense rock]]]
(11) [a [[kind of] dense] rock]
Tabor (1994) discusses how Old English has kind but few or no uses of kind of. Beginning in Middle English, ambiguous contexts, which provide a breeding ground for the reanalysis, start to appear (the 1570 example in Example (13)), and then, later, examples that are unambiguously the hedging modifier appear (the 1830 example in Example (14)):
(12) A nette sent in to the see, and of alle kind of fishis gedrynge (Wyclif, 1382)
(13) Their finest and best, is a kind of course red cloth (True Report, 1570)
(14) I was kind of provoked at the way you came up (Mass. Spy, 1830)
This is history not synchrony. Presumably kids today learn the softener use of kind/sort of first. Did the reader notice an example of it in the quote in my first paragraph?
(15) NLP is kind of like a rabbit in the headlights of the deep learning machine (Neil Lawrence, DL workshop panel, 2015)
Whitney Tabor modeled this evolution with a small, but already deep, recurrent neural network—one with two hidden layers. He did that in 1994, taking advantage of the opportunity to work with Dave Rumelhart at Stanford.
Just recently, there has started to be some new work harnessing the power of distributed representations for modeling and explaining linguistic variation and change. Sagi, Kaufmann, and Clark (2011)—actually using the more traditional method of Latent Semantic Analysis to generate distributed word representations—show how distributed representations can capture a semantic change: the broadening and narrowing of reference over time. They look at examples such as how in Old English deer was any animal, whereas in Middle and Modern English it applies to one clear animal family. The words dog and hound have swapped: In Middle English, hound was used for any kind of canine, while now it is used for a particular sub-kind, whereas the reverse is true for dog.
Kulkarni et al. (2015) use neural word embeddings to model the shift in meaning of words such as gay over the last century (exploiting the online Google Books Ngrams corpus). At a recent ACL workshop, Kim et al. (2014) use a similar approach—using word2vec—to look at recent changes in the meaning of words. For example, in Figure 1, they show how around 2000, the meaning of the word cell changed rapidly from being
close in meaning to closet and dungeon to being close in meaning to phone and cordless. The meaning of a word in this context is the average over the meanings of all senses of a word, weighted by their frequency of use.
These more scientific uses of distributed representations and Deep Learning for modeling phenomena characterize the previous boom in neural networks. There has been a bit of a kerfuffle online lately about citing and crediting work in Deep Learning, and from that perspective, it seems to me that the two people who scarcely get mentioned any more are Dave Rumelhart and Jay McClelland. Starting from the Parallel Distributed Processing Research Group in San Diego, their research program was aimed at a clearly more scientific and cognitive study of neural networks.
Now, there are indeed some good questions about the adequacy of neural network approaches for rule-governed linguistic behavior. Old timers in our community should remember that arguing against the adequacy of neural networks for rule-governed linguistic behavior was the foundation for the rise to fame of Steve Pinker—and the foundation of the career of about six of his graduate students. It would take too much space to go through the issues here, but in the end, I think it was a productive debate. It led to a vast amount of work by Paul Smolensky on how basically categorical systems can emerge and be represented in a neural substrate (Smolensky and Legendre 2006). Indeed, Paul Smolensky arguably went too far down the rabbit hole, devoting a large part of his career to developing a new categorical model of phonology, Optimality Theory (Prince and Smolensky 2004). There is a rich body of earlier scientific work that has been neglected. It would be good to return some emphasis within NLP to cognitive and scientific investigation of language rather than almost exclusively using an engineering model of research.
Overall, I think we should feel excited and glad to live in a time when Natural Language Processing is seen as so central to both the further development of machine learning and industry application problems. The future is bright. However, I would encourage everyone to think about problems, architectures, cognitive science, and the details of human language, how it is learned, processed, and how it changes, rather than just chasing state-of-the-art numbers on a benchmark task. Deep Learning waves have lapped at the shores of computational linguistics for several years now, but 2015 seems like the year when the full force of the tsunami hit the major Natural Language Processing (NLP) conferences. However, some pundits are predicting that the final damage will be even worse. Accompanying ICML 2015 in Lille, France, there was another, almost as big, event: the 2015 Deep Learning Workshop. The workshop ended with a panel discussion, and at it, Neil Lawrence said, “NLP is kind of like a rabbit in the headlights of the Deep Learning machine, waiting to be flattened.” Now that is a remark that the computational linguistics community has to take seriously! Is it the end of the road for us? Where are these predictions of steamrollering coming from? At the June 2015 opening of the Facebook AI Research Lab in Paris, its director Yann LeCun said: “The next big step for Deep Learning is natural language understanding, which aims to give machines the power to understand not just individual words but entire sentences and paragraphs.”1 In a November 2014 Reddit AMA (Ask Me Anything), Geoff Hinton said, “I think that the most exciting areas over the next five years will be really understanding text and videos. I will be disappointed if in five years’ time we do not have something that can watch a YouTube video and tell a story about what happened. In a few years time we will put [Deep Learning] on a chip that fits into someone’s ear and have an English-decoding chip that’s just like a real Babel fish.”2 And Yoshua Bengio, the third giant of modern Deep Learning, has also increasingly oriented his group’s research toward language, including recent exciting new developments in neural machine translation systems. It’s not just Deep Learning researchers. When leading machine learning researcher Michael Jordan was asked at a September 2014 AMA, “If you got a billion dollars to spend on a huge research project that you get to lead, what would you like to do?”, he answered: “I’d use the billion dollars to build a NASA-size program focusing on natural language processing, in all of its glory (semantics, pragmatics, etc.).” He went on: “Intellectually I think that NLP is fascinating, allowing us to focus on highly structured inference problems, on issues that go to the core of ‘what is thought’ but remain eminently practical, and on a technology [{""affiliations"": [], ""name"": ""Christopher D. Manning""}] SP:004e03879ab56d9a4ae9ba5b7644f5c0875d2453 [{""authors"": [""Chomsky"", ""Noam.""], ""title"": ""Remarks on nominalization"", ""venue"": ""R. Jacobs and P. Rosenbaum, editors, Readings in English Transformational Grammar. Ginn, Waltham, MA, pages 184\u2013221."", ""year"": 1970}, {""authors"": [""Houston"", ""Ann Celeste.""], ""title"": ""Continuity and Change in English Morphology: The Variable (ing)"", ""venue"": ""Ph.D. thesis, University of Pennsylvania."", ""year"": 1985}, {""authors"": [""Kim"", ""Yoon"", ""Yi-I Chiu"", ""Kentaro Hanaki"", ""Darshan Hegde"", ""Slav Petrov.""], ""title"": ""Temporal analysis of language through neural language models"", ""venue"": ""Proceedings of the ACL 2014 Workshop on Language"", ""year"": 2014}, {""authors"": [""Kulkarni"", ""Vivek"", ""Rami Al-Rfou"", ""Bryan Perozzi"", ""Steven Skiena.""], ""title"": ""Statistically significant detection of linguistic change"", ""venue"": ""Proceedings of the 24th International World Wide Web Conference"", ""year"": 2015}, {""authors"": [""Luong"", ""Minh-Thang"", ""Ilya Sutskever"", ""Quoc V. Le"", ""Oriol Vinyals"", ""Wojciech Zaremba.""], ""title"": ""Addressing the rare word problem in neural machine translation"", ""venue"": ""Proceedings of the 53rd Annual Meeting of the Association"", ""year"": 2015}, {""authors"": [""Manning"", ""Christopher D."", ""Hinrich Sch\u00fctze.""], ""title"": ""Foundations of Statistical Natural Language Processing"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 1999}, {""authors"": [""Mikolov"", ""Tomas"", ""Ilya Sutskever"", ""Kai Chen"", ""Greg S. Corrado"", ""Jeffrey Dean""], ""title"": ""Distributed representations of words"", ""year"": 2013}, {""authors"": [""Pennington"", ""Jeffrey"", ""Richard Socher"", ""Christopher D. Manning.""], ""title"": ""GloVe: Global vectors for word representation"", ""venue"": ""Proceedings of the 2014 Conference on Empirical Methods in Natural Language"", ""year"": 2014}, {""authors"": [""Prince"", ""Alan"", ""Paul Smolensky.""], ""title"": ""Optimality Theory: Constraint Interaction in Generative Grammar"", ""venue"": ""Blackwell, Oxford."", ""year"": 2004}, {""authors"": [""Ross"", ""John R.""], ""title"": ""The category squish: Endstation Hauptwort"", ""venue"": ""Papers from the Eighth Regional Meeting, pages 316\u2013328, Chicago."", ""year"": 1972}, {""authors"": [""Sagi"", ""Eyal"", ""Stefan Kaufmann"", ""Brady Clark.""], ""title"": ""Tracing semantic change with latent semantic analysis"", ""venue"": ""Kathryn Allen and Justyna Robinson, editors, Current Methods in Historical Semantics. De Gruyter"", ""year"": 2011}, {""authors"": [""Smolensky"", ""Paul"", ""G\u00e9raldine Legendre.""], ""title"": ""The Harmonic Mind: From Neural Computation to Optimality-Theoretic Grammar, volume 1"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 2006}, {""authors"": [""Sutskever"", ""Ilya"", ""Oriol Vinyals"", ""Quoc V. Le.""], ""title"": ""Sequence to sequence learning with neural networks"", ""venue"": ""Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in"", ""year"": 2014}, {""authors"": [""Tabor"", ""Whitney.""], ""title"": ""Syntactic Innovation: A Connectionist Model"", ""venue"": ""Ph.D. thesis, Stanford. 707"", ""year"": 1994}] acknowledgments  :This Last Words contribution covers part of my 2015 ACL Presidential Address. Thanks to Paola Merlo for suggesting writing it up for publication. last words : computational linguistics and : deep learning :Christopher D. Manning∗ Stanford University","4 deep learning of language :Where has Deep Learning helped NLP? The gains so far have not so much been from true Deep Learning (use of a hierarchy of more abstract representations to promote generalization) as from the use of distributed word representations—through the use of real-valued vector representations of words and concepts. Having a dense, multidimensional representation of similarity between all words is incredibly useful in NLP, but not only in NLP. Indeed, the importance of distributed representations evokes the “Parallel Distributed Processing” mantra of the earlier surge of neural network methods, which had a much more cognitive-science directed focus (Rumelhart and McClelland 1986). It can better explain human-like generalization, but also, from an engineering perspective, the use of small dimensionality and dense vectors for words allows us to model large contexts, leading to greatly improved language models. Especially seen from this new perspective, the exponentially greater sparsity that comes from increasing the order of traditional word n-gram models seems conceptually bankrupt.
I do believe that the idea of deep models will also prove useful. The sharing that occurs within deep representations can theoretically give an exponential representational advantage, and, in practice, offers improved learning systems. The general approach to building Deep Learning systems is compelling and powerful: The researcher defines a model architecture and a top-level loss function and then both the parameters and the representations of the model self-organize so as to minimize this loss, in an end-to-end learning framework. We are starting to see the power of such deep systems in recent work in neural machine translation (Sutskever, Vinyals, and Le 2014; Luong et al. 2015).
Finally, I have been an advocate for focusing more on compositionality in models, for language in particular, and for artificial intelligence in general. Intelligence requires being able to understand bigger things from knowing about smaller parts. In particular for language, understanding novel and complex sentences crucially depends on being able to construct their meaning compositionally from smaller parts—words and multiword expressions—of which they are constituted. Recently, there have been many, many papers showing how systems can be improved by using distributed word representations from “deep learning” approaches, such as word2vec (Mikolov et al. 2013) or GloVe (Pennington, Socher, and Manning 2014). However, this is not actually building Deep Learning models, and I hope in the future that more people focus on the strongly linguistic question of whether we can build meaning composition functions in Deep Learning systems. 5 scientific questions that connect computational linguistics and deep learning :I encourage people to not get into the rut of doing no more than using word vectors to make performance go up a couple of percent. Even more strongly, I would like to
6 http://universaldependencies.github.io/docs/. 7 http://amr.isi.edu.
suggest that we might return instead to some of the interesting linguistic and cognitive issues that motivated noncategorical representations and neural network approaches.
One example of noncategorical phenomena in language is the POS of words in the gerund V-ing form, such as driving. This form is classically described as ambiguous between a verbal form and a nominal gerund. In fact, however, the situation is more complex, as V-ing forms can appear in any of the four core categories of Chomsky (1970):
V + −
N + Adjective: an unassuming man Noun: the opening of the store − Verb: she is eating dinner Preposition: concerning your point
What is even more interesting is that there is evidence that there is not just an ambiguity but mixed noun–verb status. For example, a classic linguistic text for being a noun is appearing with a determiner, while a classic linguistic test for being a verb is taking a direct object. However, it is well known that the gerund nominalization can do both of these things at once:
(1) The not observing this rule is that which the world has blamed in our satorist. (Dryden, Essay Dramatick Poesy, 1684, page 310)
(2) The only mental provision she was making for the evening of life, was the collecting and transcribing all the riddles of every sort that she could meet with. (Jane Austen, Emma, 1816)
(3) The difficulty is in the getting the gold into Erewhon. (Sam Butler, Erewhon Revisited, 1902)
This is oftentimes analyzed by some sort of category-change operation within the levels of a phrase-structure tree, but there is good evidence that this is in fact a case of noncategorical behavior in language.
Indeed, this construction was used early on as an example of a “squish” by Ross (1972). Diachronically, the V-ing form shows a history of increasing verbalization, but in many periods it shows a notably non-discrete status. For example, we find clearly graded judgments in this domain:
(4) Tom’s winning the election was a big upset.
(5) ?This teasing John all the time has got to stop.
(6) ?There is no marking exams on Fridays.
(7) *The cessation hostilities was unexpected.
Various combinations of determiner and verb object do not sound so good, but still much better than trying to put a direct object after a nominalization via a derivational morpheme such as -ation. Houston (1985, page 320) shows that assignment of V-ing forms to a discrete part-of-speech classification is less successful (in a predictive sense) than a continuum in explaining the spoken alternation between -ing vs. -in’, suggesting that “grammatical categories exist along a continuum which does not exhibit sharp boundaries between the categories.”
A different, interesting example was explored by one of my graduate school classmates, Whitney Tabor. Tabor (1994) looked at the use of kind of and sort of, an example that I then used in the introductory chapter of my 1999 textbook (Manning and Schütze 1999). The nouns kind or sort can head an NP or be used as a hedging adverbial modifier:
(8) [That kind [of knife]] isn’t used much.
(9) We are [kind of] hungry.
The interesting thing is that there is a path of reanalysis through ambiguous forms, such as the following pair, which suggests how one form emerged from the other.
(10) [a [kind [of dense rock]]]
(11) [a [[kind of] dense] rock]
Tabor (1994) discusses how Old English has kind but few or no uses of kind of. Beginning in Middle English, ambiguous contexts, which provide a breeding ground for the reanalysis, start to appear (the 1570 example in Example (13)), and then, later, examples that are unambiguously the hedging modifier appear (the 1830 example in Example (14)):
(12) A nette sent in to the see, and of alle kind of fishis gedrynge (Wyclif, 1382)
(13) Their finest and best, is a kind of course red cloth (True Report, 1570)
(14) I was kind of provoked at the way you came up (Mass. Spy, 1830)
This is history not synchrony. Presumably kids today learn the softener use of kind/sort of first. Did the reader notice an example of it in the quote in my first paragraph?
(15) NLP is kind of like a rabbit in the headlights of the deep learning machine (Neil Lawrence, DL workshop panel, 2015)
Whitney Tabor modeled this evolution with a small, but already deep, recurrent neural network—one with two hidden layers. He did that in 1994, taking advantage of the opportunity to work with Dave Rumelhart at Stanford.
Just recently, there has started to be some new work harnessing the power of distributed representations for modeling and explaining linguistic variation and change. Sagi, Kaufmann, and Clark (2011)—actually using the more traditional method of Latent Semantic Analysis to generate distributed word representations—show how distributed representations can capture a semantic change: the broadening and narrowing of reference over time. They look at examples such as how in Old English deer was any animal, whereas in Middle and Modern English it applies to one clear animal family. The words dog and hound have swapped: In Middle English, hound was used for any kind of canine, while now it is used for a particular sub-kind, whereas the reverse is true for dog.
Kulkarni et al. (2015) use neural word embeddings to model the shift in meaning of words such as gay over the last century (exploiting the online Google Books Ngrams corpus). At a recent ACL workshop, Kim et al. (2014) use a similar approach—using word2vec—to look at recent changes in the meaning of words. For example, in Figure 1, they show how around 2000, the meaning of the word cell changed rapidly from being
close in meaning to closet and dungeon to being close in meaning to phone and cordless. The meaning of a word in this context is the average over the meanings of all senses of a word, weighted by their frequency of use.
These more scientific uses of distributed representations and Deep Learning for modeling phenomena characterize the previous boom in neural networks. There has been a bit of a kerfuffle online lately about citing and crediting work in Deep Learning, and from that perspective, it seems to me that the two people who scarcely get mentioned any more are Dave Rumelhart and Jay McClelland. Starting from the Parallel Distributed Processing Research Group in San Diego, their research program was aimed at a clearly more scientific and cognitive study of neural networks.
Now, there are indeed some good questions about the adequacy of neural network approaches for rule-governed linguistic behavior. Old timers in our community should remember that arguing against the adequacy of neural networks for rule-governed linguistic behavior was the foundation for the rise to fame of Steve Pinker—and the foundation of the career of about six of his graduate students. It would take too much space to go through the issues here, but in the end, I think it was a productive debate. It led to a vast amount of work by Paul Smolensky on how basically categorical systems can emerge and be represented in a neural substrate (Smolensky and Legendre 2006). Indeed, Paul Smolensky arguably went too far down the rabbit hole, devoting a large part of his career to developing a new categorical model of phonology, Optimality Theory (Prince and Smolensky 2004). There is a rich body of earlier scientific work that has been neglected. It would be good to return some emphasis within NLP to cognitive and scientific investigation of language rather than almost exclusively using an engineering model of research.
Overall, I think we should feel excited and glad to live in a time when Natural Language Processing is seen as so central to both the further development of machine learning and industry application problems. The future is bright. However, I would encourage everyone to think about problems, architectures, cognitive science, and the details of human language, how it is learned, processed, and how it changes, rather than just chasing state-of-the-art numbers on a benchmark task."
"1 introduction :A sentiment lexicon is regarded as the most valuable resource for sentiment analysis (Pang and Lee 2008), and lays the groundwork of much sentiment analysis research, for example, sentiment classification (Yu and Hatzivassiloglou 2003; Kim and Hovy 2004) and opinion summarization (Hu and Liu 2004). To avoid manually annotating sentiment words, an automatically learning sentiment lexicon has attracted considerable attention in the community of sentiment analysis. The existing work determines word sentiment polarities either by the statistical information (e.g., the co-occurrence of words with predefined sentiment seed words) derived from a large corpus (Riloff, Wiebe, and Wilson 2003; Hu and Liu 2006) or by the word semantic information (e.g., synonym relations) found in existing human-created resources (e.g., WordNet) (Takamura, Inui, and Okumura 2005; Rao and Ravichandran 2009). However, current work mainly focuses on English sentiment lexicon generation or expansion, while sentiment lexicon learning for other languages has not been well studied.
In this article, we address the issue of cross-lingual sentiment lexicon learning, which aims to generate sentiment lexicons for a non-English language (hereafter referred to as “the target language”) with the help of the available English sentiment lexicons. The underlying motivation of this task is to leverage the existing English sentiment lexicons and substantial linguistic resources to label the sentiment polarities of the words in the target language. To this end, we need an approach to transferring the sentiment information from English words to the words in the target language. The few existing approaches first build word relations between English and the target language. Then, based on the word relation and English sentiment seed words, they determine the sentiment polarities of the words in the target language. In these two steps, relation-building plays a fundamental role because it is responsible for the transfer of sentiment information between the two languages. Two approaches are often used to connect the words in different languages in the literature. One is based on translation entries in cross-lingual dictionaries (Hassan et al. 2011). The other relies on a machine translation (MT) engine as a black box to translate the sentiment words in English to the target language (Steinberger et al. 2011). The two approaches in Duh, Fujino, and Nagata (2011) and Mihalcea, Banea, and Weibe (2007) tend to use a small set of vocabularies to translate the natural language, which leads to a low coverage of generated sentiment lexicons for the target language.
To solve this problem, we propose a generic approach to addressing the task of cross-lingual sentiment lexicon learning. Specifically, we model this task with a bilingual word graph, which is composed of two intra-language subgraphs and an interlanguage subgraph. The intra-language subgraphs are used to model the semantic relations among the words in the same languages. When building them, we incorporate both synonym and antonym word relations in a novel manner, represented by positive and negative sign weights in the subgraphs, respectively. These two intra-language subgraphs are then connected by the inter-language subgraph. We propose Bilingual word graph Label Propagation (BLP), which simultaneously takes the inter-language relations and the intra-language relations into account in an iterative way. Moreover, we leverage the word alignment information derived from a parallel corpus to build the inter-language relations. We connect two words from different languages that are aligned to each other in a parallel sentence pair. Taking advantage of a large parallel corpus, this approach significantly improves the coverage of the generated sentiment lexicon. The experimental results on Chinese sentiment lexicon learning show the effectiveness of the proposed approach in terms of both precision and recall. We further
evaluate the impact of the learned sentiment lexicon on sentence-level sentiment classification. When using words in the learned sentiment lexicon as features for sentiment classification of the target language, the sentiment classification can achieve a high performance.
We make the following contributions in this article.
1. We present a generic approach to automatically learning sentiment lexicons for the target language with the available sentiment lexicon in English, and we formalize the cross-lingual sentiment learning task on a bilingual word graph.
2. We build a bilingual word graph by using synonym and antonym word relations and propose a bilingual word graph label propagation approach, which effectively leverages the inter-language relations and both types (synonym and antonym) of the intra-language relations in sentiment lexicon learning.
3. We leverage the word alignment information derived from a large number of parallel sentences in sentiment lexicon learning. We build the inter-language relation in the bilingual word graph upon word alignment, and achieve significant results.","2 related work : In general, the work on sentiment lexicon learning focuses mainly on English and can be categorized as co-occurrence–based approaches (Hatzivassiloglou and McKeown 1997; Riloff, Wiebe, and Wilson 2003; Qiu et al. 2011) and semantic-based approaches (Mihalcea, Banea, and Wiebe 2007; Takamura, Inui, and Okumura 2005; Kim and Hovy 2004).
The co-occurrence-based approaches determine the sentiment polarity of a given word according to the statistical information, like the co-occurrence of the word to predefined sentiment seed words or the co-occurrence to product features. The statistical information is mainly derived from certain corpora. One of the earliest work conducted by Hatzivassiloglou and McKeown (1997) assumes that the conjunction words can convey the polarity relation of the two words they connect. For example, the conjunction word and tends to link two words with the same polarity, whereas the conjunction word but is likely to link two words with opposite polarities. Their approach only considers adjectives, not nouns or verbs, and it is unable to extract adjectives that are not conjoined by conjunctions. Riloff et al. (2003) define several pattern templates and extract sentiment words by two bootstrapping approaches. Turney and Littman (2003) calculate the pointwise mutual information (PMI) of a given word with positive and negative sets of sentiment words. The sentiment polarity of the word is determined by average PMI values of the positive and negative sets. To obtain PMI, they provide queries (consisting of the given word and the sentiment word) to the search engine. The number of hits and the position (if the given word is near the sentiment word) are used to estimate the association of the given word to the sentiment word. Hu and Liu (2004) research sentiment word learning on customer reviews and they assume that the sentiment words tend to be correlated with product features. The frequent nouns and noun phrases are treated as product features. Then they extract the adjective words
as sentiment words from those sentences that contain one or more product features. This approach may work on a product review corpus, where one product feature may frequently appear. But for other corpora, like news articles, this approach may not be effective. Qiu et al. (2011) combine sentiment lexicon learning and opinion target extraction. A double propagation approach is proposed to learn sentiment words and to extract opinion targets simultaneously, based on eight manually defined rules.
The semantic-based approaches determine the sentiment polarity of a given word according to the word semantic relation, like the synonyms of sentiment seed words. The word semantic relation is usually obtained from dictionaries, for example, WordNet.1 Kim and Hovy (2004) assume that the synonyms of a positive (negative) word are positive (negative) and its antonyms are negative (positive). Initializing with a set of sentiment words, they expand sentiment lexicons based on these two kinds of word relations. Kamps et al. (2004) build a synonym graph according to the synonym relation (synset) derived from WordNet. The sentiment polarity of a word is calculated by the shortest path to two sentiment words good and bad. However, the shortest path cannot precisely describe the sentiment orientation, considering there are only five steps between the word good and the word bad in WordNet (Hassan et al. 2011). Takamura et al. (2005) construct a word graph with the gloss of WordNet. Words are connected if a word appears in the gloss of another. The word sentiment polarity is determined by the weight of its connections on the word graph. Based on WordNet, Rao and Ravichandran (2009) exploit several graph-based semi-supervised learning methods like Mincuts and Label Propagation. The word polarity orientations are induced by initializing some sentiment seed words in the WordNet graph. Esuli et al. (2006, 2007) and Baccianella et al. (2010) treat sentiment word learning as a machine learning problem, that is, to classify the polarity orientations of the words in WordNet. They select seven positive words and seven negative words and expand them through the see-also and antonym relations in WordNet. These expanded words are then used for training. They train a ternary classifier to predict the sentiment polarities of all the words in WordNet and use the glosses (textual definitions of the words in WordNet) as the features of classification. The sentiment lexicon generated is the well-known SentiWordNet.2 The work on cross-lingual sentiment lexicon learning is still at an early stage and can be categorized into two types, according to how they bridge the words in two languages.
Mihalcea et al. (2007) generate sentiment lexicon for Romanian by directly translating the English sentiment words into Romanian through bilingual English–Romanian dictionaries. When confronting multiword translations, they translate the multiwords word by word. Then the validated translations must occur at least three times on the Web. The approach proposed by Hassan et al. (2011) learns sentiment words based on English WordNet and WordNets in the target languages (e.g., Hindi and Arabic). Crosslingual dictionaries are used to connect the words in two languages and the polarity of a given word is determined by the average hitting time from the word to the English sentiment word set. These approaches connect words in two languages based on crosslingual dictionaries. The main concern of these approaches is the effect of morphological inflection (i.e., a word may be mapped to multiple words in cross-lingual dictionaries).
1 http://wordnet.princeton.edu/. 2 http://sentiwordnet.isti.cnr.it/.
For example, one single English word typically has four Spanish or Italian word forms (two each for gender and for number) and many Russian word forms (due to gender, number, and case distinctions) (Steinberger et al. 2011). Usually, this approach requires an additional process to disambiguate the sentiment polarities of all the morphological variants.
To improve the sentiment classification for the target language, Banea, Mihalcea, and Wiebe (2010) translate the English sentiment lexicon into the target language using Google Translator.3 Similarly, Google Translator is used by Steinberger et al. (2011). They manually produce two high-level gold-standard sentiment lexicons for two languages (e.g., English and Spanish) and then translate them into the third language (e.g., Italian) via Google Translator. They believe that those words in the third language that appear in both translation lists are likely to be sentiment words. These approaches connect the words in two languages based on MT engines. The main concern of these approaches is the low overlapping between the vocabularies of natural documents and the vocabularies of the documents translated by MT engines (Duh, Fujino, and Nagata 2011; Meng et al. 2012a). The shortcoming of these MT-based approaches inevitably leads to low coverage.
Our task resembles the task of cross-lingual sentiment classification, like Wan (2009), Lu et al. (2011), and Meng et al. (2012a), which classifies the sentiment polarities of product reviews. Generally, these studies use semi-supervised learning approaches and regard translations from labeled English sentiment reviews as the training data. The terms in each review are leveraged as the features for training, which has proven to be effective in sentiment classification (Pang and Lee 2008). We can regard the task of sentiment lexicon learning as word-level sentiment classification. However, for wordlevel sentiment classification, it is not straightforward to extract features for a single word. Without sufficient features, it is difficult for these approaches to perform well in learning. Another line of cross-lingual sentiment classification uses Latent Dirichlet Allocation (LDA) (Blei, Ng, and Jordan 2003) or its variants, like Boyd-Graber and Resnik (2010) or He, Alani, and Zhou (2010). These studies assume that each review is a mixture of sentiments and each sentiment is a probability over words. Then they apply the LDA-like approach to model the sentiment polarity of each review. Nonetheless, this assumption may not be applicable in sentiment lexicon learning because a single word can be regarded as the minimal semantic unit, and it is difficult, if not impossible, to infer the latent topics from a single word. Recall that different from the sentiment classification of product reviews where the instances are normally independent, words in sentiment lexicon learning are highly related with each other, like synonyms and antonyms. Through these relations, the words can naturally form a word graph. Thus we use the graph-based learning approach to leverage the word distributions in sentiment lexicon learning. In the next section, we will introduce our proposed graph-based cross-lingual sentiment lexicon learning.","3 cross-lingual sentiment lexicon learning :In this work, we model the task of cross-lingual sentiment lexicon learning with a bilingual word graph, where (1) the words in the two languages are represented by the nodes in two intra-language subgraphs, respectively; (2) the synonym and antonym word relations within each language are represented by the positive and negative sign
3 http://translate.google.com/.
weights in the corresponding intra-language subgraphs; and (3) the two intra-language subgraphs are connected by an inter-language subgraph. Mathematically, we build a
graph G = (XE ∪ XT, WE ∪ W̃E ∪ WT ∪ W̃T ∪ WA) that consists of two intra-language subgraphs GE = (XE, WE ∪ W̃E) and GT = (XT, WT ∪ W̃T ) as shown in Figure 1. These two subgraphs are connected by the inter-language graph GR = (XE ∪ XT, WA). The elements of WE, WT, and WA are positive real numbers, that is, WE, WT, and WA ∈ R+, and W̃E and W̃T ∈ R−. Because G incorporates the words in two languages, we call it a Bilingual Word Graph. Specifically, the positive weights, WE and WT, represent the synonym intra-language relations, and the negative weights, W̃E and W̃T, represent the antonym intra-language relations. The inter-language relations, WA, represent the connections between the words in the two languages. For cross-lingual sentiment lexicon learning, XE = {X L E, X U E } denotes the labeled and unlabeled words in English and XT denotes the unlabeled words in the target language. Given the labels YLE = {yE1 , ..., yEl} of the seeds X L E, we aim to predict the sentiment polarities of the words XT. In the remainder of this section, we will present the bilingual word graph construction and the algorithm of bilingual word graph label propagation. We represent the words in English and in the target language as the nodes of the bilingual word graph. We use the synonym and antonym relations of the words in the
same language to build W and W̃ in the intra-language graph, respectively. In the rest of this section, we will focus on how to build the inter-language relation.
Intuitively, there are two ways to connect the words in two languages. One is to insert links to the words if there exist entry mappings between the words in bilingual dictionaries (e.g., the English–Chinese dictionary). This method is simple and straightforward, but it suffers from two limitations. (1) Dictionaries are static during a certain period, whereas the sentiment lexicon evolves over time. (2) The entries in dictionaries tend to be the expressions of formal and written languages, but people prefer using colloquial language in expressing their sentiments or opinions on-line. These limitations lead to the low coverage of the links from English to the target language. An alternative way is to use an MT engine as a black box to build the inter-language relation. One can send each word in English to a publicly available MT engine and get the translations in the target language. Edges can then be inserted into the graph between the English
words and their corresponding translations. This approach suffers from the problem of low coverage as well because MT engines tend to use a small set of vocabularies (Duh, Fujino, and Nagata 2011).
In this article, we propose to leverage a large bilingual parallel corpus, which is readily available in the MT research community, to build the bilingual word graph. The parallel corpus consists of a large number of parallel sentence pairs from two different languages that have been used as the foundation of the state-of-the-art statistical MT engines. Like the example shown in Figure 2, the two sentences in English and Chinese are parallel sentences, which express the same meaning in different languages. We can easily derive the word alignment from the sentence pairs, automatically using a stateof-the-art toolkit, like GIZA++4 or BerkeleyAligner.5 In this example, the Chinese word
(happy) is linked to the English word happy and we say that these two words are
aligned. Similarly the English words best and wishes are both aligned to (wish). The word alignment information encodes the rich association information between the words from the two languages. We are therefore motivated to leverage the parallel corpus and word alignment to build the bilingual word graph for cross-lingual sentiment lexicon learning. We take the words from both languages in the bilingual parallel corpus as the nodes in the bilingual word graph, and build the inter-language relations by connecting the two words that are aligned together in a sentence pair from a parallel corpus. There are several advantages of using a parallel corpus to build the inter-language subgraph. First, large parallel corpora are extensively used for training statistical MT engines and can be easily reused in our task. The parallel sentence pairs are usually automatically collected and mined from the Web. As a result, they contain the different and practical variations of words and phrases embedded in sentiment expressions. Second, the parallel corpus can be dynamically changed when necessary because it is relatively easy to collect from the Web. Consequentially, the novel sentiment information inferred from the parallel corpus can easily update the existing sentiment lexicons. These advantages can greatly improve the coverage of the generated sentiment lexicon, as demonstrated later in our experiments. As commonly used semi-supervised approaches, label propagation (Zhu and Ghahramani 2002) and its variants (Zhu, Ghahramani, and Lafferty 2003; Zhou et al. 2004) have been applied to many applications, such as part-of-speech tagging (Das and Petrov 2011; Li, Graca, and Taskar 2012), image annotation (Wang, Huang, and Ding 2011), protein function prediction (Jiang 2011; Jiang and McQuay 2012), and so forth.
4 http://www.statmt.org/moses/giza/GIZA++.html. 5 http://nlp.cs.berkeley.edu.
The underlying idea of label propagation is that the connected nodes in the graph tend to share the same sentiment labels. In bilingual word graph label propagation, the words tend to share same sentiment labels if they are connected by synonym relations or word alignment and tend to belong to different sentiment labels if connected by antonym relations.
In this article we propose bilingual word graph label propagation for cross-lingual sentiment lexicon learning. Let F = {FE, FT} denote the predicted labels of the unlabeled words X. The loss function can be defined as
El(F) = µ
n∑
i=1
‖fEi − yEi‖ 2 + µ
m∑
i=1
‖ fTi − yTi‖ 2 (1)
where n and m denote the numbers of English words and words in the target language. Let Y = {YE, YT} denote the initial sentiment labels of all the words; the loss function means that the prediction could not change too much from the initial label assignment. Similar to Zhou et al. (2004), we define the smoothness function to indicate that if two words are connected by synonym relation or by word alignment, then they tend to share the same sentiment label. The smoothness function can be further represented by two parts, that is, the inter-language smoothness Einters (F) and the synonym intra-language smoothness Eintras (F)
Einters (F) = 1 2
n∑
i=1
m∑
j=1
wAij‖ fEi√ dALii − fTj√ dARjj ‖2 (2)
Eintras (F) = ρ1 1 2
n∑
i,j=1
wEij‖ fEi√ dEii − fEj√ dEjj ‖2 + ρ2 1 2
m∑
i,j=1
wTij‖ fTi√ dTii − fTj√ dTjj ‖2 (3)
DAL and DAR are defined as DAL = diag( ∑ j wA(1, j), . . . , ∑ j wA(n, j)) ′ and DAR = diag( ∑ i wA(i, 1), . . . , ∑ i wA(i, m)) ′
. DE and DT are the degree matrices of the synonym intra-language relations WE and WT, respectively. We then define the distance function to indicate that if two words are connected by the antonym relation they tend to belong to different sentiment labels. The distance function can be defined as
Eintrad (F) = ρ3 1 2
n∑
i,j=1 |w̃Eij |‖ fEi√ d̃Eii − fEj√ d̃Ejj ‖2 + ρ4 1 2
m∑
i,j=1 |w̃Tij |‖ fTi√ d̃Tii − fTj√ d̃Tjj ‖2 (4)
where D̃E and D̃T are the degree matrices of the absolute value of the antonym intralanguage relations W̃E and W̃T, respectively. Intuitively, for the inter-language smoothness and the synonym intra-language smoothness, the nearer the words connect with each other, the better performance will be achieved, whereas for the antonym intralanguage distance, the farther the better. The objective functions can be defined as
arg min(E(F)) = arg min(Eintras (F) + E inter s (F) + El(F))
arg max(E(F)) = arg max(Eintrad (F))
Thus, we define the whole objective function for cross-lingual sentiment lexicon learning as
arg min(E(F)) = arg min(Eintras (F) + E inter s (F) + El(F) − E intra d (F)) (5)
To obtain the solution to Equation (5), we differentiate the objective function according to FE and FT, and we have
∂E(F) ∂FE |FE=F⋆E = ρ1SEFE + 1 2 SAFT − ρ3S̃EFE + µFE − µYE = 0 (6)
∂E(F) ∂FT |FT=F⋆T = ρ2STFT + 1 2 SA ′ FE − ρ4S̃TFT + µFT − µYT = 0 (7)
where P ′
is the transpose of the matrix P. The graph Laplacians SE and ST of the
synonym intra-language relations are (I − D − 12 E WED − 12 E ) and (I − D − 12 T WTD − 12 T ), where I is the identity matrix. The graph Laplacians S̃E and S̃T of the antonym intra-language relations are (I − D̃ − 12 E W̃ED̃ − 12 E ) and (I − D̃ − 12 T W̃TD̃ − 12 T ), which has been proven to be positive semi-definite (Kunegis et al. 2010). The graph Laplacian SA of the interlanguage relation is (I − D − 12 AL WAD − 12 AR ). From Equations (6) and (7), we can obtain the optimal solutions
(ME − SAM −1 T S ′ A)FE = 2µ(YE − SAM −1 T YT ) (8)
(MT − S ′ AM −1 E SA)FT = 2µ(YT − S ′ AM −1 E YE) (9)
where ME = 2ρ1SE − 2ρ3S̃E + 2µI and MT = 2ρ2ST − 2ρ4S̃T + 2µI. To avoid computing the inverse matrix in Equations (8) and (9), we apply the Jacobi algorithm (Saad 2003) to calculate the solutions as described in Algorithm 1. In line 1, we set the label of the positive seed xi as y L E+ i = (1, 0) and the label of the negative seed xj as y L E− j = (0, 1). We set the label YUE of the unlabeled words as zero, and then generate YE with Y L E and Y U E . Line 2 sets YT as zero matrix. In line 3, we compute the matrixes SE, S̃E, ST, S̃T, SA, and then compute the matrixes ME and MT. The sentiment information is simultaneously
Algorithm 1. Bilingual word graph label propagation
Input: Given G = (XE ∪ XT, WE ∪ W̃E ∪ WT ∪ W̃T ∪ WA),XE, label YLE for X L E, initialize µ and ρ1∼4 Output: FT for XT and FE for XE
1. Initialize YE with the English sentiment seeds 2. Set YT as zero 3. Calculate SE, S̃E, ST, S̃T, and SA, then calculate ME and MT 4. Loop 5. f (t+1) Ei = 1 (ME−SAM −1 T S ′ A)ii (2µ(YE − SAM −1 T YT )i − ∑ j 6=i(ME − SAM −1 T S ′ A)ij f (t) Ei ) 6. f (t+1) Ti = 1 (MT−S ′ AM −1 E SA)ii (2µ(YT − S ′ AM −1 E YE)i − ∑ j 6=i(MT − S ′ AM −1 E SA)ij f (t) Ti ) 7. Until FE and FT convergence
propagated through lines 4–7 until the predicted labels FE and FT are converged. For an unlabeled word xi, if | f (i, 0) − f (i, 1)| < ξ (ξ is set as 1.0E
−4), xi is regarded as neutral; if ( f (i, 0) − f (i, 1)) ≥ ξ, xi is regarded as positive; and if ( f (i, 1) − f (i, 0)) ≥ ξ, xi is regarded as negative.","4 experiment : We conduct experiments on Chinese sentiment lexicon learning. As in previous work (Baccianella, Esuli, and Sebastiani 2010), the sentiment words in General Inquirer lexicon are selected as the English seeds (Stone 1997). From the GI lexicon we collect 2,005 positive words and 1,635 negative words. To build the bilingual word graph, we adopt the Chinese–English parallel corpus, which is obtained from the news articles published by Xinhua News Agency in Chinese and English collections, using the automatic parallel sentence identification approach (Munteanu and Marcu 2005). Altogether, we collect more than 25M parallel sentence pairs in English and
Chinese. We remove all the stopwords in Chinese and English (e.g., (of) and am) together with the low-frequency words that occur fewer than 5 times. After preprocessing, we finally have more than 174,000 English words, among which 3,519 words have sentiment labels and more than 146,000 Chinese words for which we need to predict the sentiment labels. To transfer sentiment information to Chinese unlabeled words more efficiently, we remove the unlabeled English words in the word graph (i.e., XUE = Φ). The unsupervised method, namely, BerkeleyAligner, is used to align the parallel sentences in this article (Liang, Taskar, and Klein 2006). As an unsupervised method, it does not require us to manually collect training data and does not need the complex training processing, and its performance is competitive with supervised methods. With these two advantages, we can focus more on our task of cross-lingual sentiment lexicon learning. Based on the word alignment derived by BerkeleyAligner, the inter-language WA is initialized with the normalized alignment frequency. The English and Chinese versions6 of WordNet are used to build the intra-language relations WE, W̃E, WT, and W̃T, respectively. WordNet (Miller 1995) groups words into synonym sets, called synsets. We collect about 117,000 synsets from the English WordNet and about 80,000 synsets from the Chinese WordNet. In total, we obtain 8,406 and 6,312 antonym synset pairs.
We first generate both positive and negative scores for each unlabeled word and then determine the word sentiment polarities based on its scores. We rank the two sets of newly labeled sentiment words according to their polarity scores. The top-ranked Chinese words are shown in Table 1. We manually label the top-ranked 1K sentiment words. For P@10K, we sequentially divide the top 10K ranked list into ten equal parts. One hundred sentiment words are randomly selected from each part for labeling. Similar to the evaluation of TREC Blog Distillation (Ounis, Macdonald, and Soboroff 2008), all the labeled words from each approach are used in the evaluation. We then evaluate the ranked lists with two metrics, Precision@K and Recall.
6 http://www.globalwordnet.org/gwa/wordnet table.html. In this set of experiments, we examine the influence of graph topologies on sentiment lexicon learning.
Mono: This approach learns the Chinese sentiment lexicon based only on the
Chinese monolingual word graph GT = (XT, WT ∪ W̃T). Because it needs labeled sentiment words, we incorporate the English labeled sentiment words XE and the interlanguage relation WA in the first iteration. Then we set XE and WA to be zero in later iterations.
BLP-WOA (bilingual word graph without antonym): This approach is based on the bilingual word graph. It only involves the inter-language relation WA and the synonym intra-language relations WE and WT. W̃E and W̃T are set to be zero. BLP: This approach is based on the bilingual word graph. It incorporates the interlanguage relation WA, the synonym intra-language relations WE and WT, and the antonym intra-language relations W̃E and W̃T.
In these approaches, µ is set to 0.1 as in Zhou et al. (2004). The precision of these approaches are shown in Figure 3. The figure shows that the approaches based on the bilingual word graph significantly outperform the one based on the monolingual word graph. The bilingual word graph can bring in more word relations and accelerate the sentiment propagation. Besides, in the bilingual word graph, the English sentiment seed words can continually provide accurate sentiment information. Thus we observe the increase in the approaches based on the bilingual word graph in term of both precision and recall (Table 2). Meanwhile, we find that adding the antonym relation in the bilingual word graph slightly enhances precision in top-ranked words and similar findings are observed in our later experiments. It appears that the antonym relations depict word relations in a more accurate way and can refine the word sentiment scores more precisely. However, the synonym relation and word alignment relation dominate, whereas the antonym relation accounts for only a small percentage of the graph. It is hard for the antonym relation to introduce new relations into the graph and thus it cannot help to further improve recall. In this set of experiments, we compare our approach with the baseline and existing approaches.
Rule: For the intra-language relation, this approach assumes that the synonyms of a positive (negative) word are always positive (negative), and the antonyms of a positive (negative) word are always negative (positive). For the inter-language relation, we regard the Chinese word aligned to positive (negative) English words as positive (negative). If a word connects to both positive and negative English words, we regard it as objective. Based on this heuristic, we generate two sets of sentiment words.
SOP: Hassan et al. (2011) present a method to predict the semantic orientation of unlabeled words based on the mean hitting time to the two sets of sentiment seed words.
Given the graph G = (XE ∪ XT, WE ∪ W̃E ∪ WT ∪ W̃T ∪ WA), it defines the transition probability from node i to node j as
p(j|i) = wi,j∑ k wi,k
The mean hitting time h(i|j) is the average number of the weighted steps from word i to word j. Starting with the word i and ending with the sentiment word k ∈ M, the mean hitting time h(i|M) can be formally defined as
h(i|M) =
{ 0, i ∈ M∑
j∈V p(j|i) × h(j|M) + 1, otherwise
Let M+ and M− denote the GI positive and negative seeds. If h(i|M+) is greater than h(i|M−), the word xi is classified as negative; otherwise it is classified as positive. The generated positive words and negative words are then ranked according to their polarity scores, respectively.
MAD: Talukdar and Crammer (2009) propose a MAD algorithm to modify the adsorption algorithm (Baluja et al. 2008) by adding a new regularization term. In particular, besides the positive and negative labels, a dummy label is assigned to each word in the MAD approach. Two additional columns, representing the scores of the dummy label, are added into Y and F, respectively. We denote these two matrices with the dummy labels as Ŷ and F̂. Meanwhile, R̂ is used to represent the initial dummy scores of all the words. For a word xi, the newly added columns in Ŷi and F̂i are set to zero (i.e., ŷ(i, 3) = f̂ (i, 3) = 0). r̂(i, 0) and r̂(i, 1) are set to zero, and r̂(i, 3) is assigned to one. Then, the predicted label F̂i of the word xi is iteratively obtained by
F̂ (t+1) i = 1 M̂ii
(λ1Ŷi + λ2 ∑
j
WijF̂ (t) j + λ3γR̂i)
M̂ii = (λ1 + λ2 ∑
j 6=i
Wji + λ3)
λ1∼3 and γ are used to tune the importance of each iteration term. We set λ1∼2 to one, λ3 to 10, and γ to 0.1, which produces reasonably good results. After propagation, f̂ (i, 0) and f̂ (i, 1) are used to determine the sentiment polarity of the word xi. We show recall of the learned Chinese sentiment words in Table 3. Compared with BLP and SOP, the Rule approach learns fewer sentiment words. The coverage of the Rule approach is inevitably low because many words in the corpus are aligned to both positive and negative words. For example, in most cases the positive Chinese word
(helpful) is aligned to the positive English word helpful. But sometimes it is aligned (or misaligned) to the negative English words, like freak. Under this situation, the word tends to be predicted as objective. In SOP, the positive and negative scores are related to the distances of the word to the positive and negative seed words, and the distance is usually coarse-grained to depict the sentiment polarity. For example, the shortest path between the word good and the word bad in WordNet is only 5 (Kamps et al. 2004). The Rule and SOP approaches find different sentiment words. We then evaluate the learned Chinese polarity word lists by precision at k. As illustrated in Figure 4, the significance test indicates that our approach significantly outperforms the Rule and SOP approaches. The major difference of our approach is that the polarity information can be transferred between English and Chinese and within each language at the same time, whereas in the other two approaches the polarity information mainly transfers from English to Chinese and once a word gets a polarity score, it is difficult to change or refine. The idea of the MAD approach is similar to bilingual graph label propagation, but the MAD
approach fails to leverage the antonym intra-language relation. We observe that the MAD approach can achieve a comparable result to the BLP approach. MAD can obtain a smoother label score by adding a dummy label. But the dummy label does not influence the sentiment labels too much because it is not used in the determination of the word sentiment polarity. Besides, MAD cannot deal with the antonym relation. As a result, these experiments demonstrate the overall superiority of our approach in cross-lingual sentiment lexicon learning. This also indicates the effectiveness of the BLP approach in Chinese sentiment lexicon learning. This set of experiments is to examine the ways to build the inter-language relation. BLP-dict: The inter-language relation is built upon the translation entries from LDC7 and Universal Dictionary (UD).8 From these dictionaries (both English–Chinese and Chinese–English dictionaries), we collect 41,034 translation entries between the English and Chinese words. If the English word xi can be translated to the Chinese word xj in UD dictionary, wA(i, j) and wA(j, i) are set to 1.
BLP-MT: All the Chinese (English) words are translated into English (Chinese) by Google Translator. If the Chinese word xi can be translated to the English word xj, the wA(i, j) and wA(j, i) are set to 1. If a Chinese word is translated to an English phrase, we assume that the Chinese word is projected to each word in the English phrase. To improve the coverage, we translate the English sentiment seed words with three other methods; they are word collocation, coordinated phrase, and punctuation, as mentioned in Meng et al. (2012b).
The learned Chinese sentiment word lists are also evaluated with precision at k. As shown in Figure 5, we find that the alignment-based approach outperforms the dictionary-based and MT-based approaches. The reason that contributes to this is that we can build more inter-language relations based on word alignment, compared
7 http://projects.ldc.upenn.edu/Chinese/LDC ch.htm. 8 http://www.dicts.info/uddl.php.
with the translation entries from the dictionary and the translation pairs from Google
Translator. For example, the English word move is often translated to (shift) and
(affect, touch) by dictionaries or MT engines. From the parallel sentences, besides
these word translation pairs, the word move can be also aligned to (plain sailing bon voyage) that is commonly used in Chinese greeting texts. This translation entry is hard to find in dictionaries or by MT engines. The words are aligned between the two parallel sentences. Sometimes the word move may be forced to be aligned
to in the parallel sentences good luck and best wishes on your career move and . Thus, when building the inter-language relations with word alignment, our approach is likely to learn more sentiment word candidates. It is also the reason why the dictionary-based and MT-based approaches learn fewer sentiment words than our approach, as indicated in Table 4. According to our statistic, on average a Chinese word is connected to 2.3 and 2.1 English words if we build the inter-language relations with the dictionary and Google Translator, respectively. By building the inter-language relation with word alignment, our approach connects a Chinese word to 16.21 English words an average, which greatly increases the coverage of the learned sentiment lexicon. The following set of experiments reveals the influence of the intra-language relation. BLP-A: As the baseline of this set of experiments, it does not build the intralanguage relations with either English or Chinese WordNet synsets. Only the
inter-language relation with word alignment is used to build the graph. That means
WE, W̃E, WT, and W̃T are defined as zero matrixes. BLP-AE: Word alignment and the English WordNet synsets are used to build the
intra-English relation WE, but the intra-Chinese relation WT and W̃T are set to zero matrixes.
BLP-AC: Word alignment and the Chinese WordNet synsets are used to build the
intra-Chinese relation WT, but the intra-English relation WE and W̃E are set to zero matrixes.
As Figure 6 shows, when combining both English and Chinese intra-language relations, the precision curves of both positive and negative predictions increase. This indicates that adding the intra-language relations has a positive influence. The improvement can be explained by the ability of the intra-language relations to refine the polarity scores. For example, the English word sophisticated can be aligned to the
positive Chinese word (delicate) as well as the negative Chinese word (wily, wicked). In the GI lexicon, the English word sophisticated is labeled as positive. In the bilingual word graph that contains only the inter-language relations, the
negative Chinese word is likely to be labeled as positive. However, with the
intra-language relation, the negative Chinese word may connect to the other
negative Chinese words, like (foxy); and the Chinese positive word
may connect to the other positive Chinese words, like (elaborate). Thus the polarity score of the word can be refined by the intra-language relation in each iteration of propagation. Another advantage of the intra-language relation is that it helps to reduce the noise introduced by the inter-language relation. For example, sometimes
the Chinese positive word (help) is misaligned to the negative English word freak
by the inter-language relation, but it is also connected to the synonyms (help)
and (salutary) (which are positive) by the intra-language relations. The polarity
score of the word can be adjusted by the intra-language relation. Thus, though the inter-language relation brings in certain noisy alignments, the intra-language relation can help to refine the polarity score of the word using its intra-language relation. ρ1 and ρ2 in Equation (3) tune the English and Chinese synonym intra-language propagation, while ρ3 and ρ4 in Equation (4) adjust the English and Chinese antonym intralanguage propagation. For simplicity, let ρ1 equal ρ2 and let ρ3 equal ρ4. Then we tune ρ1,2 and ρ3,4 together. When ρ1,2 and ρ3,4 range from {1e − 2, 1e − 1, 1, 10, 100, 1000}, Precision@1K ranges from 0.631 to 0.689 and Recall ranges from 0.651 to 0.729 on average. In general, we find that when 1 ≤ ρ3,4 < ρ1,2 ≤ 10, we can obtain better results. Sentiment classification is one of the most extensively studied tasks in the community of sentiment analysis (Pang and Lee 2008). To see whether the performance improvement in lexicon learning also improves the results of sentiment classification, we apply the generated Chinese sentiment lexicons to sentence-level sentiment classification.
Data set: The NTCIR sentiment-labeled corpus is used for sentiment classification (Seki et al. 2008, 2009). We extract the Chinese sentences that have positive, negative, or neutral labels. The numbers of extracted sentences are shown in Table 5. The learned sentiment words in the Mono and BLP approaches are used as classification features. We implement the following baselines for comparison.
BSL DF: The Chinese word unigrams and bigrams are extracted from the NTCIR data set as features. We rank the features according to their frequencies and gradually increase the value of N for the Top-N classification features.
BSL LF: The words in existing Chinese sentiment lexicons are used as features. A total of 836 positive words and 1,254 negative words are collected from HowNet.9
We use LibSVM10 and perform 10-fold cross-validation on the NTCIR polarity sentences. The accuracies over N number of features are plotted in Figure 7. Our approach achieves a very promising improvement, although the features and the sentences that need to be classified are selected from different corpora. This suggests that the generated sentiment lexicon is adaptive and qualitative enough for sentiment classification.","5 conclusions and future work :In this article, we studied the task of cross-lingual sentiment lexicon learning. We built a bilingual word graph with the words in two languages and connected them with the inter-language and intra-language relations. We proposed a bilingual word graph label propagation approach to transduce the sentiment information from English sentiment words to the words in the target language. The synonym and antonym relations among
9 http://www.keenage.com/. 10 http://www.csie.ntu.edu.tw/∼cjlin/libsvm/.
the words in the same languages are leveraged to build the intra-language relations. Word alignment derived from a large parallel corpus is used to build the inter-language relations. Experiments on Chinese sentiment lexicon learning demonstrate the effectiveness of the proposed approach. There are three main conclusions from this work. First, the bilingual word graph is suitable for sentiment information transfer and the proposed approach can iteratively improve the precision of the generated sentiment lexicon. Second, building the inter-language relations with the large parallel corpus can significantly improve the coverage. Third, by incorporating the antonym relations into the bilingual word graph, the BLP approach can achieve an improvement in precision. In the future, we will explore the opportunity of expanding or generating the sentiment lexicons for multiple languages by bootstrapping.",,,,"In this article we address the task of cross-lingual sentiment lexicon learning, which aims to automatically generate sentiment lexicons for the target languages with available English sentiment lexicons. We formalize the task as a learning problem on a bilingual word graph, in which the intra-language relations among the words in the same language and the interlanguage relations among the words between different languages are properly represented. With the words in the English sentiment lexicon as seeds, we propose a bilingual word graph label propagation approach to induce sentiment polarities of the unlabeled words in the target language. Particularly, we show that both synonym and antonym word relations can be used to build the intra-language relation, and that the word alignment information derived from bilingual parallel sentences can be effectively leveraged to build the inter-language relation. The evaluation of Chinese sentiment lexicon learning shows that the proposed approach outperforms existing approaches in both precision and recall. Experiments conducted on the NTCIR data set further demonstrate the effectiveness of the learned sentiment lexicon in sentence-level sentiment classification.","[{""affiliations"": [], ""name"": ""Dehong Gao""}, {""affiliations"": [], ""name"": ""Furu Wei""}, {""affiliations"": [], ""name"": ""Wenjie Li""}, {""affiliations"": [], ""name"": ""Xiaohua Liu""}, {""affiliations"": [], ""name"": ""Ming Zhou""}]",SP:013ad460023a35d367e90751a36dd74c493752bd,"[{""authors"": [""Banea"", ""Carmen"", ""Rada Mihalcea"", ""Janyce Wiebe""], ""title"": ""Multilingual subjectivity: Are more languages better"", ""venue"": ""In Proceedings of the 23rd International Conference on Computational Linguistics,"", ""year"": 2010}, {""authors"": [""Blei"", ""David M."", ""Andrew Y. Ng"", ""Michael I. Jordan.""], ""title"": ""Latent Dirichlet allocation"", ""venue"": ""Journal of Machine Learning Research, 3:993\u20131022."", ""year"": 2003}, {""authors"": [""Boyd-Graber"", ""Jordan"", ""Philip Resnik.""], ""title"": ""Holistic sentiment analysis across languages: Multilingual supervised latent Dirichlet allocation"", ""venue"": ""Proceedings of the 2010 Conference on Empirical Methods in"", ""year"": 2010}, {""authors"": [""Das"", ""Dipanjan"", ""Slav Petrov.""], ""title"": ""Unsupervised part-of-speech tagging with bilingual graph-based projections"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational"", ""year"": 2011}, {""authors"": [""Duh"", ""Kevin"", ""Akinori Fujino"", ""Masaaki Nagata""], ""title"": ""Is machine translation ripe"", ""year"": 2011}, {""authors"": [""Esuli"", ""Andrea"", ""Fabrizio Sebastiani.""], ""title"": ""Sentiwordnet: A publicly available lexical resource for opinion mining"", ""venue"": ""Proceedings of the 3rd International Conference on Language Resources and Evaluation,"", ""year"": 2006}, {""authors"": [""Esuli"", ""Andrea"", ""Fabrizio Sebastiani.""], ""title"": ""Random-walk models of term semantics: An application to opinionrelated properties"", ""venue"": ""Proceedings of the 3rd Language and Technology Conference,"", ""year"": 2007}, {""authors"": [""Hassan"", ""Ahmed"", ""Amjad Abu-Jbara"", ""Rahul Jha"", ""Dragomir Radev.""], ""title"": ""Identifying the semantic orientation of foreign words"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2011}, {""authors"": [""Hatzivassiloglou"", ""Vasileios"", ""Kathleen R. McKeown.""], ""title"": ""Predicting the semantic orientation of adjectives"", ""venue"": ""Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and 8th Conference"", ""year"": 1997}, {""authors"": [""He"", ""Yulan"", ""Harith Alani"", ""Deyu Zhou.""], ""title"": ""Exploring English lexicon knowledge for Chinese sentiment analysis"", ""venue"": ""Proceedings of the 2010 CIPS-SIGHAN Joint Conference on Chinese Language"", ""year"": 2010}, {""authors"": [""Hu"", ""Minqing"", ""Bing Liu.""], ""title"": ""Mining and summarizing customer reviews"", ""venue"": ""Proceedings of the 10th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 168\u2013177,"", ""year"": 2004}, {""authors"": [""Hu"", ""Minqing"", ""Bing Liu.""], ""title"": ""Opinion extraction and summarization on the Web"", ""venue"": ""Proceedings of the 21st National Conference on Artificial Intelligence, pages 1,621\u20131,624, Boston, MA."", ""year"": 2006}, {""authors"": [""Jiang"", ""Jonathan Q.""], ""title"": ""Learning protein functions from bi-relational graph of proteins and function annotations"", ""venue"": ""Algorithms in Bioinformatics, 6833:128\u2013138, Springer Berlin Heidelberg."", ""year"": 2011}, {""authors"": [""Jiang"", ""Jonathan Q."", ""Lisa J. McQuay.""], ""title"": ""Predicting protein function by multi-label correlated semi-supervised learning"", ""venue"": ""IEEE/ACM Transactions on Computational Biology and Bioinformatics, 9(4):1059\u20131069."", ""year"": 2012}, {""authors"": [""Kim"", ""Soo-Min"", ""Eduard Hovy.""], ""title"": ""Determining the sentiment of opinions"", ""venue"": ""Proceedings of the 20th International Conference on Computational Linguistics, pages 355\u2013363, Geneva."", ""year"": 2004}, {""authors"": [""Kunegis"", ""Jerome"", ""Stephan Schmidt"", ""Andreas Lommatzsch"", ""J\u00fcrgen Lerner"", ""Ernesto W. De"", ""Luca Sahin Albayrak.""], ""title"": ""Spectral analysis of signed graphs for clustering, prediction and visualization"", ""venue"": ""In"", ""year"": 2010}, {""authors"": [""Li"", ""Shen"", ""Joao V. Graca"", ""Ben Taskar.""], ""title"": ""Wiki-ly supervised part-of-speech tagging"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural"", ""year"": 2012}, {""authors"": [""Liang"", ""Percy"", ""Ben Taskar"", ""Dan Klein.""], ""title"": ""Alignment by agreement"", ""venue"": ""Proceedings of the Main Conference on Human Language Technology Conference of the North American Chapter of the Association of"", ""year"": 2006}, {""authors"": [""Lu"", ""Bin"", ""Chenhao Tan"", ""Claire Cardie"", ""Benjamin K. Tsou.""], ""title"": ""Joint bilingual sentiment classification with unlabeled parallel corpora"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for"", ""year"": 2011}, {""authors"": [""Meng"", ""Xinfan"", ""Furu Wei"", ""Xiaohua Liu"", ""Ming Zhou"", ""Ge Xu"", ""Houfeng Wang.""], ""title"": ""Cross-lingual mixture model for sentiment classification"", ""venue"": ""Proceedings of the 50th Annual Meeting of the Association"", ""year"": 2012}, {""authors"": [""Meng"", ""Xinfan"", ""Furu Wei"", ""Ge Xu"", ""Longkai Zhang"", ""Xiaohua Liu"", ""Ming Zhou"", ""Houfeng Wang""], ""title"": ""Lost in translations? Building sentiment lexicons using context-based machine translation"", ""year"": 2012}, {""authors"": [""Mihalcea"", ""Rada"", ""Carmen Banea"", ""Janyce Wiebe.""], ""title"": ""Learning multilingual subjective language via cross-lingual projections"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for"", ""year"": 2007}, {""authors"": [""Miller"", ""George A.""], ""title"": ""Wordnet: A lexical database for English"", ""venue"": ""Communications of the ACM, 38(11):39\u201341."", ""year"": 1995}, {""authors"": [""Munteanu"", ""Dragos Stefan"", ""Daniel Marcu.""], ""title"": ""Improving machine translation performance by exploiting non-parallel corpora"", ""venue"": ""Journal of Computational Linguistics, 31(4):477\u2013504."", ""year"": 2005}, {""authors"": [""Ounis"", ""Iadh"", ""Craig Macdonald"", ""Ian Soboroff.""], ""title"": ""Overview of the TREC-2010 blog track"", ""venue"": ""Proceedings of the NTCIR07 workshop, pages 104\u2013111, Tokyo."", ""year"": 2008}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""Opinion mining and sentiment analysis, volume 2"", ""venue"": ""Foundations and Trends in Information Retrieval. Now Publishers, Inc."", ""year"": 2008}, {""authors"": [""Qiu"", ""Guang"", ""Bing Liu"", ""Jiajun Bu"", ""Chun Chen.""], ""title"": ""Opinion word expansion and target extraction through double propagation"", ""venue"": ""Computational Linguistics, 37(1):9\u201327."", ""year"": 2011}, {""authors"": [""Rao"", ""Delip"", ""Deepak Ravichandran.""], ""title"": ""Semi-supervised polarity lexicon induction"", ""venue"": ""Proceedings of the 12th Conference of the European Chapter of the ACL, pages 675\u2013682, Athens."", ""year"": 2009}, {""authors"": [""Riloff"", ""Ellen"", ""Janyce Wiebe"", ""Theresa Wilson.""], ""title"": ""Learning subjective nouns using extraction pattern bootstrapping"", ""venue"": ""Proceedings of the 7th Conference on Natural Language Learning, pages 25\u201332,"", ""year"": 2003}, {""authors"": [""Saad"", ""Yousef.""], ""title"": ""Iterative Methods for Sparse Linear Systems"", ""venue"": ""Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 2nd edition."", ""year"": 2003}, {""authors"": [""Seki"", ""Yohei"", ""David Kirk Evans"", ""Lun-Wei Ku"", ""Le Sun"", ""Hsin-Hsi Chen"", ""Noriko Kando.""], ""title"": ""Overview of multilingual opinion analysis task at NTCIR-7"", ""venue"": ""Proceedings of the NTCIR07 Workshop,"", ""year"": 2008}, {""authors"": [""Seki"", ""Yohei"", ""Lun-Wei Ku"", ""Le Sun"", ""Hsin-Hsi Chen"", ""Noriko Kando.""], ""title"": ""Overview of multilingual opinion analysis task at NTCIR-8: A step toward cross lingual opinion analysis"", ""venue"": ""Proceedings of the"", ""year"": 2009}, {""authors"": [""Stone"", ""Philip J.""], ""title"": ""Thematic text analysis: New agendas for analyzing text content"", ""venue"": ""Text Analysis for the Social Sciences, chapter 2. Lawerence Erlbaum, Mahwah, NJ."", ""year"": 1997}, {""authors"": [""Takamura"", ""Hiroya"", ""Takashi Inui"", ""Manabu Okumura.""], ""title"": ""Extracting semantic orientations of words using spin model"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""Talukdar"", ""Partha Pratim"", ""Koby Crammer.""], ""title"": ""New regularized algorithms for transductive learning"", ""venue"": ""Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases,"", ""year"": 2009}, {""authors"": [""Turney"", ""Peter D."", ""Michael L. Littman.""], ""title"": ""Measuring praise and criticism: Inference of semantic orientation from association"", ""venue"": ""ACM Transactions on Information Systems, 21(4):315\u2013346."", ""year"": 2003}, {""authors"": [""Wan"", ""Xiaojun.""], ""title"": ""Co-training for cross-lingual sentiment classification"", ""venue"": ""Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural"", ""year"": 2009}, {""authors"": [""Wang"", ""Hua"", ""Heng Huang"", ""Chris Ding.""], ""title"": ""Image annotation using bi-relational graph of images and semantic labels"", ""venue"": ""the 24th IEEE Conference on Computer Vision and Pattern Recognition, pages 793\u2013800,"", ""year"": 2011}, {""authors"": [""Yu"", ""Hong"", ""Vasileios Hatzivassiloglou.""], ""title"": ""Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences"", ""venue"": ""Proceedings of the 2003"", ""year"": 2003}, {""authors"": [""Zhou"", ""Dengyong"", ""Olivier Bousquet"", ""Thomas Navin Lal"", ""Jason Weston"", ""Bernhard Scholkopf.""], ""title"": ""Learning with local and global consistency"", ""venue"": ""Proceedings of Advances in Neural Information Processing"", ""year"": 2004}, {""authors"": [""Zhu"", ""Xiaojin"", ""Zoubin Ghahramani.""], ""title"": ""Learning from labeled and unlabeled data with label propagation"", ""venue"": ""Technical report CMU-CALD-02-107, Carnegie Mellon University."", ""year"": 2002}, {""authors"": [""Zhu"", ""Xiaojin"", ""Zoubin Ghahramani"", ""John Lafferty.""], ""title"": ""Semi-supervised learning using Gaussian fields and harmonic functions"", ""venue"": ""International Conference Machine Learning,"", ""year"": 2003}]",acknowledgments  :The work described in this article was supported by a Hong Kong RGC project (PolyU no. 5202/12E) and a National Nature Science Foundation of China (NSFC no. 61272291).,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :A sentiment lexicon is regarded as the most valuable resource for sentiment analysis (Pang and Lee 2008), and lays the groundwork of much sentiment analysis research, for example, sentiment classification (Yu and Hatzivassiloglou 2003; Kim and Hovy 2004) and opinion summarization (Hu and Liu 2004). To avoid manually annotating sentiment words, an automatically learning sentiment lexicon has attracted considerable attention in the community of sentiment analysis. The existing work determines word sentiment polarities either by the statistical information (e.g., the co-occurrence of words with predefined sentiment seed words) derived from a large corpus (Riloff, Wiebe, and Wilson 2003; Hu and Liu 2006) or by the word semantic information (e.g., synonym relations) found in existing human-created resources (e.g., WordNet) (Takamura, Inui, and Okumura 2005; Rao and Ravichandran 2009). However, current work mainly focuses on English sentiment lexicon generation or expansion, while sentiment lexicon learning for other languages has not been well studied.
In this article, we address the issue of cross-lingual sentiment lexicon learning, which aims to generate sentiment lexicons for a non-English language (hereafter referred to as “the target language”) with the help of the available English sentiment lexicons. The underlying motivation of this task is to leverage the existing English sentiment lexicons and substantial linguistic resources to label the sentiment polarities of the words in the target language. To this end, we need an approach to transferring the sentiment information from English words to the words in the target language. The few existing approaches first build word relations between English and the target language. Then, based on the word relation and English sentiment seed words, they determine the sentiment polarities of the words in the target language. In these two steps, relation-building plays a fundamental role because it is responsible for the transfer of sentiment information between the two languages. Two approaches are often used to connect the words in different languages in the literature. One is based on translation entries in cross-lingual dictionaries (Hassan et al. 2011). The other relies on a machine translation (MT) engine as a black box to translate the sentiment words in English to the target language (Steinberger et al. 2011). The two approaches in Duh, Fujino, and Nagata (2011) and Mihalcea, Banea, and Weibe (2007) tend to use a small set of vocabularies to translate the natural language, which leads to a low coverage of generated sentiment lexicons for the target language.
To solve this problem, we propose a generic approach to addressing the task of cross-lingual sentiment lexicon learning. Specifically, we model this task with a bilingual word graph, which is composed of two intra-language subgraphs and an interlanguage subgraph. The intra-language subgraphs are used to model the semantic relations among the words in the same languages. When building them, we incorporate both synonym and antonym word relations in a novel manner, represented by positive and negative sign weights in the subgraphs, respectively. These two intra-language subgraphs are then connected by the inter-language subgraph. We propose Bilingual word graph Label Propagation (BLP), which simultaneously takes the inter-language relations and the intra-language relations into account in an iterative way. Moreover, we leverage the word alignment information derived from a parallel corpus to build the inter-language relations. We connect two words from different languages that are aligned to each other in a parallel sentence pair. Taking advantage of a large parallel corpus, this approach significantly improves the coverage of the generated sentiment lexicon. The experimental results on Chinese sentiment lexicon learning show the effectiveness of the proposed approach in terms of both precision and recall. We further
evaluate the impact of the learned sentiment lexicon on sentence-level sentiment classification. When using words in the learned sentiment lexicon as features for sentiment classification of the target language, the sentiment classification can achieve a high performance.
We make the following contributions in this article.
1. We present a generic approach to automatically learning sentiment lexicons for the target language with the available sentiment lexicon in English, and we formalize the cross-lingual sentiment learning task on a bilingual word graph.
2. We build a bilingual word graph by using synonym and antonym word relations and propose a bilingual word graph label propagation approach, which effectively leverages the inter-language relations and both types (synonym and antonym) of the intra-language relations in sentiment lexicon learning.
3. We leverage the word alignment information derived from a large number of parallel sentences in sentiment lexicon learning. We build the inter-language relation in the bilingual word graph upon word alignment, and achieve significant results. 2 related work : In general, the work on sentiment lexicon learning focuses mainly on English and can be categorized as co-occurrence–based approaches (Hatzivassiloglou and McKeown 1997; Riloff, Wiebe, and Wilson 2003; Qiu et al. 2011) and semantic-based approaches (Mihalcea, Banea, and Wiebe 2007; Takamura, Inui, and Okumura 2005; Kim and Hovy 2004).
The co-occurrence-based approaches determine the sentiment polarity of a given word according to the statistical information, like the co-occurrence of the word to predefined sentiment seed words or the co-occurrence to product features. The statistical information is mainly derived from certain corpora. One of the earliest work conducted by Hatzivassiloglou and McKeown (1997) assumes that the conjunction words can convey the polarity relation of the two words they connect. For example, the conjunction word and tends to link two words with the same polarity, whereas the conjunction word but is likely to link two words with opposite polarities. Their approach only considers adjectives, not nouns or verbs, and it is unable to extract adjectives that are not conjoined by conjunctions. Riloff et al. (2003) define several pattern templates and extract sentiment words by two bootstrapping approaches. Turney and Littman (2003) calculate the pointwise mutual information (PMI) of a given word with positive and negative sets of sentiment words. The sentiment polarity of the word is determined by average PMI values of the positive and negative sets. To obtain PMI, they provide queries (consisting of the given word and the sentiment word) to the search engine. The number of hits and the position (if the given word is near the sentiment word) are used to estimate the association of the given word to the sentiment word. Hu and Liu (2004) research sentiment word learning on customer reviews and they assume that the sentiment words tend to be correlated with product features. The frequent nouns and noun phrases are treated as product features. Then they extract the adjective words
as sentiment words from those sentences that contain one or more product features. This approach may work on a product review corpus, where one product feature may frequently appear. But for other corpora, like news articles, this approach may not be effective. Qiu et al. (2011) combine sentiment lexicon learning and opinion target extraction. A double propagation approach is proposed to learn sentiment words and to extract opinion targets simultaneously, based on eight manually defined rules.
The semantic-based approaches determine the sentiment polarity of a given word according to the word semantic relation, like the synonyms of sentiment seed words. The word semantic relation is usually obtained from dictionaries, for example, WordNet.1 Kim and Hovy (2004) assume that the synonyms of a positive (negative) word are positive (negative) and its antonyms are negative (positive). Initializing with a set of sentiment words, they expand sentiment lexicons based on these two kinds of word relations. Kamps et al. (2004) build a synonym graph according to the synonym relation (synset) derived from WordNet. The sentiment polarity of a word is calculated by the shortest path to two sentiment words good and bad. However, the shortest path cannot precisely describe the sentiment orientation, considering there are only five steps between the word good and the word bad in WordNet (Hassan et al. 2011). Takamura et al. (2005) construct a word graph with the gloss of WordNet. Words are connected if a word appears in the gloss of another. The word sentiment polarity is determined by the weight of its connections on the word graph. Based on WordNet, Rao and Ravichandran (2009) exploit several graph-based semi-supervised learning methods like Mincuts and Label Propagation. The word polarity orientations are induced by initializing some sentiment seed words in the WordNet graph. Esuli et al. (2006, 2007) and Baccianella et al. (2010) treat sentiment word learning as a machine learning problem, that is, to classify the polarity orientations of the words in WordNet. They select seven positive words and seven negative words and expand them through the see-also and antonym relations in WordNet. These expanded words are then used for training. They train a ternary classifier to predict the sentiment polarities of all the words in WordNet and use the glosses (textual definitions of the words in WordNet) as the features of classification. The sentiment lexicon generated is the well-known SentiWordNet.2 The work on cross-lingual sentiment lexicon learning is still at an early stage and can be categorized into two types, according to how they bridge the words in two languages.
Mihalcea et al. (2007) generate sentiment lexicon for Romanian by directly translating the English sentiment words into Romanian through bilingual English–Romanian dictionaries. When confronting multiword translations, they translate the multiwords word by word. Then the validated translations must occur at least three times on the Web. The approach proposed by Hassan et al. (2011) learns sentiment words based on English WordNet and WordNets in the target languages (e.g., Hindi and Arabic). Crosslingual dictionaries are used to connect the words in two languages and the polarity of a given word is determined by the average hitting time from the word to the English sentiment word set. These approaches connect words in two languages based on crosslingual dictionaries. The main concern of these approaches is the effect of morphological inflection (i.e., a word may be mapped to multiple words in cross-lingual dictionaries).
1 http://wordnet.princeton.edu/. 2 http://sentiwordnet.isti.cnr.it/.
For example, one single English word typically has four Spanish or Italian word forms (two each for gender and for number) and many Russian word forms (due to gender, number, and case distinctions) (Steinberger et al. 2011). Usually, this approach requires an additional process to disambiguate the sentiment polarities of all the morphological variants.
To improve the sentiment classification for the target language, Banea, Mihalcea, and Wiebe (2010) translate the English sentiment lexicon into the target language using Google Translator.3 Similarly, Google Translator is used by Steinberger et al. (2011). They manually produce two high-level gold-standard sentiment lexicons for two languages (e.g., English and Spanish) and then translate them into the third language (e.g., Italian) via Google Translator. They believe that those words in the third language that appear in both translation lists are likely to be sentiment words. These approaches connect the words in two languages based on MT engines. The main concern of these approaches is the low overlapping between the vocabularies of natural documents and the vocabularies of the documents translated by MT engines (Duh, Fujino, and Nagata 2011; Meng et al. 2012a). The shortcoming of these MT-based approaches inevitably leads to low coverage.
Our task resembles the task of cross-lingual sentiment classification, like Wan (2009), Lu et al. (2011), and Meng et al. (2012a), which classifies the sentiment polarities of product reviews. Generally, these studies use semi-supervised learning approaches and regard translations from labeled English sentiment reviews as the training data. The terms in each review are leveraged as the features for training, which has proven to be effective in sentiment classification (Pang and Lee 2008). We can regard the task of sentiment lexicon learning as word-level sentiment classification. However, for wordlevel sentiment classification, it is not straightforward to extract features for a single word. Without sufficient features, it is difficult for these approaches to perform well in learning. Another line of cross-lingual sentiment classification uses Latent Dirichlet Allocation (LDA) (Blei, Ng, and Jordan 2003) or its variants, like Boyd-Graber and Resnik (2010) or He, Alani, and Zhou (2010). These studies assume that each review is a mixture of sentiments and each sentiment is a probability over words. Then they apply the LDA-like approach to model the sentiment polarity of each review. Nonetheless, this assumption may not be applicable in sentiment lexicon learning because a single word can be regarded as the minimal semantic unit, and it is difficult, if not impossible, to infer the latent topics from a single word. Recall that different from the sentiment classification of product reviews where the instances are normally independent, words in sentiment lexicon learning are highly related with each other, like synonyms and antonyms. Through these relations, the words can naturally form a word graph. Thus we use the graph-based learning approach to leverage the word distributions in sentiment lexicon learning. In the next section, we will introduce our proposed graph-based cross-lingual sentiment lexicon learning. 3 cross-lingual sentiment lexicon learning :In this work, we model the task of cross-lingual sentiment lexicon learning with a bilingual word graph, where (1) the words in the two languages are represented by the nodes in two intra-language subgraphs, respectively; (2) the synonym and antonym word relations within each language are represented by the positive and negative sign
3 http://translate.google.com/.
weights in the corresponding intra-language subgraphs; and (3) the two intra-language subgraphs are connected by an inter-language subgraph. Mathematically, we build a
graph G = (XE ∪ XT, WE ∪ W̃E ∪ WT ∪ W̃T ∪ WA) that consists of two intra-language subgraphs GE = (XE, WE ∪ W̃E) and GT = (XT, WT ∪ W̃T ) as shown in Figure 1. These two subgraphs are connected by the inter-language graph GR = (XE ∪ XT, WA). The elements of WE, WT, and WA are positive real numbers, that is, WE, WT, and WA ∈ R+, and W̃E and W̃T ∈ R−. Because G incorporates the words in two languages, we call it a Bilingual Word Graph. Specifically, the positive weights, WE and WT, represent the synonym intra-language relations, and the negative weights, W̃E and W̃T, represent the antonym intra-language relations. The inter-language relations, WA, represent the connections between the words in the two languages. For cross-lingual sentiment lexicon learning, XE = {X L E, X U E } denotes the labeled and unlabeled words in English and XT denotes the unlabeled words in the target language. Given the labels YLE = {yE1 , ..., yEl} of the seeds X L E, we aim to predict the sentiment polarities of the words XT. In the remainder of this section, we will present the bilingual word graph construction and the algorithm of bilingual word graph label propagation. We represent the words in English and in the target language as the nodes of the bilingual word graph. We use the synonym and antonym relations of the words in the
same language to build W and W̃ in the intra-language graph, respectively. In the rest of this section, we will focus on how to build the inter-language relation.
Intuitively, there are two ways to connect the words in two languages. One is to insert links to the words if there exist entry mappings between the words in bilingual dictionaries (e.g., the English–Chinese dictionary). This method is simple and straightforward, but it suffers from two limitations. (1) Dictionaries are static during a certain period, whereas the sentiment lexicon evolves over time. (2) The entries in dictionaries tend to be the expressions of formal and written languages, but people prefer using colloquial language in expressing their sentiments or opinions on-line. These limitations lead to the low coverage of the links from English to the target language. An alternative way is to use an MT engine as a black box to build the inter-language relation. One can send each word in English to a publicly available MT engine and get the translations in the target language. Edges can then be inserted into the graph between the English
words and their corresponding translations. This approach suffers from the problem of low coverage as well because MT engines tend to use a small set of vocabularies (Duh, Fujino, and Nagata 2011).
In this article, we propose to leverage a large bilingual parallel corpus, which is readily available in the MT research community, to build the bilingual word graph. The parallel corpus consists of a large number of parallel sentence pairs from two different languages that have been used as the foundation of the state-of-the-art statistical MT engines. Like the example shown in Figure 2, the two sentences in English and Chinese are parallel sentences, which express the same meaning in different languages. We can easily derive the word alignment from the sentence pairs, automatically using a stateof-the-art toolkit, like GIZA++4 or BerkeleyAligner.5 In this example, the Chinese word
(happy) is linked to the English word happy and we say that these two words are
aligned. Similarly the English words best and wishes are both aligned to (wish). The word alignment information encodes the rich association information between the words from the two languages. We are therefore motivated to leverage the parallel corpus and word alignment to build the bilingual word graph for cross-lingual sentiment lexicon learning. We take the words from both languages in the bilingual parallel corpus as the nodes in the bilingual word graph, and build the inter-language relations by connecting the two words that are aligned together in a sentence pair from a parallel corpus. There are several advantages of using a parallel corpus to build the inter-language subgraph. First, large parallel corpora are extensively used for training statistical MT engines and can be easily reused in our task. The parallel sentence pairs are usually automatically collected and mined from the Web. As a result, they contain the different and practical variations of words and phrases embedded in sentiment expressions. Second, the parallel corpus can be dynamically changed when necessary because it is relatively easy to collect from the Web. Consequentially, the novel sentiment information inferred from the parallel corpus can easily update the existing sentiment lexicons. These advantages can greatly improve the coverage of the generated sentiment lexicon, as demonstrated later in our experiments. As commonly used semi-supervised approaches, label propagation (Zhu and Ghahramani 2002) and its variants (Zhu, Ghahramani, and Lafferty 2003; Zhou et al. 2004) have been applied to many applications, such as part-of-speech tagging (Das and Petrov 2011; Li, Graca, and Taskar 2012), image annotation (Wang, Huang, and Ding 2011), protein function prediction (Jiang 2011; Jiang and McQuay 2012), and so forth.
4 http://www.statmt.org/moses/giza/GIZA++.html. 5 http://nlp.cs.berkeley.edu.
The underlying idea of label propagation is that the connected nodes in the graph tend to share the same sentiment labels. In bilingual word graph label propagation, the words tend to share same sentiment labels if they are connected by synonym relations or word alignment and tend to belong to different sentiment labels if connected by antonym relations.
In this article we propose bilingual word graph label propagation for cross-lingual sentiment lexicon learning. Let F = {FE, FT} denote the predicted labels of the unlabeled words X. The loss function can be defined as
El(F) = µ
n∑
i=1
‖fEi − yEi‖ 2 + µ
m∑
i=1
‖ fTi − yTi‖ 2 (1)
where n and m denote the numbers of English words and words in the target language. Let Y = {YE, YT} denote the initial sentiment labels of all the words; the loss function means that the prediction could not change too much from the initial label assignment. Similar to Zhou et al. (2004), we define the smoothness function to indicate that if two words are connected by synonym relation or by word alignment, then they tend to share the same sentiment label. The smoothness function can be further represented by two parts, that is, the inter-language smoothness Einters (F) and the synonym intra-language smoothness Eintras (F)
Einters (F) = 1 2
n∑
i=1
m∑
j=1
wAij‖ fEi√ dALii − fTj√ dARjj ‖2 (2)
Eintras (F) = ρ1 1 2
n∑
i,j=1
wEij‖ fEi√ dEii − fEj√ dEjj ‖2 + ρ2 1 2
m∑
i,j=1
wTij‖ fTi√ dTii − fTj√ dTjj ‖2 (3)
DAL and DAR are defined as DAL = diag( ∑ j wA(1, j), . . . , ∑ j wA(n, j)) ′ and DAR = diag( ∑ i wA(i, 1), . . . , ∑ i wA(i, m)) ′
. DE and DT are the degree matrices of the synonym intra-language relations WE and WT, respectively. We then define the distance function to indicate that if two words are connected by the antonym relation they tend to belong to different sentiment labels. The distance function can be defined as
Eintrad (F) = ρ3 1 2
n∑
i,j=1 |w̃Eij |‖ fEi√ d̃Eii − fEj√ d̃Ejj ‖2 + ρ4 1 2
m∑
i,j=1 |w̃Tij |‖ fTi√ d̃Tii − fTj√ d̃Tjj ‖2 (4)
where D̃E and D̃T are the degree matrices of the absolute value of the antonym intralanguage relations W̃E and W̃T, respectively. Intuitively, for the inter-language smoothness and the synonym intra-language smoothness, the nearer the words connect with each other, the better performance will be achieved, whereas for the antonym intralanguage distance, the farther the better. The objective functions can be defined as
arg min(E(F)) = arg min(Eintras (F) + E inter s (F) + El(F))
arg max(E(F)) = arg max(Eintrad (F))
Thus, we define the whole objective function for cross-lingual sentiment lexicon learning as
arg min(E(F)) = arg min(Eintras (F) + E inter s (F) + El(F) − E intra d (F)) (5)
To obtain the solution to Equation (5), we differentiate the objective function according to FE and FT, and we have
∂E(F) ∂FE |FE=F⋆E = ρ1SEFE + 1 2 SAFT − ρ3S̃EFE + µFE − µYE = 0 (6)
∂E(F) ∂FT |FT=F⋆T = ρ2STFT + 1 2 SA ′ FE − ρ4S̃TFT + µFT − µYT = 0 (7)
where P ′
is the transpose of the matrix P. The graph Laplacians SE and ST of the
synonym intra-language relations are (I − D − 12 E WED − 12 E ) and (I − D − 12 T WTD − 12 T ), where I is the identity matrix. The graph Laplacians S̃E and S̃T of the antonym intra-language relations are (I − D̃ − 12 E W̃ED̃ − 12 E ) and (I − D̃ − 12 T W̃TD̃ − 12 T ), which has been proven to be positive semi-definite (Kunegis et al. 2010). The graph Laplacian SA of the interlanguage relation is (I − D − 12 AL WAD − 12 AR ). From Equations (6) and (7), we can obtain the optimal solutions
(ME − SAM −1 T S ′ A)FE = 2µ(YE − SAM −1 T YT ) (8)
(MT − S ′ AM −1 E SA)FT = 2µ(YT − S ′ AM −1 E YE) (9)
where ME = 2ρ1SE − 2ρ3S̃E + 2µI and MT = 2ρ2ST − 2ρ4S̃T + 2µI. To avoid computing the inverse matrix in Equations (8) and (9), we apply the Jacobi algorithm (Saad 2003) to calculate the solutions as described in Algorithm 1. In line 1, we set the label of the positive seed xi as y L E+ i = (1, 0) and the label of the negative seed xj as y L E− j = (0, 1). We set the label YUE of the unlabeled words as zero, and then generate YE with Y L E and Y U E . Line 2 sets YT as zero matrix. In line 3, we compute the matrixes SE, S̃E, ST, S̃T, SA, and then compute the matrixes ME and MT. The sentiment information is simultaneously
Algorithm 1. Bilingual word graph label propagation
Input: Given G = (XE ∪ XT, WE ∪ W̃E ∪ WT ∪ W̃T ∪ WA),XE, label YLE for X L E, initialize µ and ρ1∼4 Output: FT for XT and FE for XE
1. Initialize YE with the English sentiment seeds 2. Set YT as zero 3. Calculate SE, S̃E, ST, S̃T, and SA, then calculate ME and MT 4. Loop 5. f (t+1) Ei = 1 (ME−SAM −1 T S ′ A)ii (2µ(YE − SAM −1 T YT )i − ∑ j 6=i(ME − SAM −1 T S ′ A)ij f (t) Ei ) 6. f (t+1) Ti = 1 (MT−S ′ AM −1 E SA)ii (2µ(YT − S ′ AM −1 E YE)i − ∑ j 6=i(MT − S ′ AM −1 E SA)ij f (t) Ti ) 7. Until FE and FT convergence
propagated through lines 4–7 until the predicted labels FE and FT are converged. For an unlabeled word xi, if | f (i, 0) − f (i, 1)| < ξ (ξ is set as 1.0E
−4), xi is regarded as neutral; if ( f (i, 0) − f (i, 1)) ≥ ξ, xi is regarded as positive; and if ( f (i, 1) − f (i, 0)) ≥ ξ, xi is regarded as negative. 4 experiment : We conduct experiments on Chinese sentiment lexicon learning. As in previous work (Baccianella, Esuli, and Sebastiani 2010), the sentiment words in General Inquirer lexicon are selected as the English seeds (Stone 1997). From the GI lexicon we collect 2,005 positive words and 1,635 negative words. To build the bilingual word graph, we adopt the Chinese–English parallel corpus, which is obtained from the news articles published by Xinhua News Agency in Chinese and English collections, using the automatic parallel sentence identification approach (Munteanu and Marcu 2005). Altogether, we collect more than 25M parallel sentence pairs in English and
Chinese. We remove all the stopwords in Chinese and English (e.g., (of) and am) together with the low-frequency words that occur fewer than 5 times. After preprocessing, we finally have more than 174,000 English words, among which 3,519 words have sentiment labels and more than 146,000 Chinese words for which we need to predict the sentiment labels. To transfer sentiment information to Chinese unlabeled words more efficiently, we remove the unlabeled English words in the word graph (i.e., XUE = Φ). The unsupervised method, namely, BerkeleyAligner, is used to align the parallel sentences in this article (Liang, Taskar, and Klein 2006). As an unsupervised method, it does not require us to manually collect training data and does not need the complex training processing, and its performance is competitive with supervised methods. With these two advantages, we can focus more on our task of cross-lingual sentiment lexicon learning. Based on the word alignment derived by BerkeleyAligner, the inter-language WA is initialized with the normalized alignment frequency. The English and Chinese versions6 of WordNet are used to build the intra-language relations WE, W̃E, WT, and W̃T, respectively. WordNet (Miller 1995) groups words into synonym sets, called synsets. We collect about 117,000 synsets from the English WordNet and about 80,000 synsets from the Chinese WordNet. In total, we obtain 8,406 and 6,312 antonym synset pairs.
We first generate both positive and negative scores for each unlabeled word and then determine the word sentiment polarities based on its scores. We rank the two sets of newly labeled sentiment words according to their polarity scores. The top-ranked Chinese words are shown in Table 1. We manually label the top-ranked 1K sentiment words. For P@10K, we sequentially divide the top 10K ranked list into ten equal parts. One hundred sentiment words are randomly selected from each part for labeling. Similar to the evaluation of TREC Blog Distillation (Ounis, Macdonald, and Soboroff 2008), all the labeled words from each approach are used in the evaluation. We then evaluate the ranked lists with two metrics, Precision@K and Recall.
6 http://www.globalwordnet.org/gwa/wordnet table.html. In this set of experiments, we examine the influence of graph topologies on sentiment lexicon learning.
Mono: This approach learns the Chinese sentiment lexicon based only on the
Chinese monolingual word graph GT = (XT, WT ∪ W̃T). Because it needs labeled sentiment words, we incorporate the English labeled sentiment words XE and the interlanguage relation WA in the first iteration. Then we set XE and WA to be zero in later iterations.
BLP-WOA (bilingual word graph without antonym): This approach is based on the bilingual word graph. It only involves the inter-language relation WA and the synonym intra-language relations WE and WT. W̃E and W̃T are set to be zero. BLP: This approach is based on the bilingual word graph. It incorporates the interlanguage relation WA, the synonym intra-language relations WE and WT, and the antonym intra-language relations W̃E and W̃T.
In these approaches, µ is set to 0.1 as in Zhou et al. (2004). The precision of these approaches are shown in Figure 3. The figure shows that the approaches based on the bilingual word graph significantly outperform the one based on the monolingual word graph. The bilingual word graph can bring in more word relations and accelerate the sentiment propagation. Besides, in the bilingual word graph, the English sentiment seed words can continually provide accurate sentiment information. Thus we observe the increase in the approaches based on the bilingual word graph in term of both precision and recall (Table 2). Meanwhile, we find that adding the antonym relation in the bilingual word graph slightly enhances precision in top-ranked words and similar findings are observed in our later experiments. It appears that the antonym relations depict word relations in a more accurate way and can refine the word sentiment scores more precisely. However, the synonym relation and word alignment relation dominate, whereas the antonym relation accounts for only a small percentage of the graph. It is hard for the antonym relation to introduce new relations into the graph and thus it cannot help to further improve recall. In this set of experiments, we compare our approach with the baseline and existing approaches.
Rule: For the intra-language relation, this approach assumes that the synonyms of a positive (negative) word are always positive (negative), and the antonyms of a positive (negative) word are always negative (positive). For the inter-language relation, we regard the Chinese word aligned to positive (negative) English words as positive (negative). If a word connects to both positive and negative English words, we regard it as objective. Based on this heuristic, we generate two sets of sentiment words.
SOP: Hassan et al. (2011) present a method to predict the semantic orientation of unlabeled words based on the mean hitting time to the two sets of sentiment seed words.
Given the graph G = (XE ∪ XT, WE ∪ W̃E ∪ WT ∪ W̃T ∪ WA), it defines the transition probability from node i to node j as
p(j|i) = wi,j∑ k wi,k
The mean hitting time h(i|j) is the average number of the weighted steps from word i to word j. Starting with the word i and ending with the sentiment word k ∈ M, the mean hitting time h(i|M) can be formally defined as
h(i|M) =
{ 0, i ∈ M∑
j∈V p(j|i) × h(j|M) + 1, otherwise
Let M+ and M− denote the GI positive and negative seeds. If h(i|M+) is greater than h(i|M−), the word xi is classified as negative; otherwise it is classified as positive. The generated positive words and negative words are then ranked according to their polarity scores, respectively.
MAD: Talukdar and Crammer (2009) propose a MAD algorithm to modify the adsorption algorithm (Baluja et al. 2008) by adding a new regularization term. In particular, besides the positive and negative labels, a dummy label is assigned to each word in the MAD approach. Two additional columns, representing the scores of the dummy label, are added into Y and F, respectively. We denote these two matrices with the dummy labels as Ŷ and F̂. Meanwhile, R̂ is used to represent the initial dummy scores of all the words. For a word xi, the newly added columns in Ŷi and F̂i are set to zero (i.e., ŷ(i, 3) = f̂ (i, 3) = 0). r̂(i, 0) and r̂(i, 1) are set to zero, and r̂(i, 3) is assigned to one. Then, the predicted label F̂i of the word xi is iteratively obtained by
F̂ (t+1) i = 1 M̂ii
(λ1Ŷi + λ2 ∑
j
WijF̂ (t) j + λ3γR̂i)
M̂ii = (λ1 + λ2 ∑
j 6=i
Wji + λ3)
λ1∼3 and γ are used to tune the importance of each iteration term. We set λ1∼2 to one, λ3 to 10, and γ to 0.1, which produces reasonably good results. After propagation, f̂ (i, 0) and f̂ (i, 1) are used to determine the sentiment polarity of the word xi. We show recall of the learned Chinese sentiment words in Table 3. Compared with BLP and SOP, the Rule approach learns fewer sentiment words. The coverage of the Rule approach is inevitably low because many words in the corpus are aligned to both positive and negative words. For example, in most cases the positive Chinese word
(helpful) is aligned to the positive English word helpful. But sometimes it is aligned (or misaligned) to the negative English words, like freak. Under this situation, the word tends to be predicted as objective. In SOP, the positive and negative scores are related to the distances of the word to the positive and negative seed words, and the distance is usually coarse-grained to depict the sentiment polarity. For example, the shortest path between the word good and the word bad in WordNet is only 5 (Kamps et al. 2004). The Rule and SOP approaches find different sentiment words. We then evaluate the learned Chinese polarity word lists by precision at k. As illustrated in Figure 4, the significance test indicates that our approach significantly outperforms the Rule and SOP approaches. The major difference of our approach is that the polarity information can be transferred between English and Chinese and within each language at the same time, whereas in the other two approaches the polarity information mainly transfers from English to Chinese and once a word gets a polarity score, it is difficult to change or refine. The idea of the MAD approach is similar to bilingual graph label propagation, but the MAD
approach fails to leverage the antonym intra-language relation. We observe that the MAD approach can achieve a comparable result to the BLP approach. MAD can obtain a smoother label score by adding a dummy label. But the dummy label does not influence the sentiment labels too much because it is not used in the determination of the word sentiment polarity. Besides, MAD cannot deal with the antonym relation. As a result, these experiments demonstrate the overall superiority of our approach in cross-lingual sentiment lexicon learning. This also indicates the effectiveness of the BLP approach in Chinese sentiment lexicon learning. This set of experiments is to examine the ways to build the inter-language relation. BLP-dict: The inter-language relation is built upon the translation entries from LDC7 and Universal Dictionary (UD).8 From these dictionaries (both English–Chinese and Chinese–English dictionaries), we collect 41,034 translation entries between the English and Chinese words. If the English word xi can be translated to the Chinese word xj in UD dictionary, wA(i, j) and wA(j, i) are set to 1.
BLP-MT: All the Chinese (English) words are translated into English (Chinese) by Google Translator. If the Chinese word xi can be translated to the English word xj, the wA(i, j) and wA(j, i) are set to 1. If a Chinese word is translated to an English phrase, we assume that the Chinese word is projected to each word in the English phrase. To improve the coverage, we translate the English sentiment seed words with three other methods; they are word collocation, coordinated phrase, and punctuation, as mentioned in Meng et al. (2012b).
The learned Chinese sentiment word lists are also evaluated with precision at k. As shown in Figure 5, we find that the alignment-based approach outperforms the dictionary-based and MT-based approaches. The reason that contributes to this is that we can build more inter-language relations based on word alignment, compared
7 http://projects.ldc.upenn.edu/Chinese/LDC ch.htm. 8 http://www.dicts.info/uddl.php.
with the translation entries from the dictionary and the translation pairs from Google
Translator. For example, the English word move is often translated to (shift) and
(affect, touch) by dictionaries or MT engines. From the parallel sentences, besides
these word translation pairs, the word move can be also aligned to (plain sailing bon voyage) that is commonly used in Chinese greeting texts. This translation entry is hard to find in dictionaries or by MT engines. The words are aligned between the two parallel sentences. Sometimes the word move may be forced to be aligned
to in the parallel sentences good luck and best wishes on your career move and . Thus, when building the inter-language relations with word alignment, our approach is likely to learn more sentiment word candidates. It is also the reason why the dictionary-based and MT-based approaches learn fewer sentiment words than our approach, as indicated in Table 4. According to our statistic, on average a Chinese word is connected to 2.3 and 2.1 English words if we build the inter-language relations with the dictionary and Google Translator, respectively. By building the inter-language relation with word alignment, our approach connects a Chinese word to 16.21 English words an average, which greatly increases the coverage of the learned sentiment lexicon. The following set of experiments reveals the influence of the intra-language relation. BLP-A: As the baseline of this set of experiments, it does not build the intralanguage relations with either English or Chinese WordNet synsets. Only the
inter-language relation with word alignment is used to build the graph. That means
WE, W̃E, WT, and W̃T are defined as zero matrixes. BLP-AE: Word alignment and the English WordNet synsets are used to build the
intra-English relation WE, but the intra-Chinese relation WT and W̃T are set to zero matrixes.
BLP-AC: Word alignment and the Chinese WordNet synsets are used to build the
intra-Chinese relation WT, but the intra-English relation WE and W̃E are set to zero matrixes.
As Figure 6 shows, when combining both English and Chinese intra-language relations, the precision curves of both positive and negative predictions increase. This indicates that adding the intra-language relations has a positive influence. The improvement can be explained by the ability of the intra-language relations to refine the polarity scores. For example, the English word sophisticated can be aligned to the
positive Chinese word (delicate) as well as the negative Chinese word (wily, wicked). In the GI lexicon, the English word sophisticated is labeled as positive. In the bilingual word graph that contains only the inter-language relations, the
negative Chinese word is likely to be labeled as positive. However, with the
intra-language relation, the negative Chinese word may connect to the other
negative Chinese words, like (foxy); and the Chinese positive word
may connect to the other positive Chinese words, like (elaborate). Thus the polarity score of the word can be refined by the intra-language relation in each iteration of propagation. Another advantage of the intra-language relation is that it helps to reduce the noise introduced by the inter-language relation. For example, sometimes
the Chinese positive word (help) is misaligned to the negative English word freak
by the inter-language relation, but it is also connected to the synonyms (help)
and (salutary) (which are positive) by the intra-language relations. The polarity
score of the word can be adjusted by the intra-language relation. Thus, though the inter-language relation brings in certain noisy alignments, the intra-language relation can help to refine the polarity score of the word using its intra-language relation. ρ1 and ρ2 in Equation (3) tune the English and Chinese synonym intra-language propagation, while ρ3 and ρ4 in Equation (4) adjust the English and Chinese antonym intralanguage propagation. For simplicity, let ρ1 equal ρ2 and let ρ3 equal ρ4. Then we tune ρ1,2 and ρ3,4 together. When ρ1,2 and ρ3,4 range from {1e − 2, 1e − 1, 1, 10, 100, 1000}, Precision@1K ranges from 0.631 to 0.689 and Recall ranges from 0.651 to 0.729 on average. In general, we find that when 1 ≤ ρ3,4 < ρ1,2 ≤ 10, we can obtain better results. Sentiment classification is one of the most extensively studied tasks in the community of sentiment analysis (Pang and Lee 2008). To see whether the performance improvement in lexicon learning also improves the results of sentiment classification, we apply the generated Chinese sentiment lexicons to sentence-level sentiment classification.
Data set: The NTCIR sentiment-labeled corpus is used for sentiment classification (Seki et al. 2008, 2009). We extract the Chinese sentences that have positive, negative, or neutral labels. The numbers of extracted sentences are shown in Table 5. The learned sentiment words in the Mono and BLP approaches are used as classification features. We implement the following baselines for comparison.
BSL DF: The Chinese word unigrams and bigrams are extracted from the NTCIR data set as features. We rank the features according to their frequencies and gradually increase the value of N for the Top-N classification features.
BSL LF: The words in existing Chinese sentiment lexicons are used as features. A total of 836 positive words and 1,254 negative words are collected from HowNet.9
We use LibSVM10 and perform 10-fold cross-validation on the NTCIR polarity sentences. The accuracies over N number of features are plotted in Figure 7. Our approach achieves a very promising improvement, although the features and the sentences that need to be classified are selected from different corpora. This suggests that the generated sentiment lexicon is adaptive and qualitative enough for sentiment classification. 5 conclusions and future work :In this article, we studied the task of cross-lingual sentiment lexicon learning. We built a bilingual word graph with the words in two languages and connected them with the inter-language and intra-language relations. We proposed a bilingual word graph label propagation approach to transduce the sentiment information from English sentiment words to the words in the target language. The synonym and antonym relations among
9 http://www.keenage.com/. 10 http://www.csie.ntu.edu.tw/∼cjlin/libsvm/.
the words in the same languages are leveraged to build the intra-language relations. Word alignment derived from a large parallel corpus is used to build the inter-language relations. Experiments on Chinese sentiment lexicon learning demonstrate the effectiveness of the proposed approach. There are three main conclusions from this work. First, the bilingual word graph is suitable for sentiment information transfer and the proposed approach can iteratively improve the precision of the generated sentiment lexicon. Second, building the inter-language relations with the large parallel corpus can significantly improve the coverage. Third, by incorporating the antonym relations into the bilingual word graph, the BLP approach can achieve an improvement in precision. In the future, we will explore the opportunity of expanding or generating the sentiment lexicons for multiple languages by bootstrapping. In this article we address the task of cross-lingual sentiment lexicon learning, which aims to automatically generate sentiment lexicons for the target languages with available English sentiment lexicons. We formalize the task as a learning problem on a bilingual word graph, in which the intra-language relations among the words in the same language and the interlanguage relations among the words between different languages are properly represented. With the words in the English sentiment lexicon as seeds, we propose a bilingual word graph label propagation approach to induce sentiment polarities of the unlabeled words in the target language. Particularly, we show that both synonym and antonym word relations can be used to build the intra-language relation, and that the word alignment information derived from bilingual parallel sentences can be effectively leveraged to build the inter-language relation. The evaluation of Chinese sentiment lexicon learning shows that the proposed approach outperforms existing approaches in both precision and recall. Experiments conducted on the NTCIR data set further demonstrate the effectiveness of the learned sentiment lexicon in sentence-level sentiment classification. [{""affiliations"": [], ""name"": ""Dehong Gao""}, {""affiliations"": [], ""name"": ""Furu Wei""}, {""affiliations"": [], ""name"": ""Wenjie Li""}, {""affiliations"": [], ""name"": ""Xiaohua Liu""}, {""affiliations"": [], ""name"": ""Ming Zhou""}] SP:013ad460023a35d367e90751a36dd74c493752bd [{""authors"": [""Banea"", ""Carmen"", ""Rada Mihalcea"", ""Janyce Wiebe""], ""title"": ""Multilingual subjectivity: Are more languages better"", ""venue"": ""In Proceedings of the 23rd International Conference on Computational Linguistics,"", ""year"": 2010}, {""authors"": [""Blei"", ""David M."", ""Andrew Y. Ng"", ""Michael I. Jordan.""], ""title"": ""Latent Dirichlet allocation"", ""venue"": ""Journal of Machine Learning Research, 3:993\u20131022."", ""year"": 2003}, {""authors"": [""Boyd-Graber"", ""Jordan"", ""Philip Resnik.""], ""title"": ""Holistic sentiment analysis across languages: Multilingual supervised latent Dirichlet allocation"", ""venue"": ""Proceedings of the 2010 Conference on Empirical Methods in"", ""year"": 2010}, {""authors"": [""Das"", ""Dipanjan"", ""Slav Petrov.""], ""title"": ""Unsupervised part-of-speech tagging with bilingual graph-based projections"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational"", ""year"": 2011}, {""authors"": [""Duh"", ""Kevin"", ""Akinori Fujino"", ""Masaaki Nagata""], ""title"": ""Is machine translation ripe"", ""year"": 2011}, {""authors"": [""Esuli"", ""Andrea"", ""Fabrizio Sebastiani.""], ""title"": ""Sentiwordnet: A publicly available lexical resource for opinion mining"", ""venue"": ""Proceedings of the 3rd International Conference on Language Resources and Evaluation,"", ""year"": 2006}, {""authors"": [""Esuli"", ""Andrea"", ""Fabrizio Sebastiani.""], ""title"": ""Random-walk models of term semantics: An application to opinionrelated properties"", ""venue"": ""Proceedings of the 3rd Language and Technology Conference,"", ""year"": 2007}, {""authors"": [""Hassan"", ""Ahmed"", ""Amjad Abu-Jbara"", ""Rahul Jha"", ""Dragomir Radev.""], ""title"": ""Identifying the semantic orientation of foreign words"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2011}, {""authors"": [""Hatzivassiloglou"", ""Vasileios"", ""Kathleen R. McKeown.""], ""title"": ""Predicting the semantic orientation of adjectives"", ""venue"": ""Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and 8th Conference"", ""year"": 1997}, {""authors"": [""He"", ""Yulan"", ""Harith Alani"", ""Deyu Zhou.""], ""title"": ""Exploring English lexicon knowledge for Chinese sentiment analysis"", ""venue"": ""Proceedings of the 2010 CIPS-SIGHAN Joint Conference on Chinese Language"", ""year"": 2010}, {""authors"": [""Hu"", ""Minqing"", ""Bing Liu.""], ""title"": ""Mining and summarizing customer reviews"", ""venue"": ""Proceedings of the 10th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 168\u2013177,"", ""year"": 2004}, {""authors"": [""Hu"", ""Minqing"", ""Bing Liu.""], ""title"": ""Opinion extraction and summarization on the Web"", ""venue"": ""Proceedings of the 21st National Conference on Artificial Intelligence, pages 1,621\u20131,624, Boston, MA."", ""year"": 2006}, {""authors"": [""Jiang"", ""Jonathan Q.""], ""title"": ""Learning protein functions from bi-relational graph of proteins and function annotations"", ""venue"": ""Algorithms in Bioinformatics, 6833:128\u2013138, Springer Berlin Heidelberg."", ""year"": 2011}, {""authors"": [""Jiang"", ""Jonathan Q."", ""Lisa J. McQuay.""], ""title"": ""Predicting protein function by multi-label correlated semi-supervised learning"", ""venue"": ""IEEE/ACM Transactions on Computational Biology and Bioinformatics, 9(4):1059\u20131069."", ""year"": 2012}, {""authors"": [""Kim"", ""Soo-Min"", ""Eduard Hovy.""], ""title"": ""Determining the sentiment of opinions"", ""venue"": ""Proceedings of the 20th International Conference on Computational Linguistics, pages 355\u2013363, Geneva."", ""year"": 2004}, {""authors"": [""Kunegis"", ""Jerome"", ""Stephan Schmidt"", ""Andreas Lommatzsch"", ""J\u00fcrgen Lerner"", ""Ernesto W. De"", ""Luca Sahin Albayrak.""], ""title"": ""Spectral analysis of signed graphs for clustering, prediction and visualization"", ""venue"": ""In"", ""year"": 2010}, {""authors"": [""Li"", ""Shen"", ""Joao V. Graca"", ""Ben Taskar.""], ""title"": ""Wiki-ly supervised part-of-speech tagging"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural"", ""year"": 2012}, {""authors"": [""Liang"", ""Percy"", ""Ben Taskar"", ""Dan Klein.""], ""title"": ""Alignment by agreement"", ""venue"": ""Proceedings of the Main Conference on Human Language Technology Conference of the North American Chapter of the Association of"", ""year"": 2006}, {""authors"": [""Lu"", ""Bin"", ""Chenhao Tan"", ""Claire Cardie"", ""Benjamin K. Tsou.""], ""title"": ""Joint bilingual sentiment classification with unlabeled parallel corpora"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for"", ""year"": 2011}, {""authors"": [""Meng"", ""Xinfan"", ""Furu Wei"", ""Xiaohua Liu"", ""Ming Zhou"", ""Ge Xu"", ""Houfeng Wang.""], ""title"": ""Cross-lingual mixture model for sentiment classification"", ""venue"": ""Proceedings of the 50th Annual Meeting of the Association"", ""year"": 2012}, {""authors"": [""Meng"", ""Xinfan"", ""Furu Wei"", ""Ge Xu"", ""Longkai Zhang"", ""Xiaohua Liu"", ""Ming Zhou"", ""Houfeng Wang""], ""title"": ""Lost in translations? Building sentiment lexicons using context-based machine translation"", ""year"": 2012}, {""authors"": [""Mihalcea"", ""Rada"", ""Carmen Banea"", ""Janyce Wiebe.""], ""title"": ""Learning multilingual subjective language via cross-lingual projections"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for"", ""year"": 2007}, {""authors"": [""Miller"", ""George A.""], ""title"": ""Wordnet: A lexical database for English"", ""venue"": ""Communications of the ACM, 38(11):39\u201341."", ""year"": 1995}, {""authors"": [""Munteanu"", ""Dragos Stefan"", ""Daniel Marcu.""], ""title"": ""Improving machine translation performance by exploiting non-parallel corpora"", ""venue"": ""Journal of Computational Linguistics, 31(4):477\u2013504."", ""year"": 2005}, {""authors"": [""Ounis"", ""Iadh"", ""Craig Macdonald"", ""Ian Soboroff.""], ""title"": ""Overview of the TREC-2010 blog track"", ""venue"": ""Proceedings of the NTCIR07 workshop, pages 104\u2013111, Tokyo."", ""year"": 2008}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""Opinion mining and sentiment analysis, volume 2"", ""venue"": ""Foundations and Trends in Information Retrieval. Now Publishers, Inc."", ""year"": 2008}, {""authors"": [""Qiu"", ""Guang"", ""Bing Liu"", ""Jiajun Bu"", ""Chun Chen.""], ""title"": ""Opinion word expansion and target extraction through double propagation"", ""venue"": ""Computational Linguistics, 37(1):9\u201327."", ""year"": 2011}, {""authors"": [""Rao"", ""Delip"", ""Deepak Ravichandran.""], ""title"": ""Semi-supervised polarity lexicon induction"", ""venue"": ""Proceedings of the 12th Conference of the European Chapter of the ACL, pages 675\u2013682, Athens."", ""year"": 2009}, {""authors"": [""Riloff"", ""Ellen"", ""Janyce Wiebe"", ""Theresa Wilson.""], ""title"": ""Learning subjective nouns using extraction pattern bootstrapping"", ""venue"": ""Proceedings of the 7th Conference on Natural Language Learning, pages 25\u201332,"", ""year"": 2003}, {""authors"": [""Saad"", ""Yousef.""], ""title"": ""Iterative Methods for Sparse Linear Systems"", ""venue"": ""Society for Industrial and Applied Mathematics, Philadelphia, PA, USA, 2nd edition."", ""year"": 2003}, {""authors"": [""Seki"", ""Yohei"", ""David Kirk Evans"", ""Lun-Wei Ku"", ""Le Sun"", ""Hsin-Hsi Chen"", ""Noriko Kando.""], ""title"": ""Overview of multilingual opinion analysis task at NTCIR-7"", ""venue"": ""Proceedings of the NTCIR07 Workshop,"", ""year"": 2008}, {""authors"": [""Seki"", ""Yohei"", ""Lun-Wei Ku"", ""Le Sun"", ""Hsin-Hsi Chen"", ""Noriko Kando.""], ""title"": ""Overview of multilingual opinion analysis task at NTCIR-8: A step toward cross lingual opinion analysis"", ""venue"": ""Proceedings of the"", ""year"": 2009}, {""authors"": [""Stone"", ""Philip J.""], ""title"": ""Thematic text analysis: New agendas for analyzing text content"", ""venue"": ""Text Analysis for the Social Sciences, chapter 2. Lawerence Erlbaum, Mahwah, NJ."", ""year"": 1997}, {""authors"": [""Takamura"", ""Hiroya"", ""Takashi Inui"", ""Manabu Okumura.""], ""title"": ""Extracting semantic orientations of words using spin model"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""Talukdar"", ""Partha Pratim"", ""Koby Crammer.""], ""title"": ""New regularized algorithms for transductive learning"", ""venue"": ""Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases,"", ""year"": 2009}, {""authors"": [""Turney"", ""Peter D."", ""Michael L. Littman.""], ""title"": ""Measuring praise and criticism: Inference of semantic orientation from association"", ""venue"": ""ACM Transactions on Information Systems, 21(4):315\u2013346."", ""year"": 2003}, {""authors"": [""Wan"", ""Xiaojun.""], ""title"": ""Co-training for cross-lingual sentiment classification"", ""venue"": ""Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural"", ""year"": 2009}, {""authors"": [""Wang"", ""Hua"", ""Heng Huang"", ""Chris Ding.""], ""title"": ""Image annotation using bi-relational graph of images and semantic labels"", ""venue"": ""the 24th IEEE Conference on Computer Vision and Pattern Recognition, pages 793\u2013800,"", ""year"": 2011}, {""authors"": [""Yu"", ""Hong"", ""Vasileios Hatzivassiloglou.""], ""title"": ""Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences"", ""venue"": ""Proceedings of the 2003"", ""year"": 2003}, {""authors"": [""Zhou"", ""Dengyong"", ""Olivier Bousquet"", ""Thomas Navin Lal"", ""Jason Weston"", ""Bernhard Scholkopf.""], ""title"": ""Learning with local and global consistency"", ""venue"": ""Proceedings of Advances in Neural Information Processing"", ""year"": 2004}, {""authors"": [""Zhu"", ""Xiaojin"", ""Zoubin Ghahramani.""], ""title"": ""Learning from labeled and unlabeled data with label propagation"", ""venue"": ""Technical report CMU-CALD-02-107, Carnegie Mellon University."", ""year"": 2002}, {""authors"": [""Zhu"", ""Xiaojin"", ""Zoubin Ghahramani"", ""John Lafferty.""], ""title"": ""Semi-supervised learning using Gaussian fields and harmonic functions"", ""venue"": ""International Conference Machine Learning,"", ""year"": 2003}] acknowledgments  :The work described in this article was supported by a Hong Kong RGC project (PolyU no. 5202/12E) and a National Nature Science Foundation of China (NSFC no. 61272291).","5 conclusions and future work :In this article, we studied the task of cross-lingual sentiment lexicon learning. We built a bilingual word graph with the words in two languages and connected them with the inter-language and intra-language relations. We proposed a bilingual word graph label propagation approach to transduce the sentiment information from English sentiment words to the words in the target language. The synonym and antonym relations among
9 http://www.keenage.com/. 10 http://www.csie.ntu.edu.tw/∼cjlin/libsvm/.
the words in the same languages are leveraged to build the intra-language relations. Word alignment derived from a large parallel corpus is used to build the inter-language relations. Experiments on Chinese sentiment lexicon learning demonstrate the effectiveness of the proposed approach. There are three main conclusions from this work. First, the bilingual word graph is suitable for sentiment information transfer and the proposed approach can iteratively improve the precision of the generated sentiment lexicon. Second, building the inter-language relations with the large parallel corpus can significantly improve the coverage. Third, by incorporating the antonym relations into the bilingual word graph, the BLP approach can achieve an improvement in precision. In the future, we will explore the opportunity of expanding or generating the sentiment lexicons for multiple languages by bootstrapping."
"1 introduction :Interest in applying natural language processing (NLP) technology to medical information has increased in recent years. Much of this work has been focused on information retrieval and extraction from clinical notes, electronic medical records, and biomedical academic literature, but there has been some work in directly analyzing the spoken language of individuals elicited during the administration of diagnostic instruments in clinical settings. Analyzing spoken language data can reveal information not only
∗ Rochester Institute of Technology, College of Liberal Arts, 92 Lomb Memorial Dr., Rochester, NY 14623. E-mail: emilypx@rit.edu.
∗∗ Google, Inc., 1001 SW Fifth Avenue, Suite 1100, Portland OR 97204. E-mail: roarkbr@gmail.com.
Submission received: 30 December 2013; revised submission received: 21 January 2015; accepted for publication: 4 May 2015.
doi:10.1162/COLI a 00232
© 2015 Association for Computational Linguistics
about impairments in language but also about a patient’s neurological status with respect to other cognitive processes such as memory and executive function, which are often impaired in individuals with neurodevelopmental disorders, such as autism and language impairment, and neurodegenerative conditions, particularly dementia.
Many widely used instruments for diagnosing certain neurological disorders include a task in which the person must produce an uninterrupted stream of spontaneous spoken language in response to a stimulus. A person might be asked, for instance, to retell a brief narrative or to describe the events depicted in a drawing. Much of the previous work in applying NLP techniques to such clinically elicited spoken language data has relied on parsing and language modeling to enable the automatic extraction of linguistic features, such as syntactic complexity and measures of vocabulary use and diversity, which can then be used as markers for various neurological impairments (Solorio and Liu 2008; Gabani et al. 2009; Roark et al. 2011; de la Rosa et al. 2013; Fraser et al. 2014). In this article, we instead use NLP techniques to analyze the content, rather than the linguistic characteristics, of weakly structured spoken language data elicited using neuropsychological assessment instruments. We will show that the content of such spoken responses contains information that can be used for accurate screening for neurodegenerative disorders.
The features we explore are grounded in the idea that individuals recalling the same narrative are likely to use the same sorts of words and semantic concepts. In other words, a retelling of a narrative will be faithful to the source narrative and similar to other retellings. This similarity can be measured with techniques such as latent semantic analysis (LSA) cosine distance or the summary-level statistics that are widely used in evaluation of machine translation or automatic summarization, such as BLEU, Meteor, or ROUGE. Perhaps not surprisingly, however, previous work in using this type of spoken language data suggests that people with neurological impairments tend to include irrelevant or off-topic information and to exclude important pieces of information, or story elements, in their retellings that are usually included by neurotypical individuals (Hier, Hagenlocker, and Shindler 1985; Ulatowska et al. 1988; Chenery and Murdoch 1994; Chapman et al. 1995; Ehrlich, Obler, and Clark 1997; Vuorinen, Laine, and Rinne 2000; Creamer and Schmitter-Edgecombe 2010). Thus, it is often not the quantity of correctly recalled information but the quality of that information that reveals the most about a person’s diagnostic status. Summary statistics like LSA cosine distance and BLEU, which are measures of the overall degree of similarity between two texts, fail to capture these sorts of patterns. The work discussed here is an attempt to reveal these patterns and to leverage them for diagnostic classification of individuals with neurodegenerative conditions, including mild cognitive impairment and dementia of the Alzheimer’s type.
Our method for extracting the elements used in a retelling of a narrative relies on establishing a word alignment between a retelling and a source narrative. Given the correspondences between the words used in a retelling and the words used in the source narrative, we can determine with relative ease the identities of the story elements of the source narrative that were used in the retelling. These word alignments are much like those used to build machine translation models. The amount of data required to generate accurate word alignment models for machine translation, however, far exceeds the amount of monolingual source-to-retelling parallel data available to train word alignment models for our task. We therefore combine several approaches for producing reliable word alignments that exploit the peculiarities of our training data, including an entirely novel alignment approach relying on random walks on graphs.
In this article, we demonstrate that this approach to word alignment is as accurate as and more efficient than standard hidden Markov model (HMM)-based alignment (derived using the Berkeley aligner [Liang, Taskar, and Klein 2006]) for this particular data. In addition, we show that the presence or absence of specific story elements in a narrative retelling, extracted automatically from these task-specific word alignments, predicts diagnostic group membership more reliably than not only other dementia screening tools but also the lexical and semantic overlap measures widely used in NLP to evaluate pairwise language sample similarity. Finally, we apply our techniques to a picture description task that lacks an existing scoring mechanism, highlighting the generalizability and adaptability of these techniques.
The importance of accurate screening tools for neurodegenerative disorders cannot be overstated given the increased prevalence of these disorders currently being observed worldwide. In the industrialized world, for the first time in recorded history, the population over 60 years of age outnumbers the population under 15 years of age, and it is expected to be double that of children by 2050 (United Nations 2002). As the elderly population grows and as researchers find new ways to slow or halt the progression of dementia, the demand for objective, simple, and noninvasive screening tools for dementia and related disorders will grow. Although we will not discuss the application of our methods to the narratives of children, the need for simple screening protocols for neurodevelopmental disorders such as autism and language impairment is equally urgent. The results presented here indicate that the path toward these goals might include automated spoken language analysis.","2 background : Because of the variety of intact cognitive functions required to generate a narrative, the inability to coherently produce or recall a narrative is associated with many different disorders, including not only neurodegenerative conditions related to dementia, but also autism (Tager-Flusberg 1995; Diehl, Bennetto, and Young 2006), language impairment (Norbury and Bishop 2003; Bishop and Donlan 2005), attention deficit disorder (Tannock, Purvis, and Schachar 1993), and schizophrenia (Lysaker et al. 2003). The bulk of the research presented here, however, focuses on the utility of a particular narrative recall task, the Wechsler Logical Memory subtest of the Wechsler Memory Scale (Wechsler 1997), for diagnosing mild cognitive impairment (MCI). (This and other abbreviations are listed in Table 1.)
MCI is the stage of cognitive decline between the sort of decline expected in typical aging and the decline associated with dementia or Alzheimer’s disease (Petersen et al. 1999; Ritchie and Touchon 2000; Petersen 2011). MCI is characterized by subtle deficits in functions of memory and cognition that are clinically significant but do not prevent carrying out the activities of daily life. This intermediary phase of decline has been identified and named numerous times: mild cognitive decline, mild neurocognitive decline, very mild dementia, isolated memory impairment, questionable dementia, and incipient dementia. Although there continues to be disagreement about the diagnostic validity of the designation (Ritchie and Touchon 2000; Ritchie, Artero, and Touchon 2001), a number of recent studies have found evidence that seniors with some subtypes of MCI are significantly more likely to develop dementia than the population as a whole (Busse et al. 2006; Manly et al. 2008; Plassman et al. 2008). Early detection can benefit both patients and researchers investigating treatments for halting or slowing the progression of dementia, but identifying MCI can be problematic, as most dementia screening instruments, such as the Mini-Mental State Exam (MMSE) (Folstein, Folstein, and McHugh 1975), lack sufficient sensitivity to the very subtle cognitive deficits that characterize the disorder (Morris et al. 2001; Ravaglia et al. 2005; Hoops et al. 2009). Diagnosis of MCI currently requires both a lengthy neuropsychological evaluation of the patient and an interview with a family member or close associate, both of which should be repeated at regular intervals in order to have a baseline for future comparison. One goal of the work presented here is to determine whether an analysis of spoken language responses to a narrative recall task, the Wechsler Logical Memory subtest, can be used as a more efficient and less intrusive screening tool for MCI. In the Wechsler Logical Memory (WLM) narrative recall subtest of the Wechsler Memory Scale, the individual listens to a brief narrative and must verbally retell the narrative to the examiner once immediately upon hearing the story and again after a delay of 20 to 30 minutes. The examiner scores each retelling according to how many story elements the patient uses in the retelling. The standard scoring procedure, described in more detail in Section 3.2, results in a single summary score for each retelling, immediate and delayed, corresponding to the total number of story elements recalled in that retelling.
The Anna Thompson narrative, shown in Figure 1 (later in this article), has been used as the primary WLM narrative for over 70 years and has been found to be sensitive to dementia and related conditions, particularly in combination with tests of verbal fluency and memory. Multiple studies have demonstrated a significant difference in performance on the WLM between individuals with MCI and typically aging controls under the standard scoring procedure (Storandt and Hill 1989; Petersen et al. 1999; Wang and Zhou 2002; Nordlund et al. 2005). Further studies have shown that performance on the WLM can help predict whether MCI will progress into Alzheimer’s disease (Morris et al. 2001; Artero et al. 2003; Tierney et al. 2005). The WLM can also serve as a cognitive indicator of physiological characteristics associated with Alzheimer’s disease. WLM scores in the impaired range are associated with the presence of changes in Pittsburgh compound B and cerebrospinal fluid amyloid beta protein, two biomarkers of Alzheimer’s disease (Galvin et al. 2010). Poor performance on the WLM and other narrative memory tests has also been strongly correlated with increased density
of Alzheimer related lesions detected in postmortem neuropathological studies, even in the absence of previously reported or detected dementia (Schmitt et al. 2000; Bennett et al. 2006; Price et al. 2009).
We note that clinicians do not use the WLM as a diagnostic test by itself for MCI or any other type of dementia. The WLM summary score is just one of a large number of instrumentally derived scores of memory and cognitive function that, in combination with one another and with a clinician’s expert observations and examination, can indicate the presence of a dementia, aphasia, or other neurological disorder. Much of the previous work in applying automated analysis of unannotated transcripts of narratives for diagnostic purposes has focused not on evaluating properties specific to narratives but rather on using narratives as a data source from which to extract speech and language features. Solorio and Liu (2008) were able to distinguish the narratives of a small set of children with specific language impairment (SLI) from those of typically developing children using perplexity scores derived from part-ofspeech language models. In a follow-up study on a larger group of children, Gabani et al. (2009) again used part-of-speech language models in an attempt to characterize the agrammaticality that is associated with language impairment. Two part-of-speech language models were trained for that experiment: one on the language of children with SLI and one on the language of typically developing children. The perplexity of each child’s utterances was calculated according to each of the models. In addition, the authors extracted a number of other structural linguistic features including mean length of utterance, total words used in the narrative, and measures of accurate subject–verb agreement. These scores collectively performed well in distinguishing children with language impairment, achieving an F1 measure of just over 70% when used within a support vector machine (SVM) for classification. In a continuation of this work, de la Rosa et al. (2013) explored complex language-model-based lexical and syntactic features to more accurately characterize the language used in narratives by children with language impairment.
Roark et al. (2011) extracted a subset of the features used by Gabani et al. (2009), along with a much larger set of language complexity features derived from syntactic parse trees for utterances from narratives produced by elderly individuals for the diagnosis of MCI. These features included simple measures, such as words per clause, and more complex measures of tree depth, embedding, and branching, such as Frazier and Yngve scores. Selecting a subset of these features for classification with an SVM yielded a classification accuracy of 0.73, as measured by the area under the receiver operating characteristic curve (AUC). A similar approach was followed by Fraser et al. (2014) to distinguish different types of primary progressive aphasia, a group of subtypes of dementia distinct from Alzheimer’s disease and MCI, in a small group of elderly individuals. The authors considered almost 60 linguistic features, including some of those explored by Roark et al. (2011) as well as numerous others relating to part-of-speech frequencies and ratios. Using a variety of classifiers and feature combinations for three different two-way classification tasks, the authors achieved classification accuracies ranging between 0.71 and 1.0.
An alternative to analyzing narratives in terms of syntactic and lexical features is to evaluate the content of the narrative retellings themselves in terms of their fidelity to the source narrative. Hakkani-Tur, Vergyri, and Tur (2010) developed a method of
automatically evaluating an audio recording of a picture description task, in which the patient looks at a picture and narrates the events occurring in the picture, similar to the task we will be analyzing in Section 8. After using automatic speech recognition (ASR) to transcribe the recording, the authors measured unigram overlap between the ASR output transcript and a predefined list of key semantic concepts. This unigram overlap measure correlated highly with manually assigned counts of these semantic concepts. The authors did not investigate whether the scores, derived either manually or automatically, were associated with any particular diagnostic group or disorder.
Dunn et al. (2002) were among the first to apply automated methods specifically to scoring the WLM subtest and determining the relationship between these scores and measures of cognitive function. The authors used Latent Semantic Analysis (LSA) to measure the semantic distance from a retelling to the source narrative. The LSA scores correlated very highly with the scores assigned by examiners under the standard scoring guidelines and with independent measures of cognitive functioning. In subsequent work comparing individuals with and without an English-speaking background (Lautenschlager et al. 2006), the authors proposed that LSA-based scoring of the WLM as a cognitive measure is less biased against people with different linguistic and cultural backgrounds than other widely used cognitive measures. This work demonstrates not only that accurate automated scoring of narrative recall tasks is possible but also that the objectivity offered by automated measures has specific benefits for tests like the WLM, which are often administered by practitioners working in a community setting and serving a diverse population. We will compare the utility of this approach with our alignment-based approach subsequently in the article.
More recently, Lehr et al. (2013) used a supervised method for scoring the responses to the WLM, transcribed both manually and via ASR, using conditional random fields. This technique resulted in slightly higher scoring and classification accuracy than the unsupervised method described here. An unsupervised variant of their algorithm, which relied on the methods described in this article to provide training data to the conditional random field, yielded about half of the scoring gains and nearly all of the classification gains of what we report here. A hybrid method that used the methods in this article to derive features was the best performing system in that paper. Hence the methods described here are important components to that approach. We also note, however, that the supervised classifier-based approach to scoring retellings requires a significant amount of hand-labeled training data, thus rendering the technique impractical for application to a new narrative or to any picture description task. The importance of this distinction will become clear in Section 8, in which the approach outlined here is applied to a new data set lacking an existing scoring mechanism or a linguistic reference against which the responses can be scored.
In this article, we will be discussing the application of our methods to manually generated transcripts of retellings and picture descriptions produced by adults with and without neurodegenerative disorders. We note, however, that the same techniques have been applied to narratives transcribed using ASR output (Lehr et al. 2012, 2013) with little degradation in accuracy, given sufficient adaptation of the acoustic and language models to the WLM retelling domain. In addition, we have applied alignment-based scoring to the narratives of children with neurodevelopmental disorders, including autism and language impairment (Prud’hommeaux and Rouhizadeh 2012), with similarly strong diagnostic classification accuracy, further demonstrating the applicability of these methods to a variety of input formats, elicitation techniques, and diagnostic goals.","3 data : The participants for this study were drawn from an ongoing study of brain aging at the Layton Aging and Alzheimer’s Disease Center at the Oregon Health and Science University. Seventy-two of these participants had received a diagnosis of MCI, and 163 individuals served as typically aging controls. Demographic information about the experimental participants is shown in Table 2. There were no significant differences in age and years of education between the two groups. The Layton Center data included retellings for individuals who were not eligible for the present study because of their age or diagnosis. Transcriptions of 48 retellings produced by these ineligible participants were used to train and tune the word alignment model but were not used to evaluate the word alignment, scoring, or classification accuracy.
We diagnose MCI using the Clinical Dementia Rating (CDR) scale (Morris 1993), following earlier work on MCI (Petersen et al. 1999; Morris et al. 2001), as well as the work of Shankle et al. (2005) and Roark et al. (2011), who have previously attempted diagnostic classification using neuropsychological instrument subtest responses. The CDR is a numerical dementia staging scale that indicates the presence of dementia and its level of severity. The CDR score is derived from measures of cognitive function in six domains: Memory; Orientation; Judgment and Problem Solving; Community Affairs; Home and Hobbies; and Personal Care. These measures are determined during an extensive semi-structured interview with the patient and a close family member or caregiver. A CDR of 0 indicates the absence of dementia, and a CDR of 0.5 corresponds to a diagnosis of MCI (Ritchie and Touchon 2000). This measure has high expert interrater reliability (Morris 1993) and is assigned without any information derived from the WLM subtest. The WLM test, discussed in detail in Section 2.2, is a subtest of the Wechsler Memory Scale (Wechsler 1997), a neuropsychological instrument used to evaluate memory function in adults. Under standard administration of the WLM, the examiner reads a brief narrative to the participant, excerpts of which are shown in Figure 1. The participant then retells the narrative to the examiner twice: once immediately upon hearing the narrative and a second time after 20 to 30 minutes. Two retellings from one of the participants in our study are shown in Figures 2 and 3. (There are currently two narrative
retelling subtests that can be administered as part of the Wechsler Memory Scale, but the Anna Thompson narrative used in the present study is the more widely used and has appeared in every version of the Wechsler Memory Scale with only minor modifications since the instrument was first released 70 years ago.)
Following the published scoring guidelines, the examiner scores the participant’s response by counting how many of the 25 story elements are recalled in the retelling without regard to their ordering or relative importance in the story. We refer to this as the summary score. The boundaries between story elements are indicated with slashes in Figure 1. The retelling in Figure 2, produced by a participant without MCI, received a summary score of 12 for the 12 story elements recalled: Anna, Boston, employed, as a cook, and robbed of, she had four, small children, reported, station, touched by the woman’s story, took up a collection, and for her. The retelling in Figure 3, produced by the same participant after receiving a diagnosis of MCI two years later, earns a summary score of 5 for the 5 elements recalled: robbed, children, had not eaten, touched by the woman’s story, and took up a collection. Note that some of the story elements in these retellings were not recalled verbatim. The scoresheet provided with the exam indicates the lexical substitutions and degree of paraphrasing that are permitted, such as Ann or Annie for Anna, or any indication that the story evoked sympathy for touched by the woman’s story. Although the scoring guidelines have an air of arbitrariness in that paraphrasing is only sometimes permitted, they do allow the test to be scored with high inter-rater reliability (Mitchell 1987).
Recall that each participant produces two retellings for the WLM: an immediate retelling and a delayed retelling. Each participant’s two retellings were transcribed at the utterance level. The transcripts were downcased, and all pause-fillers, incomplete words, and punctuation were removed. The transcribed retellings were scored manually according to the published scoring guidelines, as described earlier in this section.","4 diagnostic classification framework : The goal of the work presented here is to demonstrate the utility of a variety of features derived from the WLM retellings for diagnostic classification of individuals with MCI. To perform this classification, we use LibSVM (Chang and Lin 2011), as implemented within the Waikato Environment for Knowledge Analysis (Weka) API (Hall et al. 2009), to train SVM classifiers, using a radial basis function kernel and default parameter settings.
We evaluate classification via receiver operating characteristic (ROC) curves, which have long been widely used to evaluate diagnostic tests (Zweig and Campbell 1993; Faraggi and Reiser 2002; Fan, Upadhye, and Worster 2006) and are also increasingly used in machine learning to evaluate classifiers in ranking scenarios (Cortes, Mohri, and Rastogi 2007; Ridgway et al. 2014). Analysis of ROC curves allows for classifier evaluation without selecting a specific, potentially arbitrary, operating point. To use standard clinical terminology, ROC curves track the tradeoff between sensitivity and specificity. Sensitivity (true positive rate) is what is commonly called recall in computational linguistics and related fields—that is, the percentage of items in the positive class that were correctly classified as positives. Specificity (true negative rate) is the percentage of items in the negative class that were correctly classified as negatives, which is equal to one minus the false positive rate. If the threshold is set so that nothing scores above threshold, the sensitivity (true positive rate, recall) is 0.0 and specificity (true negative rate) is 1.0. If the threshold is set so that everything scores above threshold, sensitivity is 1.0 and specificity is 0.0. As we sweep across intervening threshold settings, the ROC curve plots sensitivity versus one minus specificity, true positive rate versus false positive rate, providing insight into the precision/recall tradeoff at all possible operating points. Each point (tp, fp) in the curve has the true positive rate as the first dimension and false positive rate as the second dimension. Hence each curve starts at the origin (0, 0), the point corresponding to a threshold where nothing scores above threshold, and ends at (1, 1), the point where everything scores above threshold.
ROC curves can be characterized by the area underneath them (“area under curve” or AUC). A perfect classifier, with all positive items ranked above all negative items, has an ROC curve that starts at point (0, 0), goes straight up to (1, 0)—the point where true positive is 1.0 and false positive is 0.0 (since it is a perfect classifier)—before continuing straight over to the final point (1, 1). The area under this curve is 1.0, hence a perfect classifier has an AUC of 1.0. A random classifier, whose ROC curve is a straight diagonal line from the origin to (1, 1), has an AUC of 0.5. The AUC is equivalent to the probability that a randomly chosen positive example is ranked higher than a randomly chosen negative example, and is, in fact, equivalent to the Wilcoxon-Mann-Whitney statistic (Hanley and McNeil 1982). This statistic allows for classifier comparison without the need of pre-specifying arbitrary thresholds. For tasks like clinical screening, different tradeoffs between sensitivity and specificity may apply, depending on the scenario. See Fan, Upadhye, and Worster (2006) for a useful discussion of clinical use of ROC curves and the AUC score. In that paper, the authors note that there are multiple scales for interpreting the value of AUC, but that a rule-of-thumb is that AUC ≤ 0.75 is generally not clinically useful. For the present article, however, AUC mainly provides us the means for evaluating the relative quality of different classifiers.
One key issue for this sort of analysis is the estimation of the AUC for a particular classifier. Leave-pair-out cross-validation—proposed by Cortes, Mohri, and Rastogi
(2007) and extensively validated in Pahikkala et al. (2008) and Airolaa et al. (2011)— is a method for providing an unbiased estimate of the AUC, and the one we use in this article. In the leave-pair-out technique, every pairing between a negative example (i.e., a participant without MCI) and a positive example (i.e., a participant with MCI) is tested using a classifier trained on all of the remaining examples. The results of each positive/negative pair can be used to calculate the Wilcoxon-Mann-Whitney statistic as follows. Let s(e) be the score of some example e; let P be the set of positive examples and N the set of negative examples; and let [s(p) > s(n)] be 1 if true and 0 if false. Then:
AUC(s, P, n) = 1|P||N| ∑ p∈P ∑ n∈N [s(p) > s(n)] (1)
Although this method is compute-intensive, it does provide an unbiased estimate of the AUC, whereas other cross-validation setups lead to biased estimates. Another benefit of using the AUC is that standard deviation can be calculated. The standard deviation for the AUC is calculated as follows, where AUC is abbreviated as A to improve readability:
σ2a = A(1− A) + (|P| − 1)( A2−A − A 2) + (|N| − 1)( 2A21+A − A 2)
|P||N| (2) Previous work has shown that the WLM summary scores assigned during standard administration of the WLM, particularly in combination with other tests of verbal fluency and memory, are sensitive to the presence of MCI and other dementias (Storandt and Hill 1989; Petersen et al. 1999; Schmitt et al. 2000; Wang and Zhou 2002; Nordlund et al. 2005; Bennett et al. 2006; Price et al. 2009). We note, however, that the WLM test alone is not typically used as a diagnostic test. One of the goals of this work is to explore the utility of the standard WLM summary scores for diagnostic classification. A more ambitious goal is to demonstrate that using smaller units of information derived from story elements, rather than gross summary-level scores, can greatly improve diagnostic accuracy. Finally, we will show that using element-level scores automatically extracted from word alignments can achieve diagnostic classification accuracy comparable to that achieved using manually assigned scores. We therefore will compare the accuracy, measured in terms of AUC, of SVM classifiers trained on both summary-level and elementlevel WLM scores extracted from word alignments to the accuracy of classifiers built using a variety of alternative feature sets, both manually and automatically derived, shown in Table 3.
First, we consider the accuracy of classifiers using the expert-assigned WLM scores as features. For each of the 235 experimental participants, we generate two summary scores: one for the immediate retelling and one for the delayed retelling. The summary score ranges from 0, indicating that no elements were recalled, to 25, indicating that all elements were recalled. Previous work using manually assigned scores as features indicate that certain elements are more powerful in their ability to predict the presence of MCI (Prud’hommeaux 2012). In addition to the summary score, we therefore also provide the SVM with a vector of 50 story element-level scores: For each of the 25 elements in each of the two retellings per patient, there is a vector element with the value of 0 if the element was not recalled, or 1 if the element was recalled.
Classification accuracy with participants with MCI using these two manually derived feature sets is shown in Table 3.
We then present in Table 3 the classification accuracy of several summary-level features derived automatically from the WLM retellings, using standard NLP techniques for evaluating the similarity of two texts. We note that none of these features makes reference to the published WLM scoring guidelines or to the predefined element boundaries. Each of these feature sets contains two scores ranging between 0 and 1 for each participant, one for each of the two retellings: (1) cosine similarity between a retelling and the source narrative measured using LSA, proposed by Dunn et al. (2002) and calculated using the University of Colorado’s online LSA interface (available at http://lsa.colorado.edu/) with the 300-factor ninth-grade reading level topic space; (2) unigram overlap precision of a retelling relative to the source, proposed by HakkaniTur, Vergyri, and Tur (2010); (3) BLEU, the n-gram overlap metric commonly used to evaluate the quality of machine translation output (Papineni et al. 2002); and (4) the F-measure for ROUGE-SU4, the n-gram overlap metric commonly used to evaluate automatic summarization output (Lin 2004). The remaining two automatically derived features are a set of binary scores corresponding to the exact match via grep of each of the open-class unigrams in the source narrative and a summary score thereof.
Finally, in order to compare the WLM with another standard psychometric test, we also show the accuracy of a classifier trained only on the expert-assigned manual scores for the MMSE (Folstein, Folstein, and McHugh 1975), a clinician-administered 30- point questionnaire that measures a patient’s degree of cognitive impairment. Although it is widely used to screen for dementias such as Alzheimer’s disease, the MMSE is reported not to be particularly sensitive to MCI (Morris et al. 2001; Ravaglia et al. 2005; Hoops et al. 2009). The MMSE is entirely independent of the WLM and, though brief (5–10 minutes), requires more time to administer than the WLM.
In Table 3, we see that the WLM-based features yield higher accuracy than the MMSE, which is notable given the role that the MMSE plays in dementia screening. In addition, although all of the automatically derived feature sets yield higher classification than the MMSE, the manually derived WLM element-level scores are by far the most accurate feature set for diagnostic classification. Summary-level statistics, whether derived manually using established scoring mechanisms or automatically using a variety of text-similarity metrics used in the NLP community, seem not to provide sufficient power to distinguish the two diagnostic groups. In the next several sections, we describe a method for accurately automatically extracting the identities of the recalled story
elements from WLM retellings via word alignment in order to try to achieve classification accuracy comparable to that of the manually assigned WLM story elements and higher than that of the other automatic scoring methods.","5 wlm scoring via alignment :The approach presented here for automatic scoring of the WLM subtest relies on word alignments of the type used in machine translation for building phrased-based translation models. The motivation for using word alignment is the inherent similarity between narrative retelling and translation. In translation, a sentence in one language is converted into another language; the translation will have different words presented in a different order, but the meaning of the original sentence will be preserved. In narrative retelling, the source narrative is “translated” into the idiolect of the individual retelling the story. Again, the retelling will have different words, possibly presented in a different order, but at least some of the meaning will be preserved. We will show that although the algorithm for extracting scores from the alignments is simple, the process of getting high quality word alignments from the corpora of narrative retellings is challenging.
Although researchers in other NLP tasks that rely on alignments, such as textual entailment and summarization, sometimes eschew the sort of word-level alignments that are used in machine translation, we have no a priori reason to believe that this sort of alignment will be inadequate for the purposes of scoring narrative retellings. In addition, unlike many of the alignment algorithms proposed for tasks such as textual entailment, the methods for unsupervised word alignment used in machine translation require no external resources or hand-labeled data, making it simple to adapt our automated scoring techniques to new scenarios. We will show that the word alignment algorithms used in machine translation, when modified in particular ways, provide sufficient information for highly accurate scoring of narrative retellings and subsequent diagnostic classification of the individuals generating those retellings. Figure 4 shows a visual grid representation of a manually generated word alignment between the source narrative shown in Figure 1 on the vertical axis and the example WLM retelling in Figure 2 on the horizontal axis. Table 4 shows the word-index-toword-index alignment, in which the first index of each sentence is 0 and in which null alignments are not shown.
When creating these manual alignments, the labelers assigned the “possible” denotation under one of these two conditions: (1) when the alignment was ambiguous, as outlined in Och and Ney (2003); and (2) when a particular word in the retelling was a logical alignment to a word in the source narrative, but it would not have been counted as a permissible substitution under the published scoring guidelines. For this reason, we see that Taylor and sixty-seven are considered to be possible alignments because although they are logical alignments, they are not permissible substitutions according to the published scoring guidelines. Note that the word dollars is considered to be only a possible alignment, as well, since the element fifty-six dollars is not correctly recalled in this retelling under the standard scoring guidelines. In Figure 4, sure alignments are marked in black and possible alignments are marked in gray. In Figure 5, sure alignments are marked with S and possible alignments are marked with P.
Manually generated alignments like this one are the gold standard against which any automatically generated alignments can be compared to determine the accuracy
of the alignment. From an accurate word-to-word alignment, the identities of the story elements used in a retellings can be accurately extracted, and from that set of story elements, the score that is assigned under the standard scoring procedure can be calculated. As described earlier, the published scoring guidelines for the WLM specify the source words that compose each story element. Figure 5 displays the source narrative with the element IDs (A− Y) and word IDs (1− 65) explicitly labeled. Element Q, for instance, consists of the words 39 and 40, small children.
Using this information, we can determine which story elements were used in a retelling from the alignments as follows: for each word in the source narrative, if that word is aligned to a word in the retelling, the story element that it is associated with is considered to be recalled. For instance, if there is an alignment between the retelling word sympathetic and the source word touched, the story element touched by the woman’s story would be counted as correctly recalled. Note that in the WLM, every word in the source narrative is part of one of the story elements. Thus, when we convert alignments to scores in the way just described, any alignment can generate a story element. This is true even for an alignment between function words such as the and of, which would be unlikely individually to indicate that a story element had been recalled. To avoid such scoring errors, we disregard any word alignment pair containing a function word from
the source narrative. The two exceptions to this rule are the final two words, for her, which are not content words but together make a single story element.
Recall that in the manually derived word alignments, certain alignment pairs were marked as possible if the word in the retelling was logically equivalent to the word in the source but was not a permissible substitute according to the published scoring guidelines. When extracting scores from a manual alignment, only sure alignments are considered. This enables us to extract scores from a manual word alignment with 100% accuracy. The possible manual alignments are used only for calculating alignment error rate (AER) of an automatic word alignment model.
From the list of story elements extracted in this way, the summary score reported under standard scoring guidelines can be determined simply by counting the number of story elements extracted. Table 5 shows the story elements extracted from the manual word alignment in Table 4. The WLM immediate and delayed retellings for all of the 235 experimental participants and the 48 retellings from participants in the larger study who were not eligible for the present study were transcribed at the word level. Partial words, punctuation, and pause-fillers were excluded from all transcriptions used for this study. The retellings were manually scored according to published guidelines. In addition, we manually
produced word-level alignments between each retelling and the source narrative presented. These manual alignments were used to evaluate the word alignment quality and never to train the word alignment model.
Word alignment for phrase-based machine translation typically takes as input a sentence-aligned parallel corpus or bi-text, in which a sentence on one side of the corpus is a translation of the sentence in that same position on the other side of the corpus. Because we are interested in learning how to align words in the source narrative to words in the retellings, our primary parallel corpus must consist of source narrative text on one side and retelling text on the other. Because the retellings contain omissions, reorderings, and embellishments, we are obliged to consider the full text of the source narrative and of each retelling to be a “sentence” in the parallel corpus.
We compiled three parallel corpora to be used for the word alignment experiments:r Corpus 1: A 518-line source-to-retelling corpus consisting of the source narrative paired with each of the two retellings from the 235 experimental participants as well as the 48 retellings from ineligible individuals.r Corpus 2: A 268,324-line pairwise retelling-to-retelling corpus, consisting of every possible pairwise combination of the 518 available retellings.r Corpus 3: A 976-line word identity corpus, consisting of every word that appears in any retelling and the source narrative paired with itself.
The explicit parallel alignments of word identities that compose Corpus 3 are included in order to encourage the alignment of a word in a retelling to that same word in the source, if it exists.
The word alignment techniques that we use are unsupervised. Other than the transcriptions themselves, no manually generated data is used to build the word alignment models. Therefore, as in the case with most experiments involving word alignment, we build a model for the data we wish to evaluate using that same data. We do, however, use the 48 retellings from the individuals who were not experimental participants as a development set for tuning the various parameters of our word alignment system, which are described in the following. We begin by building two word alignment models using the Berkeley aligner (Liang, Taskar, and Klein 2006), a state-of-the-art word alignment package that relies on IBM Models 1 and 2 (Brown et al. 1993) and an HMM. We chose to use the Berkeley aligner, rather than the more widely used Giza++ alignment package, for this task because its joint training and posterior decoding algorithms yield lower alignment error rates on most data sets (including the data set used here [Prud’hommeaux and Roark 2011]) and because it offers functionality for testing an existing model on new data and, more crucially, for outputting posterior probabilities. The smaller of our two Berkeley-generated models is trained on Corpus 1 (the source-to-retelling parallel corpus described earlier) and ten copies of Corpus 3 (the word identity corpus). The larger model is trained on Corpus 1, Corpus 2 (the pairwise retelling corpus), and 100 copies of Corpus 3. Both models are then tested on the 470 retellings from our 235 experimental participants. In addition, we use both models to align every retelling to every other retelling so that we will have all pairwise alignments available for use in the graph-based model presented in the next section.
We note that the Berkeley aligner occasionally fails to return an alignment for a sentence pair, either because one of the sentences is too long or because the time required to perform the necessary calculations exceeds some maximum allotted time. In these cases, in order to generate alignments for all retellings and to build a complete graph that includes all retellings, we back off to the alignments and posteriors generated by IBM Model 1.
The first two rows of Table 6 show the precision, recall, and alignment error rate (AER) (Och and Ney 2003) for these two Berkeley aligner models. We note that although the AER for the larger model is lower, the time required to train the model is significantly longer. The alignments generated by the Berkeley aligner serve not only as a baseline for comparison of word alignment quality but also as a springboard for the novel graph-based method of alignment we will now discuss. Graph-based methods, in which paths or random walks are traced through an interconnected graph of nodes in order to learn more about the nodes themselves, have been used for NLP tasks in information extraction and retrieval, including Web-page ranking (PageRank; Page et al. 1999) and extractive summarization (LexRank; Erkan and Radev 2004; Otterbacher, Erkan, and Radev 2009). In the PageRank algorithm, the nodes of the graph are Web pages and the edges connecting the nodes are the hyperlinks leading from those pages to other pages. The nodes in the LexRank algorithm are sentences in a document and the edges are the similarity scores between those sentences. The number of times that a particular node is visited in a random walk reveals information about the importance of that node and its relationship to the other nodes. In many applications of random walks, the goal is to determine which node is the most central or has the highest prestige. In word alignment, however, the goal is to learn new relationships and strengthen existing relationships between words in a retelling and words in the source narrative.
In the case of our graph-based method for word alignment, each node represents a word in one of the retellings or in the source narrative. The edges are the normalized posterior-weighted alignments that the Berkeley aligner proposes between each word and (1) words in the source narrative, and (2) words in the other retellings. We generate these edges by using an existing baseline alignment model to align every retelling to every other retelling and to the source narrative. The posterior probabilities produced by the baseline alignment model serve as the weights on the edges. At each step in the walk, the choice of the next destination node can be determined according to
the strength of the outgoing edges, as measured by the posterior probability of that alignment.
Starting at a word in one of the retellings, represented by a node in the graph, the algorithm can walk from that node either to another retelling word in the graph to which it is aligned or to a word in the source narrative to which it is aligned. At each step in the walk, there is an empirically derived probability, λ, that sets the likelihood of transitioning to another retelling word versus a word in the source narrative. This probability functions similarly to the damping factor used in PageRank and LexRank, although its purpose is quite different. Once the decision whether to walk to a retelling word or source word has been made, the destination word itself is chosen according to the weights, which are the posterior probabilities assigned by the baseline alignment model. When the walk arrives at a source narrative word, that particular random walk ends, and the count for that source word as a possible alignment for the input retelling word is incremented by one.
For each word in each retelling, we perform 1,000 of these random walks, thereby generating a distribution for each retelling word over all of the words in the source narrative. The new alignment for the word is the source word with the highest frequency in that distribution. Pseudocode for this algorithm is provided in Figure 6.
Consider the following excerpts of five of the retellings. In each excerpt, the word that should align to the source word touched is rendered in bold:
the police were so moved by the story that they took up a collection for her
the fellow was sympathetic and made a collection for her so that she can feed the children
the police were touched by their story so they took up a collection
the police were so impressed with her story they took up a collection
the police felt sorry for her and took up a collection
Figure 7 presents a small idealized subgraph of the pairwise alignments of these five retellings. The arrows represent the alignments proposed by the Berkeley aligner between the relevant words in the retellings and their alignment (or lack of alignment) to the word touched in the source narrative. Thin arrows indicate alignment edges in the graph between retelling words, and bold arrows indicate alignment edges between retelling words and words in the source narrative. Words in the retellings are rendered as nodes with a single outline, and words in the source are rendered as nodes with a double outline.
We see that a number of these words were not aligned to the correct source word, touched. They are all, however, aligned to other retelling words that are in turn eventually aligned to the source word. Starting at any of the nodes in the graph, it is possible to walk from node to node and eventually reach the correct source word. Although sympathetic was not aligned to touched by the Berkeley aligner, its correct alignment can be recovered from the graph by following the path through other retelling words. After hundreds or thousands of random walks on the graph, evidence for the correct alignment will accumulate.
The approach as described might seem most beneficial to a system in need of improvements to recall rather than precision. Our baseline systems, however, are already favoring recall over precision. For this reason we include the NULL word in the list of words in the source narrative. We note that most implementations of both IBM Model 1 and HMM-based alignment also model the probability of aligning to a hidden word, NULL. In word alignment for machine translation, alignment to NULL usually indicates that a word in one language has no equivalent in the other language because the two languages express the same idea or construction in a slightly different way. Romance languages, for instance, often use prepositions before infinitival complements (e.g., Italian cerco di ridere) when English does not (e.g., I try to laugh). In the alignment of narrative retellings, however, alignment to NULL often indicates that the word in question is part of an aside or a piece of information that was not expressed in the source narrative.
Any retelling word that is not aligned to a source word by the baseline alignment system will implicitly be aligned to the hidden source word NULL, guaranteeing that every retelling word has at least one outgoing alignment edge and allowing us to
model the likelihood of being unaligned. A word that was unaligned by the original system can remain unaligned. A word that should have been left unaligned but was mistakenly aligned to a source word by the original system can recover its correct (lack of) alignment by following an edge to another retelling word that was correctly left unaligned (i.e., aligned to NULL). Figure 8 shows the graph in Figure 7 with the addition of the NULL node and the corresponding alignment edges to that node. This figure also includes two new retelling words, food and apple, and their respective alignment edges. Here we see that although the retelling word food was incorrectly aligned to the source word touched by the baseline system, its correct alignment to NULL can be recovered by traversing the edge to retelling word apple and from there, the edge to the source word NULL.
The optimal values for the following two parameters for the random walk must be determined: (1) the value of λ, the probability of walking to a retelling word node rather than a source word, and (2) the posterior probability threshold for including a particular edge in the graph. We optimize these parameters by testing the output of the graphbased approach on the development set of 48 retellings from the individuals who were not eligible for the study, discussed in Section 5.3. Recall that these additional retellings were included in the training data for the alignment model but were not included in the test set used to evaluate its performance. Tuning on this set of retellings therefore introduces no additional words, out-of-vocabulary words, or other information to the graph, while preventing overfitting.
The posterior threshold is set to 0.5 in the Berkeley aligner’s default configuration, and we found that this value did indeed yield the lowest AER for the Berkeley aligner on our data. When building the graph using Berkeley alignments and posteriors, however, we can adjust the value of this threshold to optimize the AER of the alignments produced via random walks. Using the development set of 48 retellings, we determined that the AER is minimized when the value of λ is 0.8 and the alignment inclusion posterior threshold is 0.5. Recall the two baseline alignment models generated by the Berkeley aligner, described in Section 5.4: (1) the small Berkeley model, trained on Corpus 1 (the source-to-retelling corpus) and 10 instances of Corpus 3 (the word identity corpus), and (2) the large Berkeley model (trained on Corpus 1, Corpus 2, the full pairwise retelling-to-retelling corpus, and 100 instances of Corpus 3). Using these models, we generate full retellingto-retelling alignments, on which we can then build two graph-based alignment models: the small graph-based model and the large graph-based model.
The manual gold alignments for the 235 experimental participants were evaluated against the alignments produced by each of the four models. Table 6 presents the precision, recall, and AER for the alignments of the experimental participants. Not surprisingly, the larger models yield lower error rates than the smaller models. More interestingly, each graph-based model outperforms the Berkeley model of the corresponding size by a large margin. The performance of the small graph-based model is particularly remarkable because it yields an AER superior to the large Berkeley model while requiring significantly fewer computing resources. Each of the graphbased models generated the full set of alignments in only a few minutes, whereas the large Berkeley model required 14 hours of training.","6 scoring evaluation :The element-level scores induced, as described in Section 5.2, from the four word alignments for all 235 experimental participants were evaluated against the manual per-element scores. We report the precision, recall, and F-measure for all four alignment models in Table 7. In addition, we report Cohen’s kappa as a measure of reliability between our automated scores and the manually assigned scores. We see that as AER improves, scoring accuracy also improves, with the large graph-based model outperforming all other models in terms of precision, F-measure, and inter-rater reliability. The scoring accuracy levels reported here are comparable to the levels of inter-rater agreement typically reported for the WLM, and reliability between our automated scores and the manual scores, as measured by Cohen’s kappa, is well within the ranges reported in the literature (Johnson, Storandt, and Balota 2003). As will be shown in the following section, scoring accuracy is important for achieving high classification accuracy of MCI.","7 diagnostic classification :As discussed in Section 2, poor performance on the WLM test is associated with MCI. We now use the scores we have extracted from the word alignments as features with an SVM to perform diagnostic classification for distinguishing participants with MCI from those without, as described in Section 4.1.
Table 8 shows the classification results for the scores derived from the four alignment models along with the classification results using the examiner-assigned manual scores, the MMSE, and the four alternative automated scoring approaches described in Section 4.2. It appears that, in all cases, the per-element scores are more effective than the summary scores in classifying the two diagnostic groups. In addition, we see that our automated scores have classificatory power comparable to that of the manual gold scores, and that as scoring accuracy increases from the small Berkeley model to the graph-based models and bigger models, classification accuracy improves. This suggests both that accurate scores are crucial for accurate classification and that pursuing even further improvements in word alignment is likely to result in improved diagnostic differentiation. We note that although the large Berkeley model achieved the highest classification accuracy of the automated methods, this very slight margin of difference may not justify its significantly greater computational requirements.
In addition to using summary scores and element-level scores as features for the story-element based models, we also perform feature selection over both sets of features using the chi-square statistic. Feature selection is performed separately on each training set for each fold in the cross-validation to avoid introducing bias from the testing example. We train and test the SVM using the top n story element features, from n = 1 to n = 50. We report here the accuracy for the top seven story elements (n = 7), which yielded the highest AUC measure. We note that over all of the folds, only 8 of the 50 features ever appeared among the seven most informative.
In all cases, the per-element scores are more effective than the summary scores in classifying the two diagnostic groups, and performing feature selection results in improved classification accuracy. All of the element-level feature sets automatically extracted from alignments outperform the MMSE and all of the alternative automatic
scoring procedures, which suggests that the extra complexity required to extract element-level features is well worth the time and effort.
We note that the final classification results for all four alignment models are not drastically different from one another, despite the large reductions in word alignment error rate and improvements in scoring accuracy observed in the larger models and graph-based models. This seeming disconnect between word alignment accuracy and downstream application performance has also been observed in the machine translation literature, where reductions in AER do not necessarily lead to meaningful increases in BLEU, the widely accepted measure of machine translation quality (Ayan and Dorr 2006; Lopez and Resnik 2006; Fraser and Marcu 2007). Our results, however, show that a feature set consisting of manually assigned WLM scores yields the highest classification accuracy of any of the feature sets evaluated here. As discussed in Section 5.2, our WLM score extraction method is designed such that element-level scores can be extracted with perfect accuracy from a perfect word alignment. Thus, the goal of seeking perfect or near-perfect word alignment accuracy is worthwhile because it will necessarily result in perfect or near-perfect scoring accuracy, which in turn is likely to yield classification accuracy approaching that of manually assigned scores.","8 application to task with non-linguistic reference :As we discussed earlier, one of the advantages of using an unsupervised method of scoring is the resulting generalizability to new data sets, particularly those generated from a non-linguistic stimulus. The Boston Diagnostic Aphasia Examination (BDAE) (Goodglass and Kaplan 1972), an instrument widely used to diagnose aphasia in adults, includes one such task, popularly known as the cookie theft picture description task. In this test, the person views a drawing of a lively scene in a family’s kitchen and must tell the examiner about all of the actions they see in the picture. The picture is reproduced below in Figure 9.
Describing visually presented material is quite different from a task such as the WLM, in which language comprehension and memory play a crucial role. Nevertheless, the processing and language production demands of a picture description task may lead to differences in performance in groups with certain cognitive and language problems. In fact, it is widely reported that the picture descriptions of seniors with dementia of the Alzheimer’s type differ from those of typically aging seniors in terms of information content (Hier, Hagenlocker, and Shindler 1985; Gilesa, Patterson, and Hodge 1996). Interestingly, this reduction in information is not necessarily accompanied by a reduction in the amount of language produced. Rather, it seems that seniors with Alzheimer’s dementia tend to include redundant information, repetitions, intrusions, and revisions that result in language samples of length comparable to that of typically aging seniors.
TalkBank (MacWhinney 2007), the online database of audio and transcribed speech, has made available the DementiaBank corpus of descriptions of the cookie theft picture by hundreds of individuals, some of whom have one of a number of types of dementia, including MCI, vascular dementia, possible Alzheimer’s disease, and probable Alzheimer’s disease. From this corpus were selected a subset of individuals without dementia and a subset with probable Alzheimer’s disease. We limit the set of descriptions to those with more than 25 but fewer than 100 words, yielding 130 descriptions for each diagnostic group. There was no significant difference in description word count between the two diagnostic groups.
The first task was to generate a source description to which all other narratives should be aligned. Working under the assumption that the control participants would
produce good descriptions, we calculated the BLEU score of every pair of descriptions from the control group. The description with the highest average pairwise BLEU score was selected as the source description. After confirming that this description did in fact contain all of the action portrayed in the picture, we removed all extraneous conversational asides from the description in order to ensure that it contained all and only information about the picture. The selected source description is as follows:
The boy is getting cookies out of the cookie jar. And the stool is just about to fall over. The little girl is reaching up for a cookie. And the mother is drying dishes. The water is running into the sink and the sink is running over onto the floor. And that little girl is laughing.
We then built an alignment model on the full pairwise description parallel corpus (2602 = 67,600 sentences) and a word identity corpus consisting of each word in each description reproduced 100 times. Using this trained model, which corresponds to the large Berkeley model that achieved the highest classification accuracy for the WLM data, we then aligned every description to the artificial source description. We also built a graph-based alignment model using these alignments and the parameter settings that maximized word alignment accuracy in the WLM data. Because the artificial source description is not a true linguistic reference for this task, we did not produce manual word alignments against which the alignment quality could be evaluated and against which the parameters could be tuned. Instead, we evaluated only the downstream application of diagnostic classification.
The method for scoring the WLM relies directly on the predetermined list of story elements, whereas the cookie theft picture description administration instructions do not include an explicit set of items that must be described. Recall that the automated scoring method we propose uses only the open-class or content words in the source narrative. In order to generate scores for the descriptions, we propose a scoring technique that considers each content word in the source description to be its own story element. Any word in a retelling that aligns to one of the content words in the source narrative is considered to be a match for that content word element. This results in a large number of elements, but it allows the scoring method to be easily adapted to other
narrative production scenarios that similarly do not have explicit scoring guidelines. Using these scores as features, we again used an SVM to classify the two diagnostic groups, typically aging and probable Alzheimer’s disease, and evaluated the classifier using leave-pair-out validation.
Table 9 shows the classification results using the content word scoring features produced using the Berkeley aligner alignments and the graph-based alignments. These can be compared to classification results using the summary similarity metrics BLEU and unigram precision. We see that using word-level features, regardless of which alignment model they are extracted from, results in significantly higher classification accuracy than both the simple similarity metrics and the summary scores. The alignmentbased scoring approach yields features with remarkably high classification accuracy given the somewhat ad hoc selection of the source narrative from the set of control retellings.
These results demonstrate the flexibility and utility of the alignment-based approach to scoring narratives. Not only can it be adapted to other narrative retelling instruments, but it can relatively trivially be adapted to instruments that use nonlinguistic stimuli for elicitation. All that is needed to build an alignment model is a sufficiently large collection of retellings of the same narrative or descriptions of the same picture. Procuring such a collection of descriptions or retellings can be done easily outside a clinical setting using a platform such as Amazon’s Mechanical Turk. No handlabeled data, outside lexical resources, prior knowledge of the content of the story, or existing scoring guidelines are required.","Among the more recent applications for natural language processing algorithms has been the analysis of spoken language data for diagnostic and remedial purposes, fueled by the demand for simple, objective, and unobtrusive screening tools for neurological disorders such as dementia. The automated analysis of narrative retellings in particular shows potential as a component of such a screening tool since the ability to produce accurate and meaningful narratives is noticeably impaired in individuals with dementia and its frequent precursor, mild cognitive impairment, as well as other neurodegenerative and neurodevelopmental disorders. In this article, we present a method for extracting narrative recall scores automatically and highly accurately from a word-level alignment between a retelling and the source narrative. We propose improvements to existing machine translation-based systems for word alignment, including a novel method of word alignment relying on random walks on a graph that achieves alignment accuracy superior to that of standard expectation maximization-based techniques for word alignment in a fraction of the time required for expectation maximization. In addition, the narrative recall score features extracted from these high-quality word alignments yield diagnostic classification accuracy comparable to that achieved using manually assigned scores and significantly higher than that achieved with summary-level text similarity metrics used in other areas of NLP. These methods can be trivially adapted to spontaneous language samples elicited with non-linguistic stimuli, thereby demonstrating the flexibility and generalizability of these methods.","[{""affiliations"": [], ""name"": ""Emily Prud\u2019hommeaux""}, {""affiliations"": [], ""name"": ""Brian Roark""}]",SP:9fff4c3673ff8364cd50b46aa7b757952eefa5de,"[{""authors"": [""Airolaa"", ""Antti"", ""Tapio Pahikkalaa"", ""Willem Waegemanc"", ""Bernard De Baetsc"", ""Tapio Salakoskia""], ""title"": ""An experimental comparison of cross-validation techniques for estimating the area under the ROC"", ""year"": 2011}, {""authors"": [""Artero"", ""Sylvain"", ""Mary Tierney"", ""Jacques Touchon"", ""Karen Ritchie""], ""title"": ""Prediction of transition from cognitive"", ""year"": 2003}, {""authors"": [""Ayan"", ""Necip Fazil"", ""Bonnie J. Dorr.""], ""title"": ""Going beyond AER: An extensive analysis of word alignments and their impact on MT"", ""venue"": ""Proceedings of the 21st International Conference on Computational Linguistics and"", ""year"": 2006}, {""authors"": [""D.A. Bennett"", ""J.A. Schneider"", ""Z. Arvanitakis"", ""J.F. Kelly"", ""N.T. Aggarwal"", ""R.C. Shah"", ""R.S. Wilson""], ""title"": ""Neuropathology of older persons without cognitive impairment"", ""year"": 2006}, {""authors"": [""Bishop"", ""Dorothy"", ""Chris Donlan.""], ""title"": ""The role of syntax in encoding and recall of pictorial narratives: Evidence from specific language impairment"", ""venue"": ""British Journal of Developmental Psychology,"", ""year"": 2005}, {""authors"": [""Brown"", ""Peter"", ""Vincent Della Pietra"", ""Steven Della Pietra"", ""Robert Mercer.""], ""title"": ""The mathematics of statistical machine translation: Parameter estimation"", ""venue"": ""Computational Linguistics,"", ""year"": 1993}, {""authors"": [""Chang"", ""Chih-Chung"", ""Chih-Jen Lin.""], ""title"": ""LIBSVM: A library for support vector machines"", ""venue"": ""ACM Transactions on Intelligent Systems and Technology, 2(27):1\u201327."", ""year"": 2011}, {""authors"": [""Chapman"", ""Sandra"", ""Hanna Ulatowska"", ""Kristin King"", ""Julene Johnson"", ""Donald McIntire.""], ""title"": ""Discourse in early Alzheimer\u2019s disease versus normal advanced aging"", ""venue"": ""American Journal of"", ""year"": 1995}, {""authors"": [""Chenery"", ""Helen J."", ""Bruce E. Murdoch.""], ""title"": ""The production of narrative discourse in response to animations in persons with dementia of the Alzheimer\u2019s type: Preliminary findings"", ""venue"": ""Aphasiology,"", ""year"": 1994}, {""authors"": [""Cortes"", ""Corinna"", ""Mehryar Mohri"", ""Ashish Rastogi.""], ""title"": ""An alternative ranking problem for search engines"", ""venue"": ""Proceedings of the 6th Workshop on Experimental Algorithms, volume 4525 of Lecture Notes in"", ""year"": 2007}, {""authors"": [""Creamer"", ""Scott"", ""Maureen Schmitter-Edgecombe""], ""title"": ""Narrative comprehension in Alzheimer\u2019s disease: Assessing inferences and memory operations with a think-aloud procedure"", ""year"": 2010}, {""authors"": [""de la Rosa"", ""Gabriela Ramirez"", ""Thamar Solorio"", ""Manuel Montes y Gomez"", ""Aquiles Iglesias"", ""Yang Liu"", ""Lisa Bedore"", ""Elizabeth Pena""], ""title"": ""Exploring word class n-grams to measure language"", ""year"": 2013}, {""authors"": [""Diehl"", ""Joshua J."", ""Loisa Bennetto"", ""Edna Carter Young.""], ""title"": ""Story recall and narrative coherence of high-functioning children with autism spectrum disorders"", ""venue"": ""Journal of Abnormal Child Psychology,"", ""year"": 2006}, {""authors"": [""Dunn"", ""John C."", ""Osvaldo P. Almeida"", ""Lee Barclay"", ""Anna Waterreus"", ""Leon Flicker.""], ""title"": ""Latent semantic analysis: A new method to measure prose recall"", ""venue"": ""Journal of Clinical and Experimental Neuropsychology,"", ""year"": 2002}, {""authors"": [""Ehrlich"", ""Jonathan S."", ""Loraine K. Obler"", ""Lynne Clark.""], ""title"": ""Ideational and semantic contributions to narrative production in adults with dementia of the Alzheimer\u2019s type"", ""venue"": ""Journal of Communication Disorders,"", ""year"": 1997}, {""authors"": [""Erkan"", ""G\u00fcnes"", ""Dragomir R. Radev""], ""title"": ""LexRank: Graph-based lexical centrality as salience in text summarization"", ""year"": 2004}, {""authors"": [""Fan"", ""Jerome"", ""Suneel Upadhye"", ""Andrew Worster.""], ""title"": ""Understanding receiver operating characteristic (ROC) curves"", ""venue"": ""Canadian Journal of Emergency Medicine, 8:19\u201320."", ""year"": 2006}, {""authors"": [""Faraggi"", ""David"", ""Benjamin Reiser.""], ""title"": ""Estimation of the area under the ROC curve"", ""venue"": ""Statistics in Medicine, 21:3093\u20133106."", ""year"": 2002}, {""authors"": [""M. Folstein"", ""S. Folstein"", ""P. McHugh.""], ""title"": ""Mini-mental state\u2014a practical method for grading the cognitive state of patients for the clinician"", ""venue"": ""Journal of Psychiatric Research, 12:189\u2013198."", ""year"": 1975}, {""authors"": [""Fraser"", ""Alexander"", ""Daniel Marcu.""], ""title"": ""Measuring word alignment quality for statistical machine translation"", ""venue"": ""Computational Linguistics, 33(3):293\u2013303."", ""year"": 2007}, {""authors"": [""Fraser"", ""Kathleen C"", ""Jed A. Meltzer"", ""Naida L. Graham"", ""Carol Leonard"", ""Graeme Hirst"", ""Sandra E. Black"", ""Elizabeth Rochon""], ""title"": ""Automated classification of primary progressive aphasia subtypes"", ""year"": 2014}, {""authors"": [""Gabani"", ""Keyur"", ""Melissa Sherman"", ""Thamar Solorio"", ""Yang Liu""], ""title"": ""A corpus-based approach for the prediction of language impairment in monolingual English and Spanish-English bilingual"", ""year"": 2009}, {""authors"": [""Galvin"", ""James"", ""Anne Fagan"", ""David Holtzman"", ""Mark Mintun"", ""John Morris.""], ""title"": ""Relationship of dementia screening tests with biomarkers of Alzheimer\u2019s disease"", ""venue"": ""Brain, 133:3290\u20133300."", ""year"": 2010}, {""authors"": [""Gilesa"", ""Elaine"", ""Karalyn Patterson"", ""John R. Hodge""], ""title"": ""Performance on the Boston cookie theft picture description task in patients with early dementia of the Alzheimer\u2019s type: Missing information"", ""year"": 1996}, {""authors"": [""H. Goodglass"", ""E. Kaplan.""], ""title"": ""Boston Diagnostic Aphasia Examination"", ""venue"": ""Lea and Febiger, Philadelphia, PA."", ""year"": 1972}, {""authors"": [""Hakkani-Tur"", ""Dilek"", ""Dimitra Vergyri"", ""Gokhan Tur.""], ""title"": ""Speech-based automated cognitive status assessment"", ""venue"": ""Proceedings of the Conference of the International Speech Communication"", ""year"": 2010}, {""authors"": [""Hall"", ""Mark"", ""Eibe Frank"", ""Geoffrey Holmes"", ""Bernhard Pfahringer"", ""Peter Reutemann"", ""Ian H. Witten.""], ""title"": ""The WEKA data mining software: An update"", ""venue"": ""SIGKDD Explorations, 11(1):10\u201318."", ""year"": 2009}, {""authors"": [""Hanley"", ""James"", ""Barbara McNeil.""], ""title"": ""The meaning and use of the area under a receiver operating characteristic (ROC) curve"", ""venue"": ""Radiology, 143:29\u201336."", ""year"": 1982}, {""authors"": [""D. Hier"", ""K. Hagenlocker"", ""A. Shindler.""], ""title"": ""Language disintegration in dementia: Effects of etiology and severity"", ""venue"": ""Brain and Language, 25:117\u2013133."", ""year"": 1985}, {""authors"": [""S. Hoops"", ""S. Nazem"", ""A.D. Siderowf"", ""J.E. Duda"", ""S.X. Xie"", ""M.B. Stern"", ""D. Weintraub.""], ""title"": ""Validity of the MoCA and MMSE in the detection of MCI and dementia in Parkinson disease"", ""venue"": ""Neurology,"", ""year"": 2009}, {""authors"": [""Johnson"", ""David K."", ""Martha Storandt"", ""David A. Balota.""], ""title"": ""Discourse analysis of logical memory recall in normal aging and in dementia of the Alzheimer type"", ""venue"": ""Neuropsychology, 17(1):82\u201392."", ""year"": 2003}, {""authors"": [""Lautenschlager"", ""Nicola T"", ""John C. Dunn"", ""Kathryn Bonney"", ""Leon Flicker"", ""Osvaldo P. Almeida""], ""title"": ""Latent semantic analysis: An improved method to measure cognitive performance in subjects"", ""year"": 2006}, {""authors"": [""Lehr"", ""Maider"", ""Emily Prud\u2019hommeaux"", ""Izhak Shafran"", ""Brian Roark""], ""title"": ""Fully automated neuropsychological assessment for detecting mild cognitive impairment"", ""venue"": ""In Proceedings of the 13th Annual Conference"", ""year"": 2012}, {""authors"": [""Lehr"", ""Maider"", ""Izhak Shafran"", ""Emily Prud\u2019hommeaux"", ""Brian Roark""], ""title"": ""Discriminative joint modeling of lexical variation and acoustic confusion for automated narrative retelling"", ""year"": 2013}, {""authors"": [""Liang"", ""Percy"", ""Ben Taskar"", ""Dan Klein.""], ""title"": ""Alignment by agreement"", ""venue"": ""Proceedings of the Human Language Technology Conference of the NAACL, pages 104\u2013111, New York, NY."", ""year"": 2006}, {""authors"": [""Lin"", ""Chin-Yiu.""], ""title"": ""ROUGE: A package for automatic evaluation of summaries"", ""venue"": ""Proceedings of the Workshop on Text Summarization Branches Out, pages 74\u201381, Barcelona."", ""year"": 2004}, {""authors"": [""Lopez"", ""Adam"", ""Philip Resnik.""], ""title"": ""Word-based alignment, phrase-based translation: What\u2019s the link"", ""venue"": ""Proceedings"", ""year"": 2006}, {""authors"": [""Lysaker"", ""Paul"", ""Amanda Wickett"", ""Neil Wilke"", ""John Lysaker.""], ""title"": ""Narrative incoherence in schizophrenia: The absent agent-protagonist and the collapse of internal dialogue"", ""venue"": ""American Journal of"", ""year"": 2003}, {""authors"": [""MacWhinney"", ""Brian.""], ""title"": ""The TalkBank Project"", ""venue"": ""J. C. Beal, K. P. Corrigan, and H. L. Moisl, editors, Creating and Digitizing Language Corpora: Synchronic Databases, Vol.1, pages 163\u2013180, Palgrave-Macmillan,"", ""year"": 2007}, {""authors"": [""Manly"", ""Jennifer J."", ""Ming Tang"", ""Nicole Schupf"", ""Yaakov Stern"", ""Jean-Paul G. Vonsattel"", ""Richard Mayeux.""], ""title"": ""Frequency and course of mild cognitive impairment in a multiethnic community"", ""venue"": ""Annals of"", ""year"": 2008}, {""authors"": [""Mitchell"", ""Margaret.""], ""title"": ""Scoring discrepancies on two subtests of the Wechsler memory scale"", ""venue"": ""Journal of Consulting and Clinical Psychology, 55:914\u2013915."", ""year"": 1987}, {""authors"": [""Morris"", ""John.""], ""title"": ""The Clinical Dementia Rating (CDR): Current version and scoring rules"", ""venue"": ""Neurology, 43:2412\u20132414."", ""year"": 1993}, {""authors"": [""Morris"", ""John"", ""Martha Storandt"", ""J. Phillip Miller"", ""Daniel McKeel"", ""Joseph Price"", ""Eugene Rubin"", ""Leonard Berg.""], ""title"": ""Mild cognitive impairment represents early-stage Alzheimer disease"", ""venue"": ""Archives of"", ""year"": 2001}, {""authors"": [""Norbury"", ""Courtenay"", ""Dorothy Bishop.""], ""title"": ""Narrative skills of children with communication impairments"", ""venue"": ""International Journal of Language and Communication Disorders, 38:287\u2013313."", ""year"": 2003}, {""authors"": [""A. Nordlund"", ""S. Rolstad"", ""P. Hellstrom"", ""M. Sjogren"", ""S. Hansen"", ""A. Wallin.""], ""title"": ""The Goteborg MCI study: Mild cognitive impairment is a heterogeneous condition"", ""venue"": ""Journal of Neurology, Neurosurgery and"", ""year"": 2005}, {""authors"": [""Och"", ""Franz Josef"", ""Hermann Ney.""], ""title"": ""A systematic comparison of various statistical alignment models"", ""venue"": ""Computational Linguistics, 29(1):19\u201351."", ""year"": 2003}, {""authors"": [""Otterbacher"", ""Jahna"", ""G\u00fcnes Erkan"", ""Dragomir R. Radev.""], ""title"": ""Biased LexRank: Passage retrieval using random walks with question-based priors"", ""venue"": ""Information Processing Management,"", ""year"": 2009}, {""authors"": [""Page"", ""Lawrence"", ""Sergey Brin"", ""Rajeev Motwani"", ""Terry Winograd""], ""title"": ""The PageRank citation ranking: Bringing order"", ""year"": 1999}, {""authors"": [""Prud\u2019hommeaux"", ""Roark""], ""title"": ""Graph-Based Word Alignment for Clinical Language Evaluation to the web"", ""venue"": ""Technical Report 1999-66,"", ""year"": 1999}, {""authors"": [""Pahikkala"", ""Tapio"", ""Antti Airola"", ""Jorma Boberg"", ""Tapio Salakoski.""], ""title"": ""Exact and efficient leave-pair-out cross-validation for ranking RLS"", ""venue"": ""The Second International and Interdisciplinary Conference on Adaptive"", ""year"": 2008}, {""authors"": [""Papineni"", ""Kishore"", ""Salim Roukos"", ""Todd Ward"", ""Wei jing Zhu.""], ""title"": ""BLEU: A method for automatic evaluation of machine translation"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for"", ""year"": 2002}, {""authors"": [""Petersen"", ""Ronald"", ""Glenn Smith"", ""Stephen Waring"", ""Robert Ivnik"", ""Eric Tangalos"", ""Emre Kokmen.""], ""title"": ""Mild cognitive impairment: Clinical characterizations and outcomes"", ""venue"": ""Archives of Neurology,"", ""year"": 1999}, {""authors"": [""Petersen"", ""Ronald C.""], ""title"": ""Mild cognitive impairment"", ""venue"": ""The New England Journal of Medicine, 364(23):2227\u20132234."", ""year"": 2011}, {""authors"": [""John J. McArdle"", ""Robert J. Willis"", ""Robert B. Wallace.""], ""title"": ""Prevalence of cognitive impairment without dementia in the United States"", ""venue"": ""Annals of Internal Medicine, 148:427\u201334."", ""year"": 2008}, {""authors"": [""John C. Morris.""], ""title"": ""Neuropathology of nondemented aging: Presumptive evidence for preclinical Alzheimer disease"", ""venue"": ""Neurobiology of Aging, 30(7):1026\u20131036."", ""year"": 2009}, {""authors"": [""Prud\u2019hommeaux"", ""Emily"", ""Brian Roark""], ""title"": ""Alignment of spoken narratives for automated neuropsychological assessment"", ""venue"": ""In Proceedings of the IEEE Workshop on Automatic Speech Recognition"", ""year"": 2011}, {""authors"": [""Prud\u2019hommeaux"", ""Emily"", ""Brian Roark""], ""title"": ""Graph-based alignment of narratives"", ""year"": 2012}, {""authors"": [""Prud\u2019hommeaux"", ""Emily"", ""Masoud Rouhizadeh""], ""title"": ""Automatic detection of pragmatic deficits in children with autism"", ""venue"": ""In Proceedings of the 3rd Workshop on Child, Computer and Interaction,"", ""year"": 2012}, {""authors"": [""Prud\u2019hommeaux"", ""Emily Tucker""], ""title"": ""Alignment of Narrative Retellings for Automated Neuropsychological Assessment"", ""venue"": ""Ph.D. thesis,"", ""year"": 2012}, {""authors"": [""Ravaglia"", ""Giovanni"", ""Paola Forti"", ""Fabiola Maioli"", ""Lucia Servadei"", ""Mabel Martelli"", ""Nicoletta Brunetti"", ""Luciana Bastagli"", ""Erminia Mariani""], ""title"": ""Screening for mild cognitive impairment in elderly"", ""year"": 2005}, {""authors"": [""Ridgway"", ""James"", ""Pierre Alquier"", ""Nicolas Chopin"", ""Feng Liang.""], ""title"": ""PAC-Bayesian AUC classification and scoring"", ""venue"": ""Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and"", ""year"": 2014}, {""authors"": [""Ritchie"", ""Karen"", ""Sylvaine Artero"", ""Jacques Touchon.""], ""title"": ""Classification criteria for mild cognitive impairment: A population-based validation study"", ""venue"": ""Neurology, 56:37\u201342."", ""year"": 2001}, {""authors"": [""Ritchie"", ""Karen"", ""Jacques Touchon.""], ""title"": ""Mild cognitive impairment: Conceptual basis and current nosological status"", ""venue"": ""Lancet, 355:225\u2013228."", ""year"": 2000}, {""authors"": [""Roark"", ""Brian"", ""Margaret Mitchell"", ""John-Paul Hosom"", ""Kristina Hollingshead"", ""Jeffrey Kaye.""], ""title"": ""Spoken language derived measures for detecting mild cognitive impairment"", ""venue"": ""IEEE Transactions on Audio,"", ""year"": 2011}, {""authors"": [""F.A. Schmitt"", ""D.G. Davis"", ""D.R. Wekstein"", ""C.D. Smith"", ""J.W. Ashford"", ""W.R. Markesbery.""], ""title"": ""Preclinical AD revisited: Neuropathology of cognitively normal older adults"", ""venue"": ""Neurology,"", ""year"": 2000}, {""authors"": [""Shankle"", ""William R"", ""A. Kimball Romney"", ""Junko Hara"", ""Dennis Fortier"", ""Malcolm B. Dick"", ""James M. Chen"", ""Timothy Chan"", ""Xijiang Sun""], ""title"": ""Methods to improve the detection of mild cognitive"", ""year"": 2005}, {""authors"": [""Solorio"", ""Thamar"", ""Yang Liu.""], ""title"": ""Using language models to identify language impairment in Spanish-English bilingual children"", ""venue"": ""Proceedings of the ACL 2008 Workshop on Biomedical Natural Language"", ""year"": 2008}, {""authors"": [""Storandt"", ""Martha"", ""Robert Hill.""], ""title"": ""Very mild senile dementia of the Alzheimer\u2019s type: II"", ""venue"": ""Psychometric test performance. Archives of Neurology, 46:383\u2013386."", ""year"": 1989}, {""authors"": [""Tager-Flusberg"", ""Helen.""], ""title"": ""Once upon a ribbit: Stories narrated by autistic children"", ""venue"": ""British Journal of Developmental Psychology, 13(1):45\u201359."", ""year"": 1995}, {""authors"": [""Tannock"", ""Rosemary"", ""Karen L. Purvis"", ""Russell J. Schachar.""], ""title"": ""Narrative abilities in children with attention deficit hyperactivity disorder and normal peers"", ""venue"": ""Journal of Abnormal Child Psychology,"", ""year"": 1993}, {""authors"": [""Tierney"", ""Mary"", ""Christie Yao"", ""Alex Kiss"", ""Ian McDowell.""], ""title"": ""Neuropsychological tests accurately predict incident Alzheimer disease after 5 and 10 years"", ""venue"": ""Neurology, 64:1853\u20131859."", ""year"": 2005}, {""authors"": [""Ulatowska"", ""Hanna"", ""Lee Allard"", ""Adrienne Donnell"", ""Jean Bristow"", ""Sara M. Haynes"", ""Adelaide Flower"", ""Alvin J. North""], ""title"": ""Discourse performance in subjects with dementia"", ""year"": 1988}, {""authors"": [""United Nations.""], ""title"": ""World Population Ageing 1950\u20132050"", ""venue"": ""United Nations, New York."", ""year"": 2002}, {""authors"": [""Vuorinen"", ""Elina"", ""Matti Laine"", ""Juha Rinne.""], ""title"": ""Common pattern of language impairment in vascular dementia and in Alzheimer disease"", ""venue"": ""Alzheimer Disease and Associated Disorders, 14(2):81\u201386."", ""year"": 2000}, {""authors"": [""Wang"", ""Qing-Song"", ""Jiang-Ning Zhou.""], ""title"": ""Retrieval and encoding of episodic memory in normal aging and patients with mild cognitive impairment"", ""venue"": ""Brain Research, 924:113\u2013115."", ""year"": 2002}, {""authors"": [""Wechsler"", ""David.""], ""title"": ""Wechsler Memory Scale - Third Edition"", ""venue"": ""The Psychological Corporation, San Antonio, TX."", ""year"": 1997}, {""authors"": [""Zweig"", ""Mark H."", ""Gregory Campbell.""], ""title"": ""Receiver-operating characteristic (ROC) plots: A fundamental evaluation tool in clinical medicine"", ""venue"": ""Clinical Chemistry, 39:561\u2013577."", ""year"": 1993}]","acknowledgments  :This research was conducted while both authors were at the Center for Spoken Language Understanding at the Oregon Health and Science University, in Portland, Oregon. This work was supported in part by NSF grant BCS-0826654 and NIH NIDCD grants R01DC012033-01 and R01DC007129. Any opinions, findings, conclusions or recommendations expressed in this publication are those of the authors and do not reflect the views of the NIH or NSF. Some of the results reported here appeared previously in Prud’hommeaux and Roark (2012) and the first author’s dissertation (Prud’hommeaux 2012). We thank Jan van Santen, Richard Sproat, and Chris Callison-Burch for their valuable input and the clinicians at the OHSU Layton Center for their care in collecting the data.",,,,,,,,,"9 conclusions and future work :The work presented here demonstrates the utility of adapting NLP algorithms to clinically elicited data for diagnostic purposes. In particular, the approach we describe for automatically analyzing clinically elicited language data shows promise as part of a pipeline for a screening tool for mild cognitive impairment. The methods offer the additional benefit of being general and flexible enough to be adapted to new data sets, even those without existing evaluation guidelines. In addition, the novel graph-based approach to word alignment results in large reductions in alignment error rate. These reductions in error rate in turn lead to human-level scoring accuracy and improved diagnostic classification.
The demand for simple, objective, and unobtrusive screening tools for MCI and other neurodegenerative and neurodevelopmental disorders will continue to grow
as the prevalence of these disorders increases. Although high-level measures of text similarity used in other NLP applications, such as machine translation, do achieve reasonable classification accuracy when applied to the WLM narrative data, the work presented here indicates that automated methods that approximate manual elementlevel scoring procedures yield superior results.
Although the results are quite robust, several enhancements and improvements can be made. First, although we were able to achieve decent word alignment accuracy, especially with our graph-based approach, many alignment errors remain. Exploration of the graph used here reveals that many correct alignments remain undiscovered, with an oracle AER of 11%. One clear weakness is the selection of only a single alignment from the distribution of source words at the end of 1,000 walks, since this does not allow for one-to-many mappings. We would also like to experiment with including nondirectional edges and outgoing edges on source words.
In our future work, we also plan to examine longitudinal data for individual participants to see whether our techniques can detect subtle differences in recall and coherence between a recent retelling and a series of earlier baseline retellings. Because the CDR, the dementia staging system often used to identify MCI, relies on observed changes in cognitive function over time, longitudinal analysis of performance on narrative retelling and picture description tasks might be the most promising application for this approach to analyzing clinically elicited language data.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :Interest in applying natural language processing (NLP) technology to medical information has increased in recent years. Much of this work has been focused on information retrieval and extraction from clinical notes, electronic medical records, and biomedical academic literature, but there has been some work in directly analyzing the spoken language of individuals elicited during the administration of diagnostic instruments in clinical settings. Analyzing spoken language data can reveal information not only
∗ Rochester Institute of Technology, College of Liberal Arts, 92 Lomb Memorial Dr., Rochester, NY 14623. E-mail: emilypx@rit.edu.
∗∗ Google, Inc., 1001 SW Fifth Avenue, Suite 1100, Portland OR 97204. E-mail: roarkbr@gmail.com.
Submission received: 30 December 2013; revised submission received: 21 January 2015; accepted for publication: 4 May 2015.
doi:10.1162/COLI a 00232
© 2015 Association for Computational Linguistics
about impairments in language but also about a patient’s neurological status with respect to other cognitive processes such as memory and executive function, which are often impaired in individuals with neurodevelopmental disorders, such as autism and language impairment, and neurodegenerative conditions, particularly dementia.
Many widely used instruments for diagnosing certain neurological disorders include a task in which the person must produce an uninterrupted stream of spontaneous spoken language in response to a stimulus. A person might be asked, for instance, to retell a brief narrative or to describe the events depicted in a drawing. Much of the previous work in applying NLP techniques to such clinically elicited spoken language data has relied on parsing and language modeling to enable the automatic extraction of linguistic features, such as syntactic complexity and measures of vocabulary use and diversity, which can then be used as markers for various neurological impairments (Solorio and Liu 2008; Gabani et al. 2009; Roark et al. 2011; de la Rosa et al. 2013; Fraser et al. 2014). In this article, we instead use NLP techniques to analyze the content, rather than the linguistic characteristics, of weakly structured spoken language data elicited using neuropsychological assessment instruments. We will show that the content of such spoken responses contains information that can be used for accurate screening for neurodegenerative disorders.
The features we explore are grounded in the idea that individuals recalling the same narrative are likely to use the same sorts of words and semantic concepts. In other words, a retelling of a narrative will be faithful to the source narrative and similar to other retellings. This similarity can be measured with techniques such as latent semantic analysis (LSA) cosine distance or the summary-level statistics that are widely used in evaluation of machine translation or automatic summarization, such as BLEU, Meteor, or ROUGE. Perhaps not surprisingly, however, previous work in using this type of spoken language data suggests that people with neurological impairments tend to include irrelevant or off-topic information and to exclude important pieces of information, or story elements, in their retellings that are usually included by neurotypical individuals (Hier, Hagenlocker, and Shindler 1985; Ulatowska et al. 1988; Chenery and Murdoch 1994; Chapman et al. 1995; Ehrlich, Obler, and Clark 1997; Vuorinen, Laine, and Rinne 2000; Creamer and Schmitter-Edgecombe 2010). Thus, it is often not the quantity of correctly recalled information but the quality of that information that reveals the most about a person’s diagnostic status. Summary statistics like LSA cosine distance and BLEU, which are measures of the overall degree of similarity between two texts, fail to capture these sorts of patterns. The work discussed here is an attempt to reveal these patterns and to leverage them for diagnostic classification of individuals with neurodegenerative conditions, including mild cognitive impairment and dementia of the Alzheimer’s type.
Our method for extracting the elements used in a retelling of a narrative relies on establishing a word alignment between a retelling and a source narrative. Given the correspondences between the words used in a retelling and the words used in the source narrative, we can determine with relative ease the identities of the story elements of the source narrative that were used in the retelling. These word alignments are much like those used to build machine translation models. The amount of data required to generate accurate word alignment models for machine translation, however, far exceeds the amount of monolingual source-to-retelling parallel data available to train word alignment models for our task. We therefore combine several approaches for producing reliable word alignments that exploit the peculiarities of our training data, including an entirely novel alignment approach relying on random walks on graphs.
In this article, we demonstrate that this approach to word alignment is as accurate as and more efficient than standard hidden Markov model (HMM)-based alignment (derived using the Berkeley aligner [Liang, Taskar, and Klein 2006]) for this particular data. In addition, we show that the presence or absence of specific story elements in a narrative retelling, extracted automatically from these task-specific word alignments, predicts diagnostic group membership more reliably than not only other dementia screening tools but also the lexical and semantic overlap measures widely used in NLP to evaluate pairwise language sample similarity. Finally, we apply our techniques to a picture description task that lacks an existing scoring mechanism, highlighting the generalizability and adaptability of these techniques.
The importance of accurate screening tools for neurodegenerative disorders cannot be overstated given the increased prevalence of these disorders currently being observed worldwide. In the industrialized world, for the first time in recorded history, the population over 60 years of age outnumbers the population under 15 years of age, and it is expected to be double that of children by 2050 (United Nations 2002). As the elderly population grows and as researchers find new ways to slow or halt the progression of dementia, the demand for objective, simple, and noninvasive screening tools for dementia and related disorders will grow. Although we will not discuss the application of our methods to the narratives of children, the need for simple screening protocols for neurodevelopmental disorders such as autism and language impairment is equally urgent. The results presented here indicate that the path toward these goals might include automated spoken language analysis. 2 background : Because of the variety of intact cognitive functions required to generate a narrative, the inability to coherently produce or recall a narrative is associated with many different disorders, including not only neurodegenerative conditions related to dementia, but also autism (Tager-Flusberg 1995; Diehl, Bennetto, and Young 2006), language impairment (Norbury and Bishop 2003; Bishop and Donlan 2005), attention deficit disorder (Tannock, Purvis, and Schachar 1993), and schizophrenia (Lysaker et al. 2003). The bulk of the research presented here, however, focuses on the utility of a particular narrative recall task, the Wechsler Logical Memory subtest of the Wechsler Memory Scale (Wechsler 1997), for diagnosing mild cognitive impairment (MCI). (This and other abbreviations are listed in Table 1.)
MCI is the stage of cognitive decline between the sort of decline expected in typical aging and the decline associated with dementia or Alzheimer’s disease (Petersen et al. 1999; Ritchie and Touchon 2000; Petersen 2011). MCI is characterized by subtle deficits in functions of memory and cognition that are clinically significant but do not prevent carrying out the activities of daily life. This intermediary phase of decline has been identified and named numerous times: mild cognitive decline, mild neurocognitive decline, very mild dementia, isolated memory impairment, questionable dementia, and incipient dementia. Although there continues to be disagreement about the diagnostic validity of the designation (Ritchie and Touchon 2000; Ritchie, Artero, and Touchon 2001), a number of recent studies have found evidence that seniors with some subtypes of MCI are significantly more likely to develop dementia than the population as a whole (Busse et al. 2006; Manly et al. 2008; Plassman et al. 2008). Early detection can benefit both patients and researchers investigating treatments for halting or slowing the progression of dementia, but identifying MCI can be problematic, as most dementia screening instruments, such as the Mini-Mental State Exam (MMSE) (Folstein, Folstein, and McHugh 1975), lack sufficient sensitivity to the very subtle cognitive deficits that characterize the disorder (Morris et al. 2001; Ravaglia et al. 2005; Hoops et al. 2009). Diagnosis of MCI currently requires both a lengthy neuropsychological evaluation of the patient and an interview with a family member or close associate, both of which should be repeated at regular intervals in order to have a baseline for future comparison. One goal of the work presented here is to determine whether an analysis of spoken language responses to a narrative recall task, the Wechsler Logical Memory subtest, can be used as a more efficient and less intrusive screening tool for MCI. In the Wechsler Logical Memory (WLM) narrative recall subtest of the Wechsler Memory Scale, the individual listens to a brief narrative and must verbally retell the narrative to the examiner once immediately upon hearing the story and again after a delay of 20 to 30 minutes. The examiner scores each retelling according to how many story elements the patient uses in the retelling. The standard scoring procedure, described in more detail in Section 3.2, results in a single summary score for each retelling, immediate and delayed, corresponding to the total number of story elements recalled in that retelling.
The Anna Thompson narrative, shown in Figure 1 (later in this article), has been used as the primary WLM narrative for over 70 years and has been found to be sensitive to dementia and related conditions, particularly in combination with tests of verbal fluency and memory. Multiple studies have demonstrated a significant difference in performance on the WLM between individuals with MCI and typically aging controls under the standard scoring procedure (Storandt and Hill 1989; Petersen et al. 1999; Wang and Zhou 2002; Nordlund et al. 2005). Further studies have shown that performance on the WLM can help predict whether MCI will progress into Alzheimer’s disease (Morris et al. 2001; Artero et al. 2003; Tierney et al. 2005). The WLM can also serve as a cognitive indicator of physiological characteristics associated with Alzheimer’s disease. WLM scores in the impaired range are associated with the presence of changes in Pittsburgh compound B and cerebrospinal fluid amyloid beta protein, two biomarkers of Alzheimer’s disease (Galvin et al. 2010). Poor performance on the WLM and other narrative memory tests has also been strongly correlated with increased density
of Alzheimer related lesions detected in postmortem neuropathological studies, even in the absence of previously reported or detected dementia (Schmitt et al. 2000; Bennett et al. 2006; Price et al. 2009).
We note that clinicians do not use the WLM as a diagnostic test by itself for MCI or any other type of dementia. The WLM summary score is just one of a large number of instrumentally derived scores of memory and cognitive function that, in combination with one another and with a clinician’s expert observations and examination, can indicate the presence of a dementia, aphasia, or other neurological disorder. Much of the previous work in applying automated analysis of unannotated transcripts of narratives for diagnostic purposes has focused not on evaluating properties specific to narratives but rather on using narratives as a data source from which to extract speech and language features. Solorio and Liu (2008) were able to distinguish the narratives of a small set of children with specific language impairment (SLI) from those of typically developing children using perplexity scores derived from part-ofspeech language models. In a follow-up study on a larger group of children, Gabani et al. (2009) again used part-of-speech language models in an attempt to characterize the agrammaticality that is associated with language impairment. Two part-of-speech language models were trained for that experiment: one on the language of children with SLI and one on the language of typically developing children. The perplexity of each child’s utterances was calculated according to each of the models. In addition, the authors extracted a number of other structural linguistic features including mean length of utterance, total words used in the narrative, and measures of accurate subject–verb agreement. These scores collectively performed well in distinguishing children with language impairment, achieving an F1 measure of just over 70% when used within a support vector machine (SVM) for classification. In a continuation of this work, de la Rosa et al. (2013) explored complex language-model-based lexical and syntactic features to more accurately characterize the language used in narratives by children with language impairment.
Roark et al. (2011) extracted a subset of the features used by Gabani et al. (2009), along with a much larger set of language complexity features derived from syntactic parse trees for utterances from narratives produced by elderly individuals for the diagnosis of MCI. These features included simple measures, such as words per clause, and more complex measures of tree depth, embedding, and branching, such as Frazier and Yngve scores. Selecting a subset of these features for classification with an SVM yielded a classification accuracy of 0.73, as measured by the area under the receiver operating characteristic curve (AUC). A similar approach was followed by Fraser et al. (2014) to distinguish different types of primary progressive aphasia, a group of subtypes of dementia distinct from Alzheimer’s disease and MCI, in a small group of elderly individuals. The authors considered almost 60 linguistic features, including some of those explored by Roark et al. (2011) as well as numerous others relating to part-of-speech frequencies and ratios. Using a variety of classifiers and feature combinations for three different two-way classification tasks, the authors achieved classification accuracies ranging between 0.71 and 1.0.
An alternative to analyzing narratives in terms of syntactic and lexical features is to evaluate the content of the narrative retellings themselves in terms of their fidelity to the source narrative. Hakkani-Tur, Vergyri, and Tur (2010) developed a method of
automatically evaluating an audio recording of a picture description task, in which the patient looks at a picture and narrates the events occurring in the picture, similar to the task we will be analyzing in Section 8. After using automatic speech recognition (ASR) to transcribe the recording, the authors measured unigram overlap between the ASR output transcript and a predefined list of key semantic concepts. This unigram overlap measure correlated highly with manually assigned counts of these semantic concepts. The authors did not investigate whether the scores, derived either manually or automatically, were associated with any particular diagnostic group or disorder.
Dunn et al. (2002) were among the first to apply automated methods specifically to scoring the WLM subtest and determining the relationship between these scores and measures of cognitive function. The authors used Latent Semantic Analysis (LSA) to measure the semantic distance from a retelling to the source narrative. The LSA scores correlated very highly with the scores assigned by examiners under the standard scoring guidelines and with independent measures of cognitive functioning. In subsequent work comparing individuals with and without an English-speaking background (Lautenschlager et al. 2006), the authors proposed that LSA-based scoring of the WLM as a cognitive measure is less biased against people with different linguistic and cultural backgrounds than other widely used cognitive measures. This work demonstrates not only that accurate automated scoring of narrative recall tasks is possible but also that the objectivity offered by automated measures has specific benefits for tests like the WLM, which are often administered by practitioners working in a community setting and serving a diverse population. We will compare the utility of this approach with our alignment-based approach subsequently in the article.
More recently, Lehr et al. (2013) used a supervised method for scoring the responses to the WLM, transcribed both manually and via ASR, using conditional random fields. This technique resulted in slightly higher scoring and classification accuracy than the unsupervised method described here. An unsupervised variant of their algorithm, which relied on the methods described in this article to provide training data to the conditional random field, yielded about half of the scoring gains and nearly all of the classification gains of what we report here. A hybrid method that used the methods in this article to derive features was the best performing system in that paper. Hence the methods described here are important components to that approach. We also note, however, that the supervised classifier-based approach to scoring retellings requires a significant amount of hand-labeled training data, thus rendering the technique impractical for application to a new narrative or to any picture description task. The importance of this distinction will become clear in Section 8, in which the approach outlined here is applied to a new data set lacking an existing scoring mechanism or a linguistic reference against which the responses can be scored.
In this article, we will be discussing the application of our methods to manually generated transcripts of retellings and picture descriptions produced by adults with and without neurodegenerative disorders. We note, however, that the same techniques have been applied to narratives transcribed using ASR output (Lehr et al. 2012, 2013) with little degradation in accuracy, given sufficient adaptation of the acoustic and language models to the WLM retelling domain. In addition, we have applied alignment-based scoring to the narratives of children with neurodevelopmental disorders, including autism and language impairment (Prud’hommeaux and Rouhizadeh 2012), with similarly strong diagnostic classification accuracy, further demonstrating the applicability of these methods to a variety of input formats, elicitation techniques, and diagnostic goals. 3 data : The participants for this study were drawn from an ongoing study of brain aging at the Layton Aging and Alzheimer’s Disease Center at the Oregon Health and Science University. Seventy-two of these participants had received a diagnosis of MCI, and 163 individuals served as typically aging controls. Demographic information about the experimental participants is shown in Table 2. There were no significant differences in age and years of education between the two groups. The Layton Center data included retellings for individuals who were not eligible for the present study because of their age or diagnosis. Transcriptions of 48 retellings produced by these ineligible participants were used to train and tune the word alignment model but were not used to evaluate the word alignment, scoring, or classification accuracy.
We diagnose MCI using the Clinical Dementia Rating (CDR) scale (Morris 1993), following earlier work on MCI (Petersen et al. 1999; Morris et al. 2001), as well as the work of Shankle et al. (2005) and Roark et al. (2011), who have previously attempted diagnostic classification using neuropsychological instrument subtest responses. The CDR is a numerical dementia staging scale that indicates the presence of dementia and its level of severity. The CDR score is derived from measures of cognitive function in six domains: Memory; Orientation; Judgment and Problem Solving; Community Affairs; Home and Hobbies; and Personal Care. These measures are determined during an extensive semi-structured interview with the patient and a close family member or caregiver. A CDR of 0 indicates the absence of dementia, and a CDR of 0.5 corresponds to a diagnosis of MCI (Ritchie and Touchon 2000). This measure has high expert interrater reliability (Morris 1993) and is assigned without any information derived from the WLM subtest. The WLM test, discussed in detail in Section 2.2, is a subtest of the Wechsler Memory Scale (Wechsler 1997), a neuropsychological instrument used to evaluate memory function in adults. Under standard administration of the WLM, the examiner reads a brief narrative to the participant, excerpts of which are shown in Figure 1. The participant then retells the narrative to the examiner twice: once immediately upon hearing the narrative and a second time after 20 to 30 minutes. Two retellings from one of the participants in our study are shown in Figures 2 and 3. (There are currently two narrative
retelling subtests that can be administered as part of the Wechsler Memory Scale, but the Anna Thompson narrative used in the present study is the more widely used and has appeared in every version of the Wechsler Memory Scale with only minor modifications since the instrument was first released 70 years ago.)
Following the published scoring guidelines, the examiner scores the participant’s response by counting how many of the 25 story elements are recalled in the retelling without regard to their ordering or relative importance in the story. We refer to this as the summary score. The boundaries between story elements are indicated with slashes in Figure 1. The retelling in Figure 2, produced by a participant without MCI, received a summary score of 12 for the 12 story elements recalled: Anna, Boston, employed, as a cook, and robbed of, she had four, small children, reported, station, touched by the woman’s story, took up a collection, and for her. The retelling in Figure 3, produced by the same participant after receiving a diagnosis of MCI two years later, earns a summary score of 5 for the 5 elements recalled: robbed, children, had not eaten, touched by the woman’s story, and took up a collection. Note that some of the story elements in these retellings were not recalled verbatim. The scoresheet provided with the exam indicates the lexical substitutions and degree of paraphrasing that are permitted, such as Ann or Annie for Anna, or any indication that the story evoked sympathy for touched by the woman’s story. Although the scoring guidelines have an air of arbitrariness in that paraphrasing is only sometimes permitted, they do allow the test to be scored with high inter-rater reliability (Mitchell 1987).
Recall that each participant produces two retellings for the WLM: an immediate retelling and a delayed retelling. Each participant’s two retellings were transcribed at the utterance level. The transcripts were downcased, and all pause-fillers, incomplete words, and punctuation were removed. The transcribed retellings were scored manually according to the published scoring guidelines, as described earlier in this section. 4 diagnostic classification framework : The goal of the work presented here is to demonstrate the utility of a variety of features derived from the WLM retellings for diagnostic classification of individuals with MCI. To perform this classification, we use LibSVM (Chang and Lin 2011), as implemented within the Waikato Environment for Knowledge Analysis (Weka) API (Hall et al. 2009), to train SVM classifiers, using a radial basis function kernel and default parameter settings.
We evaluate classification via receiver operating characteristic (ROC) curves, which have long been widely used to evaluate diagnostic tests (Zweig and Campbell 1993; Faraggi and Reiser 2002; Fan, Upadhye, and Worster 2006) and are also increasingly used in machine learning to evaluate classifiers in ranking scenarios (Cortes, Mohri, and Rastogi 2007; Ridgway et al. 2014). Analysis of ROC curves allows for classifier evaluation without selecting a specific, potentially arbitrary, operating point. To use standard clinical terminology, ROC curves track the tradeoff between sensitivity and specificity. Sensitivity (true positive rate) is what is commonly called recall in computational linguistics and related fields—that is, the percentage of items in the positive class that were correctly classified as positives. Specificity (true negative rate) is the percentage of items in the negative class that were correctly classified as negatives, which is equal to one minus the false positive rate. If the threshold is set so that nothing scores above threshold, the sensitivity (true positive rate, recall) is 0.0 and specificity (true negative rate) is 1.0. If the threshold is set so that everything scores above threshold, sensitivity is 1.0 and specificity is 0.0. As we sweep across intervening threshold settings, the ROC curve plots sensitivity versus one minus specificity, true positive rate versus false positive rate, providing insight into the precision/recall tradeoff at all possible operating points. Each point (tp, fp) in the curve has the true positive rate as the first dimension and false positive rate as the second dimension. Hence each curve starts at the origin (0, 0), the point corresponding to a threshold where nothing scores above threshold, and ends at (1, 1), the point where everything scores above threshold.
ROC curves can be characterized by the area underneath them (“area under curve” or AUC). A perfect classifier, with all positive items ranked above all negative items, has an ROC curve that starts at point (0, 0), goes straight up to (1, 0)—the point where true positive is 1.0 and false positive is 0.0 (since it is a perfect classifier)—before continuing straight over to the final point (1, 1). The area under this curve is 1.0, hence a perfect classifier has an AUC of 1.0. A random classifier, whose ROC curve is a straight diagonal line from the origin to (1, 1), has an AUC of 0.5. The AUC is equivalent to the probability that a randomly chosen positive example is ranked higher than a randomly chosen negative example, and is, in fact, equivalent to the Wilcoxon-Mann-Whitney statistic (Hanley and McNeil 1982). This statistic allows for classifier comparison without the need of pre-specifying arbitrary thresholds. For tasks like clinical screening, different tradeoffs between sensitivity and specificity may apply, depending on the scenario. See Fan, Upadhye, and Worster (2006) for a useful discussion of clinical use of ROC curves and the AUC score. In that paper, the authors note that there are multiple scales for interpreting the value of AUC, but that a rule-of-thumb is that AUC ≤ 0.75 is generally not clinically useful. For the present article, however, AUC mainly provides us the means for evaluating the relative quality of different classifiers.
One key issue for this sort of analysis is the estimation of the AUC for a particular classifier. Leave-pair-out cross-validation—proposed by Cortes, Mohri, and Rastogi
(2007) and extensively validated in Pahikkala et al. (2008) and Airolaa et al. (2011)— is a method for providing an unbiased estimate of the AUC, and the one we use in this article. In the leave-pair-out technique, every pairing between a negative example (i.e., a participant without MCI) and a positive example (i.e., a participant with MCI) is tested using a classifier trained on all of the remaining examples. The results of each positive/negative pair can be used to calculate the Wilcoxon-Mann-Whitney statistic as follows. Let s(e) be the score of some example e; let P be the set of positive examples and N the set of negative examples; and let [s(p) > s(n)] be 1 if true and 0 if false. Then:
AUC(s, P, n) = 1|P||N| ∑ p∈P ∑ n∈N [s(p) > s(n)] (1)
Although this method is compute-intensive, it does provide an unbiased estimate of the AUC, whereas other cross-validation setups lead to biased estimates. Another benefit of using the AUC is that standard deviation can be calculated. The standard deviation for the AUC is calculated as follows, where AUC is abbreviated as A to improve readability:
σ2a = A(1− A) + (|P| − 1)( A2−A − A 2) + (|N| − 1)( 2A21+A − A 2)
|P||N| (2) Previous work has shown that the WLM summary scores assigned during standard administration of the WLM, particularly in combination with other tests of verbal fluency and memory, are sensitive to the presence of MCI and other dementias (Storandt and Hill 1989; Petersen et al. 1999; Schmitt et al. 2000; Wang and Zhou 2002; Nordlund et al. 2005; Bennett et al. 2006; Price et al. 2009). We note, however, that the WLM test alone is not typically used as a diagnostic test. One of the goals of this work is to explore the utility of the standard WLM summary scores for diagnostic classification. A more ambitious goal is to demonstrate that using smaller units of information derived from story elements, rather than gross summary-level scores, can greatly improve diagnostic accuracy. Finally, we will show that using element-level scores automatically extracted from word alignments can achieve diagnostic classification accuracy comparable to that achieved using manually assigned scores. We therefore will compare the accuracy, measured in terms of AUC, of SVM classifiers trained on both summary-level and elementlevel WLM scores extracted from word alignments to the accuracy of classifiers built using a variety of alternative feature sets, both manually and automatically derived, shown in Table 3.
First, we consider the accuracy of classifiers using the expert-assigned WLM scores as features. For each of the 235 experimental participants, we generate two summary scores: one for the immediate retelling and one for the delayed retelling. The summary score ranges from 0, indicating that no elements were recalled, to 25, indicating that all elements were recalled. Previous work using manually assigned scores as features indicate that certain elements are more powerful in their ability to predict the presence of MCI (Prud’hommeaux 2012). In addition to the summary score, we therefore also provide the SVM with a vector of 50 story element-level scores: For each of the 25 elements in each of the two retellings per patient, there is a vector element with the value of 0 if the element was not recalled, or 1 if the element was recalled.
Classification accuracy with participants with MCI using these two manually derived feature sets is shown in Table 3.
We then present in Table 3 the classification accuracy of several summary-level features derived automatically from the WLM retellings, using standard NLP techniques for evaluating the similarity of two texts. We note that none of these features makes reference to the published WLM scoring guidelines or to the predefined element boundaries. Each of these feature sets contains two scores ranging between 0 and 1 for each participant, one for each of the two retellings: (1) cosine similarity between a retelling and the source narrative measured using LSA, proposed by Dunn et al. (2002) and calculated using the University of Colorado’s online LSA interface (available at http://lsa.colorado.edu/) with the 300-factor ninth-grade reading level topic space; (2) unigram overlap precision of a retelling relative to the source, proposed by HakkaniTur, Vergyri, and Tur (2010); (3) BLEU, the n-gram overlap metric commonly used to evaluate the quality of machine translation output (Papineni et al. 2002); and (4) the F-measure for ROUGE-SU4, the n-gram overlap metric commonly used to evaluate automatic summarization output (Lin 2004). The remaining two automatically derived features are a set of binary scores corresponding to the exact match via grep of each of the open-class unigrams in the source narrative and a summary score thereof.
Finally, in order to compare the WLM with another standard psychometric test, we also show the accuracy of a classifier trained only on the expert-assigned manual scores for the MMSE (Folstein, Folstein, and McHugh 1975), a clinician-administered 30- point questionnaire that measures a patient’s degree of cognitive impairment. Although it is widely used to screen for dementias such as Alzheimer’s disease, the MMSE is reported not to be particularly sensitive to MCI (Morris et al. 2001; Ravaglia et al. 2005; Hoops et al. 2009). The MMSE is entirely independent of the WLM and, though brief (5–10 minutes), requires more time to administer than the WLM.
In Table 3, we see that the WLM-based features yield higher accuracy than the MMSE, which is notable given the role that the MMSE plays in dementia screening. In addition, although all of the automatically derived feature sets yield higher classification than the MMSE, the manually derived WLM element-level scores are by far the most accurate feature set for diagnostic classification. Summary-level statistics, whether derived manually using established scoring mechanisms or automatically using a variety of text-similarity metrics used in the NLP community, seem not to provide sufficient power to distinguish the two diagnostic groups. In the next several sections, we describe a method for accurately automatically extracting the identities of the recalled story
elements from WLM retellings via word alignment in order to try to achieve classification accuracy comparable to that of the manually assigned WLM story elements and higher than that of the other automatic scoring methods. 5 wlm scoring via alignment :The approach presented here for automatic scoring of the WLM subtest relies on word alignments of the type used in machine translation for building phrased-based translation models. The motivation for using word alignment is the inherent similarity between narrative retelling and translation. In translation, a sentence in one language is converted into another language; the translation will have different words presented in a different order, but the meaning of the original sentence will be preserved. In narrative retelling, the source narrative is “translated” into the idiolect of the individual retelling the story. Again, the retelling will have different words, possibly presented in a different order, but at least some of the meaning will be preserved. We will show that although the algorithm for extracting scores from the alignments is simple, the process of getting high quality word alignments from the corpora of narrative retellings is challenging.
Although researchers in other NLP tasks that rely on alignments, such as textual entailment and summarization, sometimes eschew the sort of word-level alignments that are used in machine translation, we have no a priori reason to believe that this sort of alignment will be inadequate for the purposes of scoring narrative retellings. In addition, unlike many of the alignment algorithms proposed for tasks such as textual entailment, the methods for unsupervised word alignment used in machine translation require no external resources or hand-labeled data, making it simple to adapt our automated scoring techniques to new scenarios. We will show that the word alignment algorithms used in machine translation, when modified in particular ways, provide sufficient information for highly accurate scoring of narrative retellings and subsequent diagnostic classification of the individuals generating those retellings. Figure 4 shows a visual grid representation of a manually generated word alignment between the source narrative shown in Figure 1 on the vertical axis and the example WLM retelling in Figure 2 on the horizontal axis. Table 4 shows the word-index-toword-index alignment, in which the first index of each sentence is 0 and in which null alignments are not shown.
When creating these manual alignments, the labelers assigned the “possible” denotation under one of these two conditions: (1) when the alignment was ambiguous, as outlined in Och and Ney (2003); and (2) when a particular word in the retelling was a logical alignment to a word in the source narrative, but it would not have been counted as a permissible substitution under the published scoring guidelines. For this reason, we see that Taylor and sixty-seven are considered to be possible alignments because although they are logical alignments, they are not permissible substitutions according to the published scoring guidelines. Note that the word dollars is considered to be only a possible alignment, as well, since the element fifty-six dollars is not correctly recalled in this retelling under the standard scoring guidelines. In Figure 4, sure alignments are marked in black and possible alignments are marked in gray. In Figure 5, sure alignments are marked with S and possible alignments are marked with P.
Manually generated alignments like this one are the gold standard against which any automatically generated alignments can be compared to determine the accuracy
of the alignment. From an accurate word-to-word alignment, the identities of the story elements used in a retellings can be accurately extracted, and from that set of story elements, the score that is assigned under the standard scoring procedure can be calculated. As described earlier, the published scoring guidelines for the WLM specify the source words that compose each story element. Figure 5 displays the source narrative with the element IDs (A− Y) and word IDs (1− 65) explicitly labeled. Element Q, for instance, consists of the words 39 and 40, small children.
Using this information, we can determine which story elements were used in a retelling from the alignments as follows: for each word in the source narrative, if that word is aligned to a word in the retelling, the story element that it is associated with is considered to be recalled. For instance, if there is an alignment between the retelling word sympathetic and the source word touched, the story element touched by the woman’s story would be counted as correctly recalled. Note that in the WLM, every word in the source narrative is part of one of the story elements. Thus, when we convert alignments to scores in the way just described, any alignment can generate a story element. This is true even for an alignment between function words such as the and of, which would be unlikely individually to indicate that a story element had been recalled. To avoid such scoring errors, we disregard any word alignment pair containing a function word from
the source narrative. The two exceptions to this rule are the final two words, for her, which are not content words but together make a single story element.
Recall that in the manually derived word alignments, certain alignment pairs were marked as possible if the word in the retelling was logically equivalent to the word in the source but was not a permissible substitute according to the published scoring guidelines. When extracting scores from a manual alignment, only sure alignments are considered. This enables us to extract scores from a manual word alignment with 100% accuracy. The possible manual alignments are used only for calculating alignment error rate (AER) of an automatic word alignment model.
From the list of story elements extracted in this way, the summary score reported under standard scoring guidelines can be determined simply by counting the number of story elements extracted. Table 5 shows the story elements extracted from the manual word alignment in Table 4. The WLM immediate and delayed retellings for all of the 235 experimental participants and the 48 retellings from participants in the larger study who were not eligible for the present study were transcribed at the word level. Partial words, punctuation, and pause-fillers were excluded from all transcriptions used for this study. The retellings were manually scored according to published guidelines. In addition, we manually
produced word-level alignments between each retelling and the source narrative presented. These manual alignments were used to evaluate the word alignment quality and never to train the word alignment model.
Word alignment for phrase-based machine translation typically takes as input a sentence-aligned parallel corpus or bi-text, in which a sentence on one side of the corpus is a translation of the sentence in that same position on the other side of the corpus. Because we are interested in learning how to align words in the source narrative to words in the retellings, our primary parallel corpus must consist of source narrative text on one side and retelling text on the other. Because the retellings contain omissions, reorderings, and embellishments, we are obliged to consider the full text of the source narrative and of each retelling to be a “sentence” in the parallel corpus.
We compiled three parallel corpora to be used for the word alignment experiments:r Corpus 1: A 518-line source-to-retelling corpus consisting of the source narrative paired with each of the two retellings from the 235 experimental participants as well as the 48 retellings from ineligible individuals.r Corpus 2: A 268,324-line pairwise retelling-to-retelling corpus, consisting of every possible pairwise combination of the 518 available retellings.r Corpus 3: A 976-line word identity corpus, consisting of every word that appears in any retelling and the source narrative paired with itself.
The explicit parallel alignments of word identities that compose Corpus 3 are included in order to encourage the alignment of a word in a retelling to that same word in the source, if it exists.
The word alignment techniques that we use are unsupervised. Other than the transcriptions themselves, no manually generated data is used to build the word alignment models. Therefore, as in the case with most experiments involving word alignment, we build a model for the data we wish to evaluate using that same data. We do, however, use the 48 retellings from the individuals who were not experimental participants as a development set for tuning the various parameters of our word alignment system, which are described in the following. We begin by building two word alignment models using the Berkeley aligner (Liang, Taskar, and Klein 2006), a state-of-the-art word alignment package that relies on IBM Models 1 and 2 (Brown et al. 1993) and an HMM. We chose to use the Berkeley aligner, rather than the more widely used Giza++ alignment package, for this task because its joint training and posterior decoding algorithms yield lower alignment error rates on most data sets (including the data set used here [Prud’hommeaux and Roark 2011]) and because it offers functionality for testing an existing model on new data and, more crucially, for outputting posterior probabilities. The smaller of our two Berkeley-generated models is trained on Corpus 1 (the source-to-retelling parallel corpus described earlier) and ten copies of Corpus 3 (the word identity corpus). The larger model is trained on Corpus 1, Corpus 2 (the pairwise retelling corpus), and 100 copies of Corpus 3. Both models are then tested on the 470 retellings from our 235 experimental participants. In addition, we use both models to align every retelling to every other retelling so that we will have all pairwise alignments available for use in the graph-based model presented in the next section.
We note that the Berkeley aligner occasionally fails to return an alignment for a sentence pair, either because one of the sentences is too long or because the time required to perform the necessary calculations exceeds some maximum allotted time. In these cases, in order to generate alignments for all retellings and to build a complete graph that includes all retellings, we back off to the alignments and posteriors generated by IBM Model 1.
The first two rows of Table 6 show the precision, recall, and alignment error rate (AER) (Och and Ney 2003) for these two Berkeley aligner models. We note that although the AER for the larger model is lower, the time required to train the model is significantly longer. The alignments generated by the Berkeley aligner serve not only as a baseline for comparison of word alignment quality but also as a springboard for the novel graph-based method of alignment we will now discuss. Graph-based methods, in which paths or random walks are traced through an interconnected graph of nodes in order to learn more about the nodes themselves, have been used for NLP tasks in information extraction and retrieval, including Web-page ranking (PageRank; Page et al. 1999) and extractive summarization (LexRank; Erkan and Radev 2004; Otterbacher, Erkan, and Radev 2009). In the PageRank algorithm, the nodes of the graph are Web pages and the edges connecting the nodes are the hyperlinks leading from those pages to other pages. The nodes in the LexRank algorithm are sentences in a document and the edges are the similarity scores between those sentences. The number of times that a particular node is visited in a random walk reveals information about the importance of that node and its relationship to the other nodes. In many applications of random walks, the goal is to determine which node is the most central or has the highest prestige. In word alignment, however, the goal is to learn new relationships and strengthen existing relationships between words in a retelling and words in the source narrative.
In the case of our graph-based method for word alignment, each node represents a word in one of the retellings or in the source narrative. The edges are the normalized posterior-weighted alignments that the Berkeley aligner proposes between each word and (1) words in the source narrative, and (2) words in the other retellings. We generate these edges by using an existing baseline alignment model to align every retelling to every other retelling and to the source narrative. The posterior probabilities produced by the baseline alignment model serve as the weights on the edges. At each step in the walk, the choice of the next destination node can be determined according to
the strength of the outgoing edges, as measured by the posterior probability of that alignment.
Starting at a word in one of the retellings, represented by a node in the graph, the algorithm can walk from that node either to another retelling word in the graph to which it is aligned or to a word in the source narrative to which it is aligned. At each step in the walk, there is an empirically derived probability, λ, that sets the likelihood of transitioning to another retelling word versus a word in the source narrative. This probability functions similarly to the damping factor used in PageRank and LexRank, although its purpose is quite different. Once the decision whether to walk to a retelling word or source word has been made, the destination word itself is chosen according to the weights, which are the posterior probabilities assigned by the baseline alignment model. When the walk arrives at a source narrative word, that particular random walk ends, and the count for that source word as a possible alignment for the input retelling word is incremented by one.
For each word in each retelling, we perform 1,000 of these random walks, thereby generating a distribution for each retelling word over all of the words in the source narrative. The new alignment for the word is the source word with the highest frequency in that distribution. Pseudocode for this algorithm is provided in Figure 6.
Consider the following excerpts of five of the retellings. In each excerpt, the word that should align to the source word touched is rendered in bold:
the police were so moved by the story that they took up a collection for her
the fellow was sympathetic and made a collection for her so that she can feed the children
the police were touched by their story so they took up a collection
the police were so impressed with her story they took up a collection
the police felt sorry for her and took up a collection
Figure 7 presents a small idealized subgraph of the pairwise alignments of these five retellings. The arrows represent the alignments proposed by the Berkeley aligner between the relevant words in the retellings and their alignment (or lack of alignment) to the word touched in the source narrative. Thin arrows indicate alignment edges in the graph between retelling words, and bold arrows indicate alignment edges between retelling words and words in the source narrative. Words in the retellings are rendered as nodes with a single outline, and words in the source are rendered as nodes with a double outline.
We see that a number of these words were not aligned to the correct source word, touched. They are all, however, aligned to other retelling words that are in turn eventually aligned to the source word. Starting at any of the nodes in the graph, it is possible to walk from node to node and eventually reach the correct source word. Although sympathetic was not aligned to touched by the Berkeley aligner, its correct alignment can be recovered from the graph by following the path through other retelling words. After hundreds or thousands of random walks on the graph, evidence for the correct alignment will accumulate.
The approach as described might seem most beneficial to a system in need of improvements to recall rather than precision. Our baseline systems, however, are already favoring recall over precision. For this reason we include the NULL word in the list of words in the source narrative. We note that most implementations of both IBM Model 1 and HMM-based alignment also model the probability of aligning to a hidden word, NULL. In word alignment for machine translation, alignment to NULL usually indicates that a word in one language has no equivalent in the other language because the two languages express the same idea or construction in a slightly different way. Romance languages, for instance, often use prepositions before infinitival complements (e.g., Italian cerco di ridere) when English does not (e.g., I try to laugh). In the alignment of narrative retellings, however, alignment to NULL often indicates that the word in question is part of an aside or a piece of information that was not expressed in the source narrative.
Any retelling word that is not aligned to a source word by the baseline alignment system will implicitly be aligned to the hidden source word NULL, guaranteeing that every retelling word has at least one outgoing alignment edge and allowing us to
model the likelihood of being unaligned. A word that was unaligned by the original system can remain unaligned. A word that should have been left unaligned but was mistakenly aligned to a source word by the original system can recover its correct (lack of) alignment by following an edge to another retelling word that was correctly left unaligned (i.e., aligned to NULL). Figure 8 shows the graph in Figure 7 with the addition of the NULL node and the corresponding alignment edges to that node. This figure also includes two new retelling words, food and apple, and their respective alignment edges. Here we see that although the retelling word food was incorrectly aligned to the source word touched by the baseline system, its correct alignment to NULL can be recovered by traversing the edge to retelling word apple and from there, the edge to the source word NULL.
The optimal values for the following two parameters for the random walk must be determined: (1) the value of λ, the probability of walking to a retelling word node rather than a source word, and (2) the posterior probability threshold for including a particular edge in the graph. We optimize these parameters by testing the output of the graphbased approach on the development set of 48 retellings from the individuals who were not eligible for the study, discussed in Section 5.3. Recall that these additional retellings were included in the training data for the alignment model but were not included in the test set used to evaluate its performance. Tuning on this set of retellings therefore introduces no additional words, out-of-vocabulary words, or other information to the graph, while preventing overfitting.
The posterior threshold is set to 0.5 in the Berkeley aligner’s default configuration, and we found that this value did indeed yield the lowest AER for the Berkeley aligner on our data. When building the graph using Berkeley alignments and posteriors, however, we can adjust the value of this threshold to optimize the AER of the alignments produced via random walks. Using the development set of 48 retellings, we determined that the AER is minimized when the value of λ is 0.8 and the alignment inclusion posterior threshold is 0.5. Recall the two baseline alignment models generated by the Berkeley aligner, described in Section 5.4: (1) the small Berkeley model, trained on Corpus 1 (the source-to-retelling corpus) and 10 instances of Corpus 3 (the word identity corpus), and (2) the large Berkeley model (trained on Corpus 1, Corpus 2, the full pairwise retelling-to-retelling corpus, and 100 instances of Corpus 3). Using these models, we generate full retellingto-retelling alignments, on which we can then build two graph-based alignment models: the small graph-based model and the large graph-based model.
The manual gold alignments for the 235 experimental participants were evaluated against the alignments produced by each of the four models. Table 6 presents the precision, recall, and AER for the alignments of the experimental participants. Not surprisingly, the larger models yield lower error rates than the smaller models. More interestingly, each graph-based model outperforms the Berkeley model of the corresponding size by a large margin. The performance of the small graph-based model is particularly remarkable because it yields an AER superior to the large Berkeley model while requiring significantly fewer computing resources. Each of the graphbased models generated the full set of alignments in only a few minutes, whereas the large Berkeley model required 14 hours of training. 6 scoring evaluation :The element-level scores induced, as described in Section 5.2, from the four word alignments for all 235 experimental participants were evaluated against the manual per-element scores. We report the precision, recall, and F-measure for all four alignment models in Table 7. In addition, we report Cohen’s kappa as a measure of reliability between our automated scores and the manually assigned scores. We see that as AER improves, scoring accuracy also improves, with the large graph-based model outperforming all other models in terms of precision, F-measure, and inter-rater reliability. The scoring accuracy levels reported here are comparable to the levels of inter-rater agreement typically reported for the WLM, and reliability between our automated scores and the manual scores, as measured by Cohen’s kappa, is well within the ranges reported in the literature (Johnson, Storandt, and Balota 2003). As will be shown in the following section, scoring accuracy is important for achieving high classification accuracy of MCI. 7 diagnostic classification :As discussed in Section 2, poor performance on the WLM test is associated with MCI. We now use the scores we have extracted from the word alignments as features with an SVM to perform diagnostic classification for distinguishing participants with MCI from those without, as described in Section 4.1.
Table 8 shows the classification results for the scores derived from the four alignment models along with the classification results using the examiner-assigned manual scores, the MMSE, and the four alternative automated scoring approaches described in Section 4.2. It appears that, in all cases, the per-element scores are more effective than the summary scores in classifying the two diagnostic groups. In addition, we see that our automated scores have classificatory power comparable to that of the manual gold scores, and that as scoring accuracy increases from the small Berkeley model to the graph-based models and bigger models, classification accuracy improves. This suggests both that accurate scores are crucial for accurate classification and that pursuing even further improvements in word alignment is likely to result in improved diagnostic differentiation. We note that although the large Berkeley model achieved the highest classification accuracy of the automated methods, this very slight margin of difference may not justify its significantly greater computational requirements.
In addition to using summary scores and element-level scores as features for the story-element based models, we also perform feature selection over both sets of features using the chi-square statistic. Feature selection is performed separately on each training set for each fold in the cross-validation to avoid introducing bias from the testing example. We train and test the SVM using the top n story element features, from n = 1 to n = 50. We report here the accuracy for the top seven story elements (n = 7), which yielded the highest AUC measure. We note that over all of the folds, only 8 of the 50 features ever appeared among the seven most informative.
In all cases, the per-element scores are more effective than the summary scores in classifying the two diagnostic groups, and performing feature selection results in improved classification accuracy. All of the element-level feature sets automatically extracted from alignments outperform the MMSE and all of the alternative automatic
scoring procedures, which suggests that the extra complexity required to extract element-level features is well worth the time and effort.
We note that the final classification results for all four alignment models are not drastically different from one another, despite the large reductions in word alignment error rate and improvements in scoring accuracy observed in the larger models and graph-based models. This seeming disconnect between word alignment accuracy and downstream application performance has also been observed in the machine translation literature, where reductions in AER do not necessarily lead to meaningful increases in BLEU, the widely accepted measure of machine translation quality (Ayan and Dorr 2006; Lopez and Resnik 2006; Fraser and Marcu 2007). Our results, however, show that a feature set consisting of manually assigned WLM scores yields the highest classification accuracy of any of the feature sets evaluated here. As discussed in Section 5.2, our WLM score extraction method is designed such that element-level scores can be extracted with perfect accuracy from a perfect word alignment. Thus, the goal of seeking perfect or near-perfect word alignment accuracy is worthwhile because it will necessarily result in perfect or near-perfect scoring accuracy, which in turn is likely to yield classification accuracy approaching that of manually assigned scores. 8 application to task with non-linguistic reference :As we discussed earlier, one of the advantages of using an unsupervised method of scoring is the resulting generalizability to new data sets, particularly those generated from a non-linguistic stimulus. The Boston Diagnostic Aphasia Examination (BDAE) (Goodglass and Kaplan 1972), an instrument widely used to diagnose aphasia in adults, includes one such task, popularly known as the cookie theft picture description task. In this test, the person views a drawing of a lively scene in a family’s kitchen and must tell the examiner about all of the actions they see in the picture. The picture is reproduced below in Figure 9.
Describing visually presented material is quite different from a task such as the WLM, in which language comprehension and memory play a crucial role. Nevertheless, the processing and language production demands of a picture description task may lead to differences in performance in groups with certain cognitive and language problems. In fact, it is widely reported that the picture descriptions of seniors with dementia of the Alzheimer’s type differ from those of typically aging seniors in terms of information content (Hier, Hagenlocker, and Shindler 1985; Gilesa, Patterson, and Hodge 1996). Interestingly, this reduction in information is not necessarily accompanied by a reduction in the amount of language produced. Rather, it seems that seniors with Alzheimer’s dementia tend to include redundant information, repetitions, intrusions, and revisions that result in language samples of length comparable to that of typically aging seniors.
TalkBank (MacWhinney 2007), the online database of audio and transcribed speech, has made available the DementiaBank corpus of descriptions of the cookie theft picture by hundreds of individuals, some of whom have one of a number of types of dementia, including MCI, vascular dementia, possible Alzheimer’s disease, and probable Alzheimer’s disease. From this corpus were selected a subset of individuals without dementia and a subset with probable Alzheimer’s disease. We limit the set of descriptions to those with more than 25 but fewer than 100 words, yielding 130 descriptions for each diagnostic group. There was no significant difference in description word count between the two diagnostic groups.
The first task was to generate a source description to which all other narratives should be aligned. Working under the assumption that the control participants would
produce good descriptions, we calculated the BLEU score of every pair of descriptions from the control group. The description with the highest average pairwise BLEU score was selected as the source description. After confirming that this description did in fact contain all of the action portrayed in the picture, we removed all extraneous conversational asides from the description in order to ensure that it contained all and only information about the picture. The selected source description is as follows:
The boy is getting cookies out of the cookie jar. And the stool is just about to fall over. The little girl is reaching up for a cookie. And the mother is drying dishes. The water is running into the sink and the sink is running over onto the floor. And that little girl is laughing.
We then built an alignment model on the full pairwise description parallel corpus (2602 = 67,600 sentences) and a word identity corpus consisting of each word in each description reproduced 100 times. Using this trained model, which corresponds to the large Berkeley model that achieved the highest classification accuracy for the WLM data, we then aligned every description to the artificial source description. We also built a graph-based alignment model using these alignments and the parameter settings that maximized word alignment accuracy in the WLM data. Because the artificial source description is not a true linguistic reference for this task, we did not produce manual word alignments against which the alignment quality could be evaluated and against which the parameters could be tuned. Instead, we evaluated only the downstream application of diagnostic classification.
The method for scoring the WLM relies directly on the predetermined list of story elements, whereas the cookie theft picture description administration instructions do not include an explicit set of items that must be described. Recall that the automated scoring method we propose uses only the open-class or content words in the source narrative. In order to generate scores for the descriptions, we propose a scoring technique that considers each content word in the source description to be its own story element. Any word in a retelling that aligns to one of the content words in the source narrative is considered to be a match for that content word element. This results in a large number of elements, but it allows the scoring method to be easily adapted to other
narrative production scenarios that similarly do not have explicit scoring guidelines. Using these scores as features, we again used an SVM to classify the two diagnostic groups, typically aging and probable Alzheimer’s disease, and evaluated the classifier using leave-pair-out validation.
Table 9 shows the classification results using the content word scoring features produced using the Berkeley aligner alignments and the graph-based alignments. These can be compared to classification results using the summary similarity metrics BLEU and unigram precision. We see that using word-level features, regardless of which alignment model they are extracted from, results in significantly higher classification accuracy than both the simple similarity metrics and the summary scores. The alignmentbased scoring approach yields features with remarkably high classification accuracy given the somewhat ad hoc selection of the source narrative from the set of control retellings.
These results demonstrate the flexibility and utility of the alignment-based approach to scoring narratives. Not only can it be adapted to other narrative retelling instruments, but it can relatively trivially be adapted to instruments that use nonlinguistic stimuli for elicitation. All that is needed to build an alignment model is a sufficiently large collection of retellings of the same narrative or descriptions of the same picture. Procuring such a collection of descriptions or retellings can be done easily outside a clinical setting using a platform such as Amazon’s Mechanical Turk. No handlabeled data, outside lexical resources, prior knowledge of the content of the story, or existing scoring guidelines are required. Among the more recent applications for natural language processing algorithms has been the analysis of spoken language data for diagnostic and remedial purposes, fueled by the demand for simple, objective, and unobtrusive screening tools for neurological disorders such as dementia. The automated analysis of narrative retellings in particular shows potential as a component of such a screening tool since the ability to produce accurate and meaningful narratives is noticeably impaired in individuals with dementia and its frequent precursor, mild cognitive impairment, as well as other neurodegenerative and neurodevelopmental disorders. In this article, we present a method for extracting narrative recall scores automatically and highly accurately from a word-level alignment between a retelling and the source narrative. We propose improvements to existing machine translation-based systems for word alignment, including a novel method of word alignment relying on random walks on a graph that achieves alignment accuracy superior to that of standard expectation maximization-based techniques for word alignment in a fraction of the time required for expectation maximization. In addition, the narrative recall score features extracted from these high-quality word alignments yield diagnostic classification accuracy comparable to that achieved using manually assigned scores and significantly higher than that achieved with summary-level text similarity metrics used in other areas of NLP. These methods can be trivially adapted to spontaneous language samples elicited with non-linguistic stimuli, thereby demonstrating the flexibility and generalizability of these methods. [{""affiliations"": [], ""name"": ""Emily Prud\u2019hommeaux""}, {""affiliations"": [], ""name"": ""Brian Roark""}] SP:9fff4c3673ff8364cd50b46aa7b757952eefa5de [{""authors"": [""Airolaa"", ""Antti"", ""Tapio Pahikkalaa"", ""Willem Waegemanc"", ""Bernard De Baetsc"", ""Tapio Salakoskia""], ""title"": ""An experimental comparison of cross-validation techniques for estimating the area under the ROC"", ""year"": 2011}, {""authors"": [""Artero"", ""Sylvain"", ""Mary Tierney"", ""Jacques Touchon"", ""Karen Ritchie""], ""title"": ""Prediction of transition from cognitive"", ""year"": 2003}, {""authors"": [""Ayan"", ""Necip Fazil"", ""Bonnie J. Dorr.""], ""title"": ""Going beyond AER: An extensive analysis of word alignments and their impact on MT"", ""venue"": ""Proceedings of the 21st International Conference on Computational Linguistics and"", ""year"": 2006}, {""authors"": [""D.A. Bennett"", ""J.A. Schneider"", ""Z. Arvanitakis"", ""J.F. Kelly"", ""N.T. Aggarwal"", ""R.C. Shah"", ""R.S. Wilson""], ""title"": ""Neuropathology of older persons without cognitive impairment"", ""year"": 2006}, {""authors"": [""Bishop"", ""Dorothy"", ""Chris Donlan.""], ""title"": ""The role of syntax in encoding and recall of pictorial narratives: Evidence from specific language impairment"", ""venue"": ""British Journal of Developmental Psychology,"", ""year"": 2005}, {""authors"": [""Brown"", ""Peter"", ""Vincent Della Pietra"", ""Steven Della Pietra"", ""Robert Mercer.""], ""title"": ""The mathematics of statistical machine translation: Parameter estimation"", ""venue"": ""Computational Linguistics,"", ""year"": 1993}, {""authors"": [""Chang"", ""Chih-Chung"", ""Chih-Jen Lin.""], ""title"": ""LIBSVM: A library for support vector machines"", ""venue"": ""ACM Transactions on Intelligent Systems and Technology, 2(27):1\u201327."", ""year"": 2011}, {""authors"": [""Chapman"", ""Sandra"", ""Hanna Ulatowska"", ""Kristin King"", ""Julene Johnson"", ""Donald McIntire.""], ""title"": ""Discourse in early Alzheimer\u2019s disease versus normal advanced aging"", ""venue"": ""American Journal of"", ""year"": 1995}, {""authors"": [""Chenery"", ""Helen J."", ""Bruce E. Murdoch.""], ""title"": ""The production of narrative discourse in response to animations in persons with dementia of the Alzheimer\u2019s type: Preliminary findings"", ""venue"": ""Aphasiology,"", ""year"": 1994}, {""authors"": [""Cortes"", ""Corinna"", ""Mehryar Mohri"", ""Ashish Rastogi.""], ""title"": ""An alternative ranking problem for search engines"", ""venue"": ""Proceedings of the 6th Workshop on Experimental Algorithms, volume 4525 of Lecture Notes in"", ""year"": 2007}, {""authors"": [""Creamer"", ""Scott"", ""Maureen Schmitter-Edgecombe""], ""title"": ""Narrative comprehension in Alzheimer\u2019s disease: Assessing inferences and memory operations with a think-aloud procedure"", ""year"": 2010}, {""authors"": [""de la Rosa"", ""Gabriela Ramirez"", ""Thamar Solorio"", ""Manuel Montes y Gomez"", ""Aquiles Iglesias"", ""Yang Liu"", ""Lisa Bedore"", ""Elizabeth Pena""], ""title"": ""Exploring word class n-grams to measure language"", ""year"": 2013}, {""authors"": [""Diehl"", ""Joshua J."", ""Loisa Bennetto"", ""Edna Carter Young.""], ""title"": ""Story recall and narrative coherence of high-functioning children with autism spectrum disorders"", ""venue"": ""Journal of Abnormal Child Psychology,"", ""year"": 2006}, {""authors"": [""Dunn"", ""John C."", ""Osvaldo P. Almeida"", ""Lee Barclay"", ""Anna Waterreus"", ""Leon Flicker.""], ""title"": ""Latent semantic analysis: A new method to measure prose recall"", ""venue"": ""Journal of Clinical and Experimental Neuropsychology,"", ""year"": 2002}, {""authors"": [""Ehrlich"", ""Jonathan S."", ""Loraine K. Obler"", ""Lynne Clark.""], ""title"": ""Ideational and semantic contributions to narrative production in adults with dementia of the Alzheimer\u2019s type"", ""venue"": ""Journal of Communication Disorders,"", ""year"": 1997}, {""authors"": [""Erkan"", ""G\u00fcnes"", ""Dragomir R. Radev""], ""title"": ""LexRank: Graph-based lexical centrality as salience in text summarization"", ""year"": 2004}, {""authors"": [""Fan"", ""Jerome"", ""Suneel Upadhye"", ""Andrew Worster.""], ""title"": ""Understanding receiver operating characteristic (ROC) curves"", ""venue"": ""Canadian Journal of Emergency Medicine, 8:19\u201320."", ""year"": 2006}, {""authors"": [""Faraggi"", ""David"", ""Benjamin Reiser.""], ""title"": ""Estimation of the area under the ROC curve"", ""venue"": ""Statistics in Medicine, 21:3093\u20133106."", ""year"": 2002}, {""authors"": [""M. Folstein"", ""S. Folstein"", ""P. McHugh.""], ""title"": ""Mini-mental state\u2014a practical method for grading the cognitive state of patients for the clinician"", ""venue"": ""Journal of Psychiatric Research, 12:189\u2013198."", ""year"": 1975}, {""authors"": [""Fraser"", ""Alexander"", ""Daniel Marcu.""], ""title"": ""Measuring word alignment quality for statistical machine translation"", ""venue"": ""Computational Linguistics, 33(3):293\u2013303."", ""year"": 2007}, {""authors"": [""Fraser"", ""Kathleen C"", ""Jed A. Meltzer"", ""Naida L. Graham"", ""Carol Leonard"", ""Graeme Hirst"", ""Sandra E. Black"", ""Elizabeth Rochon""], ""title"": ""Automated classification of primary progressive aphasia subtypes"", ""year"": 2014}, {""authors"": [""Gabani"", ""Keyur"", ""Melissa Sherman"", ""Thamar Solorio"", ""Yang Liu""], ""title"": ""A corpus-based approach for the prediction of language impairment in monolingual English and Spanish-English bilingual"", ""year"": 2009}, {""authors"": [""Galvin"", ""James"", ""Anne Fagan"", ""David Holtzman"", ""Mark Mintun"", ""John Morris.""], ""title"": ""Relationship of dementia screening tests with biomarkers of Alzheimer\u2019s disease"", ""venue"": ""Brain, 133:3290\u20133300."", ""year"": 2010}, {""authors"": [""Gilesa"", ""Elaine"", ""Karalyn Patterson"", ""John R. Hodge""], ""title"": ""Performance on the Boston cookie theft picture description task in patients with early dementia of the Alzheimer\u2019s type: Missing information"", ""year"": 1996}, {""authors"": [""H. Goodglass"", ""E. Kaplan.""], ""title"": ""Boston Diagnostic Aphasia Examination"", ""venue"": ""Lea and Febiger, Philadelphia, PA."", ""year"": 1972}, {""authors"": [""Hakkani-Tur"", ""Dilek"", ""Dimitra Vergyri"", ""Gokhan Tur.""], ""title"": ""Speech-based automated cognitive status assessment"", ""venue"": ""Proceedings of the Conference of the International Speech Communication"", ""year"": 2010}, {""authors"": [""Hall"", ""Mark"", ""Eibe Frank"", ""Geoffrey Holmes"", ""Bernhard Pfahringer"", ""Peter Reutemann"", ""Ian H. Witten.""], ""title"": ""The WEKA data mining software: An update"", ""venue"": ""SIGKDD Explorations, 11(1):10\u201318."", ""year"": 2009}, {""authors"": [""Hanley"", ""James"", ""Barbara McNeil.""], ""title"": ""The meaning and use of the area under a receiver operating characteristic (ROC) curve"", ""venue"": ""Radiology, 143:29\u201336."", ""year"": 1982}, {""authors"": [""D. Hier"", ""K. Hagenlocker"", ""A. Shindler.""], ""title"": ""Language disintegration in dementia: Effects of etiology and severity"", ""venue"": ""Brain and Language, 25:117\u2013133."", ""year"": 1985}, {""authors"": [""S. Hoops"", ""S. Nazem"", ""A.D. Siderowf"", ""J.E. Duda"", ""S.X. Xie"", ""M.B. Stern"", ""D. Weintraub.""], ""title"": ""Validity of the MoCA and MMSE in the detection of MCI and dementia in Parkinson disease"", ""venue"": ""Neurology,"", ""year"": 2009}, {""authors"": [""Johnson"", ""David K."", ""Martha Storandt"", ""David A. Balota.""], ""title"": ""Discourse analysis of logical memory recall in normal aging and in dementia of the Alzheimer type"", ""venue"": ""Neuropsychology, 17(1):82\u201392."", ""year"": 2003}, {""authors"": [""Lautenschlager"", ""Nicola T"", ""John C. Dunn"", ""Kathryn Bonney"", ""Leon Flicker"", ""Osvaldo P. Almeida""], ""title"": ""Latent semantic analysis: An improved method to measure cognitive performance in subjects"", ""year"": 2006}, {""authors"": [""Lehr"", ""Maider"", ""Emily Prud\u2019hommeaux"", ""Izhak Shafran"", ""Brian Roark""], ""title"": ""Fully automated neuropsychological assessment for detecting mild cognitive impairment"", ""venue"": ""In Proceedings of the 13th Annual Conference"", ""year"": 2012}, {""authors"": [""Lehr"", ""Maider"", ""Izhak Shafran"", ""Emily Prud\u2019hommeaux"", ""Brian Roark""], ""title"": ""Discriminative joint modeling of lexical variation and acoustic confusion for automated narrative retelling"", ""year"": 2013}, {""authors"": [""Liang"", ""Percy"", ""Ben Taskar"", ""Dan Klein.""], ""title"": ""Alignment by agreement"", ""venue"": ""Proceedings of the Human Language Technology Conference of the NAACL, pages 104\u2013111, New York, NY."", ""year"": 2006}, {""authors"": [""Lin"", ""Chin-Yiu.""], ""title"": ""ROUGE: A package for automatic evaluation of summaries"", ""venue"": ""Proceedings of the Workshop on Text Summarization Branches Out, pages 74\u201381, Barcelona."", ""year"": 2004}, {""authors"": [""Lopez"", ""Adam"", ""Philip Resnik.""], ""title"": ""Word-based alignment, phrase-based translation: What\u2019s the link"", ""venue"": ""Proceedings"", ""year"": 2006}, {""authors"": [""Lysaker"", ""Paul"", ""Amanda Wickett"", ""Neil Wilke"", ""John Lysaker.""], ""title"": ""Narrative incoherence in schizophrenia: The absent agent-protagonist and the collapse of internal dialogue"", ""venue"": ""American Journal of"", ""year"": 2003}, {""authors"": [""MacWhinney"", ""Brian.""], ""title"": ""The TalkBank Project"", ""venue"": ""J. C. Beal, K. P. Corrigan, and H. L. Moisl, editors, Creating and Digitizing Language Corpora: Synchronic Databases, Vol.1, pages 163\u2013180, Palgrave-Macmillan,"", ""year"": 2007}, {""authors"": [""Manly"", ""Jennifer J."", ""Ming Tang"", ""Nicole Schupf"", ""Yaakov Stern"", ""Jean-Paul G. Vonsattel"", ""Richard Mayeux.""], ""title"": ""Frequency and course of mild cognitive impairment in a multiethnic community"", ""venue"": ""Annals of"", ""year"": 2008}, {""authors"": [""Mitchell"", ""Margaret.""], ""title"": ""Scoring discrepancies on two subtests of the Wechsler memory scale"", ""venue"": ""Journal of Consulting and Clinical Psychology, 55:914\u2013915."", ""year"": 1987}, {""authors"": [""Morris"", ""John.""], ""title"": ""The Clinical Dementia Rating (CDR): Current version and scoring rules"", ""venue"": ""Neurology, 43:2412\u20132414."", ""year"": 1993}, {""authors"": [""Morris"", ""John"", ""Martha Storandt"", ""J. Phillip Miller"", ""Daniel McKeel"", ""Joseph Price"", ""Eugene Rubin"", ""Leonard Berg.""], ""title"": ""Mild cognitive impairment represents early-stage Alzheimer disease"", ""venue"": ""Archives of"", ""year"": 2001}, {""authors"": [""Norbury"", ""Courtenay"", ""Dorothy Bishop.""], ""title"": ""Narrative skills of children with communication impairments"", ""venue"": ""International Journal of Language and Communication Disorders, 38:287\u2013313."", ""year"": 2003}, {""authors"": [""A. Nordlund"", ""S. Rolstad"", ""P. Hellstrom"", ""M. Sjogren"", ""S. Hansen"", ""A. Wallin.""], ""title"": ""The Goteborg MCI study: Mild cognitive impairment is a heterogeneous condition"", ""venue"": ""Journal of Neurology, Neurosurgery and"", ""year"": 2005}, {""authors"": [""Och"", ""Franz Josef"", ""Hermann Ney.""], ""title"": ""A systematic comparison of various statistical alignment models"", ""venue"": ""Computational Linguistics, 29(1):19\u201351."", ""year"": 2003}, {""authors"": [""Otterbacher"", ""Jahna"", ""G\u00fcnes Erkan"", ""Dragomir R. Radev.""], ""title"": ""Biased LexRank: Passage retrieval using random walks with question-based priors"", ""venue"": ""Information Processing Management,"", ""year"": 2009}, {""authors"": [""Page"", ""Lawrence"", ""Sergey Brin"", ""Rajeev Motwani"", ""Terry Winograd""], ""title"": ""The PageRank citation ranking: Bringing order"", ""year"": 1999}, {""authors"": [""Prud\u2019hommeaux"", ""Roark""], ""title"": ""Graph-Based Word Alignment for Clinical Language Evaluation to the web"", ""venue"": ""Technical Report 1999-66,"", ""year"": 1999}, {""authors"": [""Pahikkala"", ""Tapio"", ""Antti Airola"", ""Jorma Boberg"", ""Tapio Salakoski.""], ""title"": ""Exact and efficient leave-pair-out cross-validation for ranking RLS"", ""venue"": ""The Second International and Interdisciplinary Conference on Adaptive"", ""year"": 2008}, {""authors"": [""Papineni"", ""Kishore"", ""Salim Roukos"", ""Todd Ward"", ""Wei jing Zhu.""], ""title"": ""BLEU: A method for automatic evaluation of machine translation"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for"", ""year"": 2002}, {""authors"": [""Petersen"", ""Ronald"", ""Glenn Smith"", ""Stephen Waring"", ""Robert Ivnik"", ""Eric Tangalos"", ""Emre Kokmen.""], ""title"": ""Mild cognitive impairment: Clinical characterizations and outcomes"", ""venue"": ""Archives of Neurology,"", ""year"": 1999}, {""authors"": [""Petersen"", ""Ronald C.""], ""title"": ""Mild cognitive impairment"", ""venue"": ""The New England Journal of Medicine, 364(23):2227\u20132234."", ""year"": 2011}, {""authors"": [""John J. McArdle"", ""Robert J. Willis"", ""Robert B. Wallace.""], ""title"": ""Prevalence of cognitive impairment without dementia in the United States"", ""venue"": ""Annals of Internal Medicine, 148:427\u201334."", ""year"": 2008}, {""authors"": [""John C. Morris.""], ""title"": ""Neuropathology of nondemented aging: Presumptive evidence for preclinical Alzheimer disease"", ""venue"": ""Neurobiology of Aging, 30(7):1026\u20131036."", ""year"": 2009}, {""authors"": [""Prud\u2019hommeaux"", ""Emily"", ""Brian Roark""], ""title"": ""Alignment of spoken narratives for automated neuropsychological assessment"", ""venue"": ""In Proceedings of the IEEE Workshop on Automatic Speech Recognition"", ""year"": 2011}, {""authors"": [""Prud\u2019hommeaux"", ""Emily"", ""Brian Roark""], ""title"": ""Graph-based alignment of narratives"", ""year"": 2012}, {""authors"": [""Prud\u2019hommeaux"", ""Emily"", ""Masoud Rouhizadeh""], ""title"": ""Automatic detection of pragmatic deficits in children with autism"", ""venue"": ""In Proceedings of the 3rd Workshop on Child, Computer and Interaction,"", ""year"": 2012}, {""authors"": [""Prud\u2019hommeaux"", ""Emily Tucker""], ""title"": ""Alignment of Narrative Retellings for Automated Neuropsychological Assessment"", ""venue"": ""Ph.D. thesis,"", ""year"": 2012}, {""authors"": [""Ravaglia"", ""Giovanni"", ""Paola Forti"", ""Fabiola Maioli"", ""Lucia Servadei"", ""Mabel Martelli"", ""Nicoletta Brunetti"", ""Luciana Bastagli"", ""Erminia Mariani""], ""title"": ""Screening for mild cognitive impairment in elderly"", ""year"": 2005}, {""authors"": [""Ridgway"", ""James"", ""Pierre Alquier"", ""Nicolas Chopin"", ""Feng Liang.""], ""title"": ""PAC-Bayesian AUC classification and scoring"", ""venue"": ""Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and"", ""year"": 2014}, {""authors"": [""Ritchie"", ""Karen"", ""Sylvaine Artero"", ""Jacques Touchon.""], ""title"": ""Classification criteria for mild cognitive impairment: A population-based validation study"", ""venue"": ""Neurology, 56:37\u201342."", ""year"": 2001}, {""authors"": [""Ritchie"", ""Karen"", ""Jacques Touchon.""], ""title"": ""Mild cognitive impairment: Conceptual basis and current nosological status"", ""venue"": ""Lancet, 355:225\u2013228."", ""year"": 2000}, {""authors"": [""Roark"", ""Brian"", ""Margaret Mitchell"", ""John-Paul Hosom"", ""Kristina Hollingshead"", ""Jeffrey Kaye.""], ""title"": ""Spoken language derived measures for detecting mild cognitive impairment"", ""venue"": ""IEEE Transactions on Audio,"", ""year"": 2011}, {""authors"": [""F.A. Schmitt"", ""D.G. Davis"", ""D.R. Wekstein"", ""C.D. Smith"", ""J.W. Ashford"", ""W.R. Markesbery.""], ""title"": ""Preclinical AD revisited: Neuropathology of cognitively normal older adults"", ""venue"": ""Neurology,"", ""year"": 2000}, {""authors"": [""Shankle"", ""William R"", ""A. Kimball Romney"", ""Junko Hara"", ""Dennis Fortier"", ""Malcolm B. Dick"", ""James M. Chen"", ""Timothy Chan"", ""Xijiang Sun""], ""title"": ""Methods to improve the detection of mild cognitive"", ""year"": 2005}, {""authors"": [""Solorio"", ""Thamar"", ""Yang Liu.""], ""title"": ""Using language models to identify language impairment in Spanish-English bilingual children"", ""venue"": ""Proceedings of the ACL 2008 Workshop on Biomedical Natural Language"", ""year"": 2008}, {""authors"": [""Storandt"", ""Martha"", ""Robert Hill.""], ""title"": ""Very mild senile dementia of the Alzheimer\u2019s type: II"", ""venue"": ""Psychometric test performance. Archives of Neurology, 46:383\u2013386."", ""year"": 1989}, {""authors"": [""Tager-Flusberg"", ""Helen.""], ""title"": ""Once upon a ribbit: Stories narrated by autistic children"", ""venue"": ""British Journal of Developmental Psychology, 13(1):45\u201359."", ""year"": 1995}, {""authors"": [""Tannock"", ""Rosemary"", ""Karen L. Purvis"", ""Russell J. Schachar.""], ""title"": ""Narrative abilities in children with attention deficit hyperactivity disorder and normal peers"", ""venue"": ""Journal of Abnormal Child Psychology,"", ""year"": 1993}, {""authors"": [""Tierney"", ""Mary"", ""Christie Yao"", ""Alex Kiss"", ""Ian McDowell.""], ""title"": ""Neuropsychological tests accurately predict incident Alzheimer disease after 5 and 10 years"", ""venue"": ""Neurology, 64:1853\u20131859."", ""year"": 2005}, {""authors"": [""Ulatowska"", ""Hanna"", ""Lee Allard"", ""Adrienne Donnell"", ""Jean Bristow"", ""Sara M. Haynes"", ""Adelaide Flower"", ""Alvin J. North""], ""title"": ""Discourse performance in subjects with dementia"", ""year"": 1988}, {""authors"": [""United Nations.""], ""title"": ""World Population Ageing 1950\u20132050"", ""venue"": ""United Nations, New York."", ""year"": 2002}, {""authors"": [""Vuorinen"", ""Elina"", ""Matti Laine"", ""Juha Rinne.""], ""title"": ""Common pattern of language impairment in vascular dementia and in Alzheimer disease"", ""venue"": ""Alzheimer Disease and Associated Disorders, 14(2):81\u201386."", ""year"": 2000}, {""authors"": [""Wang"", ""Qing-Song"", ""Jiang-Ning Zhou.""], ""title"": ""Retrieval and encoding of episodic memory in normal aging and patients with mild cognitive impairment"", ""venue"": ""Brain Research, 924:113\u2013115."", ""year"": 2002}, {""authors"": [""Wechsler"", ""David.""], ""title"": ""Wechsler Memory Scale - Third Edition"", ""venue"": ""The Psychological Corporation, San Antonio, TX."", ""year"": 1997}, {""authors"": [""Zweig"", ""Mark H."", ""Gregory Campbell.""], ""title"": ""Receiver-operating characteristic (ROC) plots: A fundamental evaluation tool in clinical medicine"", ""venue"": ""Clinical Chemistry, 39:561\u2013577."", ""year"": 1993}] acknowledgments  :This research was conducted while both authors were at the Center for Spoken Language Understanding at the Oregon Health and Science University, in Portland, Oregon. This work was supported in part by NSF grant BCS-0826654 and NIH NIDCD grants R01DC012033-01 and R01DC007129. Any opinions, findings, conclusions or recommendations expressed in this publication are those of the authors and do not reflect the views of the NIH or NSF. Some of the results reported here appeared previously in Prud’hommeaux and Roark (2012) and the first author’s dissertation (Prud’hommeaux 2012). We thank Jan van Santen, Richard Sproat, and Chris Callison-Burch for their valuable input and the clinicians at the OHSU Layton Center for their care in collecting the data. 9 conclusions and future work :The work presented here demonstrates the utility of adapting NLP algorithms to clinically elicited data for diagnostic purposes. In particular, the approach we describe for automatically analyzing clinically elicited language data shows promise as part of a pipeline for a screening tool for mild cognitive impairment. The methods offer the additional benefit of being general and flexible enough to be adapted to new data sets, even those without existing evaluation guidelines. In addition, the novel graph-based approach to word alignment results in large reductions in alignment error rate. These reductions in error rate in turn lead to human-level scoring accuracy and improved diagnostic classification.
The demand for simple, objective, and unobtrusive screening tools for MCI and other neurodegenerative and neurodevelopmental disorders will continue to grow
as the prevalence of these disorders increases. Although high-level measures of text similarity used in other NLP applications, such as machine translation, do achieve reasonable classification accuracy when applied to the WLM narrative data, the work presented here indicates that automated methods that approximate manual elementlevel scoring procedures yield superior results.
Although the results are quite robust, several enhancements and improvements can be made. First, although we were able to achieve decent word alignment accuracy, especially with our graph-based approach, many alignment errors remain. Exploration of the graph used here reveals that many correct alignments remain undiscovered, with an oracle AER of 11%. One clear weakness is the selection of only a single alignment from the distribution of source words at the end of 1,000 walks, since this does not allow for one-to-many mappings. We would also like to experiment with including nondirectional edges and outgoing edges on source words.
In our future work, we also plan to examine longitudinal data for individual participants to see whether our techniques can detect subtle differences in recall and coherence between a recent retelling and a series of earlier baseline retellings. Because the CDR, the dementia staging system often used to identify MCI, relies on observed changes in cognitive function over time, longitudinal analysis of performance on narrative retelling and picture description tasks might be the most promising application for this approach to analyzing clinically elicited language data.","2 background : Because of the variety of intact cognitive functions required to generate a narrative, the inability to coherently produce or recall a narrative is associated with many different disorders, including not only neurodegenerative conditions related to dementia, but also autism (Tager-Flusberg 1995; Diehl, Bennetto, and Young 2006), language impairment (Norbury and Bishop 2003; Bishop and Donlan 2005), attention deficit disorder (Tannock, Purvis, and Schachar 1993), and schizophrenia (Lysaker et al. 2003). The bulk of the research presented here, however, focuses on the utility of a particular narrative recall task, the Wechsler Logical Memory subtest of the Wechsler Memory Scale (Wechsler 1997), for diagnosing mild cognitive impairment (MCI). (This and other abbreviations are listed in Table 1.)
MCI is the stage of cognitive decline between the sort of decline expected in typical aging and the decline associated with dementia or Alzheimer’s disease (Petersen et al. 1999; Ritchie and Touchon 2000; Petersen 2011). MCI is characterized by subtle deficits in functions of memory and cognition that are clinically significant but do not prevent carrying out the activities of daily life. This intermediary phase of decline has been identified and named numerous times: mild cognitive decline, mild neurocognitive decline, very mild dementia, isolated memory impairment, questionable dementia, and incipient dementia. Although there continues to be disagreement about the diagnostic validity of the designation (Ritchie and Touchon 2000; Ritchie, Artero, and Touchon 2001), a number of recent studies have found evidence that seniors with some subtypes of MCI are significantly more likely to develop dementia than the population as a whole (Busse et al. 2006; Manly et al. 2008; Plassman et al. 2008). Early detection can benefit both patients and researchers investigating treatments for halting or slowing the progression of dementia, but identifying MCI can be problematic, as most dementia screening instruments, such as the Mini-Mental State Exam (MMSE) (Folstein, Folstein, and McHugh 1975), lack sufficient sensitivity to the very subtle cognitive deficits that characterize the disorder (Morris et al. 2001; Ravaglia et al. 2005; Hoops et al. 2009). Diagnosis of MCI currently requires both a lengthy neuropsychological evaluation of the patient and an interview with a family member or close associate, both of which should be repeated at regular intervals in order to have a baseline for future comparison. One goal of the work presented here is to determine whether an analysis of spoken language responses to a narrative recall task, the Wechsler Logical Memory subtest, can be used as a more efficient and less intrusive screening tool for MCI. In the Wechsler Logical Memory (WLM) narrative recall subtest of the Wechsler Memory Scale, the individual listens to a brief narrative and must verbally retell the narrative to the examiner once immediately upon hearing the story and again after a delay of 20 to 30 minutes. The examiner scores each retelling according to how many story elements the patient uses in the retelling. The standard scoring procedure, described in more detail in Section 3.2, results in a single summary score for each retelling, immediate and delayed, corresponding to the total number of story elements recalled in that retelling.
The Anna Thompson narrative, shown in Figure 1 (later in this article), has been used as the primary WLM narrative for over 70 years and has been found to be sensitive to dementia and related conditions, particularly in combination with tests of verbal fluency and memory. Multiple studies have demonstrated a significant difference in performance on the WLM between individuals with MCI and typically aging controls under the standard scoring procedure (Storandt and Hill 1989; Petersen et al. 1999; Wang and Zhou 2002; Nordlund et al. 2005). Further studies have shown that performance on the WLM can help predict whether MCI will progress into Alzheimer’s disease (Morris et al. 2001; Artero et al. 2003; Tierney et al. 2005). The WLM can also serve as a cognitive indicator of physiological characteristics associated with Alzheimer’s disease. WLM scores in the impaired range are associated with the presence of changes in Pittsburgh compound B and cerebrospinal fluid amyloid beta protein, two biomarkers of Alzheimer’s disease (Galvin et al. 2010). Poor performance on the WLM and other narrative memory tests has also been strongly correlated with increased density
of Alzheimer related lesions detected in postmortem neuropathological studies, even in the absence of previously reported or detected dementia (Schmitt et al. 2000; Bennett et al. 2006; Price et al. 2009).
We note that clinicians do not use the WLM as a diagnostic test by itself for MCI or any other type of dementia. The WLM summary score is just one of a large number of instrumentally derived scores of memory and cognitive function that, in combination with one another and with a clinician’s expert observations and examination, can indicate the presence of a dementia, aphasia, or other neurological disorder. Much of the previous work in applying automated analysis of unannotated transcripts of narratives for diagnostic purposes has focused not on evaluating properties specific to narratives but rather on using narratives as a data source from which to extract speech and language features. Solorio and Liu (2008) were able to distinguish the narratives of a small set of children with specific language impairment (SLI) from those of typically developing children using perplexity scores derived from part-ofspeech language models. In a follow-up study on a larger group of children, Gabani et al. (2009) again used part-of-speech language models in an attempt to characterize the agrammaticality that is associated with language impairment. Two part-of-speech language models were trained for that experiment: one on the language of children with SLI and one on the language of typically developing children. The perplexity of each child’s utterances was calculated according to each of the models. In addition, the authors extracted a number of other structural linguistic features including mean length of utterance, total words used in the narrative, and measures of accurate subject–verb agreement. These scores collectively performed well in distinguishing children with language impairment, achieving an F1 measure of just over 70% when used within a support vector machine (SVM) for classification. In a continuation of this work, de la Rosa et al. (2013) explored complex language-model-based lexical and syntactic features to more accurately characterize the language used in narratives by children with language impairment.
Roark et al. (2011) extracted a subset of the features used by Gabani et al. (2009), along with a much larger set of language complexity features derived from syntactic parse trees for utterances from narratives produced by elderly individuals for the diagnosis of MCI. These features included simple measures, such as words per clause, and more complex measures of tree depth, embedding, and branching, such as Frazier and Yngve scores. Selecting a subset of these features for classification with an SVM yielded a classification accuracy of 0.73, as measured by the area under the receiver operating characteristic curve (AUC). A similar approach was followed by Fraser et al. (2014) to distinguish different types of primary progressive aphasia, a group of subtypes of dementia distinct from Alzheimer’s disease and MCI, in a small group of elderly individuals. The authors considered almost 60 linguistic features, including some of those explored by Roark et al. (2011) as well as numerous others relating to part-of-speech frequencies and ratios. Using a variety of classifiers and feature combinations for three different two-way classification tasks, the authors achieved classification accuracies ranging between 0.71 and 1.0.
An alternative to analyzing narratives in terms of syntactic and lexical features is to evaluate the content of the narrative retellings themselves in terms of their fidelity to the source narrative. Hakkani-Tur, Vergyri, and Tur (2010) developed a method of
automatically evaluating an audio recording of a picture description task, in which the patient looks at a picture and narrates the events occurring in the picture, similar to the task we will be analyzing in Section 8. After using automatic speech recognition (ASR) to transcribe the recording, the authors measured unigram overlap between the ASR output transcript and a predefined list of key semantic concepts. This unigram overlap measure correlated highly with manually assigned counts of these semantic concepts. The authors did not investigate whether the scores, derived either manually or automatically, were associated with any particular diagnostic group or disorder.
Dunn et al. (2002) were among the first to apply automated methods specifically to scoring the WLM subtest and determining the relationship between these scores and measures of cognitive function. The authors used Latent Semantic Analysis (LSA) to measure the semantic distance from a retelling to the source narrative. The LSA scores correlated very highly with the scores assigned by examiners under the standard scoring guidelines and with independent measures of cognitive functioning. In subsequent work comparing individuals with and without an English-speaking background (Lautenschlager et al. 2006), the authors proposed that LSA-based scoring of the WLM as a cognitive measure is less biased against people with different linguistic and cultural backgrounds than other widely used cognitive measures. This work demonstrates not only that accurate automated scoring of narrative recall tasks is possible but also that the objectivity offered by automated measures has specific benefits for tests like the WLM, which are often administered by practitioners working in a community setting and serving a diverse population. We will compare the utility of this approach with our alignment-based approach subsequently in the article.
More recently, Lehr et al. (2013) used a supervised method for scoring the responses to the WLM, transcribed both manually and via ASR, using conditional random fields. This technique resulted in slightly higher scoring and classification accuracy than the unsupervised method described here. An unsupervised variant of their algorithm, which relied on the methods described in this article to provide training data to the conditional random field, yielded about half of the scoring gains and nearly all of the classification gains of what we report here. A hybrid method that used the methods in this article to derive features was the best performing system in that paper. Hence the methods described here are important components to that approach. We also note, however, that the supervised classifier-based approach to scoring retellings requires a significant amount of hand-labeled training data, thus rendering the technique impractical for application to a new narrative or to any picture description task. The importance of this distinction will become clear in Section 8, in which the approach outlined here is applied to a new data set lacking an existing scoring mechanism or a linguistic reference against which the responses can be scored.
In this article, we will be discussing the application of our methods to manually generated transcripts of retellings and picture descriptions produced by adults with and without neurodegenerative disorders. We note, however, that the same techniques have been applied to narratives transcribed using ASR output (Lehr et al. 2012, 2013) with little degradation in accuracy, given sufficient adaptation of the acoustic and language models to the WLM retelling domain. In addition, we have applied alignment-based scoring to the narratives of children with neurodevelopmental disorders, including autism and language impairment (Prud’hommeaux and Rouhizadeh 2012), with similarly strong diagnostic classification accuracy, further demonstrating the applicability of these methods to a variety of input formats, elicitation techniques, and diagnostic goals. 9 conclusions and future work :The work presented here demonstrates the utility of adapting NLP algorithms to clinically elicited data for diagnostic purposes. In particular, the approach we describe for automatically analyzing clinically elicited language data shows promise as part of a pipeline for a screening tool for mild cognitive impairment. The methods offer the additional benefit of being general and flexible enough to be adapted to new data sets, even those without existing evaluation guidelines. In addition, the novel graph-based approach to word alignment results in large reductions in alignment error rate. These reductions in error rate in turn lead to human-level scoring accuracy and improved diagnostic classification.
The demand for simple, objective, and unobtrusive screening tools for MCI and other neurodegenerative and neurodevelopmental disorders will continue to grow
as the prevalence of these disorders increases. Although high-level measures of text similarity used in other NLP applications, such as machine translation, do achieve reasonable classification accuracy when applied to the WLM narrative data, the work presented here indicates that automated methods that approximate manual elementlevel scoring procedures yield superior results.
Although the results are quite robust, several enhancements and improvements can be made. First, although we were able to achieve decent word alignment accuracy, especially with our graph-based approach, many alignment errors remain. Exploration of the graph used here reveals that many correct alignments remain undiscovered, with an oracle AER of 11%. One clear weakness is the selection of only a single alignment from the distribution of source words at the end of 1,000 walks, since this does not allow for one-to-many mappings. We would also like to experiment with including nondirectional edges and outgoing edges on source words.
In our future work, we also plan to examine longitudinal data for individual participants to see whether our techniques can detect subtle differences in recall and coherence between a recent retelling and a series of earlier baseline retellings. Because the CDR, the dementia staging system often used to identify MCI, relies on observed changes in cognitive function over time, longitudinal analysis of performance on narrative retelling and picture description tasks might be the most promising application for this approach to analyzing clinically elicited language data."
"1 introduction :There is very little similar about coffee and cups. Coffee refers to a plant, which is a living organism or a hot brown (liquid) drink. In contrast, a cup is a man-made solid of broadly well-defined shape and size with a specific function relating to the consumption of liquids. Perhaps the only clear trait these concepts have in common is that they are concrete entities. Nevertheless, in what is currently the most popular evaluation gold standard for semantic similarity, WordSim(WS)-353 (Finkelstein et al. 2001), coffee and
∗ Computer Laboratory University of Cambridge, UK. E-mail: {felix.hill, anna.korhonen}@ cl.cam.ac.uk.
∗∗ Technion, Israel Institute of Technology, Haifa, Israel. E-mail: roiri@ie.technion.ac.il.
Submission received: 25 July 2014; revised submission received: 10 June 2015; accepted for publication: 31 August 2015.
doi:10.1162/COLI a 00237
© 2015 Association for Computational Linguistics
cup are rated as more “similar” than pairs such as car and train, which share numerous common properties (function, material, dynamic behavior, wheels, windows, etc.). Such anomalies also exist in other gold standards such as the MEN data set (Bruni et al. 2012a). As a consequence, these evaluations effectively penalize models for learning the evident truth that coffee and cup are dissimilar.
Although clearly different, coffee and cup are very much related. The psychological literature refers to the conceptual relationship between these concepts as association, although it has been given a range of names including relatedness (Budanitsky and Hirst 2006; Agirre et al. 2009), topical similarity (Hatzivassiloglou et al. 2001), and domain similarity (Turney 2012). Association contrasts with similarity, the relation connecting cup and mug (Tversky 1977). At its strongest, the similarity relation is exemplified by pairs of synonyms; words with identical referents.
Computational models that effectively capture similarity as distinct from association have numerous applications. Such models are used for the automatic generation of dictionaries, thesauri, ontologies, and language correction tools (Biemann 2005; Cimiano, Hotho, and Staab 2005; Li et al. 2006). Machine translation systems, which aim to define mappings between fragments of different languages whose meaning is similar, but not necessarily associated, are another established application (He et al. 2008; Marton, Callison-Burch, and Resnik 2009). Moreover, since, as we establish, similarity is a cognitively complex operation that can require rich, structured conceptual knowledge to compute accurately, similarity estimation constitutes an effective proxy evaluation for general-purpose representation-learning models whose ultimate application is variable or unknown (Collobert and Weston 2008; Baroni and Lenci 2010).
As we show in Section 2, the predominant gold standards for semantic evaluation in NLP do not measure the ability of models to reflect similarity. In particular, in both WS353 and MEN, pairs of words with associated meaning, such as coffee and cup (rating = 6.810), telephone and communication (7.510), or movie and theater (7.710), receive a high rating regardless of whether or not their constituents are similar. Thus, the utility of such resources to the development and application of similarity models is limited, a problem exacerbated by the fact that many researchers appear unaware of what their evaluation resources actually measure.1
Although certain smaller gold standards—those of Rubenstein and Goodenough (1965) (RG) and Agirre et al. (2009) (WS-Sim)—do focus clearly on similarity, these resources suffer from other important limitations. For instance, as we show, and as is also the case for WS-353 and MEN, state-of-the-art models have reached the average performance of a human annotator on these evaluations. It is common practice in NLP to define the upper limit for automated performance on an evaluation as the average human performance or inter-annotator agreement (Yong and Foo 1999; Cunningham 2005; Resnik and Lin 2010). Based on this established principle and the current evaluations, it would therefore be reasonable to conclude that the problem of representation learning, at least for similarity modeling, is approaching resolution. However, circumstantial evidence suggests that distributional models are far from perfect. For instance, we are some way from automatically generated dictionaries, thesauri, or ontologies that can be used with the same confidence as their manually created equivalents.
1 For instance, Huang et al. (2012, pages 1, 4, 10) and Reisinger and Mooney (2010b, page 4) refer to MEN and/or WS-353 as “similarity data sets.” Others evaluate on both these association-based and genuine similarity-based gold standards with no reference to the fact that they measure different things (Medelyan et al. 2009; Li et al. 2014).
Motivated by these observations, in Section 3 we present SimLex-999, a gold standard resource for evaluating the ability of models to reflect similarity. SimLex-999 was produced by 500 paid native English speakers, recruited via Amazon Mechanical Turk,2 who were asked to rate the similarity, as opposed to association, of concepts via a simple visual interface. The choice of evaluation pairs in SimLex-999 was motivated by empirical evidence that humans represent concepts of distinct part-of-speech (POS) (Gentner 1978) and conceptual concreteness (Hill, Korhonen, and Bentz 2014) differently. Whereas existing gold standards contain only concrete noun concepts (MEN) or cover only some of these distinctions via a random selection of items (WS-353, RG), SimLex-999 contains a principled selection of adjective, verb, and noun concept pairs covering the full concreteness spectrum. This design enables more nuanced analyses of how computational models overcome the distinct challenges of representing concepts of these types.
In Section 4 we present quantitative and qualitative analyses of the SimLex-999 ratings, which indicate that participants found it unproblematic to quantify consistently the similarity of the full range of concepts and to distinguish it from association. Unlike existing data sets, SimLex-999 therefore contains a significant number of pairs, such as [movie, theater], which are strongly associated but receive low similarity scores.
The second main contribution of this paper, presented in Section 5, is the evaluation of state-of-the-art distributional semantic models using SimLex-999. These include the well-known neural language models (NLMs) of Huang et al. (2012), Collobert and Weston (2008), and Mikolov et al. (2013a), which we compare with traditional vectorspace co-occurrence models (VSMs) (Turney and Pantel 2010) with and without dimensionality reduction (SVD) (Landauer and Dumais 1997). Our analyses demonstrate how SimLex-999 can be applied to uncover substantial differences in the ability of models to represent concepts of different types.
Despite these differences, the models we consider each share the characteristic of being better able to capture association than similarity. We show that the difficulty of estimating similarity is driven primarily by those strongly associated pairs with a high (association) rating in gold standards such as WS-353 and MEN, but a low similarity rating in SimLex-999. As a result of including these challenging cases, together with a wider diversity of lexical concepts in general, current models achieve notably lower scores on SimLex-999 than on existing gold standard evaluations, and well below the SimLex-999 inter-human agreement ceiling.
Finally, we explore ways in which distributional models might improve on this performance in similarity modeling. To do so, we evaluate the models on the SimLex999 subsets of adjectives, nouns, and verbs, as well as on abstract and concrete subsets and subsets of more and less strongly associated pairs (Sections 5.2.2–5.2.4). As part of these analyses, we confirm the hypothesis (Agirre et al. 2009; Levy and Goldberg 2014) that models learning from input informed by dependency parsing, rather than simple running-text input, yield improved similarity estimation and, specifically, clearer distinction between similarity and association. In contrast, we find no evidence for a related hypothesis (Agirre et al. 2009; Kiela and Clark 2014) that smaller context windows improve the ability of models to capture similarity. We do, however, observe clear differences in model performance on the distinct concept types included in SimLex-999. Taken together, these experiments demonstrate the benefit of the diversity of concepts
2 www.mturk.com/.
included in SimLex-999; it would not have been possible to derive similar insights by evaluating based on existing gold standards.
We conclude by discussing how observations such as these can guide future research into distributional semantic models. By facilitating better-defined evaluations and finer-grained analyses, we hope that SimLex-999 will ultimately contribute to the development of models that accurately reflect human intuitions of similarity for the full range of concepts in language.","2 design motivation :In this section, we motivate the design decisions made in developing SimLex-999. We begin (2.1) by examining the distinction between similarity and association. We then show that for a meaningful treatment of similarity it is also important to take a principled approach to both POS and conceptual concreteness (2.2). We finish by reviewing existing gold standards, and show that none enables a satisfactory evaluation of the capability of models to capture similarity (2.3). The difference between association and similarity is exemplified by the concept pairs [car, bike] and [car, petrol]. Car is said to be (semantically) similar to bike and associated with (but not similar to) petrol. Intuitively, car and bike can be understood as similar because of their common physical features (e.g., wheels), their common function (transport), or because they fall within a clearly definable category (modes of transport). In contrast, car and petrol are associated because they frequently occur together in space and language, in this case as a result of a clear functional relationship (Plaut 1995; McRae, Khalkhali, and Hare 2012).
Association and similarity are neither mutually exclusive nor independent. Bike and car, for instance, are related to some degree by both relations. Because it is common in both the physical world and in language for distinct entities to interact, it is relatively easy to conceive of concept pairs, such as car and petrol, that are strongly associated but not similar. Identifying pairs of concepts for which the converse is true is comparatively more difficult. One exception is common concepts paired with low frequency synonyms, such as camel and dromedary. Because the essence of association is co-occurrence (linguistic or otherwise [McRae, Khalkhali, and Hare 2012]), such pairs can seem, at least intuitively, to be similar but not strongly associated.
To explore the interaction between the two cognitive phenomena quantitatively, we exploited perhaps the only two existing large-scale means of quantifying similarity and association. To estimate similarity, we considered proximity in the WordNet taxonomy (Fellbaum 1998). Specifically, we applied the measure of Wu and Palmer (1994) (henceforth WupSim), which approximates similarity on a [0,1] scale reflecting the minimum distance between any two synsets of two given concepts in WordNet. WupSim has been shown to correlate well with human judgments on the similarity-focused RG data set (Wu and Palmer 1994). To estimate association, we extracted ratings directly from the University of South Florida Free Association Database (USF) (Nelson, McEvoy, and Schreiber 2004). These data were generated by presenting human subjects with one of 5,000 cue concepts and asking them to write the first word that comes into their head that is associated with or meaningfully related to that concept. Each cue concept c was normed in this way by over 10 participants, resulting in a set of associates for each cue, and a total of over 72,000 (c, a) pairs. Moreover, for each such pair, the proportion of participants
who produced associate a when presented with cue c can be used as a proxy for the strength of association between the two concepts.
By measuring WupSim between all pairs in the USF data set, we observed, as expected, a high correlation between similarity and association strength across all USF pairs (Spearman ρ = 0.65, p < 0.001). However, in line with the intuitive ubiquity of pairs such as car and petrol, of the USF pairs (all of which are associated to a greater or lesser degree) over 10% had a WupSim score of less than 0.25. These include pairs of ontologically different entities with a clear functional relationship in the world [refrigerator, food], which may be of differing concreteness [lung, disease]; pairs in which one concept is a small concrete part of a larger abstract category [sheriff, police]; pairs in a relationship of modification or subcategorization [gravy, boat]; and even those whose principal connection is phonetic [wiggle, giggle]. As we show in Section 2.2, these are precisely the sort of pairs that are not contained in existing evaluation gold standards. Table 1 lists the USF noun pairs with the lowest similarity scores overall, and also those with the largest additive discrepancy between association strength and similarity.
2.1.1 Association and Similarity in NLP. As noted in the Introduction, the similarity/association distinction is not only of interest to researchers in psychology or linguistics. Models of similarity are particularly applicable to various NLP tasks, such as lexical resource building, semantic parsing, and machine translation (Haghighi et al. 2008; He et al. 2008; Marton, Callison-Burch, and Resnik 2009; Beltagy, Erk, and Mooney 2014). Models of association, on the other hand, may be better suited to tasks such as wordsense disambiguation (Navigli 2009), and applications such as text classification (Phan, Nguyen, and Horiguchi 2008) in which the target classes correspond to topical domains such as agriculture or sport (Rose, Stevenson, and Whitehead 2002).
Much recent research in distributional semantics does not distinguish between association and similarity in a principled way (see, e.g., Reisinger and Mooney 2010b; Huang et al. 2012; Luong, Socher, and Manning 2013).3 One exception is Turney (2012), who constructs two distributional models with different features and parameter settings, explicitly designed to capture either similarity or association. Using the output of these two models as input to a logistic regression classifier, Turney predicts whether two
3 Several papers that take a knowledge-based or symbolic approach to meaning do address the similarity/association issue (Budanitsky and Hirst 2006).
concepts are associated, similar, or both, with 61% accuracy. However, in the absence of a gold standard covering the full range of similarity ratings (rather than a list of pairs identified as being similar or not) Turney cannot confirm directly that the similarityfocused model does indeed effectively quantify similarity.
Agirre et al. (2009) explicitly examine the distinction between association and similarity in relation to distributional semantic models. Their study is based on the partition of WS-353 into a subset focused on similarity, which we refer to as WS-Sim, and a subset focused on association, which we term WS-Rel. More precisely, WS-Sim is the union of the pairs in WS-353 judged by three annotators to be similar and the set U of entirely unrelated pairs, and WS-Rel is the union of U and pairs judged to be associated but not similar. Agirre et al. confirm the importance of the association/similarity distinction by showing that certain models perform relatively well on WS-Rel, whereas others perform comparatively better on WS-Sim. However, as shown in the following section, a model need not be an exemplary model of similarity in order to perform well on WS-Sim, because an important class of concept pair (associated but not similar entities) is not represented in this data set. Therefore the insights that can be drawn from the results of the Agirre et al. study are limited.
Several other authors touch on the similarity/association distinction in inspecting the output of distributional models (Andrews, Vigliocco, and Vinson 2009; Kiela and Clark 2014; Levy and Goldberg 2014). Although the strength of the conclusions that can be drawn from such qualitative analyses is clearly limited, there appear to be two broad areas of consensus concerning similarity and distributional models:r Models that learn from input annotated for syntactic or dependency
relations better reflect similarity, whereas approaches that learn from running-text or bag-of-words input better model association (Agirre et al. 2009; Levy and Goldberg 2014).r Models with larger context windows may learn representations that better capture association, whereas models with narrower windows better reflect similarity (Agirre et al. 2009; Kiela and Clark 2014). Empirical studies have shown that the performance of both humans and distributional models depends on the POS category of the concepts learned. Gentner (2006) showed that children find verb concepts harder to learn than noun concepts, and Markman and Wisniewski (1997) present evidence that different cognitive operations are used when comparing two nouns or two verbs. Hill, Reichart, and Korhonen (2014) demonstrate differences in the ability of distributional models to acquire noun and verb semantics. Further, they show that these differences are greater for models that learn from both text and perceptual input (as with humans).
In addition to POS category, differences in human and computational concept learning and representation have been attributed to the effects of concreteness, the extent to which a concept has a directly perceptible physical referent. On the cognitive side, these “concreteness effects” are well established, even if the causes are still debated (Paivio 1991; Hill, Korhonen, and Bentz 2014). Concreteness has also been associated with differential performance in computational text-based (Hill, Kiela, and Korhonen 2013) and multi-modal semantic models (Kiela et al. 2014). For brevity, we do not exhaustively review all methods that have been used to evaluate semantic models, but instead focus on the similarity or association-based gold standards that are most commonly applied in recent work in NLP. In each case, we consider how well the data set satisfies one of the three following criteria:
Representative. The resource should cover the full range of concepts that occur in natural language. In particular, it should include cases representing the different ways in which humans represent or process concepts, and cases that are both challenging and straightforward for computational models.
Clearly defined. In order for a gold standard to be diagnostic of how well a model can be applied to downstream applications, a clear understanding is needed of what exactly the gold standard measures. In particular, it must clearly distinguish between dissociable semantic relations such as association and similarity.
Consistent and reliable. Untrained native speakers must be able to quantify the target property consistently, without requiring lengthy or detailed instructions. This ensures that the data reflect a meaningful cognitive or semantic phenomenon, and also enables the data set to be scaled up or transferred to other languages at minimal cost and effort.
We begin our review of existing evaluation with the gold standard most commonly applied in current NLP research.
WordSim-353. WS-353 (Finkelstein et al. 2001) is perhaps the most commonly used evaluation gold standard for semantic models. Despite its name, and the fact that it is often referred to as a “similarity gold standard,”4 in fact, the instructions given to annotators when producing WS-353 were ambiguous with respect to similarity and association. Subjects were asked to:
Assign a numerical similarity score between 0 and 10 (0 = words totally unrelated, 10 = words VERY closely related) ... when estimating similarity of antonyms, consider them “similar” (i.e., belonging to the same domain or representing features of the same concept), not “dissimilar”.
As we confirm analytically in Section 5.2, these instructions result in pairs being rated according to association rather than similarity.5 WS-353 consequently suffers two important limitations as an evaluation of similarity (which also afflict other resources to a greater or lesser degree):
1. Many dissimilar word pairs receive a high rating.
2. No associated but dissimilar concepts receive low ratings.
As noted in the Introduction, an arguably more serious third limitation of WS-353 is low inter-annotator agreement, and the fact that state-of-the-art models such as those
4 See, e.g., Huang et al. 2012 and Bansal, Gimpel, and Livescu 2014. 5 This fact is also noted by the data set authors. See www.cs.technion.ac.il/~gabr/resources/ data/wordsim353/.
of Collobert and Weston (2008) and Huang et al. (2012) reach, or even surpass, the inter-annotator agreement ceiling in estimating the WS-353 scores. Huang et al. report a Spearman correlation of ρ = 0.713 between their model output and WS-353. This is 10 percentage points higher than inter-annotator agreement (ρ = 0.611) when defined as the average pairwise correlation between two annotators, as is common in NLP work (Padó, Padó, and Erk 2007; Reisinger and Mooney 2010a; Silberer and Lapata 2014). It could be argued that a different comparison is more appropriate: Because the model is compared to the gold-standard average across all annotators, we should compare a single annotator with the (almost) gold-standard average over all other annotators. Based on this metric the average performance of an annotator on WS-353 is ρ = 0.756, which is still only marginally better than the best automatic method.6
Thus, at least according to the established wisdom in NLP evaluation (Yong and Foo 1999; Cunningham 2005; Resnik and Lin 2010), the strength of the conclusions that can be inferred from improvements on WS-353 is limited. At the same time, however, state-of-the-art distributional models are clearly not perfect representation-learning or even similarity estimation engines, as evidenced by the fact they cannot yet be applied, for instance, to generate flawless lexical resources (Alfonseca and Manandhar 2002).
WS-Sim. WS-Sim is the set of pairs in WS-353 identified by Agirre et al. (2009) as either containing similar or unrelated (neither similar nor associated) concepts. The ratings in WS-Sim are mapped directly from WS-353, so that all concept pairs in WSSim that receive a high rating are associated and all pairs that receive a low rating are unassociated. Consequently, any model that simply reflects association would score highly on WS-Sim, irrespective of how well it captures similarity.
Such a possibility could be excluded by requiring models to perform well on WSSim and poorly on WS-Rel, the subset of WS-353 identified by Agirre et al. (2009) as containing no pairs of similar concepts. However, although this would exclude models of pure association, it would not test the ability of models to quantify the similarity of the pairs in WS-Sim. Put another way, the WS-Sim/WS-Rel partition could in theory resolve limitation (1) of WS-353 but it would not resolve limitation (2): Models are not tested on their ability to attribute low scores to associated but dissimilar pairs.
In fact, there are more fundamental limitations of WS-Sim as a similarity-based evaluation resource. It does not, strictly speaking, reflect similarity at all, since the ratings of its constituent pairs were assigned by the WS-353 annotators, who were asked to estimate association, not similarity. Moreover, it inherits the limitation of low interannotator agreement from WS-353. The average pairwise correlation between annotators on WS-Sim is ρ = 0.667, and the average correlation of a single annotator with the gold standard is only ρ = 0.651, both below the performance of automatic methods (Agirre et al. 2009). Finally, the small size of WS-Sim renders it poorly representative of the full range of concepts that semantic models may be required to learn.
Rubenstein & Goodenough. Prior to WS-353, the smaller RG data set, consisting of 65 pairs, was often used to evaluate semantic models. The 15 raters employed in the data collection were asked to rate the “similarity of meaning” of each concept pair. Thus RG does appear to reflect similarity rather than association. However, although limitation (1) of WS-353 is therefore avoided, RG still suffers from limitation (2): By inspection,
6 Individual annotator responses for WS-353 were downloaded from www.cs.technion.ac.il/~gabr/ resources/data/wordsim353.
it is clear that the low similarity pairs in RG are not associated. A further limitation is that distributional models now achieve better performance on RG (correlations of up to Pearson r = 0.86 [Hassan and Mihalcea 2011]) than the reported inter-annotator agreement of r = 0.85 (Rubenstein and Goodenough 1965). Finally, the size of RG renders it an even less comprehensive evaluation than WS-Sim.
The MEN Test Collection. A larger data set, MEN (Bruni et al. 2012a), is used in a handful of recent studies (Bruni et al. 2012b; Bernardi et al. 2013). As with WS-353, both terms similarity and relatedness are used by the authors when describing MEN, although the annotators were expressly asked to rate pairs according to relatedness.7
The construction of MEN differed from RG and WS-353 in that each pair was only considered by one rater, who ranked it for relatedness relative to 50 other pairs in the data set. An overall score out of 50 was then attributed to each pair corresponding to how many times it was ranked as more related than an alternative. However, because these rankings are based on relatedness, with respect to evaluating similarity MEN necessarily suffers from both of the limitations (1) and (2) that apply to WS-353. Further, there is a strong bias towards concrete concepts in MEN because the concepts were originally selected from those identified in an image-bank (Bruni et al. 2012a).
Synonym Detection Sets. Multiple-choice synonym detection tasks, such as the TOEFL test questions (Landauer and Dumais 1997), are an alternative means of evaluating distributional models. A question in the TOEFL task consists of a cue word and four possible answer words, only one of which is a true synonym. Models are scored on the number of true synonyms identified out of 80 questions. The questions were designed by linguists to evaluate synonymy, so, unlike the evaluations considered thus far, TOEFL-style tests effectively discriminate between similarity and association. However, because they require a zero-one classification of pairs as synonymous or not, they do not test how well models discern pairs of medium or low similarity. More generally, in opposition to the fuzzy, statistical approaches to meaning predominant in both cognitive psychology (Griffiths, Steyvers, and Tenenbaum 2007) and NLP (Turney and Pantel 2010), they do not require similarity to be measured on a continuous scale.","3 the simlex-999 data set :Having considered the limitations of existing gold standards, in this section we describe the design of SimLex-999 in detail. Separating similarity from association. To create a test of the ability of models to capture similarity as opposed to association, we started with the ≈ 72,000 pairs of concepts in the USF data set. As the output of a free-association experiment, each of these pairs is associated to a greater or lesser extent. Importantly, inspecting the pairs revealed that a good range of similarity values are represented. In particular, there were many examples of hypernym/hyponym pairs [body, abdomen], cohyponym pairs [cat, dog], synonyms or near synonyms [deodorant, antiperspirant], and antonym pairs [good, evil]. From this cohort, we excluded pairs containing a multiple-word item [hot dog, mustard],
7 http://clic.cimec.unitn.it/~elia.bruni/MEN.html.
and pairs containing a capital letter [Mexico, sun]. We ultimately sampled 900 of the SimLex-999 pairs from the resulting cohort of pairs, according to the stratification procedures outlined in the following sections.
To complement this cohort with entirely unassociated pairs, we paired up the concepts from the 900 associated pairs at random. From these random parings, we excluded those that coincidentally occurred elsewhere in USF (and therefore had a degree of association). From the remaining pairs, we accepted only those in which both concepts had been subject to the USF norming procedure, ensuring that these non-USF pairs were indeed unassociated rather than simply not normed. We sampled the remaining 99 SimLex-999 pairs from this resulting cohort of unassociated pairs.
POS category. In light of the conceptual differences outlined in Section 2.2, SimLex999 includes subsets of pairs from the three principle meaning-bearing POS categories: nouns, verbs, and adjectives. To classify potential pairs according to POS, we counted the frequency with which the items in each pair occurred with the three possible tags in the POS-tagged British National Corpus (Leech, Garside, and Bryant 1994). To minimize POS ambiguity, which could lead to inconsistent ratings, we excluded pairs containing a concept with lower than 75% tendency towards one particular POS. This yielded three sets of potential pairs : [A,A] pairs (of two concepts whose majority tag was Adjective), [N,N] pairs, and [V,V] pairs.
Given the likelihood that different cognitive operations are used in estimating the similarity between items of different POS-category (Section 2.2), concept pairs were presented to raters in batches defined according to POS. Unlike both WS-353 and MEN, pairs of concepts of mixed POS ([white, rabbit], [run,marathon]) were excluded. POS categories are generally considered to reflect very broad ontological classes (Fellbaum 1998). We thus felt it would be very difficult, or even counter-intuitive, for annotators to quantify the similarity of mixed POS pairs according to our instructions.
Concreteness. Although a clear majority of pairs in gold standards such as MEN and RG contain concrete items, perhaps surprisingly, the vast majority of adjective, noun, and verb concepts in everyday language are in fact abstract (Hill, Reichart, and Korhonen 2014; Kiela et al. 2014).8 To facilitate the evaluation of models for both concrete and abstract concept meaning, and in light of the cognitive and computational modeling differences between abstract and concrete concepts noted in Section 2.2, we aimed to include both concept types in SimLex-999.
Unlike the POS distinction, concreteness is generally considered to be a gradual phenomenon. One benefit of sampling pairs for SimLex-999 from the USF data set is that most items have been rated according to concreteness on a scale of 1–7 by at least 10 human subjects. As Figure 1 demonstrates, concreteness (as the average over these ratings) interacts with POS on these concepts: Nouns are on average more concrete than verbs, which are more concrete than adjectives. However, there is also clear variation in concreteness within each POS category. We therefore aimed to select pairs for SimLex999 that spanned the full abstract–concrete continuum within each POS category.
After excluding any pairs that contained an item with no concreteness rating, for each potential SimLex-999 pair we considered both the concreteness of the first item and the additive difference in concreteness between the two items. This enabled us
8 According to the USF concreteness ratings, 72% of noun or verb types in the British National Corpus are more abstract than the concept war, a concept many would already consider quite abstract.
●
athletefailure treebelief propertychristmas
coughscaremakeseem
liberal loud darkhappy
Nouns
Verbs
Adjectives
2 3 4 5 6 7 Concreteness Rating
Figure 1 Boxplots showing the interaction between concreteness and POS for concepts in USF. The white boxes range from the first to third quartiles and the central vertical line indicates the median.
to stratify our sampling equally across four classes: (C1) concrete first item (rating > 4) with below-median concreteness difference; (C2) concrete first item (rating> 4), second item of lower concreteness and the difference being greater than the median; (C3) abstract first item (rating≤ 4) with below-median concreteness difference; and (C4) abstract first item (rating ≤ 4) with the second item of greater concreteness and the difference being greater than the median.
Final sampling. From the associated (USF) cohort of potential pairs we selected 600 noun pairs, 200 verb pairs, and 100 adjective pairs, and from the unassociated (non-USF) cohort, we sampled 66 nouns pairs, 22 verb pairs, and 11 adjective pairs. In both cases, the sampling was stratified such that, in each POS subset, each of the four concreteness classes C1−C4 was equally represented. The annotator instructions for SimLex-999 are shown in Figure 2. We did not attempt to formalize the notion of similarity, but rather introduce it via the well-understood idea of synonymy, and in contrast to association. Even if a formal characterization of similarity existed, the evidence in Section 2 suggests that the instructions would need separate cases to cover different concept types, increasing the difficulty of the rating task. Therefore, we preferred to appeal to intuition on similarity, and to verify post hoc that subjects were able to interpret and apply the informal characterization consistently for each concept type.
Immediately following the instructions in Figure 2, participants were presented with two “checkpoint” questions, one with abstract examples and one with concrete examples. In each case the participant was required to identify the most similar pair from a set of three options, all of which were associated, but only one of which was clearly similar (e.g. [bread, butter] [bread, toast] [stale, bread]). After this, the participants began rating pairs in groups of six or seven pairs by moving a slider, as shown in Figure 3.
This group size was chosen because the (relative) rating of a set of pairs implicitly requires pairwise comparisons between all pairs in that set. Therefore, larger groups would have significantly increased the cognitive load on the annotators. Another advantage of grouping was the clear break (submitting a set of ratings and moving to
the next page) between the tasks of rating adjective, noun, and verb pairs. For better inter-group calibration, from the second group onwards the last pair of the previous group became the first pair of the present group, and participants were asked to re-assign the rating previously attributed to the first pair before rating the remaining new items. As with MEN, WS-353, and RG, SimLex-999 consists of pairs of concept words together with a numerical rating. Thus, unlike in the small evaluation constructed by Huang et al. (2012), words are not rated in a phrasal or sentential context. Such meaning-in-context evaluations are motivated by a desire to disambiguate words that otherwise might be considered to have multiple senses.
We did not attempt to construct an evaluation based on meaning-in-context for several reasons. First, determining the set of senses for a given word, and then the set of contexts that represent those senses, introduces a high degree of subjectivity into the design process. Second, ensuring that a model has learned a high quality representation of a given concept would have required evaluating that concept in each of its given contexts, necessitating many more cases and a far greater annotation effort. Third, in the (infrequent) case that some concept c1 in an evaluation pair (c1, c2) is genuinely (etymologically) polysemous, c2 can provide sufficient context to disambiguate c1.9
9 This is supported by the fact that the WordNet-based methods that perform best at modeling human ratings model the similarity between concepts c1 and c2 as the minimum of all pairwise distances between the senses of c1 and the senses of c2 (Resnik 1995; Pedersen, Patwardhan, and Michelizzi 2004).
Finally, the POS grouping of pairs in the survey can also serve to disambiguate in the case that the conflicting senses of the polysemous concept are of differing POS categories. Each participant was asked to rate 20 groups of pairs on a 0–6 scale of integers (nonintegral ratings were not possible). Checkpoint multiple-choice questions were inserted at points between the 20 groups in order to ensure the participant had retained the correct notion of similarity. In addition to the checkpoint of three noun pairs presented before the first group (which contained noun pairs), checkpoint questions containing adjective pairs were inserted before the first adjective group and checkpoints of three verb pairs were inserted before the first verb group.
From the 999 evaluation pairs, 14 noun pairs, 4 verb pairs, and 2 adjective pairs were selected as a consistency set. The data set of pairs was then partitioned into 10 tranches, each consisting of 119 pairs, of which 20 were from the consistency set and the remaining 99 unique to that tranche. To reduce workload, each annotator was asked to rate the pairs in a single tranche only. The tranche itself was divided into 20 groups, with
each group consisting of 7 pairs (with the exception of the last group of the 20, which had 6). Of these seven pairs, the first pair was the last pair from the previous group, and the second pair was taken from the consistency set. The remaining pairs were unique to that particular group and tranche. The design enabled control for possible systematic differences between annotators and tranches, which could be detected by variation on the consistency set. Five hundred residents of the United States were recruited from Amazon Mechanical Turk, each with at least 95% approval rate for work on the Web service. Each participant was required to check a box confirming that he or she was a native speaker of English and warned that work would be rejected if the pattern of responses indicated otherwise. The participants were distributed evenly to rate pairs in one of the ten question tranches, so that each pair was rated by approximately 50 subjects. Participants took between 8 and 21 minutes to rate the 119 pairs across the 20 groups, together with the checkpoint questions. In order to correct for systematic differences in the overall calibration of the rating scale between respondents, we measured the average (mean) response of each rater on the consistency set. For 32 respondents, the absolute difference between this average and the mean of all such averages was greater than 1 (though never greater than 2); that is, 32 respondents demonstrated a clear tendency to rate pairs as either more or less similar than the overall rater population. To correct for this bias, we increased (or decreased) the rating of such respondents for each pair by one, except in cases where they had given the maximum rating, 6 (or minimum rating, 0). This adjustment, which ensured that the average response of each participant was within one of the mean of all respondents on the consistency set, resulted in a small increase to the inter-rater agreement on the data set as a whole.
After controlling for systematic calibration differences, we imposed three conditions for the responses of a rater to be included in the final data collation. First, the average pairwise Spearman correlation of responses with all other responses for a participant could not be more than one standard deviation below the mean of all such averages. Second, the increase in inter-rater agreement when a rater was excluded from the analysis needed to be smaller than at least 50 other raters (i.e., 10% of raters were excluded on this criterion). Third, we excluded the six participants who got one or more of the checkpoint questions wrong. A total of 99 participants were excluded based on one or more of these conditions, but no more than 16 from any one tranche (so that each pair in the final data set was rated by a minimum of 36 raters). Finally, we computed average (mean) scores for each pair, and transformed all scores linearly from the interval [0, 6] to the interval [0, 10].","4 analysis of the data set :In this section we analyze the responses of the SimLex-999 annotators and the resulting ratings. First, by considering inter-annotator agreement, we examine the consistency with which annotators were able to apply the characterization of similarity, outlined in the instructions for the range of concept types in SimLex-999. Second, we verify that a
valid notion of similarity was understood by the annotators, in that they were able to accurately separate similarity from association. As in previous annotation or data collection for computational semantics (Padó, Padó, and Erk 2007; Reisinger and Mooney 2010a; Silberer and Lapata 2014) we computed the inter-rater agreement as the average of pairwise Spearman ρ correlations between the ratings of all respondents. Overall agreement was ρ = 0.67. This compares favorably with the agreement on WS-353 (ρ = 0.61 using the same method). The design of the MEN rating system precludes a conventional calculation of inter-rater agreement (Bruni et al. 2012b). However, two of the creators of MEN who independently rated the data set achieved an agreement of ρ = 0.68.10
The SimLex-999 inter-rater agreement suggests that participants were able to understand the (single) characterization of similarity presented in the instructions and to apply it to concepts of various types consistently. This conclusion was supported by inspection of the brief feedback offered by the majority of annotators in a final text field in the questionnaire: 78% expressed sentiment that the test was clear, easy to complete, or some similar sentiment.
Interestingly, as shown in Figure 4 (left), agreement was not uniform across the concept types. Contrary to what might be expected given established concreteness effects (Paivio 1991), we observed not only higher inter-rater agreement but also less per-pair variability for abstract rather than concrete concepts.11
Strikingly, the highest inter-rater consistency and lowest per-pair variation (defined as the inverse of the standard deviation of all ratings for that pair) was observed on adjective pairs. Although we are unsure exactly what drives this effect, a possible cause
10 Reported at http://clic.cimec.unitn.it/~elia.bruni/MEN. It is reasonable to assume that actual agreement on MEN may be somewhat lower than 0.68, given the small sample size and the expertise of the raters. 11 Per-pair variability was measured by calculating the standard deviation of responses for each pair, and averaging these scores across the pairs of each concept type.
is that many pairs of adjectives in SimLex-999 cohabit a single salient, one-dimensional scale ( freezing > cold > warm > hot). This may be a consequence of the fact that many pairs in SimLex-999 were selected (from USF) to have a degree of association. On inspection, pairs of nouns and verbs in SimLex-999 do not appear to occupy scales in the same way, possibly because concepts of these POS categories come to be associated via a more diverse range of relations. It seems plausible that humans are able to estimate the similarity of scale-based concepts more consistently than pairs of concepts related in a less uni-dimensional fashion.
Regardless of cause, however, the high agreement on adjectives is a satisfactory property of SimLex-999. Adjectives exhibit various aspects of lexical semantics that have proved challenging for computational models, including antonymy, polarity (Williams and Anand 2009), and sentiment (Wiebe 2000). To approach the high level of human confidence on the adjective pairs in SimLex-999, it may be necessary to focus particularly on developing automatic ways to capture these phenomena. Inspection of the SimLex-999 ratings indicated that pairs were indeed evaluated according to similarity rather than association. Table 2 includes examples that demonstrate a clear dissociation between the two semantic relations.
To verify this effect quantitatively, we recruited 100 additional participants to rate the WS-353 pairs, but following the SimLex-999 instructions and question format. As shown in Fig 5(a), there were clear differences between these new ratings and the original WS-353 ratings. In particular, a high proportion of pairs was given a lower rating by subjects following the SimLex-999 instructions than those following the WS-353 guidelines: The mean SimLex rating was 4.07 compared with 5.91 for WS-353.
This was consistent with our expectations that pairs of associated but dissimilar concepts would receive lower ratings based on the SimLex-999 than on the WS-353 instructions, whereas pairs that were both associated and similar would receive similar ratings in both cases. To confirm this, we compared the WS-353 and SimLex-999based ratings on the subsets WS-Rel and WS-Sim, which were hand-sorted by Agirre
et al. (2009) to include pairs connected by association (and not similarity) and those connected by similarity (but possibly also association), respectively.
As shown in Figure 5(b–c), the correlation between the SimLex-999-based and WS353 ratings was notably higher (ρ = 0.73) on the WS-Sim subset than the WS-Rel subset (ρ = 0.38). Specifically, the tendency of subjects following the SimLex-999 instructions to assign lower ratings than those following the WS-353 instructions was far more pronounced for pairs in WS-Sim (Figure 5(b)) than for those in WS-Rel (Figure 5(c)). This observation suggests that the associated but dissimilar pairs in WS-353 were an important driver of the overall lower mean for SimLex-999-based ratings, and thus provide strong evidence that the SimLex-999 instructions do indeed enable subjects to distinguish similarity from association effectively. We have established the validity of similarity as a notion understood by human raters and distinct from association. However, much theoretical semantics focuses on relations between words or concepts that are finer-grained than similarity and association. These include meronymy (a part to its whole, e.g., blade–knife), hypernymy (a category concept to a member of that category, e.g., animal–dog), and cohyponymy (two members of the same implicit category, e.g., the pair of animals dog–cat) (Cruse 1986). Beyond theoretical interest, these relations can have practical relevance. For instance, hypernymy can form the basis of semantic entailment and therefore textual inference: The proposition a cat is on the table entails that an animal is on the table precisely because of the hypernymy relation from animal to cat.
We chose not to make these finer-grained relations the basis of our evaluation for several reasons. At present, detecting relations such as hypernymy using distributional methods is challenging, even when supported by supervised classifiers with access to labeled pairs (Levy et al. 2015). Such a designation can seem to require specific
world-knowledge (is a snale a reptile?), can be gradual, as evidenced by typicality effects (Rosch, Simpson, and Miller 1976), or simply highly subjective. Moreover, a fine-grained relation R will only be attested (to any degree) between a small subset of all possible word pairs, whereas similarity can in theory be quantified for any two words chosen at random. We thus considered a focus on fine-grained semantic relations to be less appropriate for a general-purpose evaluation of representation quality.
Nevertheless, post hoc analysis of the SimLex annotator responses and fine-grained relation classes, as defined by lexicographers, yields further interesting insights into the nature of both similarity and association. Of the 999 word pairs in SimLex, 382 are also connected by one of the common finer-grained semantic relations in WordNet. For each of these relations, Figure 6 shows the average similarity rating and average USF free association score for all pairs that exhibit that relation.
In cases where a relationship of hypernymy/hyponymy exists between the words in a pair (not necessarily immediate : 1 hypernym, 2 hypernym, etc.) similarity and association coincide. Hyper/hyponym pairs that are separated by fewer levels in the WordNet hierarchy are both more strongly associated and rated as more similar. However, there are also interesting discrepancies between similarity and association. Unsurprisingly, pairs that are classed as synonyms in WordNet (i.e., having at least one sense in some common synset) are rated as more similar than pairs of any other relation type by SimLex annotators. In contrast, antonyms are the most strongly associated word pairs among these finer-grained relations. Further, pairs consisting of a meronym and holonym (part and whole) are comparatively strongly associated but not judged to be similar.
The analysis also highlights a case that can be particularly problematic when rating similarity: cohyponyms, or members of the same salient category (such as knife and fork). We gave no specific guidelines for how to rate such pairs in the SimLex annotator instructions, and whether they are considered similar or not seems to be a matter of perspective. On one hand, their membership of a common category could make them appear similar, particularly if the category is relatively specific. On the other hand, in
the case of knife and fork, for instance, the underlying category cultery might provide a backdrop against which the differences of distinct members become particularly salient.","5 evaluating models with simlex-999 :In this section, we demonstrate the applicability of SimLex-999 by analyzing the performance of various distributional semantic models in estimating the new ratings. The models were selected to cover the main classes of representation learning architectures (Baroni, Dinu, and Kruszewski 2014): Vector space co-occurrence (counting) models and NLMs (Bengio et al. 2003). We first show that SimLex-999 is notably more difficult for state-of-the-art models to estimate than existing gold standards. We then conduct more focused analyses on the various concept subsets defined in SimLex-999, exploring possible causes for the comparatively low performance of current models and, in turn, demonstrating how SimLex-999 can be applied to investigate such questions. Collobert & Weston. Collobert and Weston (2008) apply the architecture of an NLM to learn a word representations vw for each word w in some corpus vocabulary V. Each sentence s in the input text is represented by a matrix containing the vector representations of the words in s in order. The model then computes output scores f (s) and f (sw), where sw denotes an “incorrect” sentence created from s by replacing its last word with some other word w from V. Training involves updating the parameters of the function f and the entries of the vector representations vw such that f (s) is larger than f (sw) for any w in V, other than the correct final word of s. This corresponds to minimizing the sum of the following sentence objectives Cs over all sentences in the input corpus, which is achieved via (mini-batch) stochastic gradient descent:
Cs = ∑ w∈V max(0, 1− f (s) + f (sw))
The relatively low-dimension, dense (vector) representations learned by this model and the other NLMs introduced in this section are sometimes referred to as embeddings (Turian, Ratinov, and Bengio 2010). Collobert and Weston (2008) train their models on 852 million words of text from a 2007 dump of Wikipedia and the RCV1 Corpus (Lewis et al. 2004) and use their embeddings to achieve state-of-the-art results on a variety of NLP tasks. We downloaded the embeddings directly from the authors’ Web page.12
Huang et al. Huang et al. (2012) present a NLM that learns word embeddings to maximize the likelihood of predicting the last word in a sentence s based on (i) the previous words in that sentence (local context, as with Collobert and Weston [2008]) and (ii) the document d in which that word occurs (global context). As with Collobert and Weston (2008), the model represents input sentences as a matrix of word embeddings. In addition, it represents documents in the input corpus as single-vector averages over all word embeddings in that document. It can then compute scores g(s, d) and g(sw, d), whereas before sw is a sentence with an “incorrect” randomly selected last word.
12 http://ml.nec-labs.com/senna/.
Training is again by stochastic gradient descent, and corresponds to minimizing the sum of the sentence objectives Cs,d over all of the sentences in the corpus:
Cs,d = ∑ w∈V max(0, 1− g(s, d) + g(sw, d))
The combination of local and global contexts in the objective encourages the final word embeddings to reflect aspects of both the meaning of nearby words and of the documents in which those words appear. When learning from 990M words of Wikipedia text, Huang et al. report a Spearman correlation of ρ = 71.3 between the cosine similarity of their model embeddings and the WS-353 scores, which constitutes state-of-the-art performance for a NLM model on that data set. We downloaded these embeddings from the authors’ Web page.13
Mikolov et al. Mikolov et al. (2013a) present an architecture that learns word embeddings similar to those of standard NLMs but with no nonlinear hidden layer (resulting in a simpler scoring function). This enables faster representation learning for large vocabularies. Despite this simplification, the embeddings achieve state-of-theart performance on several semantic tasks including sentence completion and analogy modeling (Mikolov et al. 2013a, 2013b).
For each word type w in the vocabulary V, the model learns both a “targetembedding” rw ∈ Rd and a “context-embedding” r̂w ∈ Rd such that, given a target word, its ability to predict nearby context words is maximized. The probability of seeing context word c given target w is defined as:
p(c|w) = e r̂c·rw∑
v∈V er̂v·rw
The model learns from a set of (target-word, context-word) pairs, extracted from a corpus of sentences as follows. In a given sentence s (of length N), for each position n ≤ N, each word wn is treated in turn as a target word. An integer t(n) is then sampled from a uniform distribution on {1, . . . k}, where k > 0 is a predefined maximum contextwindow parameter. The pair tokens {(wn, wn+j) : −t(n) ≤ j ≤ t(n), wi ∈ s} are then appended to the training data. Thus, target/context training pairs are such that (i) only words within a k-window of the target are selected as context words for that target, and (ii) words closer to the target are more likely to be selected than those further away.
The training objective is then to maximize the log probability T, defined here, across all such examples from s, and then across all sentences in the corpus. This is achieved by stochastic gradient descent.
T = 1N N∑ n=1 ∑ −t(n)≤j≤t(n),j 6=0 log(p(wn+j|wn))
As with other NLMs, Mikolov et al.’s model captures conceptual semantics by exploiting the fact that words appearing in similar linguistic contexts are likely to
13 www.socher.org/index.php/Main/ImprovingWordRepresentationsViaGlobalContextAndMultiple WordPrototypes.
have similar meanings. Informally, the model adjusts its embeddings to increase the probability of observing the training corpus. Because this probability increases with p(c|w), and p(c|w) increases with the dot product r̂c · rw, the updates have the effect of moving each target-embedding incrementally “closer” to the context-embeddings of its collocates. In the target-embedding space, this results in embeddings of concept words that regularly occur in similar contexts moving closer together.
We use the author’s Word2vec software in order to train their model and use the target embeddings in our evaluations. We experimented with embeddings of dimension 100, 200, 300, 400, and 500 and found that 200 gave the best performance on both WS-353 and SimLex-999.
Vector Space Model (VSM). As an alternative to NLMs, we constructed a vector space model following the guidelines for optimal performance outlined by Kiela and Clark (2014). After extracting the 2,000 most frequent word tokens in the corpus that are not in a common list of stopwords14 as features, we populated a matrix of co-occurrence counts with a row for each of the concepts in some pair in our evaluation sets, and a column for each of the features. Co-occurrence was counted within a specified window size, although never across a sentence boundary. This resulting matrix was then weighted according to Pointwise Mutual Information (PMI) (Recchia and Jones 2009). The rows of the resulting matrix constitute the vector representations of the concepts.
SVD. As proposed initially in Landauer and Dumais (1997), we also experimented with models in which SVD (Golub and Reinsch 1970) is applied to the PMI-weighted VSM matrix, reducing the dimension of each concept representation to 300 (which yielded best results after experimenting, as before, with 100–500 dimension vectors).
For each model described in this section, we calculate similarity as the cosine similarity between the (vector) representations learned by that model. In experimenting with different models on SimLex-999, we aimed to answer the following questions: (i) How well do the established models perform on SimLex-999 versus on existing gold standards? (ii) Are any observed differences caused by the potential of different models to measure similarity vs. association? (iii) Are there interesting differences in ability of models to capture similarity between adjectives vs. nouns vs. verbs? (iv) In this case, are the observed differences driven by concreteness, and its interaction with POS, or are other factors also relevant?
Overall Performance on SimLex-999. Figure 7 shows the performance of the NLMs on SimLex-999 versus on comparable data sets, measured by Spearman’s ρ correlation. All models estimate the ratings of MEN and WS-353 more accurately than SimLex-999. The Huang et al. (2012) model performs well on WS-353,15 but is not very robust to changes in evaluation gold standard, and performs worst of all the models on SimLex-999. Given the focus of the WS-353 ratings, it is tempting to explain this by concluding that the
14 Taken from the Python Natural Language Toolkit (Bird 2006). 15 This score, based on embeddings downloaded from the authors’ webpage, is notably lower than the score
reported in Huang et al. (2012), mentioned in Section 5.1.
global context objective leads the Huang et al. (2012) model to focus on association rather than similarity. However, the true explanation may be less simple, since the Huang et al. (2012) model performs weakly on the association-based MEN data set. The Collobert and Weston (2008) model is more robust across WS-353 and MEN, but still does not match the performance of the Mikolov et al. (2013a) model on SimLex-999.
Figure 8 compares the best performing NLM model (Mikolov et al. 2013a) with the VSM and SVD models.16 In contrast to recent results that emphasize the superiority of NLMs over alternatives (Baroni, Dinu, and Kruszewski 2014), we observed no clear advantage for the NLM over the VSM or SVD when considering the associationbased gold standards WS-353 and MEN together. While the NLM is the strongest performer on WS-353, SVD is the strongest performer on MEN. However, the NLM model performs notably better than the alternatives at modeling similarity, as measured by SimLex-999.
Comparing all models in Figures 7 and 8 suggests that SimLex-999 is notably more challenging to model than the alternative data sets, with correlation scores ranging from 0.098 to 0.414. Thus, even when state-of-the-art models are trained for several days on massive text corpora,17 their performance on SimLex-999 is well below the interannotator agreement (Figure 7). This suggests that there is ample scope for SimLex-999 to guide the development of improved models.
Modeling Similarity vs. Association. The comparatively low performance of NLM, VSM, and SVD models on SimLex-999 compared with MEN and WS-353 is consistent with our hypothesis that modeling similarity is more difficult than modeling association. Indeed, given that many strongly associated but dissimilar pairs, such as [coffee, cup], are likely to have high co-occurrence in the training data, and that all models infer connections between concepts from linguistic co-occurrence in some form or another,
16 We conduct this comparison on the smaller RCV1 Corpus (Lewis et al. 2004) because training the VSM and SVD models is comparatively slow. 17 Training times reported by Huang et al. (2012) and for Collobert and Weston (2008) at http://ronan.collobert.com/senna/.
it seems plausible that models may overestimate the similarity of such pairs because they are “distracted” by association.
To test this hypothesis more precisely, we compared the performance of models on the whole of SimLex-999 versus its 333 most associated pairs (according to the USF free association scores). Importantly, pairs in this strongly associated subset still span the full range of possible similarity scores (min similarity = 0.23 [shrink, grow], max similarity = 9.80 [vanish, disappear]).
As shown in Figure 9, all models performed worse when the evaluation was restricted to pairs of strongly associated concepts, which was consistent with our hypothesis. The Collobert and Weston (2008) model was better than the Huang et al. (2012) model at estimating similarity in the face of high association. This is not entirely surprising given the global-context objective in the latter model, which may have encouraged more association-based connections between concepts. The Mikolov et al. model, however, performed notably better than both other NLMs. Moreover, this superiority is proportionally greater when evaluating on the most associated pairs only (as
indicated by the difference between the red and gray bars), suggesting that the improvement is driven at least in part by an increased ability to “distinguish” similarity from association.
To understand better how the architecture of models captures information pertinent to similarity modeling, we performed two additional experiments using SimLex-999. These comparisons were also motivated by the hypotheses, made in previous studies and outlined in Section 2.1.2, that both dependency-informed input and smaller context windows encourage models to capture similarity rather than association.
We tested the first hypothesis using the embeddings of Levy and Goldberg (2014), whose model extends the Mikolov et al. (2013a) model so that target-context training instances are extracted based on dependency-parsed rather than simple running text. As illustrated in Figure 9, the dependency-based embeddings outperform the original (running text) embeddings trained on the same corpus. Moreover, the comparatively large increase in the red bar compared to the gray bar suggests that an important part of the improvement of the dependency-based model derives from a greater ability to discern similarity from association.
Our comparisons provided less support for the second (window size) hypothesis. As shown in Figure 10, there is a negligible improvement in the performance of the model when the window size is reduced from 10 to 2. However, for the SVD model we observed the converse. The SVD model with window size 10 slightly outperforms the SVD model with window 2, and this improvement is quite pronounced on the most associated pairs in SimLex-999.
Learning Concepts of Different POS. Given the theoretical likelihood of variation in model performance across POS categories noted in Section 2.2, we evaluated the Mikolov et al. (2013a), VSM, and SVD models on the subsets of SimLex-999 containing adjective, noun, and verb concept pairs.
The analyses yield two notable conclusions, as shown in Figure 11. First, perhaps contrary to intuition, all models estimate the similarity of adjectives better than other concept categories. This aligns with the (also unexpected) observation that humans rate the similarity of adjectives more consistently and with more agreement than other parts of speech (see the dashed lines). However, the parallels between human raters and the models do not extend to verbs and nouns; verb similarity is rated more consistently
than noun similarity by humans, but models estimate these ratings more accurately for nouns than for verbs.
To better understand the linguistic information exploited by models when acquiring concepts of different POS, we also computed performance on the POS subsets of SimLex-999 of the dependency-based model of Levy and Goldberg (2014) and the standard skipgram model, in which linguistic contexts are encoded as simple bagsof-words (BOW) (Mikolov et al. (2013a) [trained on the same Wikipedia text]). As shown in Figure 12, dependency-aware contexts yield the largest improvements for capturing verb similarity. This aligns with the cognitive theory of verbs as relational concepts (Markman and Wisniewski 1997) whose meanings rely on their interaction with (or dependency on) other words or concepts. It is also consistent with research on the automatic acquisition of verb semantics, in which syntactic features have proven particularly important (Sun, Korhonen, and Krymolowski 2008). Although a deeper exploration of these effects is beyond the scope of this work, this preliminary analysis
again highlights how the word classes integrated into SimLex-999 are pertinent to a range of questions concerning lexical semantics.
Learning Concrete and Abstract Concepts. Given the strong interdependence between POS and conceptual concreteness (Figure 1), we aimed to explore whether the variation in model performance on different POS categories was in fact driven by an underlying effect of concreteness. To do so, we ranked each pair in the SimLex-999 data set according to the sum of the concreteness of the two words, and compared performance of models on the most concrete and least concrete quartiles according to this ranking (Figure 13).
Interestingly, the performance of models on the most abstract and most concrete pairs suggests that the distinction characterized by concreteness is at least partially independent of POS. Specifically, while the Mikolov et al. model was the highest performer of all POS categories, its performance was worse than both the simple VSM and SVD models (of window size 10) on the most concrete concept pairs.
This finding supports the growing evidence for systematic differences in representation and/or similarity operations between abstract and concrete concepts (Hill, Kiela, and Korhonen 2013), and suggests that at least part of these concreteness effects are independent of POS. In particular, it appears that models built from underlying vectors of co-occurrence counts, such as VSMs and SVD, are better equipped to capture the semantics of concrete entities, whereas the embeddings learned by NLMs can better capture abstract semantics.","6 conclusion :Although the ultimate test of semantic models should be their utility in downstream applications, the research community can undoubtedly benefit from ways to evaluate the general quality of the representations learned by such models, prior to their integration in any particular system. We have presented SimLex-999, a gold standard resource for the evaluation of semantic representations containing similarity ratings of word pairs of different POS categories and concreteness levels.
The development of SimLex-999 was principally motivated by two factors. First, as we demonstrated, several existing gold standards measure the ability of models to capture association rather than similarity, and others do not adequately test their ability to discriminate similarity from association. This is despite the many potential applications for accurate similarity-focused representation learning models. Analysis of the ratings of the 500 SimLex-999 annotators showed that subjects can consistently quantify similarity, as distinct from association, and apply it to various concept types, based on minimal intuitive instructions.
Second, as we showed, state-of-the-art models trained solely on running-text corpora have now reached or surpassed the human agreement ceiling on WordSim-353 and MEN, the most popular existing gold standards, as well as on RG and WS-Sim. These evaluations may therefore have limited use in guiding or moderating future improvements to distributional semantic models. Nevertheless, there is clearly still room for improvement in terms of the use of distributional models in functional applications. We therefore consider the comparatively low performance of state-of-the-art models on SimLex-999 to be one of its principal strengths. There is clear room under the inter-rating ceiling to guide the development of the next generation of distributional models.
We conducted a brief exploration of how models might improve on this performance, and verified the hypotheses that models trained on dependency-based input capture similarity more effectively than those trained on running-text input. The evidence that smaller context windows are also beneficial for similarity models was mixed, however. Indeed, we showed that the optimal window size depends on both the general model architecture and the part-of-speech and concreteness of the target concepts.
Our analysis of these hypotheses illustrates how the design of SimLex-999— covering a principled set of concept categories and including meta-information on concreteness and free-association strength—enables fine-grained analyses of the performance and parameterization of semantic models. However, these experiments only scratch the surface in terms of the possible analyses. We hope that researchers will adopt the resource as a robust means of answering a diverse range of questions pertinent to similarity modeling, distributional semantics, and representation learning in general.
In particular, for models to learn high-quality representations for all linguistic concepts, we believe that future work must uncover ways to explicitly or implicitly infer “deeper,” more general, conceptual properties such as intentionality, polarity, subjectivity, or concreteness (Gershman and Dyer 2014). However, although improving corpusbased models in this direction is certainly realistic, models that learn exclusively via the linguistic modality may never reach human-level performance on evaluations such as SimLex-999. This is because much conceptual knowledge, and particularly that which underlines similarity computations for concrete concepts, appears to be grounded in the perceptual modalities as much as in language (Barsalou et al. 2003).
Whatever the means by which the improvements are achieved, accurate conceptlevel representation is likely to constitute a necessary first step towards learning informative, language-neutral phrasal and sentential representations. Such representations would be hugely valuable for fundamental NLP applications such as language understanding tools and machine translation.
Distributional semantics aims to infer the meaning of words based on the company they keep (Firth 1957). However, although words that occur together in text often have associated meanings, these meanings may be very similar or indeed very different. Thus, possibly excepting the population of Argentina, most people would agree that, strictly speaking, Maradona is not synonymous with football (despite their high rating of 8.62 in WordSim-353). The challenge for the next generation of distributional models
may therefore be to infer what is useful from the co-occurrence signal and to overlook what is not. Perhaps only then will models capture most, or even all, of what humans know when they know how to use a language.",,,"We present SimLex-999, a gold standard resource for evaluating distributional semantic models that improves on existing resources in several important ways. First, in contrast to gold standards such as WordSim-353 and MEN, it explicitly quantifies similarity rather than association or relatedness so that pairs of entities that are associated but not actually similar (Freud, psychology) have a low rating. We show that, via this focus on similarity, SimLex-999 incentivizes the development of models with a different, and arguably wider, range of applications than those which reflect conceptual association. Second, SimLex-999 contains a range of concrete and abstract adjective, noun, and verb pairs, together with an independent rating of concreteness and (free) association strength for each pair. This diversity enables fine-grained analyses of the performance of models on concepts of different types, and consequently greater insight into how architectures can be improved. Further, unlike existing gold standard evaluations, for which automatic approaches have reached or surpassed the inter-annotator agreement ceiling, state-ofthe-art models perform well below this ceiling on SimLex-999. There is therefore plenty of scope for SimLex-999 to quantify future improvements to distributional semantic models, guiding the development of the next generation of representation-learning architectures.","[{""affiliations"": [], ""name"": ""Felix Hill""}, {""affiliations"": [], ""name"": ""Roi Reichart""}, {""affiliations"": [], ""name"": ""Anna Korhonen""}]",SP:bc23a2bd2507750d9ea8f0a4937c6bc8162c549c,"[{""authors"": [""Agirre"", ""Eneko"", ""Enrique Alfonseca"", ""Keith Hall"", ""Jana Kravalova"", ""Marius Pa\u015fca"", ""Aitor Soroa.""], ""title"": ""A study on similarity and relatedness using distributional and Wordnet-based approaches"", ""venue"": ""Proceedings"", ""year"": 2009}, {""authors"": [""Alfonseca"", ""Enrique"", ""Suresh Manandhar.""], ""title"": ""Extending a lexical ontology by a combination of distributional semantics signatures"", ""venue"": ""G. Schrieber et al., Knowledge Engineering and Knowledge"", ""year"": 2002}, {""authors"": [""Andrews"", ""Mark"", ""Gabriella Vigliocco"", ""David Vinson.""], ""title"": ""Integrating experiential and distributional data to learn semantic representations"", ""venue"": ""Psychological Review, 116(3):463."", ""year"": 2009}, {""authors"": [""Bansal"", ""Mohit"", ""Kevin Gimpel"", ""Karen Livescu.""], ""title"": ""Tailoring continuous word representations for dependency parsing"", ""venue"": ""Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Baroni"", ""Marco"", ""Georgiana Dinu"", ""Germ\u00e1n Kruszewski.""], ""title"": ""Don\u2019t count, predict! A systematic comparison of context-counting vs"", ""venue"": ""context-predicting semantic vectors. In Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Baroni"", ""Marco"", ""Alessandro Lenci.""], ""title"": ""Distributional memory: A general framework for corpus-based semantics"", ""venue"": ""Computational Linguistics, 36(4):673\u2013721."", ""year"": 2010}, {""authors"": [""Barsalou"", ""Lawrence W."", ""W. Kyle Simmons"", ""Aron K. Barbey"", ""Christine D. Wilson.""], ""title"": ""Grounding conceptual knowledge in modality-specific systems"", ""venue"": ""Trends in Cognitive Sciences, 7(2):84\u201391."", ""year"": 2003}, {""authors"": [""Beltagy"", ""Islam"", ""Katrin Erk"", ""Raymond Mooney.""], ""title"": ""Semantic parsing using distributional semantics and probabilistic logic"", ""venue"": ""ACL 2014 Workshop on Semantic Parsing."", ""year"": 2014}, {""authors"": [""Bengio"", ""Yoshua"", ""R\u00e9jean Ducharme"", ""Pascal Vincent"", ""Christian Jauvin.""], ""title"": ""A neural probabilistic language model"", ""venue"": ""The Journal of Machine Learning Research, 3:1137\u20131155."", ""year"": 2003}, {""authors"": [""Bernardi"", ""Raffaella"", ""Georgiana Dinu"", ""Marco Marelli"", ""Marco Baroni.""], ""title"": ""A relatedness benchmark to test the role of determiners in compositional distributional semantics"", ""venue"": ""Proceedings of"", ""year"": 2013}, {""authors"": [""Biemann"", ""Chris.""], ""title"": ""Ontology learning from text: A survey of methods"", ""venue"": ""LDV Forum, 20(2):75\u201393."", ""year"": 2005}, {""authors"": [""Bird"", ""Steven.""], ""title"": ""Nltk: the natural language toolkit"", ""venue"": ""Proceedings of the COLING/ACL on Interactive Presentation sessions, pages 69\u201372, Sydney."", ""year"": 2006}, {""authors"": [""Bruni"", ""Elia"", ""Gemma Boleda"", ""Marco Baroni"", ""Nam-Khanh Tran.""], ""title"": ""Distributional semantics in technicolor"", ""venue"": ""Proceedings of ACL, Jeju Island."", ""year"": 2012}, {""authors"": [""Bruni"", ""Elia"", ""Jasper Uijlings"", ""Marco Baroni"", ""Nicu Sebe.""], ""title"": ""Distributional semantics with eyes: Using image analysis to improve computational representations of word meaning"", ""venue"": ""Proceedings of the 20th"", ""year"": 2012}, {""authors"": [""Budanitsky"", ""Alexander"", ""Graeme Hirst.""], ""title"": ""Evaluating Wordnet-based measures of lexical semantic relatedness"", ""venue"": ""Computational Linguistics, 32(1):13\u201347."", ""year"": 2006}, {""authors"": [""Cimiano"", ""Philipp"", ""Andreas Hotho"", ""Steffen Staab.""], ""title"": ""Learning concept hierarchies from text corpora using formal concept analysis"", ""venue"": ""J. Artif. Intell. Res. (JAIR), 24:305\u2013339."", ""year"": 2005}, {""authors"": [""R. Collobert"", ""J. Weston.""], ""title"": ""A unified architecture for natural language processing: Deep neural networks with multitask learning"", ""venue"": ""International Conference on Machine Learning, ICML, Helsinki."", ""year"": 2008}, {""authors"": [""Cruse"", ""D. Alan.""], ""title"": ""Lexical semantics"", ""venue"": ""Cambridge University Press."", ""year"": 1986}, {""authors"": [""Cunningham"", ""Hamish.""], ""title"": ""Information extraction, automatic"", ""venue"": ""Encyclopedia of language and linguistics, pages 665\u2013677."", ""year"": 2005}, {""authors"": [""Fellbaum"", ""Christiane""], ""title"": ""WordNet"", ""venue"": ""Wiley Online Library."", ""year"": 1998}, {""authors"": [""Finkelstein"", ""Lev"", ""Evgeniy Gabrilovich"", ""Yossi Matias"", ""Ehud Rivlin"", ""Zach Solan"", ""Gadi Wolfman"", ""Eytan Ruppin.""], ""title"": ""Placing search in context: The concept revisited"", ""venue"": ""Proceedings of the"", ""year"": 2001}, {""authors"": [""J.R. Firth""], ""title"": ""Papers in Linguistics 1934\u20131951"", ""venue"": ""Oxford University Press."", ""year"": 1957}, {""authors"": [""Gentner"", ""Dedre.""], ""title"": ""On relational meaning: The acquisition of verb meaning"", ""venue"": ""Child Development, pages 988\u2013998."", ""year"": 1978}, {""authors"": [""Gentner"", ""Dedre.""], ""title"": ""Why verbs are hard to learn"", ""venue"": ""Action meets word: How Children Learn Verbs, pages 544\u2013564. 692"", ""year"": 2006}, {""authors"": [""Gershman"", ""Anatole"", ""Yulia Tsvetkov"", ""Leonid Boytsov"", ""Eric Nyberg"", ""Chris Dyer.""], ""title"": ""Metaphor detection with cross-lingual model transfer"", ""venue"": ""Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Golub"", ""Gene H."", ""Christian Reinsch.""], ""title"": ""Singular value decomposition and least squares solutions"", ""venue"": ""Numerische Mathematik, 14(5):403\u2013420."", ""year"": 1970}, {""authors"": [""Griffiths"", ""Thomas L."", ""Mark Steyvers"", ""Joshua B. Tenenbaum.""], ""title"": ""Topics in semantic representation"", ""venue"": ""Psychological Review, 114(2):211."", ""year"": 2007}, {""authors"": [""Haghighi"", ""Aria"", ""Percy Liang"", ""Taylor Berg-Kirkpatrick"", ""Dan Klein.""], ""title"": ""Learning bilingual lexicons from monolingual corpora"", ""venue"": ""Proceedings of ACL 2008, Columbus, OH."", ""year"": 2008}, {""authors"": [""Hassan"", ""Samer"", ""Rada Mihalcea.""], ""title"": ""Semantic relatedness using salient semantic analysis"", ""venue"": ""AAAI, San Francisco, CA."", ""year"": 2011}, {""authors"": [""Hatzivassiloglou"", ""Vasileios"", ""Judith L. Klavans"", ""Melissa L. Holcombe"", ""Regina Barzilay"", ""Min-Yen Kan"", ""Kathleen McKeown.""], ""title"": ""Simfinder: A flexible clustering tool for summarization"", ""venue"": ""In"", ""year"": 2001}, {""authors"": [""He"", ""Xiaodong"", ""Mei Yang"", ""Jianfeng Gao"", ""Patrick Nguyen"", ""Robert Moore.""], ""title"": ""Indirect-HMM-based hypothesis alignment for combining outputs from machine translation systems"", ""venue"": ""Proceedings"", ""year"": 2008}, {""authors"": [""Hill"", ""Felix"", ""Douwe Kiela"", ""Anna Korhonen.""], ""title"": ""Concreteness and corpora: A theoretical and practical analysis"", ""venue"": ""CMCL 2013, page 75, Sofia."", ""year"": 2013}, {""authors"": [""Hill"", ""Felix"", ""Anna Korhonen"", ""Christian Bentz.""], ""title"": ""A quantitative empirical analysis of the abstract/concrete distinction"", ""venue"": ""Cognitive Science, 38(1):162\u2013177."", ""year"": 2014}, {""authors"": [""Hill"", ""Felix"", ""Roi Reichart"", ""Anna Korhonen.""], ""title"": ""Multi-modal models for concrete and abstract concept meaning"", ""venue"": ""Transactions of the Association for Computational Linguistics (TACL), 2:285\u2013296."", ""year"": 2014}, {""authors"": [""Huang"", ""Eric H."", ""Richard Socher"", ""Christopher D. Manning"", ""Andrew Y. Ng.""], ""title"": ""Improving word representations via global context and multiple word prototypes"", ""venue"": ""Proceedings of ACL, pages 873\u2013882,"", ""year"": 2012}, {""authors"": [""Kiela"", ""Douwe"", ""Stephen Clark.""], ""title"": ""A systematic study of semantic vector space model parameters"", ""venue"": ""Proceedings of the 2nd Workshop on Continuous Vector"", ""year"": 2014}, {""authors"": [""Kiela"", ""Douwe"", ""Felix Hill"", ""Anna Korhonen"", ""Stephen Clark.""], ""title"": ""Improving multi-modal representations using image dispersion: Why less is sometimes more"", ""venue"": ""Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Landauer"", ""Thomas K."", ""Susan T. Dumais.""], ""title"": ""A solution to Plato\u2019s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge"", ""venue"": ""Psychological Review,"", ""year"": 1997}, {""authors"": [""Leech"", ""Geoffrey"", ""Roger Garside"", ""Michael Bryant.""], ""title"": ""Claws4: The tagging of the British National Corpus"", ""venue"": ""Proceedings of COLING, pages 622\u2013628, Kyoto."", ""year"": 1994}, {""authors"": [""Levy"", ""Omer"", ""Yoav Goldberg.""], ""title"": ""Dependency-based word embeddings"", ""venue"": ""Proceedings of ACL, volume 2."", ""year"": 2014}, {""authors"": [""Levy"", ""Omer"", ""Steffen Remus"", ""Chris Biemann"", ""Idol Dagan""], ""title"": ""Do supervised distributional methods really learn lexical inference relations"", ""venue"": ""Proceedings of NAACL,"", ""year"": 2015}, {""authors"": [""Lewis"", ""David D."", ""Yiming Yang"", ""Tony G. Rose"", ""Fan Li.""], ""title"": ""Rcv1: A new benchmark collection for text categorization research"", ""venue"": ""The Journal of Machine Learning Research, 5:361\u2013397."", ""year"": 2004}, {""authors"": [""Li"", ""Changliang"", ""Bo Xu"", ""Gaowei Wu"", ""Xiuying Wang"", ""Wendong Ge"", ""Yan Li.""], ""title"": ""Obtaining better word representations via language transfer"", ""venue"": ""A. Gelbukh, editor, Computational Linguistics and Intelligent"", ""year"": 2014}, {""authors"": [""Li"", ""Mu"", ""Yang Zhang"", ""Muhua Zhu"", ""Ming Zhou.""], ""title"": ""Exploring distributional similarity based models for query spelling correction"", ""venue"": ""Proceedings of ALC, pages 1025\u20131032."", ""year"": 2006}, {""authors"": [""Luong"", ""Minh-Thang"", ""Richard Socher"", ""Christopher D. Manning.""], ""title"": ""Better word representations with recursive neural networks for morphology"", ""venue"": ""CoNLL-2013, page 104, Sofia."", ""year"": 2013}, {""authors"": [""Markman"", ""Arthur B."", ""Edward J. Wisniewski.""], ""title"": ""Similar and different: The differentiation of basic-level categories"", ""venue"": ""Journal of Experimental Psychology: Learning, Memory, and Cognition, 23(1):54."", ""year"": 1997}, {""authors"": [""Marton"", ""Yuval"", ""Chris Callison-Burch"", ""Philip Resnik.""], ""title"": ""Improved statistical machine translation using monolinguallyderived paraphrases"", ""venue"": ""Proceedings of EMNLP, pages 381\u2013390, Edinburgh."", ""year"": 2009}, {""authors"": [""McRae"", ""Ken"", ""Saman Khalkhali"", ""Mary Hare""], ""title"": ""Semantic and associative relations in adolescents and young"", ""year"": 2012}, {""authors"": [""Medelyan"", ""Olena"", ""David Milne"", ""Catherine Legg"", ""Ian H. Witten.""], ""title"": ""Mining meaning from Wikipedia"", ""venue"": ""International Journal of Human-Computer Studies, 67(9):716\u2013754."", ""year"": 2009}, {""authors"": [""Mikolov"", ""Tomas"", ""Kai Chen"", ""Greg Corrado"", ""Jeffrey Dean.""], ""title"": ""Efficient estimation of word representations in vector space"", ""venue"": ""Proceedings of International Conference of Learning Representations,"", ""year"": 2013}, {""authors"": [""Mikolov"", ""Tomas"", ""Ilya Sutskever"", ""Kai Chen"", ""Greg S. Corrado"", ""Jeff Dean.""], ""title"": ""Distributed representations of words and phrases and their compositionality"", ""venue"": ""Advances in Neural Information"", ""year"": 2013}, {""authors"": [""Navigli"", ""Roberto.""], ""title"": ""Word sense disambiguation: A survey"", ""venue"": ""ACM Computing Surveys (CSUR), 41(2):10."", ""year"": 2009}, {""authors"": [""Nelson"", ""Douglas L."", ""Cathy L. McEvoy"", ""Thomas A. Schreiber.""], ""title"": ""The University of South Florida free association, rhyme, and word fragment norms"", ""venue"": ""Behavior Research Methods, Instruments, & Computers,"", ""year"": 2004}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Ulrike Pad\u00f3"", ""Katrin Erk.""], ""title"": ""Flexible, corpus-based modelling of human plausibility judgements"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 400\u2013409, Prague."", ""year"": 2007}, {""authors"": [""Paivio"", ""Allan.""], ""title"": ""Dual coding theory: Retrospect and current status"", ""venue"": ""Canadian Journal of Psychology/Revue canadienne de psychologie, 45(3):255."", ""year"": 1991}, {""authors"": [""Pedersen"", ""Ted"", ""Siddharth Patwardhan"", ""Jason Michelizzi.""], ""title"": ""Wordnet:: Similarity: Measuring the relatedness of concepts"", ""venue"": ""Demonstration Papers at HLT-NAACL 2004, pages 38\u201341, New York,"", ""year"": 2004}, {""authors"": [""Phan"", ""Xuan-Hieu"", ""Le-Minh Nguyen"", ""Susumu Horiguchi.""], ""title"": ""Learning to classify short and sparse text & Web with hidden topics from large-scale data collections"", ""venue"": ""Proceedings of the 17th"", ""year"": 2008}, {""authors"": [""Plaut"", ""David C.""], ""title"": ""Semantic and associative priming in a distributed attractor network"", ""venue"": ""Proceedings"", ""year"": 1995}, {""authors"": [""Recchia"", ""Gabriel"", ""Michael N. Jones.""], ""title"": ""More data trumps smarter algorithms: Comparing pointwise mutual information with latent semantic analysis"", ""venue"": ""Behavior Research Methods, 41(3):647\u2013656."", ""year"": 2009}, {""authors"": [""Reisinger"", ""Joseph"", ""Raymond Mooney.""], ""title"": ""A mixture model with sharing for lexical semantics"", ""venue"": ""Proceedings of EMNLP, pages 1173\u20131182, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Reisinger"", ""Joseph"", ""Raymond J. Mooney.""], ""title"": ""Multi-prototype vector-space models of word meaning"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of"", ""year"": 2010}, {""authors"": [""Resnik"", ""Philip.""], ""title"": ""Using information content to evaluate semantic similarity in a taxonomy"", ""venue"": ""Proceedings of IJCAI."", ""year"": 1995}, {""authors"": [""Resnik"", ""Philip"", ""Jimmy Lin.""], ""title"": ""11 evaluations of NLP systems"", ""venue"": ""The handbook of computational linguistics and natural language processing, 57:271."", ""year"": 2010}, {""authors"": [""Rosch"", ""Eleanor"", ""Carol Simpson"", ""R. Scott Miller.""], ""title"": ""Structural bases of typicality effects"", ""venue"": ""Journal of Experimental Psychology: Human Perception and Performance, 2(4):491."", ""year"": 1976}, {""authors"": [""Rose"", ""Tony"", ""Mark Stevenson"", ""Miles Whitehead.""], ""title"": ""The Reuters corpus volume 1\u2014from yesterday\u2019s news to tomorrow\u2019s language resources"", ""venue"": ""LREC, volume 2, pages 827\u2013832,"", ""year"": 2002}, {""authors"": [""Rubenstein"", ""Herbert"", ""John B. Goodenough.""], ""title"": ""Contextual correlates of synonymy"", ""venue"": ""Communications of the ACM, 8(10):627\u2013633."", ""year"": 1965}, {""authors"": [""Silberer"", ""Carina"", ""Mirella Lapata.""], ""title"": ""Learning grounded meaning representations with autoencoders"", ""venue"": ""Proceedings of ACL, Sofia."", ""year"": 2014}, {""authors"": [""Sun"", ""Lin"", ""Anna Korhonen"", ""Yuval Krymolowski.""], ""title"": ""Verb class discovery from rich syntactic data"", ""venue"": ""A. Gelbukh, editor, Computational Linguistics and Intelligent Text processing. Springer,"", ""year"": 2008}, {""authors"": [""Turian"", ""Joseph"", ""Lev Ratinov"", ""Yoshua Bengio.""], ""title"": ""Word representations: A simple and general method for semi-supervised learning"", ""venue"": ""Proceedings of ACL, pages 384\u2013394,"", ""year"": 2010}, {""authors"": [""Turney"", ""Peter D.""], ""title"": ""Domain and function: A dual-space model of semantic relations and compositions"", ""venue"": ""Journal 694"", ""year"": 2012}, {""authors"": [""Turney"", ""Peter D."", ""Patrick Pantel.""], ""title"": ""From frequency to meaning: Vector space models of semantics"", ""venue"": ""Journal of Artificial Intelligence Research, 37(1):141\u2013188."", ""year"": 2010}, {""authors"": [""Tversky"", ""Amos.""], ""title"": ""Features of similarity"", ""venue"": ""Psychological Review, 84(4):327."", ""year"": 1977}, {""authors"": [""Wiebe"", ""Janyce.""], ""title"": ""Learning subjective adjectives from corpora"", ""venue"": ""AAAI/IAAI, pages 735\u2013740, Austin, TX."", ""year"": 2000}, {""authors"": [""Williams"", ""Gbolahan K"", ""Sarabjot Singh Anand""], ""title"": ""Predicting the polarity"", ""year"": 2009}, {""authors"": [""Wu"", ""Zhibiao"", ""Martha Palmer.""], ""title"": ""Verbs, semantics and lexical selection"", ""venue"": ""Proceedings of ACL, pages 133\u2013138, Las Cruces, NM."", ""year"": 1994}, {""authors"": [""Yong"", ""Chung"", ""Shou King Foo.""], ""title"": ""A case study on inter-annotator agreement for word sense disambiguation"", ""venue"": ""Proceedings of the ACL SIGLEX Workshop on Standardizing"", ""year"": 1999}]",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :There is very little similar about coffee and cups. Coffee refers to a plant, which is a living organism or a hot brown (liquid) drink. In contrast, a cup is a man-made solid of broadly well-defined shape and size with a specific function relating to the consumption of liquids. Perhaps the only clear trait these concepts have in common is that they are concrete entities. Nevertheless, in what is currently the most popular evaluation gold standard for semantic similarity, WordSim(WS)-353 (Finkelstein et al. 2001), coffee and
∗ Computer Laboratory University of Cambridge, UK. E-mail: {felix.hill, anna.korhonen}@ cl.cam.ac.uk.
∗∗ Technion, Israel Institute of Technology, Haifa, Israel. E-mail: roiri@ie.technion.ac.il.
Submission received: 25 July 2014; revised submission received: 10 June 2015; accepted for publication: 31 August 2015.
doi:10.1162/COLI a 00237
© 2015 Association for Computational Linguistics
cup are rated as more “similar” than pairs such as car and train, which share numerous common properties (function, material, dynamic behavior, wheels, windows, etc.). Such anomalies also exist in other gold standards such as the MEN data set (Bruni et al. 2012a). As a consequence, these evaluations effectively penalize models for learning the evident truth that coffee and cup are dissimilar.
Although clearly different, coffee and cup are very much related. The psychological literature refers to the conceptual relationship between these concepts as association, although it has been given a range of names including relatedness (Budanitsky and Hirst 2006; Agirre et al. 2009), topical similarity (Hatzivassiloglou et al. 2001), and domain similarity (Turney 2012). Association contrasts with similarity, the relation connecting cup and mug (Tversky 1977). At its strongest, the similarity relation is exemplified by pairs of synonyms; words with identical referents.
Computational models that effectively capture similarity as distinct from association have numerous applications. Such models are used for the automatic generation of dictionaries, thesauri, ontologies, and language correction tools (Biemann 2005; Cimiano, Hotho, and Staab 2005; Li et al. 2006). Machine translation systems, which aim to define mappings between fragments of different languages whose meaning is similar, but not necessarily associated, are another established application (He et al. 2008; Marton, Callison-Burch, and Resnik 2009). Moreover, since, as we establish, similarity is a cognitively complex operation that can require rich, structured conceptual knowledge to compute accurately, similarity estimation constitutes an effective proxy evaluation for general-purpose representation-learning models whose ultimate application is variable or unknown (Collobert and Weston 2008; Baroni and Lenci 2010).
As we show in Section 2, the predominant gold standards for semantic evaluation in NLP do not measure the ability of models to reflect similarity. In particular, in both WS353 and MEN, pairs of words with associated meaning, such as coffee and cup (rating = 6.810), telephone and communication (7.510), or movie and theater (7.710), receive a high rating regardless of whether or not their constituents are similar. Thus, the utility of such resources to the development and application of similarity models is limited, a problem exacerbated by the fact that many researchers appear unaware of what their evaluation resources actually measure.1
Although certain smaller gold standards—those of Rubenstein and Goodenough (1965) (RG) and Agirre et al. (2009) (WS-Sim)—do focus clearly on similarity, these resources suffer from other important limitations. For instance, as we show, and as is also the case for WS-353 and MEN, state-of-the-art models have reached the average performance of a human annotator on these evaluations. It is common practice in NLP to define the upper limit for automated performance on an evaluation as the average human performance or inter-annotator agreement (Yong and Foo 1999; Cunningham 2005; Resnik and Lin 2010). Based on this established principle and the current evaluations, it would therefore be reasonable to conclude that the problem of representation learning, at least for similarity modeling, is approaching resolution. However, circumstantial evidence suggests that distributional models are far from perfect. For instance, we are some way from automatically generated dictionaries, thesauri, or ontologies that can be used with the same confidence as their manually created equivalents.
1 For instance, Huang et al. (2012, pages 1, 4, 10) and Reisinger and Mooney (2010b, page 4) refer to MEN and/or WS-353 as “similarity data sets.” Others evaluate on both these association-based and genuine similarity-based gold standards with no reference to the fact that they measure different things (Medelyan et al. 2009; Li et al. 2014).
Motivated by these observations, in Section 3 we present SimLex-999, a gold standard resource for evaluating the ability of models to reflect similarity. SimLex-999 was produced by 500 paid native English speakers, recruited via Amazon Mechanical Turk,2 who were asked to rate the similarity, as opposed to association, of concepts via a simple visual interface. The choice of evaluation pairs in SimLex-999 was motivated by empirical evidence that humans represent concepts of distinct part-of-speech (POS) (Gentner 1978) and conceptual concreteness (Hill, Korhonen, and Bentz 2014) differently. Whereas existing gold standards contain only concrete noun concepts (MEN) or cover only some of these distinctions via a random selection of items (WS-353, RG), SimLex-999 contains a principled selection of adjective, verb, and noun concept pairs covering the full concreteness spectrum. This design enables more nuanced analyses of how computational models overcome the distinct challenges of representing concepts of these types.
In Section 4 we present quantitative and qualitative analyses of the SimLex-999 ratings, which indicate that participants found it unproblematic to quantify consistently the similarity of the full range of concepts and to distinguish it from association. Unlike existing data sets, SimLex-999 therefore contains a significant number of pairs, such as [movie, theater], which are strongly associated but receive low similarity scores.
The second main contribution of this paper, presented in Section 5, is the evaluation of state-of-the-art distributional semantic models using SimLex-999. These include the well-known neural language models (NLMs) of Huang et al. (2012), Collobert and Weston (2008), and Mikolov et al. (2013a), which we compare with traditional vectorspace co-occurrence models (VSMs) (Turney and Pantel 2010) with and without dimensionality reduction (SVD) (Landauer and Dumais 1997). Our analyses demonstrate how SimLex-999 can be applied to uncover substantial differences in the ability of models to represent concepts of different types.
Despite these differences, the models we consider each share the characteristic of being better able to capture association than similarity. We show that the difficulty of estimating similarity is driven primarily by those strongly associated pairs with a high (association) rating in gold standards such as WS-353 and MEN, but a low similarity rating in SimLex-999. As a result of including these challenging cases, together with a wider diversity of lexical concepts in general, current models achieve notably lower scores on SimLex-999 than on existing gold standard evaluations, and well below the SimLex-999 inter-human agreement ceiling.
Finally, we explore ways in which distributional models might improve on this performance in similarity modeling. To do so, we evaluate the models on the SimLex999 subsets of adjectives, nouns, and verbs, as well as on abstract and concrete subsets and subsets of more and less strongly associated pairs (Sections 5.2.2–5.2.4). As part of these analyses, we confirm the hypothesis (Agirre et al. 2009; Levy and Goldberg 2014) that models learning from input informed by dependency parsing, rather than simple running-text input, yield improved similarity estimation and, specifically, clearer distinction between similarity and association. In contrast, we find no evidence for a related hypothesis (Agirre et al. 2009; Kiela and Clark 2014) that smaller context windows improve the ability of models to capture similarity. We do, however, observe clear differences in model performance on the distinct concept types included in SimLex-999. Taken together, these experiments demonstrate the benefit of the diversity of concepts
2 www.mturk.com/.
included in SimLex-999; it would not have been possible to derive similar insights by evaluating based on existing gold standards.
We conclude by discussing how observations such as these can guide future research into distributional semantic models. By facilitating better-defined evaluations and finer-grained analyses, we hope that SimLex-999 will ultimately contribute to the development of models that accurately reflect human intuitions of similarity for the full range of concepts in language. 2 design motivation :In this section, we motivate the design decisions made in developing SimLex-999. We begin (2.1) by examining the distinction between similarity and association. We then show that for a meaningful treatment of similarity it is also important to take a principled approach to both POS and conceptual concreteness (2.2). We finish by reviewing existing gold standards, and show that none enables a satisfactory evaluation of the capability of models to capture similarity (2.3). The difference between association and similarity is exemplified by the concept pairs [car, bike] and [car, petrol]. Car is said to be (semantically) similar to bike and associated with (but not similar to) petrol. Intuitively, car and bike can be understood as similar because of their common physical features (e.g., wheels), their common function (transport), or because they fall within a clearly definable category (modes of transport). In contrast, car and petrol are associated because they frequently occur together in space and language, in this case as a result of a clear functional relationship (Plaut 1995; McRae, Khalkhali, and Hare 2012).
Association and similarity are neither mutually exclusive nor independent. Bike and car, for instance, are related to some degree by both relations. Because it is common in both the physical world and in language for distinct entities to interact, it is relatively easy to conceive of concept pairs, such as car and petrol, that are strongly associated but not similar. Identifying pairs of concepts for which the converse is true is comparatively more difficult. One exception is common concepts paired with low frequency synonyms, such as camel and dromedary. Because the essence of association is co-occurrence (linguistic or otherwise [McRae, Khalkhali, and Hare 2012]), such pairs can seem, at least intuitively, to be similar but not strongly associated.
To explore the interaction between the two cognitive phenomena quantitatively, we exploited perhaps the only two existing large-scale means of quantifying similarity and association. To estimate similarity, we considered proximity in the WordNet taxonomy (Fellbaum 1998). Specifically, we applied the measure of Wu and Palmer (1994) (henceforth WupSim), which approximates similarity on a [0,1] scale reflecting the minimum distance between any two synsets of two given concepts in WordNet. WupSim has been shown to correlate well with human judgments on the similarity-focused RG data set (Wu and Palmer 1994). To estimate association, we extracted ratings directly from the University of South Florida Free Association Database (USF) (Nelson, McEvoy, and Schreiber 2004). These data were generated by presenting human subjects with one of 5,000 cue concepts and asking them to write the first word that comes into their head that is associated with or meaningfully related to that concept. Each cue concept c was normed in this way by over 10 participants, resulting in a set of associates for each cue, and a total of over 72,000 (c, a) pairs. Moreover, for each such pair, the proportion of participants
who produced associate a when presented with cue c can be used as a proxy for the strength of association between the two concepts.
By measuring WupSim between all pairs in the USF data set, we observed, as expected, a high correlation between similarity and association strength across all USF pairs (Spearman ρ = 0.65, p < 0.001). However, in line with the intuitive ubiquity of pairs such as car and petrol, of the USF pairs (all of which are associated to a greater or lesser degree) over 10% had a WupSim score of less than 0.25. These include pairs of ontologically different entities with a clear functional relationship in the world [refrigerator, food], which may be of differing concreteness [lung, disease]; pairs in which one concept is a small concrete part of a larger abstract category [sheriff, police]; pairs in a relationship of modification or subcategorization [gravy, boat]; and even those whose principal connection is phonetic [wiggle, giggle]. As we show in Section 2.2, these are precisely the sort of pairs that are not contained in existing evaluation gold standards. Table 1 lists the USF noun pairs with the lowest similarity scores overall, and also those with the largest additive discrepancy between association strength and similarity.
2.1.1 Association and Similarity in NLP. As noted in the Introduction, the similarity/association distinction is not only of interest to researchers in psychology or linguistics. Models of similarity are particularly applicable to various NLP tasks, such as lexical resource building, semantic parsing, and machine translation (Haghighi et al. 2008; He et al. 2008; Marton, Callison-Burch, and Resnik 2009; Beltagy, Erk, and Mooney 2014). Models of association, on the other hand, may be better suited to tasks such as wordsense disambiguation (Navigli 2009), and applications such as text classification (Phan, Nguyen, and Horiguchi 2008) in which the target classes correspond to topical domains such as agriculture or sport (Rose, Stevenson, and Whitehead 2002).
Much recent research in distributional semantics does not distinguish between association and similarity in a principled way (see, e.g., Reisinger and Mooney 2010b; Huang et al. 2012; Luong, Socher, and Manning 2013).3 One exception is Turney (2012), who constructs two distributional models with different features and parameter settings, explicitly designed to capture either similarity or association. Using the output of these two models as input to a logistic regression classifier, Turney predicts whether two
3 Several papers that take a knowledge-based or symbolic approach to meaning do address the similarity/association issue (Budanitsky and Hirst 2006).
concepts are associated, similar, or both, with 61% accuracy. However, in the absence of a gold standard covering the full range of similarity ratings (rather than a list of pairs identified as being similar or not) Turney cannot confirm directly that the similarityfocused model does indeed effectively quantify similarity.
Agirre et al. (2009) explicitly examine the distinction between association and similarity in relation to distributional semantic models. Their study is based on the partition of WS-353 into a subset focused on similarity, which we refer to as WS-Sim, and a subset focused on association, which we term WS-Rel. More precisely, WS-Sim is the union of the pairs in WS-353 judged by three annotators to be similar and the set U of entirely unrelated pairs, and WS-Rel is the union of U and pairs judged to be associated but not similar. Agirre et al. confirm the importance of the association/similarity distinction by showing that certain models perform relatively well on WS-Rel, whereas others perform comparatively better on WS-Sim. However, as shown in the following section, a model need not be an exemplary model of similarity in order to perform well on WS-Sim, because an important class of concept pair (associated but not similar entities) is not represented in this data set. Therefore the insights that can be drawn from the results of the Agirre et al. study are limited.
Several other authors touch on the similarity/association distinction in inspecting the output of distributional models (Andrews, Vigliocco, and Vinson 2009; Kiela and Clark 2014; Levy and Goldberg 2014). Although the strength of the conclusions that can be drawn from such qualitative analyses is clearly limited, there appear to be two broad areas of consensus concerning similarity and distributional models:r Models that learn from input annotated for syntactic or dependency
relations better reflect similarity, whereas approaches that learn from running-text or bag-of-words input better model association (Agirre et al. 2009; Levy and Goldberg 2014).r Models with larger context windows may learn representations that better capture association, whereas models with narrower windows better reflect similarity (Agirre et al. 2009; Kiela and Clark 2014). Empirical studies have shown that the performance of both humans and distributional models depends on the POS category of the concepts learned. Gentner (2006) showed that children find verb concepts harder to learn than noun concepts, and Markman and Wisniewski (1997) present evidence that different cognitive operations are used when comparing two nouns or two verbs. Hill, Reichart, and Korhonen (2014) demonstrate differences in the ability of distributional models to acquire noun and verb semantics. Further, they show that these differences are greater for models that learn from both text and perceptual input (as with humans).
In addition to POS category, differences in human and computational concept learning and representation have been attributed to the effects of concreteness, the extent to which a concept has a directly perceptible physical referent. On the cognitive side, these “concreteness effects” are well established, even if the causes are still debated (Paivio 1991; Hill, Korhonen, and Bentz 2014). Concreteness has also been associated with differential performance in computational text-based (Hill, Kiela, and Korhonen 2013) and multi-modal semantic models (Kiela et al. 2014). For brevity, we do not exhaustively review all methods that have been used to evaluate semantic models, but instead focus on the similarity or association-based gold standards that are most commonly applied in recent work in NLP. In each case, we consider how well the data set satisfies one of the three following criteria:
Representative. The resource should cover the full range of concepts that occur in natural language. In particular, it should include cases representing the different ways in which humans represent or process concepts, and cases that are both challenging and straightforward for computational models.
Clearly defined. In order for a gold standard to be diagnostic of how well a model can be applied to downstream applications, a clear understanding is needed of what exactly the gold standard measures. In particular, it must clearly distinguish between dissociable semantic relations such as association and similarity.
Consistent and reliable. Untrained native speakers must be able to quantify the target property consistently, without requiring lengthy or detailed instructions. This ensures that the data reflect a meaningful cognitive or semantic phenomenon, and also enables the data set to be scaled up or transferred to other languages at minimal cost and effort.
We begin our review of existing evaluation with the gold standard most commonly applied in current NLP research.
WordSim-353. WS-353 (Finkelstein et al. 2001) is perhaps the most commonly used evaluation gold standard for semantic models. Despite its name, and the fact that it is often referred to as a “similarity gold standard,”4 in fact, the instructions given to annotators when producing WS-353 were ambiguous with respect to similarity and association. Subjects were asked to:
Assign a numerical similarity score between 0 and 10 (0 = words totally unrelated, 10 = words VERY closely related) ... when estimating similarity of antonyms, consider them “similar” (i.e., belonging to the same domain or representing features of the same concept), not “dissimilar”.
As we confirm analytically in Section 5.2, these instructions result in pairs being rated according to association rather than similarity.5 WS-353 consequently suffers two important limitations as an evaluation of similarity (which also afflict other resources to a greater or lesser degree):
1. Many dissimilar word pairs receive a high rating.
2. No associated but dissimilar concepts receive low ratings.
As noted in the Introduction, an arguably more serious third limitation of WS-353 is low inter-annotator agreement, and the fact that state-of-the-art models such as those
4 See, e.g., Huang et al. 2012 and Bansal, Gimpel, and Livescu 2014. 5 This fact is also noted by the data set authors. See www.cs.technion.ac.il/~gabr/resources/ data/wordsim353/.
of Collobert and Weston (2008) and Huang et al. (2012) reach, or even surpass, the inter-annotator agreement ceiling in estimating the WS-353 scores. Huang et al. report a Spearman correlation of ρ = 0.713 between their model output and WS-353. This is 10 percentage points higher than inter-annotator agreement (ρ = 0.611) when defined as the average pairwise correlation between two annotators, as is common in NLP work (Padó, Padó, and Erk 2007; Reisinger and Mooney 2010a; Silberer and Lapata 2014). It could be argued that a different comparison is more appropriate: Because the model is compared to the gold-standard average across all annotators, we should compare a single annotator with the (almost) gold-standard average over all other annotators. Based on this metric the average performance of an annotator on WS-353 is ρ = 0.756, which is still only marginally better than the best automatic method.6
Thus, at least according to the established wisdom in NLP evaluation (Yong and Foo 1999; Cunningham 2005; Resnik and Lin 2010), the strength of the conclusions that can be inferred from improvements on WS-353 is limited. At the same time, however, state-of-the-art distributional models are clearly not perfect representation-learning or even similarity estimation engines, as evidenced by the fact they cannot yet be applied, for instance, to generate flawless lexical resources (Alfonseca and Manandhar 2002).
WS-Sim. WS-Sim is the set of pairs in WS-353 identified by Agirre et al. (2009) as either containing similar or unrelated (neither similar nor associated) concepts. The ratings in WS-Sim are mapped directly from WS-353, so that all concept pairs in WSSim that receive a high rating are associated and all pairs that receive a low rating are unassociated. Consequently, any model that simply reflects association would score highly on WS-Sim, irrespective of how well it captures similarity.
Such a possibility could be excluded by requiring models to perform well on WSSim and poorly on WS-Rel, the subset of WS-353 identified by Agirre et al. (2009) as containing no pairs of similar concepts. However, although this would exclude models of pure association, it would not test the ability of models to quantify the similarity of the pairs in WS-Sim. Put another way, the WS-Sim/WS-Rel partition could in theory resolve limitation (1) of WS-353 but it would not resolve limitation (2): Models are not tested on their ability to attribute low scores to associated but dissimilar pairs.
In fact, there are more fundamental limitations of WS-Sim as a similarity-based evaluation resource. It does not, strictly speaking, reflect similarity at all, since the ratings of its constituent pairs were assigned by the WS-353 annotators, who were asked to estimate association, not similarity. Moreover, it inherits the limitation of low interannotator agreement from WS-353. The average pairwise correlation between annotators on WS-Sim is ρ = 0.667, and the average correlation of a single annotator with the gold standard is only ρ = 0.651, both below the performance of automatic methods (Agirre et al. 2009). Finally, the small size of WS-Sim renders it poorly representative of the full range of concepts that semantic models may be required to learn.
Rubenstein & Goodenough. Prior to WS-353, the smaller RG data set, consisting of 65 pairs, was often used to evaluate semantic models. The 15 raters employed in the data collection were asked to rate the “similarity of meaning” of each concept pair. Thus RG does appear to reflect similarity rather than association. However, although limitation (1) of WS-353 is therefore avoided, RG still suffers from limitation (2): By inspection,
6 Individual annotator responses for WS-353 were downloaded from www.cs.technion.ac.il/~gabr/ resources/data/wordsim353.
it is clear that the low similarity pairs in RG are not associated. A further limitation is that distributional models now achieve better performance on RG (correlations of up to Pearson r = 0.86 [Hassan and Mihalcea 2011]) than the reported inter-annotator agreement of r = 0.85 (Rubenstein and Goodenough 1965). Finally, the size of RG renders it an even less comprehensive evaluation than WS-Sim.
The MEN Test Collection. A larger data set, MEN (Bruni et al. 2012a), is used in a handful of recent studies (Bruni et al. 2012b; Bernardi et al. 2013). As with WS-353, both terms similarity and relatedness are used by the authors when describing MEN, although the annotators were expressly asked to rate pairs according to relatedness.7
The construction of MEN differed from RG and WS-353 in that each pair was only considered by one rater, who ranked it for relatedness relative to 50 other pairs in the data set. An overall score out of 50 was then attributed to each pair corresponding to how many times it was ranked as more related than an alternative. However, because these rankings are based on relatedness, with respect to evaluating similarity MEN necessarily suffers from both of the limitations (1) and (2) that apply to WS-353. Further, there is a strong bias towards concrete concepts in MEN because the concepts were originally selected from those identified in an image-bank (Bruni et al. 2012a).
Synonym Detection Sets. Multiple-choice synonym detection tasks, such as the TOEFL test questions (Landauer and Dumais 1997), are an alternative means of evaluating distributional models. A question in the TOEFL task consists of a cue word and four possible answer words, only one of which is a true synonym. Models are scored on the number of true synonyms identified out of 80 questions. The questions were designed by linguists to evaluate synonymy, so, unlike the evaluations considered thus far, TOEFL-style tests effectively discriminate between similarity and association. However, because they require a zero-one classification of pairs as synonymous or not, they do not test how well models discern pairs of medium or low similarity. More generally, in opposition to the fuzzy, statistical approaches to meaning predominant in both cognitive psychology (Griffiths, Steyvers, and Tenenbaum 2007) and NLP (Turney and Pantel 2010), they do not require similarity to be measured on a continuous scale. 3 the simlex-999 data set :Having considered the limitations of existing gold standards, in this section we describe the design of SimLex-999 in detail. Separating similarity from association. To create a test of the ability of models to capture similarity as opposed to association, we started with the ≈ 72,000 pairs of concepts in the USF data set. As the output of a free-association experiment, each of these pairs is associated to a greater or lesser extent. Importantly, inspecting the pairs revealed that a good range of similarity values are represented. In particular, there were many examples of hypernym/hyponym pairs [body, abdomen], cohyponym pairs [cat, dog], synonyms or near synonyms [deodorant, antiperspirant], and antonym pairs [good, evil]. From this cohort, we excluded pairs containing a multiple-word item [hot dog, mustard],
7 http://clic.cimec.unitn.it/~elia.bruni/MEN.html.
and pairs containing a capital letter [Mexico, sun]. We ultimately sampled 900 of the SimLex-999 pairs from the resulting cohort of pairs, according to the stratification procedures outlined in the following sections.
To complement this cohort with entirely unassociated pairs, we paired up the concepts from the 900 associated pairs at random. From these random parings, we excluded those that coincidentally occurred elsewhere in USF (and therefore had a degree of association). From the remaining pairs, we accepted only those in which both concepts had been subject to the USF norming procedure, ensuring that these non-USF pairs were indeed unassociated rather than simply not normed. We sampled the remaining 99 SimLex-999 pairs from this resulting cohort of unassociated pairs.
POS category. In light of the conceptual differences outlined in Section 2.2, SimLex999 includes subsets of pairs from the three principle meaning-bearing POS categories: nouns, verbs, and adjectives. To classify potential pairs according to POS, we counted the frequency with which the items in each pair occurred with the three possible tags in the POS-tagged British National Corpus (Leech, Garside, and Bryant 1994). To minimize POS ambiguity, which could lead to inconsistent ratings, we excluded pairs containing a concept with lower than 75% tendency towards one particular POS. This yielded three sets of potential pairs : [A,A] pairs (of two concepts whose majority tag was Adjective), [N,N] pairs, and [V,V] pairs.
Given the likelihood that different cognitive operations are used in estimating the similarity between items of different POS-category (Section 2.2), concept pairs were presented to raters in batches defined according to POS. Unlike both WS-353 and MEN, pairs of concepts of mixed POS ([white, rabbit], [run,marathon]) were excluded. POS categories are generally considered to reflect very broad ontological classes (Fellbaum 1998). We thus felt it would be very difficult, or even counter-intuitive, for annotators to quantify the similarity of mixed POS pairs according to our instructions.
Concreteness. Although a clear majority of pairs in gold standards such as MEN and RG contain concrete items, perhaps surprisingly, the vast majority of adjective, noun, and verb concepts in everyday language are in fact abstract (Hill, Reichart, and Korhonen 2014; Kiela et al. 2014).8 To facilitate the evaluation of models for both concrete and abstract concept meaning, and in light of the cognitive and computational modeling differences between abstract and concrete concepts noted in Section 2.2, we aimed to include both concept types in SimLex-999.
Unlike the POS distinction, concreteness is generally considered to be a gradual phenomenon. One benefit of sampling pairs for SimLex-999 from the USF data set is that most items have been rated according to concreteness on a scale of 1–7 by at least 10 human subjects. As Figure 1 demonstrates, concreteness (as the average over these ratings) interacts with POS on these concepts: Nouns are on average more concrete than verbs, which are more concrete than adjectives. However, there is also clear variation in concreteness within each POS category. We therefore aimed to select pairs for SimLex999 that spanned the full abstract–concrete continuum within each POS category.
After excluding any pairs that contained an item with no concreteness rating, for each potential SimLex-999 pair we considered both the concreteness of the first item and the additive difference in concreteness between the two items. This enabled us
8 According to the USF concreteness ratings, 72% of noun or verb types in the British National Corpus are more abstract than the concept war, a concept many would already consider quite abstract.
●
athletefailure treebelief propertychristmas
coughscaremakeseem
liberal loud darkhappy
Nouns
Verbs
Adjectives
2 3 4 5 6 7 Concreteness Rating
Figure 1 Boxplots showing the interaction between concreteness and POS for concepts in USF. The white boxes range from the first to third quartiles and the central vertical line indicates the median.
to stratify our sampling equally across four classes: (C1) concrete first item (rating > 4) with below-median concreteness difference; (C2) concrete first item (rating> 4), second item of lower concreteness and the difference being greater than the median; (C3) abstract first item (rating≤ 4) with below-median concreteness difference; and (C4) abstract first item (rating ≤ 4) with the second item of greater concreteness and the difference being greater than the median.
Final sampling. From the associated (USF) cohort of potential pairs we selected 600 noun pairs, 200 verb pairs, and 100 adjective pairs, and from the unassociated (non-USF) cohort, we sampled 66 nouns pairs, 22 verb pairs, and 11 adjective pairs. In both cases, the sampling was stratified such that, in each POS subset, each of the four concreteness classes C1−C4 was equally represented. The annotator instructions for SimLex-999 are shown in Figure 2. We did not attempt to formalize the notion of similarity, but rather introduce it via the well-understood idea of synonymy, and in contrast to association. Even if a formal characterization of similarity existed, the evidence in Section 2 suggests that the instructions would need separate cases to cover different concept types, increasing the difficulty of the rating task. Therefore, we preferred to appeal to intuition on similarity, and to verify post hoc that subjects were able to interpret and apply the informal characterization consistently for each concept type.
Immediately following the instructions in Figure 2, participants were presented with two “checkpoint” questions, one with abstract examples and one with concrete examples. In each case the participant was required to identify the most similar pair from a set of three options, all of which were associated, but only one of which was clearly similar (e.g. [bread, butter] [bread, toast] [stale, bread]). After this, the participants began rating pairs in groups of six or seven pairs by moving a slider, as shown in Figure 3.
This group size was chosen because the (relative) rating of a set of pairs implicitly requires pairwise comparisons between all pairs in that set. Therefore, larger groups would have significantly increased the cognitive load on the annotators. Another advantage of grouping was the clear break (submitting a set of ratings and moving to
the next page) between the tasks of rating adjective, noun, and verb pairs. For better inter-group calibration, from the second group onwards the last pair of the previous group became the first pair of the present group, and participants were asked to re-assign the rating previously attributed to the first pair before rating the remaining new items. As with MEN, WS-353, and RG, SimLex-999 consists of pairs of concept words together with a numerical rating. Thus, unlike in the small evaluation constructed by Huang et al. (2012), words are not rated in a phrasal or sentential context. Such meaning-in-context evaluations are motivated by a desire to disambiguate words that otherwise might be considered to have multiple senses.
We did not attempt to construct an evaluation based on meaning-in-context for several reasons. First, determining the set of senses for a given word, and then the set of contexts that represent those senses, introduces a high degree of subjectivity into the design process. Second, ensuring that a model has learned a high quality representation of a given concept would have required evaluating that concept in each of its given contexts, necessitating many more cases and a far greater annotation effort. Third, in the (infrequent) case that some concept c1 in an evaluation pair (c1, c2) is genuinely (etymologically) polysemous, c2 can provide sufficient context to disambiguate c1.9
9 This is supported by the fact that the WordNet-based methods that perform best at modeling human ratings model the similarity between concepts c1 and c2 as the minimum of all pairwise distances between the senses of c1 and the senses of c2 (Resnik 1995; Pedersen, Patwardhan, and Michelizzi 2004).
Finally, the POS grouping of pairs in the survey can also serve to disambiguate in the case that the conflicting senses of the polysemous concept are of differing POS categories. Each participant was asked to rate 20 groups of pairs on a 0–6 scale of integers (nonintegral ratings were not possible). Checkpoint multiple-choice questions were inserted at points between the 20 groups in order to ensure the participant had retained the correct notion of similarity. In addition to the checkpoint of three noun pairs presented before the first group (which contained noun pairs), checkpoint questions containing adjective pairs were inserted before the first adjective group and checkpoints of three verb pairs were inserted before the first verb group.
From the 999 evaluation pairs, 14 noun pairs, 4 verb pairs, and 2 adjective pairs were selected as a consistency set. The data set of pairs was then partitioned into 10 tranches, each consisting of 119 pairs, of which 20 were from the consistency set and the remaining 99 unique to that tranche. To reduce workload, each annotator was asked to rate the pairs in a single tranche only. The tranche itself was divided into 20 groups, with
each group consisting of 7 pairs (with the exception of the last group of the 20, which had 6). Of these seven pairs, the first pair was the last pair from the previous group, and the second pair was taken from the consistency set. The remaining pairs were unique to that particular group and tranche. The design enabled control for possible systematic differences between annotators and tranches, which could be detected by variation on the consistency set. Five hundred residents of the United States were recruited from Amazon Mechanical Turk, each with at least 95% approval rate for work on the Web service. Each participant was required to check a box confirming that he or she was a native speaker of English and warned that work would be rejected if the pattern of responses indicated otherwise. The participants were distributed evenly to rate pairs in one of the ten question tranches, so that each pair was rated by approximately 50 subjects. Participants took between 8 and 21 minutes to rate the 119 pairs across the 20 groups, together with the checkpoint questions. In order to correct for systematic differences in the overall calibration of the rating scale between respondents, we measured the average (mean) response of each rater on the consistency set. For 32 respondents, the absolute difference between this average and the mean of all such averages was greater than 1 (though never greater than 2); that is, 32 respondents demonstrated a clear tendency to rate pairs as either more or less similar than the overall rater population. To correct for this bias, we increased (or decreased) the rating of such respondents for each pair by one, except in cases where they had given the maximum rating, 6 (or minimum rating, 0). This adjustment, which ensured that the average response of each participant was within one of the mean of all respondents on the consistency set, resulted in a small increase to the inter-rater agreement on the data set as a whole.
After controlling for systematic calibration differences, we imposed three conditions for the responses of a rater to be included in the final data collation. First, the average pairwise Spearman correlation of responses with all other responses for a participant could not be more than one standard deviation below the mean of all such averages. Second, the increase in inter-rater agreement when a rater was excluded from the analysis needed to be smaller than at least 50 other raters (i.e., 10% of raters were excluded on this criterion). Third, we excluded the six participants who got one or more of the checkpoint questions wrong. A total of 99 participants were excluded based on one or more of these conditions, but no more than 16 from any one tranche (so that each pair in the final data set was rated by a minimum of 36 raters). Finally, we computed average (mean) scores for each pair, and transformed all scores linearly from the interval [0, 6] to the interval [0, 10]. 4 analysis of the data set :In this section we analyze the responses of the SimLex-999 annotators and the resulting ratings. First, by considering inter-annotator agreement, we examine the consistency with which annotators were able to apply the characterization of similarity, outlined in the instructions for the range of concept types in SimLex-999. Second, we verify that a
valid notion of similarity was understood by the annotators, in that they were able to accurately separate similarity from association. As in previous annotation or data collection for computational semantics (Padó, Padó, and Erk 2007; Reisinger and Mooney 2010a; Silberer and Lapata 2014) we computed the inter-rater agreement as the average of pairwise Spearman ρ correlations between the ratings of all respondents. Overall agreement was ρ = 0.67. This compares favorably with the agreement on WS-353 (ρ = 0.61 using the same method). The design of the MEN rating system precludes a conventional calculation of inter-rater agreement (Bruni et al. 2012b). However, two of the creators of MEN who independently rated the data set achieved an agreement of ρ = 0.68.10
The SimLex-999 inter-rater agreement suggests that participants were able to understand the (single) characterization of similarity presented in the instructions and to apply it to concepts of various types consistently. This conclusion was supported by inspection of the brief feedback offered by the majority of annotators in a final text field in the questionnaire: 78% expressed sentiment that the test was clear, easy to complete, or some similar sentiment.
Interestingly, as shown in Figure 4 (left), agreement was not uniform across the concept types. Contrary to what might be expected given established concreteness effects (Paivio 1991), we observed not only higher inter-rater agreement but also less per-pair variability for abstract rather than concrete concepts.11
Strikingly, the highest inter-rater consistency and lowest per-pair variation (defined as the inverse of the standard deviation of all ratings for that pair) was observed on adjective pairs. Although we are unsure exactly what drives this effect, a possible cause
10 Reported at http://clic.cimec.unitn.it/~elia.bruni/MEN. It is reasonable to assume that actual agreement on MEN may be somewhat lower than 0.68, given the small sample size and the expertise of the raters. 11 Per-pair variability was measured by calculating the standard deviation of responses for each pair, and averaging these scores across the pairs of each concept type.
is that many pairs of adjectives in SimLex-999 cohabit a single salient, one-dimensional scale ( freezing > cold > warm > hot). This may be a consequence of the fact that many pairs in SimLex-999 were selected (from USF) to have a degree of association. On inspection, pairs of nouns and verbs in SimLex-999 do not appear to occupy scales in the same way, possibly because concepts of these POS categories come to be associated via a more diverse range of relations. It seems plausible that humans are able to estimate the similarity of scale-based concepts more consistently than pairs of concepts related in a less uni-dimensional fashion.
Regardless of cause, however, the high agreement on adjectives is a satisfactory property of SimLex-999. Adjectives exhibit various aspects of lexical semantics that have proved challenging for computational models, including antonymy, polarity (Williams and Anand 2009), and sentiment (Wiebe 2000). To approach the high level of human confidence on the adjective pairs in SimLex-999, it may be necessary to focus particularly on developing automatic ways to capture these phenomena. Inspection of the SimLex-999 ratings indicated that pairs were indeed evaluated according to similarity rather than association. Table 2 includes examples that demonstrate a clear dissociation between the two semantic relations.
To verify this effect quantitatively, we recruited 100 additional participants to rate the WS-353 pairs, but following the SimLex-999 instructions and question format. As shown in Fig 5(a), there were clear differences between these new ratings and the original WS-353 ratings. In particular, a high proportion of pairs was given a lower rating by subjects following the SimLex-999 instructions than those following the WS-353 guidelines: The mean SimLex rating was 4.07 compared with 5.91 for WS-353.
This was consistent with our expectations that pairs of associated but dissimilar concepts would receive lower ratings based on the SimLex-999 than on the WS-353 instructions, whereas pairs that were both associated and similar would receive similar ratings in both cases. To confirm this, we compared the WS-353 and SimLex-999based ratings on the subsets WS-Rel and WS-Sim, which were hand-sorted by Agirre
et al. (2009) to include pairs connected by association (and not similarity) and those connected by similarity (but possibly also association), respectively.
As shown in Figure 5(b–c), the correlation between the SimLex-999-based and WS353 ratings was notably higher (ρ = 0.73) on the WS-Sim subset than the WS-Rel subset (ρ = 0.38). Specifically, the tendency of subjects following the SimLex-999 instructions to assign lower ratings than those following the WS-353 instructions was far more pronounced for pairs in WS-Sim (Figure 5(b)) than for those in WS-Rel (Figure 5(c)). This observation suggests that the associated but dissimilar pairs in WS-353 were an important driver of the overall lower mean for SimLex-999-based ratings, and thus provide strong evidence that the SimLex-999 instructions do indeed enable subjects to distinguish similarity from association effectively. We have established the validity of similarity as a notion understood by human raters and distinct from association. However, much theoretical semantics focuses on relations between words or concepts that are finer-grained than similarity and association. These include meronymy (a part to its whole, e.g., blade–knife), hypernymy (a category concept to a member of that category, e.g., animal–dog), and cohyponymy (two members of the same implicit category, e.g., the pair of animals dog–cat) (Cruse 1986). Beyond theoretical interest, these relations can have practical relevance. For instance, hypernymy can form the basis of semantic entailment and therefore textual inference: The proposition a cat is on the table entails that an animal is on the table precisely because of the hypernymy relation from animal to cat.
We chose not to make these finer-grained relations the basis of our evaluation for several reasons. At present, detecting relations such as hypernymy using distributional methods is challenging, even when supported by supervised classifiers with access to labeled pairs (Levy et al. 2015). Such a designation can seem to require specific
world-knowledge (is a snale a reptile?), can be gradual, as evidenced by typicality effects (Rosch, Simpson, and Miller 1976), or simply highly subjective. Moreover, a fine-grained relation R will only be attested (to any degree) between a small subset of all possible word pairs, whereas similarity can in theory be quantified for any two words chosen at random. We thus considered a focus on fine-grained semantic relations to be less appropriate for a general-purpose evaluation of representation quality.
Nevertheless, post hoc analysis of the SimLex annotator responses and fine-grained relation classes, as defined by lexicographers, yields further interesting insights into the nature of both similarity and association. Of the 999 word pairs in SimLex, 382 are also connected by one of the common finer-grained semantic relations in WordNet. For each of these relations, Figure 6 shows the average similarity rating and average USF free association score for all pairs that exhibit that relation.
In cases where a relationship of hypernymy/hyponymy exists between the words in a pair (not necessarily immediate : 1 hypernym, 2 hypernym, etc.) similarity and association coincide. Hyper/hyponym pairs that are separated by fewer levels in the WordNet hierarchy are both more strongly associated and rated as more similar. However, there are also interesting discrepancies between similarity and association. Unsurprisingly, pairs that are classed as synonyms in WordNet (i.e., having at least one sense in some common synset) are rated as more similar than pairs of any other relation type by SimLex annotators. In contrast, antonyms are the most strongly associated word pairs among these finer-grained relations. Further, pairs consisting of a meronym and holonym (part and whole) are comparatively strongly associated but not judged to be similar.
The analysis also highlights a case that can be particularly problematic when rating similarity: cohyponyms, or members of the same salient category (such as knife and fork). We gave no specific guidelines for how to rate such pairs in the SimLex annotator instructions, and whether they are considered similar or not seems to be a matter of perspective. On one hand, their membership of a common category could make them appear similar, particularly if the category is relatively specific. On the other hand, in
the case of knife and fork, for instance, the underlying category cultery might provide a backdrop against which the differences of distinct members become particularly salient. 5 evaluating models with simlex-999 :In this section, we demonstrate the applicability of SimLex-999 by analyzing the performance of various distributional semantic models in estimating the new ratings. The models were selected to cover the main classes of representation learning architectures (Baroni, Dinu, and Kruszewski 2014): Vector space co-occurrence (counting) models and NLMs (Bengio et al. 2003). We first show that SimLex-999 is notably more difficult for state-of-the-art models to estimate than existing gold standards. We then conduct more focused analyses on the various concept subsets defined in SimLex-999, exploring possible causes for the comparatively low performance of current models and, in turn, demonstrating how SimLex-999 can be applied to investigate such questions. Collobert & Weston. Collobert and Weston (2008) apply the architecture of an NLM to learn a word representations vw for each word w in some corpus vocabulary V. Each sentence s in the input text is represented by a matrix containing the vector representations of the words in s in order. The model then computes output scores f (s) and f (sw), where sw denotes an “incorrect” sentence created from s by replacing its last word with some other word w from V. Training involves updating the parameters of the function f and the entries of the vector representations vw such that f (s) is larger than f (sw) for any w in V, other than the correct final word of s. This corresponds to minimizing the sum of the following sentence objectives Cs over all sentences in the input corpus, which is achieved via (mini-batch) stochastic gradient descent:
Cs = ∑ w∈V max(0, 1− f (s) + f (sw))
The relatively low-dimension, dense (vector) representations learned by this model and the other NLMs introduced in this section are sometimes referred to as embeddings (Turian, Ratinov, and Bengio 2010). Collobert and Weston (2008) train their models on 852 million words of text from a 2007 dump of Wikipedia and the RCV1 Corpus (Lewis et al. 2004) and use their embeddings to achieve state-of-the-art results on a variety of NLP tasks. We downloaded the embeddings directly from the authors’ Web page.12
Huang et al. Huang et al. (2012) present a NLM that learns word embeddings to maximize the likelihood of predicting the last word in a sentence s based on (i) the previous words in that sentence (local context, as with Collobert and Weston [2008]) and (ii) the document d in which that word occurs (global context). As with Collobert and Weston (2008), the model represents input sentences as a matrix of word embeddings. In addition, it represents documents in the input corpus as single-vector averages over all word embeddings in that document. It can then compute scores g(s, d) and g(sw, d), whereas before sw is a sentence with an “incorrect” randomly selected last word.
12 http://ml.nec-labs.com/senna/.
Training is again by stochastic gradient descent, and corresponds to minimizing the sum of the sentence objectives Cs,d over all of the sentences in the corpus:
Cs,d = ∑ w∈V max(0, 1− g(s, d) + g(sw, d))
The combination of local and global contexts in the objective encourages the final word embeddings to reflect aspects of both the meaning of nearby words and of the documents in which those words appear. When learning from 990M words of Wikipedia text, Huang et al. report a Spearman correlation of ρ = 71.3 between the cosine similarity of their model embeddings and the WS-353 scores, which constitutes state-of-the-art performance for a NLM model on that data set. We downloaded these embeddings from the authors’ Web page.13
Mikolov et al. Mikolov et al. (2013a) present an architecture that learns word embeddings similar to those of standard NLMs but with no nonlinear hidden layer (resulting in a simpler scoring function). This enables faster representation learning for large vocabularies. Despite this simplification, the embeddings achieve state-of-theart performance on several semantic tasks including sentence completion and analogy modeling (Mikolov et al. 2013a, 2013b).
For each word type w in the vocabulary V, the model learns both a “targetembedding” rw ∈ Rd and a “context-embedding” r̂w ∈ Rd such that, given a target word, its ability to predict nearby context words is maximized. The probability of seeing context word c given target w is defined as:
p(c|w) = e r̂c·rw∑
v∈V er̂v·rw
The model learns from a set of (target-word, context-word) pairs, extracted from a corpus of sentences as follows. In a given sentence s (of length N), for each position n ≤ N, each word wn is treated in turn as a target word. An integer t(n) is then sampled from a uniform distribution on {1, . . . k}, where k > 0 is a predefined maximum contextwindow parameter. The pair tokens {(wn, wn+j) : −t(n) ≤ j ≤ t(n), wi ∈ s} are then appended to the training data. Thus, target/context training pairs are such that (i) only words within a k-window of the target are selected as context words for that target, and (ii) words closer to the target are more likely to be selected than those further away.
The training objective is then to maximize the log probability T, defined here, across all such examples from s, and then across all sentences in the corpus. This is achieved by stochastic gradient descent.
T = 1N N∑ n=1 ∑ −t(n)≤j≤t(n),j 6=0 log(p(wn+j|wn))
As with other NLMs, Mikolov et al.’s model captures conceptual semantics by exploiting the fact that words appearing in similar linguistic contexts are likely to
13 www.socher.org/index.php/Main/ImprovingWordRepresentationsViaGlobalContextAndMultiple WordPrototypes.
have similar meanings. Informally, the model adjusts its embeddings to increase the probability of observing the training corpus. Because this probability increases with p(c|w), and p(c|w) increases with the dot product r̂c · rw, the updates have the effect of moving each target-embedding incrementally “closer” to the context-embeddings of its collocates. In the target-embedding space, this results in embeddings of concept words that regularly occur in similar contexts moving closer together.
We use the author’s Word2vec software in order to train their model and use the target embeddings in our evaluations. We experimented with embeddings of dimension 100, 200, 300, 400, and 500 and found that 200 gave the best performance on both WS-353 and SimLex-999.
Vector Space Model (VSM). As an alternative to NLMs, we constructed a vector space model following the guidelines for optimal performance outlined by Kiela and Clark (2014). After extracting the 2,000 most frequent word tokens in the corpus that are not in a common list of stopwords14 as features, we populated a matrix of co-occurrence counts with a row for each of the concepts in some pair in our evaluation sets, and a column for each of the features. Co-occurrence was counted within a specified window size, although never across a sentence boundary. This resulting matrix was then weighted according to Pointwise Mutual Information (PMI) (Recchia and Jones 2009). The rows of the resulting matrix constitute the vector representations of the concepts.
SVD. As proposed initially in Landauer and Dumais (1997), we also experimented with models in which SVD (Golub and Reinsch 1970) is applied to the PMI-weighted VSM matrix, reducing the dimension of each concept representation to 300 (which yielded best results after experimenting, as before, with 100–500 dimension vectors).
For each model described in this section, we calculate similarity as the cosine similarity between the (vector) representations learned by that model. In experimenting with different models on SimLex-999, we aimed to answer the following questions: (i) How well do the established models perform on SimLex-999 versus on existing gold standards? (ii) Are any observed differences caused by the potential of different models to measure similarity vs. association? (iii) Are there interesting differences in ability of models to capture similarity between adjectives vs. nouns vs. verbs? (iv) In this case, are the observed differences driven by concreteness, and its interaction with POS, or are other factors also relevant?
Overall Performance on SimLex-999. Figure 7 shows the performance of the NLMs on SimLex-999 versus on comparable data sets, measured by Spearman’s ρ correlation. All models estimate the ratings of MEN and WS-353 more accurately than SimLex-999. The Huang et al. (2012) model performs well on WS-353,15 but is not very robust to changes in evaluation gold standard, and performs worst of all the models on SimLex-999. Given the focus of the WS-353 ratings, it is tempting to explain this by concluding that the
14 Taken from the Python Natural Language Toolkit (Bird 2006). 15 This score, based on embeddings downloaded from the authors’ webpage, is notably lower than the score
reported in Huang et al. (2012), mentioned in Section 5.1.
global context objective leads the Huang et al. (2012) model to focus on association rather than similarity. However, the true explanation may be less simple, since the Huang et al. (2012) model performs weakly on the association-based MEN data set. The Collobert and Weston (2008) model is more robust across WS-353 and MEN, but still does not match the performance of the Mikolov et al. (2013a) model on SimLex-999.
Figure 8 compares the best performing NLM model (Mikolov et al. 2013a) with the VSM and SVD models.16 In contrast to recent results that emphasize the superiority of NLMs over alternatives (Baroni, Dinu, and Kruszewski 2014), we observed no clear advantage for the NLM over the VSM or SVD when considering the associationbased gold standards WS-353 and MEN together. While the NLM is the strongest performer on WS-353, SVD is the strongest performer on MEN. However, the NLM model performs notably better than the alternatives at modeling similarity, as measured by SimLex-999.
Comparing all models in Figures 7 and 8 suggests that SimLex-999 is notably more challenging to model than the alternative data sets, with correlation scores ranging from 0.098 to 0.414. Thus, even when state-of-the-art models are trained for several days on massive text corpora,17 their performance on SimLex-999 is well below the interannotator agreement (Figure 7). This suggests that there is ample scope for SimLex-999 to guide the development of improved models.
Modeling Similarity vs. Association. The comparatively low performance of NLM, VSM, and SVD models on SimLex-999 compared with MEN and WS-353 is consistent with our hypothesis that modeling similarity is more difficult than modeling association. Indeed, given that many strongly associated but dissimilar pairs, such as [coffee, cup], are likely to have high co-occurrence in the training data, and that all models infer connections between concepts from linguistic co-occurrence in some form or another,
16 We conduct this comparison on the smaller RCV1 Corpus (Lewis et al. 2004) because training the VSM and SVD models is comparatively slow. 17 Training times reported by Huang et al. (2012) and for Collobert and Weston (2008) at http://ronan.collobert.com/senna/.
it seems plausible that models may overestimate the similarity of such pairs because they are “distracted” by association.
To test this hypothesis more precisely, we compared the performance of models on the whole of SimLex-999 versus its 333 most associated pairs (according to the USF free association scores). Importantly, pairs in this strongly associated subset still span the full range of possible similarity scores (min similarity = 0.23 [shrink, grow], max similarity = 9.80 [vanish, disappear]).
As shown in Figure 9, all models performed worse when the evaluation was restricted to pairs of strongly associated concepts, which was consistent with our hypothesis. The Collobert and Weston (2008) model was better than the Huang et al. (2012) model at estimating similarity in the face of high association. This is not entirely surprising given the global-context objective in the latter model, which may have encouraged more association-based connections between concepts. The Mikolov et al. model, however, performed notably better than both other NLMs. Moreover, this superiority is proportionally greater when evaluating on the most associated pairs only (as
indicated by the difference between the red and gray bars), suggesting that the improvement is driven at least in part by an increased ability to “distinguish” similarity from association.
To understand better how the architecture of models captures information pertinent to similarity modeling, we performed two additional experiments using SimLex-999. These comparisons were also motivated by the hypotheses, made in previous studies and outlined in Section 2.1.2, that both dependency-informed input and smaller context windows encourage models to capture similarity rather than association.
We tested the first hypothesis using the embeddings of Levy and Goldberg (2014), whose model extends the Mikolov et al. (2013a) model so that target-context training instances are extracted based on dependency-parsed rather than simple running text. As illustrated in Figure 9, the dependency-based embeddings outperform the original (running text) embeddings trained on the same corpus. Moreover, the comparatively large increase in the red bar compared to the gray bar suggests that an important part of the improvement of the dependency-based model derives from a greater ability to discern similarity from association.
Our comparisons provided less support for the second (window size) hypothesis. As shown in Figure 10, there is a negligible improvement in the performance of the model when the window size is reduced from 10 to 2. However, for the SVD model we observed the converse. The SVD model with window size 10 slightly outperforms the SVD model with window 2, and this improvement is quite pronounced on the most associated pairs in SimLex-999.
Learning Concepts of Different POS. Given the theoretical likelihood of variation in model performance across POS categories noted in Section 2.2, we evaluated the Mikolov et al. (2013a), VSM, and SVD models on the subsets of SimLex-999 containing adjective, noun, and verb concept pairs.
The analyses yield two notable conclusions, as shown in Figure 11. First, perhaps contrary to intuition, all models estimate the similarity of adjectives better than other concept categories. This aligns with the (also unexpected) observation that humans rate the similarity of adjectives more consistently and with more agreement than other parts of speech (see the dashed lines). However, the parallels between human raters and the models do not extend to verbs and nouns; verb similarity is rated more consistently
than noun similarity by humans, but models estimate these ratings more accurately for nouns than for verbs.
To better understand the linguistic information exploited by models when acquiring concepts of different POS, we also computed performance on the POS subsets of SimLex-999 of the dependency-based model of Levy and Goldberg (2014) and the standard skipgram model, in which linguistic contexts are encoded as simple bagsof-words (BOW) (Mikolov et al. (2013a) [trained on the same Wikipedia text]). As shown in Figure 12, dependency-aware contexts yield the largest improvements for capturing verb similarity. This aligns with the cognitive theory of verbs as relational concepts (Markman and Wisniewski 1997) whose meanings rely on their interaction with (or dependency on) other words or concepts. It is also consistent with research on the automatic acquisition of verb semantics, in which syntactic features have proven particularly important (Sun, Korhonen, and Krymolowski 2008). Although a deeper exploration of these effects is beyond the scope of this work, this preliminary analysis
again highlights how the word classes integrated into SimLex-999 are pertinent to a range of questions concerning lexical semantics.
Learning Concrete and Abstract Concepts. Given the strong interdependence between POS and conceptual concreteness (Figure 1), we aimed to explore whether the variation in model performance on different POS categories was in fact driven by an underlying effect of concreteness. To do so, we ranked each pair in the SimLex-999 data set according to the sum of the concreteness of the two words, and compared performance of models on the most concrete and least concrete quartiles according to this ranking (Figure 13).
Interestingly, the performance of models on the most abstract and most concrete pairs suggests that the distinction characterized by concreteness is at least partially independent of POS. Specifically, while the Mikolov et al. model was the highest performer of all POS categories, its performance was worse than both the simple VSM and SVD models (of window size 10) on the most concrete concept pairs.
This finding supports the growing evidence for systematic differences in representation and/or similarity operations between abstract and concrete concepts (Hill, Kiela, and Korhonen 2013), and suggests that at least part of these concreteness effects are independent of POS. In particular, it appears that models built from underlying vectors of co-occurrence counts, such as VSMs and SVD, are better equipped to capture the semantics of concrete entities, whereas the embeddings learned by NLMs can better capture abstract semantics. 6 conclusion :Although the ultimate test of semantic models should be their utility in downstream applications, the research community can undoubtedly benefit from ways to evaluate the general quality of the representations learned by such models, prior to their integration in any particular system. We have presented SimLex-999, a gold standard resource for the evaluation of semantic representations containing similarity ratings of word pairs of different POS categories and concreteness levels.
The development of SimLex-999 was principally motivated by two factors. First, as we demonstrated, several existing gold standards measure the ability of models to capture association rather than similarity, and others do not adequately test their ability to discriminate similarity from association. This is despite the many potential applications for accurate similarity-focused representation learning models. Analysis of the ratings of the 500 SimLex-999 annotators showed that subjects can consistently quantify similarity, as distinct from association, and apply it to various concept types, based on minimal intuitive instructions.
Second, as we showed, state-of-the-art models trained solely on running-text corpora have now reached or surpassed the human agreement ceiling on WordSim-353 and MEN, the most popular existing gold standards, as well as on RG and WS-Sim. These evaluations may therefore have limited use in guiding or moderating future improvements to distributional semantic models. Nevertheless, there is clearly still room for improvement in terms of the use of distributional models in functional applications. We therefore consider the comparatively low performance of state-of-the-art models on SimLex-999 to be one of its principal strengths. There is clear room under the inter-rating ceiling to guide the development of the next generation of distributional models.
We conducted a brief exploration of how models might improve on this performance, and verified the hypotheses that models trained on dependency-based input capture similarity more effectively than those trained on running-text input. The evidence that smaller context windows are also beneficial for similarity models was mixed, however. Indeed, we showed that the optimal window size depends on both the general model architecture and the part-of-speech and concreteness of the target concepts.
Our analysis of these hypotheses illustrates how the design of SimLex-999— covering a principled set of concept categories and including meta-information on concreteness and free-association strength—enables fine-grained analyses of the performance and parameterization of semantic models. However, these experiments only scratch the surface in terms of the possible analyses. We hope that researchers will adopt the resource as a robust means of answering a diverse range of questions pertinent to similarity modeling, distributional semantics, and representation learning in general.
In particular, for models to learn high-quality representations for all linguistic concepts, we believe that future work must uncover ways to explicitly or implicitly infer “deeper,” more general, conceptual properties such as intentionality, polarity, subjectivity, or concreteness (Gershman and Dyer 2014). However, although improving corpusbased models in this direction is certainly realistic, models that learn exclusively via the linguistic modality may never reach human-level performance on evaluations such as SimLex-999. This is because much conceptual knowledge, and particularly that which underlines similarity computations for concrete concepts, appears to be grounded in the perceptual modalities as much as in language (Barsalou et al. 2003).
Whatever the means by which the improvements are achieved, accurate conceptlevel representation is likely to constitute a necessary first step towards learning informative, language-neutral phrasal and sentential representations. Such representations would be hugely valuable for fundamental NLP applications such as language understanding tools and machine translation.
Distributional semantics aims to infer the meaning of words based on the company they keep (Firth 1957). However, although words that occur together in text often have associated meanings, these meanings may be very similar or indeed very different. Thus, possibly excepting the population of Argentina, most people would agree that, strictly speaking, Maradona is not synonymous with football (despite their high rating of 8.62 in WordSim-353). The challenge for the next generation of distributional models
may therefore be to infer what is useful from the co-occurrence signal and to overlook what is not. Perhaps only then will models capture most, or even all, of what humans know when they know how to use a language. We present SimLex-999, a gold standard resource for evaluating distributional semantic models that improves on existing resources in several important ways. First, in contrast to gold standards such as WordSim-353 and MEN, it explicitly quantifies similarity rather than association or relatedness so that pairs of entities that are associated but not actually similar (Freud, psychology) have a low rating. We show that, via this focus on similarity, SimLex-999 incentivizes the development of models with a different, and arguably wider, range of applications than those which reflect conceptual association. Second, SimLex-999 contains a range of concrete and abstract adjective, noun, and verb pairs, together with an independent rating of concreteness and (free) association strength for each pair. This diversity enables fine-grained analyses of the performance of models on concepts of different types, and consequently greater insight into how architectures can be improved. Further, unlike existing gold standard evaluations, for which automatic approaches have reached or surpassed the inter-annotator agreement ceiling, state-ofthe-art models perform well below this ceiling on SimLex-999. There is therefore plenty of scope for SimLex-999 to quantify future improvements to distributional semantic models, guiding the development of the next generation of representation-learning architectures. [{""affiliations"": [], ""name"": ""Felix Hill""}, {""affiliations"": [], ""name"": ""Roi Reichart""}, {""affiliations"": [], ""name"": ""Anna Korhonen""}] SP:bc23a2bd2507750d9ea8f0a4937c6bc8162c549c [{""authors"": [""Agirre"", ""Eneko"", ""Enrique Alfonseca"", ""Keith Hall"", ""Jana Kravalova"", ""Marius Pa\u015fca"", ""Aitor Soroa.""], ""title"": ""A study on similarity and relatedness using distributional and Wordnet-based approaches"", ""venue"": ""Proceedings"", ""year"": 2009}, {""authors"": [""Alfonseca"", ""Enrique"", ""Suresh Manandhar.""], ""title"": ""Extending a lexical ontology by a combination of distributional semantics signatures"", ""venue"": ""G. Schrieber et al., Knowledge Engineering and Knowledge"", ""year"": 2002}, {""authors"": [""Andrews"", ""Mark"", ""Gabriella Vigliocco"", ""David Vinson.""], ""title"": ""Integrating experiential and distributional data to learn semantic representations"", ""venue"": ""Psychological Review, 116(3):463."", ""year"": 2009}, {""authors"": [""Bansal"", ""Mohit"", ""Kevin Gimpel"", ""Karen Livescu.""], ""title"": ""Tailoring continuous word representations for dependency parsing"", ""venue"": ""Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Baroni"", ""Marco"", ""Georgiana Dinu"", ""Germ\u00e1n Kruszewski.""], ""title"": ""Don\u2019t count, predict! A systematic comparison of context-counting vs"", ""venue"": ""context-predicting semantic vectors. In Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Baroni"", ""Marco"", ""Alessandro Lenci.""], ""title"": ""Distributional memory: A general framework for corpus-based semantics"", ""venue"": ""Computational Linguistics, 36(4):673\u2013721."", ""year"": 2010}, {""authors"": [""Barsalou"", ""Lawrence W."", ""W. Kyle Simmons"", ""Aron K. Barbey"", ""Christine D. Wilson.""], ""title"": ""Grounding conceptual knowledge in modality-specific systems"", ""venue"": ""Trends in Cognitive Sciences, 7(2):84\u201391."", ""year"": 2003}, {""authors"": [""Beltagy"", ""Islam"", ""Katrin Erk"", ""Raymond Mooney.""], ""title"": ""Semantic parsing using distributional semantics and probabilistic logic"", ""venue"": ""ACL 2014 Workshop on Semantic Parsing."", ""year"": 2014}, {""authors"": [""Bengio"", ""Yoshua"", ""R\u00e9jean Ducharme"", ""Pascal Vincent"", ""Christian Jauvin.""], ""title"": ""A neural probabilistic language model"", ""venue"": ""The Journal of Machine Learning Research, 3:1137\u20131155."", ""year"": 2003}, {""authors"": [""Bernardi"", ""Raffaella"", ""Georgiana Dinu"", ""Marco Marelli"", ""Marco Baroni.""], ""title"": ""A relatedness benchmark to test the role of determiners in compositional distributional semantics"", ""venue"": ""Proceedings of"", ""year"": 2013}, {""authors"": [""Biemann"", ""Chris.""], ""title"": ""Ontology learning from text: A survey of methods"", ""venue"": ""LDV Forum, 20(2):75\u201393."", ""year"": 2005}, {""authors"": [""Bird"", ""Steven.""], ""title"": ""Nltk: the natural language toolkit"", ""venue"": ""Proceedings of the COLING/ACL on Interactive Presentation sessions, pages 69\u201372, Sydney."", ""year"": 2006}, {""authors"": [""Bruni"", ""Elia"", ""Gemma Boleda"", ""Marco Baroni"", ""Nam-Khanh Tran.""], ""title"": ""Distributional semantics in technicolor"", ""venue"": ""Proceedings of ACL, Jeju Island."", ""year"": 2012}, {""authors"": [""Bruni"", ""Elia"", ""Jasper Uijlings"", ""Marco Baroni"", ""Nicu Sebe.""], ""title"": ""Distributional semantics with eyes: Using image analysis to improve computational representations of word meaning"", ""venue"": ""Proceedings of the 20th"", ""year"": 2012}, {""authors"": [""Budanitsky"", ""Alexander"", ""Graeme Hirst.""], ""title"": ""Evaluating Wordnet-based measures of lexical semantic relatedness"", ""venue"": ""Computational Linguistics, 32(1):13\u201347."", ""year"": 2006}, {""authors"": [""Cimiano"", ""Philipp"", ""Andreas Hotho"", ""Steffen Staab.""], ""title"": ""Learning concept hierarchies from text corpora using formal concept analysis"", ""venue"": ""J. Artif. Intell. Res. (JAIR), 24:305\u2013339."", ""year"": 2005}, {""authors"": [""R. Collobert"", ""J. Weston.""], ""title"": ""A unified architecture for natural language processing: Deep neural networks with multitask learning"", ""venue"": ""International Conference on Machine Learning, ICML, Helsinki."", ""year"": 2008}, {""authors"": [""Cruse"", ""D. Alan.""], ""title"": ""Lexical semantics"", ""venue"": ""Cambridge University Press."", ""year"": 1986}, {""authors"": [""Cunningham"", ""Hamish.""], ""title"": ""Information extraction, automatic"", ""venue"": ""Encyclopedia of language and linguistics, pages 665\u2013677."", ""year"": 2005}, {""authors"": [""Fellbaum"", ""Christiane""], ""title"": ""WordNet"", ""venue"": ""Wiley Online Library."", ""year"": 1998}, {""authors"": [""Finkelstein"", ""Lev"", ""Evgeniy Gabrilovich"", ""Yossi Matias"", ""Ehud Rivlin"", ""Zach Solan"", ""Gadi Wolfman"", ""Eytan Ruppin.""], ""title"": ""Placing search in context: The concept revisited"", ""venue"": ""Proceedings of the"", ""year"": 2001}, {""authors"": [""J.R. Firth""], ""title"": ""Papers in Linguistics 1934\u20131951"", ""venue"": ""Oxford University Press."", ""year"": 1957}, {""authors"": [""Gentner"", ""Dedre.""], ""title"": ""On relational meaning: The acquisition of verb meaning"", ""venue"": ""Child Development, pages 988\u2013998."", ""year"": 1978}, {""authors"": [""Gentner"", ""Dedre.""], ""title"": ""Why verbs are hard to learn"", ""venue"": ""Action meets word: How Children Learn Verbs, pages 544\u2013564. 692"", ""year"": 2006}, {""authors"": [""Gershman"", ""Anatole"", ""Yulia Tsvetkov"", ""Leonid Boytsov"", ""Eric Nyberg"", ""Chris Dyer.""], ""title"": ""Metaphor detection with cross-lingual model transfer"", ""venue"": ""Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Golub"", ""Gene H."", ""Christian Reinsch.""], ""title"": ""Singular value decomposition and least squares solutions"", ""venue"": ""Numerische Mathematik, 14(5):403\u2013420."", ""year"": 1970}, {""authors"": [""Griffiths"", ""Thomas L."", ""Mark Steyvers"", ""Joshua B. Tenenbaum.""], ""title"": ""Topics in semantic representation"", ""venue"": ""Psychological Review, 114(2):211."", ""year"": 2007}, {""authors"": [""Haghighi"", ""Aria"", ""Percy Liang"", ""Taylor Berg-Kirkpatrick"", ""Dan Klein.""], ""title"": ""Learning bilingual lexicons from monolingual corpora"", ""venue"": ""Proceedings of ACL 2008, Columbus, OH."", ""year"": 2008}, {""authors"": [""Hassan"", ""Samer"", ""Rada Mihalcea.""], ""title"": ""Semantic relatedness using salient semantic analysis"", ""venue"": ""AAAI, San Francisco, CA."", ""year"": 2011}, {""authors"": [""Hatzivassiloglou"", ""Vasileios"", ""Judith L. Klavans"", ""Melissa L. Holcombe"", ""Regina Barzilay"", ""Min-Yen Kan"", ""Kathleen McKeown.""], ""title"": ""Simfinder: A flexible clustering tool for summarization"", ""venue"": ""In"", ""year"": 2001}, {""authors"": [""He"", ""Xiaodong"", ""Mei Yang"", ""Jianfeng Gao"", ""Patrick Nguyen"", ""Robert Moore.""], ""title"": ""Indirect-HMM-based hypothesis alignment for combining outputs from machine translation systems"", ""venue"": ""Proceedings"", ""year"": 2008}, {""authors"": [""Hill"", ""Felix"", ""Douwe Kiela"", ""Anna Korhonen.""], ""title"": ""Concreteness and corpora: A theoretical and practical analysis"", ""venue"": ""CMCL 2013, page 75, Sofia."", ""year"": 2013}, {""authors"": [""Hill"", ""Felix"", ""Anna Korhonen"", ""Christian Bentz.""], ""title"": ""A quantitative empirical analysis of the abstract/concrete distinction"", ""venue"": ""Cognitive Science, 38(1):162\u2013177."", ""year"": 2014}, {""authors"": [""Hill"", ""Felix"", ""Roi Reichart"", ""Anna Korhonen.""], ""title"": ""Multi-modal models for concrete and abstract concept meaning"", ""venue"": ""Transactions of the Association for Computational Linguistics (TACL), 2:285\u2013296."", ""year"": 2014}, {""authors"": [""Huang"", ""Eric H."", ""Richard Socher"", ""Christopher D. Manning"", ""Andrew Y. Ng.""], ""title"": ""Improving word representations via global context and multiple word prototypes"", ""venue"": ""Proceedings of ACL, pages 873\u2013882,"", ""year"": 2012}, {""authors"": [""Kiela"", ""Douwe"", ""Stephen Clark.""], ""title"": ""A systematic study of semantic vector space model parameters"", ""venue"": ""Proceedings of the 2nd Workshop on Continuous Vector"", ""year"": 2014}, {""authors"": [""Kiela"", ""Douwe"", ""Felix Hill"", ""Anna Korhonen"", ""Stephen Clark.""], ""title"": ""Improving multi-modal representations using image dispersion: Why less is sometimes more"", ""venue"": ""Proceedings of ACL, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Landauer"", ""Thomas K."", ""Susan T. Dumais.""], ""title"": ""A solution to Plato\u2019s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge"", ""venue"": ""Psychological Review,"", ""year"": 1997}, {""authors"": [""Leech"", ""Geoffrey"", ""Roger Garside"", ""Michael Bryant.""], ""title"": ""Claws4: The tagging of the British National Corpus"", ""venue"": ""Proceedings of COLING, pages 622\u2013628, Kyoto."", ""year"": 1994}, {""authors"": [""Levy"", ""Omer"", ""Yoav Goldberg.""], ""title"": ""Dependency-based word embeddings"", ""venue"": ""Proceedings of ACL, volume 2."", ""year"": 2014}, {""authors"": [""Levy"", ""Omer"", ""Steffen Remus"", ""Chris Biemann"", ""Idol Dagan""], ""title"": ""Do supervised distributional methods really learn lexical inference relations"", ""venue"": ""Proceedings of NAACL,"", ""year"": 2015}, {""authors"": [""Lewis"", ""David D."", ""Yiming Yang"", ""Tony G. Rose"", ""Fan Li.""], ""title"": ""Rcv1: A new benchmark collection for text categorization research"", ""venue"": ""The Journal of Machine Learning Research, 5:361\u2013397."", ""year"": 2004}, {""authors"": [""Li"", ""Changliang"", ""Bo Xu"", ""Gaowei Wu"", ""Xiuying Wang"", ""Wendong Ge"", ""Yan Li.""], ""title"": ""Obtaining better word representations via language transfer"", ""venue"": ""A. Gelbukh, editor, Computational Linguistics and Intelligent"", ""year"": 2014}, {""authors"": [""Li"", ""Mu"", ""Yang Zhang"", ""Muhua Zhu"", ""Ming Zhou.""], ""title"": ""Exploring distributional similarity based models for query spelling correction"", ""venue"": ""Proceedings of ALC, pages 1025\u20131032."", ""year"": 2006}, {""authors"": [""Luong"", ""Minh-Thang"", ""Richard Socher"", ""Christopher D. Manning.""], ""title"": ""Better word representations with recursive neural networks for morphology"", ""venue"": ""CoNLL-2013, page 104, Sofia."", ""year"": 2013}, {""authors"": [""Markman"", ""Arthur B."", ""Edward J. Wisniewski.""], ""title"": ""Similar and different: The differentiation of basic-level categories"", ""venue"": ""Journal of Experimental Psychology: Learning, Memory, and Cognition, 23(1):54."", ""year"": 1997}, {""authors"": [""Marton"", ""Yuval"", ""Chris Callison-Burch"", ""Philip Resnik.""], ""title"": ""Improved statistical machine translation using monolinguallyderived paraphrases"", ""venue"": ""Proceedings of EMNLP, pages 381\u2013390, Edinburgh."", ""year"": 2009}, {""authors"": [""McRae"", ""Ken"", ""Saman Khalkhali"", ""Mary Hare""], ""title"": ""Semantic and associative relations in adolescents and young"", ""year"": 2012}, {""authors"": [""Medelyan"", ""Olena"", ""David Milne"", ""Catherine Legg"", ""Ian H. Witten.""], ""title"": ""Mining meaning from Wikipedia"", ""venue"": ""International Journal of Human-Computer Studies, 67(9):716\u2013754."", ""year"": 2009}, {""authors"": [""Mikolov"", ""Tomas"", ""Kai Chen"", ""Greg Corrado"", ""Jeffrey Dean.""], ""title"": ""Efficient estimation of word representations in vector space"", ""venue"": ""Proceedings of International Conference of Learning Representations,"", ""year"": 2013}, {""authors"": [""Mikolov"", ""Tomas"", ""Ilya Sutskever"", ""Kai Chen"", ""Greg S. Corrado"", ""Jeff Dean.""], ""title"": ""Distributed representations of words and phrases and their compositionality"", ""venue"": ""Advances in Neural Information"", ""year"": 2013}, {""authors"": [""Navigli"", ""Roberto.""], ""title"": ""Word sense disambiguation: A survey"", ""venue"": ""ACM Computing Surveys (CSUR), 41(2):10."", ""year"": 2009}, {""authors"": [""Nelson"", ""Douglas L."", ""Cathy L. McEvoy"", ""Thomas A. Schreiber.""], ""title"": ""The University of South Florida free association, rhyme, and word fragment norms"", ""venue"": ""Behavior Research Methods, Instruments, & Computers,"", ""year"": 2004}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Ulrike Pad\u00f3"", ""Katrin Erk.""], ""title"": ""Flexible, corpus-based modelling of human plausibility judgements"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 400\u2013409, Prague."", ""year"": 2007}, {""authors"": [""Paivio"", ""Allan.""], ""title"": ""Dual coding theory: Retrospect and current status"", ""venue"": ""Canadian Journal of Psychology/Revue canadienne de psychologie, 45(3):255."", ""year"": 1991}, {""authors"": [""Pedersen"", ""Ted"", ""Siddharth Patwardhan"", ""Jason Michelizzi.""], ""title"": ""Wordnet:: Similarity: Measuring the relatedness of concepts"", ""venue"": ""Demonstration Papers at HLT-NAACL 2004, pages 38\u201341, New York,"", ""year"": 2004}, {""authors"": [""Phan"", ""Xuan-Hieu"", ""Le-Minh Nguyen"", ""Susumu Horiguchi.""], ""title"": ""Learning to classify short and sparse text & Web with hidden topics from large-scale data collections"", ""venue"": ""Proceedings of the 17th"", ""year"": 2008}, {""authors"": [""Plaut"", ""David C.""], ""title"": ""Semantic and associative priming in a distributed attractor network"", ""venue"": ""Proceedings"", ""year"": 1995}, {""authors"": [""Recchia"", ""Gabriel"", ""Michael N. Jones.""], ""title"": ""More data trumps smarter algorithms: Comparing pointwise mutual information with latent semantic analysis"", ""venue"": ""Behavior Research Methods, 41(3):647\u2013656."", ""year"": 2009}, {""authors"": [""Reisinger"", ""Joseph"", ""Raymond Mooney.""], ""title"": ""A mixture model with sharing for lexical semantics"", ""venue"": ""Proceedings of EMNLP, pages 1173\u20131182, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Reisinger"", ""Joseph"", ""Raymond J. Mooney.""], ""title"": ""Multi-prototype vector-space models of word meaning"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of"", ""year"": 2010}, {""authors"": [""Resnik"", ""Philip.""], ""title"": ""Using information content to evaluate semantic similarity in a taxonomy"", ""venue"": ""Proceedings of IJCAI."", ""year"": 1995}, {""authors"": [""Resnik"", ""Philip"", ""Jimmy Lin.""], ""title"": ""11 evaluations of NLP systems"", ""venue"": ""The handbook of computational linguistics and natural language processing, 57:271."", ""year"": 2010}, {""authors"": [""Rosch"", ""Eleanor"", ""Carol Simpson"", ""R. Scott Miller.""], ""title"": ""Structural bases of typicality effects"", ""venue"": ""Journal of Experimental Psychology: Human Perception and Performance, 2(4):491."", ""year"": 1976}, {""authors"": [""Rose"", ""Tony"", ""Mark Stevenson"", ""Miles Whitehead.""], ""title"": ""The Reuters corpus volume 1\u2014from yesterday\u2019s news to tomorrow\u2019s language resources"", ""venue"": ""LREC, volume 2, pages 827\u2013832,"", ""year"": 2002}, {""authors"": [""Rubenstein"", ""Herbert"", ""John B. Goodenough.""], ""title"": ""Contextual correlates of synonymy"", ""venue"": ""Communications of the ACM, 8(10):627\u2013633."", ""year"": 1965}, {""authors"": [""Silberer"", ""Carina"", ""Mirella Lapata.""], ""title"": ""Learning grounded meaning representations with autoencoders"", ""venue"": ""Proceedings of ACL, Sofia."", ""year"": 2014}, {""authors"": [""Sun"", ""Lin"", ""Anna Korhonen"", ""Yuval Krymolowski.""], ""title"": ""Verb class discovery from rich syntactic data"", ""venue"": ""A. Gelbukh, editor, Computational Linguistics and Intelligent Text processing. Springer,"", ""year"": 2008}, {""authors"": [""Turian"", ""Joseph"", ""Lev Ratinov"", ""Yoshua Bengio.""], ""title"": ""Word representations: A simple and general method for semi-supervised learning"", ""venue"": ""Proceedings of ACL, pages 384\u2013394,"", ""year"": 2010}, {""authors"": [""Turney"", ""Peter D.""], ""title"": ""Domain and function: A dual-space model of semantic relations and compositions"", ""venue"": ""Journal 694"", ""year"": 2012}, {""authors"": [""Turney"", ""Peter D."", ""Patrick Pantel.""], ""title"": ""From frequency to meaning: Vector space models of semantics"", ""venue"": ""Journal of Artificial Intelligence Research, 37(1):141\u2013188."", ""year"": 2010}, {""authors"": [""Tversky"", ""Amos.""], ""title"": ""Features of similarity"", ""venue"": ""Psychological Review, 84(4):327."", ""year"": 1977}, {""authors"": [""Wiebe"", ""Janyce.""], ""title"": ""Learning subjective adjectives from corpora"", ""venue"": ""AAAI/IAAI, pages 735\u2013740, Austin, TX."", ""year"": 2000}, {""authors"": [""Williams"", ""Gbolahan K"", ""Sarabjot Singh Anand""], ""title"": ""Predicting the polarity"", ""year"": 2009}, {""authors"": [""Wu"", ""Zhibiao"", ""Martha Palmer.""], ""title"": ""Verbs, semantics and lexical selection"", ""venue"": ""Proceedings of ACL, pages 133\u2013138, Las Cruces, NM."", ""year"": 1994}, {""authors"": [""Yong"", ""Chung"", ""Shou King Foo.""], ""title"": ""A case study on inter-annotator agreement for word sense disambiguation"", ""venue"": ""Proceedings of the ACL SIGLEX Workshop on Standardizing"", ""year"": 1999}]","1 introduction :There is very little similar about coffee and cups. Coffee refers to a plant, which is a living organism or a hot brown (liquid) drink. In contrast, a cup is a man-made solid of broadly well-defined shape and size with a specific function relating to the consumption of liquids. Perhaps the only clear trait these concepts have in common is that they are concrete entities. Nevertheless, in what is currently the most popular evaluation gold standard for semantic similarity, WordSim(WS)-353 (Finkelstein et al. 2001), coffee and
∗ Computer Laboratory University of Cambridge, UK. E-mail: {felix.hill, anna.korhonen}@ cl.cam.ac.uk.
∗∗ Technion, Israel Institute of Technology, Haifa, Israel. E-mail: roiri@ie.technion.ac.il.
Submission received: 25 July 2014; revised submission received: 10 June 2015; accepted for publication: 31 August 2015.
doi:10.1162/COLI a 00237
© 2015 Association for Computational Linguistics
cup are rated as more “similar” than pairs such as car and train, which share numerous common properties (function, material, dynamic behavior, wheels, windows, etc.). Such anomalies also exist in other gold standards such as the MEN data set (Bruni et al. 2012a). As a consequence, these evaluations effectively penalize models for learning the evident truth that coffee and cup are dissimilar.
Although clearly different, coffee and cup are very much related. The psychological literature refers to the conceptual relationship between these concepts as association, although it has been given a range of names including relatedness (Budanitsky and Hirst 2006; Agirre et al. 2009), topical similarity (Hatzivassiloglou et al. 2001), and domain similarity (Turney 2012). Association contrasts with similarity, the relation connecting cup and mug (Tversky 1977). At its strongest, the similarity relation is exemplified by pairs of synonyms; words with identical referents.
Computational models that effectively capture similarity as distinct from association have numerous applications. Such models are used for the automatic generation of dictionaries, thesauri, ontologies, and language correction tools (Biemann 2005; Cimiano, Hotho, and Staab 2005; Li et al. 2006). Machine translation systems, which aim to define mappings between fragments of different languages whose meaning is similar, but not necessarily associated, are another established application (He et al. 2008; Marton, Callison-Burch, and Resnik 2009). Moreover, since, as we establish, similarity is a cognitively complex operation that can require rich, structured conceptual knowledge to compute accurately, similarity estimation constitutes an effective proxy evaluation for general-purpose representation-learning models whose ultimate application is variable or unknown (Collobert and Weston 2008; Baroni and Lenci 2010).
As we show in Section 2, the predominant gold standards for semantic evaluation in NLP do not measure the ability of models to reflect similarity. In particular, in both WS353 and MEN, pairs of words with associated meaning, such as coffee and cup (rating = 6.810), telephone and communication (7.510), or movie and theater (7.710), receive a high rating regardless of whether or not their constituents are similar. Thus, the utility of such resources to the development and application of similarity models is limited, a problem exacerbated by the fact that many researchers appear unaware of what their evaluation resources actually measure.1
Although certain smaller gold standards—those of Rubenstein and Goodenough (1965) (RG) and Agirre et al. (2009) (WS-Sim)—do focus clearly on similarity, these resources suffer from other important limitations. For instance, as we show, and as is also the case for WS-353 and MEN, state-of-the-art models have reached the average performance of a human annotator on these evaluations. It is common practice in NLP to define the upper limit for automated performance on an evaluation as the average human performance or inter-annotator agreement (Yong and Foo 1999; Cunningham 2005; Resnik and Lin 2010). Based on this established principle and the current evaluations, it would therefore be reasonable to conclude that the problem of representation learning, at least for similarity modeling, is approaching resolution. However, circumstantial evidence suggests that distributional models are far from perfect. For instance, we are some way from automatically generated dictionaries, thesauri, or ontologies that can be used with the same confidence as their manually created equivalents.
1 For instance, Huang et al. (2012, pages 1, 4, 10) and Reisinger and Mooney (2010b, page 4) refer to MEN and/or WS-353 as “similarity data sets.” Others evaluate on both these association-based and genuine similarity-based gold standards with no reference to the fact that they measure different things (Medelyan et al. 2009; Li et al. 2014).
Motivated by these observations, in Section 3 we present SimLex-999, a gold standard resource for evaluating the ability of models to reflect similarity. SimLex-999 was produced by 500 paid native English speakers, recruited via Amazon Mechanical Turk,2 who were asked to rate the similarity, as opposed to association, of concepts via a simple visual interface. The choice of evaluation pairs in SimLex-999 was motivated by empirical evidence that humans represent concepts of distinct part-of-speech (POS) (Gentner 1978) and conceptual concreteness (Hill, Korhonen, and Bentz 2014) differently. Whereas existing gold standards contain only concrete noun concepts (MEN) or cover only some of these distinctions via a random selection of items (WS-353, RG), SimLex-999 contains a principled selection of adjective, verb, and noun concept pairs covering the full concreteness spectrum. This design enables more nuanced analyses of how computational models overcome the distinct challenges of representing concepts of these types.
In Section 4 we present quantitative and qualitative analyses of the SimLex-999 ratings, which indicate that participants found it unproblematic to quantify consistently the similarity of the full range of concepts and to distinguish it from association. Unlike existing data sets, SimLex-999 therefore contains a significant number of pairs, such as [movie, theater], which are strongly associated but receive low similarity scores.
The second main contribution of this paper, presented in Section 5, is the evaluation of state-of-the-art distributional semantic models using SimLex-999. These include the well-known neural language models (NLMs) of Huang et al. (2012), Collobert and Weston (2008), and Mikolov et al. (2013a), which we compare with traditional vectorspace co-occurrence models (VSMs) (Turney and Pantel 2010) with and without dimensionality reduction (SVD) (Landauer and Dumais 1997). Our analyses demonstrate how SimLex-999 can be applied to uncover substantial differences in the ability of models to represent concepts of different types.
Despite these differences, the models we consider each share the characteristic of being better able to capture association than similarity. We show that the difficulty of estimating similarity is driven primarily by those strongly associated pairs with a high (association) rating in gold standards such as WS-353 and MEN, but a low similarity rating in SimLex-999. As a result of including these challenging cases, together with a wider diversity of lexical concepts in general, current models achieve notably lower scores on SimLex-999 than on existing gold standard evaluations, and well below the SimLex-999 inter-human agreement ceiling.
Finally, we explore ways in which distributional models might improve on this performance in similarity modeling. To do so, we evaluate the models on the SimLex999 subsets of adjectives, nouns, and verbs, as well as on abstract and concrete subsets and subsets of more and less strongly associated pairs (Sections 5.2.2–5.2.4). As part of these analyses, we confirm the hypothesis (Agirre et al. 2009; Levy and Goldberg 2014) that models learning from input informed by dependency parsing, rather than simple running-text input, yield improved similarity estimation and, specifically, clearer distinction between similarity and association. In contrast, we find no evidence for a related hypothesis (Agirre et al. 2009; Kiela and Clark 2014) that smaller context windows improve the ability of models to capture similarity. We do, however, observe clear differences in model performance on the distinct concept types included in SimLex-999. Taken together, these experiments demonstrate the benefit of the diversity of concepts
2 www.mturk.com/.
included in SimLex-999; it would not have been possible to derive similar insights by evaluating based on existing gold standards.
We conclude by discussing how observations such as these can guide future research into distributional semantic models. By facilitating better-defined evaluations and finer-grained analyses, we hope that SimLex-999 will ultimately contribute to the development of models that accurately reflect human intuitions of similarity for the full range of concepts in language. 6 conclusion :Although the ultimate test of semantic models should be their utility in downstream applications, the research community can undoubtedly benefit from ways to evaluate the general quality of the representations learned by such models, prior to their integration in any particular system. We have presented SimLex-999, a gold standard resource for the evaluation of semantic representations containing similarity ratings of word pairs of different POS categories and concreteness levels.
The development of SimLex-999 was principally motivated by two factors. First, as we demonstrated, several existing gold standards measure the ability of models to capture association rather than similarity, and others do not adequately test their ability to discriminate similarity from association. This is despite the many potential applications for accurate similarity-focused representation learning models. Analysis of the ratings of the 500 SimLex-999 annotators showed that subjects can consistently quantify similarity, as distinct from association, and apply it to various concept types, based on minimal intuitive instructions.
Second, as we showed, state-of-the-art models trained solely on running-text corpora have now reached or surpassed the human agreement ceiling on WordSim-353 and MEN, the most popular existing gold standards, as well as on RG and WS-Sim. These evaluations may therefore have limited use in guiding or moderating future improvements to distributional semantic models. Nevertheless, there is clearly still room for improvement in terms of the use of distributional models in functional applications. We therefore consider the comparatively low performance of state-of-the-art models on SimLex-999 to be one of its principal strengths. There is clear room under the inter-rating ceiling to guide the development of the next generation of distributional models.
We conducted a brief exploration of how models might improve on this performance, and verified the hypotheses that models trained on dependency-based input capture similarity more effectively than those trained on running-text input. The evidence that smaller context windows are also beneficial for similarity models was mixed, however. Indeed, we showed that the optimal window size depends on both the general model architecture and the part-of-speech and concreteness of the target concepts.
Our analysis of these hypotheses illustrates how the design of SimLex-999— covering a principled set of concept categories and including meta-information on concreteness and free-association strength—enables fine-grained analyses of the performance and parameterization of semantic models. However, these experiments only scratch the surface in terms of the possible analyses. We hope that researchers will adopt the resource as a robust means of answering a diverse range of questions pertinent to similarity modeling, distributional semantics, and representation learning in general.
In particular, for models to learn high-quality representations for all linguistic concepts, we believe that future work must uncover ways to explicitly or implicitly infer “deeper,” more general, conceptual properties such as intentionality, polarity, subjectivity, or concreteness (Gershman and Dyer 2014). However, although improving corpusbased models in this direction is certainly realistic, models that learn exclusively via the linguistic modality may never reach human-level performance on evaluations such as SimLex-999. This is because much conceptual knowledge, and particularly that which underlines similarity computations for concrete concepts, appears to be grounded in the perceptual modalities as much as in language (Barsalou et al. 2003).
Whatever the means by which the improvements are achieved, accurate conceptlevel representation is likely to constitute a necessary first step towards learning informative, language-neutral phrasal and sentential representations. Such representations would be hugely valuable for fundamental NLP applications such as language understanding tools and machine translation.
Distributional semantics aims to infer the meaning of words based on the company they keep (Firth 1957). However, although words that occur together in text often have associated meanings, these meanings may be very similar or indeed very different. Thus, possibly excepting the population of Argentina, most people would agree that, strictly speaking, Maradona is not synonymous with football (despite their high rating of 8.62 in WordSim-353). The challenge for the next generation of distributional models
may therefore be to infer what is useful from the co-occurrence signal and to overlook what is not. Perhaps only then will models capture most, or even all, of what humans know when they know how to use a language."
"1 introduction :With the emergence of popular social media services such as Twitter and Facebook, many studies in the area of Natural Language Processing (NLP) have been published that analyze the text data from these services for a variety of applications, such as opinion mining, sentiment analysis, event detection, or crisis management (Culotta 2010; Sriram et al. 2010; Yin et al. 2012). Many of these studies have primarily relied on building classification models for different learning tasks, such as text classification or Named Entity Recognition. The effectiveness of these models is often evaluated using cross-validation techniques.
Cross-validation, first introduced by Geisser (1974), has been acclaimed as the most popular evaluation method for estimating prediction errors in regression and
∗ CSIRO, Sydney, New South Wales, Australia. E-mail: {sarvnaz.karimi, jie.jin}@csiro.au. ∗∗ Sabik Software Solutions Pty Ltd, Sydney, New South Wales, Australia. E-mail: jiri@baum.com.au.
Submission received: 4 October 2013; revised submission received: 9 September 2014; accepted for publication: 10 November 2014.
doi:10.1162/COLI a 00230
© 2015 Association for Computational Linguistics
classification problems. In that method, the data D are randomly partitioned into k non-overlapping subsets (folds) Dk of approximately equal size. The validation step is repeated k times, using a different Dv = Dk as the validation data and Dt = D \ Dk as the training data each time. The final evaluation is the average over the k validation steps. Cross-validation is found to have a lower variance than a single hold-out set validation and thus it is commonly used on both moderate and large amounts of data without introducing efficiency concerns (Arlot and Celisse 2010). Compared with other choices of k, 10-fold cross-validation has been accepted as the most reliable method, which gives a highly accurate estimate of the generalization error of a given model for a variety of algorithms and applications.
Despite its wide applications, debates on the appropriateness of cross-validation have been raised in a number of areas, particularly in time series analysis (Bergmeir and Benı́tez 2012) and chemical engineering (Sheridan 2013). A fundamental assumption of cross-validation is that the data need to be independent and identically distributed (i.i.d.) between folds (Arlot and Celisse 2010). Therefore, if the data points used for training and validation are not independent, cross-validation is no longer valid and would usually overestimate the validity of the model. For time series forecasting, because the data are comprised of correlated observations and might be generated by a process that evolves over time, the training set and the validation set are not independent if randomly chosen for cross-validation. Researchers since the early 1990s have used modified variants of cross-validation to compensate for time dependence within times series (Chu and Marron 1991; Bergmeir and Benı́tez 2012). In the area of chemical engineering, the work of (Sheridan 2013) investigates the dependence in chemical data and observes that the existence of similar compounds or molecules across the data set leads to overoptimistic results using standard k-fold cross-validation.
We take such observations from the time series and chemical domains as a warning to investigate the data dependence in computational linguistics. We argue that even when the data appear to be independently generated and when there is no reason to believe that temporal dependencies are present, unexpected statistical dependencies may be induced through an incorrect application of cross-validation. Once there is a chance of having similar or otherwise dependent data points, random split of the data without taking this factor into account would cause incorrect or at least unreliable evaluation, which may lead to invalid or at least unjustified conclusions. Although similar concerns have been raised by a prior study (Lerman et al. 2008) that cross-validation might not be suitable for measuring the accuracy of public opinion forecasting, there is a lack of systematic analysis on how potential data dependency might invalidate cross-validation and what alternative evaluation methods exist. With the aim of gaining further insight into this issue, we perform a detailed empirical study based on text classification in Twitter, and show that inappropriate choice of cross-validation techniques could potentially lead to misleading conclusions. This concern could apply more generally to other data types of similar nature. We also explore several evaluation methods, mostly borrowed from research on time series, that are better suited for statistically dependent data and that could be adapted by researchers working in the NLP area.","2 unexpected statistical dependence in microblog data :We argue that microblog data, such as Twitter messages, known as tweets, are statistically dependent in nature. It has been demonstrated that there is redundancy in
Twitter language (Zanzotto, Pennacchiotti, and Tsioutsiouliklis 2011), which in turn can translate to evidence against statistical independence of microblog posts in given periods of time or occurrences of the same event. In summary, statistical dependence in tweets can arise for the following reasons:
Events The events of interest, or events being discussed by the microbloggers at large, may be temporally limited within certain specific time horizons; Textual Links Hashtags and other idiomatic expressions may be invented, gain and lose popularity, fall out of use, or be reused with a different meaning. In addition, particular microbloggers may actively share information on certain types of events or express their opinions on certain topics in similar contexts; Twinning There may be “twinning” of tweets (and therefore data points) because of various forms of retweeting, where different microbloggers post substantially the same tweets, in response to one another.
Many existing studies that use microblog data (e.g., for tweet classification)—especially those related to detecting or monitoring events—have adopted k-fold cross-validation to evaluate the effectiveness of the learned classification models (e.g., Culotta 2010; Sriram et al. 2010; Jiang et al. 2011; Uysal and Croft 2011; Takemura and Tajima 2012; Kumar et al. 2014). However, they overlook the possible statistical dependence among microblog data in terms of content (e.g., sharing hashtags in Twitter), relevance to current events, and time of publishing that could potentially impact their evaluation results. In Table 1 we use examples of these studies to illustrate how potential statistical dependence of microblog data is omitted, which might have impacted the validity of the results in these studies. The potential source of the ignored statistical dependence shown in the last column is categorized into Events, Textual Links, and Twinning (described above), based on the list of features that the authors have used in their machine learning approaches. For example, if one is interested in detecting specific events (e.g., diseases [Culotta 2010] or disasters [Verma et al. 2011; Kumar, Jiang, and Fang 2014]), using tweets from the middle of an event to train for and evaluate the detection of the onset of that same event is clearly invalid. In addition, certain users (e.g., authorities that announce disease outbreaks such as @ECDC Outbreaks, or car accidents such as @emergencyAUS) may always post in the same way [Verma et al. 2011; Kumar, Jiang, and Fang 2014], making tweets share similar contexts; or microbloggers always use the same hashtags to indicate similar topics (Jiang et al. 2011). More generally, if substantially verbatim “twin” copies of tweets find their way into both the training and validation data sets (e.g., [Uysal and Croft 2011; Verma et al. 2011; Kumar, Jiang, and Fang 2014]), model validity will be overestimated. We note that, as we do not have access to the data sets used in these studies, we cannot verify the level of influence that the data dependence and their choice of evaluation have on their reported results. However, we raise a warning that there might be an overlooked unexpected dependence influencing their results.","3 validation for statistically dependent data :Where the data is not independent, dependence-aware validation needs to be used. We describe four dependence-aware validation methods that take data dependence into consideration in the evaluation. In the following discussion, we refer to the data set used in the experiments as D = {d1, d2, · · · , dn}. The training data is referred to as Dt, and testing, evaluation or validation data as Dv, where both Dt and Dv are subsets of D.","4 a case study on tweet classification :We focus on two tweet classification tasks in our case study. The first is a binary classification, where tweets are classified into disaster-related or not; the second is a disaster type classification, where tweets are predicted to one of the following six classes: nondisaster, storm, earthquake, flooding, fire, and other disasters. In our experiments, we use LibSVM (Chang and Lin 2011) to build discriminative classifiers for our classification tasks.
Tweets are short texts currently limited up to 140 characters. Often, microbloggers use tweets in reply to others by using mentions (which are Twitter usernames and are preceded by @), or use hashtags such as #CycloneEvan to make the grouping of similar messages easier, or to increase the visibility of their posts to others interested in the
same topic. Use of links to Web pages, mostly full stories of what is briefly reported in the tweet, is also popular. Selecting features for a text classifier that is built on Twitter data can therefore benefit from both conventional textual features, such as n-grams, and Twitter-specific features, such as hashtags and mentions.
In our experiments, we investigate the effect of the following features and their combinations on classification of tweets for disasters: (1) n-grams: Unigrams and bigrams of tweet text at the word level, excluding any hashtag or mention in the text. To find n-grams we pre-process tweets to remove stopwords; (2) Hashtag: Two different features are explored. First, a binary feature of the hashtags in the tweets, which indicates whether a hashtag exists in the tweet or not; second, the total number of hashtags in a tweet; (3) Mention: Two types of features (binary and mention count) are explored, exactly the same as hashtags explained above; and (4) Link: A binary feature that specifies whether or not a tweet contains any link to a Web page. We randomly sampled a total of 7,500 English tweets published in two years from December 2010 to December 2012 from a system (Yin et al. 2012) that stores billions of tweets from the Twitter streaming API. Explicit retweets were excluded. This set contained a number of disasters such as earthquakes in Christchurch, New Zealand 2011; York floods, England 2012; and Hurricane Sandy, United States 2012.
For a machine learner such as a classifier to work, we need to present it with a representative set of labeled training data. We therefore annotated our tweet data manually to identify disaster tweets and their types. We annotated the data set based on two main questions: Is this tweet talking about a specific disaster? What type of disaster is it talking about? Types of disasters were defined as earthquake, fire, flooding, storm, other, and non-disaster. Annotations were done by three annotators for each tweet, who were hired through the crowd-sourcing service Crowdflower. After taking majority votes where at least two out of the three annotators agreed on both questions, we ended up with a set of 6,580 annotated tweets, of which 2,897 tweets were identified as disaster-related and 3,683 as non-disaster. In disaster tweets, 37% were annotated with earthquake, followed by fire, flooding, and storm constituting 23%, 21%, and 12%, respectively. For our tweet classification tasks, we set up experiments to compare five validation methods—standard 10-fold cross-validation, border-split cross-validation, neighborsplit validation, time-split validation, and time-border-split validation—for identifying tweets that are relevant to a disaster, and whether or not we can broadly identify the type of disaster. We evaluate classification effectiveness using the accuracy metric, which is the percentage of correctly classified tweets. We used the following settings for our evaluation schemes:
r k-fold cross-validation: We used k = 10 folds.r Border-split cross-validation: We used k = 10 folds for border-split and a radius h = 20 days. We assume that for most events, the social media activity on the topic dies after three weeks of their occurrence.
r Neighbor-split validation: We used cosine similarity to find a subset of the data that has the least number of neighbors in the data set. We weighted hashtags double and disregarded mentions in calculating the similarity. Neighborhood was decided using a minimum threshold of 0.25 on the resulting cosine similarity. The size of the test data set was the same as in time-split (below) but the size of the training data set was 5,868.r Time-split validation: We chose the cut-off time t∗ so that 90% of the data was used as training (5,922 tweets) and 10% as validation (658 tweets).r Time-border-split validation: We used the same t∗ and h as time-split and border-split. This resulted in a training data set containing 87.1% of the data and a validation data set identical to the one in time-split at 10% of the data (658 tweets). In removing the border, 2.9% of the data were discarded. The size of the training set was 5,750 tweets. We run two sets of experiments: Discriminating disaster tweets from non-disaster (Disaster or Not), and classifying tweets into the six classes of earthquake, fire, flooding, storm, other, and non-disaster (Disaster Type). Table 2 compares the classification results for SVM on a range of feature combinations using five different validation methods. We aim to show that inappropriate choice of cross-validation could possibly lead to misleading conclusions, including overestimated classification accuracies, and suboptimal feature sets that yield the highest accuracies.
k-fold Cross-Validation. The first set of experiments was conducted using standard 10-fold cross-validation. For discriminating disaster tweets from non-disaster, SVM achieved a maximum of 92.8% accuracy when unigrams and hashtags were used. A similar result of 92.7% accuracy was recorded for classifying tweets to their disaster types. Unsurprisingly, having hashtags as additional features was the most
helpful. As a side note, the standard deviations of 10-fold cross-validation are quite small; although it is well-known that these are not an unbiased estimate of the true variance (see, for instance, Bengio and Grandvalet [2004]), it nevertheless seems to suggest a degree of confidence. If we would assume that tweets are statistically independent, we could conclude that the SVM classifier using unigrams plus hashtags is the best performer for classifying disaster tweets with a high accuracy of over 92%. This is what most previous studies have done. However, this result is overoptimistic. For example, during the cross-validation, tweets, with the same hashtags are distributed over all folds, making it easier for the classifier to associate labels with known hashtags.
Border-Split Cross-Validation. In contrast to 10-fold cross-validation, border-split cross-validation gives a much lower performance score across all of the results. The standard deviations are also much larger.
Neighbor-split Validation. The neighbour-split results are very different. Unlike previous work (Sheridan 2013), we find that neighbor-split judges the effectiveness of the models quite highly, near the scores from 10-fold cross-validation but ranking the feature combinations differently. We believe the high scores are because it tends to pick non-disaster tweets into the validation set, as those tweets have fewest neighbors. This makes it easier to classify all validation tweets into the majority class (i.e., non-disasters).
Time-split and Time-border-split Validation. Neither of these validations leads to results similar to 10-fold cross-validation; they are substantially lower, in one case by over 20 percentage points (bottom row of table). Numerically, they are similar to the border-split cross-validation results, but again with a different ranking of the feature combinations. Time-split and time-border-split methods gave similar results to each other, with different rankings but only small numerical differences. Coincidentally, in our data set, the time t∗ fell largely between events of interest, so the elimination of the border resulted in only a small correction, confounded with the effect of the small reduction in training set size. The validation methods provide very different results, with a number of the differences around 20 percentage points. In addition, they disagree on the ranking of the feature combinations, which may be the more important problem in many studies. This is particularly visible with the Disaster or Not classification; the best combination of features (bold in Table 2) is different, depending on the validation method used. A study based on 10-fold cross-validation would suggest Unigram+Hashtag for this problem, which is the second-worst combination of features, according to time-border-split validation. Further, the confidence intervals, if used, would suggest confidence in this choice.
Given our experiments, we believe that the standard k-fold cross-validation substantially overestimates the performance of the models in our case study. Time-split and time-border-split validation are more likely to represent accurate evaluation when temporally dependent data is involved, as they simulate true prospective prediction. However, they both have the downside that they only rely on one pair of training and validation data sets. This means that they will have a larger variance than a
cross-validation method, and do not give any measure of confidence in their own results.
Border-split cross-validation may also be acceptable, provided that its assumptions are satisfied—primarily, that the data points are independent beyond a certain radius. Depending on a particular task, this may be easy to decide (e.g., separate events) or hard (e.g., a number of overlapping events with no clear time difference).
We also find that neighbor-split overestimates rather than underestimates the performance of the models relative to time-split and time-border-split. This represents a hazard to its use: The neighborhood measure may interact in unexpected ways with other features of the data, rendering the validation unpredictable.
Ideally, one would use multiple large test data sets, all collected during separate, non-overlapping periods of time, after the training data set. This would be the most reliable measure of the performance of the models, but also the most expensive in terms of required data collection and annotation. Failing that, we recommend time-split or time-border-split validation when temporal statistical dependence cannot be ruled out.","5 conclusions :We used a common task in NLP, text classification, on a relatively recent but widely used data source, Twitter streams, to show that blindly following the same evaluation method for tasks of similar nature could lead to invalid conclusions. In particular, we investigated the most common evaluation method for machine learning applications, standard 10-fold cross-validation, and compared it with other validation methods that take the statistical dependence in the data into account, including time-split, bordersplit, neighbor-split, and a combination of time- and border-split validation. We showed how cross-validation can overestimate the effectiveness of a tweet classification application. We argued that text in microblogs or other similar text from social media (e.g., Web forums or even online news) can be statistically dependent for specific studies, such as those looking at events. Researchers therefore need to be careful in choosing evaluation methodologies based on the nature of the data at hand to avoid bias in their results.",,,,"In recent years, many studies have been published on data collected from social media, especially microblogs such as Twitter. However, rather few of these studies have considered evaluation methodologies that take into account the statistically dependent nature of such data, which breaks the theoretical conditions for using cross-validation. Despite concerns raised in the past about using cross-validation for data of similar characteristics, such as time series, some of these studies evaluate their work using standard k-fold cross-validation. Through experiments on Twitter data collected during a two-year period that includes disastrous events, we show that by ignoring the statistical dependence of the text messages published in social media, standard cross-validation can result in misleading conclusions in a machine learning task. We explore alternative evaluation methods that explicitly deal with statistical dependence in text. Our work also raises concerns for any other data for which similar conditions might hold.","[{""affiliations"": [], ""name"": ""Sarvnaz Karimi""}, {""affiliations"": [], ""name"": ""Jie Yin""}, {""affiliations"": [], ""name"": ""Jiri Baum""}]",SP:17ced766be8b5ad0aeadb2cbd25ee9d31eb0e63a,"[{""authors"": [""S. Arlot"", ""A. Celisse.""], ""title"": ""A survey of cross-validation procedures for model selection"", ""venue"": ""Statistics Surveys, 4:40\u201379. Bengio, Y. and Y. Grandvalet. 2004. No"", ""year"": 2010}, {""authors"": [""C. Lin""], ""title"": ""LIBSVM: A library for support vector machines"", ""venue"": ""Information Sciences,"", ""year"": 2011}, {""authors"": [""Chu"", ""C.-K"", ""J.S. Marron""], ""title"": ""Comparison of two bandwidth selectors"", ""year"": 1991}, {""authors"": [""S. Geisser""], ""title"": ""A predictive approach to the random effect mode"", ""venue"": ""Biometrika Trust, 61(9):101\u2013107."", ""year"": 1974}, {""authors"": [""L. Jiang"", ""M. Yu"", ""M. Zhou"", ""X. Liu"", ""T. Zhao.""], ""title"": ""Target-dependent Twitter sentiment classification"", ""venue"": ""ACL-HLT, pages 151\u2013160, Portland, OR."", ""year"": 2011}, {""authors"": [""A. Kumar"", ""M. Jiang"", ""Y. Fang.""], ""title"": ""Where not to go?: Detecting road hazards using Twitter"", ""venue"": ""SIGIR, pages 1223\u20131226, Gold Coast."", ""year"": 2014}, {""authors"": [""K. Lerman"", ""A. Gilder"", ""M. Dredze"", ""F. Pereira.""], ""title"": ""Reading the markets: Forecasting public opinion of political candidates by news analysis"", ""venue"": ""COLING, pages 473\u2013480, Manchester."", ""year"": 2008}, {""authors"": [""R.P. Sheridan""], ""title"": ""Time-split cross-validation as a method for estimating the goodness of prospective prediction"", ""venue"": ""Journal of Chemical Information and Modeling,"", ""year"": 2013}, {""authors"": [""B. Sriram"", ""D. Fuhry"", ""E. Demir"", ""H. Ferhatosmanoglu"", ""M. Demirbas.""], ""title"": ""Short text classification in Twitter to improve information filtering"", ""venue"": ""SIGIR, pages 841\u2013842, Geneva."", ""year"": 2010}, {""authors"": [""H. Takemura"", ""K. Tajima.""], ""title"": ""Tweet classification based on their lifetime duration"", ""venue"": ""CIKM, pages 2367\u20132370, Maui, HI."", ""year"": 2012}, {""authors"": [""I. Uysal"", ""W.B. Croft.""], ""title"": ""User oriented tweet ranking: A filtering approach to microblogs"", ""venue"": ""CIKM, pages 2261\u20132264, Glasgow."", ""year"": 2011}, {""authors"": [""S. Verma"", ""S. Vieweg"", ""W.J. Corvey"", ""L. Palen"", ""J.H. Martin"", ""M. Palmer"", ""A. Schram"", ""K.M. Anderson""], ""title"": ""Natural language processing to the rescue? Extracting \u201csituational awareness"", ""year"": 2011}, {""authors"": [""J. Yin"", ""S. Karimi"", ""B. Robinson"", ""M. Cameron.""], ""title"": ""ESA: Emergency situation awareness via microbloggers"", ""venue"": ""CIKM, pages 2701\u20132703, Maui, HI."", ""year"": 2012}, {""authors"": [""F.M. Zanzotto"", ""M. Pennacchiotti"", ""K. Tsioutsiouliklis.""], ""title"": ""Linguistic redundancy in Twitter"", ""venue"": ""EMNLP, pages 659\u2013669, Edinburgh. 548"", ""year"": 2011}]",,,,,,,,,,,squibs  :,evaluation methods for statistically :,dependent text :Sarvnaz Karimi∗,"csiro  :Jie Yin∗ Jiri Baum∗∗ Sabik Software Solutions
In recent years, many studies have been published on data collected from social media, especially microblogs such as Twitter. However, rather few of these studies have considered evaluation methodologies that take into account the statistically dependent nature of such data, which breaks the theoretical conditions for using cross-validation. Despite concerns raised in the past about using cross-validation for data of similar characteristics, such as time series, some of these studies evaluate their work using standard k-fold cross-validation. Through experiments on Twitter data collected during a two-year period that includes disastrous events, we show that by ignoring the statistical dependence of the text messages published in social media, standard cross-validation can result in misleading conclusions in a machine learning task. We explore alternative evaluation methods that explicitly deal with statistical dependence in text. Our work also raises concerns for any other data for which similar conditions might hold.","reference study data dependence concern :Sriram et al. (2010) Classification of tweets into five categories of events, news, deals, opinions, and private messages 5,407 tweets, no mention of removing retweets Events, Textual Links, Twinning
Culotta (2010) Classification of tweets using regression models into flu or not flu
500,000 tweets from a 10-week period Events, Textual Links, Twinning
Verma et al. (2011) Classification of tweets to specific disaster events, and identification of tweets that contained situational awareness content 1,965 tweets collected from specific disasters in the U.S. by keyword search
Twinning, Textual Links
Jiang et al. (2011) Sentiment classification of tweets; hashtags were mentioned as one of the features, retweets were considered to share the same sentiment Tweets found using keyword search, without removing retweets
Twinning, Textual Links
Uysal and Croft (2011)
Personalized tweet ranking using retweet behavior. Decision tree classifiers were used based on features from the tweet content, user behavior, and tweet author 24,200 tweets, from which 2,547 were retweeted by the seed users
Twinning, Textual Links
Takemura and Tajima (2012) Classification of tweets into three categories based on whether they should be read now, later, or outdated
9,890 tweets from a fixed period of time, annotated for time-(in)dependency Events, Textual Links, Twinning
Kumar, Jiang, and Fang (2014)
Classification of tweets into two categories of road hazard and non-hazard
30,876 tweets, retweets were not removed
Events, Textual Links, Twinning
Border-Split Cross-Validation. This method, proposed by Chu and Marron (1991), is a modification of k-fold cross-validation for time series data. Data are partitioned into the folds in time sequence rather than randomly; then in each validation step, data points within a time distance h of any point in the validation data set are excluded from the training data. That is, for each validation step k, the validation data is Dv = Dk, there is a border data set Db = {db ∈ D \ Dv : (∃dv∈Dv) |tdb − tdv | ≤ h}, which is disregarded, and the training data is Dt = D \ (Dk ∪ Db). This method assumes that the data dependence is confined to some known radius h, with data points beyond that radius being independent.
Time-Split Validation. In time-split validation (Sheridan 2013), a particular time t∗ is chosen; data points prior to this time are allocated to the training data set and data points after t∗ are used as validation data; that is, Dt = {dt ∈ D : tdt ≤ t
∗} and Dv = {dv ∈ D : tdv > t
∗}. Note that this is not a cross-validation method, as no “crossing” takes place. The motivation is to emulate prospective prediction: A model is built using only the information available up to time t∗ and evaluated on data that are collected after that time.
Time-Border-Split Validation. This is a combination of border-split and time-split validation. Both a time t∗ and a radius h are chosen; data points prior to time t∗ − h are allocated to the training data set and data points after time t∗ are used as validation data; that is, Dt = {dt ∈ D : tdt ≤ t ∗ − h} and Dv = {dv ∈ D : tdv > t ∗}. The remaining data points are not used. The motivation is to combine the conservative aspects of both time-split and border-split, emulating prospective prediction more carefully.
Neighbor-Split Validation. Proposed by (Sheridan 2013), this method assumes the existence of some similarity metric for the data points. The number of neighbors for each data point is calculated, using a threshold on the similarity metric. A desired fraction of data points with the fewest neighbors are then allocated to the validation data Dv and the rest to the training data Dt. The motivation of this approach is to deliberately reduce the similarity between training and validation data. It is inspired by leave-class-out validation or cross-validation, which assumes a pre-existing classification (rather than similarity metric) and allocates data points to the validation data Dv or cross-validation folds Dk, according to this classification. The advantage of neighborsplit over leave-class-out is that the size of the validation data is a parameter that can be chosen, rather than being dependent on the (potentially unbalanced) sizes of the classes in the classification.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :With the emergence of popular social media services such as Twitter and Facebook, many studies in the area of Natural Language Processing (NLP) have been published that analyze the text data from these services for a variety of applications, such as opinion mining, sentiment analysis, event detection, or crisis management (Culotta 2010; Sriram et al. 2010; Yin et al. 2012). Many of these studies have primarily relied on building classification models for different learning tasks, such as text classification or Named Entity Recognition. The effectiveness of these models is often evaluated using cross-validation techniques.
Cross-validation, first introduced by Geisser (1974), has been acclaimed as the most popular evaluation method for estimating prediction errors in regression and
∗ CSIRO, Sydney, New South Wales, Australia. E-mail: {sarvnaz.karimi, jie.jin}@csiro.au. ∗∗ Sabik Software Solutions Pty Ltd, Sydney, New South Wales, Australia. E-mail: jiri@baum.com.au.
Submission received: 4 October 2013; revised submission received: 9 September 2014; accepted for publication: 10 November 2014.
doi:10.1162/COLI a 00230
© 2015 Association for Computational Linguistics
classification problems. In that method, the data D are randomly partitioned into k non-overlapping subsets (folds) Dk of approximately equal size. The validation step is repeated k times, using a different Dv = Dk as the validation data and Dt = D \ Dk as the training data each time. The final evaluation is the average over the k validation steps. Cross-validation is found to have a lower variance than a single hold-out set validation and thus it is commonly used on both moderate and large amounts of data without introducing efficiency concerns (Arlot and Celisse 2010). Compared with other choices of k, 10-fold cross-validation has been accepted as the most reliable method, which gives a highly accurate estimate of the generalization error of a given model for a variety of algorithms and applications.
Despite its wide applications, debates on the appropriateness of cross-validation have been raised in a number of areas, particularly in time series analysis (Bergmeir and Benı́tez 2012) and chemical engineering (Sheridan 2013). A fundamental assumption of cross-validation is that the data need to be independent and identically distributed (i.i.d.) between folds (Arlot and Celisse 2010). Therefore, if the data points used for training and validation are not independent, cross-validation is no longer valid and would usually overestimate the validity of the model. For time series forecasting, because the data are comprised of correlated observations and might be generated by a process that evolves over time, the training set and the validation set are not independent if randomly chosen for cross-validation. Researchers since the early 1990s have used modified variants of cross-validation to compensate for time dependence within times series (Chu and Marron 1991; Bergmeir and Benı́tez 2012). In the area of chemical engineering, the work of (Sheridan 2013) investigates the dependence in chemical data and observes that the existence of similar compounds or molecules across the data set leads to overoptimistic results using standard k-fold cross-validation.
We take such observations from the time series and chemical domains as a warning to investigate the data dependence in computational linguistics. We argue that even when the data appear to be independently generated and when there is no reason to believe that temporal dependencies are present, unexpected statistical dependencies may be induced through an incorrect application of cross-validation. Once there is a chance of having similar or otherwise dependent data points, random split of the data without taking this factor into account would cause incorrect or at least unreliable evaluation, which may lead to invalid or at least unjustified conclusions. Although similar concerns have been raised by a prior study (Lerman et al. 2008) that cross-validation might not be suitable for measuring the accuracy of public opinion forecasting, there is a lack of systematic analysis on how potential data dependency might invalidate cross-validation and what alternative evaluation methods exist. With the aim of gaining further insight into this issue, we perform a detailed empirical study based on text classification in Twitter, and show that inappropriate choice of cross-validation techniques could potentially lead to misleading conclusions. This concern could apply more generally to other data types of similar nature. We also explore several evaluation methods, mostly borrowed from research on time series, that are better suited for statistically dependent data and that could be adapted by researchers working in the NLP area. 2 unexpected statistical dependence in microblog data :We argue that microblog data, such as Twitter messages, known as tweets, are statistically dependent in nature. It has been demonstrated that there is redundancy in
Twitter language (Zanzotto, Pennacchiotti, and Tsioutsiouliklis 2011), which in turn can translate to evidence against statistical independence of microblog posts in given periods of time or occurrences of the same event. In summary, statistical dependence in tweets can arise for the following reasons:
Events The events of interest, or events being discussed by the microbloggers at large, may be temporally limited within certain specific time horizons; Textual Links Hashtags and other idiomatic expressions may be invented, gain and lose popularity, fall out of use, or be reused with a different meaning. In addition, particular microbloggers may actively share information on certain types of events or express their opinions on certain topics in similar contexts; Twinning There may be “twinning” of tweets (and therefore data points) because of various forms of retweeting, where different microbloggers post substantially the same tweets, in response to one another.
Many existing studies that use microblog data (e.g., for tweet classification)—especially those related to detecting or monitoring events—have adopted k-fold cross-validation to evaluate the effectiveness of the learned classification models (e.g., Culotta 2010; Sriram et al. 2010; Jiang et al. 2011; Uysal and Croft 2011; Takemura and Tajima 2012; Kumar et al. 2014). However, they overlook the possible statistical dependence among microblog data in terms of content (e.g., sharing hashtags in Twitter), relevance to current events, and time of publishing that could potentially impact their evaluation results. In Table 1 we use examples of these studies to illustrate how potential statistical dependence of microblog data is omitted, which might have impacted the validity of the results in these studies. The potential source of the ignored statistical dependence shown in the last column is categorized into Events, Textual Links, and Twinning (described above), based on the list of features that the authors have used in their machine learning approaches. For example, if one is interested in detecting specific events (e.g., diseases [Culotta 2010] or disasters [Verma et al. 2011; Kumar, Jiang, and Fang 2014]), using tweets from the middle of an event to train for and evaluate the detection of the onset of that same event is clearly invalid. In addition, certain users (e.g., authorities that announce disease outbreaks such as @ECDC Outbreaks, or car accidents such as @emergencyAUS) may always post in the same way [Verma et al. 2011; Kumar, Jiang, and Fang 2014], making tweets share similar contexts; or microbloggers always use the same hashtags to indicate similar topics (Jiang et al. 2011). More generally, if substantially verbatim “twin” copies of tweets find their way into both the training and validation data sets (e.g., [Uysal and Croft 2011; Verma et al. 2011; Kumar, Jiang, and Fang 2014]), model validity will be overestimated. We note that, as we do not have access to the data sets used in these studies, we cannot verify the level of influence that the data dependence and their choice of evaluation have on their reported results. However, we raise a warning that there might be an overlooked unexpected dependence influencing their results. 3 validation for statistically dependent data :Where the data is not independent, dependence-aware validation needs to be used. We describe four dependence-aware validation methods that take data dependence into consideration in the evaluation. In the following discussion, we refer to the data set used in the experiments as D = {d1, d2, · · · , dn}. The training data is referred to as Dt, and testing, evaluation or validation data as Dv, where both Dt and Dv are subsets of D. 4 a case study on tweet classification :We focus on two tweet classification tasks in our case study. The first is a binary classification, where tweets are classified into disaster-related or not; the second is a disaster type classification, where tweets are predicted to one of the following six classes: nondisaster, storm, earthquake, flooding, fire, and other disasters. In our experiments, we use LibSVM (Chang and Lin 2011) to build discriminative classifiers for our classification tasks.
Tweets are short texts currently limited up to 140 characters. Often, microbloggers use tweets in reply to others by using mentions (which are Twitter usernames and are preceded by @), or use hashtags such as #CycloneEvan to make the grouping of similar messages easier, or to increase the visibility of their posts to others interested in the
same topic. Use of links to Web pages, mostly full stories of what is briefly reported in the tweet, is also popular. Selecting features for a text classifier that is built on Twitter data can therefore benefit from both conventional textual features, such as n-grams, and Twitter-specific features, such as hashtags and mentions.
In our experiments, we investigate the effect of the following features and their combinations on classification of tweets for disasters: (1) n-grams: Unigrams and bigrams of tweet text at the word level, excluding any hashtag or mention in the text. To find n-grams we pre-process tweets to remove stopwords; (2) Hashtag: Two different features are explored. First, a binary feature of the hashtags in the tweets, which indicates whether a hashtag exists in the tweet or not; second, the total number of hashtags in a tweet; (3) Mention: Two types of features (binary and mention count) are explored, exactly the same as hashtags explained above; and (4) Link: A binary feature that specifies whether or not a tweet contains any link to a Web page. We randomly sampled a total of 7,500 English tweets published in two years from December 2010 to December 2012 from a system (Yin et al. 2012) that stores billions of tweets from the Twitter streaming API. Explicit retweets were excluded. This set contained a number of disasters such as earthquakes in Christchurch, New Zealand 2011; York floods, England 2012; and Hurricane Sandy, United States 2012.
For a machine learner such as a classifier to work, we need to present it with a representative set of labeled training data. We therefore annotated our tweet data manually to identify disaster tweets and their types. We annotated the data set based on two main questions: Is this tweet talking about a specific disaster? What type of disaster is it talking about? Types of disasters were defined as earthquake, fire, flooding, storm, other, and non-disaster. Annotations were done by three annotators for each tweet, who were hired through the crowd-sourcing service Crowdflower. After taking majority votes where at least two out of the three annotators agreed on both questions, we ended up with a set of 6,580 annotated tweets, of which 2,897 tweets were identified as disaster-related and 3,683 as non-disaster. In disaster tweets, 37% were annotated with earthquake, followed by fire, flooding, and storm constituting 23%, 21%, and 12%, respectively. For our tweet classification tasks, we set up experiments to compare five validation methods—standard 10-fold cross-validation, border-split cross-validation, neighborsplit validation, time-split validation, and time-border-split validation—for identifying tweets that are relevant to a disaster, and whether or not we can broadly identify the type of disaster. We evaluate classification effectiveness using the accuracy metric, which is the percentage of correctly classified tweets. We used the following settings for our evaluation schemes:
r k-fold cross-validation: We used k = 10 folds.r Border-split cross-validation: We used k = 10 folds for border-split and a radius h = 20 days. We assume that for most events, the social media activity on the topic dies after three weeks of their occurrence.
r Neighbor-split validation: We used cosine similarity to find a subset of the data that has the least number of neighbors in the data set. We weighted hashtags double and disregarded mentions in calculating the similarity. Neighborhood was decided using a minimum threshold of 0.25 on the resulting cosine similarity. The size of the test data set was the same as in time-split (below) but the size of the training data set was 5,868.r Time-split validation: We chose the cut-off time t∗ so that 90% of the data was used as training (5,922 tweets) and 10% as validation (658 tweets).r Time-border-split validation: We used the same t∗ and h as time-split and border-split. This resulted in a training data set containing 87.1% of the data and a validation data set identical to the one in time-split at 10% of the data (658 tweets). In removing the border, 2.9% of the data were discarded. The size of the training set was 5,750 tweets. We run two sets of experiments: Discriminating disaster tweets from non-disaster (Disaster or Not), and classifying tweets into the six classes of earthquake, fire, flooding, storm, other, and non-disaster (Disaster Type). Table 2 compares the classification results for SVM on a range of feature combinations using five different validation methods. We aim to show that inappropriate choice of cross-validation could possibly lead to misleading conclusions, including overestimated classification accuracies, and suboptimal feature sets that yield the highest accuracies.
k-fold Cross-Validation. The first set of experiments was conducted using standard 10-fold cross-validation. For discriminating disaster tweets from non-disaster, SVM achieved a maximum of 92.8% accuracy when unigrams and hashtags were used. A similar result of 92.7% accuracy was recorded for classifying tweets to their disaster types. Unsurprisingly, having hashtags as additional features was the most
helpful. As a side note, the standard deviations of 10-fold cross-validation are quite small; although it is well-known that these are not an unbiased estimate of the true variance (see, for instance, Bengio and Grandvalet [2004]), it nevertheless seems to suggest a degree of confidence. If we would assume that tweets are statistically independent, we could conclude that the SVM classifier using unigrams plus hashtags is the best performer for classifying disaster tweets with a high accuracy of over 92%. This is what most previous studies have done. However, this result is overoptimistic. For example, during the cross-validation, tweets, with the same hashtags are distributed over all folds, making it easier for the classifier to associate labels with known hashtags.
Border-Split Cross-Validation. In contrast to 10-fold cross-validation, border-split cross-validation gives a much lower performance score across all of the results. The standard deviations are also much larger.
Neighbor-split Validation. The neighbour-split results are very different. Unlike previous work (Sheridan 2013), we find that neighbor-split judges the effectiveness of the models quite highly, near the scores from 10-fold cross-validation but ranking the feature combinations differently. We believe the high scores are because it tends to pick non-disaster tweets into the validation set, as those tweets have fewest neighbors. This makes it easier to classify all validation tweets into the majority class (i.e., non-disasters).
Time-split and Time-border-split Validation. Neither of these validations leads to results similar to 10-fold cross-validation; they are substantially lower, in one case by over 20 percentage points (bottom row of table). Numerically, they are similar to the border-split cross-validation results, but again with a different ranking of the feature combinations. Time-split and time-border-split methods gave similar results to each other, with different rankings but only small numerical differences. Coincidentally, in our data set, the time t∗ fell largely between events of interest, so the elimination of the border resulted in only a small correction, confounded with the effect of the small reduction in training set size. The validation methods provide very different results, with a number of the differences around 20 percentage points. In addition, they disagree on the ranking of the feature combinations, which may be the more important problem in many studies. This is particularly visible with the Disaster or Not classification; the best combination of features (bold in Table 2) is different, depending on the validation method used. A study based on 10-fold cross-validation would suggest Unigram+Hashtag for this problem, which is the second-worst combination of features, according to time-border-split validation. Further, the confidence intervals, if used, would suggest confidence in this choice.
Given our experiments, we believe that the standard k-fold cross-validation substantially overestimates the performance of the models in our case study. Time-split and time-border-split validation are more likely to represent accurate evaluation when temporally dependent data is involved, as they simulate true prospective prediction. However, they both have the downside that they only rely on one pair of training and validation data sets. This means that they will have a larger variance than a
cross-validation method, and do not give any measure of confidence in their own results.
Border-split cross-validation may also be acceptable, provided that its assumptions are satisfied—primarily, that the data points are independent beyond a certain radius. Depending on a particular task, this may be easy to decide (e.g., separate events) or hard (e.g., a number of overlapping events with no clear time difference).
We also find that neighbor-split overestimates rather than underestimates the performance of the models relative to time-split and time-border-split. This represents a hazard to its use: The neighborhood measure may interact in unexpected ways with other features of the data, rendering the validation unpredictable.
Ideally, one would use multiple large test data sets, all collected during separate, non-overlapping periods of time, after the training data set. This would be the most reliable measure of the performance of the models, but also the most expensive in terms of required data collection and annotation. Failing that, we recommend time-split or time-border-split validation when temporal statistical dependence cannot be ruled out. 5 conclusions :We used a common task in NLP, text classification, on a relatively recent but widely used data source, Twitter streams, to show that blindly following the same evaluation method for tasks of similar nature could lead to invalid conclusions. In particular, we investigated the most common evaluation method for machine learning applications, standard 10-fold cross-validation, and compared it with other validation methods that take the statistical dependence in the data into account, including time-split, bordersplit, neighbor-split, and a combination of time- and border-split validation. We showed how cross-validation can overestimate the effectiveness of a tweet classification application. We argued that text in microblogs or other similar text from social media (e.g., Web forums or even online news) can be statistically dependent for specific studies, such as those looking at events. Researchers therefore need to be careful in choosing evaluation methodologies based on the nature of the data at hand to avoid bias in their results. In recent years, many studies have been published on data collected from social media, especially microblogs such as Twitter. However, rather few of these studies have considered evaluation methodologies that take into account the statistically dependent nature of such data, which breaks the theoretical conditions for using cross-validation. Despite concerns raised in the past about using cross-validation for data of similar characteristics, such as time series, some of these studies evaluate their work using standard k-fold cross-validation. Through experiments on Twitter data collected during a two-year period that includes disastrous events, we show that by ignoring the statistical dependence of the text messages published in social media, standard cross-validation can result in misleading conclusions in a machine learning task. We explore alternative evaluation methods that explicitly deal with statistical dependence in text. Our work also raises concerns for any other data for which similar conditions might hold. [{""affiliations"": [], ""name"": ""Sarvnaz Karimi""}, {""affiliations"": [], ""name"": ""Jie Yin""}, {""affiliations"": [], ""name"": ""Jiri Baum""}] SP:17ced766be8b5ad0aeadb2cbd25ee9d31eb0e63a [{""authors"": [""S. Arlot"", ""A. Celisse.""], ""title"": ""A survey of cross-validation procedures for model selection"", ""venue"": ""Statistics Surveys, 4:40\u201379. Bengio, Y. and Y. Grandvalet. 2004. No"", ""year"": 2010}, {""authors"": [""C. Lin""], ""title"": ""LIBSVM: A library for support vector machines"", ""venue"": ""Information Sciences,"", ""year"": 2011}, {""authors"": [""Chu"", ""C.-K"", ""J.S. Marron""], ""title"": ""Comparison of two bandwidth selectors"", ""year"": 1991}, {""authors"": [""S. Geisser""], ""title"": ""A predictive approach to the random effect mode"", ""venue"": ""Biometrika Trust, 61(9):101\u2013107."", ""year"": 1974}, {""authors"": [""L. Jiang"", ""M. Yu"", ""M. Zhou"", ""X. Liu"", ""T. Zhao.""], ""title"": ""Target-dependent Twitter sentiment classification"", ""venue"": ""ACL-HLT, pages 151\u2013160, Portland, OR."", ""year"": 2011}, {""authors"": [""A. Kumar"", ""M. Jiang"", ""Y. Fang.""], ""title"": ""Where not to go?: Detecting road hazards using Twitter"", ""venue"": ""SIGIR, pages 1223\u20131226, Gold Coast."", ""year"": 2014}, {""authors"": [""K. Lerman"", ""A. Gilder"", ""M. Dredze"", ""F. Pereira.""], ""title"": ""Reading the markets: Forecasting public opinion of political candidates by news analysis"", ""venue"": ""COLING, pages 473\u2013480, Manchester."", ""year"": 2008}, {""authors"": [""R.P. Sheridan""], ""title"": ""Time-split cross-validation as a method for estimating the goodness of prospective prediction"", ""venue"": ""Journal of Chemical Information and Modeling,"", ""year"": 2013}, {""authors"": [""B. Sriram"", ""D. Fuhry"", ""E. Demir"", ""H. Ferhatosmanoglu"", ""M. Demirbas.""], ""title"": ""Short text classification in Twitter to improve information filtering"", ""venue"": ""SIGIR, pages 841\u2013842, Geneva."", ""year"": 2010}, {""authors"": [""H. Takemura"", ""K. Tajima.""], ""title"": ""Tweet classification based on their lifetime duration"", ""venue"": ""CIKM, pages 2367\u20132370, Maui, HI."", ""year"": 2012}, {""authors"": [""I. Uysal"", ""W.B. Croft.""], ""title"": ""User oriented tweet ranking: A filtering approach to microblogs"", ""venue"": ""CIKM, pages 2261\u20132264, Glasgow."", ""year"": 2011}, {""authors"": [""S. Verma"", ""S. Vieweg"", ""W.J. Corvey"", ""L. Palen"", ""J.H. Martin"", ""M. Palmer"", ""A. Schram"", ""K.M. Anderson""], ""title"": ""Natural language processing to the rescue? Extracting \u201csituational awareness"", ""year"": 2011}, {""authors"": [""J. Yin"", ""S. Karimi"", ""B. Robinson"", ""M. Cameron.""], ""title"": ""ESA: Emergency situation awareness via microbloggers"", ""venue"": ""CIKM, pages 2701\u20132703, Maui, HI."", ""year"": 2012}, {""authors"": [""F.M. Zanzotto"", ""M. Pennacchiotti"", ""K. Tsioutsiouliklis.""], ""title"": ""Linguistic redundancy in Twitter"", ""venue"": ""EMNLP, pages 659\u2013669, Edinburgh. 548"", ""year"": 2011}] squibs  : evaluation methods for statistically : dependent text :Sarvnaz Karimi∗ csiro  :Jie Yin∗ Jiri Baum∗∗ Sabik Software Solutions
In recent years, many studies have been published on data collected from social media, especially microblogs such as Twitter. However, rather few of these studies have considered evaluation methodologies that take into account the statistically dependent nature of such data, which breaks the theoretical conditions for using cross-validation. Despite concerns raised in the past about using cross-validation for data of similar characteristics, such as time series, some of these studies evaluate their work using standard k-fold cross-validation. Through experiments on Twitter data collected during a two-year period that includes disastrous events, we show that by ignoring the statistical dependence of the text messages published in social media, standard cross-validation can result in misleading conclusions in a machine learning task. We explore alternative evaluation methods that explicitly deal with statistical dependence in text. Our work also raises concerns for any other data for which similar conditions might hold. reference study data dependence concern :Sriram et al. (2010) Classification of tweets into five categories of events, news, deals, opinions, and private messages 5,407 tweets, no mention of removing retweets Events, Textual Links, Twinning
Culotta (2010) Classification of tweets using regression models into flu or not flu
500,000 tweets from a 10-week period Events, Textual Links, Twinning
Verma et al. (2011) Classification of tweets to specific disaster events, and identification of tweets that contained situational awareness content 1,965 tweets collected from specific disasters in the U.S. by keyword search
Twinning, Textual Links
Jiang et al. (2011) Sentiment classification of tweets; hashtags were mentioned as one of the features, retweets were considered to share the same sentiment Tweets found using keyword search, without removing retweets
Twinning, Textual Links
Uysal and Croft (2011)
Personalized tweet ranking using retweet behavior. Decision tree classifiers were used based on features from the tweet content, user behavior, and tweet author 24,200 tweets, from which 2,547 were retweeted by the seed users
Twinning, Textual Links
Takemura and Tajima (2012) Classification of tweets into three categories based on whether they should be read now, later, or outdated
9,890 tweets from a fixed period of time, annotated for time-(in)dependency Events, Textual Links, Twinning
Kumar, Jiang, and Fang (2014)
Classification of tweets into two categories of road hazard and non-hazard
30,876 tweets, retweets were not removed
Events, Textual Links, Twinning
Border-Split Cross-Validation. This method, proposed by Chu and Marron (1991), is a modification of k-fold cross-validation for time series data. Data are partitioned into the folds in time sequence rather than randomly; then in each validation step, data points within a time distance h of any point in the validation data set are excluded from the training data. That is, for each validation step k, the validation data is Dv = Dk, there is a border data set Db = {db ∈ D \ Dv : (∃dv∈Dv) |tdb − tdv | ≤ h}, which is disregarded, and the training data is Dt = D \ (Dk ∪ Db). This method assumes that the data dependence is confined to some known radius h, with data points beyond that radius being independent.
Time-Split Validation. In time-split validation (Sheridan 2013), a particular time t∗ is chosen; data points prior to this time are allocated to the training data set and data points after t∗ are used as validation data; that is, Dt = {dt ∈ D : tdt ≤ t
∗} and Dv = {dv ∈ D : tdv > t
∗}. Note that this is not a cross-validation method, as no “crossing” takes place. The motivation is to emulate prospective prediction: A model is built using only the information available up to time t∗ and evaluated on data that are collected after that time.
Time-Border-Split Validation. This is a combination of border-split and time-split validation. Both a time t∗ and a radius h are chosen; data points prior to time t∗ − h are allocated to the training data set and data points after time t∗ are used as validation data; that is, Dt = {dt ∈ D : tdt ≤ t ∗ − h} and Dv = {dv ∈ D : tdv > t ∗}. The remaining data points are not used. The motivation is to combine the conservative aspects of both time-split and border-split, emulating prospective prediction more carefully.
Neighbor-Split Validation. Proposed by (Sheridan 2013), this method assumes the existence of some similarity metric for the data points. The number of neighbors for each data point is calculated, using a threshold on the similarity metric. A desired fraction of data points with the fewest neighbors are then allocated to the validation data Dv and the rest to the training data Dt. The motivation of this approach is to deliberately reduce the similarity between training and validation data. It is inspired by leave-class-out validation or cross-validation, which assumes a pre-existing classification (rather than similarity metric) and allocates data points to the validation data Dv or cross-validation folds Dk, according to this classification. The advantage of neighborsplit over leave-class-out is that the size of the validation data is a parameter that can be chosen, rather than being dependent on the (potentially unbalanced) sizes of the classes in the classification.",
,,,,,,,,"I volunteered to review this book because I found Dr. Markowitz’s 1995 book, Using Speech Recognition, extremely helpful when I first became involved in medical applications of speech. In fact, I made all my students in that area read it. This book is very different, but in its own way, just as useful. In fact, it is a must-read for anyone contemplating a new speech application or enhancing a current one, as well as for anyone searching for new directions in dialogue research.","[{""affiliations"": [], ""name"": ""Judith A. Markowitz""}]",SP:1f0848033050983612eb42224be414133f2cf75b,"[{""authors"": [""K. Bhatt"", ""S. Argamon"", ""M.W. Evens""], ""title"": ""Hedged responses and expressions of affect in human/human and human/ computer tutorial interactions"", ""venue"": ""In Proceedings of COGSCI"", ""year"": 2004}, {""authors"": [""B. Reeves"", ""C. Nass""], ""title"": ""The media equation: How people treat computers, television, and new media like real people and places"", ""year"": 1996}]",,,,,,,,,,,,,,,,book reviews :,robots that talk and listen :,"judith a. markowitz (editor) :(President of J. Markowitz, Consultants, Chicago, IL)
Berlin/Boston/Munich: Walter de Gruyter, 2015, hardbound, ISBN 978-1-61451-603-3; PDF, e-ISBN 978-1-61451-440-4 EPUB; e-ISBN 978-1-61451-915-7","reviewed by martha evens :Illinois Institute of Technology
I volunteered to review this book because I found Dr. Markowitz’s 1995 book, Using Speech Recognition, extremely helpful when I first became involved in medical applications of speech. In fact, I made all my students in that area read it. This book is very different, but in its own way, just as useful. In fact, it is a must-read for anyone contemplating a new speech application or enhancing a current one, as well as for anyone searching for new directions in dialogue research.
© 2015 Association for Computational Linguistics
This book is divided into four parts, followed by a separate conclusion. It is always hard to do justice to all the papers and authors in a collection like this—with twelve varied papers addressing many different parts of the complex problem of enabling a robot to carry on a spoken dialogue—but I will try to give at least a brief taste of each.
Part I is about “Images.” In the first chapter of Part I, Steve Mushkin, the founder and president of Latitude Research describes an experiment in which 348 children aged 8–12 from six digitally-advanced countries were asked to draw and describe in words how they would like to interact with a personal robot in the future. In the second chapter Markowitz herself reviews the language capabilities and behavior of powerful cultural icons, such as Frankenstein, the Hebrew Golem, the Japanese karakuri, and Pygmalion’s beloved statue, Galatea. This is highly relevant, because science fiction, whether on paper or in the movies, part of traditional mythology, or local social history, has largely shaped our expectations about what new technology promises or threatens. In the third chapter, David Dufty, author of the well-known book How to Build an Android: The True Story of Philip K. Dick’s Robotic Resurrection, believes that robots are intrinsically part of art and entertainment, and that is the major reason why we should build them.
Part II is called “Frameworks and Guidelines.” The first of the three chapters in this part, written by Bilge Mutlu and two of his students at the University of Wisconsin, Madison, discusses the requirements for designing a robot capable of sustaining productive dialogue. They test their framework in robot-human interactions; first using a robot as a teacher, then to assess linguistic factors that contribute to expertise. Director of Columbia University’s Japanese Language Program, focuses on the knowledge that any teacher must have to be effective in teaching a foreign language. Based on her series of research studies on the teaching of Japanese, her concern is that a robot (or human) teacher must understand the bond between language and culture, which she sees as inseparable. She emphasizes the importance of using human-like facial expressions in the teaching machine, at first for teaching pronunciation, for communicating emotional
doi:10.1162/COLI r 00224
reactions to the student’s contributions, and making the student feel comfortable with the machine. In this excellent paper, chock full of valuable references, Nazikian unintentionally reminds us of the huge and inexplicable divide between the world of Computer-Aided Language Learning (CALL) and the world of AI in education. Art Graesser and his group at the University of Memphis (Link et al., 2001) have carried out a long series of experiments studying the effect of using expressive faces on the screen coordinated with the dialogue, but the work at Memphis is never mentioned in this paper and there are no references to papers in Discourse Processes or the Journal of the Learning Sciences. The third paper in Part II, written by Manfred Tscheligi and his student, Nicole Mirnig, at the University of Salzburg, focuses on the problems of Comprehension, Coherence, and Consistency. These qualities depend, they argue, on the use and transfer of mental models, which they find essential to effective dialogue with any robot capable of learning what human beings want and giving it to them.
Part III is about constructing a speech-enabled robot capable of useful, real-world tasks. In Chapter 7, Jonathan Connell of IBM’s T. J. Watson Research Center explains how he defined these capabilities for a robot named ELI (the Extensible Language Interface). First of all, such a robot needs to be able to accept and synthesize multimodal information. It needs to recognize speech, gestures, and object manipulations and put this information together into an algorithm for getting you a cup of coffee? It has to find your kitchen, pick up the right coffee pot, heat the water in a microwave safe measuring cup, and put in just the right amount of coffee, cream, and sugar. In the process, it will need to learn new nouns (the names that you use for objects and rooms in your house) and new verbs (for activities like opening bottles and boiling water). Their robot parser is a finite state machine, but it can handle new names and new activities. The emphasis is on the grounding of new language in a new environment. In Chapter 8, Alan Wagner, a Senior Research Scientist at Georgia Tech, outlines what a robot has to know before you can teach it to tell lies or recognize lies produced by others. Unlike ELI, this robot does not yet exist and I find it hard to imagine wanting to build it or even use it myself, but the description of deceptive language in this chapter has charm. Joerg Wolf and Guido Bugmann, at the University of Plymouth in the UK, recount their experience getting university students to teach their off-the-shelf robot how to play a simple card game. They began by collecting a corpus of sessions in which one university student taught another to play the game. They used this corpus to develop the rules for their robot, grammar rules, rules for anaphora, dialogue rules, rules for dealing cards, and then they carried out a series of experiments in which university students tried to teach the robot the game. One of many problems that they had not anticipated was that when the robot did not understand what the student wanted him to do, the student raised his or her voice and tried to simplify the explanation. The result was many out-of-grammar and out-of-vocabulary errors.
Part IV contains two very different papers, one about improving audition (so that the robot can understand what you are telling it, even when several people are talking at once or machines are clanking) and the other about how design factors, such as robot voices and gesture speed affect memory and engagement in both children and adults. The experts in audition, François Grondin and François Michaud from the Robotics Laboratory at the Canadian Université de Sherbrooke, have developed algorithms for localization (figuring out where a sound is coming from), tracking (following sounds as people or robots move), and separation (extracting sounds from a given person or machine when several are making noises at once). They have achieved remarkable success in a number of experiments run on a system with a bank of separate microphones and
parallel hardware capable of processing all this information at once. Sandra Okita, then at Stanford, now at Columbia University, and Victor Ng-Thow-Hing from the Honda Research Institute USA in Mountain View teamed up to carry out a series of experiments with robot voices. The objective of their studies was to determine how design choices like robot voices and gesture speed effect how humans of all ages respond to a robot. They tried out two voices with young children, one robot-like (monotone) and one more human-like. Both robots followed exactly the same script, but the children remembered much more of what the humanoid voice said and they were much more likely to be willing to talk to that robot again. Teenagers preferred the humanoid voice, but there was much less difference between the two groups interacting with the different voices. Adults seemed to be affected even less. Similar experiments with gesture speed showed that robots who gestured rapidly were judged to be happier and more pleasant to spend time with. Again, small children were affected the most, but teenagers also had strong preferences here. Again, adults were affected less.
We badly need more work of the kind described by Okita and Ng-Thow-Hing. Too many people believed Reeves and Nass’s (1996) Media Equation, which is the claim that people interact with computers in the same way that they interact with people; and decided that they could substitute research on human interaction for research on interactions between people and machines. Our own experiments (Bhatt, Argamon, and Evens, 2004) have convinced us that the Media Equation is certainly not true for the students interacting with our intelligent tutoring system, who are constantly polite to human tutors, whether their professors or their peers, but often very rude to our computer tutor.
The concluding chapter in the book, written by Roger K. Moore, Professor of Computer Science at the University of Sheffield and the Editor-in-Chief of Computer Speech and Language, takes us back to the future where the first chapter began. He argues that we have a long way to go to meet the goal of intelligent communication with machines. We need to build robots that understand human behavior and are capable of generating the language necessary to change it.
Moore essentially focuses this chapter on his own view of the major conflict of opinion and approach separating the experts featured in this book. It will never be enough to simply tack on a multi-purpose language module to an existing robot, he argues. Spoken communication is a fundamental and integral part of ordinary human beings; and if we are to succeed in our goal of constructing “intelligent communicative machines,” we need to make the communication function a central part of the machine design and not simply an add-on function.
There is a major gulf between Moore’s argument and the experts quoted by Dufty back in the third chapter (page 55). Dufty quotes Sylvia Solon, deputy editor of Wired in the UK, as saying, “There is no point making robots look and act like humans.” Dufty adds that Martin Robbins, who blogs for the Guardian, under the name “The Lay Scientist,” wrote: “Humanoid robots are seen as the future, but for almost any specific task you can think of, a focused simple design will work better.” Of course, Dufty goes on to tell us about the use of robots in art and entertainment and about the workings of the Philip K. Dick android, so he may not agree entirely with Solon and Robbins.
There is a cogent case on the opposite shore of this gulf in at least three of the other chapters (4, 7, and 9). In Chapter 4, Mutlu’s group at the University of Wisconsin describe the simple semantic grammar used by their robot (a Wakamaru), along with more sophisticated dialogue and behavior models and algorithms to control gaze and gestures. Jonathan Connell begins Chapter 7 with “Suppose you buy a general fetch-and-carry robot from Sears and take it home” and continues in this vein. In
Chapter 9, Wolf and Bugmann designed their system to operate on components made of production rules, but they still have a lot to teach us.
Personally, I am on a raft drifting in the middle of this gulf. I think that we need thoughtful experiments with existing robots now, to help us discover the problems, but I believe that really satisfactory solutions will require much more research. I cannot agree with Moore, however, when he argues that it is unethical to attempt to imitate human communication by stitching together the limited technologies we have today. It seems to me that much of what we now know about human communication comes from attempting to do just that.
In the process of reading this book I discovered that I had not heard about two other collections edited by Markowitz and published by Springer in 2013. Both of these were edited by Neustein & Markowitz jointly. Dr. Amy Neustein is founder and CEO of Linguistic Technology Systems and Editor-in-Chief of the International Journal of Speech Technology. The first is entitled Mobile Speech and Advanced Natural Language Solutions (hardbound ISBN 978-1-4614-6017-6; e-Book ISBN 978-1-4614-6018-3); the second, Where Humans Meet Machines: Innovative Solutions for Knotty Natural Language Problems (ISBN 978-1-4614-6933-9, eBook ISBN 978-1-4614-6934-6).",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"I volunteered to review this book because I found Dr. Markowitz’s 1995 book, Using Speech Recognition, extremely helpful when I first became involved in medical applications of speech. In fact, I made all my students in that area read it. This book is very different, but in its own way, just as useful. In fact, it is a must-read for anyone contemplating a new speech application or enhancing a current one, as well as for anyone searching for new directions in dialogue research. [{""affiliations"": [], ""name"": ""Judith A. Markowitz""}] SP:1f0848033050983612eb42224be414133f2cf75b [{""authors"": [""K. Bhatt"", ""S. Argamon"", ""M.W. Evens""], ""title"": ""Hedged responses and expressions of affect in human/human and human/ computer tutorial interactions"", ""venue"": ""In Proceedings of COGSCI"", ""year"": 2004}, {""authors"": [""B. Reeves"", ""C. Nass""], ""title"": ""The media equation: How people treat computers, television, and new media like real people and places"", ""year"": 1996}] book reviews : robots that talk and listen : judith a. markowitz (editor) :(President of J. Markowitz, Consultants, Chicago, IL)
Berlin/Boston/Munich: Walter de Gruyter, 2015, hardbound, ISBN 978-1-61451-603-3; PDF, e-ISBN 978-1-61451-440-4 EPUB; e-ISBN 978-1-61451-915-7 reviewed by martha evens :Illinois Institute of Technology
I volunteered to review this book because I found Dr. Markowitz’s 1995 book, Using Speech Recognition, extremely helpful when I first became involved in medical applications of speech. In fact, I made all my students in that area read it. This book is very different, but in its own way, just as useful. In fact, it is a must-read for anyone contemplating a new speech application or enhancing a current one, as well as for anyone searching for new directions in dialogue research.
© 2015 Association for Computational Linguistics
This book is divided into four parts, followed by a separate conclusion. It is always hard to do justice to all the papers and authors in a collection like this—with twelve varied papers addressing many different parts of the complex problem of enabling a robot to carry on a spoken dialogue—but I will try to give at least a brief taste of each.
Part I is about “Images.” In the first chapter of Part I, Steve Mushkin, the founder and president of Latitude Research describes an experiment in which 348 children aged 8–12 from six digitally-advanced countries were asked to draw and describe in words how they would like to interact with a personal robot in the future. In the second chapter Markowitz herself reviews the language capabilities and behavior of powerful cultural icons, such as Frankenstein, the Hebrew Golem, the Japanese karakuri, and Pygmalion’s beloved statue, Galatea. This is highly relevant, because science fiction, whether on paper or in the movies, part of traditional mythology, or local social history, has largely shaped our expectations about what new technology promises or threatens. In the third chapter, David Dufty, author of the well-known book How to Build an Android: The True Story of Philip K. Dick’s Robotic Resurrection, believes that robots are intrinsically part of art and entertainment, and that is the major reason why we should build them.
Part II is called “Frameworks and Guidelines.” The first of the three chapters in this part, written by Bilge Mutlu and two of his students at the University of Wisconsin, Madison, discusses the requirements for designing a robot capable of sustaining productive dialogue. They test their framework in robot-human interactions; first using a robot as a teacher, then to assess linguistic factors that contribute to expertise. Director of Columbia University’s Japanese Language Program, focuses on the knowledge that any teacher must have to be effective in teaching a foreign language. Based on her series of research studies on the teaching of Japanese, her concern is that a robot (or human) teacher must understand the bond between language and culture, which she sees as inseparable. She emphasizes the importance of using human-like facial expressions in the teaching machine, at first for teaching pronunciation, for communicating emotional
doi:10.1162/COLI r 00224
reactions to the student’s contributions, and making the student feel comfortable with the machine. In this excellent paper, chock full of valuable references, Nazikian unintentionally reminds us of the huge and inexplicable divide between the world of Computer-Aided Language Learning (CALL) and the world of AI in education. Art Graesser and his group at the University of Memphis (Link et al., 2001) have carried out a long series of experiments studying the effect of using expressive faces on the screen coordinated with the dialogue, but the work at Memphis is never mentioned in this paper and there are no references to papers in Discourse Processes or the Journal of the Learning Sciences. The third paper in Part II, written by Manfred Tscheligi and his student, Nicole Mirnig, at the University of Salzburg, focuses on the problems of Comprehension, Coherence, and Consistency. These qualities depend, they argue, on the use and transfer of mental models, which they find essential to effective dialogue with any robot capable of learning what human beings want and giving it to them.
Part III is about constructing a speech-enabled robot capable of useful, real-world tasks. In Chapter 7, Jonathan Connell of IBM’s T. J. Watson Research Center explains how he defined these capabilities for a robot named ELI (the Extensible Language Interface). First of all, such a robot needs to be able to accept and synthesize multimodal information. It needs to recognize speech, gestures, and object manipulations and put this information together into an algorithm for getting you a cup of coffee? It has to find your kitchen, pick up the right coffee pot, heat the water in a microwave safe measuring cup, and put in just the right amount of coffee, cream, and sugar. In the process, it will need to learn new nouns (the names that you use for objects and rooms in your house) and new verbs (for activities like opening bottles and boiling water). Their robot parser is a finite state machine, but it can handle new names and new activities. The emphasis is on the grounding of new language in a new environment. In Chapter 8, Alan Wagner, a Senior Research Scientist at Georgia Tech, outlines what a robot has to know before you can teach it to tell lies or recognize lies produced by others. Unlike ELI, this robot does not yet exist and I find it hard to imagine wanting to build it or even use it myself, but the description of deceptive language in this chapter has charm. Joerg Wolf and Guido Bugmann, at the University of Plymouth in the UK, recount their experience getting university students to teach their off-the-shelf robot how to play a simple card game. They began by collecting a corpus of sessions in which one university student taught another to play the game. They used this corpus to develop the rules for their robot, grammar rules, rules for anaphora, dialogue rules, rules for dealing cards, and then they carried out a series of experiments in which university students tried to teach the robot the game. One of many problems that they had not anticipated was that when the robot did not understand what the student wanted him to do, the student raised his or her voice and tried to simplify the explanation. The result was many out-of-grammar and out-of-vocabulary errors.
Part IV contains two very different papers, one about improving audition (so that the robot can understand what you are telling it, even when several people are talking at once or machines are clanking) and the other about how design factors, such as robot voices and gesture speed affect memory and engagement in both children and adults. The experts in audition, François Grondin and François Michaud from the Robotics Laboratory at the Canadian Université de Sherbrooke, have developed algorithms for localization (figuring out where a sound is coming from), tracking (following sounds as people or robots move), and separation (extracting sounds from a given person or machine when several are making noises at once). They have achieved remarkable success in a number of experiments run on a system with a bank of separate microphones and
parallel hardware capable of processing all this information at once. Sandra Okita, then at Stanford, now at Columbia University, and Victor Ng-Thow-Hing from the Honda Research Institute USA in Mountain View teamed up to carry out a series of experiments with robot voices. The objective of their studies was to determine how design choices like robot voices and gesture speed effect how humans of all ages respond to a robot. They tried out two voices with young children, one robot-like (monotone) and one more human-like. Both robots followed exactly the same script, but the children remembered much more of what the humanoid voice said and they were much more likely to be willing to talk to that robot again. Teenagers preferred the humanoid voice, but there was much less difference between the two groups interacting with the different voices. Adults seemed to be affected even less. Similar experiments with gesture speed showed that robots who gestured rapidly were judged to be happier and more pleasant to spend time with. Again, small children were affected the most, but teenagers also had strong preferences here. Again, adults were affected less.
We badly need more work of the kind described by Okita and Ng-Thow-Hing. Too many people believed Reeves and Nass’s (1996) Media Equation, which is the claim that people interact with computers in the same way that they interact with people; and decided that they could substitute research on human interaction for research on interactions between people and machines. Our own experiments (Bhatt, Argamon, and Evens, 2004) have convinced us that the Media Equation is certainly not true for the students interacting with our intelligent tutoring system, who are constantly polite to human tutors, whether their professors or their peers, but often very rude to our computer tutor.
The concluding chapter in the book, written by Roger K. Moore, Professor of Computer Science at the University of Sheffield and the Editor-in-Chief of Computer Speech and Language, takes us back to the future where the first chapter began. He argues that we have a long way to go to meet the goal of intelligent communication with machines. We need to build robots that understand human behavior and are capable of generating the language necessary to change it.
Moore essentially focuses this chapter on his own view of the major conflict of opinion and approach separating the experts featured in this book. It will never be enough to simply tack on a multi-purpose language module to an existing robot, he argues. Spoken communication is a fundamental and integral part of ordinary human beings; and if we are to succeed in our goal of constructing “intelligent communicative machines,” we need to make the communication function a central part of the machine design and not simply an add-on function.
There is a major gulf between Moore’s argument and the experts quoted by Dufty back in the third chapter (page 55). Dufty quotes Sylvia Solon, deputy editor of Wired in the UK, as saying, “There is no point making robots look and act like humans.” Dufty adds that Martin Robbins, who blogs for the Guardian, under the name “The Lay Scientist,” wrote: “Humanoid robots are seen as the future, but for almost any specific task you can think of, a focused simple design will work better.” Of course, Dufty goes on to tell us about the use of robots in art and entertainment and about the workings of the Philip K. Dick android, so he may not agree entirely with Solon and Robbins.
There is a cogent case on the opposite shore of this gulf in at least three of the other chapters (4, 7, and 9). In Chapter 4, Mutlu’s group at the University of Wisconsin describe the simple semantic grammar used by their robot (a Wakamaru), along with more sophisticated dialogue and behavior models and algorithms to control gaze and gestures. Jonathan Connell begins Chapter 7 with “Suppose you buy a general fetch-and-carry robot from Sears and take it home” and continues in this vein. In
Chapter 9, Wolf and Bugmann designed their system to operate on components made of production rules, but they still have a lot to teach us.
Personally, I am on a raft drifting in the middle of this gulf. I think that we need thoughtful experiments with existing robots now, to help us discover the problems, but I believe that really satisfactory solutions will require much more research. I cannot agree with Moore, however, when he argues that it is unethical to attempt to imitate human communication by stitching together the limited technologies we have today. It seems to me that much of what we now know about human communication comes from attempting to do just that.
In the process of reading this book I discovered that I had not heard about two other collections edited by Markowitz and published by Springer in 2013. Both of these were edited by Neustein & Markowitz jointly. Dr. Amy Neustein is founder and CEO of Linguistic Technology Systems and Editor-in-Chief of the International Journal of Speech Technology. The first is entitled Mobile Speech and Advanced Natural Language Solutions (hardbound ISBN 978-1-4614-6017-6; e-Book ISBN 978-1-4614-6018-3); the second, Where Humans Meet Machines: Innovative Solutions for Knotty Natural Language Problems (ISBN 978-1-4614-6933-9, eBook ISBN 978-1-4614-6934-6).",
,,,,,,,,"A long time ago now (maybe 1988?), Gerald (Gazdar) and I supervised Adam’s DPhil at the University of Sussex. Adam was my age, give or take a year, having come to academia a little late, and was my first doctoral student. Adam’s topic was polysemy, and I’m not really sure that much supervision was actually required, though I recall fun exchanges trying to model the subtleties of word meaning using symbolic knowledge representation techniques—an experience that was clearly enough to convince Adam later that this was a bad idea. In fact, Adam’s thesis title itself was Polysemy. Much as we encourage short thesis titles, pulling off the one-word title is a tall order, requiring a unique combination of focus and coverage, breadth and depth, and, most of all, authority. Adam completely nailed it, at least from the perspective of the pre-empirical Computational Linguistics of the early 1990s. Three years later, after a spell working for dictionary publishers, Adam joined me as a research fellow, now at the University of Brighton. I had a project to explore the automatic enrichment of lexical databases to support the latest trends in language analysis, and, in particular, task-specific lexical resources. I was really pleased and excited to recruit Adam—he had lost none of his intellectual independence, a quality I particularly valued. Within a few weeks he came to me with his own plan for the research—a “detour,” as he put it, from the original workplan. I still have the e-mail, dated 6 April 1995, in which he proposed that, instead of chasing a prescriptive notion of a single lexical resource that needed to be customized to each domain, we should let the domain determine the lexicon, providing lexicographic tools to explore words, and particularly word senses, that were significant for that domain. In that e-mail, Computational Lexicography at Brighton was born. Over the next eight years or so, Computational Lexicography became a key part of our group’s success, increasingly under Adam’s direct leadership. The key project, WASPS, developed the WASPbench—the direct precursor of the Sketch Engine, recruiting David (Tugwell) to the team. In addition, Adam was one of the founding organizers of SENSEVAL, an initiative to bring international teams of researchers together to work in friendly competition on a pre-determined word sense disambiguation task (and which has now transformed into SEMEVAL). Together we secured funding to support the first two rounds of SENSEVAL; each round required the preparation of standardized data sets, guided by Adam’s highly tuned intuitions about lexical data preparation and management. And we engaged somewhat in the European funding merry-go-round, most fondly in the CONCEDE project, working on dictionaries for Central European languages with amazing teams from the MULTEXT-EAST consortium, and with Georgian and German colleagues in the GREG project.","[{""affiliations"": [], ""name"": ""Adam Kilgarriff""}, {""affiliations"": [], ""name"": ""Roger Evans""}]",SP:abc692872b31b28c2a71c7bbf580d44025f08604,[],,,,,,,,,,,,,,,,,,,,"adam kilgarriff :Roger Evans∗ University of Brighton
A long time ago now (maybe 1988?), Gerald (Gazdar) and I supervised Adam’s DPhil at the University of Sussex. Adam was my age, give or take a year, having come to academia a little late, and was my first doctoral student. Adam’s topic was polysemy, and I’m not really sure that much supervision was actually required, though I recall fun exchanges trying to model the subtleties of word meaning using symbolic knowledge representation techniques—an experience that was clearly enough to convince Adam later that this was a bad idea. In fact, Adam’s thesis title itself was Polysemy. Much as we encourage short thesis titles, pulling off the one-word title is a tall order, requiring a unique combination of focus and coverage, breadth and depth, and, most of all, authority. Adam completely nailed it, at least from the perspective of the pre-empirical Computational Linguistics of the early 1990s.
Three years later, after a spell working for dictionary publishers, Adam joined me as a research fellow, now at the University of Brighton. I had a project to explore the automatic enrichment of lexical databases to support the latest trends in language analysis, and, in particular, task-specific lexical resources. I was really pleased and excited to recruit Adam—he had lost none of his intellectual independence, a quality I particularly valued. Within a few weeks he came to me with his own plan for the research—a “detour,” as he put it, from the original workplan. I still have the e-mail, dated 6 April 1995, in which he proposed that, instead of chasing a prescriptive notion of a single lexical resource that needed to be customized to each domain, we should let the domain determine the lexicon, providing lexicographic tools to explore words, and particularly word senses, that were significant for that domain. In that e-mail, Computational Lexicography at Brighton was born.
Over the next eight years or so, Computational Lexicography became a key part of our group’s success, increasingly under Adam’s direct leadership. The key project, WASPS, developed the WASPbench—the direct precursor of the Sketch Engine, recruiting David (Tugwell) to the team. In addition, Adam was one of the founding organizers of SENSEVAL, an initiative to bring international teams of researchers together to work in friendly competition on a pre-determined word sense disambiguation task (and which has now transformed into SEMEVAL). Together we secured funding to support the first two rounds of SENSEVAL; each round required the preparation of standardized data sets, guided by Adam’s highly tuned intuitions about lexical data preparation and management. And we engaged somewhat in the European funding merry-go-round, most fondly in the CONCEDE project, working on dictionaries for Central European languages with amazing teams from the MULTEXT-EAST consortium, and with Georgian and German colleagues in the GREG project.
∗ E-mail: R.P.Evans@brighton.ac.uk; Twitter: @rogerevansbton.
doi:10.1162/COLI a 00234
© 2015 Association for Computational Linguistics
But Adam was not entirely comfortable in academia, or at least not in a version of academia that didn’t share his drive for the practical as well as the theoretical. He didn’t have tenure, nor any clear route to achieve tenure, which meant that he could not apply for and hold grants in his own right (although he freely admitted he was happy not to have the associated administrative responsibilities); he set up a high quality masters program in Computational Lexicography, which ran for a couple of years, but the funding model didn’t really work, and it quickly evolved into the highly successful, but independent, Lexicom workshop series, still running today; and he couldn’t engage university support for developing the WASPbench as a commercial product. So in 2003, he spread his wings, left the university, and set up Lexical Computing Ltd.
For many people, Lexical Computing and the Sketch Engine are what Adam is best known for. He spent eleven years tirelessly developing the company, the software, the methodology, the resources, the discipline. It was an environment in which he seemed completely at ease, sometimes the shameless promoter of his wares, sometimes the astute academic authority, often the source of practical solutions to real problems, and the instigator of new initiatives, and always the generous facilitator, educator, and friend. For me personally, though, this was a time when our friendship was more prominent than our professional relationship. We would meet for the odd drink, usually in the Constant Service pub (Adam’s favorite), and chat about life, family, sometimes work, and occasional schemes for new collaborations, though the company didn’t leave him very much time for that. It was one of those relaxed, undemanding friendships that just picks up whenever and wherever we find the time to meet, but remains strong nevertheless.
Adam’s illness was as unexpected to him as to anyone. Over the summer of 2014, he was making plans for new directions and projects. And then, there was a brief hiatus in communication before we heard the news in early November. And yet, even then, he seemed reconciled—not resigned, but resolved, calm, dignified. I was upset, angry, helpless—useless really, and feeling very selfish in my distress. I saw Adam three times after he became ill and they are all good, strong memories, and that is more to his credit than mine.
The first was in his kitchen, with early spring sunshine, drinking strong coffee he had made very meticulously, watching the winter birds scavenging in the garden, just chatting about nothing in particular, and gossiping about work for a couple of hours. The second was a surprise trip to the pub—the surprise being that Adam was strong enough to get there (and back) on his own, and drink a couple of pints, too. We went to the Constant Service, as always, and it was one of our occasional “NLP group” outings, so a good crowd was there. The third was back in his kitchen, this time for work a few weeks later. Ironically, the university system that struggled to engage with Adam’s practical drive is now fully signed up to demonstrating the “impact” of its research. Adam’s work on Computational Lexicography at Brighton and afterwards through Lexical Computing, featured as an “Impact Case Study” in recent national evaluations, has subsequently been selected for a wider national initiative showcasing UK Computer Science research, currently in development1. Adam was happy to cooperate with this, in part to alleviate boredom, and we arranged a Skype call with a technical author for the initiative from his kitchen. Adam was in excellent form describing his work, his passion, and still full of ideas for gentle academic engagement if his “retirement” would allow it.
1 http://cs-academic-impact.uk.
Shortly after that meeting we heard the news of Adam’s relapse and decision not to continue treatment. Like everyone else, I followed his blog, and also emailed a little privately. I arranged to go and visit again, but Adam wasn’t well enough so we cancelled. Like everyone else, I waited for the inevitable blog post.
Adam’s funeral was in a modest church in the village of Rottingdean just along the coast from Brighton. A beautiful setting and a sunny afternoon. The church was absolutely packed—standing room only—we estimate about 250 people; family, friends, and colleagues from far and wide. A committed atheist, the service focused on fond memories of Adam from those closest to him, with just one hymn, Immortal Invisible, as all his blog readers will understand. A beautiful and fitting farewell to a man who, it seems, was to everyone a friend first, and a colleague, boss, or antagonist, second.
There have been many comments on Adam’s blog, on twitter, in academic forums, which say much more and so much better than I can. Some have said that Adam will be remembered for the Sketch Engine and the amazing data resources that have been built up around it. I would say that his real legacy is much more deeply intellectual than that. Adam would probably smile with satisfaction that the two things can coexist so comfortably—a rare combination of the intellectual and practitioner, a real giant of the field.
Rest now in peace, Adam. Roger Brighton, June 2015",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"A long time ago now (maybe 1988?), Gerald (Gazdar) and I supervised Adam’s DPhil at the University of Sussex. Adam was my age, give or take a year, having come to academia a little late, and was my first doctoral student. Adam’s topic was polysemy, and I’m not really sure that much supervision was actually required, though I recall fun exchanges trying to model the subtleties of word meaning using symbolic knowledge representation techniques—an experience that was clearly enough to convince Adam later that this was a bad idea. In fact, Adam’s thesis title itself was Polysemy. Much as we encourage short thesis titles, pulling off the one-word title is a tall order, requiring a unique combination of focus and coverage, breadth and depth, and, most of all, authority. Adam completely nailed it, at least from the perspective of the pre-empirical Computational Linguistics of the early 1990s. Three years later, after a spell working for dictionary publishers, Adam joined me as a research fellow, now at the University of Brighton. I had a project to explore the automatic enrichment of lexical databases to support the latest trends in language analysis, and, in particular, task-specific lexical resources. I was really pleased and excited to recruit Adam—he had lost none of his intellectual independence, a quality I particularly valued. Within a few weeks he came to me with his own plan for the research—a “detour,” as he put it, from the original workplan. I still have the e-mail, dated 6 April 1995, in which he proposed that, instead of chasing a prescriptive notion of a single lexical resource that needed to be customized to each domain, we should let the domain determine the lexicon, providing lexicographic tools to explore words, and particularly word senses, that were significant for that domain. In that e-mail, Computational Lexicography at Brighton was born. Over the next eight years or so, Computational Lexicography became a key part of our group’s success, increasingly under Adam’s direct leadership. The key project, WASPS, developed the WASPbench—the direct precursor of the Sketch Engine, recruiting David (Tugwell) to the team. In addition, Adam was one of the founding organizers of SENSEVAL, an initiative to bring international teams of researchers together to work in friendly competition on a pre-determined word sense disambiguation task (and which has now transformed into SEMEVAL). Together we secured funding to support the first two rounds of SENSEVAL; each round required the preparation of standardized data sets, guided by Adam’s highly tuned intuitions about lexical data preparation and management. And we engaged somewhat in the European funding merry-go-round, most fondly in the CONCEDE project, working on dictionaries for Central European languages with amazing teams from the MULTEXT-EAST consortium, and with Georgian and German colleagues in the GREG project. [{""affiliations"": [], ""name"": ""Adam Kilgarriff""}, {""affiliations"": [], ""name"": ""Roger Evans""}] SP:abc692872b31b28c2a71c7bbf580d44025f08604 [] adam kilgarriff :Roger Evans∗ University of Brighton
A long time ago now (maybe 1988?), Gerald (Gazdar) and I supervised Adam’s DPhil at the University of Sussex. Adam was my age, give or take a year, having come to academia a little late, and was my first doctoral student. Adam’s topic was polysemy, and I’m not really sure that much supervision was actually required, though I recall fun exchanges trying to model the subtleties of word meaning using symbolic knowledge representation techniques—an experience that was clearly enough to convince Adam later that this was a bad idea. In fact, Adam’s thesis title itself was Polysemy. Much as we encourage short thesis titles, pulling off the one-word title is a tall order, requiring a unique combination of focus and coverage, breadth and depth, and, most of all, authority. Adam completely nailed it, at least from the perspective of the pre-empirical Computational Linguistics of the early 1990s.
Three years later, after a spell working for dictionary publishers, Adam joined me as a research fellow, now at the University of Brighton. I had a project to explore the automatic enrichment of lexical databases to support the latest trends in language analysis, and, in particular, task-specific lexical resources. I was really pleased and excited to recruit Adam—he had lost none of his intellectual independence, a quality I particularly valued. Within a few weeks he came to me with his own plan for the research—a “detour,” as he put it, from the original workplan. I still have the e-mail, dated 6 April 1995, in which he proposed that, instead of chasing a prescriptive notion of a single lexical resource that needed to be customized to each domain, we should let the domain determine the lexicon, providing lexicographic tools to explore words, and particularly word senses, that were significant for that domain. In that e-mail, Computational Lexicography at Brighton was born.
Over the next eight years or so, Computational Lexicography became a key part of our group’s success, increasingly under Adam’s direct leadership. The key project, WASPS, developed the WASPbench—the direct precursor of the Sketch Engine, recruiting David (Tugwell) to the team. In addition, Adam was one of the founding organizers of SENSEVAL, an initiative to bring international teams of researchers together to work in friendly competition on a pre-determined word sense disambiguation task (and which has now transformed into SEMEVAL). Together we secured funding to support the first two rounds of SENSEVAL; each round required the preparation of standardized data sets, guided by Adam’s highly tuned intuitions about lexical data preparation and management. And we engaged somewhat in the European funding merry-go-round, most fondly in the CONCEDE project, working on dictionaries for Central European languages with amazing teams from the MULTEXT-EAST consortium, and with Georgian and German colleagues in the GREG project.
∗ E-mail: R.P.Evans@brighton.ac.uk; Twitter: @rogerevansbton.
doi:10.1162/COLI a 00234
© 2015 Association for Computational Linguistics
But Adam was not entirely comfortable in academia, or at least not in a version of academia that didn’t share his drive for the practical as well as the theoretical. He didn’t have tenure, nor any clear route to achieve tenure, which meant that he could not apply for and hold grants in his own right (although he freely admitted he was happy not to have the associated administrative responsibilities); he set up a high quality masters program in Computational Lexicography, which ran for a couple of years, but the funding model didn’t really work, and it quickly evolved into the highly successful, but independent, Lexicom workshop series, still running today; and he couldn’t engage university support for developing the WASPbench as a commercial product. So in 2003, he spread his wings, left the university, and set up Lexical Computing Ltd.
For many people, Lexical Computing and the Sketch Engine are what Adam is best known for. He spent eleven years tirelessly developing the company, the software, the methodology, the resources, the discipline. It was an environment in which he seemed completely at ease, sometimes the shameless promoter of his wares, sometimes the astute academic authority, often the source of practical solutions to real problems, and the instigator of new initiatives, and always the generous facilitator, educator, and friend. For me personally, though, this was a time when our friendship was more prominent than our professional relationship. We would meet for the odd drink, usually in the Constant Service pub (Adam’s favorite), and chat about life, family, sometimes work, and occasional schemes for new collaborations, though the company didn’t leave him very much time for that. It was one of those relaxed, undemanding friendships that just picks up whenever and wherever we find the time to meet, but remains strong nevertheless.
Adam’s illness was as unexpected to him as to anyone. Over the summer of 2014, he was making plans for new directions and projects. And then, there was a brief hiatus in communication before we heard the news in early November. And yet, even then, he seemed reconciled—not resigned, but resolved, calm, dignified. I was upset, angry, helpless—useless really, and feeling very selfish in my distress. I saw Adam three times after he became ill and they are all good, strong memories, and that is more to his credit than mine.
The first was in his kitchen, with early spring sunshine, drinking strong coffee he had made very meticulously, watching the winter birds scavenging in the garden, just chatting about nothing in particular, and gossiping about work for a couple of hours. The second was a surprise trip to the pub—the surprise being that Adam was strong enough to get there (and back) on his own, and drink a couple of pints, too. We went to the Constant Service, as always, and it was one of our occasional “NLP group” outings, so a good crowd was there. The third was back in his kitchen, this time for work a few weeks later. Ironically, the university system that struggled to engage with Adam’s practical drive is now fully signed up to demonstrating the “impact” of its research. Adam’s work on Computational Lexicography at Brighton and afterwards through Lexical Computing, featured as an “Impact Case Study” in recent national evaluations, has subsequently been selected for a wider national initiative showcasing UK Computer Science research, currently in development1. Adam was happy to cooperate with this, in part to alleviate boredom, and we arranged a Skype call with a technical author for the initiative from his kitchen. Adam was in excellent form describing his work, his passion, and still full of ideas for gentle academic engagement if his “retirement” would allow it.
1 http://cs-academic-impact.uk.
Shortly after that meeting we heard the news of Adam’s relapse and decision not to continue treatment. Like everyone else, I followed his blog, and also emailed a little privately. I arranged to go and visit again, but Adam wasn’t well enough so we cancelled. Like everyone else, I waited for the inevitable blog post.
Adam’s funeral was in a modest church in the village of Rottingdean just along the coast from Brighton. A beautiful setting and a sunny afternoon. The church was absolutely packed—standing room only—we estimate about 250 people; family, friends, and colleagues from far and wide. A committed atheist, the service focused on fond memories of Adam from those closest to him, with just one hymn, Immortal Invisible, as all his blog readers will understand. A beautiful and fitting farewell to a man who, it seems, was to everyone a friend first, and a colleague, boss, or antagonist, second.
There have been many comments on Adam’s blog, on twitter, in academic forums, which say much more and so much better than I can. Some have said that Adam will be remembered for the Sketch Engine and the amazing data resources that have been built up around it. I would say that his real legacy is much more deeply intellectual than that. Adam would probably smile with satisfaction that the two things can coexist so comfortably—a rare combination of the intellectual and practitioner, a real giant of the field.
Rest now in peace, Adam. Roger Brighton, June 2015",
"1 early machine translation in china :The history of machine translation (MT) in China dates back to 1956. At that time the new country had immense construction projects to recover what had been ruined in the war. However, the government precisely recognized the significance of machine translation, and started to explore this area, as the fourth country following the United States, the United Kingdom, and the Soviet Union. In 1959, Russian–Chinese machine translation was demonstrated on a Type-104 general-purpose computer made in China. This first MT system had a dictionary of 2,030 entries, and 29 groups of rules for lexical analysis. Programmed by machine instructions, the system was able to translate nine different types of sentences. It used punched tape as the input, and the output was a special kind of code for Chinese characters, since there was no Chinese character output device at the time. As the pioneer in Chinese MT, the system touched the issues of word sense disambiguation, word reordering, and proposed the idea of predicate-focused sentence analysis and pivot language for multilingual translation. In the same year, machine translation research at the Harbin Institute of Technology (HIT) was started by Prof. Zhen Wang (and later Prof. Kaizhu Wang), focusing on the Russian–Chinese MT group. The pursuit for MT has never halted after these forerunners.
© Association for Computational Linguistics
doi:10.1162/COLI a 00240
© 2015 Association for Computational Linguistics","2 the cemt series :In 1960, I was admitted to HIT. Five years later, I graduated and became a faculty member in the computer department of HIT, which was probably the first computer discipline among Chinese universities. I started my research, however, not from machine translation but from information retrieval (IR). I was fully occupied by how to effectively store books and documents on computers, and then retrieve them quickly and accurately. The start of my research in MT was incidentally caused by IR problems.
At that time, Ming Zhou was my Ph.D. student. He is now the principal researcher of Natural Language Computing at Microsoft Research Asia (MSRA), and many of you may be acquainted with him. In 1985, at the beginning of his graduate study, he was aiming to address the topic of word extraction for Chinese documents to boost IR performance. For an exhaustive survey, Ming went to Beijing from Harbin alone, and buried himself at the National Library for over a month. He came back disappointed, finding that the related work was some language-dependent solutions for English. Actually, many research directions encountered this problem at that time. That’s why Ming and I decided to develop an MT system through which we could first translate Chinese materials into English, so as to take advantage of the solutions proposed for English, and finally translate the results back into Chinese, if necessary.
In those years, the translation from Chinese to other foreign languages was less studied in China. Everything was hard in the beginning. We had to build everything from scratch, such as collecting and inputting each entry of the translation dictionary. Fortunately, we were not alone. I came to know many peer scholars, including Prof. Weitian Wu, Zhiwei Feng, Prof. Zhendong Dong, Prof. Shiwen Yu, and Prof. Changning Huang, as well as Dr. Zhaoxiong Chen. Although we didn’t work together, we could always learn from each other and inspire each other in MT research.
After three years’ effort, we accomplished a rule-based MT system named CEMT-I (Li et al. 1988). It ran on an IBM PC XT1 and was capable of translating eight kinds of Chinese sentence patterns with fewer than one thousand rules. It had a dictionary of 30,000 Chinese-English entries. Simple or even crude as it now seems, it really encouraged every member of our team. After that, we developed CEMT-II (Zhou et al. 1990) and CEMT-III (Zhao, Li, and Zhang 1995) successively. The CEMT series seemed to have a special kind of magic. Almost all the students who participated in these projects devoted themselves to machine translation in their following careers, including Ming Zhou, Min Zhang, and Tiejun Zhao.","3 dear and bt863 :Inspired by the success of the CEMT series, we also developed a computer-aided translation system called “DEAR.” DEAR was put to market via a software chain store. Although it did not sell much, it was our first effort to commercialize the MT technology. I still remember how excited I was when I saw DEAR placed on the shelves for the first time. Today, it still reminds me that research work cannot just stay in the lab.
Also in the 1980s, China’s NLP field was marked by a milestone event: the establishment of the Chinese Information Processing Society of China (CIPS). From then on, NLP researchers throughout the country have been connected and the academic exchange has been facilitated at the national scale. It was far beyond my imagination then that,
1 https://en.wikipedia.org/wiki/IBM_Personal_Computer_XT.
thirty years later, I would have the honor to be the president of this society, leading it to keep on contributing to the development of world-level NLP technology.
I usually regard the series of MT systems that we developed as a large family. In 1994, BT863 joined this family with some new features (Zhao, Li, and Wang 1995; Wang et al. 1997). First, BT863 was distinguished as a bi-direction translation between Chinese and English under a uniform architecture. Second, in addition to the rules, it was augmented with examples and templates learned from a corpus. Finally, this system is remembered for its top performance in the early national MT evaluation organized by the 863 High Tech Program of China.","4 syntactic and semantic parsing :Time passed quickly. The rising of the Internet made communication more convenient, and our research was gradually connected with international peers. We concentrated on the mining and accumulation of bilingual and multilingual corpora. We explored how to integrate rule-based and example-based MT models under a unifying statistical framework. However, as more and more work was conducted, I found it more difficult to go deeper. I began to realize that translation problems cannot rely only on translation methods.
From word segmentation, morphology, word meaning, to named entity, syntax, and semantics, every step in this procedure affects the quality of translation. I remember an interesting story. One day, my student Wanxiang Che input his name into our machine translation system. The system literally translated his name into ‘thousands of cars flying in the sky’. This was rated as the joke of the year in my lab, but the underlying problem is worth pondering.
Traditional Chinese medicine advocates the treatment of both symptoms and root causes. The same principle applies to MT research, in which models for word alignment, decoding, reordering, and so forth, can solve the surface problems of machine translation, whereas understanding the word sense, sentence structure, and semantics is the solution to the fundamental problems. We therefore carried out research on syntactic analysis, including phrase-structure parsing and dependency parsing.
In those days, dependency parsing on Chinese was not widely studied. There was no well-accepted annotation or transformed standard. Therefore, we referred to a large number of linguistic studies, developed a Chinese syntactic dependency annotation standard, and annotated a 50,000-sentence Chinese syntactic dependency treebank on this basis. This is the largest Chinese dependency treebank available. Differently from those transformed from phrase-structure treebanks, our dependency structure uses native dependency structure, which can handle a large number of specific grammatical phenomena in dependency structures. This treebank has been released by the Linguistic Data Consortium (LDC) (Che, Li, and Liu 2012). We hope that more researchers can benefit from it.
Based on syntactic parsing, we hoped to further explore the semantic structure and the relationship of sentences. Therefore, we carried out research on semantic role labeling, and worked on the semantic role labeling methods based on the tree kernel, including the hybrid convolution tree kernel (Che et al. 2008) and the grammar-driven tree kernel (Zhang et al. 2007). In addition, we further broadened our mind and tried to analyze the semantics of Chinese directly. We proposed semantic dependency parsing tasks that directly establish semantic-level dependencies between content words, ignoring auxiliaries and prepositions. Meanwhile, we violated the tree structure constraints, allowing one word to depend on more than one parent node, so as to form semantic
dependency graph structures. At this point, the semantic dependency treebank that we have already labeled has reached more than 30,000 sentences. Much ongoing research is based on these data. Figure 1 shows an example of syntactic dependency parsing, semantic role labeling, and semantic dependency parsing for an input sentence “现在 / 她 /脸色 /难看 /， /好像 /病了 /。 [Now she looks terrible, seems to be sick]” .","5 ltp and ltp-cloud :Every summer, HIT and MSRA would jointly organize a summer school for NLP research students. We invited domestic and foreign experts to give lectures to Chinese students engaged in this field. Because the summer school was free, students from all over the country came together every year, listening to lectures and conducting experiments. When I communicated with these students, I found that many of them came from labs that lacked fundamental NLP tools, such as word segmentors, partof-speech taggers, and syntactic parsers. It would have been very difficult for them to implement their research ideas without these tools. I felt bad when I saw that. They are all students with dreams and innovative ideas. We must create a level playing field for all of them, I thought.
After coming back from the summer school, I met Ting Liu. He is a strong supporter of the idea of sharing. We decided to release an open-source NLP system: Language Technology Platform (LTP). This platform integrates several Chinese NLP basic technologies, including Chinese word segmentation, part-of-speech tagging, named entity recognition, dependency parsing, and semantic role labeling, which has made great contributions to the development of further applications.
In recent years, we realized that cloud computing and the mobile Internet have brought great opportunities and challenges to the NLP field. Therefore, we developed
LTP-cloud2 in 2013, which provides accurate and fast NLP service via the Internet. Currently, the number of LTP-cloud registered users has exceeded 3,000, and most of them are NLP beginners. As I had wished, they no longer need to build an NLP basic processing system from scratch for their research. Every time I see the thank-you notes to the LTP and LTP-cloud in the acknowledgments of their papers, I am proud and grateful.","6 machine translation on the internet :As more and more papers were published in top conferences and journals, our lab made a name in the academic world. Many people in the lab were satisfied, but I felt differently, since publishing papers should not be the major objective for research. New models and techniques should be applied to solve real-world problems and improve people’s daily lives. Particularly since we have moved into the era of the Internet. Many new concepts and ideas have come into being, such as big data and cloud computing. In such a new era, machine translation research should no longer be restricted to the labs, running experiments on a small parallel corpus. Instead, it should embrace the Internet, and embrace big data. We paid great attention to the cooperation with IT and Internet companies. We established a joint lab with MSRA right after it was founded. After that, we have also established joint labs with other companies, like IBM, Baidu, and Tencent.
My student Haifeng Wang is the vice president of Baidu. He is in charge of NLP research and development, as well as Web search. We decided to collaborate in MT shortly after he joined Baidu, since Baidu can provide a huge platform for us to verify our ideas. Together with Tsinghua University, Zhejiang University, the Institute of Computing Technology, and the Institute of Automation of the Chinese Academy of Science, we successfully applied for an 863 project titled “Machine Translation on the Internet.” All the members participating in this project have great passion for MT technologies and products.
Chinese people accept the principle that “取之于民，用之于民” [what is taken from the people should be used for the interests of the people]. Internet-based machine translation also follows this principle, which mines a large volume of translation data from the Internet, trains the translation model, and then provides high-quality services for Internet users. In our online translation system, taking Chinese–English translation, for example, there are hundreds of millions of parallel data for this language pair, which were filtered from billions of raw parallel data. We collected a large amount of data from hundreds of billions of Web pages. So I should say our MT service is actually built upon the whole Internet.
We have designed various mining models for these heterogeneous Internet data sources, including bilingual parallel pages, bilingual comparable pages, Web pages containing aligned sentence pairs, as well as plain texts containing entity and terminology translations. The mined translation data are filtered and refined. We set different updating frequencies for different Web sites, so as to guarantee that the latest data can be included. I often observe the mined translation data by myself, and I can find plenty of wonderful translations generated by ordinary Internet users. Their wisdom is perfectly integrated into the translation system. However, how to make use of such a big corpus?
2 http://www.ltp-cloud.com/demo/.
This is a sweet annoyance. To handle big data, we have developed fast training and parallel decoding techniques in our project.
With such big data and frequent updates, even Internet buzzwords can be correctly translated. My students often post the so-called “magic translation” on microblogs. After the machine translation service came online, I began to realize that it would not only influence those Ph.D. students who are reading and writing research papers, or those businessmen who are studying materials from foreign countries. It also makes a huge difference to ordinary people’s lives. Figure 2 shows some examples of Chinese– English machine translation from the Baidu online translation service,3 which integrates the research work of the 863 project “Machine Translation on the Internet.”
I once met a 50-year-old Chinese lady on a flight to Japan. She could not speak Japanese, but she had finally decided to marry her Japanese husband, with whom she had chatted online using machine translation. Another story comes from my neighbors, who are a couple my age. Their children have lived in Germany for years. The first
3 http://fanyi.baidu.com/.
time the old couple met their grandson when their family came back to China, they were thrilled. However, on meeting their grandson, who can only speak German, they had no way to express their love, which made them sad. The grandma blamed herself and even wept when she was alone. With my recommendation, they started to use the online speech translation app in their smart phone. Now, they can finally talk to their grandson.","7 integration of mt models :I have been working in machine translation for several decades, going through almost all the streams of technologies, from rule-based MT (RBMT) models at the very beginning, example-based MT (EBMT) methods, to statistical MT (SMT) methods, as well as the research hotspot today—neural network machine translation (NMT). Actually, we tried the neural network–based models in NLP tasks, such as dialogue act analysis and word sense disambiguation, more than 15 years ago (Wang, Gao, and Li 1999a, 1999b). It is big data and computing power today that help neural network–based models significantly outperform traditional ones. I know that every method has its advantages and disadvantages. Although the new model and its methodology surpasses the old ones overall, it does not mean that the old methods are useless. There’s an old saying in Chinese, “the silly bear keeps picking corn,” which describes that when a bear is stealing corn from a peasant’s field, it would always throw away the old one in its hand when it picked a new one; the silly bear would always end up with only one corn in his hand. I hoped that my team and I wouldn’t become the “silly bears.” Therefore, when we decided to develop an Internet MT system, we all agreed on the idea that we needed a hybrid approach, with which we could integrate all translation models and subsystems, on each of which we have all spent great effort. It is just like an orchestra, in which all instruments, such as piano, violin, cello, trumpet, and so on, are arranged perfectly together. Only in this way can the orchestra present a wonderful performance. As shown in Figure 3, in our MT system today, different models work together perfectly.
The rule-based method is used to translate expressions like date, time, and number. The example-based method is applied to translate buzzwords, especially the new emerging Internet expressions. On the other hand, those complicated long sentences are translated using the syntax-based statistical model, while those sentences that can be covered by a predefined vocabulary are translated with an NMT model. Finally, the
sentences left are all translated with a classical SMT model. The conductor of such an orchestra is a discriminative distributing module, which decides what subsystem an input sentence should be distributed to, based on a variety of statistical and linguistic features.","8 translation for resource-poor languages :Shortly after the release of Chinese–English and English–Chinese translation services, we also released translation services between Chinese and Japanese, Korean, and other daily-used foreign languages. However, with translation directions expanded, users’ expectations for the translation between the resource-poor languages became higher and higher. Especially in recent years, China has been doing business more frequently with many countries, such as Thailand and Portugal, among others, and the destinations for Chinese tourists have become more diverse. One of my friends told me a story after he came back from a tour in Southeast Asia. He ordered three kinds of salads in a restaurant, since he did not understand or speak the local language. He could not communicate with the waiters or even read the menu. These incidents told us that solving translation problems for these resource-poor languages is urgent. Therefore, we have successively released translation services between Chinese and over 20 foreign languages. Now, we have covered languages in eight of the top ten destinations for Chinese tourists, and all the top ten foreign cities where Chinese tourists spend the most money.
On this basis, we took a further step. We built translation systems between any two languages using the pivot approach (Wang, Wu, and Liu 2006; Wu and Wang 2007). For resource-poor language pairs, we use English or Chinese as the pivot language. Translation models are trained for source-pivot and pivot-target, respectively, which are then combined to form the translation model from the source to the target language. Using this model, Baidu online translation services successfully realized pairwise translation between any two of 27 languages; in total, 702 translation directions.","Good afternoon, ladies and gentlemen. I am standing here, grateful, excited, and proud. I see so many friends, my colleagues, students, and many more researchers in this room. I see that the work we started 50 years ago is now flourishing and is embedded in people’s everyday lives. I see for the first time that the ACL conference is held here in Beijing, China. And I am deeply honored to be awarded the Lifetime Achievement Award of 2015. I want to thank the ACL for giving me the Lifetime Achievement Award of 2015. It is the appreciation of not only my work, but also of the work that my fellow researchers, my colleagues, and my students have done through all these years. It is an honor for all of us. As a veteran of NLP research, I am fortunate to witness and be a part of its long yet inspiring journey in China. So today, to everyone here, my friends, colleagues, and students, either well-known scientists or young researchers: I’d like to share my experience and thoughts with you.","[{""affiliations"": [], ""name"": ""Sheng Li""}]",SP:33d2ba530a9f1861cb08be9db34d54b1396d0270,"[{""authors"": [""Ting Liu""], ""title"": ""Leveraging multiple MT"", ""year"": 2010}, {""authors"": [""Seoul. Zhao"", ""Tiejun"", ""Sheng Li"", ""Min Zhang""], ""title"": ""CEMT-III: A fully automatic"", ""venue"": ""Proceedings of the NLPPRS,"", ""year"": 1995}, {""authors"": [""Zhou"", ""Ming"", ""Sheng Li"", ""Mingzeng Hu"", ""Shi Miao.""], ""title"": ""An interactive Chinese\u2013English machine translation system: CEMT-II"", ""venue"": ""Journal of the China Society for Scientific and"", ""year"": 1990}]",,,,,,,,,,"9 mt methodology for other areas :“他山之石，可以攻玉” [Stones from other hills may serve to polish the jade at hand]. This is a Chinese old saying from “《诗经》” [The Book of Songs], which was written 2,500 years ago. It suggests that one may benefit from other people’s opinions and methods for their task. Machine translation technology is now a “stone from another hill,” which has been used in many other areas. For instance, some researchers recast paraphrasing as a monolingual translation problem and use MT models to generate paraphrases of the input sentences (Zhao et al. 2009, 2010). There are also researchers who regard query reformulation as the translation from the original query to the rewritten one (Riezler and Liu 2010). However, what interests me the most is the encounter between translation technology and Chinese traditional culture. For example, MSRA uses the translation model to automatically generate couplets (Jiang and Zhou 2008), which are posted on the doors of every house during Chinese New Year. Baidu applies translation methods to compose poems. Given a picture and the first line of a poem, the system can generate another three lines of the poem that describe the content of the picture. In addition, I have heard recently that both Microsoft and Baidu have released their chatting robots, which are named Microsoft XiaoIce and Baidu Xiaodu, respectively. They both use translation techniques in the searching and generation of chatting responses. It is fair to say that machine translation has become more than a specific
method. Instead, it has evolved into a methodology and could make a contribution to other similar or related areas.",,,,,,,,,,,"10 conclusion :There is an ancient story in China called “愚公移山” [Yugong moves the mountain]. In the story, an old man called Yugong—meaning an unwise man—lived in a mountain area. He decided to build a road to the outside world by moving two huge mountains away. Other people all thought it was impossible and laughed at him. However, Yugong said to the people calmly: “Even if I die, I have children; and my children would have children in the future. As the mountain wouldn’t grow, we would move the mountain away eventually.” Today, when facing the ambitious goal of automatic high-quality machine translation, and even the whole NLP field, I cannot help thinking of Yugong’s spirit. I have been, and I still am, trying to solve the questions and obstacles along the way. Even if one day I will no longer be able to keep exploring MT, I believe that the younger generations will keep on going until the dream of making a computer truly understand languages eventually comes true.
My friends, especially the young ones, to share what I have learned from my career, I’d like to say: Make yourself a good translation system: Input diligence today, and it will definitely translate into an amazing tomorrow!",,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 early machine translation in china :The history of machine translation (MT) in China dates back to 1956. At that time the new country had immense construction projects to recover what had been ruined in the war. However, the government precisely recognized the significance of machine translation, and started to explore this area, as the fourth country following the United States, the United Kingdom, and the Soviet Union. In 1959, Russian–Chinese machine translation was demonstrated on a Type-104 general-purpose computer made in China. This first MT system had a dictionary of 2,030 entries, and 29 groups of rules for lexical analysis. Programmed by machine instructions, the system was able to translate nine different types of sentences. It used punched tape as the input, and the output was a special kind of code for Chinese characters, since there was no Chinese character output device at the time. As the pioneer in Chinese MT, the system touched the issues of word sense disambiguation, word reordering, and proposed the idea of predicate-focused sentence analysis and pivot language for multilingual translation. In the same year, machine translation research at the Harbin Institute of Technology (HIT) was started by Prof. Zhen Wang (and later Prof. Kaizhu Wang), focusing on the Russian–Chinese MT group. The pursuit for MT has never halted after these forerunners.
© Association for Computational Linguistics
doi:10.1162/COLI a 00240
© 2015 Association for Computational Linguistics 2 the cemt series :In 1960, I was admitted to HIT. Five years later, I graduated and became a faculty member in the computer department of HIT, which was probably the first computer discipline among Chinese universities. I started my research, however, not from machine translation but from information retrieval (IR). I was fully occupied by how to effectively store books and documents on computers, and then retrieve them quickly and accurately. The start of my research in MT was incidentally caused by IR problems.
At that time, Ming Zhou was my Ph.D. student. He is now the principal researcher of Natural Language Computing at Microsoft Research Asia (MSRA), and many of you may be acquainted with him. In 1985, at the beginning of his graduate study, he was aiming to address the topic of word extraction for Chinese documents to boost IR performance. For an exhaustive survey, Ming went to Beijing from Harbin alone, and buried himself at the National Library for over a month. He came back disappointed, finding that the related work was some language-dependent solutions for English. Actually, many research directions encountered this problem at that time. That’s why Ming and I decided to develop an MT system through which we could first translate Chinese materials into English, so as to take advantage of the solutions proposed for English, and finally translate the results back into Chinese, if necessary.
In those years, the translation from Chinese to other foreign languages was less studied in China. Everything was hard in the beginning. We had to build everything from scratch, such as collecting and inputting each entry of the translation dictionary. Fortunately, we were not alone. I came to know many peer scholars, including Prof. Weitian Wu, Zhiwei Feng, Prof. Zhendong Dong, Prof. Shiwen Yu, and Prof. Changning Huang, as well as Dr. Zhaoxiong Chen. Although we didn’t work together, we could always learn from each other and inspire each other in MT research.
After three years’ effort, we accomplished a rule-based MT system named CEMT-I (Li et al. 1988). It ran on an IBM PC XT1 and was capable of translating eight kinds of Chinese sentence patterns with fewer than one thousand rules. It had a dictionary of 30,000 Chinese-English entries. Simple or even crude as it now seems, it really encouraged every member of our team. After that, we developed CEMT-II (Zhou et al. 1990) and CEMT-III (Zhao, Li, and Zhang 1995) successively. The CEMT series seemed to have a special kind of magic. Almost all the students who participated in these projects devoted themselves to machine translation in their following careers, including Ming Zhou, Min Zhang, and Tiejun Zhao. 3 dear and bt863 :Inspired by the success of the CEMT series, we also developed a computer-aided translation system called “DEAR.” DEAR was put to market via a software chain store. Although it did not sell much, it was our first effort to commercialize the MT technology. I still remember how excited I was when I saw DEAR placed on the shelves for the first time. Today, it still reminds me that research work cannot just stay in the lab.
Also in the 1980s, China’s NLP field was marked by a milestone event: the establishment of the Chinese Information Processing Society of China (CIPS). From then on, NLP researchers throughout the country have been connected and the academic exchange has been facilitated at the national scale. It was far beyond my imagination then that,
1 https://en.wikipedia.org/wiki/IBM_Personal_Computer_XT.
thirty years later, I would have the honor to be the president of this society, leading it to keep on contributing to the development of world-level NLP technology.
I usually regard the series of MT systems that we developed as a large family. In 1994, BT863 joined this family with some new features (Zhao, Li, and Wang 1995; Wang et al. 1997). First, BT863 was distinguished as a bi-direction translation between Chinese and English under a uniform architecture. Second, in addition to the rules, it was augmented with examples and templates learned from a corpus. Finally, this system is remembered for its top performance in the early national MT evaluation organized by the 863 High Tech Program of China. 4 syntactic and semantic parsing :Time passed quickly. The rising of the Internet made communication more convenient, and our research was gradually connected with international peers. We concentrated on the mining and accumulation of bilingual and multilingual corpora. We explored how to integrate rule-based and example-based MT models under a unifying statistical framework. However, as more and more work was conducted, I found it more difficult to go deeper. I began to realize that translation problems cannot rely only on translation methods.
From word segmentation, morphology, word meaning, to named entity, syntax, and semantics, every step in this procedure affects the quality of translation. I remember an interesting story. One day, my student Wanxiang Che input his name into our machine translation system. The system literally translated his name into ‘thousands of cars flying in the sky’. This was rated as the joke of the year in my lab, but the underlying problem is worth pondering.
Traditional Chinese medicine advocates the treatment of both symptoms and root causes. The same principle applies to MT research, in which models for word alignment, decoding, reordering, and so forth, can solve the surface problems of machine translation, whereas understanding the word sense, sentence structure, and semantics is the solution to the fundamental problems. We therefore carried out research on syntactic analysis, including phrase-structure parsing and dependency parsing.
In those days, dependency parsing on Chinese was not widely studied. There was no well-accepted annotation or transformed standard. Therefore, we referred to a large number of linguistic studies, developed a Chinese syntactic dependency annotation standard, and annotated a 50,000-sentence Chinese syntactic dependency treebank on this basis. This is the largest Chinese dependency treebank available. Differently from those transformed from phrase-structure treebanks, our dependency structure uses native dependency structure, which can handle a large number of specific grammatical phenomena in dependency structures. This treebank has been released by the Linguistic Data Consortium (LDC) (Che, Li, and Liu 2012). We hope that more researchers can benefit from it.
Based on syntactic parsing, we hoped to further explore the semantic structure and the relationship of sentences. Therefore, we carried out research on semantic role labeling, and worked on the semantic role labeling methods based on the tree kernel, including the hybrid convolution tree kernel (Che et al. 2008) and the grammar-driven tree kernel (Zhang et al. 2007). In addition, we further broadened our mind and tried to analyze the semantics of Chinese directly. We proposed semantic dependency parsing tasks that directly establish semantic-level dependencies between content words, ignoring auxiliaries and prepositions. Meanwhile, we violated the tree structure constraints, allowing one word to depend on more than one parent node, so as to form semantic
dependency graph structures. At this point, the semantic dependency treebank that we have already labeled has reached more than 30,000 sentences. Much ongoing research is based on these data. Figure 1 shows an example of syntactic dependency parsing, semantic role labeling, and semantic dependency parsing for an input sentence “现在 / 她 /脸色 /难看 /， /好像 /病了 /。 [Now she looks terrible, seems to be sick]” . 5 ltp and ltp-cloud :Every summer, HIT and MSRA would jointly organize a summer school for NLP research students. We invited domestic and foreign experts to give lectures to Chinese students engaged in this field. Because the summer school was free, students from all over the country came together every year, listening to lectures and conducting experiments. When I communicated with these students, I found that many of them came from labs that lacked fundamental NLP tools, such as word segmentors, partof-speech taggers, and syntactic parsers. It would have been very difficult for them to implement their research ideas without these tools. I felt bad when I saw that. They are all students with dreams and innovative ideas. We must create a level playing field for all of them, I thought.
After coming back from the summer school, I met Ting Liu. He is a strong supporter of the idea of sharing. We decided to release an open-source NLP system: Language Technology Platform (LTP). This platform integrates several Chinese NLP basic technologies, including Chinese word segmentation, part-of-speech tagging, named entity recognition, dependency parsing, and semantic role labeling, which has made great contributions to the development of further applications.
In recent years, we realized that cloud computing and the mobile Internet have brought great opportunities and challenges to the NLP field. Therefore, we developed
LTP-cloud2 in 2013, which provides accurate and fast NLP service via the Internet. Currently, the number of LTP-cloud registered users has exceeded 3,000, and most of them are NLP beginners. As I had wished, they no longer need to build an NLP basic processing system from scratch for their research. Every time I see the thank-you notes to the LTP and LTP-cloud in the acknowledgments of their papers, I am proud and grateful. 6 machine translation on the internet :As more and more papers were published in top conferences and journals, our lab made a name in the academic world. Many people in the lab were satisfied, but I felt differently, since publishing papers should not be the major objective for research. New models and techniques should be applied to solve real-world problems and improve people’s daily lives. Particularly since we have moved into the era of the Internet. Many new concepts and ideas have come into being, such as big data and cloud computing. In such a new era, machine translation research should no longer be restricted to the labs, running experiments on a small parallel corpus. Instead, it should embrace the Internet, and embrace big data. We paid great attention to the cooperation with IT and Internet companies. We established a joint lab with MSRA right after it was founded. After that, we have also established joint labs with other companies, like IBM, Baidu, and Tencent.
My student Haifeng Wang is the vice president of Baidu. He is in charge of NLP research and development, as well as Web search. We decided to collaborate in MT shortly after he joined Baidu, since Baidu can provide a huge platform for us to verify our ideas. Together with Tsinghua University, Zhejiang University, the Institute of Computing Technology, and the Institute of Automation of the Chinese Academy of Science, we successfully applied for an 863 project titled “Machine Translation on the Internet.” All the members participating in this project have great passion for MT technologies and products.
Chinese people accept the principle that “取之于民，用之于民” [what is taken from the people should be used for the interests of the people]. Internet-based machine translation also follows this principle, which mines a large volume of translation data from the Internet, trains the translation model, and then provides high-quality services for Internet users. In our online translation system, taking Chinese–English translation, for example, there are hundreds of millions of parallel data for this language pair, which were filtered from billions of raw parallel data. We collected a large amount of data from hundreds of billions of Web pages. So I should say our MT service is actually built upon the whole Internet.
We have designed various mining models for these heterogeneous Internet data sources, including bilingual parallel pages, bilingual comparable pages, Web pages containing aligned sentence pairs, as well as plain texts containing entity and terminology translations. The mined translation data are filtered and refined. We set different updating frequencies for different Web sites, so as to guarantee that the latest data can be included. I often observe the mined translation data by myself, and I can find plenty of wonderful translations generated by ordinary Internet users. Their wisdom is perfectly integrated into the translation system. However, how to make use of such a big corpus?
2 http://www.ltp-cloud.com/demo/.
This is a sweet annoyance. To handle big data, we have developed fast training and parallel decoding techniques in our project.
With such big data and frequent updates, even Internet buzzwords can be correctly translated. My students often post the so-called “magic translation” on microblogs. After the machine translation service came online, I began to realize that it would not only influence those Ph.D. students who are reading and writing research papers, or those businessmen who are studying materials from foreign countries. It also makes a huge difference to ordinary people’s lives. Figure 2 shows some examples of Chinese– English machine translation from the Baidu online translation service,3 which integrates the research work of the 863 project “Machine Translation on the Internet.”
I once met a 50-year-old Chinese lady on a flight to Japan. She could not speak Japanese, but she had finally decided to marry her Japanese husband, with whom she had chatted online using machine translation. Another story comes from my neighbors, who are a couple my age. Their children have lived in Germany for years. The first
3 http://fanyi.baidu.com/.
time the old couple met their grandson when their family came back to China, they were thrilled. However, on meeting their grandson, who can only speak German, they had no way to express their love, which made them sad. The grandma blamed herself and even wept when she was alone. With my recommendation, they started to use the online speech translation app in their smart phone. Now, they can finally talk to their grandson. 7 integration of mt models :I have been working in machine translation for several decades, going through almost all the streams of technologies, from rule-based MT (RBMT) models at the very beginning, example-based MT (EBMT) methods, to statistical MT (SMT) methods, as well as the research hotspot today—neural network machine translation (NMT). Actually, we tried the neural network–based models in NLP tasks, such as dialogue act analysis and word sense disambiguation, more than 15 years ago (Wang, Gao, and Li 1999a, 1999b). It is big data and computing power today that help neural network–based models significantly outperform traditional ones. I know that every method has its advantages and disadvantages. Although the new model and its methodology surpasses the old ones overall, it does not mean that the old methods are useless. There’s an old saying in Chinese, “the silly bear keeps picking corn,” which describes that when a bear is stealing corn from a peasant’s field, it would always throw away the old one in its hand when it picked a new one; the silly bear would always end up with only one corn in his hand. I hoped that my team and I wouldn’t become the “silly bears.” Therefore, when we decided to develop an Internet MT system, we all agreed on the idea that we needed a hybrid approach, with which we could integrate all translation models and subsystems, on each of which we have all spent great effort. It is just like an orchestra, in which all instruments, such as piano, violin, cello, trumpet, and so on, are arranged perfectly together. Only in this way can the orchestra present a wonderful performance. As shown in Figure 3, in our MT system today, different models work together perfectly.
The rule-based method is used to translate expressions like date, time, and number. The example-based method is applied to translate buzzwords, especially the new emerging Internet expressions. On the other hand, those complicated long sentences are translated using the syntax-based statistical model, while those sentences that can be covered by a predefined vocabulary are translated with an NMT model. Finally, the
sentences left are all translated with a classical SMT model. The conductor of such an orchestra is a discriminative distributing module, which decides what subsystem an input sentence should be distributed to, based on a variety of statistical and linguistic features. 8 translation for resource-poor languages :Shortly after the release of Chinese–English and English–Chinese translation services, we also released translation services between Chinese and Japanese, Korean, and other daily-used foreign languages. However, with translation directions expanded, users’ expectations for the translation between the resource-poor languages became higher and higher. Especially in recent years, China has been doing business more frequently with many countries, such as Thailand and Portugal, among others, and the destinations for Chinese tourists have become more diverse. One of my friends told me a story after he came back from a tour in Southeast Asia. He ordered three kinds of salads in a restaurant, since he did not understand or speak the local language. He could not communicate with the waiters or even read the menu. These incidents told us that solving translation problems for these resource-poor languages is urgent. Therefore, we have successively released translation services between Chinese and over 20 foreign languages. Now, we have covered languages in eight of the top ten destinations for Chinese tourists, and all the top ten foreign cities where Chinese tourists spend the most money.
On this basis, we took a further step. We built translation systems between any two languages using the pivot approach (Wang, Wu, and Liu 2006; Wu and Wang 2007). For resource-poor language pairs, we use English or Chinese as the pivot language. Translation models are trained for source-pivot and pivot-target, respectively, which are then combined to form the translation model from the source to the target language. Using this model, Baidu online translation services successfully realized pairwise translation between any two of 27 languages; in total, 702 translation directions. Good afternoon, ladies and gentlemen. I am standing here, grateful, excited, and proud. I see so many friends, my colleagues, students, and many more researchers in this room. I see that the work we started 50 years ago is now flourishing and is embedded in people’s everyday lives. I see for the first time that the ACL conference is held here in Beijing, China. And I am deeply honored to be awarded the Lifetime Achievement Award of 2015. I want to thank the ACL for giving me the Lifetime Achievement Award of 2015. It is the appreciation of not only my work, but also of the work that my fellow researchers, my colleagues, and my students have done through all these years. It is an honor for all of us. As a veteran of NLP research, I am fortunate to witness and be a part of its long yet inspiring journey in China. So today, to everyone here, my friends, colleagues, and students, either well-known scientists or young researchers: I’d like to share my experience and thoughts with you. [{""affiliations"": [], ""name"": ""Sheng Li""}] SP:33d2ba530a9f1861cb08be9db34d54b1396d0270 [{""authors"": [""Ting Liu""], ""title"": ""Leveraging multiple MT"", ""year"": 2010}, {""authors"": [""Seoul. Zhao"", ""Tiejun"", ""Sheng Li"", ""Min Zhang""], ""title"": ""CEMT-III: A fully automatic"", ""venue"": ""Proceedings of the NLPPRS,"", ""year"": 1995}, {""authors"": [""Zhou"", ""Ming"", ""Sheng Li"", ""Mingzeng Hu"", ""Shi Miao.""], ""title"": ""An interactive Chinese\u2013English machine translation system: CEMT-II"", ""venue"": ""Journal of the China Society for Scientific and"", ""year"": 1990}] 9 mt methodology for other areas :“他山之石，可以攻玉” [Stones from other hills may serve to polish the jade at hand]. This is a Chinese old saying from “《诗经》” [The Book of Songs], which was written 2,500 years ago. It suggests that one may benefit from other people’s opinions and methods for their task. Machine translation technology is now a “stone from another hill,” which has been used in many other areas. For instance, some researchers recast paraphrasing as a monolingual translation problem and use MT models to generate paraphrases of the input sentences (Zhao et al. 2009, 2010). There are also researchers who regard query reformulation as the translation from the original query to the rewritten one (Riezler and Liu 2010). However, what interests me the most is the encounter between translation technology and Chinese traditional culture. For example, MSRA uses the translation model to automatically generate couplets (Jiang and Zhou 2008), which are posted on the doors of every house during Chinese New Year. Baidu applies translation methods to compose poems. Given a picture and the first line of a poem, the system can generate another three lines of the poem that describe the content of the picture. In addition, I have heard recently that both Microsoft and Baidu have released their chatting robots, which are named Microsoft XiaoIce and Baidu Xiaodu, respectively. They both use translation techniques in the searching and generation of chatting responses. It is fair to say that machine translation has become more than a specific
method. Instead, it has evolved into a methodology and could make a contribution to other similar or related areas. 10 conclusion :There is an ancient story in China called “愚公移山” [Yugong moves the mountain]. In the story, an old man called Yugong—meaning an unwise man—lived in a mountain area. He decided to build a road to the outside world by moving two huge mountains away. Other people all thought it was impossible and laughed at him. However, Yugong said to the people calmly: “Even if I die, I have children; and my children would have children in the future. As the mountain wouldn’t grow, we would move the mountain away eventually.” Today, when facing the ambitious goal of automatic high-quality machine translation, and even the whole NLP field, I cannot help thinking of Yugong’s spirit. I have been, and I still am, trying to solve the questions and obstacles along the way. Even if one day I will no longer be able to keep exploring MT, I believe that the younger generations will keep on going until the dream of making a computer truly understand languages eventually comes true.
My friends, especially the young ones, to share what I have learned from my career, I’d like to say: Make yourself a good translation system: Input diligence today, and it will definitely translate into an amazing tomorrow!","10 conclusion :There is an ancient story in China called “愚公移山” [Yugong moves the mountain]. In the story, an old man called Yugong—meaning an unwise man—lived in a mountain area. He decided to build a road to the outside world by moving two huge mountains away. Other people all thought it was impossible and laughed at him. However, Yugong said to the people calmly: “Even if I die, I have children; and my children would have children in the future. As the mountain wouldn’t grow, we would move the mountain away eventually.” Today, when facing the ambitious goal of automatic high-quality machine translation, and even the whole NLP field, I cannot help thinking of Yugong’s spirit. I have been, and I still am, trying to solve the questions and obstacles along the way. Even if one day I will no longer be able to keep exploring MT, I believe that the younger generations will keep on going until the dream of making a computer truly understand languages eventually comes true.
My friends, especially the young ones, to share what I have learned from my career, I’d like to say: Make yourself a good translation system: Input diligence today, and it will definitely translate into an amazing tomorrow!"
"1 introduction :Much of statistical NLP research relies on some sorts of manually annotated corpora to train models, but annotated resources are extremely expensive to build, especially on a large scale. The creation of treebanks is a prime example (Marcus, Santorini, and Marcinkiewicz 1993). However, the linguistic theories motivating these annotation efforts are often heavily debated, and as a result there often exist multiple corpora for the same task with vastly different and incompatible annotation philosophies. For example, there are several treebanks for English, including the Chomskian-style Penn Treebank (Marcus, Santorini, and Marcinkiewicz 1993), the HPSG LinGo Redwoods Treebank (Oepen et al. 2002), and a smaller dependency treebank (Buchholz and Marsi 2006). From the perspective of resource accumulation, it seems a waste in human efforts.1
A second, related problem is that the raw texts are also drawn from different domains, which for the above example range from financial news (Penn Treebank/Wall Street Journal) to transcribed dialog (LinGo). It would be nice if a system could be automatically ported from one set of guidelines and/or domain to another, in order to exploit a much larger data set. The second problem, domain adaptation, is very well studied (e.g., Blitzer, McDonald, & Pereira 2006; Daumé III 2007). This work focuses on the widely existing and equally important problem, annotation adaptation, in order to adapt the divergence between different annotation guidelines and integrate linguistic knowledge in corpora with incongruent annotation formats.
In this article, we describe the problem of annotation adaptation and the intrinsic principles of the solutions, and present a series of successively improved concrete models, the goal being to transfer the annotations of a corpus (source corpus) to the annotation format of another corpus (target corpus). The transfer classifier is the fundamental component for annotation adaptation algorithms. It learns correspondence regularities between annotation guidelines from a parallel annotated corpus, which has two kinds of annotations for the same data. In the simplest model (Model 1), the source classifier trained on the source corpus gives its predications to the transfer classifier trained on the parallel annotated corpus, so as to integrate the knowledge in the two corpora. In a variant of the simplest model (Model 2), the transfer classifier is used to transform the annotations in the source corpus into the annotation format of the target corpus; then the transformed source corpus and the target corpus are merged in order to train a more accurate classifier. Based on the second model, we finally develop an optimized model (Model 3), where two optimization strategies, iterative training and predict-self re-estimation, are integrated to further improve the efficiency of annotation adaptation.
We experiment on Chinese word segmentation and dependency parsing to test the efficacy of our methods. For word segmentation, the problem of incompatible annotation guidelines is one of the most glaring: No segmentation guideline has been widely accepted due to the lack of a clear definition of Chinese word morphology. For dependency parsing there also exist multiple disparate annotation guidelines. For
1 Different annotated corpora for the same task facilitate the comparison of linguistic theories. From this perspective, having multiple standards is not necessarily a waste but rather a blessing, because it is a necessary phase in coming to a consensus if there is one.
example, the dependency relations extracted from a constituency treebank follow syntactic principles, whereas the semantic dependency treebank is annotated in a semantic perspective.
The two corpora for word segmentation are the much larger People’s Daily corpus (PD) (5.86M words) (Yu et al. 2001) and the smaller but more popular Penn Chinese Treebank (CTB) (0.47M words) (Xue et al. 2005). They utilize very different segmentation guidelines; for example, as shown in Figure 1, PD breaks VicePresident into two words and combines the phrase visited-China as a compound, compared with the segmentation following the CTB annotation guideline. It is preferable to transfer knowledge from PD to CTB because the latter also annotates tree structures, which are useful for downstream applications like parsing, summarization, and machine translation, yet it is much smaller in size. For dependency parsing, we use the dependency treebank (DCTB) extracted from CTB according to the rules of Yamada and Matsumoto (2003), and the Semantic Dependency Treebank (SDT) built on a small part of the CTB text (Che et al. 2012). Compared with the automatically extracted dependencies in DCTB, semantic dependencies in SDT reveal semantic relationships between words, rather than the syntactic relationships in syntactic dependencies. Figure 2 shows an example. Experiments on both word segmentation and dependency parsing show that annotation adaptation results in significant improvement over the baselines, and achieves the state-of-the-art with only local features.
The rest of the article is organized as follows. Section 2 gives a description of the problem of annotation adaptation. Section 3 briefly introduces the tasks of word segmentation and dependency parsing as well as their state-of-the-art models. In Section 4 we first describe the transfer classifier that indicates the
intrinsic principles of annotation adaptation, and then depict the three successively enhanced models for automatic adaptation of annotations. After the description of experimental results in Section 5 and the discussion of application scenarios in Section 6, we give a brief review of related work in Section 7, drawing conclusions in Section 8.","2 automatic annotation adaptation :We define annotation adaptation as a task aimed at automatically adapting the divergence between different annotation guidelines. Statistical models can be designed to learn the relevance of two annotation guidelines in order to transform a corpus from one annotation guideline to another. From this point of view, annotation adaptation can be seen as a special case of transfer learning. Through annotation adaptation, the linguistic knowledge in different corpora is integrated, resulting in enhanced NLP systems without complicated models and features.
Much research has considered the problem of domain adaptation (Blitzer, McDonald, and Pereira 2006; Daumé III 2007), which also can be seen as a special case of transfer learning. It aims to adapt models trained in one domain (e.g., chemistry) to work well in other domains (e.g., medicine). Despite superficial similarities between domain adaptation and annotation adaptation, we argue that the underlying problems are quite different. Domain adaptation assumes that the labeling guidelines are preserved between the two domains—for example, an adjective is always labeled as JJ regardless of whether it is from the Wall Street Journal (WSJ) or a biomedical text, and only the distributions are different—for example, the word control is most likely a verb in WSJ but often a noun in biomedical texts (as in control experiment). Annotation adaptation, however, tackles the problem where the guideline itself is changed, for example, one treebank might distinguish between transitive and intransitive verbs, while merging the different noun types (NN, NNS, etc.), or one treebank (PTB) might be much flatter than the other (LinGo), not to mention the fundamental disparities between their underlying linguistic representations (CFG vs. HPSG).
A more formal description will allow us to make these claims more precise. Let X be the data and Y be the annotation. Annotation adaptation can be understood as a change of P(Y) due to the change in annotation guidelines while P(X) remains constant. Through annotation adaptation, we want to change the annotations of the data from one guideline to another, leaving the data itself unchanged. However, in domain adaptation, P(X) changes, but P(Y) is assumed to be constant. The word assumed means that the distributions P(Y, X) and P(Y|X) are actually changed because P(X) is changed. Domain adaptation aims to make the model better adapt to a different domain with the same annotation guidelines.
According to this analysis, annotation adaptation seems more motivated from a linguistic (rather than statistical) point of view, and tackles a serious problem fundamentally different from domain adaptation, which is also a serious problem (often leading to >10% loss in accuracy). More interestingly, annotation adaptation, without any assumptions about distributions, can be simultaneously applied to both domain and annotation adaptation problems, which is very appealing in practice because the latter problem often implies the former.","3 case studies: word segmentation and dependency parsing : In many Asian languages there are no explicit word boundaries, thus word segmentation is a fundamental task for the processing and understanding of these languages. Given a sentence as a sequence of n characters:
x = x1 x2 .. xn
where xi is a character, word segmentation aims to split the sequence into m(≤ n) words:
x1:e1 xe1+1:e2 .. xem−1+1:em
where each subsequence xi:j indicates a Chinese word spanning from characters xi to xj.
Word segmentation can be formalized as a sequence labeling problem (Xue and Shen 2003), where each character in the sentence is given a boundary tag representing its position in a word. Following Ng and Low (2004), joint word segmentation and partof-speech (POS) tagging can also be solved using a character classification approach by extending boundary tags to include POS information. For word segmentation we adopt the four boundary tags of Ng and Low (2004), B, M, E, and S, where B, M, and E mean the beginning, the middle, and the end of a word, respectively, and S indicates a singlecharacter word. The word segmentation result can be generated by splitting the labeled character sequence into subsequences of pattern S or BM∗E, indicating single-character words or multi-character words, respectively.
Given the character sequence x, the decoder finds the output ỹ that maximizes the score function:
ỹ = argmax y f(x, y) ·w
= argmax y
∑
xi∈x,yi∈y
f(xi, yi) ·w (1)
Where function f maps (x, y) into a feature vector, w is the parameter vector generated by the training algorithm, and f(x, y) ·w is the inner product of f(x, y) and w. The score of the sentence is further factorized into each character, where yi is the character classification label of character xi.
The training procedure of perceptron learns a discriminative model mapping from the inputs x to the outputs y. Algorithm 1 shows the perceptron algorithm for tuning the parameter w. The “averaged parameters” technology (Collins 2002) is used for better performance. The feature templates of the classifier is shown in Table 1. The function Pu(·) returns true for a punctuation character and false for others; the function T(·) classifies a character into four types: number, date, English letter, and others, corresponding to function values 1, 2, 3, and 4, respectively. Dependency parsing aims to link each word to its arguments so as to form a directed graph spanning the whole sentence. Normally, the directed graph is restricted to a
Algorithm 1 Perceptron training algorithm.
1: Input: Training set C 2: w← 0 3: for t← 1 .. T do ⊲ T iterations 4: for (x, y) ∈ C do 5: z̃← argmaxz f(x, z) ·w 6: if z̃ 6= y then 7: w← w + f(x, y)− f(x, z̃) ⊲ update the parameters 8: end if 9: end for
10: end for 11: Output: Parameters w
dependency tree where each word depends on exactly one parent, and all words find their parents. Given a sentence as a sequence n words:
x = x1 x2 .. xn
dependency parsing finds a dependency tree y, where (i, j) ∈ y is an edge from the head word xi to the modifier word xj. The root r ∈ x in the tree y has no head word, and each of the other words, j(j ∈ x and j 6= r), depends on a head word i(i ∈ x and i 6= j).
For many languages, the dependency structures are supposed to be projective. If xj is dependent on xi, then all the words between i and j must be directly or indirectly dependent on xi. Therefore, if we put the words in their linear order, preceded by the root, all edges can be drawn above the words without crossing. We follow this constraint because the dependency treebanks in our experiments are projective.
Following the edge-based factorization method (Eisner 1996), the score of a dependency tree can be factorized into the dependency edges in the tree. The spanning
tree method (McDonald, Crammer, and Pereira 2005) factorizes the score of the tree as the sum of the scores of all its edges, and the score of an edge is defined as the inner product of the feature vector f and the weight vector w. Given a sentence x, the parsing procedure searches for the candidate dependency tree with the maximum score:
ỹ = argmax y f(x, y) ·w
= argmax y
∑
(i,j)∈y
f(i, j) ·w (2)
The averaged perceptron algorithm is used again to train the parameter vector. A bottom–up dynamic programming algorithm is designed to search for the candidate parse with the maximum score as shown in Algorithm 2, where V[i, j] contains the candidate dependency fragments of the span [i, j]. The feature templates are similar to those of the first-ordered MST model (McDonald, Crammer, and Pereira 2005). Each feature is composed of some words and POS tags surround word i and/or word j, as well as an optional distance representation between the two words. Table 2 shows the feature templates without distance representations.","4 models for automatic annotation adaptation :In this section, we present a series of discriminative learning algorithms for the automatic adaptation of annotation guidelines. To facilitate the description of the algorithms, several shortened forms are adopted for convenience of description. We use source corpus to denote the corpus with the annotation guideline that we do not require, which is of course the source side of adaptation, and target corpus denotes the corpus with the desired guideline. Correspondingly, the annotation guidelines of the two corpora are denoted as source guideline and target guideline, and the classifiers following the two annotation guidelines are respectively named as source classifier and target classifier. Given a parallel annotated corpus, that is, a corpus labeled with two annotation
Algorithm 2 Dependency parsing algorithm.
1: Input: sentence x to be parsed 2: for [i, j] ⊆ [1, |x|] in topological order do 3: buf← ∅ 4: for k← i..j− 1 do ⊲ all partitions 5: for l ∈ V[i, k] and r ∈ V[k + 1, j] do 6: insert DERIV(l, r) into buf 7: insert DERIV(r, l) into buf 8: end for 9: end for 10: V[i, j]← best K in buf 11: end for 12: Output: the best of V[1, |x|] 13: function DERIV(p, c) 14: return p ∪ c ∪ {(p · root, c · root)} ⊲ new derivation 15: end function
guidelines, a transfer classifier can be trained to capture the regularity of transformation from the source annotation to the target annotation. The classifiers mentioned here are normal discriminative classifiers that take a set of features as input and give a classification label as output. For the POS tagging problem, the classification label is a POS tag, and for the parsing task, the classification label is a dependency edge, a constituency span, or a shift-reduce action.
The parallel annotated corpus is the knowledge source of annotation adaptation. The annotation quality and data size of the parallel annotated corpus determine the accuracy of the transfer classifier. Such a corpus is difficult to build manually, although we can generate a noisy one automatically from the source corpus and the target corpus. For example, we can apply the source classifier on the target corpus, thus generating
a parallel annotated corpus with noisy source annotations and accurate target annotations. The training procedure of the transfer classifier predicts the target annotations with guiding features extracted from the source annotations. This approach can alleviate the effect of the noise in source annotations, and learn annotation adaptation regularities accurately. By reducing the noise in the automatically generated parallel annotated corpus, a higher accuracy of annotation adaptation can be achieved.
In the following sections, we first describe the transfer classifier that reveals the intrinsic principles of annotation adaptation, then describe a series of successively enhanced models that are developed from our previous investigation (Jiang, Huang, and Liu 2009; Jiang et al. 2012). In the simplest model (Model 1), two classifiers, a source classifier and a transfer classifier, are used in a cascade. The classification results of the lower source classifier provide additional guiding features to the upper transfer classifier, yielding an improved classification result. A variant of the first model (Model 2) uses the transfer classifier to transform the source corpus from source guideline to target guideline first, then merges the transformed source corpus into the target corpus in order to train an improved target classifier on the enlarged corpus. An optimized model (Model 3) is further proposed based on Model 2. Two optimization strategies, iterative training and predict-self re-estimation, are adopted to improve the efficiency of annotation adaptation, in order to fully utilize the knowledge in heterogeneous corpora. In order to learn the regularity of the adaptation from one annotation guideline to another, a parallel annotated corpus is needed to train the transfer classifier. The parallel annotated corpus is a corpus with two different annotation guidelines, the source guideline and the target guideline. With the target annotation labels as learning objectives and the source annotation labels as guiding information, the transfer classifier learns the statistical regularity of the adaptation from the source annotations to the target annotations.
The training procedure of the transfer classifier is analogous to the training of a normal classifier except for the introduction of additional guiding features. For word segmentation, the most intuitive guiding feature is the source annotation label itself. For dependency parsing, an effective guiding feature is the dependency path between the hypothetic head and modifier, as shown in Figure 3. However, our effort is not limited to this, and more special features are introduced: A classification label or dependency path is attached to each feature of the baseline classifier to generate combined guiding features. This is similar to the feature design in discriminative dependency parsing (McDonald, Crammer, and Pereira 2005; McDonald and Pereira 2006), where the basic features, composed of words and POSs in the context, are also conjoined with link direction and distance in order to generate more special features.
Table 3 shows an example of guide features (as well as baseline features) for word segmentation, where “α = B” indicates that the source classification label of the current character is B, demarcating the beginning of a word. The combination strategy derives a series of specific features, helping the transfer classifier to produce more precise classifications. The parameter-tuning procedure of the transfer classifier will automatically learn the regularity of using the source annotations to guide the classification decision. In decoding, if the current character shares some basic features in Table 3 and it is classified as B in the source annotation, then the transfer classifier will probably classify it as M.
In addition, the original features used in the normal classifier are also used in order to leverage the knowledge from the target annotation of the parallel annotated corpus, and the training procedure of the transfer classifier also learns the relative weights between the guiding features and the original features. Therefore, the knowledge from both the source annotation and the target annotation are automatically integrated together, and higher and more stable prediction accuracy can be achieved. The most intuitive model for annotation adaptation uses two cascaded classifiers, the source classifier and the transfer classifier, to integrate the knowledge in corpora with different annotation guidelines. In the training procedure, a source classifier is trained on the source corpus and is used to process the target corpus, generating a parallel annotated corpus (albeit a noisy one). Then, the transfer classifier is trained on the parallel annotated corpus,with the target annotations as the classification labels, and the source annotation as guiding information. Figure 4 depicts the training pipeline. The best training iterations for the source classifier and the transfer classifier are determined on the development sets of the source corpus and target corpus.
In the decoding procedure, a sequence of characters (for word segmentation) or words (for dependency parsing) is input into the source classifier to obtain a classification result under the source guideline; then it is input into the transfer classifier with this classification result as the guiding information to get the final result following the target guideline. This coincides with the stacking method for combining dependency parsers (Martins et al. 2008; Nivre and McDonald 2008), and is also similar to the Pred baseline for domain adaptation in Daumé et al. (Daumé III and Marcu 2006; Daumé III 2007). Figure 5 shows the pipeline for decoding. The previous model has a drawback: It has to cascade two classifiers in decoding to integrate the knowledge in two corpora, which seriously degrades the processing speed. Here we describe a variant of the previous model, aiming at automatic transformation (rather than integration as in Model 1) between annotation guidelines of humanannotated corpora. The source classifier and the transfer classifier are trained in the same way as in the previous model. The transfer classifier is used to process the source corpus, with the source annotations as guiding information, so as to relabel the source corpus with the target annotation guideline. By merging the target corpus and the transformed source corpus for the training of the final classifier, improved classification accuracy can be achieved.
From this point on we describe the pipelines of annotation transformation in pseudo-codes for simplicity and convenience of extensions. Algorithm 3 shows the overall training algorithm for the variant model. Cs and Ct denote the source corpus and the target corpus. Ms and Ms→t denote the source classifier and the transfer classifier. C q p denotes the p corpus relabeled in q annotation guideline; for example, C t s is a corpus that labels the text of the source corpus with the target guideline. Functions TRAIN and TRANSTRAIN train the source classifier and the transfer classifier, respectively, both invoking the perceptron algorithm, but with different feature sets. Functions ANNOTATE and TRANSANNOTATE call the function DECODE with different models (source/transfer classifiers), feature functions (without/with guiding features), and inputs (raw/source-annotated sentences).
In the algorithm the parameters corresponding to development sets are omitted for simplicity. Compared to the online knowledge integration methodology of the previous model, annotation transformation leads to improved performance in an offline manner by integrating corpora before the training procedure. This approach enables processing speeds several times faster than the cascaded classifiers in the previous model. It also has another advantage in that we can integrate the knowledge in more than two corpora without slowing down the processing of the final classifier. The training of the transfer classifier is based on an automatically generated (rather than a gold standard) parallel annotated corpus, where the source annotations are
Algorithm 3 Baseline annotation adaptation.
1: function ANNOTRANS(Cs, Ct) 2: Ms ← TRAIN(Cs) ⊲ source classifier 3: Cst ← ANNOTATE(Ms, Ct) 4: Ms→t ← TRANSTRAIN(C s t , Ct) ⊲ transfer classifier 5: Cts ← TRANSANNOTATE(Ms→t, Cs) 6: Ct∗ ← C t s ∪ Ct ⊲ integrated corpus with target guideline 7: return Ct∗ 8: end function 9: function DECODE(M, Φ, x)
10: return argmaxy∈GEN(x) S(y|M,Φ, x) 11: end function
provided by the source classifier. Therefore, the performance of annotation transformation is correspondingly determined by the accuracy of the source classifier, and we can generate a more accurate parallel annotated corpus for better annotation adaptation if an improved source classifier can be obtained. Based on Model 2, two optimization strategies—iterative bidirectional training and predict-self hypothesis—are introduced to optimize the parallel annotated corpora for better annotation adaptation.
We first use an iterative training procedure to gradually improve the transformation accuracy by iteratively optimizing the parallel annotated corpora. In each training iteration, both source-to-target and target-to-source annotation transformations are performed, and the transformed corpora are used to provide better annotations for the parallel annotated corpora of the next iteration. Then in the new iteration, the better parallel annotated corpora will result in more accurate transfer classifiers, resulting in the generation of better transformed corpora.
Algorithm 4 shows the overall procedure of the iterative training method. The loop of lines 6–13 iteratively performs source-to-target and target-to-source annotation transformations. The source annotations of the parallelly annotated corpora, Cst and Cts, are initialized by applying the source and target classifiers on the target and source corpora, respectively (lines 2–5). In each training iteration, the transfer classifiers are trained on the current parallel annotated corpora (lines 7–8); they are used to produce the transformed corpora (lines 9–10), which provide better annotations for the parallel annotated corpora of the next iteration. The iterative training terminates when the performance of the classifier trained on the merged corpus Cts ∪ Ct converges (line 13).
The discriminative training of TRANSTRAIN predicts the target annotations with the guidance of source annotations. In the first iteration, the transformed corpora generated by the transfer classifiers are better than the initial ones generated by the source and target classifiers, due to the assistance of the guiding features. In the following iterations,
Algorithm 4 Iterative annotation transformation.
1: function ITERANNOTRANS(Cs, Ct) 2: Ms ← TRAIN(Cs) ⊲ source classifier 3: Cst ← ANNOTATE(Ms, Ct) 4: Mt ← TRAIN(Ct) ⊲ target classifier 5: Cts ← ANNOTATE(Mt, Cs) 6: repeat 7: Ms→t ← TRANSTRAIN(C s t , Ct) ⊲ source-to-target transfer classifier 8: Mt→s ← TRANSTRAIN(C t s, Cs) ⊲ target-to-source transfer classifier
9: Cts ← TRANSANNOTATE(Ms→t, Cs) 10: Cst ← TRANSANNOTATE(Mt→s, Ct) 11: Ct∗ ← C t s ∪ Ct 12: M∗ ← TRAIN(C t ∗) ⊲ enhanced classifier trained on merged corpus 13: until EVAL(M∗) converges 14: return Ct∗ 15: end function 16: function DECODE(M, Φ, x) 17: return argmaxy∈GEN(x) S(y|M,Φ, x) 18: end function
the transformed corpora provide better annotations for the parallel annotated corpora of the subsequent iteration; the transformation accuracy will improve gradually along with the optimization of the parallel annotated corpora until convergence.
The predict-self hypothesis is introduced to improve the transformation accuracy from another perspective. This hypothesis is implicit in many unsupervised learning approaches, such as Markov random field; it has also been successfully used by Daumé III (2009) in unsupervised dependency parsing. The basic idea of predict-self is, if a prediction is a better candidate for an input, it would be easier to convert it back to the original input by a reverse procedure. If applied to annotation transformation, predictself indicates that a better transformation candidate following the target guideline can be more easily transformed back to the original form following the source guideline.
The most intuitive strategy to introduce the predict-self methodology into annotation transformation is using a reversed annotation transformation procedure to filter out unreliable predictions of the previous transformation. In detail, a source-to-target annotation transformation procedure is performed on the source corpus to obtain a prediction that follows the target guideline; then a second, target-to-source transformation procedure is performed on this prediction result to check whether it can be transformed back to the original source annotation. The source-to-target prediction results that fail in this reverse-verification step are discarded, so this strategy can be called predict-self filtering.
A more sophisticated strategy can be called predict-self re-estimation. Instead of using the reversed transformation procedure for filtering, the re-estimation strategy integrates the scores given by the source-to-target and target-to-source annotation transformation models when evaluating the transformation candidates. By properly tuning the relative weights of the two transformation directions, better transformation performance is achieved. The scores of the two transformation models are weighted, integrated in a log-linear manner:
S+(y|Ms→t,Mt→s,Φ, x) = (1− λ)× S(y|Ms→t,Φ, x) + λ× S(x|Mt→s,Φ, y) (3)
The weight parameter λ is tuned on the development set. To integrate the predictself reestimation into the iterative transformation training, a reversed transformation model is introduced and the enhanced scoring function is used when the function TRANSANNOTATE invokes the function DECODE.","5 experiments :To evaluate the performance of annotation adaptation, we experiment on two important NLP tasks, Chinese word segmentation and dependency parsing, both of which can be modeled as discriminative classification problems. For both tasks, we give the performances of the baseline models and the annotation adaptation algorithms. We perform annotation adaptation for word segmentation from People’s Daily (PD) (Yu et al. 2001) to Penn Chinese Treebank 5.0 (CTB) (Xue et al. 2005). The two corpora are built according to different segmentation guidelines and differ largely in quantity of data. CTB is smaller in size with about 0.5M words, whereas PD is much larger, containing nearly 6M words. Table 4 shows the data partitioning for the two corpora. We train the baseline perceptron classifiers for Chinese word segmentation on the training sets of CTB and SPD, using corresponding development sets to determine the best As a variant of Model 1, Model 2 shares the same transfer classifier, and differs only in training and decoding of the final classifier. Tables 8 and 9 show the performances of systems resulting from Models 1 and 2, as well as the classifiers trained on the directly merged corpora. The time costs for decoding are also listed to facilitate the practical comparison.
We find that the simple corpus merging strategy leads to a dramatic decrease in accuracy, due to the different and incompatible annotation guidelines. Model 1, the simplest model for annotation adaptation, gives significant improvement over the baseline classifiers for word segmentation and dependency parsing. This indicates that the statistical regularities for annotation adaptation learned by the transfer classifiers
bring performance improvement, utilizing guided decisions in the cascaded classifiers. Model 2 leads to classifiers with accuracy increments comparable to those of Model 1, while consuming only one third of the decoding time. It is inconsistent with our expectation. The strategy of directly transforming the source corpus to the target guideline also facilitates the utilization of more than one source corpus.
We first introduce the iterative training strategy to Model 2. The corresponding development sets are used to determine the best training iterations for the iterative annotation transformations. After each iteration, we test the performance of the classifiers trained on the merged corpora. Figures 7 and 8 show the performance curves for Chinese word segmentation and semantic dependency parsing, respectively, with iterations ranging from 1 to 10. The performance of Model 2 is naturally included
in the curves (located at iteration 1). The curves show that, for both segmentation and parsing, the accuracies of the classifiers trained on the merged corpora consistently improve in the earlier iterations (e.g., from iteration 2 to iteration 5 for word segmentation).
Experiments for introduction of predict-self filtering and predict-self re-estimation are shown in Figures 9 and 10. The curves show the performances of the predictself re-estimation with a series of weight parameters, ranging from 0 to 1 with step 0.05. Note that in both figures, the points at λ = 0 show the performances of Model 2. We find that predict-self filtering brings a slight improvement over the baseline for
word segmentation, but even decreases the accuracy for dependency parsing. An initial analysis on the experimental results reveals that the filtering strategy discards some complicated sentences in the source corpora, and the discarded sentences would bring further improvement if properly used. For example, in word segmentation, predict-self filtering discards 5% of sentences from the source corpus, containing nearly 10% of the training words. For the two tasks, the predict-self re-estimation outperforms the filtering strategy. With properly tuned weights, predict-self re-estimation can make better use of the training data. The largest accuracy improvements achieved over Model 2 for word segmentation and dependency parsing are 0.3 points and 0.6 points.
Figures 11 and 12 show the performance curves after the introduction of both iterative training and predict-self re-estimation on the basis of Model 2 (this enhanced","6 discussion: application situations :Automatic annotation adaptation aims to transform the annotations in a corpus to the annotations following other guidelines. The models for annotation adaptation use a transfer classifier to learn the statistical correspondence regularities between different annotation guidelines. These statistical regularities are learned from a parallel annotated corpus, which does not need to be manually annotated. In fact, the models for annotation adaptation train the transfer classifier on an automatically generated parallel annotated corpus, which is generated by processing a corpus with a classifier trained on another corpus. That is to say, if we want to conduct annotation adaptation across several corpora, no additional corpora need to be manually annotated. This setting makes the strategy of annotation adaptation more general, because it is much harder to manually annotate a parallel annotated corpus, regardless of the language or the NLP problem under consideration. To tackle the problem of noise in automatically generated annotations, the advanced models we designed generate a better parallel annotated corpus by making use of strategies such as iterative optimization.
Automatic annotation adaptation can be applied in any situation where we have multiple corpora with different and incompatible annotation philosophies for the same task. As our case studies, both Chinese word segmentation and dependency parsing have more than one corpora with different annotation guidelines, such as the People’s Daily and the Penn Chinese Treebank for Chinese word segmentation. In a more abstract view, constituency grammar and dependency grammar can be treated as two annotation guidelines for parsing. The syntactic knowledge in a constituency treebank and a dependency treebank, therefore, can be integrated by automatic annotation adaptation. For example, the LinGo Redwoods Treebank can also be transformed to the annotation guideline of the Semantic Dependency Treebank.
Furthermore, the annotations (such as a grammar) given by bilingual projection or unsupervised induction can be seen as following a special annotation philosophy. For bilingually projected annotations, the annotation guideline would be similar to that of the counterpart language. For unsupervised induced annotations, the annotation guideline reflects the statistical structural distribution of a specific data set. In both situations, the underlying annotation guidelines may be largely different from that of the testing sets, which usually come from human-annotated corpora. The system trained on a bilingually projected or unsupervised induced corpus may perform poorly on an existing
testing set, but if the projected or induced corpus has high inner consistency, it could improve a system trained on an existing corpus by automatic annotation adaptation. In this point of view, the practical value of the current work on bilingual projection and unsupervised induction may be underestimated, and annotation adaptation could make better use of the projected or induced knowledge.3","7 related work :There has already been some preliminary work tackling the divergence between different annotation guidelines. Gao et al. (2004) described a transformation-based converter to transfer a certain word segmentation result to another annotation guideline. They designed class-type transformation templates and used the transformation-based error-driven learning method of Brill (1995) to learn what word delimiters should be modified. Many efforts have been devoted to manual treebank transformation, where PTB is adapted to other grammar formalisms, such as CCG and LFG (Cahill et al. 2002; Hockenmaier and Steedman 2007). However, all these are heuristic-based—that is, they need manually designed transformation templates and involve heavy human engineering. Such strategies are hard to be generalized to POS tagging, not to mention other complicated structural prediction tasks.
We investigated the automatic integration of word segmentation knowledge in differently annotated corpora (Jiang, Huang, and Liu 2009; Jiang et al. 2012), which can be seen as the preliminary work of automatic annotation adaptation. Motivated by our initial investigation, researchers applied similar methodologies to constituency parsing (Sun, Wang, and Zhang 2010; Zhu, Zhu, and Hu 2011) and word segmentation (Sun and Wan 2012). This previous work verified the effectiveness of automatic annotation adaptation, but did not reveal the essential definition of the problem nor the intrinsic principles of the solutions. Instead, this work clearly defines the problem of annotation adaptation, reveals the intrinsic principles of the solutions, and systematically describes a series of gradually improved models. The most advanced model learns transformation regularities much better and achieves significant higher accuracy for both word segmentation and dependency parsing, without slowing down the final language processors.
The problem of automatic annotation adaptation can be seen as a special case of transfer learning (Pan and Yang 2010), where the source and target tasks are similar, but not identical. More specifically, the problem related to annotation adaptation assumes that the labeling mechanism across the source and target tasks are the same, but the predictive functions are different. The goal of annotation adaptation is to adapt the source predictive function to be used for the target task by exploiting the labeled data of the source task and the target task. Furthermore, automatic annotation adaptation approximately falls into the spectrum of relational-knowledge-transfer problems (Mihalkova, Huynh, and Mooney 2007; Mihalkova and Mooney 2008; Davis and Domingos 2009), but it tackles problems where the relations among data between the source and target domains can be largely isomerous—or, in other words, with different and incompatible annotation schemes. This work enriches the research of transfer learning by proposing and solving an NLP problem different from previous situations. For more details of transfer learning please refer to the survey of Pan and Yang (2010).
3 We have performed preliminary experiments on word segmentation. Bilingual projection was conducted from English to Chinese with the Chinese–English FBIS as the bilingual corpus. By annotation adaptation, the projected corpus for word segmentation brings a significant F-measure increment of nearly 0.6 points over the baseline trained on CTB only.
The training procedure for an annotation adaptation model requires a parallel annotated corpus (which may be automatically generated); this fact puts the method into the neighborhood of the family of approaches known as annotation projection (Hwa et al. 2002, 2005; Ganchev, Gillenwater, and Taskar 2009; Smith and Eisner 2009; Jiang and Liu 2010; Das and Petrov 2011). Essentially, annotation adaptation and annotation projection tackle different problems; the former aims to transform the annotations from one guideline to another (of course in the same language), whereas the latter aims to project the annotation (as well as the annotation guideline) from one language to another. Therefore, the machine learning methods for annotation adaptation pay attention to automatic transformation of annotations, while for annotation projection, the machine learning methods focus on the bilingual projection across languages.
Co-training (Sarkar 2001) and classifier combination (Nivre and McDonald 2008) are two techniques for training improved dependency parsers. The co-training technique lets two different parsing models learn from each other during the parsing of unlabeled text: One model selects some unlabeled sentences it can confidently parse, and provides them to the other model as additional training data in order to train more powerful parsers. The classifier combination lets graph-based and transition-based dependency parsers utilize the features extracted from each other’s parsing results, to obtain combined, enhanced parsers. The two techniques aim to let two models learn from each other on the same corpus with the same distribution and annotation guideline, whereas our strategy aims to integrate the knowledge in multiple corpora with different annotation guidelines.
The iterative training procedure used in the optimized model shares some similarities with the co-training algorithm in parsing (Sarkar 2001), where the training procedure lets two different models learn from each other during parsing of the raw text. The key idea of co-training is to utilize the complementarity of different parsing models to mine additional training data from raw text, whereas iterative training for annotation adaptation emphasizes the iterative optimization of the parallel annotated corpora used to train the transfer classifiers. The predict-self methodology is implicit in many unsupervised learning approaches; it has been successfully used in unsupervised dependency parsing (Daumé III 2009). We adapt this idea to the scenario of annotation adaptation to improve transformation accuracy.
In recent years much effort has been devoted to the improvement of word segmentation and dependency parsing. For example, the introduction of global training or complicated features (Zhang and Clark 2007, 2010); the investigation of word structures (Li 2011); the strategies of hybrid, joint, or stacked modeling (Nakagawa and Uchimoto 2007; Kruengkrai et al. 2009; Wang, Zong, and Su 2010; Sun 2011); and the semi-supervised and unsupervised technologies utilizing raw text (Zhao and Kit 2008; Johnson and Goldwater 2009; Mochihashi, Yamada, and Ueda 2009; Hewlett and Cohen 2011). We believe that the annotation adaptation technologies can be adopted jointly with complicated features, system combination, and semi-supervised/unsupervised technologies to further improve the performance of word segmentation and dependency parsing.","8 conclusion and future work :We have described the problem of annotation adaptation and the intrinsic principles of its solutions, and proposed a series of successively enhanced models that can automatically adapt the divergence between different annotation formats. These models learn the statistical regularities of adaptation between different annotation guidelines,
and integrate the knowledge in corpora with different annotation guidelines. In the problems of Chinese word segmentation and semantic dependency parsing, annotation adaptation algorithms bring significant improvements by integrating the knowledge in differently annotated corpora, People’s Daily and Penn Chinese Treebank for word segmentation, and Penn Chinese Treebank and Semantic Dependency Parsing for dependency parsing. For both tasks, annotation adaptation leads to a segmenter and a parser achieving the state-of-the-art, despite using only local features in single classifiers.
Many aspects related to annotation adaptation deserve further investigation in the future. First, models for annotation adaptation can be adapted to other NLP tasks such as semantic analysis. Second, jointly tackling the divergences in both annotations and domains is an important problem. In addition, an unsupervised-induced or bilingually projected corpus, despite performing poorly on the specified testing data, may have high inner annotation consistency. That is to say, the induced corpora can be treated as a knowledge source following another annotation guideline, and the performance of current unsupervised or bilingually projected models may be seriously underestimated. Annotation adaptation may give us a new perspective on knowledge induction and measurement for such methods.","Manually annotated corpora are indispensable resources, yet for many annotation tasks, such as the creation of treebanks, there exist multiple corpora with different and incompatible annotation guidelines. This leads to an inefficient use of human expertise, but it could be remedied by integrating knowledge across corpora with different annotation guidelines. In this article we describe the problem of annotation adaptation and the intrinsic principles of the solutions, and present a series of successively enhanced models that can automatically adapt the divergence between different annotation formats. We evaluate our algorithms on the tasks of Chinese word segmentation and dependency parsing. For word segmentation, where there are no universal segmentation guidelines because of the lack of morphology in Chinese, we perform annotation adaptation from the much larger People’s Daily corpus to the smaller but more popular Penn Chinese Treebank. For dependency parsing, we perform annotation adaptation from the Penn Chinese Treebank to a semantics-oriented Dependency Treebank, which is annotated using significantly different annotation guidelines. In both experiments, automatic annotation adaptation brings significant improvement, achieving state-of-the-art performance despite the use of purely local features in training.","[{""affiliations"": [], ""name"": ""Wenbin Jiang""}, {""affiliations"": [], ""name"": ""Yajuan L\u00fc""}, {""affiliations"": [], ""name"": ""Liang Huang""}, {""affiliations"": [], ""name"": ""Qun Liu""}]",SP:376fabb797ac14f81d1ea7f54ed5432a0a7bf244,"[{""authors"": [""Blitzer"", ""John"", ""Ryan McDonald"", ""Fernando Pereira.""], ""title"": ""Domain adaptation with structural correspondence learning"", ""venue"": ""Proceedings of EMNLP, pages 120\u2013128, Sydney."", ""year"": 2006}, {""authors"": [""Brill"", ""Eric.""], ""title"": ""Transformation-based error-driven learning and natural language processing: A case study in part-of-speech tagging"", ""venue"": ""Computational Linguistics, 21(4):543\u2013565."", ""year"": 1995}, {""authors"": [""Buchholz"", ""Sabine"", ""Erwin Marsi.""], ""title"": ""CONLL-X shared task on multilingual dependency parsing"", ""venue"": ""Proceedings of CoNLL, pages 149\u2013164, New York, NY."", ""year"": 2006}, {""authors"": [""Cahill"", ""Aoife"", ""Mairead McCarthy"", ""Josef van Genabith"", ""Andy Way.""], ""title"": ""Automatic annotation of the Penn treebank with LFG F-structure information"", ""venue"": ""Proceedings of the LREC Workshop, Las Palmas."", ""year"": 2002}, {""authors"": [""Che"", ""Wanxiang"", ""Meishan Zhang"", ""Yanqiu Shao"", ""Ting Liu.""], ""title"": ""Semeval-2012 task 5: Chinese semantic dependency parsing"", ""venue"": ""Proceedings of SemEval, pages 378\u2013384, Montreal."", ""year"": 2012}, {""authors"": [""Collins"", ""Michael.""], ""title"": ""Discriminative training methods for hidden Markov models: Theory and experiments with perceptron algorithms"", ""venue"": ""Proceedings of EMNLP, pages 1\u20138, Philadelphia, PA."", ""year"": 2002}, {""authors"": [""Das"", ""Dipanjan"", ""Slav Petrov.""], ""title"": ""Unsupervised part-of-speech tagging with bilingual graph-based projections"", ""venue"": ""Proceedings of ACL, pages 600\u2013609, Portland, OR."", ""year"": 2011}, {""authors"": [""Daum\u00e9 III"", ""Hal.""], ""title"": ""Frustratingly easy domain adaptation"", ""venue"": ""Proceedings of ACL, pages 256\u2013263, Prague."", ""year"": 2007}, {""authors"": [""Daum\u00e9 III"", ""Hal.""], ""title"": ""Unsupervised searchbased structured prediction"", ""venue"": ""Proceedings of ICML, pages 209\u2013216, Montreal."", ""year"": 2009}, {""authors"": [""Daum\u00e9 III"", ""Hal"", ""Daniel Marcu.""], ""title"": ""Domain adaptation for statistical classifiers"", ""venue"": ""Journal of Artificial Intelligence Research, 26:101\u2013126."", ""year"": 2006}, {""authors"": [""Davis"", ""Jesse"", ""Pedro Domingos.""], ""title"": ""Deep transfer via second-order Markov logic"", ""venue"": ""Proceedings of ICML, pages 217\u2013224, Montreal."", ""year"": 2009}, {""authors"": [""Eisner"", ""Jason M.""], ""title"": ""Three new probabilistic models for dependency parsing: An exploration"", ""venue"": ""Proceedings of COLING, pages 340\u2013345, Copenhagen."", ""year"": 1996}, {""authors"": [""Ganchev"", ""Kuzman"", ""Jennifer Gillenwater"", ""Ben Taskar.""], ""title"": ""Dependency grammar induction via bitext projection constraints"", ""venue"": ""Proceedings of ACL, pages 369\u2013377, Singapore."", ""year"": 2009}, {""authors"": [""Gao"", ""Jianfeng"", ""Andi Wu"", ""Mu Li"", ""Chang-Ning Huang"", ""Hongqiao Li"", ""Xinsong Xia"", ""Haowei Qin.""], ""title"": ""Adaptive Chinese word segmentation"", ""venue"": ""Proceedings of ACL, pages 462\u2013469, Barcelona."", ""year"": 2004}, {""authors"": [""Hewlett"", ""Daniel"", ""Paul Cohen.""], ""title"": ""Fully unsupervised word segmentation with BVE and MDL"", ""venue"": ""Proceedings of ACL, pages 540\u2013545, Portland, OR."", ""year"": 2011}, {""authors"": [""Hockenmaier"", ""Julia"", ""Mark Steedman.""], ""title"": ""CCGBank: A corpus of CCG derivations and dependency structures extracted from the Penn treebank"", ""venue"": ""Computational Linguistics, 33(3):355\u2013396."", ""year"": 2007}, {""authors"": [""Hwa"", ""Rebecca"", ""Philip Resnik"", ""Amy Weinberg"", ""Clara Cabezas"", ""Okan Kolak.""], ""title"": ""Bootstrapping parsers via syntactic projection across parallel texts"", ""venue"": ""Natural Language Engineering, 11(3):311\u2013325."", ""year"": 2005}, {""authors"": [""Hwa"", ""Rebecca"", ""Philip Resnik"", ""Amy Weinberg"", ""Okan Kolak.""], ""title"": ""Evaluating translational correspondence using annotation projection"", ""venue"": ""Proceedings of ACL, pages 392\u2013399, Philadephia, PA."", ""year"": 2002}, {""authors"": [""Jiang"", ""Wenbin"", ""Liang Huang"", ""Qun Liu.""], ""title"": ""Automatic adaptation of annotation standards: Chinese word segmentation and POS tagging \u2013 A case study"", ""venue"": ""Proceedings of ACL, pages 522\u2013530,"", ""year"": 2009}, {""authors"": [""Jiang"", ""Wenbin"", ""Liang Huang"", ""Yajuan L\u00fc"", ""Qun Liu.""], ""title"": ""A cascaded linear model for joint Chinese word segmentation and part-of-speech tagging"", ""venue"": ""Proceedings of ACL, pages 897\u2013904, Columbus, OH."", ""year"": 2008}, {""authors"": [""Jiang"", ""Wenbin"", ""Qun Liu.""], ""title"": ""Dependency parsing and projection based on word-pair classification"", ""venue"": ""Proceedings of the ACL, pages 12\u201320, Uppsala."", ""year"": 2010}, {""authors"": [""Jiang"", ""Wenbin"", ""Fandong Meng"", ""Qun Liu"", ""Yajuan L\u00fc.""], ""title"": ""Iterative annotation transformation with predict-self reestimation for Chinese word segmentation"", ""venue"": ""Proceedings of EMNLP,"", ""year"": 2012}, {""authors"": [""Johnson"", ""Mark"", ""Sharon Goldwater.""], ""title"": ""Improving nonparameteric Bayesian inference: Experiments on unsupervised word segmentation with adaptor grammars"", ""venue"": ""Proceedings of NAACL,"", ""year"": 2009}, {""authors"": [""Kruengkrai"", ""Canasai"", ""Kiyotaka Uchimoto"", ""Junichi Kazama"", ""Yiou Wang"", ""Kentaro Torisawa"", ""Hitoshi Isahara""], ""title"": ""An error-driven word-character hybrid model for joint Chinese word segmentation"", ""year"": 2009}, {""authors"": [""Li"", ""Zhongguo.""], ""title"": ""Parsing the internal structure of words: A new paradigm for Chinese word segmentation"", ""venue"": ""Proceedings of ACL, pages 1,405\u20131,414, Portland, OR."", ""year"": 2011}, {""authors"": [""Marcus"", ""Mitchell P."", ""Beatrice Santorini"", ""Mary Ann Marcinkiewicz.""], ""title"": ""Building a large annotated corpus of English: The Penn treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1993}, {""authors"": [""Martins"", ""Andr\u00e9 F.T."", ""Dipanjan Das"", ""Noah A. Smith"", ""Eric P. Xing.""], ""title"": ""Stacking dependency parsers"", ""venue"": ""Proceedings of EMNLP, pages 157\u2013166, Honolulu, HI."", ""year"": 2008}, {""authors"": [""McDonald"", ""Ryan"", ""Koby Crammer"", ""Fernando Pereira.""], ""title"": ""Online large-margin training of dependency parsers"", ""venue"": ""Proceedings of ACL, pages 91\u201398, Ann Arbor, MI."", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Fernando Pereira.""], ""title"": ""Online learning of approximate dependency parsing algorithms"", ""venue"": ""Proceedings of EACL, pages 81\u201388, Trento."", ""year"": 2006}, {""authors"": [""Mihalkova"", ""Lilyana"", ""Tuyen Huynh"", ""Raymond J. Mooney.""], ""title"": ""Mapping and revising Markov logic networks for transfer learning"", ""venue"": ""Proceedings of AAAI, volume 7, pages 608\u2013614, Vancouver."", ""year"": 2007}, {""authors"": [""Mihalkova"", ""Lilyana"", ""Raymond J. Mooney.""], ""title"": ""Transfer learning by mapping with minimal target data"", ""venue"": ""Proceedings of AAAI Workshop Transfer Learning for Complex Tasks, Chicago, IL."", ""year"": 2008}, {""authors"": [""Mochihashi"", ""Daichi"", ""Takeshi Yamada"", ""Naonori Ueda.""], ""title"": ""Bayesian unsupervised word segmentation with nested Pitman-Yor language modeling"", ""venue"": ""Proceedings of ACL-IJCNLP,"", ""year"": 2009}, {""authors"": [""Nakagawa"", ""Tetsuji"", ""Kiyotaka Uchimoto.""], ""title"": ""A hybrid approach to word segmentation and POS tagging"", ""venue"": ""Proceedings of ACL, pages 217\u2013220, Prague."", ""year"": 2007}, {""authors"": [""Ng"", ""Hwee Tou"", ""Jin Kiat Low""], ""title"": ""Chinese part-of-speech tagging: One-at-a-time or all-at-once? Word-based or character-based"", ""venue"": ""In Proceedings of EMNLP,"", ""year"": 2004}, {""authors"": [""Nivre"", ""Joakim"", ""Ryan McDonald.""], ""title"": ""Integrating graph-based and transition-based dependency parsers"", ""venue"": ""Proceedings of ACL, pages 950\u2013958, Columbus, OH."", ""year"": 2008}, {""authors"": [""Oepen"", ""Stephan"", ""Kristina Toutanova"", ""Stuart Shieber"", ""Thorsten Brants""], ""title"": ""The LinGo Redwoods treebank: Motivation and preliminary applications"", ""year"": 2002}, {""authors"": [""Pan"", ""Sinno Jialin"", ""Qiang Yang.""], ""title"": ""A survey on transfer learning"", ""venue"": ""IEEE TKDE, 22(10):1345\u20131359."", ""year"": 2010}, {""authors"": [""Sarkar"", ""Anoop.""], ""title"": ""Applying co-training methods to statistical parsing"", ""venue"": ""Proceedings of NAACL, pages 1\u20138, Pittsburgh, PA."", ""year"": 2001}, {""authors"": [""Smith"", ""David"", ""Jason Eisner.""], ""title"": ""Parser adaptation and projection with quasi-synchronous grammar features"", ""venue"": ""Proceedings of EMNLP, volume 2, pages 822\u2013831, Singapore."", ""year"": 2009}, {""authors"": [""Sun"", ""Weiwei.""], ""title"": ""A stacked sub-word model for joint Chinese word segmentation and part-of-speech tagging"", ""venue"": ""Proceedings of ACL, pages 1,385\u20131,394, Portland, OR."", ""year"": 2011}, {""authors"": [""Sun"", ""Weiwei"", ""Xiaojun Wan.""], ""title"": ""Reducing approximation and estimation errors for Chinese lexical processing with heterogeneous annotations"", ""venue"": ""Proceedings of ACL, volume 1, pages 232\u2013241,"", ""year"": 2012}, {""authors"": [""Sun"", ""Weiwei"", ""Rui Wang"", ""Yi Zhang.""], ""title"": ""Discriminative parse reranking for Chinese with homogeneous and heterogeneous annotations"", ""venue"": ""Proceedings of CIPS-SIGHAN, Beijing. Available at"", ""year"": 2010}, {""authors"": [""Wang"", ""Kun"", ""Chengqing Zong"", ""Keh-Yih Su.""], ""title"": ""A character-based joint model for Chinese word segmentation"", ""venue"": ""Proceedings of COLING, pages 1,173\u20131,181, Beijing."", ""year"": 2010}, {""authors"": [""Xue"", ""Nianwen"", ""Libin Shen.""], ""title"": ""Chinese word segmentation as LMR tagging"", ""venue"": ""Proceedings of SIGHAN Workshop, volume 17, pages 176\u2013179, Sapporo."", ""year"": 2003}, {""authors"": [""Xue"", ""Nianwen"", ""Fei Xia"", ""Fu-Dong Chiou"", ""Martha Palmer.""], ""title"": ""The Penn Chinese treebank: Phrase structure annotation of a large corpus"", ""venue"": ""Natural Language Engineering, 11(2):207\u2013238."", ""year"": 2005}, {""authors"": [""H. Yamada"", ""Y. Matsumoto.""], ""title"": ""Statistical dependency analysis with support vector machines"", ""venue"": ""Proceedings of IWPT, pages 195\u2013206, Nancy."", ""year"": 2003}, {""authors"": [""Yu"", ""Shiwen"", ""Jianming Lu"", ""Xuefeng Zhu"", ""Huiming Duan"", ""Shiyong Kang"", ""Honglin Sun"", ""Hui Wang"", ""Qiang Zhao"", ""Weidong Zhan.""], ""title"": ""Processing norms of modern Chinese corpus"", ""venue"": ""Technical"", ""year"": 2001}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""Chinese segmentation with a word-based perceptron algorithm"", ""venue"": ""Proceedings of ACL, pages 840\u2013847, Prague."", ""year"": 2007}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""A fast decoder for joint word segmentation and POS-tagging using a single discriminative model"", ""venue"": ""Proceedings of EMNLP, pages 843\u2013852, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Zhao"", ""Hai"", ""Chunyu Kit.""], ""title"": ""Unsupervised segmentation helps supervised learning of character tagging for word segmentation and named entity recognition"", ""venue"": ""Proceedings of IJCNLP,"", ""year"": 2008}, {""authors"": [""Zhu"", ""Muhua"", ""Jingbo Zhu"", ""Minghan Hu.""], ""title"": ""Better automatic treebank conversion using a feature-based approach"", ""venue"": ""Proceedings of ACL, volume 2, pages 715\u2013719, Portland, OR."", ""year"": 2011}]","acknowledgments  :Jiang, Lü, and Liu were supported by National Natural Science Foundation of China (contract 61202216) and the National Key Technology R&D Program (no. 2012BAH39B03). Huang was supported in part by the DARPA DEFT Project (FA8750-13-2-0041). Liu was partially supported by the Science Foundation Ireland (grant no. 07/CE/I1142) as part of the CNGL at Dublin City University. We also thank the anonymous reviewers for their insightful comments. Finally, we want to thank Chris Hokamp for proofreading.",,,,,,,,,,,,,,,,,,,,,"automatic adaptation of annotations :Wenbin Jiang∗
Chinese Academy of Sciences
Yajuan Lü∗
Chinese Academy of Sciences
Liang Huang∗∗
Queens College and Graduate Center,
The City University of New York
Qun Liu∗†
Dublin City University
Chinese Academy of Sciences
Manually annotated corpora are indispensable resources, yet for many annotation tasks, such as
the creation of treebanks, there exist multiple corpora with different and incompatible annotation
guidelines. This leads to an inefficient use of human expertise, but it could be remedied by
integrating knowledge across corpora with different annotation guidelines. In this article we
describe the problem of annotation adaptation and the intrinsic principles of the solutions, and
present a series of successively enhanced models that can automatically adapt the divergence
between different annotation formats.
We evaluate our algorithms on the tasks of Chinese word segmentation and dependency
parsing. For word segmentation, where there are no universal segmentation guidelines be-
cause of the lack of morphology in Chinese, we perform annotation adaptation from the much
larger People’s Daily corpus to the smaller but more popular Penn Chinese Treebank. For
dependency parsing, we perform annotation adaptation from the Penn Chinese Treebank to
a semantics-oriented Dependency Treebank, which is annotated using significantly different
annotation guidelines. In both experiments, automatic annotation adaptation brings significant
improvement, achieving state-of-the-art performance despite the use of purely local features in
training.
∗ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, No. 6 Kexueyuan South Road, Haidian District, P.O. Box 2704, Beijing 100190, China. E-mail: {jiangwenbin, liuqun, lvyajuan}@ict.ac.cn.
∗∗ Department of Computer Science, Queens College / CUNY, 65-30 Kissena Blvd., Queens, NY 11367. E-mail: liang.huang.sh@gmail.com. † Centre for Next Generation Localisation, Faculty of Engineering and Computing, Dublin City University.
E-mail: qliu@computing.dcu.ie.
Submission received: 24 April 2013; revised version received: 6 March 2014; accepted for publication: 18 April 2014.
doi:10.1162/COLI a 00210
© 2015 Association for Computational Linguistics",type templates instances :Guiding α α=B α ◦ C−2 α=B ◦ C−2=美 α ◦ C−1 α=B ◦ C−1=副 α ◦ C0 α=B ◦ C0=总 α ◦ C1 α=B ◦ C1=统 α ◦ C2 α=B ◦ C2=访 α ◦ C−2C−1 α=B ◦ C−2C−1=美副 α ◦ C−1C0 α=B ◦ C−1C0=副总 α ◦ C0C1 α=B ◦ C0C1=总统 α ◦ C1C2 α=B ◦ C1C2=统访 α ◦ C−1C1 α=B ◦ C−1C1=副统 α ◦ Pu(C0) α=B ◦ Pu(C0)=true α ◦ T(C−2:2) α=B ◦ T(C−2:2)= 44444,"partition sections # of words : training iterations. The performance measurement indicator for word segmentation is the balanced F-measure, F = 2PR/(P + R), a function of Precision P and Recall R, where P is the percentage of words in segmentation results that are segmented correctly, and R is the percentage of correctly segmented words in the gold standard words.
For both syntactic and semantic dependency parsing, we concentrate on nonlabeled parsing that predicts the graphic dependency structure for the input sentence without considering dependency labels. The perceptron-based baseline dependency models are trained on the training sets of DCTB and SDT, using the development sets to determine the best training iterations. The performance measurement indicator for dependency parsing is the Unlabeled Attachment Score, denoted as Precision P, indicating the percentage of words in predicted dependency structure that are correctly attached to their head words.
Figure 6 shows the learning curve of the averaged perceptron for word segmentation on the development set. Accuracies of the baseline classifiers are listed in Table 6. We also report the performance of the classifiers on the testing sets of the opposite corpora. Experimental results are in line with our expectations. A classifier performs better in its corresponding testing set, and performs significantly worse on testing data following a different annotation guideline.
Table 7 shows the accuracies of the baseline syntactic and semantic parsers, as well as the performance of the parsers on the testing sets of the opposite corpora. Similar
to the situations in word segmentation, two parsers give state-of-the-art accuracies on their own testing sets, but perform poorly on the other testing sets. This indicates the degree of divergence between the annotation guidelines of DCTB and SDT.","ctb  :To approximate more general scenarios of annotation adaptation problems, we extract from PD a subset that is comparable to CTB in size. Because there are many extremely long sentences in the original PD corpus, we first split them into normal sentences according to the full-stop punctuation symbol. We randomly select 20, 000 sentences (0.45M words) from the PD training data as the new training set, and 1, 000/1, 000 sentences from the PD test data as the new testing/developing set. We label the smaller version of PD as SPD. The balanced source corpus and target corpus also facilitate the investigation of annotation adaptation.
Annotation adaptation for dependency parsing is performed from the CTB-derived syntactic dependency treebank (DCTB) (Yamada and Matsumoto 2003) to the Semantic Dependency Treebank (SDT) (Che et al. 2012). Semantic dependency encodes the semantic relationships between words, which are very different from syntactic dependencies. SDT is annotated on a small portion of the CTB text as depicted in Table 5; therefore, we use the subset of DCTB covering the remaining CTB text as the source corpus. We still denote the source corpus as DCTB in the following for simplicity.",previous work :,"semeval-2012 contest :model is denoted as Model 3). We find that the predict-self re-estimation brings improvement to the iterative training at each iteration, for both word segmentation and dependency parsing. The maximum performance is achieved at iteration 4 for word segmentation, and at iteration 5 for dependency parsing. The corresponding models are evaluated on the corresponding testing sets, and the experimental results are also shown in Tables 8 and 9. Compared to Model 1, the optimized annotation adaptation strategy, Model 3, leads to classifiers with significantly higher accuracy and to processing speeds that are several times faster. Tables 10 and 11 show the experimental results compared with previous work. For both Chinese word segmentation and semantic dependency parsing, automatic annotation adaptation yields state-of-the-art performance, despite using single classifiers with only local features. Note that for the systems in the SemEval contest (Che et al. 2012), many other technologies including clause segmentation, system combination, and complicated features were adopted, as well as elaborate engineering. We also performed significance tests2 to verify the effectiveness of annotation adaptation.We find that for both Chinese word segmentation and semantic dependency parsing, annotation adaptation brings significant improvement (p < 0.001) over the baselines trained on the target corpora only.
2 http://www.cis.upenn.edu/∼dbikel/download/compare.pl.","word type proportion baseline anno. ada. trend :To evaluate the stability of annotation adaptation, we perform quantitative analysis on the results of annotation adaptation. For word segmentation, the words are grouped according to POS tags. For dependency parsing, the dependency edges are grouped according to POS tag pairs. For each category, the recall values of baseline and annotation adaptation are reported. To filter the lists, we set two significance thresholds with respect to the proportion of a category and the performance fluctuation between two systems. For word segmentation, only the categories with proportions of more than 1% and with fluctuations of more than 0.1 points are reserved, and for dependency parsing, the two thresholds are 1% and 0.5. Tables 12 and 13 show the analysis results for word segmentation and dependency parsing, respectively. For both tasks, annotation adaptation brings improvement for most of the situations.","edge type proportion baseline anno. ada. trend :We further investigate the effect of varying the sizes of the target corpora. Experiments are conducted for word segmentation and dependency parsing with fixed-size source corpora and varying-size target corpora. We use SPD and DCTB as the source corpora for word segmentation and dependency parsing, respectively. Figures 13 and 14 show the performance curves on the testing sets. We find that, for both word segmentation and dependency parsing, the improvements brought by annotation adaptation are more significant when the target corpora are smaller. It means that the automatic annotation adaptation is more valuable when the size of the target corpus is small,
which is good news for the situation where the corpus we are concerned with is smaller but a larger differently annotated corpus exists.
Of course, the comparison between automatic annotation adaptation and previous strategies without using additional training data is unfair. Our work aims to find another way to improve NLP tasks: focusing on the collection of more training data instead of making full use of a certain corpus. We believe that the performance of automatic annotation adaptation can be further improved by adopting the advanced technologies of previous work, such as complicated features and model combination. It would be useful to conduct experiments with more source-annotated training data, such as the SIGHAN data set for word segmentation, to investigate the trend of improvement along with the further increment of annotated sentences. It would also be valuable to evaluate the improved word segmenter and dependency parser on the out-of-domain data sets. However, currently most corpora for word segmentation and dependency parsing do not explicitly distinguish the domains of their data sections, making such evaluations difficult to conduct.",,,,,,,,,,,,,,,,,,,,,"1 introduction :Much of statistical NLP research relies on some sorts of manually annotated corpora to train models, but annotated resources are extremely expensive to build, especially on a large scale. The creation of treebanks is a prime example (Marcus, Santorini, and Marcinkiewicz 1993). However, the linguistic theories motivating these annotation efforts are often heavily debated, and as a result there often exist multiple corpora for the same task with vastly different and incompatible annotation philosophies. For example, there are several treebanks for English, including the Chomskian-style Penn Treebank (Marcus, Santorini, and Marcinkiewicz 1993), the HPSG LinGo Redwoods Treebank (Oepen et al. 2002), and a smaller dependency treebank (Buchholz and Marsi 2006). From the perspective of resource accumulation, it seems a waste in human efforts.1
A second, related problem is that the raw texts are also drawn from different domains, which for the above example range from financial news (Penn Treebank/Wall Street Journal) to transcribed dialog (LinGo). It would be nice if a system could be automatically ported from one set of guidelines and/or domain to another, in order to exploit a much larger data set. The second problem, domain adaptation, is very well studied (e.g., Blitzer, McDonald, & Pereira 2006; Daumé III 2007). This work focuses on the widely existing and equally important problem, annotation adaptation, in order to adapt the divergence between different annotation guidelines and integrate linguistic knowledge in corpora with incongruent annotation formats.
In this article, we describe the problem of annotation adaptation and the intrinsic principles of the solutions, and present a series of successively improved concrete models, the goal being to transfer the annotations of a corpus (source corpus) to the annotation format of another corpus (target corpus). The transfer classifier is the fundamental component for annotation adaptation algorithms. It learns correspondence regularities between annotation guidelines from a parallel annotated corpus, which has two kinds of annotations for the same data. In the simplest model (Model 1), the source classifier trained on the source corpus gives its predications to the transfer classifier trained on the parallel annotated corpus, so as to integrate the knowledge in the two corpora. In a variant of the simplest model (Model 2), the transfer classifier is used to transform the annotations in the source corpus into the annotation format of the target corpus; then the transformed source corpus and the target corpus are merged in order to train a more accurate classifier. Based on the second model, we finally develop an optimized model (Model 3), where two optimization strategies, iterative training and predict-self re-estimation, are integrated to further improve the efficiency of annotation adaptation.
We experiment on Chinese word segmentation and dependency parsing to test the efficacy of our methods. For word segmentation, the problem of incompatible annotation guidelines is one of the most glaring: No segmentation guideline has been widely accepted due to the lack of a clear definition of Chinese word morphology. For dependency parsing there also exist multiple disparate annotation guidelines. For
1 Different annotated corpora for the same task facilitate the comparison of linguistic theories. From this perspective, having multiple standards is not necessarily a waste but rather a blessing, because it is a necessary phase in coming to a consensus if there is one.
example, the dependency relations extracted from a constituency treebank follow syntactic principles, whereas the semantic dependency treebank is annotated in a semantic perspective.
The two corpora for word segmentation are the much larger People’s Daily corpus (PD) (5.86M words) (Yu et al. 2001) and the smaller but more popular Penn Chinese Treebank (CTB) (0.47M words) (Xue et al. 2005). They utilize very different segmentation guidelines; for example, as shown in Figure 1, PD breaks VicePresident into two words and combines the phrase visited-China as a compound, compared with the segmentation following the CTB annotation guideline. It is preferable to transfer knowledge from PD to CTB because the latter also annotates tree structures, which are useful for downstream applications like parsing, summarization, and machine translation, yet it is much smaller in size. For dependency parsing, we use the dependency treebank (DCTB) extracted from CTB according to the rules of Yamada and Matsumoto (2003), and the Semantic Dependency Treebank (SDT) built on a small part of the CTB text (Che et al. 2012). Compared with the automatically extracted dependencies in DCTB, semantic dependencies in SDT reveal semantic relationships between words, rather than the syntactic relationships in syntactic dependencies. Figure 2 shows an example. Experiments on both word segmentation and dependency parsing show that annotation adaptation results in significant improvement over the baselines, and achieves the state-of-the-art with only local features.
The rest of the article is organized as follows. Section 2 gives a description of the problem of annotation adaptation. Section 3 briefly introduces the tasks of word segmentation and dependency parsing as well as their state-of-the-art models. In Section 4 we first describe the transfer classifier that indicates the
intrinsic principles of annotation adaptation, and then depict the three successively enhanced models for automatic adaptation of annotations. After the description of experimental results in Section 5 and the discussion of application scenarios in Section 6, we give a brief review of related work in Section 7, drawing conclusions in Section 8. 2 automatic annotation adaptation :We define annotation adaptation as a task aimed at automatically adapting the divergence between different annotation guidelines. Statistical models can be designed to learn the relevance of two annotation guidelines in order to transform a corpus from one annotation guideline to another. From this point of view, annotation adaptation can be seen as a special case of transfer learning. Through annotation adaptation, the linguistic knowledge in different corpora is integrated, resulting in enhanced NLP systems without complicated models and features.
Much research has considered the problem of domain adaptation (Blitzer, McDonald, and Pereira 2006; Daumé III 2007), which also can be seen as a special case of transfer learning. It aims to adapt models trained in one domain (e.g., chemistry) to work well in other domains (e.g., medicine). Despite superficial similarities between domain adaptation and annotation adaptation, we argue that the underlying problems are quite different. Domain adaptation assumes that the labeling guidelines are preserved between the two domains—for example, an adjective is always labeled as JJ regardless of whether it is from the Wall Street Journal (WSJ) or a biomedical text, and only the distributions are different—for example, the word control is most likely a verb in WSJ but often a noun in biomedical texts (as in control experiment). Annotation adaptation, however, tackles the problem where the guideline itself is changed, for example, one treebank might distinguish between transitive and intransitive verbs, while merging the different noun types (NN, NNS, etc.), or one treebank (PTB) might be much flatter than the other (LinGo), not to mention the fundamental disparities between their underlying linguistic representations (CFG vs. HPSG).
A more formal description will allow us to make these claims more precise. Let X be the data and Y be the annotation. Annotation adaptation can be understood as a change of P(Y) due to the change in annotation guidelines while P(X) remains constant. Through annotation adaptation, we want to change the annotations of the data from one guideline to another, leaving the data itself unchanged. However, in domain adaptation, P(X) changes, but P(Y) is assumed to be constant. The word assumed means that the distributions P(Y, X) and P(Y|X) are actually changed because P(X) is changed. Domain adaptation aims to make the model better adapt to a different domain with the same annotation guidelines.
According to this analysis, annotation adaptation seems more motivated from a linguistic (rather than statistical) point of view, and tackles a serious problem fundamentally different from domain adaptation, which is also a serious problem (often leading to >10% loss in accuracy). More interestingly, annotation adaptation, without any assumptions about distributions, can be simultaneously applied to both domain and annotation adaptation problems, which is very appealing in practice because the latter problem often implies the former. 3 case studies: word segmentation and dependency parsing : In many Asian languages there are no explicit word boundaries, thus word segmentation is a fundamental task for the processing and understanding of these languages. Given a sentence as a sequence of n characters:
x = x1 x2 .. xn
where xi is a character, word segmentation aims to split the sequence into m(≤ n) words:
x1:e1 xe1+1:e2 .. xem−1+1:em
where each subsequence xi:j indicates a Chinese word spanning from characters xi to xj.
Word segmentation can be formalized as a sequence labeling problem (Xue and Shen 2003), where each character in the sentence is given a boundary tag representing its position in a word. Following Ng and Low (2004), joint word segmentation and partof-speech (POS) tagging can also be solved using a character classification approach by extending boundary tags to include POS information. For word segmentation we adopt the four boundary tags of Ng and Low (2004), B, M, E, and S, where B, M, and E mean the beginning, the middle, and the end of a word, respectively, and S indicates a singlecharacter word. The word segmentation result can be generated by splitting the labeled character sequence into subsequences of pattern S or BM∗E, indicating single-character words or multi-character words, respectively.
Given the character sequence x, the decoder finds the output ỹ that maximizes the score function:
ỹ = argmax y f(x, y) ·w
= argmax y
∑
xi∈x,yi∈y
f(xi, yi) ·w (1)
Where function f maps (x, y) into a feature vector, w is the parameter vector generated by the training algorithm, and f(x, y) ·w is the inner product of f(x, y) and w. The score of the sentence is further factorized into each character, where yi is the character classification label of character xi.
The training procedure of perceptron learns a discriminative model mapping from the inputs x to the outputs y. Algorithm 1 shows the perceptron algorithm for tuning the parameter w. The “averaged parameters” technology (Collins 2002) is used for better performance. The feature templates of the classifier is shown in Table 1. The function Pu(·) returns true for a punctuation character and false for others; the function T(·) classifies a character into four types: number, date, English letter, and others, corresponding to function values 1, 2, 3, and 4, respectively. Dependency parsing aims to link each word to its arguments so as to form a directed graph spanning the whole sentence. Normally, the directed graph is restricted to a
Algorithm 1 Perceptron training algorithm.
1: Input: Training set C 2: w← 0 3: for t← 1 .. T do ⊲ T iterations 4: for (x, y) ∈ C do 5: z̃← argmaxz f(x, z) ·w 6: if z̃ 6= y then 7: w← w + f(x, y)− f(x, z̃) ⊲ update the parameters 8: end if 9: end for
10: end for 11: Output: Parameters w
dependency tree where each word depends on exactly one parent, and all words find their parents. Given a sentence as a sequence n words:
x = x1 x2 .. xn
dependency parsing finds a dependency tree y, where (i, j) ∈ y is an edge from the head word xi to the modifier word xj. The root r ∈ x in the tree y has no head word, and each of the other words, j(j ∈ x and j 6= r), depends on a head word i(i ∈ x and i 6= j).
For many languages, the dependency structures are supposed to be projective. If xj is dependent on xi, then all the words between i and j must be directly or indirectly dependent on xi. Therefore, if we put the words in their linear order, preceded by the root, all edges can be drawn above the words without crossing. We follow this constraint because the dependency treebanks in our experiments are projective.
Following the edge-based factorization method (Eisner 1996), the score of a dependency tree can be factorized into the dependency edges in the tree. The spanning
tree method (McDonald, Crammer, and Pereira 2005) factorizes the score of the tree as the sum of the scores of all its edges, and the score of an edge is defined as the inner product of the feature vector f and the weight vector w. Given a sentence x, the parsing procedure searches for the candidate dependency tree with the maximum score:
ỹ = argmax y f(x, y) ·w
= argmax y
∑
(i,j)∈y
f(i, j) ·w (2)
The averaged perceptron algorithm is used again to train the parameter vector. A bottom–up dynamic programming algorithm is designed to search for the candidate parse with the maximum score as shown in Algorithm 2, where V[i, j] contains the candidate dependency fragments of the span [i, j]. The feature templates are similar to those of the first-ordered MST model (McDonald, Crammer, and Pereira 2005). Each feature is composed of some words and POS tags surround word i and/or word j, as well as an optional distance representation between the two words. Table 2 shows the feature templates without distance representations. 4 models for automatic annotation adaptation :In this section, we present a series of discriminative learning algorithms for the automatic adaptation of annotation guidelines. To facilitate the description of the algorithms, several shortened forms are adopted for convenience of description. We use source corpus to denote the corpus with the annotation guideline that we do not require, which is of course the source side of adaptation, and target corpus denotes the corpus with the desired guideline. Correspondingly, the annotation guidelines of the two corpora are denoted as source guideline and target guideline, and the classifiers following the two annotation guidelines are respectively named as source classifier and target classifier. Given a parallel annotated corpus, that is, a corpus labeled with two annotation
Algorithm 2 Dependency parsing algorithm.
1: Input: sentence x to be parsed 2: for [i, j] ⊆ [1, |x|] in topological order do 3: buf← ∅ 4: for k← i..j− 1 do ⊲ all partitions 5: for l ∈ V[i, k] and r ∈ V[k + 1, j] do 6: insert DERIV(l, r) into buf 7: insert DERIV(r, l) into buf 8: end for 9: end for 10: V[i, j]← best K in buf 11: end for 12: Output: the best of V[1, |x|] 13: function DERIV(p, c) 14: return p ∪ c ∪ {(p · root, c · root)} ⊲ new derivation 15: end function
guidelines, a transfer classifier can be trained to capture the regularity of transformation from the source annotation to the target annotation. The classifiers mentioned here are normal discriminative classifiers that take a set of features as input and give a classification label as output. For the POS tagging problem, the classification label is a POS tag, and for the parsing task, the classification label is a dependency edge, a constituency span, or a shift-reduce action.
The parallel annotated corpus is the knowledge source of annotation adaptation. The annotation quality and data size of the parallel annotated corpus determine the accuracy of the transfer classifier. Such a corpus is difficult to build manually, although we can generate a noisy one automatically from the source corpus and the target corpus. For example, we can apply the source classifier on the target corpus, thus generating
a parallel annotated corpus with noisy source annotations and accurate target annotations. The training procedure of the transfer classifier predicts the target annotations with guiding features extracted from the source annotations. This approach can alleviate the effect of the noise in source annotations, and learn annotation adaptation regularities accurately. By reducing the noise in the automatically generated parallel annotated corpus, a higher accuracy of annotation adaptation can be achieved.
In the following sections, we first describe the transfer classifier that reveals the intrinsic principles of annotation adaptation, then describe a series of successively enhanced models that are developed from our previous investigation (Jiang, Huang, and Liu 2009; Jiang et al. 2012). In the simplest model (Model 1), two classifiers, a source classifier and a transfer classifier, are used in a cascade. The classification results of the lower source classifier provide additional guiding features to the upper transfer classifier, yielding an improved classification result. A variant of the first model (Model 2) uses the transfer classifier to transform the source corpus from source guideline to target guideline first, then merges the transformed source corpus into the target corpus in order to train an improved target classifier on the enlarged corpus. An optimized model (Model 3) is further proposed based on Model 2. Two optimization strategies, iterative training and predict-self re-estimation, are adopted to improve the efficiency of annotation adaptation, in order to fully utilize the knowledge in heterogeneous corpora. In order to learn the regularity of the adaptation from one annotation guideline to another, a parallel annotated corpus is needed to train the transfer classifier. The parallel annotated corpus is a corpus with two different annotation guidelines, the source guideline and the target guideline. With the target annotation labels as learning objectives and the source annotation labels as guiding information, the transfer classifier learns the statistical regularity of the adaptation from the source annotations to the target annotations.
The training procedure of the transfer classifier is analogous to the training of a normal classifier except for the introduction of additional guiding features. For word segmentation, the most intuitive guiding feature is the source annotation label itself. For dependency parsing, an effective guiding feature is the dependency path between the hypothetic head and modifier, as shown in Figure 3. However, our effort is not limited to this, and more special features are introduced: A classification label or dependency path is attached to each feature of the baseline classifier to generate combined guiding features. This is similar to the feature design in discriminative dependency parsing (McDonald, Crammer, and Pereira 2005; McDonald and Pereira 2006), where the basic features, composed of words and POSs in the context, are also conjoined with link direction and distance in order to generate more special features.
Table 3 shows an example of guide features (as well as baseline features) for word segmentation, where “α = B” indicates that the source classification label of the current character is B, demarcating the beginning of a word. The combination strategy derives a series of specific features, helping the transfer classifier to produce more precise classifications. The parameter-tuning procedure of the transfer classifier will automatically learn the regularity of using the source annotations to guide the classification decision. In decoding, if the current character shares some basic features in Table 3 and it is classified as B in the source annotation, then the transfer classifier will probably classify it as M.
In addition, the original features used in the normal classifier are also used in order to leverage the knowledge from the target annotation of the parallel annotated corpus, and the training procedure of the transfer classifier also learns the relative weights between the guiding features and the original features. Therefore, the knowledge from both the source annotation and the target annotation are automatically integrated together, and higher and more stable prediction accuracy can be achieved. The most intuitive model for annotation adaptation uses two cascaded classifiers, the source classifier and the transfer classifier, to integrate the knowledge in corpora with different annotation guidelines. In the training procedure, a source classifier is trained on the source corpus and is used to process the target corpus, generating a parallel annotated corpus (albeit a noisy one). Then, the transfer classifier is trained on the parallel annotated corpus,with the target annotations as the classification labels, and the source annotation as guiding information. Figure 4 depicts the training pipeline. The best training iterations for the source classifier and the transfer classifier are determined on the development sets of the source corpus and target corpus.
In the decoding procedure, a sequence of characters (for word segmentation) or words (for dependency parsing) is input into the source classifier to obtain a classification result under the source guideline; then it is input into the transfer classifier with this classification result as the guiding information to get the final result following the target guideline. This coincides with the stacking method for combining dependency parsers (Martins et al. 2008; Nivre and McDonald 2008), and is also similar to the Pred baseline for domain adaptation in Daumé et al. (Daumé III and Marcu 2006; Daumé III 2007). Figure 5 shows the pipeline for decoding. The previous model has a drawback: It has to cascade two classifiers in decoding to integrate the knowledge in two corpora, which seriously degrades the processing speed. Here we describe a variant of the previous model, aiming at automatic transformation (rather than integration as in Model 1) between annotation guidelines of humanannotated corpora. The source classifier and the transfer classifier are trained in the same way as in the previous model. The transfer classifier is used to process the source corpus, with the source annotations as guiding information, so as to relabel the source corpus with the target annotation guideline. By merging the target corpus and the transformed source corpus for the training of the final classifier, improved classification accuracy can be achieved.
From this point on we describe the pipelines of annotation transformation in pseudo-codes for simplicity and convenience of extensions. Algorithm 3 shows the overall training algorithm for the variant model. Cs and Ct denote the source corpus and the target corpus. Ms and Ms→t denote the source classifier and the transfer classifier. C q p denotes the p corpus relabeled in q annotation guideline; for example, C t s is a corpus that labels the text of the source corpus with the target guideline. Functions TRAIN and TRANSTRAIN train the source classifier and the transfer classifier, respectively, both invoking the perceptron algorithm, but with different feature sets. Functions ANNOTATE and TRANSANNOTATE call the function DECODE with different models (source/transfer classifiers), feature functions (without/with guiding features), and inputs (raw/source-annotated sentences).
In the algorithm the parameters corresponding to development sets are omitted for simplicity. Compared to the online knowledge integration methodology of the previous model, annotation transformation leads to improved performance in an offline manner by integrating corpora before the training procedure. This approach enables processing speeds several times faster than the cascaded classifiers in the previous model. It also has another advantage in that we can integrate the knowledge in more than two corpora without slowing down the processing of the final classifier. The training of the transfer classifier is based on an automatically generated (rather than a gold standard) parallel annotated corpus, where the source annotations are
Algorithm 3 Baseline annotation adaptation.
1: function ANNOTRANS(Cs, Ct) 2: Ms ← TRAIN(Cs) ⊲ source classifier 3: Cst ← ANNOTATE(Ms, Ct) 4: Ms→t ← TRANSTRAIN(C s t , Ct) ⊲ transfer classifier 5: Cts ← TRANSANNOTATE(Ms→t, Cs) 6: Ct∗ ← C t s ∪ Ct ⊲ integrated corpus with target guideline 7: return Ct∗ 8: end function 9: function DECODE(M, Φ, x)
10: return argmaxy∈GEN(x) S(y|M,Φ, x) 11: end function
provided by the source classifier. Therefore, the performance of annotation transformation is correspondingly determined by the accuracy of the source classifier, and we can generate a more accurate parallel annotated corpus for better annotation adaptation if an improved source classifier can be obtained. Based on Model 2, two optimization strategies—iterative bidirectional training and predict-self hypothesis—are introduced to optimize the parallel annotated corpora for better annotation adaptation.
We first use an iterative training procedure to gradually improve the transformation accuracy by iteratively optimizing the parallel annotated corpora. In each training iteration, both source-to-target and target-to-source annotation transformations are performed, and the transformed corpora are used to provide better annotations for the parallel annotated corpora of the next iteration. Then in the new iteration, the better parallel annotated corpora will result in more accurate transfer classifiers, resulting in the generation of better transformed corpora.
Algorithm 4 shows the overall procedure of the iterative training method. The loop of lines 6–13 iteratively performs source-to-target and target-to-source annotation transformations. The source annotations of the parallelly annotated corpora, Cst and Cts, are initialized by applying the source and target classifiers on the target and source corpora, respectively (lines 2–5). In each training iteration, the transfer classifiers are trained on the current parallel annotated corpora (lines 7–8); they are used to produce the transformed corpora (lines 9–10), which provide better annotations for the parallel annotated corpora of the next iteration. The iterative training terminates when the performance of the classifier trained on the merged corpus Cts ∪ Ct converges (line 13).
The discriminative training of TRANSTRAIN predicts the target annotations with the guidance of source annotations. In the first iteration, the transformed corpora generated by the transfer classifiers are better than the initial ones generated by the source and target classifiers, due to the assistance of the guiding features. In the following iterations,
Algorithm 4 Iterative annotation transformation.
1: function ITERANNOTRANS(Cs, Ct) 2: Ms ← TRAIN(Cs) ⊲ source classifier 3: Cst ← ANNOTATE(Ms, Ct) 4: Mt ← TRAIN(Ct) ⊲ target classifier 5: Cts ← ANNOTATE(Mt, Cs) 6: repeat 7: Ms→t ← TRANSTRAIN(C s t , Ct) ⊲ source-to-target transfer classifier 8: Mt→s ← TRANSTRAIN(C t s, Cs) ⊲ target-to-source transfer classifier
9: Cts ← TRANSANNOTATE(Ms→t, Cs) 10: Cst ← TRANSANNOTATE(Mt→s, Ct) 11: Ct∗ ← C t s ∪ Ct 12: M∗ ← TRAIN(C t ∗) ⊲ enhanced classifier trained on merged corpus 13: until EVAL(M∗) converges 14: return Ct∗ 15: end function 16: function DECODE(M, Φ, x) 17: return argmaxy∈GEN(x) S(y|M,Φ, x) 18: end function
the transformed corpora provide better annotations for the parallel annotated corpora of the subsequent iteration; the transformation accuracy will improve gradually along with the optimization of the parallel annotated corpora until convergence.
The predict-self hypothesis is introduced to improve the transformation accuracy from another perspective. This hypothesis is implicit in many unsupervised learning approaches, such as Markov random field; it has also been successfully used by Daumé III (2009) in unsupervised dependency parsing. The basic idea of predict-self is, if a prediction is a better candidate for an input, it would be easier to convert it back to the original input by a reverse procedure. If applied to annotation transformation, predictself indicates that a better transformation candidate following the target guideline can be more easily transformed back to the original form following the source guideline.
The most intuitive strategy to introduce the predict-self methodology into annotation transformation is using a reversed annotation transformation procedure to filter out unreliable predictions of the previous transformation. In detail, a source-to-target annotation transformation procedure is performed on the source corpus to obtain a prediction that follows the target guideline; then a second, target-to-source transformation procedure is performed on this prediction result to check whether it can be transformed back to the original source annotation. The source-to-target prediction results that fail in this reverse-verification step are discarded, so this strategy can be called predict-self filtering.
A more sophisticated strategy can be called predict-self re-estimation. Instead of using the reversed transformation procedure for filtering, the re-estimation strategy integrates the scores given by the source-to-target and target-to-source annotation transformation models when evaluating the transformation candidates. By properly tuning the relative weights of the two transformation directions, better transformation performance is achieved. The scores of the two transformation models are weighted, integrated in a log-linear manner:
S+(y|Ms→t,Mt→s,Φ, x) = (1− λ)× S(y|Ms→t,Φ, x) + λ× S(x|Mt→s,Φ, y) (3)
The weight parameter λ is tuned on the development set. To integrate the predictself reestimation into the iterative transformation training, a reversed transformation model is introduced and the enhanced scoring function is used when the function TRANSANNOTATE invokes the function DECODE. 5 experiments :To evaluate the performance of annotation adaptation, we experiment on two important NLP tasks, Chinese word segmentation and dependency parsing, both of which can be modeled as discriminative classification problems. For both tasks, we give the performances of the baseline models and the annotation adaptation algorithms. We perform annotation adaptation for word segmentation from People’s Daily (PD) (Yu et al. 2001) to Penn Chinese Treebank 5.0 (CTB) (Xue et al. 2005). The two corpora are built according to different segmentation guidelines and differ largely in quantity of data. CTB is smaller in size with about 0.5M words, whereas PD is much larger, containing nearly 6M words. Table 4 shows the data partitioning for the two corpora. We train the baseline perceptron classifiers for Chinese word segmentation on the training sets of CTB and SPD, using corresponding development sets to determine the best As a variant of Model 1, Model 2 shares the same transfer classifier, and differs only in training and decoding of the final classifier. Tables 8 and 9 show the performances of systems resulting from Models 1 and 2, as well as the classifiers trained on the directly merged corpora. The time costs for decoding are also listed to facilitate the practical comparison.
We find that the simple corpus merging strategy leads to a dramatic decrease in accuracy, due to the different and incompatible annotation guidelines. Model 1, the simplest model for annotation adaptation, gives significant improvement over the baseline classifiers for word segmentation and dependency parsing. This indicates that the statistical regularities for annotation adaptation learned by the transfer classifiers
bring performance improvement, utilizing guided decisions in the cascaded classifiers. Model 2 leads to classifiers with accuracy increments comparable to those of Model 1, while consuming only one third of the decoding time. It is inconsistent with our expectation. The strategy of directly transforming the source corpus to the target guideline also facilitates the utilization of more than one source corpus.
We first introduce the iterative training strategy to Model 2. The corresponding development sets are used to determine the best training iterations for the iterative annotation transformations. After each iteration, we test the performance of the classifiers trained on the merged corpora. Figures 7 and 8 show the performance curves for Chinese word segmentation and semantic dependency parsing, respectively, with iterations ranging from 1 to 10. The performance of Model 2 is naturally included
in the curves (located at iteration 1). The curves show that, for both segmentation and parsing, the accuracies of the classifiers trained on the merged corpora consistently improve in the earlier iterations (e.g., from iteration 2 to iteration 5 for word segmentation).
Experiments for introduction of predict-self filtering and predict-self re-estimation are shown in Figures 9 and 10. The curves show the performances of the predictself re-estimation with a series of weight parameters, ranging from 0 to 1 with step 0.05. Note that in both figures, the points at λ = 0 show the performances of Model 2. We find that predict-self filtering brings a slight improvement over the baseline for
word segmentation, but even decreases the accuracy for dependency parsing. An initial analysis on the experimental results reveals that the filtering strategy discards some complicated sentences in the source corpora, and the discarded sentences would bring further improvement if properly used. For example, in word segmentation, predict-self filtering discards 5% of sentences from the source corpus, containing nearly 10% of the training words. For the two tasks, the predict-self re-estimation outperforms the filtering strategy. With properly tuned weights, predict-self re-estimation can make better use of the training data. The largest accuracy improvements achieved over Model 2 for word segmentation and dependency parsing are 0.3 points and 0.6 points.
Figures 11 and 12 show the performance curves after the introduction of both iterative training and predict-self re-estimation on the basis of Model 2 (this enhanced 6 discussion: application situations :Automatic annotation adaptation aims to transform the annotations in a corpus to the annotations following other guidelines. The models for annotation adaptation use a transfer classifier to learn the statistical correspondence regularities between different annotation guidelines. These statistical regularities are learned from a parallel annotated corpus, which does not need to be manually annotated. In fact, the models for annotation adaptation train the transfer classifier on an automatically generated parallel annotated corpus, which is generated by processing a corpus with a classifier trained on another corpus. That is to say, if we want to conduct annotation adaptation across several corpora, no additional corpora need to be manually annotated. This setting makes the strategy of annotation adaptation more general, because it is much harder to manually annotate a parallel annotated corpus, regardless of the language or the NLP problem under consideration. To tackle the problem of noise in automatically generated annotations, the advanced models we designed generate a better parallel annotated corpus by making use of strategies such as iterative optimization.
Automatic annotation adaptation can be applied in any situation where we have multiple corpora with different and incompatible annotation philosophies for the same task. As our case studies, both Chinese word segmentation and dependency parsing have more than one corpora with different annotation guidelines, such as the People’s Daily and the Penn Chinese Treebank for Chinese word segmentation. In a more abstract view, constituency grammar and dependency grammar can be treated as two annotation guidelines for parsing. The syntactic knowledge in a constituency treebank and a dependency treebank, therefore, can be integrated by automatic annotation adaptation. For example, the LinGo Redwoods Treebank can also be transformed to the annotation guideline of the Semantic Dependency Treebank.
Furthermore, the annotations (such as a grammar) given by bilingual projection or unsupervised induction can be seen as following a special annotation philosophy. For bilingually projected annotations, the annotation guideline would be similar to that of the counterpart language. For unsupervised induced annotations, the annotation guideline reflects the statistical structural distribution of a specific data set. In both situations, the underlying annotation guidelines may be largely different from that of the testing sets, which usually come from human-annotated corpora. The system trained on a bilingually projected or unsupervised induced corpus may perform poorly on an existing
testing set, but if the projected or induced corpus has high inner consistency, it could improve a system trained on an existing corpus by automatic annotation adaptation. In this point of view, the practical value of the current work on bilingual projection and unsupervised induction may be underestimated, and annotation adaptation could make better use of the projected or induced knowledge.3 7 related work :There has already been some preliminary work tackling the divergence between different annotation guidelines. Gao et al. (2004) described a transformation-based converter to transfer a certain word segmentation result to another annotation guideline. They designed class-type transformation templates and used the transformation-based error-driven learning method of Brill (1995) to learn what word delimiters should be modified. Many efforts have been devoted to manual treebank transformation, where PTB is adapted to other grammar formalisms, such as CCG and LFG (Cahill et al. 2002; Hockenmaier and Steedman 2007). However, all these are heuristic-based—that is, they need manually designed transformation templates and involve heavy human engineering. Such strategies are hard to be generalized to POS tagging, not to mention other complicated structural prediction tasks.
We investigated the automatic integration of word segmentation knowledge in differently annotated corpora (Jiang, Huang, and Liu 2009; Jiang et al. 2012), which can be seen as the preliminary work of automatic annotation adaptation. Motivated by our initial investigation, researchers applied similar methodologies to constituency parsing (Sun, Wang, and Zhang 2010; Zhu, Zhu, and Hu 2011) and word segmentation (Sun and Wan 2012). This previous work verified the effectiveness of automatic annotation adaptation, but did not reveal the essential definition of the problem nor the intrinsic principles of the solutions. Instead, this work clearly defines the problem of annotation adaptation, reveals the intrinsic principles of the solutions, and systematically describes a series of gradually improved models. The most advanced model learns transformation regularities much better and achieves significant higher accuracy for both word segmentation and dependency parsing, without slowing down the final language processors.
The problem of automatic annotation adaptation can be seen as a special case of transfer learning (Pan and Yang 2010), where the source and target tasks are similar, but not identical. More specifically, the problem related to annotation adaptation assumes that the labeling mechanism across the source and target tasks are the same, but the predictive functions are different. The goal of annotation adaptation is to adapt the source predictive function to be used for the target task by exploiting the labeled data of the source task and the target task. Furthermore, automatic annotation adaptation approximately falls into the spectrum of relational-knowledge-transfer problems (Mihalkova, Huynh, and Mooney 2007; Mihalkova and Mooney 2008; Davis and Domingos 2009), but it tackles problems where the relations among data between the source and target domains can be largely isomerous—or, in other words, with different and incompatible annotation schemes. This work enriches the research of transfer learning by proposing and solving an NLP problem different from previous situations. For more details of transfer learning please refer to the survey of Pan and Yang (2010).
3 We have performed preliminary experiments on word segmentation. Bilingual projection was conducted from English to Chinese with the Chinese–English FBIS as the bilingual corpus. By annotation adaptation, the projected corpus for word segmentation brings a significant F-measure increment of nearly 0.6 points over the baseline trained on CTB only.
The training procedure for an annotation adaptation model requires a parallel annotated corpus (which may be automatically generated); this fact puts the method into the neighborhood of the family of approaches known as annotation projection (Hwa et al. 2002, 2005; Ganchev, Gillenwater, and Taskar 2009; Smith and Eisner 2009; Jiang and Liu 2010; Das and Petrov 2011). Essentially, annotation adaptation and annotation projection tackle different problems; the former aims to transform the annotations from one guideline to another (of course in the same language), whereas the latter aims to project the annotation (as well as the annotation guideline) from one language to another. Therefore, the machine learning methods for annotation adaptation pay attention to automatic transformation of annotations, while for annotation projection, the machine learning methods focus on the bilingual projection across languages.
Co-training (Sarkar 2001) and classifier combination (Nivre and McDonald 2008) are two techniques for training improved dependency parsers. The co-training technique lets two different parsing models learn from each other during the parsing of unlabeled text: One model selects some unlabeled sentences it can confidently parse, and provides them to the other model as additional training data in order to train more powerful parsers. The classifier combination lets graph-based and transition-based dependency parsers utilize the features extracted from each other’s parsing results, to obtain combined, enhanced parsers. The two techniques aim to let two models learn from each other on the same corpus with the same distribution and annotation guideline, whereas our strategy aims to integrate the knowledge in multiple corpora with different annotation guidelines.
The iterative training procedure used in the optimized model shares some similarities with the co-training algorithm in parsing (Sarkar 2001), where the training procedure lets two different models learn from each other during parsing of the raw text. The key idea of co-training is to utilize the complementarity of different parsing models to mine additional training data from raw text, whereas iterative training for annotation adaptation emphasizes the iterative optimization of the parallel annotated corpora used to train the transfer classifiers. The predict-self methodology is implicit in many unsupervised learning approaches; it has been successfully used in unsupervised dependency parsing (Daumé III 2009). We adapt this idea to the scenario of annotation adaptation to improve transformation accuracy.
In recent years much effort has been devoted to the improvement of word segmentation and dependency parsing. For example, the introduction of global training or complicated features (Zhang and Clark 2007, 2010); the investigation of word structures (Li 2011); the strategies of hybrid, joint, or stacked modeling (Nakagawa and Uchimoto 2007; Kruengkrai et al. 2009; Wang, Zong, and Su 2010; Sun 2011); and the semi-supervised and unsupervised technologies utilizing raw text (Zhao and Kit 2008; Johnson and Goldwater 2009; Mochihashi, Yamada, and Ueda 2009; Hewlett and Cohen 2011). We believe that the annotation adaptation technologies can be adopted jointly with complicated features, system combination, and semi-supervised/unsupervised technologies to further improve the performance of word segmentation and dependency parsing. 8 conclusion and future work :We have described the problem of annotation adaptation and the intrinsic principles of its solutions, and proposed a series of successively enhanced models that can automatically adapt the divergence between different annotation formats. These models learn the statistical regularities of adaptation between different annotation guidelines,
and integrate the knowledge in corpora with different annotation guidelines. In the problems of Chinese word segmentation and semantic dependency parsing, annotation adaptation algorithms bring significant improvements by integrating the knowledge in differently annotated corpora, People’s Daily and Penn Chinese Treebank for word segmentation, and Penn Chinese Treebank and Semantic Dependency Parsing for dependency parsing. For both tasks, annotation adaptation leads to a segmenter and a parser achieving the state-of-the-art, despite using only local features in single classifiers.
Many aspects related to annotation adaptation deserve further investigation in the future. First, models for annotation adaptation can be adapted to other NLP tasks such as semantic analysis. Second, jointly tackling the divergences in both annotations and domains is an important problem. In addition, an unsupervised-induced or bilingually projected corpus, despite performing poorly on the specified testing data, may have high inner annotation consistency. That is to say, the induced corpora can be treated as a knowledge source following another annotation guideline, and the performance of current unsupervised or bilingually projected models may be seriously underestimated. Annotation adaptation may give us a new perspective on knowledge induction and measurement for such methods. Manually annotated corpora are indispensable resources, yet for many annotation tasks, such as the creation of treebanks, there exist multiple corpora with different and incompatible annotation guidelines. This leads to an inefficient use of human expertise, but it could be remedied by integrating knowledge across corpora with different annotation guidelines. In this article we describe the problem of annotation adaptation and the intrinsic principles of the solutions, and present a series of successively enhanced models that can automatically adapt the divergence between different annotation formats. We evaluate our algorithms on the tasks of Chinese word segmentation and dependency parsing. For word segmentation, where there are no universal segmentation guidelines because of the lack of morphology in Chinese, we perform annotation adaptation from the much larger People’s Daily corpus to the smaller but more popular Penn Chinese Treebank. For dependency parsing, we perform annotation adaptation from the Penn Chinese Treebank to a semantics-oriented Dependency Treebank, which is annotated using significantly different annotation guidelines. In both experiments, automatic annotation adaptation brings significant improvement, achieving state-of-the-art performance despite the use of purely local features in training. [{""affiliations"": [], ""name"": ""Wenbin Jiang""}, {""affiliations"": [], ""name"": ""Yajuan L\u00fc""}, {""affiliations"": [], ""name"": ""Liang Huang""}, {""affiliations"": [], ""name"": ""Qun Liu""}] SP:376fabb797ac14f81d1ea7f54ed5432a0a7bf244 [{""authors"": [""Blitzer"", ""John"", ""Ryan McDonald"", ""Fernando Pereira.""], ""title"": ""Domain adaptation with structural correspondence learning"", ""venue"": ""Proceedings of EMNLP, pages 120\u2013128, Sydney."", ""year"": 2006}, {""authors"": [""Brill"", ""Eric.""], ""title"": ""Transformation-based error-driven learning and natural language processing: A case study in part-of-speech tagging"", ""venue"": ""Computational Linguistics, 21(4):543\u2013565."", ""year"": 1995}, {""authors"": [""Buchholz"", ""Sabine"", ""Erwin Marsi.""], ""title"": ""CONLL-X shared task on multilingual dependency parsing"", ""venue"": ""Proceedings of CoNLL, pages 149\u2013164, New York, NY."", ""year"": 2006}, {""authors"": [""Cahill"", ""Aoife"", ""Mairead McCarthy"", ""Josef van Genabith"", ""Andy Way.""], ""title"": ""Automatic annotation of the Penn treebank with LFG F-structure information"", ""venue"": ""Proceedings of the LREC Workshop, Las Palmas."", ""year"": 2002}, {""authors"": [""Che"", ""Wanxiang"", ""Meishan Zhang"", ""Yanqiu Shao"", ""Ting Liu.""], ""title"": ""Semeval-2012 task 5: Chinese semantic dependency parsing"", ""venue"": ""Proceedings of SemEval, pages 378\u2013384, Montreal."", ""year"": 2012}, {""authors"": [""Collins"", ""Michael.""], ""title"": ""Discriminative training methods for hidden Markov models: Theory and experiments with perceptron algorithms"", ""venue"": ""Proceedings of EMNLP, pages 1\u20138, Philadelphia, PA."", ""year"": 2002}, {""authors"": [""Das"", ""Dipanjan"", ""Slav Petrov.""], ""title"": ""Unsupervised part-of-speech tagging with bilingual graph-based projections"", ""venue"": ""Proceedings of ACL, pages 600\u2013609, Portland, OR."", ""year"": 2011}, {""authors"": [""Daum\u00e9 III"", ""Hal.""], ""title"": ""Frustratingly easy domain adaptation"", ""venue"": ""Proceedings of ACL, pages 256\u2013263, Prague."", ""year"": 2007}, {""authors"": [""Daum\u00e9 III"", ""Hal.""], ""title"": ""Unsupervised searchbased structured prediction"", ""venue"": ""Proceedings of ICML, pages 209\u2013216, Montreal."", ""year"": 2009}, {""authors"": [""Daum\u00e9 III"", ""Hal"", ""Daniel Marcu.""], ""title"": ""Domain adaptation for statistical classifiers"", ""venue"": ""Journal of Artificial Intelligence Research, 26:101\u2013126."", ""year"": 2006}, {""authors"": [""Davis"", ""Jesse"", ""Pedro Domingos.""], ""title"": ""Deep transfer via second-order Markov logic"", ""venue"": ""Proceedings of ICML, pages 217\u2013224, Montreal."", ""year"": 2009}, {""authors"": [""Eisner"", ""Jason M.""], ""title"": ""Three new probabilistic models for dependency parsing: An exploration"", ""venue"": ""Proceedings of COLING, pages 340\u2013345, Copenhagen."", ""year"": 1996}, {""authors"": [""Ganchev"", ""Kuzman"", ""Jennifer Gillenwater"", ""Ben Taskar.""], ""title"": ""Dependency grammar induction via bitext projection constraints"", ""venue"": ""Proceedings of ACL, pages 369\u2013377, Singapore."", ""year"": 2009}, {""authors"": [""Gao"", ""Jianfeng"", ""Andi Wu"", ""Mu Li"", ""Chang-Ning Huang"", ""Hongqiao Li"", ""Xinsong Xia"", ""Haowei Qin.""], ""title"": ""Adaptive Chinese word segmentation"", ""venue"": ""Proceedings of ACL, pages 462\u2013469, Barcelona."", ""year"": 2004}, {""authors"": [""Hewlett"", ""Daniel"", ""Paul Cohen.""], ""title"": ""Fully unsupervised word segmentation with BVE and MDL"", ""venue"": ""Proceedings of ACL, pages 540\u2013545, Portland, OR."", ""year"": 2011}, {""authors"": [""Hockenmaier"", ""Julia"", ""Mark Steedman.""], ""title"": ""CCGBank: A corpus of CCG derivations and dependency structures extracted from the Penn treebank"", ""venue"": ""Computational Linguistics, 33(3):355\u2013396."", ""year"": 2007}, {""authors"": [""Hwa"", ""Rebecca"", ""Philip Resnik"", ""Amy Weinberg"", ""Clara Cabezas"", ""Okan Kolak.""], ""title"": ""Bootstrapping parsers via syntactic projection across parallel texts"", ""venue"": ""Natural Language Engineering, 11(3):311\u2013325."", ""year"": 2005}, {""authors"": [""Hwa"", ""Rebecca"", ""Philip Resnik"", ""Amy Weinberg"", ""Okan Kolak.""], ""title"": ""Evaluating translational correspondence using annotation projection"", ""venue"": ""Proceedings of ACL, pages 392\u2013399, Philadephia, PA."", ""year"": 2002}, {""authors"": [""Jiang"", ""Wenbin"", ""Liang Huang"", ""Qun Liu.""], ""title"": ""Automatic adaptation of annotation standards: Chinese word segmentation and POS tagging \u2013 A case study"", ""venue"": ""Proceedings of ACL, pages 522\u2013530,"", ""year"": 2009}, {""authors"": [""Jiang"", ""Wenbin"", ""Liang Huang"", ""Yajuan L\u00fc"", ""Qun Liu.""], ""title"": ""A cascaded linear model for joint Chinese word segmentation and part-of-speech tagging"", ""venue"": ""Proceedings of ACL, pages 897\u2013904, Columbus, OH."", ""year"": 2008}, {""authors"": [""Jiang"", ""Wenbin"", ""Qun Liu.""], ""title"": ""Dependency parsing and projection based on word-pair classification"", ""venue"": ""Proceedings of the ACL, pages 12\u201320, Uppsala."", ""year"": 2010}, {""authors"": [""Jiang"", ""Wenbin"", ""Fandong Meng"", ""Qun Liu"", ""Yajuan L\u00fc.""], ""title"": ""Iterative annotation transformation with predict-self reestimation for Chinese word segmentation"", ""venue"": ""Proceedings of EMNLP,"", ""year"": 2012}, {""authors"": [""Johnson"", ""Mark"", ""Sharon Goldwater.""], ""title"": ""Improving nonparameteric Bayesian inference: Experiments on unsupervised word segmentation with adaptor grammars"", ""venue"": ""Proceedings of NAACL,"", ""year"": 2009}, {""authors"": [""Kruengkrai"", ""Canasai"", ""Kiyotaka Uchimoto"", ""Junichi Kazama"", ""Yiou Wang"", ""Kentaro Torisawa"", ""Hitoshi Isahara""], ""title"": ""An error-driven word-character hybrid model for joint Chinese word segmentation"", ""year"": 2009}, {""authors"": [""Li"", ""Zhongguo.""], ""title"": ""Parsing the internal structure of words: A new paradigm for Chinese word segmentation"", ""venue"": ""Proceedings of ACL, pages 1,405\u20131,414, Portland, OR."", ""year"": 2011}, {""authors"": [""Marcus"", ""Mitchell P."", ""Beatrice Santorini"", ""Mary Ann Marcinkiewicz.""], ""title"": ""Building a large annotated corpus of English: The Penn treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1993}, {""authors"": [""Martins"", ""Andr\u00e9 F.T."", ""Dipanjan Das"", ""Noah A. Smith"", ""Eric P. Xing.""], ""title"": ""Stacking dependency parsers"", ""venue"": ""Proceedings of EMNLP, pages 157\u2013166, Honolulu, HI."", ""year"": 2008}, {""authors"": [""McDonald"", ""Ryan"", ""Koby Crammer"", ""Fernando Pereira.""], ""title"": ""Online large-margin training of dependency parsers"", ""venue"": ""Proceedings of ACL, pages 91\u201398, Ann Arbor, MI."", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Fernando Pereira.""], ""title"": ""Online learning of approximate dependency parsing algorithms"", ""venue"": ""Proceedings of EACL, pages 81\u201388, Trento."", ""year"": 2006}, {""authors"": [""Mihalkova"", ""Lilyana"", ""Tuyen Huynh"", ""Raymond J. Mooney.""], ""title"": ""Mapping and revising Markov logic networks for transfer learning"", ""venue"": ""Proceedings of AAAI, volume 7, pages 608\u2013614, Vancouver."", ""year"": 2007}, {""authors"": [""Mihalkova"", ""Lilyana"", ""Raymond J. Mooney.""], ""title"": ""Transfer learning by mapping with minimal target data"", ""venue"": ""Proceedings of AAAI Workshop Transfer Learning for Complex Tasks, Chicago, IL."", ""year"": 2008}, {""authors"": [""Mochihashi"", ""Daichi"", ""Takeshi Yamada"", ""Naonori Ueda.""], ""title"": ""Bayesian unsupervised word segmentation with nested Pitman-Yor language modeling"", ""venue"": ""Proceedings of ACL-IJCNLP,"", ""year"": 2009}, {""authors"": [""Nakagawa"", ""Tetsuji"", ""Kiyotaka Uchimoto.""], ""title"": ""A hybrid approach to word segmentation and POS tagging"", ""venue"": ""Proceedings of ACL, pages 217\u2013220, Prague."", ""year"": 2007}, {""authors"": [""Ng"", ""Hwee Tou"", ""Jin Kiat Low""], ""title"": ""Chinese part-of-speech tagging: One-at-a-time or all-at-once? Word-based or character-based"", ""venue"": ""In Proceedings of EMNLP,"", ""year"": 2004}, {""authors"": [""Nivre"", ""Joakim"", ""Ryan McDonald.""], ""title"": ""Integrating graph-based and transition-based dependency parsers"", ""venue"": ""Proceedings of ACL, pages 950\u2013958, Columbus, OH."", ""year"": 2008}, {""authors"": [""Oepen"", ""Stephan"", ""Kristina Toutanova"", ""Stuart Shieber"", ""Thorsten Brants""], ""title"": ""The LinGo Redwoods treebank: Motivation and preliminary applications"", ""year"": 2002}, {""authors"": [""Pan"", ""Sinno Jialin"", ""Qiang Yang.""], ""title"": ""A survey on transfer learning"", ""venue"": ""IEEE TKDE, 22(10):1345\u20131359."", ""year"": 2010}, {""authors"": [""Sarkar"", ""Anoop.""], ""title"": ""Applying co-training methods to statistical parsing"", ""venue"": ""Proceedings of NAACL, pages 1\u20138, Pittsburgh, PA."", ""year"": 2001}, {""authors"": [""Smith"", ""David"", ""Jason Eisner.""], ""title"": ""Parser adaptation and projection with quasi-synchronous grammar features"", ""venue"": ""Proceedings of EMNLP, volume 2, pages 822\u2013831, Singapore."", ""year"": 2009}, {""authors"": [""Sun"", ""Weiwei.""], ""title"": ""A stacked sub-word model for joint Chinese word segmentation and part-of-speech tagging"", ""venue"": ""Proceedings of ACL, pages 1,385\u20131,394, Portland, OR."", ""year"": 2011}, {""authors"": [""Sun"", ""Weiwei"", ""Xiaojun Wan.""], ""title"": ""Reducing approximation and estimation errors for Chinese lexical processing with heterogeneous annotations"", ""venue"": ""Proceedings of ACL, volume 1, pages 232\u2013241,"", ""year"": 2012}, {""authors"": [""Sun"", ""Weiwei"", ""Rui Wang"", ""Yi Zhang.""], ""title"": ""Discriminative parse reranking for Chinese with homogeneous and heterogeneous annotations"", ""venue"": ""Proceedings of CIPS-SIGHAN, Beijing. Available at"", ""year"": 2010}, {""authors"": [""Wang"", ""Kun"", ""Chengqing Zong"", ""Keh-Yih Su.""], ""title"": ""A character-based joint model for Chinese word segmentation"", ""venue"": ""Proceedings of COLING, pages 1,173\u20131,181, Beijing."", ""year"": 2010}, {""authors"": [""Xue"", ""Nianwen"", ""Libin Shen.""], ""title"": ""Chinese word segmentation as LMR tagging"", ""venue"": ""Proceedings of SIGHAN Workshop, volume 17, pages 176\u2013179, Sapporo."", ""year"": 2003}, {""authors"": [""Xue"", ""Nianwen"", ""Fei Xia"", ""Fu-Dong Chiou"", ""Martha Palmer.""], ""title"": ""The Penn Chinese treebank: Phrase structure annotation of a large corpus"", ""venue"": ""Natural Language Engineering, 11(2):207\u2013238."", ""year"": 2005}, {""authors"": [""H. Yamada"", ""Y. Matsumoto.""], ""title"": ""Statistical dependency analysis with support vector machines"", ""venue"": ""Proceedings of IWPT, pages 195\u2013206, Nancy."", ""year"": 2003}, {""authors"": [""Yu"", ""Shiwen"", ""Jianming Lu"", ""Xuefeng Zhu"", ""Huiming Duan"", ""Shiyong Kang"", ""Honglin Sun"", ""Hui Wang"", ""Qiang Zhao"", ""Weidong Zhan.""], ""title"": ""Processing norms of modern Chinese corpus"", ""venue"": ""Technical"", ""year"": 2001}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""Chinese segmentation with a word-based perceptron algorithm"", ""venue"": ""Proceedings of ACL, pages 840\u2013847, Prague."", ""year"": 2007}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""A fast decoder for joint word segmentation and POS-tagging using a single discriminative model"", ""venue"": ""Proceedings of EMNLP, pages 843\u2013852, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Zhao"", ""Hai"", ""Chunyu Kit.""], ""title"": ""Unsupervised segmentation helps supervised learning of character tagging for word segmentation and named entity recognition"", ""venue"": ""Proceedings of IJCNLP,"", ""year"": 2008}, {""authors"": [""Zhu"", ""Muhua"", ""Jingbo Zhu"", ""Minghan Hu.""], ""title"": ""Better automatic treebank conversion using a feature-based approach"", ""venue"": ""Proceedings of ACL, volume 2, pages 715\u2013719, Portland, OR."", ""year"": 2011}] acknowledgments  :Jiang, Lü, and Liu were supported by National Natural Science Foundation of China (contract 61202216) and the National Key Technology R&D Program (no. 2012BAH39B03). Huang was supported in part by the DARPA DEFT Project (FA8750-13-2-0041). Liu was partially supported by the Science Foundation Ireland (grant no. 07/CE/I1142) as part of the CNGL at Dublin City University. We also thank the anonymous reviewers for their insightful comments. Finally, we want to thank Chris Hokamp for proofreading. automatic adaptation of annotations :Wenbin Jiang∗
Chinese Academy of Sciences
Yajuan Lü∗
Chinese Academy of Sciences
Liang Huang∗∗
Queens College and Graduate Center,
The City University of New York
Qun Liu∗†
Dublin City University
Chinese Academy of Sciences
Manually annotated corpora are indispensable resources, yet for many annotation tasks, such as
the creation of treebanks, there exist multiple corpora with different and incompatible annotation
guidelines. This leads to an inefficient use of human expertise, but it could be remedied by
integrating knowledge across corpora with different annotation guidelines. In this article we
describe the problem of annotation adaptation and the intrinsic principles of the solutions, and
present a series of successively enhanced models that can automatically adapt the divergence
between different annotation formats.
We evaluate our algorithms on the tasks of Chinese word segmentation and dependency
parsing. For word segmentation, where there are no universal segmentation guidelines be-
cause of the lack of morphology in Chinese, we perform annotation adaptation from the much
larger People’s Daily corpus to the smaller but more popular Penn Chinese Treebank. For
dependency parsing, we perform annotation adaptation from the Penn Chinese Treebank to
a semantics-oriented Dependency Treebank, which is annotated using significantly different
annotation guidelines. In both experiments, automatic annotation adaptation brings significant
improvement, achieving state-of-the-art performance despite the use of purely local features in
training.
∗ Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences, No. 6 Kexueyuan South Road, Haidian District, P.O. Box 2704, Beijing 100190, China. E-mail: {jiangwenbin, liuqun, lvyajuan}@ict.ac.cn.
∗∗ Department of Computer Science, Queens College / CUNY, 65-30 Kissena Blvd., Queens, NY 11367. E-mail: liang.huang.sh@gmail.com. † Centre for Next Generation Localisation, Faculty of Engineering and Computing, Dublin City University.
E-mail: qliu@computing.dcu.ie.
Submission received: 24 April 2013; revised version received: 6 March 2014; accepted for publication: 18 April 2014.
doi:10.1162/COLI a 00210
© 2015 Association for Computational Linguistics type templates instances :Guiding α α=B α ◦ C−2 α=B ◦ C−2=美 α ◦ C−1 α=B ◦ C−1=副 α ◦ C0 α=B ◦ C0=总 α ◦ C1 α=B ◦ C1=统 α ◦ C2 α=B ◦ C2=访 α ◦ C−2C−1 α=B ◦ C−2C−1=美副 α ◦ C−1C0 α=B ◦ C−1C0=副总 α ◦ C0C1 α=B ◦ C0C1=总统 α ◦ C1C2 α=B ◦ C1C2=统访 α ◦ C−1C1 α=B ◦ C−1C1=副统 α ◦ Pu(C0) α=B ◦ Pu(C0)=true α ◦ T(C−2:2) α=B ◦ T(C−2:2)= 44444 partition sections # of words : training iterations. The performance measurement indicator for word segmentation is the balanced F-measure, F = 2PR/(P + R), a function of Precision P and Recall R, where P is the percentage of words in segmentation results that are segmented correctly, and R is the percentage of correctly segmented words in the gold standard words.
For both syntactic and semantic dependency parsing, we concentrate on nonlabeled parsing that predicts the graphic dependency structure for the input sentence without considering dependency labels. The perceptron-based baseline dependency models are trained on the training sets of DCTB and SDT, using the development sets to determine the best training iterations. The performance measurement indicator for dependency parsing is the Unlabeled Attachment Score, denoted as Precision P, indicating the percentage of words in predicted dependency structure that are correctly attached to their head words.
Figure 6 shows the learning curve of the averaged perceptron for word segmentation on the development set. Accuracies of the baseline classifiers are listed in Table 6. We also report the performance of the classifiers on the testing sets of the opposite corpora. Experimental results are in line with our expectations. A classifier performs better in its corresponding testing set, and performs significantly worse on testing data following a different annotation guideline.
Table 7 shows the accuracies of the baseline syntactic and semantic parsers, as well as the performance of the parsers on the testing sets of the opposite corpora. Similar
to the situations in word segmentation, two parsers give state-of-the-art accuracies on their own testing sets, but perform poorly on the other testing sets. This indicates the degree of divergence between the annotation guidelines of DCTB and SDT. ctb  :To approximate more general scenarios of annotation adaptation problems, we extract from PD a subset that is comparable to CTB in size. Because there are many extremely long sentences in the original PD corpus, we first split them into normal sentences according to the full-stop punctuation symbol. We randomly select 20, 000 sentences (0.45M words) from the PD training data as the new training set, and 1, 000/1, 000 sentences from the PD test data as the new testing/developing set. We label the smaller version of PD as SPD. The balanced source corpus and target corpus also facilitate the investigation of annotation adaptation.
Annotation adaptation for dependency parsing is performed from the CTB-derived syntactic dependency treebank (DCTB) (Yamada and Matsumoto 2003) to the Semantic Dependency Treebank (SDT) (Che et al. 2012). Semantic dependency encodes the semantic relationships between words, which are very different from syntactic dependencies. SDT is annotated on a small portion of the CTB text as depicted in Table 5; therefore, we use the subset of DCTB covering the remaining CTB text as the source corpus. We still denote the source corpus as DCTB in the following for simplicity. previous work : semeval-2012 contest :model is denoted as Model 3). We find that the predict-self re-estimation brings improvement to the iterative training at each iteration, for both word segmentation and dependency parsing. The maximum performance is achieved at iteration 4 for word segmentation, and at iteration 5 for dependency parsing. The corresponding models are evaluated on the corresponding testing sets, and the experimental results are also shown in Tables 8 and 9. Compared to Model 1, the optimized annotation adaptation strategy, Model 3, leads to classifiers with significantly higher accuracy and to processing speeds that are several times faster. Tables 10 and 11 show the experimental results compared with previous work. For both Chinese word segmentation and semantic dependency parsing, automatic annotation adaptation yields state-of-the-art performance, despite using single classifiers with only local features. Note that for the systems in the SemEval contest (Che et al. 2012), many other technologies including clause segmentation, system combination, and complicated features were adopted, as well as elaborate engineering. We also performed significance tests2 to verify the effectiveness of annotation adaptation.We find that for both Chinese word segmentation and semantic dependency parsing, annotation adaptation brings significant improvement (p < 0.001) over the baselines trained on the target corpora only.
2 http://www.cis.upenn.edu/∼dbikel/download/compare.pl. word type proportion baseline anno. ada. trend :To evaluate the stability of annotation adaptation, we perform quantitative analysis on the results of annotation adaptation. For word segmentation, the words are grouped according to POS tags. For dependency parsing, the dependency edges are grouped according to POS tag pairs. For each category, the recall values of baseline and annotation adaptation are reported. To filter the lists, we set two significance thresholds with respect to the proportion of a category and the performance fluctuation between two systems. For word segmentation, only the categories with proportions of more than 1% and with fluctuations of more than 0.1 points are reserved, and for dependency parsing, the two thresholds are 1% and 0.5. Tables 12 and 13 show the analysis results for word segmentation and dependency parsing, respectively. For both tasks, annotation adaptation brings improvement for most of the situations. edge type proportion baseline anno. ada. trend :We further investigate the effect of varying the sizes of the target corpora. Experiments are conducted for word segmentation and dependency parsing with fixed-size source corpora and varying-size target corpora. We use SPD and DCTB as the source corpora for word segmentation and dependency parsing, respectively. Figures 13 and 14 show the performance curves on the testing sets. We find that, for both word segmentation and dependency parsing, the improvements brought by annotation adaptation are more significant when the target corpora are smaller. It means that the automatic annotation adaptation is more valuable when the size of the target corpus is small,
which is good news for the situation where the corpus we are concerned with is smaller but a larger differently annotated corpus exists.
Of course, the comparison between automatic annotation adaptation and previous strategies without using additional training data is unfair. Our work aims to find another way to improve NLP tasks: focusing on the collection of more training data instead of making full use of a certain corpus. We believe that the performance of automatic annotation adaptation can be further improved by adopting the advanced technologies of previous work, such as complicated features and model combination. It would be useful to conduct experiments with more source-annotated training data, such as the SIGHAN data set for word segmentation, to investigate the trend of improvement along with the further increment of annotated sentences. It would also be valuable to evaluate the improved word segmenter and dependency parser on the out-of-domain data sets. However, currently most corpora for word segmentation and dependency parsing do not explicitly distinguish the domains of their data sections, making such evaluations difficult to conduct.","8 conclusion and future work :We have described the problem of annotation adaptation and the intrinsic principles of its solutions, and proposed a series of successively enhanced models that can automatically adapt the divergence between different annotation formats. These models learn the statistical regularities of adaptation between different annotation guidelines,
and integrate the knowledge in corpora with different annotation guidelines. In the problems of Chinese word segmentation and semantic dependency parsing, annotation adaptation algorithms bring significant improvements by integrating the knowledge in differently annotated corpora, People’s Daily and Penn Chinese Treebank for word segmentation, and Penn Chinese Treebank and Semantic Dependency Parsing for dependency parsing. For both tasks, annotation adaptation leads to a segmenter and a parser achieving the state-of-the-art, despite using only local features in single classifiers.
Many aspects related to annotation adaptation deserve further investigation in the future. First, models for annotation adaptation can be adapted to other NLP tasks such as semantic analysis. Second, jointly tackling the divergences in both annotations and domains is an important problem. In addition, an unsupervised-induced or bilingually projected corpus, despite performing poorly on the specified testing data, may have high inner annotation consistency. That is to say, the induced corpora can be treated as a knowledge source following another annotation guideline, and the performance of current unsupervised or bilingually projected models may be seriously underestimated. Annotation adaptation may give us a new perspective on knowledge induction and measurement for such methods."
"1 introduction :Human evaluation is a key aspect of many NLP technologies. Automatic metrics that correlate with human judgments have been developed, especially in Machine Translation, to relieve some of the burden. Neverthess, Callison-Burch et al. (2007) note in their meta-evaluation that in MT they still “consider the human evaluation to be primary.” Whereas MT has traditionally used a Likert scale score for the criteria of adequacy and fluency, this meta-evaluation noted that these are “seemingly difficult things for judges to agree on”; consequently, asking judges to express a preference between alternative translations is increasingly used on the grounds of ease and intuitiveness. Further, where the major empirical results of a paper are from automatic metrics, it is still useful to supplement them: As two examples, Collins, Koehn, and Kucerova (2005) and Lewis and Steedman (2013), in addition to a metric-based evaluation, present human judgments of preferences for their systems with respect to a baseline (Fig. 1). For results in published work, the reader is typically left to draw inferences from the numbers. For the data in Figure 1, is there a strong preference for the non-baseline system overall, or do null preferences count against that? Is anything about the results statistically significant? There has been work in various areas of NLP in assessing statistical significance of human judgment results. However, to our knowledge, the field has not taken advantage of a body of work dedicated to analyzing human preferences—predominantly in the context of sensory discrimination testing, and consequent consumer behavior—which is supported by a great deal of statistical theory. It is linked to the mixed-effect models that are increasingly prominent in psycholinguistics and elsewhere, it has associated freely available R software, and it permits questions like the following to be asked: Can we say that the judges are expressing a preference at all, as opposed to no preference? Is there an effect from judge disagreement or inconsistency?
∗ Department of Computing, Macquarie University, NSW 2109, Australia. E-mail: mark.dras@mq.edu.au.
Submission received: 10 April 2014; accepted for publication: 18 July 2014.
doi:10.1162/COLI a 00222
© 2015 Association for Computational Linguistics
We describe our sample data (Section 2), sketch a classical non-parametric approach (Section 3) and discuss the issues that arise from this, and look at some of the approaches used in MT (Section 4). We then (Section 5) introduce ideas from human sensory preference testing, where we review log-linear Bradley-Terry models of preferences, and apply this to our data, including discussion of ties, of subject effects, and of multiple pairwise comparisons.","2 two data sets :Single Pairwise Comparison. Our basic single pairwise comparisons are those presented in Figure 1(a) and (b). Figure 1(c) contains the counts we will be using in later analysis: We refer to counts in favor of the new system by n+, those in favor of the baseline by n−, and those reflecting no preference as n0; the Lewis and Steedman results were over 87 pairwise judgments. We add some further artificial data to illustrate how the Log-Linear Bradley-Terry (LLBT) models of Section 5 behave in accordance with intuition for data where the conclusion should be clear. These comprise a distribution with a moderate preference for + over − and not too many null preferences (ModPref), a distribution of equal preferences over all three categories (EqualPref), a distribution with mostly null preferences and equal n+ and n− (NoPref), and a distribution with very few null preferences (StrongPref).
Multiple Pairwise Comparison. As noted in Section 1, there has been a trend to using human preference judgments, particularly in the workshops on statistical machine translation from Callison-Burch et al. (2007) onwards. Schemes have included asking humans to rank random samples of five translations, each from a different system. Vilar et al. (2007) propose using binary rather than n-ary rankings, arguing that this is a natural and easy task for judges. Here we present some artificial data of pairwise (binary) rankings to illustrate the techniques we discuss in Section 5, although these techniques can be extended to n-ary comparisons. In our example, there are four systems A, B, C, D and four judges J1, J2, J3, J4. The judges have pairwise ranked 240 translation pairs from systems x and y, indicating whether the translation of x is better than y (x y), worse than y (x ≺ y), or similar in quality to y (x = y); see Table 1. An overall impression, totalling all pairwise first preferences for each system (Section 4), gives a ranking of systems A–D–B–C. It can also be seen that there is little in the way of undecidedness, and also that judge J3 differs from the general judge opinion in pairwise ranking of AD and BC.","3 classical non-parametric methods :A classical approach to evaluating preferences is the non-parametric sign test (Sprent and Smeeton 2007). The first issue in applying this test here is ties, or expressions of no preference—these are often ignored when the proportion of ties is small, but for our typical examples of Figure 1, this is not true. Randles (2001) observes, regarding the approach most widely recommended by textbooks of just ignoring ties, that “the constrained number of possible p values and its ‘elimination of zeroes’ has caused concern and controversy through the years.” Randles (2001) and Rayner and Best (2001, chapter 2), reviewing several approaches to handling ties, both advocate splitting ties in various ways depending on the problem setting, for (in Randles’s characterization) “it is desirable that zeros have a conservative influence on declaring preference, but not to the same degree as negative responses.” The key point is that modeling of ties explicitly can be important, although there is no consensus on how this should be done; no approach apart from ignoring ties appears to be in widespread use. The second issue with the sign test is that of multiple judges, where data points are related (e.g., the same items are given to all judges). The Friedman test (Sprent and Smeeton 2007, Section 7.3.1) can be viewed as an extension that can be applied to multiple subjects ranking multiple items (see Bi 2006, Section 5.1.3, for an example). However, Francis, Dittrich, and Hatzinger (2010) note that
[the Friedman test] simply examines the null hypothesis that the median ranks for all items are equal, and does not consider any differences in ranking between respondents. . . . Moreover, if the Friedman test rejects the null hypothesis, no quantitative interpretation, such as the odds of preferring one item over another, is provided. [Further, this] fail[s] both to consider the underlying psychological mechanism for ranking, and to formulate correct statistical models for this mechanism.","4 methods in machine translation :Human evaluation in NLP is a pervasive issue, but here we focus on MT and its shared tasks. The 2007 shared task (Callison-Burch et al. 2007) was the first to investigate a range of approaches that specifically included ranking of n translations, from best to worst, allowing ties (which were ignored); from this they defined an aggregate “rank,” “the average number of times that a system was judged to be better than any other system in the sentence ranking evaluation.” They assessed inter-annotator agreement, and—with a key goal of the meta-evaluation being to find the automatic evaluation metric that best matched human evaluations—calculated Spearman’s rank correlation coefficient between the two types of assessment. The 2008 shared task (Callison-Burch
et al. 2008) took the same approach, but noted that in ranking, “[h]ow best to treat these is an open discussion, and certainly warrants further thought,” in particular because of ties “further complicating matters.” Pado et al. (2009) modified the systemlevel predictions approach to become “tie-aware,” and noted that that this “makes a considerable practical difference, improving correlation figures by 5–10 points.” At around the same time Vilar et al. (2007) examined the use of pairwise comparisons in MT evaluation. They pose the problem as one where, given an order relationship is-better-than between pairs of systems, the goal is to find an ordering of all the systems: They see this as the fundamental computer science problem of sorting. They define an aggregate evaluation score for comparing systems, estimating expected value and standard error for hypothesis testing. However, in aggregating this way information about ties is lost.
Bojar et al. (2011) critique the earlier WMT evaluations, citing issues with the ignoring of non-top ranks (noted in Section 3 herein also), with ties and also with interannotator agreement. Lopez (2012) extends the analysis of Bojar et al. and casts the problem as “finding the minimum feedback arc set in a tournament, a well-known NPcomplete problem.” He advocates using the pairwise rankings themselves, rather than aggregate statistics like Vilar et al. (2007), and aims to minimize the number of violations among these. Koehn (2012) evaluates empirically the approaches of both Bojar et al. (2011) and Lopez (2012), with a focus on determining which systems are statistically distinguishable in terms of performance, defining confidence bounds for this purpose.
Hopkins and May (2013) recently advocated a focus on finding the extent to which particular rankings could be trusted. They proposed a model based on Item Response Theory (IRT), which underlies many standardized tests. They draw an analogy with judges assessing students on the basis of an underlying distribution of the student’s ability, with items authored by students having a quality drawn from the student’s ability distribution. They note in passing that a Gaussian parameterization of their IRT models resembles Thurstone and Bradley-Terry models; this leads us to the topic of Section 5.
Overall, then, there are ongoing discussions about what kind of analysis is appropriate for preference judgments. Some of this involves moderately heavy-duty computation for bootstrapping; this is suitable for large-scale WMT evaluations with dozens of competing systems, but perhaps less so for the scenarios we envisage in Section 1. Moreover, examining what techniques other fields have developed could be useful, especially when they come with ready-made, easy-to-use tools for smaller-scale evaluation.","5 preferences and log-linear bradley-terry methods :The statistical analysis of human perception and preferences dates back at least to the psychophysics work of German physiologist E. H. Weber in the nineteenth century. A progression from the way humans perceive differences between physical stimuli to more general analysis of human preferences has occurred particularly in the context of investigating consumer behavior—dealing with questions like whether there is a definite preference for a food with a particular type of ingredient, for example—and this is now a fully fledged area of research. Sources like Lawless and Heymann (2010) give overviews of the field and relevant statistical techniques. The earliest generally cited models for pairwise comparisons are the Thurstone model (Thurstone 1927) and the closely related Bradley-Terry (BT) model (Bradley and Terry 1952); these have connections to the IRT models, widely used in analyzing responses to questionnaires, which Hopkins and May (2013) drew on. Here we only look at BT models.
In a basic BT model, the probability that object j (Oj) is preferred to object k (Ok) from a set of J objects in a particular pairwise comparison jk is given by p(Oj Ok |πj,πk) =
πj πj+πk
for all j = k, where πj and πk are non-negative “worth” parameters describing the location of the object on the scale of preferences for some attribute. For n objects, there will be ( n 2 ) pairwise comparisons.
Log-Linear Models. It is now standard to fit BT models as log-linear models (Agresti 2007, for example), which allows them to be treated in a uniform way with much of modern statistical analysis. Log-linear models are a variety of generalized linear models (GLM), as is, for example, the logistic regression used throughout NLP. GLMs consist of a random component that identifies the response variable Y and selects a probability distribution for it; a systematic component that specifies some linear combination of the explanatory variables xi; and a link function g(·) applied to the mean μ of Y relating μ to this linear combination. They thus have the form g(μ) = α+ β1x1 + . . .+ βkxk. For log-linear models, the response variables are counts that are assumed to follow a Poisson distribution, and the link function is g(μ) = log(μ) (compare logistic regression’s g(μ) = log μ1−μ ). As an example, Y might be counts of people who hold some belief, and the various xi might be gender, socioeconomic status, and so forth. GLMs are a key tool for modern categorical data analysis, Agresti (2007, p. 65) noting that using models rather than the non-parametric approaches of Section 3 has several benefits:
The structural form of the model describes the patterns of association and interaction. The sizes of the model parameters determine the strength and importance of the effects. Inferences about the parameters evaluate which explanatory variables affect the response variable Y, while controlling effects of possible confounding variables. Finally, the model’s predicted values smooth the data and provide improved estimates of the mean of Y at possible explanatory variable values.
In a log-linear model, intuitive log-odds interpretations of making one response relative to another can be derived from the parameters. (Typically, software chooses a reference parameter and other parameter values are relative to that.) Statistical significance scores and standard errors can be calculated for these parameters. In addition, GLMs allow for testing of model fit. There are various model choices (e.g., should we include ties? should we include terms representing interactions?) and goodnessof-fit tests can assess the alternatives (see, e.g., Agresti 2007, Section 7.2.1). The model with a separate parameter for each cell in the associated contingency table is called the saturated model, and fits the data perfectly, making it a suitable comparator for alternatives. Deviance is a likelihood ratio statistic comparing a proposed model to the saturated one, allowing a test of the hypothesis that parameters not included in the model are zero, via goodness of fit tests; large test statistics and small p-values provide evidence of model lack of fit.
Models with Ties. To set out the representation of LLBT models, we follow the formulation of Dittrich and Hatzinger (2009). Let n(jk) be the number of comparisons between objects j and k; and let Y(jk)j be the number of preferences for object j with respect to k (similarly, Y(jk)k). The outcome of a paired comparison experiment can also be regarded as a (J 2
)× J incomplete two-dimensional contingency table: There are(J 2 ) rows of pairwise comparisons, and J columns recording choices of the jth object. As with log-linear models in general, the distribution of random variables Y(jk)j and Y(jk)k is assumed to be Poisson. Conditional on fixed n(jk) = Y(jk)j + Y(jk)k, (Y(jk)j, Y(jk)k)
follow a binomial (more generally, multinomial) distribution. The expected number of preferences of object j with respect to object k is denoted m(jk)j and given by n(jk)p(jk)j, with p(jk)j the binomial probability. So far this is only for binary preferences; there are various ways to account for ties. We describe the approach of Davidson and Beaver (1977), which appears quite widely used, where there is a common null preference effect for all pairwise comparisons. Then
log m(jk)j = μ(jk) + λOj − λOk log m(jk)0 = μ(jk) + γ log m(jk)k = μ(jk) − λOj + λOk
(1)
where the μ’s are “nuisance” parameters that fix the n(jk) marginal distributions, and the λO’s represent object parameters, m(jk)0 is the expected number of null preferences for pair (jk), and γ is the undecided effect. The object parameters are related to the worth parameters of the original definition by logπ = 2λO: These represent the log-odds.
In addition to the theoretical reasons for using LLBTs for modeling pairwise comparisons, a key benefit is the availability of packages in R for doing the modeling. Two candidates allowing a variety of sophisticated models are by Turner and Firth (2012) and Hatzinger and Dittrich (2012); we use the latter as the current version of the former does not handle ties. We first apply the model described by Equations (1) to the single pairwise data with ties from Section 2 using R. We refer the reader to the associated data bundle1 for the full output; we only excerpt it in the discussion below. Immediately following is a snippet of the R output for the ModPref data from Figure 1. o1 is the variable λOj for the + category, o2 for the – category, g1 for the null preferences.
Estimate Std. Error z value Pr(>|z|) o1 0.2778 0.1060 2.620 0.0088 ** o2 0.0000 NA NA NA g1 -0.6551 0.2300 -2.848 0.0044 ** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual deviance: -6.6614e-15 on 0 degrees of freedom
In the R output, o2 is the reference object, with parameter value set to zero; the negative value of the estimate for g1 combined with its statistical significance says that there is a strong tendency for an expression of preference. The positive value of the o1 parameter and its significance indicate that the + group is strongly preferred: The odds in favor of this group with respect to the – group is exp(2 × 0.2778) = 1.74 to 1. Relating this to the description of the data in Section 2, then, there is a strong preference for translations by the proposed system relative to the baseline, even taking into account null preferences. The LLBT model confirms that even small data sets like this can produce meaningful and statistically significant results. For the other artificial preference data of Figure 1, the parameters behave as expected: for EqualPref, parameter estimates are all zero, signifying that they all have the same odds; for NoPref, the positive g1 indicates a strong tendency towards no preference; for StrongPref, the negative g1 indicates a strong tendency towards some preference, but with + or – equally likely. Note that all of these are saturated models: there are three objects and three parameters, so the model fits perfectly (indicated also by zero residual deviance). When we apply them
1 Data files and all R commands and output are at https://purl.org/NET/cl-llbt-data.
to the real count data of Figure 1 (c), the results indicate that for the Collins et al. data there is a weak to moderate tendency not to choose (g1 estimate 0.303, p = 0.0432), but, given that, there is a significant (0.0001) preference in favor of the reordered system. For the Lewis and Steedman results, the model gives similar results, albeit with a much stronger disposition to null preferences. In the data bundle we also carry out the sign test ignoring ties for each data set for comparison; it gives the same results in each case for the relation of + than –, but does not allow an evaluation of the effect of ties.
We now apply the model described by Equations (1) to the multiple pairwise data of Table 1. In the R output, the four systems A, B, C, D correspond to objects o1, o2, o3, o4, and g1 again to null preferences. As per the overview of the MT data in Section 2, there is little undecidedness (large negative g1). The coefficients show that object o1 (system A) is most preferred, followed by o4 (D), then o2 (B) and o3 (C). Note also that in this case, the model is not saturated: There is a non-zero residual deviance. As mentioned, log-linear models can be compared in terms of goodness of fit: Dittrich, Hatzinger, and Katzenbeisser (1998) and Dittrich and Hatzinger (2009) discuss this in some detail for LLBT models. Chi-squared statistics can be used to assess goodness of fit based on the residual deviance; the degrees of freedom (d.f.) equal the number of cell counts minus the number of model parameters; both deviance and d.f. are given in the R output. For this data deviance is 30.646 on 8 d.f., whereas by contrast if the ties (g1) are left out, it is 221.22 on 9 d.f. A chi-squared test would establish the goodness of fit for each model; but even without consulting the test it can be seen that leaving out the one parameter related to ties (1 d.f.) gives a seven-fold increase in deviance, so clearly inclusion of ties produces a much better model.
Introducing Subject Covariates. The model can also incorporate a range of other factors, a possibility not easily open to non-parametric methods. The one we look at here is the notion of a categorical covariate, introduced into LLBT models in Dittrich, Hatzinger, and Katzenbeisser (1998): This allows the objects (items) to vary with characteristics of the subject (judge). Many types of subject covariates could be added, grouping subjects by native language of the speaker, source of judges (e.g., Mechanical Turk, university), and so forth. Here we add just one, the identity of the subject. (Typically in a GLM this would be a random effect; we treat it as a covariate just for our simple illustration.) We define our categorical covariate S to have levels l, l = 1, 2, . . . , L. Let m(jk)j|l be the expected number of preferences for object j with respect to object k for subjects in covariate class l. The log-linear representation is then as follows:
log m( jk)j|l = μ(jk)l + λOj − λOk + λSl + λOSjl − λOSkl log m( jk)0|l = μ(jk)l + λSl + γ log m( jk)k|l = μ(jk)l − λOj + λOk + λSl − λOSjl + λOSkl
(2)
As do Dittrich and Hatzinger (2009), we define a reference group, with the λOj ’s representing the ordering for that group; the orderings for other groups are obtained by adding the λOSjl ’s specific to group l to the λ O l ’s for the reference group. μ(jk)l and λ S l are again “nuisance” parameters, the latter representing the main effect of the subject covariate measured on the lth level; λOSjl ’s are the (useful) subject-object interaction parameters describing the effect of the subject covariate on the preference for object j (similarly λOSkl and object k). We apply the model described by Equations (2) to the multiple pairwise data, with the subject covariate SUBJ with four levels (one per judge Ji of Table 1). There are a few complexities in interpreting the output, beyond the scope
of this article to discuss but covered in Dittrich, Hatzinger, and Katzenbeisser (1998). The broad interpretations to draw from the output are that interactions o1:SUBJ3 and o2:SUBJ3 are large and significant, and contribute to the model, unlike any others. These correspond to the different pairwise rankings given by judge J3 to system A (relative to D) and to B (relative to C): This is how subject effects are indicated in these LLBT models.
There are many other extensions to these models. Cattelan (2012) gives a state-ofthe-art overview of such extensions across a range of approaches, with an emphasis on dependent data. We only note two extensions here that are incorporated into prefmod and relevant to NLP. With categorical object covariates, items can be grouped as well, to investigate effects of grouping there, for example, different origins for translation sources. With non-pairwise rankings, judges can rank over more than two elements, as in the standard WMT evaluations, although this needs a special treatment in the models.","6 conclusions :We have looked at the sort of (pairwise) preference data that is encountered often in NLP. A particular characteristic of NLP data is that ties or undecided results may be frequent, and there is often a concern with inter-judge consistency. Reviewing classical non-parametric approaches, we note the opinion that it is important to model ties, and also note that approaches to looking at subject (judge) effects have several issues, such as a lack of quantitative interpretation of results. Among NLP approaches, especially within MT, new techniques are still being derived, which could benefit from views from outside the field. What we present are techniques from the field of sensory preference evaluation, where there has been a long history of development by statistics researchers. Recently, log-linear models have attracted attention. Applying them to sample data, we find that they provide the sort of information and uniform framework for analysis that NLP researchers could find useful. Given both extensive theoretical underpinings and freely available statistical software, we recommend LLBT models as a potential tool.",,,"Human evaluation plays an important role in NLP, often in the form of preference judgments. Although there has been some use of classical non-parametric and bespoke approaches to evaluating these sorts of judgments, there is an entire body of work on this in the context of sensory discrimination testing and the human judgments that are central to it, backed by rigorous statistical theory and freely available software, that NLP can draw on. We investigate one approach, Log-Linear Bradley-Terry models, and apply it to sample NLP data.","[{""affiliations"": [], ""name"": ""Mark Dras""}]",SP:dea01ec1a2d1ac7f75876dd5e599522271536f62,"[{""authors"": [""Agresti"", ""Alan.""], ""title"": ""An Introduction to Categorical Data Analysis"", ""venue"": ""John Wiley, 2nd edition."", ""year"": 2007}, {""authors"": [""Bi"", ""Jian.""], ""title"": ""Sensory Discrimination Tests and Measurements: Statistical Principles, Procedures and Tables"", ""venue"": ""Blackwell, Oxford, UK."", ""year"": 2006}, {""authors"": [""Bojar"", ""Ond\u0159ej"", ""Milo\u0161 Ercegov\u010devi\u0107"", ""Martin Popel"", ""Omar Zaidan.""], ""title"": ""A grain of salt for the WMT manual evaluation"", ""venue"": ""Proceedings of WMT, pages 1\u201311."", ""year"": 2011}, {""authors"": [""Bradley"", ""Ralph"", ""Milton Terry.""], ""title"": ""Rank analysis of incomplete block designs, I"", ""venue"": ""The method of paired comparisons. Biometrika, 39:324\u2013345."", ""year"": 1952}, {""authors"": [""Callison-Burch"", ""Chris"", ""Cameron Fordyce"", ""Philipp Koehn"", ""Christof Monz"", ""Josh Schroeder.""], ""title"": ""Meta-) Evaluation of machine translation"", ""venue"": ""Proceedings of WMT, pages 136\u2013158."", ""year"": 2007}, {""authors"": [""Josh Schroeder""], ""title"": ""Further meta-evaluation of machine translation"", ""venue"": ""In Proceedings of WMT,"", ""year"": 2008}, {""authors"": [""Cattelan"", ""Manuela.""], ""title"": ""Models for paired comparison data: A review with emphasis on dependent data"", ""venue"": ""Statistical Science, 27(3):412\u2013423."", ""year"": 2012}, {""authors"": [""Collins"", ""Michael"", ""Philipp Koehn"", ""Ivona Kucerova.""], ""title"": ""Clause restructuring for statistical machine translation"", ""venue"": ""Proceedings of ACL, pages 531\u2013540."", ""year"": 2005}, {""authors"": [""R.R. Davidson"", ""R.J. Beaver.""], ""title"": ""On extending the Bradley-Terry model to incorporate within-pair order effects"", ""venue"": ""Biometrics, 33:693\u2013702."", ""year"": 1977}, {""authors"": [""Dittrich"", ""Regina"", ""Reinhold Hatzinger.""], ""title"": ""Fitting loglinear Bradley-Terry models (LLBT) for paired comparisons using the R package prefmod"", ""venue"": ""Psychological Science Quarterly,"", ""year"": 2009}, {""authors"": [""Dittrich"", ""Regina"", ""Reinhold Hatzinger"", ""W. Katzenbeisser""], ""title"": ""Modelling the effect of subject-specific covariates in paired comparison studies with an application to university rankings"", ""year"": 1998}, {""authors"": [""Francis"", ""Brian"", ""Regina Dittrich"", ""Reinhold Hatzinger""], ""title"": ""Modeling heterogeneity in ranked responses by nonparametric maximum likelihood: How do Europeans"", ""year"": 2010}, {""authors"": [""Hatzinger"", ""Reinhold"", ""Regina Dittrich.""], ""title"": ""prefmod: A package for modeling preferences based on paired comparisons, rankings, or ratings"", ""venue"": ""Journal of Statistical Software, 48(10):1\u201331."", ""year"": 2012}, {""authors"": [""Hopkins"", ""Mark"", ""Jonathan May.""], ""title"": ""Models of translation competitions"", ""venue"": ""Proceedings of ACL, pages 1,416\u20131,424."", ""year"": 2013}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Simulating human judgment in machine translation evaluation campaigns"", ""venue"": ""Proceedings of IWSLT, pages 179\u2013184."", ""year"": 2012}, {""authors"": [""Lawless"", ""Harry T."", ""Hildegarde Heymann.""], ""title"": ""Sensory Evaluation of Food: Principles and Practices"", ""venue"": ""Springer, New York, NY 2nd edition."", ""year"": 2010}, {""authors"": [""Lewis"", ""Mike"", ""Mark Steedman""], ""title"": ""Unsupervised induction of cross-lingual"", ""year"": 2013}, {""authors"": [""Lopez"", ""Adam.""], ""title"": ""Putting human assessments of machine translation systems in order"", ""venue"": ""Proceedings of WMT, pages 1\u20139."", ""year"": 2012}, {""authors"": [""Pado"", ""Sebastian"", ""Michel Galley"", ""Dan Jurafsky"", ""Christopher D. Manning.""], ""title"": ""Robust machine translation evaluation with entailment features"", ""venue"": ""Proceedings of ACL / AFNLP, pages 297\u2013305."", ""year"": 2009}, {""authors"": [""Randles"", ""Ronald H.""], ""title"": ""On Neutral Responses (zeros) in the sign test and ties in the Wilcoxon-Mann-Whitney test"", ""venue"": ""The American Statistician, 55(2):96\u2013101."", ""year"": 2001}, {""authors"": [""J.C.W. Rayner"", ""D.J. Best.""], ""title"": ""A Contingency Table Approach to Nonparametric Testing"", ""venue"": ""Chapman and Hall/CRC, Boca Raton, FL."", ""year"": 2001}, {""authors"": [""Sprent"", ""Peter"", ""Nigel Smeeton.""], ""title"": ""Applied Nonparametric Statistical Methods"", ""venue"": ""Chapman and Hall, London, UK."", ""year"": 2007}, {""authors"": [""L.L. Thurstone""], ""title"": ""A law of comparative Judgement"", ""venue"": ""Psychological Review, 34:278\u2013286."", ""year"": 1927}, {""authors"": [""Turner"", ""Heather"", ""David Firth.""], ""title"": ""Bradley-Terry Models in R: The BradleyTerry2 Package"", ""venue"": ""Journal of Statistical Software, 48(9):1\u201321."", ""year"": 2012}, {""authors"": [""Vilar"", ""David"", ""Gregor Leusch"", ""Hermann Ney"", ""Rafael E. Banchs.""], ""title"": ""Human evaluation of machine translation through binary system comparisons"", ""venue"": ""Proceedings of WMT, pages 96\u2013103."", ""year"": 2007}]",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :Human evaluation is a key aspect of many NLP technologies. Automatic metrics that correlate with human judgments have been developed, especially in Machine Translation, to relieve some of the burden. Neverthess, Callison-Burch et al. (2007) note in their meta-evaluation that in MT they still “consider the human evaluation to be primary.” Whereas MT has traditionally used a Likert scale score for the criteria of adequacy and fluency, this meta-evaluation noted that these are “seemingly difficult things for judges to agree on”; consequently, asking judges to express a preference between alternative translations is increasingly used on the grounds of ease and intuitiveness. Further, where the major empirical results of a paper are from automatic metrics, it is still useful to supplement them: As two examples, Collins, Koehn, and Kucerova (2005) and Lewis and Steedman (2013), in addition to a metric-based evaluation, present human judgments of preferences for their systems with respect to a baseline (Fig. 1). For results in published work, the reader is typically left to draw inferences from the numbers. For the data in Figure 1, is there a strong preference for the non-baseline system overall, or do null preferences count against that? Is anything about the results statistically significant? There has been work in various areas of NLP in assessing statistical significance of human judgment results. However, to our knowledge, the field has not taken advantage of a body of work dedicated to analyzing human preferences—predominantly in the context of sensory discrimination testing, and consequent consumer behavior—which is supported by a great deal of statistical theory. It is linked to the mixed-effect models that are increasingly prominent in psycholinguistics and elsewhere, it has associated freely available R software, and it permits questions like the following to be asked: Can we say that the judges are expressing a preference at all, as opposed to no preference? Is there an effect from judge disagreement or inconsistency?
∗ Department of Computing, Macquarie University, NSW 2109, Australia. E-mail: mark.dras@mq.edu.au.
Submission received: 10 April 2014; accepted for publication: 18 July 2014.
doi:10.1162/COLI a 00222
© 2015 Association for Computational Linguistics
We describe our sample data (Section 2), sketch a classical non-parametric approach (Section 3) and discuss the issues that arise from this, and look at some of the approaches used in MT (Section 4). We then (Section 5) introduce ideas from human sensory preference testing, where we review log-linear Bradley-Terry models of preferences, and apply this to our data, including discussion of ties, of subject effects, and of multiple pairwise comparisons. 2 two data sets :Single Pairwise Comparison. Our basic single pairwise comparisons are those presented in Figure 1(a) and (b). Figure 1(c) contains the counts we will be using in later analysis: We refer to counts in favor of the new system by n+, those in favor of the baseline by n−, and those reflecting no preference as n0; the Lewis and Steedman results were over 87 pairwise judgments. We add some further artificial data to illustrate how the Log-Linear Bradley-Terry (LLBT) models of Section 5 behave in accordance with intuition for data where the conclusion should be clear. These comprise a distribution with a moderate preference for + over − and not too many null preferences (ModPref), a distribution of equal preferences over all three categories (EqualPref), a distribution with mostly null preferences and equal n+ and n− (NoPref), and a distribution with very few null preferences (StrongPref).
Multiple Pairwise Comparison. As noted in Section 1, there has been a trend to using human preference judgments, particularly in the workshops on statistical machine translation from Callison-Burch et al. (2007) onwards. Schemes have included asking humans to rank random samples of five translations, each from a different system. Vilar et al. (2007) propose using binary rather than n-ary rankings, arguing that this is a natural and easy task for judges. Here we present some artificial data of pairwise (binary) rankings to illustrate the techniques we discuss in Section 5, although these techniques can be extended to n-ary comparisons. In our example, there are four systems A, B, C, D and four judges J1, J2, J3, J4. The judges have pairwise ranked 240 translation pairs from systems x and y, indicating whether the translation of x is better than y (x y), worse than y (x ≺ y), or similar in quality to y (x = y); see Table 1. An overall impression, totalling all pairwise first preferences for each system (Section 4), gives a ranking of systems A–D–B–C. It can also be seen that there is little in the way of undecidedness, and also that judge J3 differs from the general judge opinion in pairwise ranking of AD and BC. 3 classical non-parametric methods :A classical approach to evaluating preferences is the non-parametric sign test (Sprent and Smeeton 2007). The first issue in applying this test here is ties, or expressions of no preference—these are often ignored when the proportion of ties is small, but for our typical examples of Figure 1, this is not true. Randles (2001) observes, regarding the approach most widely recommended by textbooks of just ignoring ties, that “the constrained number of possible p values and its ‘elimination of zeroes’ has caused concern and controversy through the years.” Randles (2001) and Rayner and Best (2001, chapter 2), reviewing several approaches to handling ties, both advocate splitting ties in various ways depending on the problem setting, for (in Randles’s characterization) “it is desirable that zeros have a conservative influence on declaring preference, but not to the same degree as negative responses.” The key point is that modeling of ties explicitly can be important, although there is no consensus on how this should be done; no approach apart from ignoring ties appears to be in widespread use. The second issue with the sign test is that of multiple judges, where data points are related (e.g., the same items are given to all judges). The Friedman test (Sprent and Smeeton 2007, Section 7.3.1) can be viewed as an extension that can be applied to multiple subjects ranking multiple items (see Bi 2006, Section 5.1.3, for an example). However, Francis, Dittrich, and Hatzinger (2010) note that
[the Friedman test] simply examines the null hypothesis that the median ranks for all items are equal, and does not consider any differences in ranking between respondents. . . . Moreover, if the Friedman test rejects the null hypothesis, no quantitative interpretation, such as the odds of preferring one item over another, is provided. [Further, this] fail[s] both to consider the underlying psychological mechanism for ranking, and to formulate correct statistical models for this mechanism. 4 methods in machine translation :Human evaluation in NLP is a pervasive issue, but here we focus on MT and its shared tasks. The 2007 shared task (Callison-Burch et al. 2007) was the first to investigate a range of approaches that specifically included ranking of n translations, from best to worst, allowing ties (which were ignored); from this they defined an aggregate “rank,” “the average number of times that a system was judged to be better than any other system in the sentence ranking evaluation.” They assessed inter-annotator agreement, and—with a key goal of the meta-evaluation being to find the automatic evaluation metric that best matched human evaluations—calculated Spearman’s rank correlation coefficient between the two types of assessment. The 2008 shared task (Callison-Burch
et al. 2008) took the same approach, but noted that in ranking, “[h]ow best to treat these is an open discussion, and certainly warrants further thought,” in particular because of ties “further complicating matters.” Pado et al. (2009) modified the systemlevel predictions approach to become “tie-aware,” and noted that that this “makes a considerable practical difference, improving correlation figures by 5–10 points.” At around the same time Vilar et al. (2007) examined the use of pairwise comparisons in MT evaluation. They pose the problem as one where, given an order relationship is-better-than between pairs of systems, the goal is to find an ordering of all the systems: They see this as the fundamental computer science problem of sorting. They define an aggregate evaluation score for comparing systems, estimating expected value and standard error for hypothesis testing. However, in aggregating this way information about ties is lost.
Bojar et al. (2011) critique the earlier WMT evaluations, citing issues with the ignoring of non-top ranks (noted in Section 3 herein also), with ties and also with interannotator agreement. Lopez (2012) extends the analysis of Bojar et al. and casts the problem as “finding the minimum feedback arc set in a tournament, a well-known NPcomplete problem.” He advocates using the pairwise rankings themselves, rather than aggregate statistics like Vilar et al. (2007), and aims to minimize the number of violations among these. Koehn (2012) evaluates empirically the approaches of both Bojar et al. (2011) and Lopez (2012), with a focus on determining which systems are statistically distinguishable in terms of performance, defining confidence bounds for this purpose.
Hopkins and May (2013) recently advocated a focus on finding the extent to which particular rankings could be trusted. They proposed a model based on Item Response Theory (IRT), which underlies many standardized tests. They draw an analogy with judges assessing students on the basis of an underlying distribution of the student’s ability, with items authored by students having a quality drawn from the student’s ability distribution. They note in passing that a Gaussian parameterization of their IRT models resembles Thurstone and Bradley-Terry models; this leads us to the topic of Section 5.
Overall, then, there are ongoing discussions about what kind of analysis is appropriate for preference judgments. Some of this involves moderately heavy-duty computation for bootstrapping; this is suitable for large-scale WMT evaluations with dozens of competing systems, but perhaps less so for the scenarios we envisage in Section 1. Moreover, examining what techniques other fields have developed could be useful, especially when they come with ready-made, easy-to-use tools for smaller-scale evaluation. 5 preferences and log-linear bradley-terry methods :The statistical analysis of human perception and preferences dates back at least to the psychophysics work of German physiologist E. H. Weber in the nineteenth century. A progression from the way humans perceive differences between physical stimuli to more general analysis of human preferences has occurred particularly in the context of investigating consumer behavior—dealing with questions like whether there is a definite preference for a food with a particular type of ingredient, for example—and this is now a fully fledged area of research. Sources like Lawless and Heymann (2010) give overviews of the field and relevant statistical techniques. The earliest generally cited models for pairwise comparisons are the Thurstone model (Thurstone 1927) and the closely related Bradley-Terry (BT) model (Bradley and Terry 1952); these have connections to the IRT models, widely used in analyzing responses to questionnaires, which Hopkins and May (2013) drew on. Here we only look at BT models.
In a basic BT model, the probability that object j (Oj) is preferred to object k (Ok) from a set of J objects in a particular pairwise comparison jk is given by p(Oj Ok |πj,πk) =
πj πj+πk
for all j = k, where πj and πk are non-negative “worth” parameters describing the location of the object on the scale of preferences for some attribute. For n objects, there will be ( n 2 ) pairwise comparisons.
Log-Linear Models. It is now standard to fit BT models as log-linear models (Agresti 2007, for example), which allows them to be treated in a uniform way with much of modern statistical analysis. Log-linear models are a variety of generalized linear models (GLM), as is, for example, the logistic regression used throughout NLP. GLMs consist of a random component that identifies the response variable Y and selects a probability distribution for it; a systematic component that specifies some linear combination of the explanatory variables xi; and a link function g(·) applied to the mean μ of Y relating μ to this linear combination. They thus have the form g(μ) = α+ β1x1 + . . .+ βkxk. For log-linear models, the response variables are counts that are assumed to follow a Poisson distribution, and the link function is g(μ) = log(μ) (compare logistic regression’s g(μ) = log μ1−μ ). As an example, Y might be counts of people who hold some belief, and the various xi might be gender, socioeconomic status, and so forth. GLMs are a key tool for modern categorical data analysis, Agresti (2007, p. 65) noting that using models rather than the non-parametric approaches of Section 3 has several benefits:
The structural form of the model describes the patterns of association and interaction. The sizes of the model parameters determine the strength and importance of the effects. Inferences about the parameters evaluate which explanatory variables affect the response variable Y, while controlling effects of possible confounding variables. Finally, the model’s predicted values smooth the data and provide improved estimates of the mean of Y at possible explanatory variable values.
In a log-linear model, intuitive log-odds interpretations of making one response relative to another can be derived from the parameters. (Typically, software chooses a reference parameter and other parameter values are relative to that.) Statistical significance scores and standard errors can be calculated for these parameters. In addition, GLMs allow for testing of model fit. There are various model choices (e.g., should we include ties? should we include terms representing interactions?) and goodnessof-fit tests can assess the alternatives (see, e.g., Agresti 2007, Section 7.2.1). The model with a separate parameter for each cell in the associated contingency table is called the saturated model, and fits the data perfectly, making it a suitable comparator for alternatives. Deviance is a likelihood ratio statistic comparing a proposed model to the saturated one, allowing a test of the hypothesis that parameters not included in the model are zero, via goodness of fit tests; large test statistics and small p-values provide evidence of model lack of fit.
Models with Ties. To set out the representation of LLBT models, we follow the formulation of Dittrich and Hatzinger (2009). Let n(jk) be the number of comparisons between objects j and k; and let Y(jk)j be the number of preferences for object j with respect to k (similarly, Y(jk)k). The outcome of a paired comparison experiment can also be regarded as a (J 2
)× J incomplete two-dimensional contingency table: There are(J 2 ) rows of pairwise comparisons, and J columns recording choices of the jth object. As with log-linear models in general, the distribution of random variables Y(jk)j and Y(jk)k is assumed to be Poisson. Conditional on fixed n(jk) = Y(jk)j + Y(jk)k, (Y(jk)j, Y(jk)k)
follow a binomial (more generally, multinomial) distribution. The expected number of preferences of object j with respect to object k is denoted m(jk)j and given by n(jk)p(jk)j, with p(jk)j the binomial probability. So far this is only for binary preferences; there are various ways to account for ties. We describe the approach of Davidson and Beaver (1977), which appears quite widely used, where there is a common null preference effect for all pairwise comparisons. Then
log m(jk)j = μ(jk) + λOj − λOk log m(jk)0 = μ(jk) + γ log m(jk)k = μ(jk) − λOj + λOk
(1)
where the μ’s are “nuisance” parameters that fix the n(jk) marginal distributions, and the λO’s represent object parameters, m(jk)0 is the expected number of null preferences for pair (jk), and γ is the undecided effect. The object parameters are related to the worth parameters of the original definition by logπ = 2λO: These represent the log-odds.
In addition to the theoretical reasons for using LLBTs for modeling pairwise comparisons, a key benefit is the availability of packages in R for doing the modeling. Two candidates allowing a variety of sophisticated models are by Turner and Firth (2012) and Hatzinger and Dittrich (2012); we use the latter as the current version of the former does not handle ties. We first apply the model described by Equations (1) to the single pairwise data with ties from Section 2 using R. We refer the reader to the associated data bundle1 for the full output; we only excerpt it in the discussion below. Immediately following is a snippet of the R output for the ModPref data from Figure 1. o1 is the variable λOj for the + category, o2 for the – category, g1 for the null preferences.
Estimate Std. Error z value Pr(>|z|) o1 0.2778 0.1060 2.620 0.0088 ** o2 0.0000 NA NA NA g1 -0.6551 0.2300 -2.848 0.0044 ** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 Residual deviance: -6.6614e-15 on 0 degrees of freedom
In the R output, o2 is the reference object, with parameter value set to zero; the negative value of the estimate for g1 combined with its statistical significance says that there is a strong tendency for an expression of preference. The positive value of the o1 parameter and its significance indicate that the + group is strongly preferred: The odds in favor of this group with respect to the – group is exp(2 × 0.2778) = 1.74 to 1. Relating this to the description of the data in Section 2, then, there is a strong preference for translations by the proposed system relative to the baseline, even taking into account null preferences. The LLBT model confirms that even small data sets like this can produce meaningful and statistically significant results. For the other artificial preference data of Figure 1, the parameters behave as expected: for EqualPref, parameter estimates are all zero, signifying that they all have the same odds; for NoPref, the positive g1 indicates a strong tendency towards no preference; for StrongPref, the negative g1 indicates a strong tendency towards some preference, but with + or – equally likely. Note that all of these are saturated models: there are three objects and three parameters, so the model fits perfectly (indicated also by zero residual deviance). When we apply them
1 Data files and all R commands and output are at https://purl.org/NET/cl-llbt-data.
to the real count data of Figure 1 (c), the results indicate that for the Collins et al. data there is a weak to moderate tendency not to choose (g1 estimate 0.303, p = 0.0432), but, given that, there is a significant (0.0001) preference in favor of the reordered system. For the Lewis and Steedman results, the model gives similar results, albeit with a much stronger disposition to null preferences. In the data bundle we also carry out the sign test ignoring ties for each data set for comparison; it gives the same results in each case for the relation of + than –, but does not allow an evaluation of the effect of ties.
We now apply the model described by Equations (1) to the multiple pairwise data of Table 1. In the R output, the four systems A, B, C, D correspond to objects o1, o2, o3, o4, and g1 again to null preferences. As per the overview of the MT data in Section 2, there is little undecidedness (large negative g1). The coefficients show that object o1 (system A) is most preferred, followed by o4 (D), then o2 (B) and o3 (C). Note also that in this case, the model is not saturated: There is a non-zero residual deviance. As mentioned, log-linear models can be compared in terms of goodness of fit: Dittrich, Hatzinger, and Katzenbeisser (1998) and Dittrich and Hatzinger (2009) discuss this in some detail for LLBT models. Chi-squared statistics can be used to assess goodness of fit based on the residual deviance; the degrees of freedom (d.f.) equal the number of cell counts minus the number of model parameters; both deviance and d.f. are given in the R output. For this data deviance is 30.646 on 8 d.f., whereas by contrast if the ties (g1) are left out, it is 221.22 on 9 d.f. A chi-squared test would establish the goodness of fit for each model; but even without consulting the test it can be seen that leaving out the one parameter related to ties (1 d.f.) gives a seven-fold increase in deviance, so clearly inclusion of ties produces a much better model.
Introducing Subject Covariates. The model can also incorporate a range of other factors, a possibility not easily open to non-parametric methods. The one we look at here is the notion of a categorical covariate, introduced into LLBT models in Dittrich, Hatzinger, and Katzenbeisser (1998): This allows the objects (items) to vary with characteristics of the subject (judge). Many types of subject covariates could be added, grouping subjects by native language of the speaker, source of judges (e.g., Mechanical Turk, university), and so forth. Here we add just one, the identity of the subject. (Typically in a GLM this would be a random effect; we treat it as a covariate just for our simple illustration.) We define our categorical covariate S to have levels l, l = 1, 2, . . . , L. Let m(jk)j|l be the expected number of preferences for object j with respect to object k for subjects in covariate class l. The log-linear representation is then as follows:
log m( jk)j|l = μ(jk)l + λOj − λOk + λSl + λOSjl − λOSkl log m( jk)0|l = μ(jk)l + λSl + γ log m( jk)k|l = μ(jk)l − λOj + λOk + λSl − λOSjl + λOSkl
(2)
As do Dittrich and Hatzinger (2009), we define a reference group, with the λOj ’s representing the ordering for that group; the orderings for other groups are obtained by adding the λOSjl ’s specific to group l to the λ O l ’s for the reference group. μ(jk)l and λ S l are again “nuisance” parameters, the latter representing the main effect of the subject covariate measured on the lth level; λOSjl ’s are the (useful) subject-object interaction parameters describing the effect of the subject covariate on the preference for object j (similarly λOSkl and object k). We apply the model described by Equations (2) to the multiple pairwise data, with the subject covariate SUBJ with four levels (one per judge Ji of Table 1). There are a few complexities in interpreting the output, beyond the scope
of this article to discuss but covered in Dittrich, Hatzinger, and Katzenbeisser (1998). The broad interpretations to draw from the output are that interactions o1:SUBJ3 and o2:SUBJ3 are large and significant, and contribute to the model, unlike any others. These correspond to the different pairwise rankings given by judge J3 to system A (relative to D) and to B (relative to C): This is how subject effects are indicated in these LLBT models.
There are many other extensions to these models. Cattelan (2012) gives a state-ofthe-art overview of such extensions across a range of approaches, with an emphasis on dependent data. We only note two extensions here that are incorporated into prefmod and relevant to NLP. With categorical object covariates, items can be grouped as well, to investigate effects of grouping there, for example, different origins for translation sources. With non-pairwise rankings, judges can rank over more than two elements, as in the standard WMT evaluations, although this needs a special treatment in the models. 6 conclusions :We have looked at the sort of (pairwise) preference data that is encountered often in NLP. A particular characteristic of NLP data is that ties or undecided results may be frequent, and there is often a concern with inter-judge consistency. Reviewing classical non-parametric approaches, we note the opinion that it is important to model ties, and also note that approaches to looking at subject (judge) effects have several issues, such as a lack of quantitative interpretation of results. Among NLP approaches, especially within MT, new techniques are still being derived, which could benefit from views from outside the field. What we present are techniques from the field of sensory preference evaluation, where there has been a long history of development by statistics researchers. Recently, log-linear models have attracted attention. Applying them to sample data, we find that they provide the sort of information and uniform framework for analysis that NLP researchers could find useful. Given both extensive theoretical underpinings and freely available statistical software, we recommend LLBT models as a potential tool. Human evaluation plays an important role in NLP, often in the form of preference judgments. Although there has been some use of classical non-parametric and bespoke approaches to evaluating these sorts of judgments, there is an entire body of work on this in the context of sensory discrimination testing and the human judgments that are central to it, backed by rigorous statistical theory and freely available software, that NLP can draw on. We investigate one approach, Log-Linear Bradley-Terry models, and apply it to sample NLP data. [{""affiliations"": [], ""name"": ""Mark Dras""}] SP:dea01ec1a2d1ac7f75876dd5e599522271536f62 [{""authors"": [""Agresti"", ""Alan.""], ""title"": ""An Introduction to Categorical Data Analysis"", ""venue"": ""John Wiley, 2nd edition."", ""year"": 2007}, {""authors"": [""Bi"", ""Jian.""], ""title"": ""Sensory Discrimination Tests and Measurements: Statistical Principles, Procedures and Tables"", ""venue"": ""Blackwell, Oxford, UK."", ""year"": 2006}, {""authors"": [""Bojar"", ""Ond\u0159ej"", ""Milo\u0161 Ercegov\u010devi\u0107"", ""Martin Popel"", ""Omar Zaidan.""], ""title"": ""A grain of salt for the WMT manual evaluation"", ""venue"": ""Proceedings of WMT, pages 1\u201311."", ""year"": 2011}, {""authors"": [""Bradley"", ""Ralph"", ""Milton Terry.""], ""title"": ""Rank analysis of incomplete block designs, I"", ""venue"": ""The method of paired comparisons. Biometrika, 39:324\u2013345."", ""year"": 1952}, {""authors"": [""Callison-Burch"", ""Chris"", ""Cameron Fordyce"", ""Philipp Koehn"", ""Christof Monz"", ""Josh Schroeder.""], ""title"": ""Meta-) Evaluation of machine translation"", ""venue"": ""Proceedings of WMT, pages 136\u2013158."", ""year"": 2007}, {""authors"": [""Josh Schroeder""], ""title"": ""Further meta-evaluation of machine translation"", ""venue"": ""In Proceedings of WMT,"", ""year"": 2008}, {""authors"": [""Cattelan"", ""Manuela.""], ""title"": ""Models for paired comparison data: A review with emphasis on dependent data"", ""venue"": ""Statistical Science, 27(3):412\u2013423."", ""year"": 2012}, {""authors"": [""Collins"", ""Michael"", ""Philipp Koehn"", ""Ivona Kucerova.""], ""title"": ""Clause restructuring for statistical machine translation"", ""venue"": ""Proceedings of ACL, pages 531\u2013540."", ""year"": 2005}, {""authors"": [""R.R. Davidson"", ""R.J. Beaver.""], ""title"": ""On extending the Bradley-Terry model to incorporate within-pair order effects"", ""venue"": ""Biometrics, 33:693\u2013702."", ""year"": 1977}, {""authors"": [""Dittrich"", ""Regina"", ""Reinhold Hatzinger.""], ""title"": ""Fitting loglinear Bradley-Terry models (LLBT) for paired comparisons using the R package prefmod"", ""venue"": ""Psychological Science Quarterly,"", ""year"": 2009}, {""authors"": [""Dittrich"", ""Regina"", ""Reinhold Hatzinger"", ""W. Katzenbeisser""], ""title"": ""Modelling the effect of subject-specific covariates in paired comparison studies with an application to university rankings"", ""year"": 1998}, {""authors"": [""Francis"", ""Brian"", ""Regina Dittrich"", ""Reinhold Hatzinger""], ""title"": ""Modeling heterogeneity in ranked responses by nonparametric maximum likelihood: How do Europeans"", ""year"": 2010}, {""authors"": [""Hatzinger"", ""Reinhold"", ""Regina Dittrich.""], ""title"": ""prefmod: A package for modeling preferences based on paired comparisons, rankings, or ratings"", ""venue"": ""Journal of Statistical Software, 48(10):1\u201331."", ""year"": 2012}, {""authors"": [""Hopkins"", ""Mark"", ""Jonathan May.""], ""title"": ""Models of translation competitions"", ""venue"": ""Proceedings of ACL, pages 1,416\u20131,424."", ""year"": 2013}, {""authors"": [""Koehn"", ""Philipp.""], ""title"": ""Simulating human judgment in machine translation evaluation campaigns"", ""venue"": ""Proceedings of IWSLT, pages 179\u2013184."", ""year"": 2012}, {""authors"": [""Lawless"", ""Harry T."", ""Hildegarde Heymann.""], ""title"": ""Sensory Evaluation of Food: Principles and Practices"", ""venue"": ""Springer, New York, NY 2nd edition."", ""year"": 2010}, {""authors"": [""Lewis"", ""Mike"", ""Mark Steedman""], ""title"": ""Unsupervised induction of cross-lingual"", ""year"": 2013}, {""authors"": [""Lopez"", ""Adam.""], ""title"": ""Putting human assessments of machine translation systems in order"", ""venue"": ""Proceedings of WMT, pages 1\u20139."", ""year"": 2012}, {""authors"": [""Pado"", ""Sebastian"", ""Michel Galley"", ""Dan Jurafsky"", ""Christopher D. Manning.""], ""title"": ""Robust machine translation evaluation with entailment features"", ""venue"": ""Proceedings of ACL / AFNLP, pages 297\u2013305."", ""year"": 2009}, {""authors"": [""Randles"", ""Ronald H.""], ""title"": ""On Neutral Responses (zeros) in the sign test and ties in the Wilcoxon-Mann-Whitney test"", ""venue"": ""The American Statistician, 55(2):96\u2013101."", ""year"": 2001}, {""authors"": [""J.C.W. Rayner"", ""D.J. Best.""], ""title"": ""A Contingency Table Approach to Nonparametric Testing"", ""venue"": ""Chapman and Hall/CRC, Boca Raton, FL."", ""year"": 2001}, {""authors"": [""Sprent"", ""Peter"", ""Nigel Smeeton.""], ""title"": ""Applied Nonparametric Statistical Methods"", ""venue"": ""Chapman and Hall, London, UK."", ""year"": 2007}, {""authors"": [""L.L. Thurstone""], ""title"": ""A law of comparative Judgement"", ""venue"": ""Psychological Review, 34:278\u2013286."", ""year"": 1927}, {""authors"": [""Turner"", ""Heather"", ""David Firth.""], ""title"": ""Bradley-Terry Models in R: The BradleyTerry2 Package"", ""venue"": ""Journal of Statistical Software, 48(9):1\u201321."", ""year"": 2012}, {""authors"": [""Vilar"", ""David"", ""Gregor Leusch"", ""Hermann Ney"", ""Rafael E. Banchs.""], ""title"": ""Human evaluation of machine translation through binary system comparisons"", ""venue"": ""Proceedings of WMT, pages 96\u2013103."", ""year"": 2007}]",
,,,,,,,,"The Web is the main source of data in modern computational linguistics. Other volumes in the same series, for example, Introductions to Opinion Mining (Liu 2012) and Semisupervised Machine Learning (Søgaard 2013), start their problem statements by referring to data from the Web. This volume starts its own introduction by praising Web corpora for their size, ease of construction, and availability as a source of new text types. A random check of papers from the most recent ACL meeting also shows that the majority of them use Web data in one way or another. Our field definitely needs a comprehensive overview and a DIY manual for the task of constructing a corpus from the Web. This book is, to the best of my knowledge, the first attempt at providing such an overview.","[{""affiliations"": [], ""name"": ""Roland Sch\u00e4fer""}, {""affiliations"": [], ""name"": ""Felix Bildhauer""}]",SP:661b841860b253255a303d552a5d5d1396e9824e,"[{""authors"": [""Baeza-Yates"", ""Ricardo"", ""Carlos Castillo"", ""Efthimis N. Efthimiadis.""], ""title"": ""Characterization of national Web domains"", ""venue"": ""ACM Transactions on Internet Technology (TOIT), 7(2):9."", ""year"": 2007}, {""authors"": [""Baroni"", ""Marco"", ""Silvia Bernardini"", ""Adriano Ferraresi"", ""Eros Zanchetta.""], ""title"": ""The WaCky wide Web: A collection of very large linguistically processed Web-crawled corpora"", ""venue"": ""Language Resources and Evaluation,"", ""year"": 2009}, {""authors"": [""Baroni"", ""Marco"", ""Francis Chantree"", ""Adam Kilgarriff"", ""Serge Sharoff""], ""title"": ""Cleaneval: A competition for cleaning"", ""year"": 2008}, {""authors"": [""Studies. Springer"", ""Berlin/New York. S\u00f8gaard"", ""Anders""], ""title"": ""Sentiment Analysis and Opinion Mining. Synthesis Lectures on Human Language Technologies"", ""year"": 2013}]",,,,,,,,,,,,,,,,,,,"reviewed by serge sharoff university of leeds :The Web is the main source of data in modern computational linguistics. Other volumes in the same series, for example, Introductions to Opinion Mining (Liu 2012) and Semisupervised Machine Learning (Søgaard 2013), start their problem statements by referring to data from the Web. This volume starts its own introduction by praising Web corpora for their size, ease of construction, and availability as a source of new text types. A random check of papers from the most recent ACL meeting also shows that the majority of them use Web data in one way or another. Our field definitely needs a comprehensive overview and a DIY manual for the task of constructing a corpus from the Web. This book is, to the best of my knowledge, the first attempt at providing such an overview.
© 2015 Association for Computational Linguistics
The book consists of an introduction and four chapters outlining the four main steps of Web corpus construction. They include: “Data Collection” (Chapter 2), “Basic Corpus Cleaning” (Chapter 3), “Linguistic Processing” (Chapter 4), and “Corpus Evaluation” (Chapter 5).
Chapter 2 provides a very useful outline of the main properties of the Web and the crawling strategies. The chapter starts with an overview of a large-scale study of Web connectivity from Baeza-Yates, Castillo, and Efthimiadis (2007), listing various parameters of connectivity for a range of Top-Level Domains. However, there is little discussion of the implications for the corpus development task; for example, does the difference of the in-degree parameter of the Web pages from Chile and the UK have any implications for the Web corpora crawled from those domains? The chapter then proceeds to another important topic, which concerns the parameters of crawling; for example, the crawl bias and the number of seeds, and their influence on the final corpus. Section 2.4.1 illustrates the problems with the crawl bias by an example of deWac, a large commonly used corpus of German (Baroni et al. 2009). The second most frequent proper name bigram in this corpus is found to be Falun Gong. However, more analysis into the nature of the bias should have been beneficial. It is less likely to be related to the PageRank bias, the main bias discussed in Section 2.4.2. Other most frequent bigrams from deWac are not presented in the book, but it is interesting to note that the fourth place in it is occupied by Hartz IV, and the tenth place by Digital Eyes. This suggests that the bias comes from frequency spikes (i.e, a large number of instances collected from a small number of Web sites). Another shortcoming of this chapter is that nothing is said specifically about obtaining data from such resources as Twitter or Facebook, which need access via APIs rather than direct crawling.
doi:10.1162/COLI r 00214
Computational Linguistics Volume 41, Number 1
Chapter 3 introduces methods for basic cleaning of the corpus content, such as processing of text formats (primarily HTML tags), language identification, boilerplate removal, and deduplication. Such low-level tasks are not considered to be glamorous from the view of computational linguistics, but they are extremely important for making Web-derived corpora usable (Baroni et al. 2008). The introduction offered in this chapter is reasonably complete, with good explanations of the sources of problems as well as with suggestions for the tools to be used in each task. An important bit which is missing in this chapter concerns the suggestions for choosing a particular cleaning pipeline. Although the choice indeed depends on the purposes of corpus collection, an indication of which pipeline suits which purpose is desirable.
Chapter 4 is devoted to basic steps for linguistic processing of Web corpora, such as tokenization, POS tagging, and lemmatization, as well as orthographic normalization. Even though the processing pipeline is roughly the same for all NLP tasks, it becomes harder for Web corpora because they exhibit greater diversity in comparison with more homogeneous text collections (e.g., WSJ texts). Web texts are also considerably noisier, in the sense of containing nonstandard linguistic expressions, which are likely to be a challenge to the tools trained on more standard texts. The chapter presents some interesting case studies—in particular, the sources of POS tagging errors and nonstandard orthography.
Chapter 5 describes ways for evaluating and comparing corpora. It gives examples of checking for word and sentence length and for sentence-level duplication. It also introduces methods for comparing frequency lists. Like other chapters it includes many interesting observations, such as the methods for extrinsic evaluation of corpora. However, the chapter does not address many issues important for corpus evaluation and comparison. Given that the previous chapters introduced a number of pipelines and corpora, this chapter would have been an ideal place to illustrate all the aspects of the pipelines by evaluating them in a consistent way. There are occasional references to this goal, such as the frequency lists of French nouns in Section 5.3.1, but this particular comparison is fairly impressionistic, and it concludes with a declaration of basic similarity of the underlying corpora. Does this mean that the crawling, cleaning, and linguistic processing pipelines do not matter? In any case, not even an impressionistic comparison of the pipelines is performed for other evaluation methods. Some illustrations are also not informative (e.g., Table 5.1.1 shows two frequency lists with the identical ranks for their words, which leads to the trivial rank correlation value of 1). The chapter contains a single paragraph devoted to composition of Web corpora. Given the size of such corpora, their evaluation crucially depends on understanding what has been crawled. The task has been approached by a number of models, such as supervised and semisupervised classification, clustering, topic modeling, and so forth, which should have been included in the discussion. The discussion does contain a relevant reference to Mehler, Sharoff, and Santini (2010), which surveys approaches to the genres of the Web, but other aspects of corpus composition need to be addressed, too.
Overall, it is very useful to have a book that introduces all the aspects of Web corpus construction in a single volume with coherent presentation. The volume under review does cover the entire range of topics relevant to Web corpus construction and illustrates them via numerous examples. I would recommend it to students just starting their corpus development experiments. As for the drawbacks of the volume, there is a need to improve the structure of argumentation for the next edition. Bits of information are sometimes introduced in an incomplete way and re-introduced again in subsequent sections. For examples, two tools for crawling are discussed towards the end of Section 2.3.3, while more tools are mentioned as the discussion of crawling
162
Book Reviews
strategies progresses. Chapter 1 starts with a fairly random list of non-Web corpora, whereas an overview of the book structure is confined to a short paragraph. Often, frustratingly little information is provided besides an annotated bibliography, rather than a presentation of the relevant methods and issues. In some cases this is accompanied with a statement that “covering this topic is beyond the scope of this volume,” even if the nature of the problem and the solutions could have been easily explained in a one-page summary. Another minor concern is an (understandable) emphasis on the tools and corpora developed by the authors, primarily on their German corpus.
I have to admit ambivalence in my final verdict: The book is a useful introduction to an important topic, but it definitely warrants a new edition, which eliminates the shortcomings of the current one.",,,,,,,,,,,"web corpus construction :Roland Schäfer and Felix Bildhauer (Freie Universität Berlin)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 22), 2013, 145 pages, paper-bound, ISBN 9781608459834, doi:10.2200/S00508ED1V01Y201305HLT022",,,,,,,,,,,,,,,,,,,,"The Web is the main source of data in modern computational linguistics. Other volumes in the same series, for example, Introductions to Opinion Mining (Liu 2012) and Semisupervised Machine Learning (Søgaard 2013), start their problem statements by referring to data from the Web. This volume starts its own introduction by praising Web corpora for their size, ease of construction, and availability as a source of new text types. A random check of papers from the most recent ACL meeting also shows that the majority of them use Web data in one way or another. Our field definitely needs a comprehensive overview and a DIY manual for the task of constructing a corpus from the Web. This book is, to the best of my knowledge, the first attempt at providing such an overview. [{""affiliations"": [], ""name"": ""Roland Sch\u00e4fer""}, {""affiliations"": [], ""name"": ""Felix Bildhauer""}] SP:661b841860b253255a303d552a5d5d1396e9824e [{""authors"": [""Baeza-Yates"", ""Ricardo"", ""Carlos Castillo"", ""Efthimis N. Efthimiadis.""], ""title"": ""Characterization of national Web domains"", ""venue"": ""ACM Transactions on Internet Technology (TOIT), 7(2):9."", ""year"": 2007}, {""authors"": [""Baroni"", ""Marco"", ""Silvia Bernardini"", ""Adriano Ferraresi"", ""Eros Zanchetta.""], ""title"": ""The WaCky wide Web: A collection of very large linguistically processed Web-crawled corpora"", ""venue"": ""Language Resources and Evaluation,"", ""year"": 2009}, {""authors"": [""Baroni"", ""Marco"", ""Francis Chantree"", ""Adam Kilgarriff"", ""Serge Sharoff""], ""title"": ""Cleaneval: A competition for cleaning"", ""year"": 2008}, {""authors"": [""Studies. Springer"", ""Berlin/New York. S\u00f8gaard"", ""Anders""], ""title"": ""Sentiment Analysis and Opinion Mining. Synthesis Lectures on Human Language Technologies"", ""year"": 2013}] reviewed by serge sharoff university of leeds :The Web is the main source of data in modern computational linguistics. Other volumes in the same series, for example, Introductions to Opinion Mining (Liu 2012) and Semisupervised Machine Learning (Søgaard 2013), start their problem statements by referring to data from the Web. This volume starts its own introduction by praising Web corpora for their size, ease of construction, and availability as a source of new text types. A random check of papers from the most recent ACL meeting also shows that the majority of them use Web data in one way or another. Our field definitely needs a comprehensive overview and a DIY manual for the task of constructing a corpus from the Web. This book is, to the best of my knowledge, the first attempt at providing such an overview.
© 2015 Association for Computational Linguistics
The book consists of an introduction and four chapters outlining the four main steps of Web corpus construction. They include: “Data Collection” (Chapter 2), “Basic Corpus Cleaning” (Chapter 3), “Linguistic Processing” (Chapter 4), and “Corpus Evaluation” (Chapter 5).
Chapter 2 provides a very useful outline of the main properties of the Web and the crawling strategies. The chapter starts with an overview of a large-scale study of Web connectivity from Baeza-Yates, Castillo, and Efthimiadis (2007), listing various parameters of connectivity for a range of Top-Level Domains. However, there is little discussion of the implications for the corpus development task; for example, does the difference of the in-degree parameter of the Web pages from Chile and the UK have any implications for the Web corpora crawled from those domains? The chapter then proceeds to another important topic, which concerns the parameters of crawling; for example, the crawl bias and the number of seeds, and their influence on the final corpus. Section 2.4.1 illustrates the problems with the crawl bias by an example of deWac, a large commonly used corpus of German (Baroni et al. 2009). The second most frequent proper name bigram in this corpus is found to be Falun Gong. However, more analysis into the nature of the bias should have been beneficial. It is less likely to be related to the PageRank bias, the main bias discussed in Section 2.4.2. Other most frequent bigrams from deWac are not presented in the book, but it is interesting to note that the fourth place in it is occupied by Hartz IV, and the tenth place by Digital Eyes. This suggests that the bias comes from frequency spikes (i.e, a large number of instances collected from a small number of Web sites). Another shortcoming of this chapter is that nothing is said specifically about obtaining data from such resources as Twitter or Facebook, which need access via APIs rather than direct crawling.
doi:10.1162/COLI r 00214
Computational Linguistics Volume 41, Number 1
Chapter 3 introduces methods for basic cleaning of the corpus content, such as processing of text formats (primarily HTML tags), language identification, boilerplate removal, and deduplication. Such low-level tasks are not considered to be glamorous from the view of computational linguistics, but they are extremely important for making Web-derived corpora usable (Baroni et al. 2008). The introduction offered in this chapter is reasonably complete, with good explanations of the sources of problems as well as with suggestions for the tools to be used in each task. An important bit which is missing in this chapter concerns the suggestions for choosing a particular cleaning pipeline. Although the choice indeed depends on the purposes of corpus collection, an indication of which pipeline suits which purpose is desirable.
Chapter 4 is devoted to basic steps for linguistic processing of Web corpora, such as tokenization, POS tagging, and lemmatization, as well as orthographic normalization. Even though the processing pipeline is roughly the same for all NLP tasks, it becomes harder for Web corpora because they exhibit greater diversity in comparison with more homogeneous text collections (e.g., WSJ texts). Web texts are also considerably noisier, in the sense of containing nonstandard linguistic expressions, which are likely to be a challenge to the tools trained on more standard texts. The chapter presents some interesting case studies—in particular, the sources of POS tagging errors and nonstandard orthography.
Chapter 5 describes ways for evaluating and comparing corpora. It gives examples of checking for word and sentence length and for sentence-level duplication. It also introduces methods for comparing frequency lists. Like other chapters it includes many interesting observations, such as the methods for extrinsic evaluation of corpora. However, the chapter does not address many issues important for corpus evaluation and comparison. Given that the previous chapters introduced a number of pipelines and corpora, this chapter would have been an ideal place to illustrate all the aspects of the pipelines by evaluating them in a consistent way. There are occasional references to this goal, such as the frequency lists of French nouns in Section 5.3.1, but this particular comparison is fairly impressionistic, and it concludes with a declaration of basic similarity of the underlying corpora. Does this mean that the crawling, cleaning, and linguistic processing pipelines do not matter? In any case, not even an impressionistic comparison of the pipelines is performed for other evaluation methods. Some illustrations are also not informative (e.g., Table 5.1.1 shows two frequency lists with the identical ranks for their words, which leads to the trivial rank correlation value of 1). The chapter contains a single paragraph devoted to composition of Web corpora. Given the size of such corpora, their evaluation crucially depends on understanding what has been crawled. The task has been approached by a number of models, such as supervised and semisupervised classification, clustering, topic modeling, and so forth, which should have been included in the discussion. The discussion does contain a relevant reference to Mehler, Sharoff, and Santini (2010), which surveys approaches to the genres of the Web, but other aspects of corpus composition need to be addressed, too.
Overall, it is very useful to have a book that introduces all the aspects of Web corpus construction in a single volume with coherent presentation. The volume under review does cover the entire range of topics relevant to Web corpus construction and illustrates them via numerous examples. I would recommend it to students just starting their corpus development experiments. As for the drawbacks of the volume, there is a need to improve the structure of argumentation for the next edition. Bits of information are sometimes introduced in an incomplete way and re-introduced again in subsequent sections. For examples, two tools for crawling are discussed towards the end of Section 2.3.3, while more tools are mentioned as the discussion of crawling
162
Book Reviews
strategies progresses. Chapter 1 starts with a fairly random list of non-Web corpora, whereas an overview of the book structure is confined to a short paragraph. Often, frustratingly little information is provided besides an annotated bibliography, rather than a presentation of the relevant methods and issues. In some cases this is accompanied with a statement that “covering this topic is beyond the scope of this volume,” even if the nature of the problem and the solutions could have been easily explained in a one-page summary. Another minor concern is an (understandable) emphasis on the tools and corpora developed by the authors, primarily on their German corpus.
I have to admit ambivalence in my final verdict: The book is a useful introduction to an important topic, but it definitely warrants a new edition, which eliminates the shortcomings of the current one. web corpus construction :Roland Schäfer and Felix Bildhauer (Freie Universität Berlin)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 22), 2013, 145 pages, paper-bound, ISBN 9781608459834, doi:10.2200/S00508ED1V01Y201305HLT022",
"1 introduction :A constancy measure for a natural language text is defined, in this article, as a computational measure that converges to a value for a certain amount of text and remains invariant for any larger size. Because such a measure exhibits the same value for any size of text larger than a certain amount, its value could be considered as a text characteristic.
The concept of such a text constancy measure was introduced by Yule (1944) in the form of his measure K. Since Yule, there has been a continuous quest for such measures, and various formulae have been proposed. They can be broadly categorized into three types, namely, those measuring (1) repetitiveness, (2) power law character, and (3) complexity.
∗ Kyushu University, 744 Motooka Nishiku, Fukuoka City, Fukuoka, Japan. E-mail: kumiko@ait.kyushu-u.ac.jp.
∗∗ JST-PRESTO, 4-1-8 Honcho, Kawaguchi, Saitama 332-0012, Japan. † Gunosy Inc., 6-10-1 Roppongi, Minato-ku, Tokyo, Japan.
Submission received: 11 July 2013; revised version received: 17 February 2015; accepted for publication: 18 March 2015.
doi:10.1162/COLI a 00228
© 2015 Association for Computational Linguistics
Yule’s original intention for K’s utility lay in author identification, assuming that it would differ for texts written by different authors. State-of-the-art multivariate machine learning techniques are powerful, however, for solving such language engineering tasks, in which Yule’s K is used only as one variable among many, as reported in Stamatatos, Fakotakis, and Kokkinakis (2001) and Stein, Lipka, and Prettenhofer (2010).
We believe that constancy measures today, however, have greater importance in understanding the mathematical nature of language. Although mathematical models of language have been studied in the computational linguistics milieu, via Markov models (Manning and Schuetze 1999), Zipf’s law and its modifications (Mandelbrot 1953; Zipf 1965; Bell, Cleary, and Witten 1990), and Pitman-Yor models (Teh 2006) more recently, the true mathematical model of linguistic processes is ultimately unknown. Therefore, the convergence of a constancy measure must be examined through empirical verification. Because some constancy measures have a mathematical theory of convergence for a known process, discrepancies in the behavior of real linguistic data from such a theory would shed light on the nature of linguistic processes and give hints towards improving the mathematical models. Furthermore, as one application, a convergent measure would allow for comparison of different texts through a common, stable norm, provided that the measure converges for a sufficiently small amount of text. One of our goals is to discover a non-trivial measure with a certain convergence speed that distinguishes the different natures of texts.
The objective of this article is thus to provide a potential explanation of what the study of constancy measures over 70 years has been about, by answering the three following questions mathematically and empirically:
Question 1 Does a measure exhibit constancy?
Question 2 If so, how fast is the convergence speed?
Question 3 How discriminatory is the measure?
We seek answers by first showing the meaning of Yule’s K in relation to the Rényi higher-order entropy, and by then empirically examining constancy across large-scale texts of different kinds. We finally provide an application by considering the natures of two unknown scripts, the Voynich manuscript and Rongorongo, in order to show the possible utility of a constancy measure.
The most important and closest previous work was reported in Tweedie and Baayen (1998), the first paper to have examined the empirical behavior of constancy measures on real texts. The authors used English literary texts to test constancy measure candidates proposed prior to their work. Today, the coverage and abundance of language corpora allow us to conduct a larger-scale investigation across multiple languages. Recently, Golcher (2007) tested his measure V (discussed later in this paper) with IndoEuropean languages and also programming language sources. Our papers (Kimura and Tanaka-Ishii 2011, 2014) also precede this one, presenting results preliminary to this article but with only part of our data, and neither of those provides mathematical analysis with respect to the Rényi entropy. Compared with these previous reports, our contribution here can be summarized as follows:
Our work elucidates the mathematical relation of Yule’s K to Rényi’s higher-order entropy and explains why K converges.
Our work vastly extends the corpora used for empirical examination in terms of both size and language.
Our work compares the convergent values for these corpora. Our work also presents results for unknown language data, specifically from the Voynich manuscript and Rongorongo.
We start by summarizing the potential constancy measures proposed so far.","2 constancy measures :The measures proposed so far can broadly be categorized into three types, calculating the repetitiveness, power-law distribution, or complexity of text. This section mathematically analyzes these measures and summarizes them. The study of text constancy started with proposals for simple text measures of vocabulary repetitiveness. The representative example is Yule’s K (Yule 1944), while Golcher recently proposed V as another candidate (Golcher 2007).
2.1.1 Yule’s K. To the best of our knowledge, the oldest mention of constancy values was made by Yule with his notion of K (Yule 1944). Let N be the total number of words in a text, V(N) be the number of distinct words, V(m, N) be the number of words appearing m times in the text, and mmax be the largest frequency of a word. Yule’s K is then defined as follows, through the first and second moments of the vocabulary population distribution of V(m, N), where S1 = N = ∑ m mV(m, N), and S2 = ∑ m m
2V(m, N) (Yule 1944; Herdan 1964):
K = CS2 − S1 S21
= C [ − 1N + mmax∑ m=1 V(m, N)( mN ) 2 ]
(1)
where C is a constant enlarging of the value of K, defined by Yule as C = 104. K is designed to measure the vocabulary richness of a text: The larger Yule’s K, the less rich the vocabulary is. The formula can be intuitively understood from the main term of the sum in the formula. Because the square of ( mN )
2 indicates the degree of recurrence of a word, the sum of such degrees for all words is small if the vocabulary is rich, or large in the opposite case. Another simple example can be given in terms of S2 in this formula. Suppose a text is 10 words long: if each of the 10 tokens is distinct (high diversity), then S2 = 1 × 1 × 10 = 10; whereas, if each of the 10 tokens is identical (low diversity), then S2 = 10 × 10 × 1 = 100.
Measures that are slightly different but essentially equivalent to Yule’s K have appeared here and there. For example, Herdan defined Vm as follows (Herdan 1964, pp. 67, 79):
Vm = √√√√ mmax∑ m=1 V(m, N)( mN ) 2 − 1 V(N)
Likewise, Simpson (1949) derived the following formula as a measure to capture the diversity of a population:
D = mmax∑ m=1 V(m, N) mN m − 1 N − 1
which is equivalent to Yule’s K, as Simpson noted.
2.1.2 Other Measures Based on Simple Text Statistics. Apart from Yule’s K, various measures have been proposed from simple statistical observation of text, as detailed in Tweedie and Baayen (1998). One genre is based on the so-called token-type relation (i.e., the ratio of the vocabulary size V(N) and the text size N, in log) as formulated by Guiraud (1954) and Herdan (1964) as a law. Because this simple ratio is not stable, the measure was modified numerous times to formulate Herdan’s C (Herdan 1964), Dugast’s k and U (Dugast 1979), Maas’ a2 (Maas 1972), Tuldava’s LN (Tuldava 1977), and Brunet’s W (Brunet 1978).
Another genre of measures concerns the proportion of hapax legomena, that is V(1, N). Honoré noted that V(1, N) increases linearly with respect to the log of a text’s vocabulary size V(N) (Honoré 1979). Another ratio, of V(2, N) to V(N), was proposed as a text characteristic by Sichel (1975) and Maas (1972).
Each of these values, however, was found not to be convergent according to the extensive study conducted by Tweedie and Baayen (1998). In common with Yule’s intention to apply such measures for author identification, they examined all of the measures discussed here, in addition to two measures explained later: Orlov’s Z, and the Shannon entropy upper bound obtained from the relative frequencies of unigrams. They examined these measures with English novels (such as Alice’s Adventures in Wonderland) and empirically found that only Yule’s K and Orlov’s Z were convergent. Given their report, we consider K the only true candidate among the constancy measures examined so far.
2.1.3 Golcher’s V. Golcher’s V is a string-based measure calculated on the suffix tree of a text (Golcher 2007). Letting the length of the string be N and the number of inner nodes of the (Patricia) suffix tree (Gusfield 1997) be k, V is defined as:
V = kN (2)
Golcher empirically showed how this measure converges to almost the same value across Indo-European languages for about 30 megabytes of data. He also showed how the convergent values differ from those calculated for programming language texts.
Golcher explains in his paper that the possibility of constancy of V does not yet have mathematical grounding and has only been shown empirically. He does not report values for texts larger than about 30 megabytes nor for those of non-Indo-European languages. A simple conjecture on this measure is that because a suffix tree for a string of length N has at most N − 1 inner nodes, V must end up at some value 0 ≤ V < 1, for any given text.
Our group tested V with larger-scale data and concluded that V could be a constancy measure, although we admitted to observing a gradual increase (Kimura and Tanaka-Ishii 2014). Because V requires further verification on larger-scale data before ruling it out, we include it as a constancy measure candidate. Since Zipf (1965), power laws have been reported as an underlying statistical characteristic of text. The famous Zipf’s law is defined as:
f (n) ∝ n−γ (3)
where γ ≈ 1, and f (n) is the frequency of the nth most frequent word in a text. Various studies have sought to explain mathematically how the exponent could differ depending on the kind of text. To the best of our knowledge, however, there has been a limited number of reports related to text constancy.
An exception is the study on Orlov’s Z (Orlov and Chitashvili 1983). Orlov and Chitashvili attempted to obtain explicit mathematical forms for V(N) and V(m, N) by more finely considering the long tails of vocabulary distributions for which Zipf’s law does not hold. They obtained these forms through a parameter Z, defined as the potential text length minimizing the square error of the estimated V(m, N), with its actual value as follows:
Z = arg min N 1 mmax mmax∑ m=1 {E[V(m, N)] − V(m, N) V(N) }2 (4)
Thus defining Z, they mathematically deduced for V(N) the following formula:
V(N) = Z log(mmaxZ) N N − Z log ( N Z ) (5)
Two ways to obtain Z can be formulated through approximation: one through GoodTuring smoothing (Good 1953), which assumes Zipf’s law to hold, and the other using Newton’s method. Tweedie and Baayen showed how the value of Z is stable at the size of an English novel by a single author and thus suggested that it could form a text characteristic. The empirical results, however, were not significantly convergent with respect to text size, and, moreover, Tweedie and Baayen provided their results without giving an estimation method (Tweedie and Baayen 1998). Calculation using Good-Turing smoothing, which is derived directly from Zipf’s law, would cause Z to converge, but this does not take Orlov’s original intention into consideration. Alternatively, our group (Kimura and Tanaka-Ishii 2014) verified Z through Newton’s method by setting g(Z) = 0, where g(Z) is the following function:
g(Z) = Z log(mmaxZ) N N − Z log ( N Z ) − V(N) (6)
We also showed how the value of Z increases rapidly when the text size is larger than 10 megabytes.
The major problem with measures based on power laws lies in the skewed head and tail of the vocabulary population distribution. Because these exceptions constitute important parts of the population, parameter estimation by fitting to Equation (3) is sensitive to the estimation method. For example, the estimated value of the exponent for Zipf’s law depends on the method used for dealing with these exceptions. We tested several simple methods of estimating the Zipf law’s exponent γ with different ways of handling the head and tail of a distribution. There were settings that led to
convergence, but the convergence depended on the settings. Such difficulty could be one reason why there has been no direct proposal for γ as a text constancy measure. Hence, due care must be taken in relating text constancy to a power law. We chose another path by considering text constancy through a random Zipf distribution, as described later in the experimental section. With respect to measures based on complexity, multiple reports have already examined the Shannon entropy (Shannon 1948; Cover and Thomas 2006). In addition, we introduce the Rényi higher-order entropy (Rényi 1960) as another possible measure.
2.3.1 Shannon Entropy Upper Bound. Let X be the random variable of a sequence X = X1, X2, . . . , Xi, . . . , where Xi represents the ith element of X: Xi = x ∈ X, and where X represents a given set (e.g., a set of words or characters) whose members constitute the sequence. Let Xji (i < j) denote the random variable indicating its subsequence Xi, Xi+1, Xi+2, . . . , Xj. Let P(X) indicate the probability function of a sequence X. The Shannon entropy is then defined as:
H(X) = − ∑
X
P(X) log P(X) (7)
Tweedie and Baayen directly calculated an approximation of this formula in terms of the relative frequencies (for P) of unigrams (for X), and they concluded that the measure would continue increasing with respect to text size and would not converge for short, literary texts (Tweedie and Baayen 1998). Because we are interested in the measure’s behavior on a larger scale, we replicated their experiment, as discussed later in the section on empirical constancy. We denote this measure as H1 in this article.
Apart from that report, many have studied the entropy rate, defined as:
h∗ = lim n→∞ H(Xn1 ) n (8)
Theoretically, the behavior of the entropy rate with respect to text size has been controversial. On the one hand, there have been indications of entropy rate constancy (Genzel and Charniak 2002; Levy and Jaeger 2007). These reports argue that the entropy rate of natural language could be constant. Due to the inherent difficulty in obtaining the true value of h∗ from a text, however, these arguments are based only on indirect clues with respect to convergence. On the other hand, Hilberg conjectured a decrease in the human conditional entropy, as follows (Hilberg 1990):
H(Xn|Xn−11 ) ∝ n−1+β
He obtained this through an examination of Shannon’s original experimental data and suggested that β ≈ 0.5. From this formula, Dȩbowski induces that H(Xn1 ) ∝ nβ and that the entropy rate can be formulated generally as follows (Dȩbowski 2014):
H(Xn1 ) n ≈ An−1+β + h∗ (9)
Note that at the limit of n → ∞, this rate goes to h∗, a constant, provided that β < 1.0. Hilberg’s conjecture is deemed compatible with entropy rate constancy at its asymptotic limit, provided that h∗ > 0 holds.1 We are therefore interested in whether this h∗ forms a text characteristic, and if so, whether h∗ > 0.
Empirically, many have attempted to calculate the upper bound of the entropy rate. Brown’s report (Brown et al. 1992) is representative in showing a good estimation of the entropy rate for English from texts, as compared with values obtained from humans (Cover and King 1978). Subsequently, there have been important studies on calculating the entropy rate, as reported thoroughly in Schümann and Grassberger (1996). The questions related to h∗, however, remain unsolved. Recently, Dȩbowski used a Lempel-Ziv compressor and examined Hilberg’s conjecture for texts by single authors (Dȩbowski 2013). He showed an exponential decrease in the entropy rate with respect to text size, supporting the validity of Equation (9). Following these previous works, we examine the entropy rate by using an algorithm proposed by Grassberger (1989) and later on by Farach et al. (1995). This method is based on universal coding. The algorithm has a theoretical background of convergence to the true h∗, provided the sequence is stationary, but has been proved by Shields (1992) to be inconsistent—that is, it does not converge to the entropy rate for certain nonMarkovian processes. We still chose to apply this method, because it requires no arbitrary parameters for calculation and is applicable to large-scale data within a reasonable time.
The Grassberger algorithm (Grassberger 1989; Farach et al. 1995) can be summarized as follows. Consider a sequence X of length N. The maximum matching length Li is defined as:
Li = max{k : Xj+kj = Xi+ki }
for j ∈ {1, . . . , i − 1}, 1 ≤ j ≤ j + k ≤ i − 1. In other words, Li is the maximum common subsequence before and after i. If L̄ is the average length of Li, given by
L̄ = 1N i=N∑ i=1 Li
then the method obtains the entropy rate h1 as
h1 = log2N
L̄ (10)
Given the true entropy rate h∗, convergence has been mathematically proven for a stationary process, such that |h∗ − h1| = O(1) when N → ∞. In this article, we consider this entropy rate h1 as a constancy measure candidate.
1 According to Dȩbowski (2009), h∗ = 0 suggests that the next element of a linguistic process is deterministic, that is, a function of the corpus observed before, under the two conditions that (1) the number of possible choices for the element is finite, and (2) the corpus observed before is infinite. In reality, the finiteness of linguistic sequences has the opposite tendency (i.e., the size of the observed corpus is finite, and the possible vocabulary size is infinite).
2.3.2 Approximation of Rényi Entropy Hα. The Rényi entropy is a generalization of the Shannon entropy, defined as follows (Rényi 1960; Rényi 1970; Cover and Thomas 2006; Bromiley, Thacker, and Bouhova-Thacker 2010):
Hα(X) = 11 − α log( ∑
X
Pα(X)) (11)
where α ≥ 0,α = 1. Hα(X) represents different ideas of sequence complexity for different α. For example:
When α = 0, H0(X) indicates the number of distinct occurrences of X. When the limit α → 1 is taken, Equation (11) reduces to the Shannon entropy.
The formula for α = 0 becomes equivalent to the so-called topological entropy (hence, it is another notion of entropy) for certain probability functions (Kitchens 1998) (Cover and Thomas 2006). Note that the number of distinct tokens (i.e., the cardinality of a set) has been used widely as a rough approximation of complexity in computational linguistics. Indeed, in Section 2.1.2, we saw how some candidate constancy measures are based on a token-type relation, such that the number of types is related to the complexity of a text. For texts, note also that the value grows with respect to the text size, unless X is considered, for example, in terms of unigrams of a phonographic alphabet.
For α → 1, there is controversy regarding convergence, as noted in the previous section. Such difficulty in convergence for these α values lies in the nature of linguistic processes, in which the vocabulary set evolves.
This view motivates us to consider α > 1 for Hα(X), since the formula captures complexity by considering linguistic hapax legomena to a lesser degree, thus giving the possibility of convergence. In fact, an approximation of the probability by the relative frequencies of unigrams at α = 2 immediately shows the essential equivalence to Yule’s K, since K from Equation (1) can be rewritten as follows:
mmax∑ m=1 V(m, N)( mN ) 2 = ∑ x∈X ( freq(x) N ) 2
where freq(x) is the frequency of x ∈ X. Therefore, Yule’s K has significance within the context of complexity.
This relation of Yule’s K to the Rényi entropy H2 is reported for the first time here, to the best of our knowledge. This mathematical relation clarifies both why Yule’s K should converge and what the convergent value means; specifically, the value represents the gross complexity underlying the language system. As noted earlier, the higher-order entropy considers hapax legomena to a lesser degree and calculates the gross entropy only from the representative vocabulary population. This simple argument shows that Yule’s K captures not only the simple repetitiveness of vocabulary but also the more profound signification of its equivalence with the approximated second-order entropy. Because K has been previously reported as a stable text constancy measure, we consider it here once again, but this time within the broader context of Hα. Based on the previous reports (Tweedie and Baayen 1998; Kimura and Tanaka-Ishii 2014) and the discussion so far, we consider the following four measures as candidates for text constancy measures.
Repetitiveness-based measures: Yule’s K (Equation (1)); and Golcher’s V (Equation (2)).
Complexity-based measures: The Shannon entropy upper bound (h1 as the entropy rate (Equations (10) and (8)) and H1 (Equation (7), with X in terms of unigrams and the probability function in terms of relative frequencies); and the approximated Rényi entropy, denoted as Hα (α > 1) (Equation (11), again with X and the probability function in terms of unigrams and relative frequencies, respectively).
In addition, we empirically consider how these measures can be understood in the context of the power-law feature of language. As noted in the Introduction, for the convergent measures the speed of attaining convergence with respect to text size is examined as well. Among the candidates, K and H1 have been previously applied in a word-based manner, whereas V is string based. The Shannon entropy rate h1 has been considered in both ways. Because we should be able to consider a text in terms of both words and characters, we examine the constancy of each measure in both ways.
Furthermore, because we have seen the mathematical equivalence of Yule’s K and H2, in the following we only consider H2. As for Hα, we consider α = 3, 4 only in comparison with H2. Because H1 is based on relative frequencies and can be considered together with H2, we first focus on the convergence of the three measures V, h1, and H2, and then we consider H1 in comparison with H2, H3, and H4.","3 data : Table 1 lists the data used in our experimental examination. The table indicates the data identifier (by which we refer to the data in the rest of the article), language, source, number of distinct tokens, data length by total number of tokens, and size in bytes. The first block contains relatively large-scale natural language corpora consisting of texts written by multiple authors, and the second block contains smaller corpora consisting of texts by single authors. The third block contains programming language corpora, and the fourth block contains corpora of unknown scripts, which we examine at the end of this article in Section 4.3.
For the large-scale natural language data, we considered five languages: English, Japanese, Chinese, Arabic, and Thai. These languages were chosen to represent different language families and writing systems. The large-scale corpora in English, Japanese, and Chinese consist of newspapers in chronological order, and the Thai and Arabic corpora include other kinds of texts. The markers ‘w’, ‘c’, and ‘cr’ appearing at the end of every identifier in Table 1 (e.g., Enews-c, Enews-w, and Jnews-cr) indicate text processed through words, characters, and transliterated Roman characters, respectively.
As for the small-scale corpora in the second block, the texts were only considered in terms of words, since verification via characters produced findings consistent with those obtained with the large-scale corpora. The texts were chosen because each was written by a single author but is relatively large.
Here, we summarize our preprocessing procedures. For the annotated Thai NECTEC corpus, texts were tokenized according to the annotation. The preprocessing methods for the other corpora were as follows:
English: NLTK2 was used to tokenize text into words. Japanese: Mecab3 was used for tokenization, and KAKASI4 was used for romanization. Chinese: ICTCLAS20135 was used for tokenization, and the pinyin Python library was used for pinyin romanization. Other European Languages: PunktWordTokenizer6 was used for tokenization.
All the other natural language corpora were tokenized simply using spaces. Following Golcher (2007), who first suggested testing constancy on programming languages, we also collected program sources from different languages (third block in Table 1). The programs were also considered solely in terms of words, not characters. C++ and Python were chosen to represent different abstraction levels, and Lisp was chosen because of its different ordering for function arguments. Source code was collected from language libraries. The programming language texts were preprocessed as follows. Comments in natural language were eliminated (although strings remained in the programs, where each was a literal token). Identical files and copies of sources in large chunks were carefully eliminated, although this process did not completely eliminate redundancy since most programs reuse some previous code. Finally, the programs were tokenized according to the language specifications.7
The last block of the table lists two corpora of unknown scripts. We consider these scripts at the end of this article in Section 4.3, through Figure 5, to show one possible application of the text constancy measures. The first unknown script is that of the Voynich manuscript, a famous text that is undeciphered but hypothesized to have been written in natural language. This corpus is considered in terms of both characters and words, where words were defined via the white space separation in the original text. Given the common understanding that the manuscript seems to have two different parts (Reddy and Knight 2011), we separated it into two parts according to the Currier annotation (identified as A and B, respectively). The second corpus of unknown text consists of the Rongorongo script of Easter Island (Daniels and Bright 1996, Section 13; Orliac 2005; Barthel 2013). This script’s status as natural language is debatable, but if so, it is considered to possess characteristics of both phonographs and ideograms (Pozdniakov and Pozdniakov 2007). Because there are several ways to consider what constitutes a character in this script (Barthel 2013), we calculate values for the two most extreme cases as follows. For corpus RongoA-c, we consider a character inclusive of all adjoining parts (i.e., including accents and ornamental parts). On the other hand, for
2 http://nltk.org. 3 http://mecab.googlecode.com/svn/trunk/mecab/doc/index.html. 4 http://kakasi.namazu.org. 5 http://ictclas.nlpir.org. 6 http://nltk.org. 7 With respect to the Lisp programming language, its culture favors long, hyphenated variable names that
can be almost as long as a sentence. For this work, therefore, Lisp variable names were tokenized by splitting at the hyphens.
corpus RongoB-c, we separate parts as reasonably as possible, among multiple possible separation methods. Because the unit of word in this script is unknown, the Rongorongo script is only considered in terms of characters. The empirical verification of convergence for real data is controversial. We must first note that it does not conform with the standard approach to statistical testing. In the domain of statistics, it is a common understanding that “convergence” cannot be tested. A statistical test raises two contrasting hypotheses—called the null and alternative hypotheses—and calculates a p-value indicating the probability of the null hypothesis to occur. When this p-value is smaller than a certain value, the null hypothesis is considered unable to occur and is thus rejected. For convergence, the null hypothesis corresponds to “not converging,” and the alternative hypothesis, to “converging.” The problem here is that the null hypothesis is always related to the alternative hypothesis to a certain extent, because the difference between convergence and non-convergence is merely a matter of degree. In other words, the notion of convergence for a constancy measure does not conform with the philosophy of statistical testing. Convergence is therefore considered in terms of the distance from convergent values, or in terms of the error with respect to some parameter (such as data size). Such a distance cannot be calculated for real data, however, since the underlying mathematical model is unknown.
To sum up, verification of the convergence of real data must be considered by some other means. Our proposal is to consider convergence in comparison to a set of random data whose process is known. For this random data, we considered two kinds.
The first kind is used to examine data convergence in Section 4.1. This random data was generated from real data by shuffling the original text with respect to certain linguistic units. Tweedie and Baayen (1998) presented results by shuffling words, where the original texts were literary texts by single authors. Here, we generated random data by shuffling (1) words/characters, (2) sentences, or (3) documents. Because these options greatly increased the number of combinations of results, we mainly present the results with option (1) for large-scale data in this article. There are three reasons for this: Convergence must be verified especially at large scale; the most important convergence findings for randomized small-scale data were already reported in Tweedie and Baayen (1998); and the results for options (2) and (3) were situated within the range of option (1) and the original texts.
Randomization of the words and characters of original texts will destroy various linguistic characteristics, such as n-grams and long-range correlation. The convergence properties of the three measures V, h1, and H2 are as follows. The convergence of V is unknown, because it lacks a mathematical background. Even if the value of V did converge, the convergent value for randomized data would differ from that of the original text, since the measure is based on repeated n-grams in the text. h1 converges to the entropy rate of the randomized text, if the data size suffices. This is supported by the mathematical background of the algorithm, which converges to the true entropy rate for stationary data. Even when h1 converges for random data, the convergent value will be larger than that of the original text, because h1 considers the probabilities of n-grams. Lastly, H2 converges to the same point for a randomized text and the original text, because it is the approximated higher-order entropy, such that words and characters are considered to occur independently.
The second kind of random data is used to compare the convergent values of different texts for a constancy measure, as considered in Section 4.2. Random corpora
were generated according to four different distributions: one uniform, and the other three following Zipf distributions with exponents of γ = 0.8, 1.0, and 1.3, respectively, for Equation (3). Because each set of real data consists of different numbers of distinct tokens, ranging from tens to billions, random data sets consisting of 2n distinct tokens for every n = 4 . . .19, were randomly generated for each of the four distributions. We only consider the measures H2 and H0 for these data sets. Both of these measures have convergent values, given a sufficient data size.","4 experimental results :From the previous discussion, we applied the three measures V, h1, and H2 with five large-scale and eight small-scale natural language corpora, three programming language corpora, and two unknown script corpora, in terms of words and characters. Because there were many results for different combinations of measure, data, and token (word or character), this section is structured so that it best highlights our findings. Figures 1, 2, and 3 in this section can be examined in the following manner. The horizontal axis indicates the text size of each corpus, in terms of the number of tokens, on a log scale. Chunks of different text sizes were always taken from the head of the corpus.8 The vertical axis indicates the values of the different measures: V, h1, or H2. Each figure contains multiple lines, each corresponding to a corpus, as indicated in the legends.
First, we consider the results for the large-scale data. Figure 1 shows the different measures for words (left three graphs) and characters (right three graphs). We can see that V increased for both words and characters (top two graphs). Golcher tested his measure on up to 30 megabytes of text in terms of characters (Golcher 2007). We also observed a stable tendency up to around 107 characters. The increase in V became apparent, however, for larger text sizes. Thus, it is difficult to consider V as a constancy measure.
As for the results for h1 (middle graphs), both graphs show a gradual decrease. The tendency was clearer for words than for characters. For some corpora, especially for characters, it was possible to observe some values converging towards h∗. The overall tendency, however, could not be concluded as converging. This result suggests the difficulty in attaining convergence of the entropy rate, even with gigabyte-scale data. From the theoretical background of the Grassberger algorithm, the values would possibly converge with larger-scale data. The continued decrease could be due to multiple reasons, including the possibility of requiring far larger data than that used here, or a discrepancy between linguistic processes and the mathematical model assumed for the Grassberger algorithm.
We tried to estimate h∗ by fitting the Equation (9). For the corpora with good fitting, all of the estimated values were larger than zero, but many of the results could not
8 For real data, this was done without any randomization of the order of texts for all corpora besides Atext-w and Atext-c. The Watan corpus is distributed not in the chronological order of the publishing dates, but as a set of articles grouped into categories (i.e., all articles of one category, then all articles of another category, and so on). Because of this, there is a large skew in the vocabulary distribution, depending on the section of the corpus. We thus randomly reshuffled the articles by categories for the whole corpus before taking chunks of different sizes (always from the beginning) to generate our results. Apart from this, we avoided any arbitrary randomization with respect to the original data summarized in Table 1.
be fitted easily, and the estimated values were unstable due to fluctuation of the lines. Whether a value for h∗ is reached asymptotically and also whether h∗ > 0 remain important questions requiring separate, more extensive mathematical and empirical studies.
In contrast, H2 (or Yule’s K, bottom graphs) showed convergence, already at the level of 105 tokens, for both words and characters. From the previous verification of Yule’s K, we can conclude that H2 is convergent. The final convergent values, however, differed for the various writing systems. We return to this issue in the next section.
To better understand the convergence, Figure 2 shows the results for the corresponding randomized data. As mentioned in Section 3.2, the original texts were
randomized by shuffling words and characters for the data examined by words and characters, respectively. Therefore, all n-gram characteristics existing in the text were destroyed, and what remained were the different words and characters appearing in a random order. Here, we see how the random data’s behavior has some of the theoretical properties of convergence, as summarized in Section 3.2.
As mentioned previously, because V has no mathematical background, its behavior even for uniform random data is unknown, and even if it converged, the convergent value would be smaller than that of the original text. The top two graphs in Figure 2 exhibit some oscillation, especially for randomized Chinese (Cnews-c,w). Such peculiar
oscillation was already reported by Golcher himself (Golcher 2007) for uniformly random data. This was easy to replicate, as reported in Kimura and Tanaka-Ishii (2014), for uniformly random data with the number of distinct tokens up to a hundred. Because the word distribution almost follows Zipf’s law, the vocabulary is not uniformly distributed, yet oscillating results occur for some randomized data in the top left figure. Moreover, the values seem to increase for Japanese and English for words at a larger scale. Although the plots for some scripts seem convergent (top right graph), these convergent values are theoretically different from those of the original texts, if they exist, and this stability is not universal across the different data sets. Given this result, it is doubtful that V is convergent across languages.
In contrast, h1 is mathematically proven to be convergent given infinite-length randomized data, but to larger values than those of the original texts, as mentioned in Section 3.2. The middle two graphs of Figure 2 show the results for h1. The majority of the plots do not reach convergence even at the largest data sizes, but for certain results with characters, especially in the Roman alphabet, the plots seem to go to a convergent value (middle right). All the plots can be extrapolated to converge to a certain entropy rate above zero, although these values are larger than the convergent values—if they ever exist—of the real data. These results confirm the difficulty of judging whether the entropy rates of the original texts are convergent and whether they remain above zero.
Lastly, it is easy to see that H2 is convergent for a randomized text (bottom two graphs), and the convergent values are the same for the cases with and without randomization. In fact, the plots converge to exactly the same points faster and more stably, which shows the effect of randomization.
As for the other randomization options, by sentences and documents, the findings—both the tendencies of the lines and the changes in the values—can be situated in the middle of what we have seen so far. The plots should increasingly fluctuate more like the real data because of the incomplete randomization, in the order of sentences and then documents.
Returning to inspection of the remaining real data, Figure 3 shows V, h1, and H2 in terms of words for the small-scale corpora (left column) and for the programming language texts (right column). For the small-scale corpora, in general, the plots are bunched together, and the results shared the tendencies noted previously for the largescale corpora. V again showed an increase, while h1 showed a tendency to decrease. H2 converged rapidly and was already almost stable at 104 tokens. This again shows how H2 exhibits stable constancy, especially with texts written by single authors.
As for the programming language results, the plots fluctuate more than for the natural language texts because of the redundancy within the program sources. Still, the global tendencies noted so far were just discernible. V had relatively larger values but h1 and H2 had smaller values for programs, as compared to the natural language texts. The differences in value indicate the larger degree of repetitiveness in programs.
Lastly, Figure 4 shows the Hα results for the Wall Street Journal in terms of words in unigrams (Enews-w). The horizontal axis indicates the corpus size, and the vertical axis indicates the approximated entropy value. The different lines represent the results for Hα with α = 1, 2, 3, 4. The two H1 plots represent calculations with and without Laplace smoothing (Manning and Schuetze 1999). We can see that without smoothing, H1 increased, as Tweedie and Baayen (1998) reported, but in contrast to their conclusion, we observe a tendency of convergence for larger-scale data. The increase was due to the influence of low-frequency vocabulary pushing up the entropy. The opposite tendency to decrease was observed for the smoothed probabilities, with the plot eventually converging to the same point as that for the unsmoothed H1 values. The convergence
was by far slower for H1 as compared with that for H2, H3, and H4, which all had attained convergence already at 102 tokens. The convergence values naturally decreased for larger α, although the amount of decrease itself rapidly decreased with larger α.
In answer to Questions 1 and 2 raised in the Introduction—which measures show constancy, with sufficient convergence speed—the empirical conclusion from our data is that Hα with α > 1 showed stable constancy when the values were approximated using relative frequencies. For H1, the convergence was much slower because of the strong influence of low-frequency words. Consequently, the constancy of Hα with α > 1 is attained by representing the gross complexity underlying a text. Now we turn to Question 3 raised in the Introduction and examine the discriminatory power of H2. As Yule intended, does H2 identify authors? Given the influence of different writing systems, as seen previously in Figure 1, we examine the relation between H2 and the number of distinct tokens (the alphabet/vocabulary size). Note that because this number corresponds to H0 in Equation (11), this analysis effectively considers texts on the H0-H2 plane. Since H0 grows according to the text size, unlike H2, the same text size must be used for all corpora in order to meaningfully compare H0 values.9 Given that H2 converges fast, we chose a size of 104 tokens to handle all of the small- and large-scale corpora.
For each of the corpora listed in Table 1 and the second kind of random corpora explained at the end of Section 3.2, Figure 5 plots the values of H2 (vertical axis) and the number of distinct tokens H0 (horizontal) measured for each corpus at a size of 104 tokens. The three large circles are groupings of points. The leftmost group represents news sources in alphabetic characters. All of the romanized Chinese, Japanese, and Arabic texts are located almost at the same vertical location as the English text. This indicates the difficulty for H2 to distinguish natural languages if measured in terms of alphabetic characters. The middle group represents the programming language texts in terms of words. This group is located separately (vertically lower than the natural language corpora in terms of words), so H2 is likely to distinguish between natural languages and programming languages. The rightmost group represents the small-scale corpora. Considering the proximity of these points despite the variety of the content, it is unlikely that H2 can distinguish authors, in contrast to Yule’s hope. Still, these points are located lower than those for news text. Therefore, H2 has the potential to distinguish genre or maybe writing style.
9 Because H0 is not convergent, the horizontal locations remain unstable, unless the tokens are of a phonographic alphabet. In other words, for all word-based results and character-based results not based on a phonographic alphabet, the resulting horizontal locations are changed by increasing the corpus size. As for the random data, the H0 values are convergent, because these data sets have a finite number of distinct tokens. Since H0 is measured only for the first 104 tokens, however, the horizontal locations are underestimated, especially for random data following a Zipf distribution.
The natural language texts located near the line for a Zipf exponent of 0.8 are those of the non-alphabetic writing systems.10 Note that Chinese characters have morphological features, and the Arabic and Thai languages also have flexibility in terms of which units are considered words and morphemes. In other words, the plots closer to the random data with a smaller Zipf exponent are for language corpora of morphemic sequences. The group of plots measured for phonographic scripts is located near the line for a Zipf exponent of 1.0 (the grouping of points in the leftmost circle), which could suggest that morphemes are more randomized units than words. The nature of unknown scripts can also be considered through our understanding thus far. Figure 5 includes plots for the Voynich manuscript in terms of words and characters, and for the Rongorongo script in terms of characters. Like all the data seen in this figure, the points are placed at the H2 values (vertically) for the number of distinct tokens (horizontally) at the specified size of 104 tokens, with the exception of Voynich-A in terms of words. Because this corpus consists of fewer than 104 words (refer to the data length by tokens listed for VoynichA-w in Table 1), its point is located horizontally at the vocabulary size corresponding to the corpus’ maximum size.
For the two Voynich manuscript parts, the plots in terms of words appear near the Arabic corpus for words (Abook-w). For characters, on the other hand, the plots are at the leftmost end of the figure. This was due to overestimation of the total number
10 Note that here we use the values of the Zipf exponent for the random data, and not the estimated exponents for the real data. The rank-frequency distributions of characters, especially for phonetic alphabets, often do not follow a power law.
of characters for the alphabetic texts (e.g., both English and other, romanized language texts), since all ASCII characters, such as colons, periods, and question marks, are counted. Still, the H2 values are located almost at the same position as for the other romanized texts, indicating that the Voinich manuscript has approximately similar complexity. These results suggest the possibility that the Voynich manuscript could have been generated from a source in natural language, possibly written in some script of the abjad type. This supports previous findings (Reddy and Knight 2011; Montemurro and Zanette 2013), which reported the possibility of the Voynich manuscript being in a natural language and the coincidence of its word length distribution with that of Arabic.
On the other hand, the plots for the Rongorongo script appear near the line for a Zipf exponent of 0.8, with RongoA near Arabic in terms of words but RongoB somewhat further down from Japanese in terms of characters. The status of Rongorongo as natural language has been controversial (Pozdniakov and Pozdniakov 2007). Both points in the graph, however, are near many other natural language texts (and not widely separated), making it reasonable to hypothesize that Rongorongo is indeed natural language. The characters can be deemed morphologically rich, because both plots are close to the line for a Zipf exponent of 0.8. In the case of RongoA, for which a character was considered inclusive of all parts (i.e., including accents and ornamental parts), the morphological richness is comparable to that of the words of an abjad script. On the other hand, when considering the different character parts as distinct (RongoB), the location drifts towards the plot for Thai, a phonographic script, in terms of characters. Therefore, the Rongorongo script could be considered basically morphemic, with some parts functioning phonographically. This conclusion again supports a previous hypothesis proposed by a domain specialist (Pozdniakov and Pozdniakov 2007).
This analysis of two unknown scripts supports previous conjectures. Our results, however, only add a small bit of evidence to those conjectures; clearly, reaching a reasonable conclusion would require further study. Moreover, the analysis of unknown scripts introduced here could provide another possible application of text constancy measures, from a broader context.","5 conclusion :We have discussed text constancy measures, whose values are invariant across different sizes of text, for a given text. Such measures have a 70-year history, since Yule originally proposed K as a text characteristic, potentially with language engineering utility for problems such as author identification. We consider text constancy measures today to have scientific importance in understanding language universals from a computational view.
After overviewing measures proposed so far and previous studies on text constancy, we explained how K essentially has a mathematical equivalence to the Rényi higher-order entropy. We then empirically examined various measures across different languages and kinds of corpora. Our results showed that only the approximated higherorder Rényi entropy exhibits stable, rapid constancy. Examining the nature of the convergent values revealed that K does not possess the discriminatory power of author identification as Yule had hoped. We also applied our understanding to two unknown scripts, the Voynich manuscript and Rongorongo, and showed how our constancy results support previous hypotheses about each of these scripts.
Our future work will include application of K to other kinds of data besides natural language. There, too, we will consider the questions raised in the Introduction, of
whether K converges and of how discriminatory it is. We are especially interested in considering the relation between the value of K and the meaningfulness of data.",,,,"This article presents a mathematical and empirical verification of computational constancy measures for natural language text. A constancy measure characterizes a given text by having an invariant value for any size larger than a certain amount. The study of such measures has a 70-year history dating back to Yule’s K, with the original intended application of author identification. We examine various measures proposed since Yule and reconsider reports made so far, thus overviewing the study of constancy measures. We then explain how K is essentially equivalent to an approximation of the second-order Rényi entropy, thus indicating its signification within language science. We then empirically examine constancy measure candidates within this new, broader context. The approximated higher-order entropy exhibits stable convergence across different languages and kinds of text. We also show, however, that it cannot identify authors, contrary to Yule’s intention. Lastly, we apply K to two unknown scripts, the Voynich manuscript and Rongorongo, and show how the results support previous hypotheses about these scripts.","[{""affiliations"": [], ""name"": ""Kumiko Tanaka-Ishii""}, {""affiliations"": [], ""name"": ""Shunsuke Aihara""}]",SP:54cf25939e5f7b648f0a97837673bc1f8e10a4be,"[{""authors"": [""T. Barthel""], ""title"": ""The Rongorongo of Easter Island: Thomas Barthel\u2019s transliteration system"", ""venue"": ""Available at http://kohaumotu. org/rongorongo org/corpus/ codes.html. Accessed June 2015."", ""year"": 2013}, {""authors"": [""T.C. Bell"", ""J.G. Cleary"", ""I.H. Witten.""], ""title"": ""Text Compression"", ""venue"": ""Prentice Hall."", ""year"": 1990}, {""authors"": [""P.A. Bromiley"", ""N.A. Thacker"", ""E. Bouhova-Thacker.""], ""title"": ""Shannon entropy, Renyi entropy, and information"", ""venue"": ""Available at http://www.tina-vision.net/ docs/memos/2004-004.pdf. Accessed"", ""year"": 2010}, {""authors"": [""P.F. Brown"", ""S.A. Della Pietra"", ""V.J. Della Pietra"", ""J.C. Lai"", ""R.L. Mercer.""], ""title"": ""An estimate of an upper bound for the entropy of English"", ""venue"": ""Computational Linguistics, 18(1):31\u201340."", ""year"": 1992}, {""authors"": [""E. Brunet""], ""title"": ""Vocabulaire de Jean Giraudoux: Structure et Evolution"", ""venue"": ""Slatkine."", ""year"": 1978}, {""authors"": [""T. Cover"", ""R. King.""], ""title"": ""A convergent gambling estimate of the entropy of English"", ""venue"": ""IEEE Transactions on Information Theory, 24(4):413\u2013421."", ""year"": 1978}, {""authors"": [""T.M. Cover"", ""J.A. Thomas.""], ""title"": ""Elements of Information Theory"", ""venue"": ""Wiley-Interscience."", ""year"": 2006}, {""authors"": [""\u0141. D\u0229bowski""], ""title"": ""A general definition of conditional information and its application to ergodic decomposition"", ""venue"": ""Statistics and Probability Letters, 79(9):1260\u20131268."", ""year"": 2009}, {""authors"": [""\u0141. D\u0229bowski""], ""title"": ""Empirical evidence for Hilberg\u2019s conjecture in single author texts"", ""venue"": ""Methods and Applications of Quantitative Linguistics: Selected papers of the 8th International Conference on"", ""year"": 2013}, {""authors"": [""\u0141. D\u0229bowski""], ""title"": ""The relaxed Hilberg conjecture: A review and new experimental support"", ""venue"": ""Available at http://www.ipipan.waw.pl/ldebowsk/. Accessed June 2015."", ""year"": 2014}, {""authors"": [""D. Dugast""], ""title"": ""Vocabulaire et Stylistique"", ""venue"": ""I Th\u00e9\u00e2tre et Dialogue. Slatkine-Champion. Travaux de Linguistique Quantitative."", ""year"": 1979}, {""authors"": [""M. Farach"", ""M. Noordewier"", ""S. Savari"", ""L. Shepp"", ""A. Wyner"", ""J. Ziv.""], ""title"": ""On the entropy of DNA: Algorithms and measurements based on memory and rapid convergence"", ""venue"": ""Proceedings of the"", ""year"": 1995}, {""authors"": [""D. Genzel"", ""E. Charniak.""], ""title"": ""Entropy rate constancy in text"", ""venue"": ""Annual Meeting of the Association for the ACL, pages 199\u2013206, Philadelphia, PA."", ""year"": 2002}, {""authors"": [""F. Golcher""], ""title"": ""A stable statistical constant specific for human language texts"", ""venue"": ""Recent Advances in Natural Language Processing, Borovets."", ""year"": 2007}, {""authors"": [""I.J. Good""], ""title"": ""The population frequencies of species and the estimation of population parameters"", ""venue"": ""Biometrika, 40(3\u20134):237\u2013264."", ""year"": 1953}, {""authors"": [""P. Grassberger""], ""title"": ""Estimating the information content of symbol sequences and efficient codes"", ""venue"": ""IEEE Transactions on Information Theory, 35:669\u2013675."", ""year"": 1989}, {""authors"": [""H. Guiraud""], ""title"": ""Les Charact\u00e8res Statistique du Vocabulaire"", ""venue"": ""Universitaires de France Press."", ""year"": 1954}, {""authors"": [""D. Gusfield""], ""title"": ""Algorithms on Strings, and Sequences: Computer Science and Computational Biology"", ""venue"": ""Cambridge University Press."", ""year"": 1997}, {""authors"": [""G. Herdan""], ""title"": ""Quantitative Linguistics"", ""venue"": ""Butterworths."", ""year"": 1964}, {""authors"": [""W. Hilberg""], ""title"": ""Der bekannte grenzwert der redundanzfreien information in texten eine fehlinterpretation der shannonschen experimente? Frequenz, 44(9\u201310):243\u2013248"", ""year"": 1990}, {""authors"": [""A. Honor\u00e9""], ""title"": ""Some simple measures of richness of vocabulary"", ""venue"": ""Association for Literary and Linguistic Computing Bulletin, 7:172\u2013177."", ""year"": 1979}, {""authors"": [""D. Kimura"", ""K. Tanaka-Ishii.""], ""title"": ""A study on constants of natural language texts"", ""venue"": ""Journal of Natural Language Processing, 18(2):119\u2013137."", ""year"": 2011}, {""authors"": [""D. Kimura"", ""K. Tanaka-Ishii.""], ""title"": ""A study on constants of natural language texts"", ""venue"": ""Journal of Natural Language Processing, 21:877\u2013895. Special issue of awarded papers. [The English translated version"", ""year"": 2014}, {""authors"": [""B. Kitchens""], ""title"": ""Symbolic Dynamics: One-sided, Two-sided and Countable State Markov Shifts"", ""venue"": ""Springer. 501"", ""year"": 1998}, {""authors"": [""R. Levy"", ""T.F. Jaeger.""], ""title"": ""Speakers optimize information density through information density through syntactic reduction"", ""venue"": ""Annual Conference on Neural Information"", ""year"": 2007}, {""authors"": [""H.D. Maas""], ""title"": ""Zusammenhang zwischen wortschatzumfang und l\u00e4nge eines textes [Relationship between vocabulary and text length"", ""venue"": ""Zeitschrift f\u00fcr Literaturwissenschaft und Linguistik, 8:73\u201370."", ""year"": 1972}, {""authors"": [""B. Mandelbrot""], ""title"": ""An informational theory of the statistical structure of language"", ""venue"": ""Communication Theory, 486\u2013500."", ""year"": 1953}, {""authors"": [""C. Manning"", ""H. Schuetze.""], ""title"": ""Foundations of Statistical Natural Language Processing"", ""venue"": ""MIT Press."", ""year"": 1999}, {""authors"": [""M. Montemurro"", ""D. Zanette.""], ""title"": ""Keywords and co-occurrence patterns in the Voynich Manuscript: An information-theoretic analysis"", ""venue"": ""PLOS One. doi: 10.1371/journal.pone.0066344."", ""year"": 2013}, {""authors"": [""C. Orliac""], ""title"": ""The Rongorongo tablets from Easter Island: Botanical identification and 14c dating"", ""venue"": ""Archaeology in Oceania, 40(3):115\u2013119."", ""year"": 2005}, {""authors"": [""J.K. Orlov"", ""R.Y. Chitashvili.""], ""title"": ""Generalized z-distribution generating the well-known \u2018rank-distributions"", ""venue"": ""Bulletin of the Academy of Sciences of Georgia, 110:269\u2013272."", ""year"": 1983}, {""authors"": [""K. Pozdniakov"", ""I. Pozdniakov.""], ""title"": ""Rapanui writing and the Rapanui language: Preliminary results of a statistical analysis"", ""venue"": ""Forum for Anthropology and Culture, 3:3\u201336."", ""year"": 2007}, {""authors"": [""S. Reddy"", ""K. Knight.""], ""title"": ""What we know about the Voynich Manuscript"", ""venue"": ""ACL Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities, Portland, OR."", ""year"": 2011}, {""authors"": [""A. R\u00e9nyi""], ""title"": ""On measures of entropy and information"", ""venue"": ""Proceedings of the Fourth Berkeley Symposium on Mathematics,"", ""year"": 1960}, {""authors"": [""A. R\u00e9nyi""], ""title"": ""Foundations of Probability"", ""venue"": ""Dover Publications."", ""year"": 1970}, {""authors"": [""T. Sch\u00fcmann"", ""P. Grassberger.""], ""title"": ""Entropy estimation of symbol sequences"", ""venue"": ""Chaos, 6(3):414\u2013427."", ""year"": 1996}, {""authors"": [""C. Shannon""], ""title"": ""A mathematical theory of communication"", ""venue"": ""Bell System Technical Journal, 27:379\u2013423, 623\u2013656."", ""year"": 1948}, {""authors"": [""P.C. Shields""], ""title"": ""Entropy and prefixes"", ""venue"": ""Annals of Probability, 20(1):403\u2013409."", ""year"": 1992}, {""authors"": [""H.S. Sichel""], ""title"": ""On a distribution law for word frequencies"", ""venue"": ""Journal of the American Statistical Association, 70(351):542\u2013547."", ""year"": 1975}, {""authors"": [""E.H. Simpson""], ""title"": ""Measurement of diversity"", ""venue"": ""Nature, 163:688."", ""year"": 1949}, {""authors"": [""E. Stamatatos"", ""N. Fakotakis"", ""G. Kokkinakis.""], ""title"": ""Automatic text categorization in terms of genre and author"", ""venue"": ""Computational Linguistics, 26(4):471\u2013495."", ""year"": 2001}, {""authors"": [""B. Stein"", ""N. Lipka"", ""P. Prettenhofer.""], ""title"": ""Intrinsic plagiarism analysis"", ""venue"": ""Language Resources and Evaluation, 45(1):63\u201382."", ""year"": 2010}, {""authors"": [""Y.W. Teh""], ""title"": ""A hierarchical Bayesian language model based on Pitman-Yor processes"", ""venue"": ""Proceedings of the 21st International Conference On Computational Linguistics and 44th Annual Meeting of the"", ""year"": 2006}, {""authors"": [""J. Tuldava""], ""title"": ""Quantitative relations between the size of the text and lexical richness"", ""venue"": ""SMIL Quarterly, Journal of Linguistic Calculus, 4:28\u201335."", ""year"": 1977}, {""authors"": [""F.J. Tweedie"", ""R.H. Baayen""], ""title"": ""How variable may a constant be? Measures of lexical richness in perspective"", ""venue"": ""Computers and the Humanities, 32:323\u2013352."", ""year"": 1998}, {""authors"": [""G.U. Yule""], ""title"": ""The Statistical Study of Literary Vocabulary"", ""venue"": ""Cambridge University Press."", ""year"": 1944}, {""authors"": [""G.K. Zipf""], ""title"": ""Human Behavior and the Principle of Least Effort: An Introduction to Human Ecology"", ""venue"": ""Hafner, New York. 502"", ""year"": 1965}]",acknowledgments  :This research was supported by JST’s PRESTO program.,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :A constancy measure for a natural language text is defined, in this article, as a computational measure that converges to a value for a certain amount of text and remains invariant for any larger size. Because such a measure exhibits the same value for any size of text larger than a certain amount, its value could be considered as a text characteristic.
The concept of such a text constancy measure was introduced by Yule (1944) in the form of his measure K. Since Yule, there has been a continuous quest for such measures, and various formulae have been proposed. They can be broadly categorized into three types, namely, those measuring (1) repetitiveness, (2) power law character, and (3) complexity.
∗ Kyushu University, 744 Motooka Nishiku, Fukuoka City, Fukuoka, Japan. E-mail: kumiko@ait.kyushu-u.ac.jp.
∗∗ JST-PRESTO, 4-1-8 Honcho, Kawaguchi, Saitama 332-0012, Japan. † Gunosy Inc., 6-10-1 Roppongi, Minato-ku, Tokyo, Japan.
Submission received: 11 July 2013; revised version received: 17 February 2015; accepted for publication: 18 March 2015.
doi:10.1162/COLI a 00228
© 2015 Association for Computational Linguistics
Yule’s original intention for K’s utility lay in author identification, assuming that it would differ for texts written by different authors. State-of-the-art multivariate machine learning techniques are powerful, however, for solving such language engineering tasks, in which Yule’s K is used only as one variable among many, as reported in Stamatatos, Fakotakis, and Kokkinakis (2001) and Stein, Lipka, and Prettenhofer (2010).
We believe that constancy measures today, however, have greater importance in understanding the mathematical nature of language. Although mathematical models of language have been studied in the computational linguistics milieu, via Markov models (Manning and Schuetze 1999), Zipf’s law and its modifications (Mandelbrot 1953; Zipf 1965; Bell, Cleary, and Witten 1990), and Pitman-Yor models (Teh 2006) more recently, the true mathematical model of linguistic processes is ultimately unknown. Therefore, the convergence of a constancy measure must be examined through empirical verification. Because some constancy measures have a mathematical theory of convergence for a known process, discrepancies in the behavior of real linguistic data from such a theory would shed light on the nature of linguistic processes and give hints towards improving the mathematical models. Furthermore, as one application, a convergent measure would allow for comparison of different texts through a common, stable norm, provided that the measure converges for a sufficiently small amount of text. One of our goals is to discover a non-trivial measure with a certain convergence speed that distinguishes the different natures of texts.
The objective of this article is thus to provide a potential explanation of what the study of constancy measures over 70 years has been about, by answering the three following questions mathematically and empirically:
Question 1 Does a measure exhibit constancy?
Question 2 If so, how fast is the convergence speed?
Question 3 How discriminatory is the measure?
We seek answers by first showing the meaning of Yule’s K in relation to the Rényi higher-order entropy, and by then empirically examining constancy across large-scale texts of different kinds. We finally provide an application by considering the natures of two unknown scripts, the Voynich manuscript and Rongorongo, in order to show the possible utility of a constancy measure.
The most important and closest previous work was reported in Tweedie and Baayen (1998), the first paper to have examined the empirical behavior of constancy measures on real texts. The authors used English literary texts to test constancy measure candidates proposed prior to their work. Today, the coverage and abundance of language corpora allow us to conduct a larger-scale investigation across multiple languages. Recently, Golcher (2007) tested his measure V (discussed later in this paper) with IndoEuropean languages and also programming language sources. Our papers (Kimura and Tanaka-Ishii 2011, 2014) also precede this one, presenting results preliminary to this article but with only part of our data, and neither of those provides mathematical analysis with respect to the Rényi entropy. Compared with these previous reports, our contribution here can be summarized as follows:
Our work elucidates the mathematical relation of Yule’s K to Rényi’s higher-order entropy and explains why K converges.
Our work vastly extends the corpora used for empirical examination in terms of both size and language.
Our work compares the convergent values for these corpora. Our work also presents results for unknown language data, specifically from the Voynich manuscript and Rongorongo.
We start by summarizing the potential constancy measures proposed so far. 2 constancy measures :The measures proposed so far can broadly be categorized into three types, calculating the repetitiveness, power-law distribution, or complexity of text. This section mathematically analyzes these measures and summarizes them. The study of text constancy started with proposals for simple text measures of vocabulary repetitiveness. The representative example is Yule’s K (Yule 1944), while Golcher recently proposed V as another candidate (Golcher 2007).
2.1.1 Yule’s K. To the best of our knowledge, the oldest mention of constancy values was made by Yule with his notion of K (Yule 1944). Let N be the total number of words in a text, V(N) be the number of distinct words, V(m, N) be the number of words appearing m times in the text, and mmax be the largest frequency of a word. Yule’s K is then defined as follows, through the first and second moments of the vocabulary population distribution of V(m, N), where S1 = N = ∑ m mV(m, N), and S2 = ∑ m m
2V(m, N) (Yule 1944; Herdan 1964):
K = CS2 − S1 S21
= C [ − 1N + mmax∑ m=1 V(m, N)( mN ) 2 ]
(1)
where C is a constant enlarging of the value of K, defined by Yule as C = 104. K is designed to measure the vocabulary richness of a text: The larger Yule’s K, the less rich the vocabulary is. The formula can be intuitively understood from the main term of the sum in the formula. Because the square of ( mN )
2 indicates the degree of recurrence of a word, the sum of such degrees for all words is small if the vocabulary is rich, or large in the opposite case. Another simple example can be given in terms of S2 in this formula. Suppose a text is 10 words long: if each of the 10 tokens is distinct (high diversity), then S2 = 1 × 1 × 10 = 10; whereas, if each of the 10 tokens is identical (low diversity), then S2 = 10 × 10 × 1 = 100.
Measures that are slightly different but essentially equivalent to Yule’s K have appeared here and there. For example, Herdan defined Vm as follows (Herdan 1964, pp. 67, 79):
Vm = √√√√ mmax∑ m=1 V(m, N)( mN ) 2 − 1 V(N)
Likewise, Simpson (1949) derived the following formula as a measure to capture the diversity of a population:
D = mmax∑ m=1 V(m, N) mN m − 1 N − 1
which is equivalent to Yule’s K, as Simpson noted.
2.1.2 Other Measures Based on Simple Text Statistics. Apart from Yule’s K, various measures have been proposed from simple statistical observation of text, as detailed in Tweedie and Baayen (1998). One genre is based on the so-called token-type relation (i.e., the ratio of the vocabulary size V(N) and the text size N, in log) as formulated by Guiraud (1954) and Herdan (1964) as a law. Because this simple ratio is not stable, the measure was modified numerous times to formulate Herdan’s C (Herdan 1964), Dugast’s k and U (Dugast 1979), Maas’ a2 (Maas 1972), Tuldava’s LN (Tuldava 1977), and Brunet’s W (Brunet 1978).
Another genre of measures concerns the proportion of hapax legomena, that is V(1, N). Honoré noted that V(1, N) increases linearly with respect to the log of a text’s vocabulary size V(N) (Honoré 1979). Another ratio, of V(2, N) to V(N), was proposed as a text characteristic by Sichel (1975) and Maas (1972).
Each of these values, however, was found not to be convergent according to the extensive study conducted by Tweedie and Baayen (1998). In common with Yule’s intention to apply such measures for author identification, they examined all of the measures discussed here, in addition to two measures explained later: Orlov’s Z, and the Shannon entropy upper bound obtained from the relative frequencies of unigrams. They examined these measures with English novels (such as Alice’s Adventures in Wonderland) and empirically found that only Yule’s K and Orlov’s Z were convergent. Given their report, we consider K the only true candidate among the constancy measures examined so far.
2.1.3 Golcher’s V. Golcher’s V is a string-based measure calculated on the suffix tree of a text (Golcher 2007). Letting the length of the string be N and the number of inner nodes of the (Patricia) suffix tree (Gusfield 1997) be k, V is defined as:
V = kN (2)
Golcher empirically showed how this measure converges to almost the same value across Indo-European languages for about 30 megabytes of data. He also showed how the convergent values differ from those calculated for programming language texts.
Golcher explains in his paper that the possibility of constancy of V does not yet have mathematical grounding and has only been shown empirically. He does not report values for texts larger than about 30 megabytes nor for those of non-Indo-European languages. A simple conjecture on this measure is that because a suffix tree for a string of length N has at most N − 1 inner nodes, V must end up at some value 0 ≤ V < 1, for any given text.
Our group tested V with larger-scale data and concluded that V could be a constancy measure, although we admitted to observing a gradual increase (Kimura and Tanaka-Ishii 2014). Because V requires further verification on larger-scale data before ruling it out, we include it as a constancy measure candidate. Since Zipf (1965), power laws have been reported as an underlying statistical characteristic of text. The famous Zipf’s law is defined as:
f (n) ∝ n−γ (3)
where γ ≈ 1, and f (n) is the frequency of the nth most frequent word in a text. Various studies have sought to explain mathematically how the exponent could differ depending on the kind of text. To the best of our knowledge, however, there has been a limited number of reports related to text constancy.
An exception is the study on Orlov’s Z (Orlov and Chitashvili 1983). Orlov and Chitashvili attempted to obtain explicit mathematical forms for V(N) and V(m, N) by more finely considering the long tails of vocabulary distributions for which Zipf’s law does not hold. They obtained these forms through a parameter Z, defined as the potential text length minimizing the square error of the estimated V(m, N), with its actual value as follows:
Z = arg min N 1 mmax mmax∑ m=1 {E[V(m, N)] − V(m, N) V(N) }2 (4)
Thus defining Z, they mathematically deduced for V(N) the following formula:
V(N) = Z log(mmaxZ) N N − Z log ( N Z ) (5)
Two ways to obtain Z can be formulated through approximation: one through GoodTuring smoothing (Good 1953), which assumes Zipf’s law to hold, and the other using Newton’s method. Tweedie and Baayen showed how the value of Z is stable at the size of an English novel by a single author and thus suggested that it could form a text characteristic. The empirical results, however, were not significantly convergent with respect to text size, and, moreover, Tweedie and Baayen provided their results without giving an estimation method (Tweedie and Baayen 1998). Calculation using Good-Turing smoothing, which is derived directly from Zipf’s law, would cause Z to converge, but this does not take Orlov’s original intention into consideration. Alternatively, our group (Kimura and Tanaka-Ishii 2014) verified Z through Newton’s method by setting g(Z) = 0, where g(Z) is the following function:
g(Z) = Z log(mmaxZ) N N − Z log ( N Z ) − V(N) (6)
We also showed how the value of Z increases rapidly when the text size is larger than 10 megabytes.
The major problem with measures based on power laws lies in the skewed head and tail of the vocabulary population distribution. Because these exceptions constitute important parts of the population, parameter estimation by fitting to Equation (3) is sensitive to the estimation method. For example, the estimated value of the exponent for Zipf’s law depends on the method used for dealing with these exceptions. We tested several simple methods of estimating the Zipf law’s exponent γ with different ways of handling the head and tail of a distribution. There were settings that led to
convergence, but the convergence depended on the settings. Such difficulty could be one reason why there has been no direct proposal for γ as a text constancy measure. Hence, due care must be taken in relating text constancy to a power law. We chose another path by considering text constancy through a random Zipf distribution, as described later in the experimental section. With respect to measures based on complexity, multiple reports have already examined the Shannon entropy (Shannon 1948; Cover and Thomas 2006). In addition, we introduce the Rényi higher-order entropy (Rényi 1960) as another possible measure.
2.3.1 Shannon Entropy Upper Bound. Let X be the random variable of a sequence X = X1, X2, . . . , Xi, . . . , where Xi represents the ith element of X: Xi = x ∈ X, and where X represents a given set (e.g., a set of words or characters) whose members constitute the sequence. Let Xji (i < j) denote the random variable indicating its subsequence Xi, Xi+1, Xi+2, . . . , Xj. Let P(X) indicate the probability function of a sequence X. The Shannon entropy is then defined as:
H(X) = − ∑
X
P(X) log P(X) (7)
Tweedie and Baayen directly calculated an approximation of this formula in terms of the relative frequencies (for P) of unigrams (for X), and they concluded that the measure would continue increasing with respect to text size and would not converge for short, literary texts (Tweedie and Baayen 1998). Because we are interested in the measure’s behavior on a larger scale, we replicated their experiment, as discussed later in the section on empirical constancy. We denote this measure as H1 in this article.
Apart from that report, many have studied the entropy rate, defined as:
h∗ = lim n→∞ H(Xn1 ) n (8)
Theoretically, the behavior of the entropy rate with respect to text size has been controversial. On the one hand, there have been indications of entropy rate constancy (Genzel and Charniak 2002; Levy and Jaeger 2007). These reports argue that the entropy rate of natural language could be constant. Due to the inherent difficulty in obtaining the true value of h∗ from a text, however, these arguments are based only on indirect clues with respect to convergence. On the other hand, Hilberg conjectured a decrease in the human conditional entropy, as follows (Hilberg 1990):
H(Xn|Xn−11 ) ∝ n−1+β
He obtained this through an examination of Shannon’s original experimental data and suggested that β ≈ 0.5. From this formula, Dȩbowski induces that H(Xn1 ) ∝ nβ and that the entropy rate can be formulated generally as follows (Dȩbowski 2014):
H(Xn1 ) n ≈ An−1+β + h∗ (9)
Note that at the limit of n → ∞, this rate goes to h∗, a constant, provided that β < 1.0. Hilberg’s conjecture is deemed compatible with entropy rate constancy at its asymptotic limit, provided that h∗ > 0 holds.1 We are therefore interested in whether this h∗ forms a text characteristic, and if so, whether h∗ > 0.
Empirically, many have attempted to calculate the upper bound of the entropy rate. Brown’s report (Brown et al. 1992) is representative in showing a good estimation of the entropy rate for English from texts, as compared with values obtained from humans (Cover and King 1978). Subsequently, there have been important studies on calculating the entropy rate, as reported thoroughly in Schümann and Grassberger (1996). The questions related to h∗, however, remain unsolved. Recently, Dȩbowski used a Lempel-Ziv compressor and examined Hilberg’s conjecture for texts by single authors (Dȩbowski 2013). He showed an exponential decrease in the entropy rate with respect to text size, supporting the validity of Equation (9). Following these previous works, we examine the entropy rate by using an algorithm proposed by Grassberger (1989) and later on by Farach et al. (1995). This method is based on universal coding. The algorithm has a theoretical background of convergence to the true h∗, provided the sequence is stationary, but has been proved by Shields (1992) to be inconsistent—that is, it does not converge to the entropy rate for certain nonMarkovian processes. We still chose to apply this method, because it requires no arbitrary parameters for calculation and is applicable to large-scale data within a reasonable time.
The Grassberger algorithm (Grassberger 1989; Farach et al. 1995) can be summarized as follows. Consider a sequence X of length N. The maximum matching length Li is defined as:
Li = max{k : Xj+kj = Xi+ki }
for j ∈ {1, . . . , i − 1}, 1 ≤ j ≤ j + k ≤ i − 1. In other words, Li is the maximum common subsequence before and after i. If L̄ is the average length of Li, given by
L̄ = 1N i=N∑ i=1 Li
then the method obtains the entropy rate h1 as
h1 = log2N
L̄ (10)
Given the true entropy rate h∗, convergence has been mathematically proven for a stationary process, such that |h∗ − h1| = O(1) when N → ∞. In this article, we consider this entropy rate h1 as a constancy measure candidate.
1 According to Dȩbowski (2009), h∗ = 0 suggests that the next element of a linguistic process is deterministic, that is, a function of the corpus observed before, under the two conditions that (1) the number of possible choices for the element is finite, and (2) the corpus observed before is infinite. In reality, the finiteness of linguistic sequences has the opposite tendency (i.e., the size of the observed corpus is finite, and the possible vocabulary size is infinite).
2.3.2 Approximation of Rényi Entropy Hα. The Rényi entropy is a generalization of the Shannon entropy, defined as follows (Rényi 1960; Rényi 1970; Cover and Thomas 2006; Bromiley, Thacker, and Bouhova-Thacker 2010):
Hα(X) = 11 − α log( ∑
X
Pα(X)) (11)
where α ≥ 0,α = 1. Hα(X) represents different ideas of sequence complexity for different α. For example:
When α = 0, H0(X) indicates the number of distinct occurrences of X. When the limit α → 1 is taken, Equation (11) reduces to the Shannon entropy.
The formula for α = 0 becomes equivalent to the so-called topological entropy (hence, it is another notion of entropy) for certain probability functions (Kitchens 1998) (Cover and Thomas 2006). Note that the number of distinct tokens (i.e., the cardinality of a set) has been used widely as a rough approximation of complexity in computational linguistics. Indeed, in Section 2.1.2, we saw how some candidate constancy measures are based on a token-type relation, such that the number of types is related to the complexity of a text. For texts, note also that the value grows with respect to the text size, unless X is considered, for example, in terms of unigrams of a phonographic alphabet.
For α → 1, there is controversy regarding convergence, as noted in the previous section. Such difficulty in convergence for these α values lies in the nature of linguistic processes, in which the vocabulary set evolves.
This view motivates us to consider α > 1 for Hα(X), since the formula captures complexity by considering linguistic hapax legomena to a lesser degree, thus giving the possibility of convergence. In fact, an approximation of the probability by the relative frequencies of unigrams at α = 2 immediately shows the essential equivalence to Yule’s K, since K from Equation (1) can be rewritten as follows:
mmax∑ m=1 V(m, N)( mN ) 2 = ∑ x∈X ( freq(x) N ) 2
where freq(x) is the frequency of x ∈ X. Therefore, Yule’s K has significance within the context of complexity.
This relation of Yule’s K to the Rényi entropy H2 is reported for the first time here, to the best of our knowledge. This mathematical relation clarifies both why Yule’s K should converge and what the convergent value means; specifically, the value represents the gross complexity underlying the language system. As noted earlier, the higher-order entropy considers hapax legomena to a lesser degree and calculates the gross entropy only from the representative vocabulary population. This simple argument shows that Yule’s K captures not only the simple repetitiveness of vocabulary but also the more profound signification of its equivalence with the approximated second-order entropy. Because K has been previously reported as a stable text constancy measure, we consider it here once again, but this time within the broader context of Hα. Based on the previous reports (Tweedie and Baayen 1998; Kimura and Tanaka-Ishii 2014) and the discussion so far, we consider the following four measures as candidates for text constancy measures.
Repetitiveness-based measures: Yule’s K (Equation (1)); and Golcher’s V (Equation (2)).
Complexity-based measures: The Shannon entropy upper bound (h1 as the entropy rate (Equations (10) and (8)) and H1 (Equation (7), with X in terms of unigrams and the probability function in terms of relative frequencies); and the approximated Rényi entropy, denoted as Hα (α > 1) (Equation (11), again with X and the probability function in terms of unigrams and relative frequencies, respectively).
In addition, we empirically consider how these measures can be understood in the context of the power-law feature of language. As noted in the Introduction, for the convergent measures the speed of attaining convergence with respect to text size is examined as well. Among the candidates, K and H1 have been previously applied in a word-based manner, whereas V is string based. The Shannon entropy rate h1 has been considered in both ways. Because we should be able to consider a text in terms of both words and characters, we examine the constancy of each measure in both ways.
Furthermore, because we have seen the mathematical equivalence of Yule’s K and H2, in the following we only consider H2. As for Hα, we consider α = 3, 4 only in comparison with H2. Because H1 is based on relative frequencies and can be considered together with H2, we first focus on the convergence of the three measures V, h1, and H2, and then we consider H1 in comparison with H2, H3, and H4. 3 data : Table 1 lists the data used in our experimental examination. The table indicates the data identifier (by which we refer to the data in the rest of the article), language, source, number of distinct tokens, data length by total number of tokens, and size in bytes. The first block contains relatively large-scale natural language corpora consisting of texts written by multiple authors, and the second block contains smaller corpora consisting of texts by single authors. The third block contains programming language corpora, and the fourth block contains corpora of unknown scripts, which we examine at the end of this article in Section 4.3.
For the large-scale natural language data, we considered five languages: English, Japanese, Chinese, Arabic, and Thai. These languages were chosen to represent different language families and writing systems. The large-scale corpora in English, Japanese, and Chinese consist of newspapers in chronological order, and the Thai and Arabic corpora include other kinds of texts. The markers ‘w’, ‘c’, and ‘cr’ appearing at the end of every identifier in Table 1 (e.g., Enews-c, Enews-w, and Jnews-cr) indicate text processed through words, characters, and transliterated Roman characters, respectively.
As for the small-scale corpora in the second block, the texts were only considered in terms of words, since verification via characters produced findings consistent with those obtained with the large-scale corpora. The texts were chosen because each was written by a single author but is relatively large.
Here, we summarize our preprocessing procedures. For the annotated Thai NECTEC corpus, texts were tokenized according to the annotation. The preprocessing methods for the other corpora were as follows:
English: NLTK2 was used to tokenize text into words. Japanese: Mecab3 was used for tokenization, and KAKASI4 was used for romanization. Chinese: ICTCLAS20135 was used for tokenization, and the pinyin Python library was used for pinyin romanization. Other European Languages: PunktWordTokenizer6 was used for tokenization.
All the other natural language corpora were tokenized simply using spaces. Following Golcher (2007), who first suggested testing constancy on programming languages, we also collected program sources from different languages (third block in Table 1). The programs were also considered solely in terms of words, not characters. C++ and Python were chosen to represent different abstraction levels, and Lisp was chosen because of its different ordering for function arguments. Source code was collected from language libraries. The programming language texts were preprocessed as follows. Comments in natural language were eliminated (although strings remained in the programs, where each was a literal token). Identical files and copies of sources in large chunks were carefully eliminated, although this process did not completely eliminate redundancy since most programs reuse some previous code. Finally, the programs were tokenized according to the language specifications.7
The last block of the table lists two corpora of unknown scripts. We consider these scripts at the end of this article in Section 4.3, through Figure 5, to show one possible application of the text constancy measures. The first unknown script is that of the Voynich manuscript, a famous text that is undeciphered but hypothesized to have been written in natural language. This corpus is considered in terms of both characters and words, where words were defined via the white space separation in the original text. Given the common understanding that the manuscript seems to have two different parts (Reddy and Knight 2011), we separated it into two parts according to the Currier annotation (identified as A and B, respectively). The second corpus of unknown text consists of the Rongorongo script of Easter Island (Daniels and Bright 1996, Section 13; Orliac 2005; Barthel 2013). This script’s status as natural language is debatable, but if so, it is considered to possess characteristics of both phonographs and ideograms (Pozdniakov and Pozdniakov 2007). Because there are several ways to consider what constitutes a character in this script (Barthel 2013), we calculate values for the two most extreme cases as follows. For corpus RongoA-c, we consider a character inclusive of all adjoining parts (i.e., including accents and ornamental parts). On the other hand, for
2 http://nltk.org. 3 http://mecab.googlecode.com/svn/trunk/mecab/doc/index.html. 4 http://kakasi.namazu.org. 5 http://ictclas.nlpir.org. 6 http://nltk.org. 7 With respect to the Lisp programming language, its culture favors long, hyphenated variable names that
can be almost as long as a sentence. For this work, therefore, Lisp variable names were tokenized by splitting at the hyphens.
corpus RongoB-c, we separate parts as reasonably as possible, among multiple possible separation methods. Because the unit of word in this script is unknown, the Rongorongo script is only considered in terms of characters. The empirical verification of convergence for real data is controversial. We must first note that it does not conform with the standard approach to statistical testing. In the domain of statistics, it is a common understanding that “convergence” cannot be tested. A statistical test raises two contrasting hypotheses—called the null and alternative hypotheses—and calculates a p-value indicating the probability of the null hypothesis to occur. When this p-value is smaller than a certain value, the null hypothesis is considered unable to occur and is thus rejected. For convergence, the null hypothesis corresponds to “not converging,” and the alternative hypothesis, to “converging.” The problem here is that the null hypothesis is always related to the alternative hypothesis to a certain extent, because the difference between convergence and non-convergence is merely a matter of degree. In other words, the notion of convergence for a constancy measure does not conform with the philosophy of statistical testing. Convergence is therefore considered in terms of the distance from convergent values, or in terms of the error with respect to some parameter (such as data size). Such a distance cannot be calculated for real data, however, since the underlying mathematical model is unknown.
To sum up, verification of the convergence of real data must be considered by some other means. Our proposal is to consider convergence in comparison to a set of random data whose process is known. For this random data, we considered two kinds.
The first kind is used to examine data convergence in Section 4.1. This random data was generated from real data by shuffling the original text with respect to certain linguistic units. Tweedie and Baayen (1998) presented results by shuffling words, where the original texts were literary texts by single authors. Here, we generated random data by shuffling (1) words/characters, (2) sentences, or (3) documents. Because these options greatly increased the number of combinations of results, we mainly present the results with option (1) for large-scale data in this article. There are three reasons for this: Convergence must be verified especially at large scale; the most important convergence findings for randomized small-scale data were already reported in Tweedie and Baayen (1998); and the results for options (2) and (3) were situated within the range of option (1) and the original texts.
Randomization of the words and characters of original texts will destroy various linguistic characteristics, such as n-grams and long-range correlation. The convergence properties of the three measures V, h1, and H2 are as follows. The convergence of V is unknown, because it lacks a mathematical background. Even if the value of V did converge, the convergent value for randomized data would differ from that of the original text, since the measure is based on repeated n-grams in the text. h1 converges to the entropy rate of the randomized text, if the data size suffices. This is supported by the mathematical background of the algorithm, which converges to the true entropy rate for stationary data. Even when h1 converges for random data, the convergent value will be larger than that of the original text, because h1 considers the probabilities of n-grams. Lastly, H2 converges to the same point for a randomized text and the original text, because it is the approximated higher-order entropy, such that words and characters are considered to occur independently.
The second kind of random data is used to compare the convergent values of different texts for a constancy measure, as considered in Section 4.2. Random corpora
were generated according to four different distributions: one uniform, and the other three following Zipf distributions with exponents of γ = 0.8, 1.0, and 1.3, respectively, for Equation (3). Because each set of real data consists of different numbers of distinct tokens, ranging from tens to billions, random data sets consisting of 2n distinct tokens for every n = 4 . . .19, were randomly generated for each of the four distributions. We only consider the measures H2 and H0 for these data sets. Both of these measures have convergent values, given a sufficient data size. 4 experimental results :From the previous discussion, we applied the three measures V, h1, and H2 with five large-scale and eight small-scale natural language corpora, three programming language corpora, and two unknown script corpora, in terms of words and characters. Because there were many results for different combinations of measure, data, and token (word or character), this section is structured so that it best highlights our findings. Figures 1, 2, and 3 in this section can be examined in the following manner. The horizontal axis indicates the text size of each corpus, in terms of the number of tokens, on a log scale. Chunks of different text sizes were always taken from the head of the corpus.8 The vertical axis indicates the values of the different measures: V, h1, or H2. Each figure contains multiple lines, each corresponding to a corpus, as indicated in the legends.
First, we consider the results for the large-scale data. Figure 1 shows the different measures for words (left three graphs) and characters (right three graphs). We can see that V increased for both words and characters (top two graphs). Golcher tested his measure on up to 30 megabytes of text in terms of characters (Golcher 2007). We also observed a stable tendency up to around 107 characters. The increase in V became apparent, however, for larger text sizes. Thus, it is difficult to consider V as a constancy measure.
As for the results for h1 (middle graphs), both graphs show a gradual decrease. The tendency was clearer for words than for characters. For some corpora, especially for characters, it was possible to observe some values converging towards h∗. The overall tendency, however, could not be concluded as converging. This result suggests the difficulty in attaining convergence of the entropy rate, even with gigabyte-scale data. From the theoretical background of the Grassberger algorithm, the values would possibly converge with larger-scale data. The continued decrease could be due to multiple reasons, including the possibility of requiring far larger data than that used here, or a discrepancy between linguistic processes and the mathematical model assumed for the Grassberger algorithm.
We tried to estimate h∗ by fitting the Equation (9). For the corpora with good fitting, all of the estimated values were larger than zero, but many of the results could not
8 For real data, this was done without any randomization of the order of texts for all corpora besides Atext-w and Atext-c. The Watan corpus is distributed not in the chronological order of the publishing dates, but as a set of articles grouped into categories (i.e., all articles of one category, then all articles of another category, and so on). Because of this, there is a large skew in the vocabulary distribution, depending on the section of the corpus. We thus randomly reshuffled the articles by categories for the whole corpus before taking chunks of different sizes (always from the beginning) to generate our results. Apart from this, we avoided any arbitrary randomization with respect to the original data summarized in Table 1.
be fitted easily, and the estimated values were unstable due to fluctuation of the lines. Whether a value for h∗ is reached asymptotically and also whether h∗ > 0 remain important questions requiring separate, more extensive mathematical and empirical studies.
In contrast, H2 (or Yule’s K, bottom graphs) showed convergence, already at the level of 105 tokens, for both words and characters. From the previous verification of Yule’s K, we can conclude that H2 is convergent. The final convergent values, however, differed for the various writing systems. We return to this issue in the next section.
To better understand the convergence, Figure 2 shows the results for the corresponding randomized data. As mentioned in Section 3.2, the original texts were
randomized by shuffling words and characters for the data examined by words and characters, respectively. Therefore, all n-gram characteristics existing in the text were destroyed, and what remained were the different words and characters appearing in a random order. Here, we see how the random data’s behavior has some of the theoretical properties of convergence, as summarized in Section 3.2.
As mentioned previously, because V has no mathematical background, its behavior even for uniform random data is unknown, and even if it converged, the convergent value would be smaller than that of the original text. The top two graphs in Figure 2 exhibit some oscillation, especially for randomized Chinese (Cnews-c,w). Such peculiar
oscillation was already reported by Golcher himself (Golcher 2007) for uniformly random data. This was easy to replicate, as reported in Kimura and Tanaka-Ishii (2014), for uniformly random data with the number of distinct tokens up to a hundred. Because the word distribution almost follows Zipf’s law, the vocabulary is not uniformly distributed, yet oscillating results occur for some randomized data in the top left figure. Moreover, the values seem to increase for Japanese and English for words at a larger scale. Although the plots for some scripts seem convergent (top right graph), these convergent values are theoretically different from those of the original texts, if they exist, and this stability is not universal across the different data sets. Given this result, it is doubtful that V is convergent across languages.
In contrast, h1 is mathematically proven to be convergent given infinite-length randomized data, but to larger values than those of the original texts, as mentioned in Section 3.2. The middle two graphs of Figure 2 show the results for h1. The majority of the plots do not reach convergence even at the largest data sizes, but for certain results with characters, especially in the Roman alphabet, the plots seem to go to a convergent value (middle right). All the plots can be extrapolated to converge to a certain entropy rate above zero, although these values are larger than the convergent values—if they ever exist—of the real data. These results confirm the difficulty of judging whether the entropy rates of the original texts are convergent and whether they remain above zero.
Lastly, it is easy to see that H2 is convergent for a randomized text (bottom two graphs), and the convergent values are the same for the cases with and without randomization. In fact, the plots converge to exactly the same points faster and more stably, which shows the effect of randomization.
As for the other randomization options, by sentences and documents, the findings—both the tendencies of the lines and the changes in the values—can be situated in the middle of what we have seen so far. The plots should increasingly fluctuate more like the real data because of the incomplete randomization, in the order of sentences and then documents.
Returning to inspection of the remaining real data, Figure 3 shows V, h1, and H2 in terms of words for the small-scale corpora (left column) and for the programming language texts (right column). For the small-scale corpora, in general, the plots are bunched together, and the results shared the tendencies noted previously for the largescale corpora. V again showed an increase, while h1 showed a tendency to decrease. H2 converged rapidly and was already almost stable at 104 tokens. This again shows how H2 exhibits stable constancy, especially with texts written by single authors.
As for the programming language results, the plots fluctuate more than for the natural language texts because of the redundancy within the program sources. Still, the global tendencies noted so far were just discernible. V had relatively larger values but h1 and H2 had smaller values for programs, as compared to the natural language texts. The differences in value indicate the larger degree of repetitiveness in programs.
Lastly, Figure 4 shows the Hα results for the Wall Street Journal in terms of words in unigrams (Enews-w). The horizontal axis indicates the corpus size, and the vertical axis indicates the approximated entropy value. The different lines represent the results for Hα with α = 1, 2, 3, 4. The two H1 plots represent calculations with and without Laplace smoothing (Manning and Schuetze 1999). We can see that without smoothing, H1 increased, as Tweedie and Baayen (1998) reported, but in contrast to their conclusion, we observe a tendency of convergence for larger-scale data. The increase was due to the influence of low-frequency vocabulary pushing up the entropy. The opposite tendency to decrease was observed for the smoothed probabilities, with the plot eventually converging to the same point as that for the unsmoothed H1 values. The convergence
was by far slower for H1 as compared with that for H2, H3, and H4, which all had attained convergence already at 102 tokens. The convergence values naturally decreased for larger α, although the amount of decrease itself rapidly decreased with larger α.
In answer to Questions 1 and 2 raised in the Introduction—which measures show constancy, with sufficient convergence speed—the empirical conclusion from our data is that Hα with α > 1 showed stable constancy when the values were approximated using relative frequencies. For H1, the convergence was much slower because of the strong influence of low-frequency words. Consequently, the constancy of Hα with α > 1 is attained by representing the gross complexity underlying a text. Now we turn to Question 3 raised in the Introduction and examine the discriminatory power of H2. As Yule intended, does H2 identify authors? Given the influence of different writing systems, as seen previously in Figure 1, we examine the relation between H2 and the number of distinct tokens (the alphabet/vocabulary size). Note that because this number corresponds to H0 in Equation (11), this analysis effectively considers texts on the H0-H2 plane. Since H0 grows according to the text size, unlike H2, the same text size must be used for all corpora in order to meaningfully compare H0 values.9 Given that H2 converges fast, we chose a size of 104 tokens to handle all of the small- and large-scale corpora.
For each of the corpora listed in Table 1 and the second kind of random corpora explained at the end of Section 3.2, Figure 5 plots the values of H2 (vertical axis) and the number of distinct tokens H0 (horizontal) measured for each corpus at a size of 104 tokens. The three large circles are groupings of points. The leftmost group represents news sources in alphabetic characters. All of the romanized Chinese, Japanese, and Arabic texts are located almost at the same vertical location as the English text. This indicates the difficulty for H2 to distinguish natural languages if measured in terms of alphabetic characters. The middle group represents the programming language texts in terms of words. This group is located separately (vertically lower than the natural language corpora in terms of words), so H2 is likely to distinguish between natural languages and programming languages. The rightmost group represents the small-scale corpora. Considering the proximity of these points despite the variety of the content, it is unlikely that H2 can distinguish authors, in contrast to Yule’s hope. Still, these points are located lower than those for news text. Therefore, H2 has the potential to distinguish genre or maybe writing style.
9 Because H0 is not convergent, the horizontal locations remain unstable, unless the tokens are of a phonographic alphabet. In other words, for all word-based results and character-based results not based on a phonographic alphabet, the resulting horizontal locations are changed by increasing the corpus size. As for the random data, the H0 values are convergent, because these data sets have a finite number of distinct tokens. Since H0 is measured only for the first 104 tokens, however, the horizontal locations are underestimated, especially for random data following a Zipf distribution.
The natural language texts located near the line for a Zipf exponent of 0.8 are those of the non-alphabetic writing systems.10 Note that Chinese characters have morphological features, and the Arabic and Thai languages also have flexibility in terms of which units are considered words and morphemes. In other words, the plots closer to the random data with a smaller Zipf exponent are for language corpora of morphemic sequences. The group of plots measured for phonographic scripts is located near the line for a Zipf exponent of 1.0 (the grouping of points in the leftmost circle), which could suggest that morphemes are more randomized units than words. The nature of unknown scripts can also be considered through our understanding thus far. Figure 5 includes plots for the Voynich manuscript in terms of words and characters, and for the Rongorongo script in terms of characters. Like all the data seen in this figure, the points are placed at the H2 values (vertically) for the number of distinct tokens (horizontally) at the specified size of 104 tokens, with the exception of Voynich-A in terms of words. Because this corpus consists of fewer than 104 words (refer to the data length by tokens listed for VoynichA-w in Table 1), its point is located horizontally at the vocabulary size corresponding to the corpus’ maximum size.
For the two Voynich manuscript parts, the plots in terms of words appear near the Arabic corpus for words (Abook-w). For characters, on the other hand, the plots are at the leftmost end of the figure. This was due to overestimation of the total number
10 Note that here we use the values of the Zipf exponent for the random data, and not the estimated exponents for the real data. The rank-frequency distributions of characters, especially for phonetic alphabets, often do not follow a power law.
of characters for the alphabetic texts (e.g., both English and other, romanized language texts), since all ASCII characters, such as colons, periods, and question marks, are counted. Still, the H2 values are located almost at the same position as for the other romanized texts, indicating that the Voinich manuscript has approximately similar complexity. These results suggest the possibility that the Voynich manuscript could have been generated from a source in natural language, possibly written in some script of the abjad type. This supports previous findings (Reddy and Knight 2011; Montemurro and Zanette 2013), which reported the possibility of the Voynich manuscript being in a natural language and the coincidence of its word length distribution with that of Arabic.
On the other hand, the plots for the Rongorongo script appear near the line for a Zipf exponent of 0.8, with RongoA near Arabic in terms of words but RongoB somewhat further down from Japanese in terms of characters. The status of Rongorongo as natural language has been controversial (Pozdniakov and Pozdniakov 2007). Both points in the graph, however, are near many other natural language texts (and not widely separated), making it reasonable to hypothesize that Rongorongo is indeed natural language. The characters can be deemed morphologically rich, because both plots are close to the line for a Zipf exponent of 0.8. In the case of RongoA, for which a character was considered inclusive of all parts (i.e., including accents and ornamental parts), the morphological richness is comparable to that of the words of an abjad script. On the other hand, when considering the different character parts as distinct (RongoB), the location drifts towards the plot for Thai, a phonographic script, in terms of characters. Therefore, the Rongorongo script could be considered basically morphemic, with some parts functioning phonographically. This conclusion again supports a previous hypothesis proposed by a domain specialist (Pozdniakov and Pozdniakov 2007).
This analysis of two unknown scripts supports previous conjectures. Our results, however, only add a small bit of evidence to those conjectures; clearly, reaching a reasonable conclusion would require further study. Moreover, the analysis of unknown scripts introduced here could provide another possible application of text constancy measures, from a broader context. 5 conclusion :We have discussed text constancy measures, whose values are invariant across different sizes of text, for a given text. Such measures have a 70-year history, since Yule originally proposed K as a text characteristic, potentially with language engineering utility for problems such as author identification. We consider text constancy measures today to have scientific importance in understanding language universals from a computational view.
After overviewing measures proposed so far and previous studies on text constancy, we explained how K essentially has a mathematical equivalence to the Rényi higher-order entropy. We then empirically examined various measures across different languages and kinds of corpora. Our results showed that only the approximated higherorder Rényi entropy exhibits stable, rapid constancy. Examining the nature of the convergent values revealed that K does not possess the discriminatory power of author identification as Yule had hoped. We also applied our understanding to two unknown scripts, the Voynich manuscript and Rongorongo, and showed how our constancy results support previous hypotheses about each of these scripts.
Our future work will include application of K to other kinds of data besides natural language. There, too, we will consider the questions raised in the Introduction, of
whether K converges and of how discriminatory it is. We are especially interested in considering the relation between the value of K and the meaningfulness of data. This article presents a mathematical and empirical verification of computational constancy measures for natural language text. A constancy measure characterizes a given text by having an invariant value for any size larger than a certain amount. The study of such measures has a 70-year history dating back to Yule’s K, with the original intended application of author identification. We examine various measures proposed since Yule and reconsider reports made so far, thus overviewing the study of constancy measures. We then explain how K is essentially equivalent to an approximation of the second-order Rényi entropy, thus indicating its signification within language science. We then empirically examine constancy measure candidates within this new, broader context. The approximated higher-order entropy exhibits stable convergence across different languages and kinds of text. We also show, however, that it cannot identify authors, contrary to Yule’s intention. Lastly, we apply K to two unknown scripts, the Voynich manuscript and Rongorongo, and show how the results support previous hypotheses about these scripts. [{""affiliations"": [], ""name"": ""Kumiko Tanaka-Ishii""}, {""affiliations"": [], ""name"": ""Shunsuke Aihara""}] SP:54cf25939e5f7b648f0a97837673bc1f8e10a4be [{""authors"": [""T. Barthel""], ""title"": ""The Rongorongo of Easter Island: Thomas Barthel\u2019s transliteration system"", ""venue"": ""Available at http://kohaumotu. org/rongorongo org/corpus/ codes.html. Accessed June 2015."", ""year"": 2013}, {""authors"": [""T.C. Bell"", ""J.G. Cleary"", ""I.H. Witten.""], ""title"": ""Text Compression"", ""venue"": ""Prentice Hall."", ""year"": 1990}, {""authors"": [""P.A. Bromiley"", ""N.A. Thacker"", ""E. Bouhova-Thacker.""], ""title"": ""Shannon entropy, Renyi entropy, and information"", ""venue"": ""Available at http://www.tina-vision.net/ docs/memos/2004-004.pdf. Accessed"", ""year"": 2010}, {""authors"": [""P.F. Brown"", ""S.A. Della Pietra"", ""V.J. Della Pietra"", ""J.C. Lai"", ""R.L. Mercer.""], ""title"": ""An estimate of an upper bound for the entropy of English"", ""venue"": ""Computational Linguistics, 18(1):31\u201340."", ""year"": 1992}, {""authors"": [""E. Brunet""], ""title"": ""Vocabulaire de Jean Giraudoux: Structure et Evolution"", ""venue"": ""Slatkine."", ""year"": 1978}, {""authors"": [""T. Cover"", ""R. King.""], ""title"": ""A convergent gambling estimate of the entropy of English"", ""venue"": ""IEEE Transactions on Information Theory, 24(4):413\u2013421."", ""year"": 1978}, {""authors"": [""T.M. Cover"", ""J.A. Thomas.""], ""title"": ""Elements of Information Theory"", ""venue"": ""Wiley-Interscience."", ""year"": 2006}, {""authors"": [""\u0141. D\u0229bowski""], ""title"": ""A general definition of conditional information and its application to ergodic decomposition"", ""venue"": ""Statistics and Probability Letters, 79(9):1260\u20131268."", ""year"": 2009}, {""authors"": [""\u0141. D\u0229bowski""], ""title"": ""Empirical evidence for Hilberg\u2019s conjecture in single author texts"", ""venue"": ""Methods and Applications of Quantitative Linguistics: Selected papers of the 8th International Conference on"", ""year"": 2013}, {""authors"": [""\u0141. D\u0229bowski""], ""title"": ""The relaxed Hilberg conjecture: A review and new experimental support"", ""venue"": ""Available at http://www.ipipan.waw.pl/ldebowsk/. Accessed June 2015."", ""year"": 2014}, {""authors"": [""D. Dugast""], ""title"": ""Vocabulaire et Stylistique"", ""venue"": ""I Th\u00e9\u00e2tre et Dialogue. Slatkine-Champion. Travaux de Linguistique Quantitative."", ""year"": 1979}, {""authors"": [""M. Farach"", ""M. Noordewier"", ""S. Savari"", ""L. Shepp"", ""A. Wyner"", ""J. Ziv.""], ""title"": ""On the entropy of DNA: Algorithms and measurements based on memory and rapid convergence"", ""venue"": ""Proceedings of the"", ""year"": 1995}, {""authors"": [""D. Genzel"", ""E. Charniak.""], ""title"": ""Entropy rate constancy in text"", ""venue"": ""Annual Meeting of the Association for the ACL, pages 199\u2013206, Philadelphia, PA."", ""year"": 2002}, {""authors"": [""F. Golcher""], ""title"": ""A stable statistical constant specific for human language texts"", ""venue"": ""Recent Advances in Natural Language Processing, Borovets."", ""year"": 2007}, {""authors"": [""I.J. Good""], ""title"": ""The population frequencies of species and the estimation of population parameters"", ""venue"": ""Biometrika, 40(3\u20134):237\u2013264."", ""year"": 1953}, {""authors"": [""P. Grassberger""], ""title"": ""Estimating the information content of symbol sequences and efficient codes"", ""venue"": ""IEEE Transactions on Information Theory, 35:669\u2013675."", ""year"": 1989}, {""authors"": [""H. Guiraud""], ""title"": ""Les Charact\u00e8res Statistique du Vocabulaire"", ""venue"": ""Universitaires de France Press."", ""year"": 1954}, {""authors"": [""D. Gusfield""], ""title"": ""Algorithms on Strings, and Sequences: Computer Science and Computational Biology"", ""venue"": ""Cambridge University Press."", ""year"": 1997}, {""authors"": [""G. Herdan""], ""title"": ""Quantitative Linguistics"", ""venue"": ""Butterworths."", ""year"": 1964}, {""authors"": [""W. Hilberg""], ""title"": ""Der bekannte grenzwert der redundanzfreien information in texten eine fehlinterpretation der shannonschen experimente? Frequenz, 44(9\u201310):243\u2013248"", ""year"": 1990}, {""authors"": [""A. Honor\u00e9""], ""title"": ""Some simple measures of richness of vocabulary"", ""venue"": ""Association for Literary and Linguistic Computing Bulletin, 7:172\u2013177."", ""year"": 1979}, {""authors"": [""D. Kimura"", ""K. Tanaka-Ishii.""], ""title"": ""A study on constants of natural language texts"", ""venue"": ""Journal of Natural Language Processing, 18(2):119\u2013137."", ""year"": 2011}, {""authors"": [""D. Kimura"", ""K. Tanaka-Ishii.""], ""title"": ""A study on constants of natural language texts"", ""venue"": ""Journal of Natural Language Processing, 21:877\u2013895. Special issue of awarded papers. [The English translated version"", ""year"": 2014}, {""authors"": [""B. Kitchens""], ""title"": ""Symbolic Dynamics: One-sided, Two-sided and Countable State Markov Shifts"", ""venue"": ""Springer. 501"", ""year"": 1998}, {""authors"": [""R. Levy"", ""T.F. Jaeger.""], ""title"": ""Speakers optimize information density through information density through syntactic reduction"", ""venue"": ""Annual Conference on Neural Information"", ""year"": 2007}, {""authors"": [""H.D. Maas""], ""title"": ""Zusammenhang zwischen wortschatzumfang und l\u00e4nge eines textes [Relationship between vocabulary and text length"", ""venue"": ""Zeitschrift f\u00fcr Literaturwissenschaft und Linguistik, 8:73\u201370."", ""year"": 1972}, {""authors"": [""B. Mandelbrot""], ""title"": ""An informational theory of the statistical structure of language"", ""venue"": ""Communication Theory, 486\u2013500."", ""year"": 1953}, {""authors"": [""C. Manning"", ""H. Schuetze.""], ""title"": ""Foundations of Statistical Natural Language Processing"", ""venue"": ""MIT Press."", ""year"": 1999}, {""authors"": [""M. Montemurro"", ""D. Zanette.""], ""title"": ""Keywords and co-occurrence patterns in the Voynich Manuscript: An information-theoretic analysis"", ""venue"": ""PLOS One. doi: 10.1371/journal.pone.0066344."", ""year"": 2013}, {""authors"": [""C. Orliac""], ""title"": ""The Rongorongo tablets from Easter Island: Botanical identification and 14c dating"", ""venue"": ""Archaeology in Oceania, 40(3):115\u2013119."", ""year"": 2005}, {""authors"": [""J.K. Orlov"", ""R.Y. Chitashvili.""], ""title"": ""Generalized z-distribution generating the well-known \u2018rank-distributions"", ""venue"": ""Bulletin of the Academy of Sciences of Georgia, 110:269\u2013272."", ""year"": 1983}, {""authors"": [""K. Pozdniakov"", ""I. Pozdniakov.""], ""title"": ""Rapanui writing and the Rapanui language: Preliminary results of a statistical analysis"", ""venue"": ""Forum for Anthropology and Culture, 3:3\u201336."", ""year"": 2007}, {""authors"": [""S. Reddy"", ""K. Knight.""], ""title"": ""What we know about the Voynich Manuscript"", ""venue"": ""ACL Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities, Portland, OR."", ""year"": 2011}, {""authors"": [""A. R\u00e9nyi""], ""title"": ""On measures of entropy and information"", ""venue"": ""Proceedings of the Fourth Berkeley Symposium on Mathematics,"", ""year"": 1960}, {""authors"": [""A. R\u00e9nyi""], ""title"": ""Foundations of Probability"", ""venue"": ""Dover Publications."", ""year"": 1970}, {""authors"": [""T. Sch\u00fcmann"", ""P. Grassberger.""], ""title"": ""Entropy estimation of symbol sequences"", ""venue"": ""Chaos, 6(3):414\u2013427."", ""year"": 1996}, {""authors"": [""C. Shannon""], ""title"": ""A mathematical theory of communication"", ""venue"": ""Bell System Technical Journal, 27:379\u2013423, 623\u2013656."", ""year"": 1948}, {""authors"": [""P.C. Shields""], ""title"": ""Entropy and prefixes"", ""venue"": ""Annals of Probability, 20(1):403\u2013409."", ""year"": 1992}, {""authors"": [""H.S. Sichel""], ""title"": ""On a distribution law for word frequencies"", ""venue"": ""Journal of the American Statistical Association, 70(351):542\u2013547."", ""year"": 1975}, {""authors"": [""E.H. Simpson""], ""title"": ""Measurement of diversity"", ""venue"": ""Nature, 163:688."", ""year"": 1949}, {""authors"": [""E. Stamatatos"", ""N. Fakotakis"", ""G. Kokkinakis.""], ""title"": ""Automatic text categorization in terms of genre and author"", ""venue"": ""Computational Linguistics, 26(4):471\u2013495."", ""year"": 2001}, {""authors"": [""B. Stein"", ""N. Lipka"", ""P. Prettenhofer.""], ""title"": ""Intrinsic plagiarism analysis"", ""venue"": ""Language Resources and Evaluation, 45(1):63\u201382."", ""year"": 2010}, {""authors"": [""Y.W. Teh""], ""title"": ""A hierarchical Bayesian language model based on Pitman-Yor processes"", ""venue"": ""Proceedings of the 21st International Conference On Computational Linguistics and 44th Annual Meeting of the"", ""year"": 2006}, {""authors"": [""J. Tuldava""], ""title"": ""Quantitative relations between the size of the text and lexical richness"", ""venue"": ""SMIL Quarterly, Journal of Linguistic Calculus, 4:28\u201335."", ""year"": 1977}, {""authors"": [""F.J. Tweedie"", ""R.H. Baayen""], ""title"": ""How variable may a constant be? Measures of lexical richness in perspective"", ""venue"": ""Computers and the Humanities, 32:323\u2013352."", ""year"": 1998}, {""authors"": [""G.U. Yule""], ""title"": ""The Statistical Study of Literary Vocabulary"", ""venue"": ""Cambridge University Press."", ""year"": 1944}, {""authors"": [""G.K. Zipf""], ""title"": ""Human Behavior and the Principle of Least Effort: An Introduction to Human Ecology"", ""venue"": ""Hafner, New York. 502"", ""year"": 1965}] acknowledgments  :This research was supported by JST’s PRESTO program.","5 conclusion :We have discussed text constancy measures, whose values are invariant across different sizes of text, for a given text. Such measures have a 70-year history, since Yule originally proposed K as a text characteristic, potentially with language engineering utility for problems such as author identification. We consider text constancy measures today to have scientific importance in understanding language universals from a computational view.
After overviewing measures proposed so far and previous studies on text constancy, we explained how K essentially has a mathematical equivalence to the Rényi higher-order entropy. We then empirically examined various measures across different languages and kinds of corpora. Our results showed that only the approximated higherorder Rényi entropy exhibits stable, rapid constancy. Examining the nature of the convergent values revealed that K does not possess the discriminatory power of author identification as Yule had hoped. We also applied our understanding to two unknown scripts, the Voynich manuscript and Rongorongo, and showed how our constancy results support previous hypotheses about each of these scripts.
Our future work will include application of K to other kinds of data besides natural language. There, too, we will consider the questions raised in the Introduction, of
whether K converges and of how discriminatory it is. We are especially interested in considering the relation between the value of K and the meaningfulness of data."
"1 introduction :Distributional semantics approximates word meanings with vectors tracking cooccurrence in corpora (Turney and Pantel 2010). Recent work has extended this approach to phrases and sentences through vector composition (Clark 2015). Resulting compositional distributional semantic models (CDSMs) estimate degrees of semantic similarity (or, more generally, relatedness) between two phrases: A good CDSM might tell us that green bird is closer to parrot than to pigeon, useful for tasks such as paraphrasing.
We take a mathematical look1 at how the composition operations postulated by CDSMs affect similarity measurements involving the vectors they produce for phrases or sentences. We show that, for an important class of composition methods, encompassing at least those based on linear transformations, the similarity equations can be decomposed into operations performed on the subparts of the input phrases,
∗ Department of Enterprise Engineering, University of Rome “Tor Vergata,” Viale del Politecnico, 1, 00133 Rome, Italy. E-mail: fabio.massimo.zanzotto@uniroma2.it. 1 Ganesalingam and Herbelot (2013) also present a mathematical investigation of CDSMs. However, except for the tensor product (a composition method we do not consider here as it is not empirically effective), they do not look at how composition strategies affect similarity comparisons.
Original submission received: 10 December 2013; revision received: 26 May 2014; accepted for publication: 1 August 2014.
doi:10.1162/COLI a 00215
© 2015 Association for Computational Linguistics
and typically factorized into terms that reflect the linguistic structure of the input. This establishes a strong link between CDSMs and convolution kernels (Haussler 1999), which act in the same way. We thus refer to our claim as the “Convolution Conjecture.”
We focus on the models in Table 1. These CDSMs all apply linear methods, and we suspect that linearity is a sufficient (but not necessary) condition to ensure that the Convolution Conjecture holds. We will first illustrate the conjecture for linear methods, and then briefly consider two nonlinear approaches: the dual space model of Turney (2012), for which it does, and a representative of the recent strand of work on neuralnetwork models of composition, for which it does not.","2 mathematical preliminaries :Vectors are represented as small letters with an arrow a and their elements are ai, matrices as capital letters in bold A and their elements are Aij, and third-order or fourth-order tensors as capital letters in the form A and their elements are Aijk or Aijkh. The symbol represents the element-wise product and ⊗ is the tensor product. The dot product is 〈 a, b〉 and the Frobenius product—that is, the generalization of the dot product to matrices and high-order tensors—is represented as 〈A, B〉F and 〈A,B〉F. The Frobenius product acts on vectors, matrices, and third-order tensors as follows:
〈 a, b〉F = ∑
i
aibi = 〈 a, b〉 〈A, B〉F = ∑ ij AijBij 〈A,B〉F = ∑
ijk
AijkBijk (1)
A simple property that relates the dot product between two vectors and the Frobenius product between two general tensors is the following:
〈 a, b〉 = 〈I, a bT〉F (2)
where I is the identity matrix. The dot product of A x and B y can be rewritten as:
〈A x, B y〉 = 〈ATB, x yT〉F (3)
Let A and B be two third-order tensors and x, y, a, c four vectors. It can be shown that:
〈 xA y, aB c 〉 = 〈 ∑
j
(A ⊗ B)j, x ⊗ y ⊗ a ⊗ c 〉F (4)
where C = ∑
j(A ⊗ B)j is a non-standard way to indicate the tensor contraction of the tensor product between two third-order tensors. In this particular tensor contraction, the elements Ciknm of the resulting fourth-order tensor C are Ciknm = ∑ j AijkBnjm. The elements Diknm of the tensor D = x ⊗ y ⊗ a ⊗ c are Diknm = xiykancm.","3 formalizing the convolution conjecture :Structured Objects. In line with Haussler (1999), a structured object x ∈ X is either a terminal object that cannot be furthermore decomposed, or a non-terminal object that can be decomposed into n subparts. We indicate with x = (x1, . . . , xn) one such decomposition, where the subparts xi ∈ X are structured objects themselves. The set X is the set of the structured objects and TX ⊆ X is the set of the terminal objects. A structured object x can be anything according to the representational needs. Here, x is a representation of a text fragment, and so it can be a sequence of words, a sequence of words along with their part of speech, a tree structure, and so on. The set R(x) is the set of decompositions of x relevant to define a specific CDSM. Note that a given decomposition of a structured object x does not need to contain all the subparts of the original object. For example, let us consider the phrase x = tall boy. We can then define R(x) = {(tall, boy), (tall), (boy)}. This set contains the three possible decompositions of the phrase: ( tall︸︷︷︸
x1 , boy︸︷︷︸ x2 ), ( tall︸︷︷︸ x1 ), and (boy︸︷︷︸ x1 ).
Recursive formulation of CDSM. A CDSM can be viewed as a function f that acts recursively on a structured object x. If x is a non-terminal object
f (x) = ⊙
x∈R(x) γ( f (x1), f (x2), . . . , f (xn)) (5)
where R(x) is the set of relevant decompositions, ⊙ is a repeated operation on this set, γ is a function defined on f (xi) where xi are the subparts of a decomposition of x. If x is a terminal object, f (x) is directly mapped to a tensor. The function f may operate differently on different kinds of structured objects, with tensor degree varying accordingly. The set R(x) and the functions f , γ, and ⊙ depend on the specific CDSM, and the same CDSM might be susceptible to alternative analyses satisfying the form in Equation (5). As an example, under Additive, x is a sequence of words and f is
f (x) = ⎧⎨ ⎩ ∑ y∈R(x) f (y) if x /∈ TX
x if x ∈ TX (6)
where R((w1, . . . , wn)) = {(w1), . . . , (wn)}. The repeated operation ⊙
corresponds to summing and γ is identity. For Multiplicative we have
f (x) = ⎧⎨ ⎩ y∈R(x) f (y) if x /∈ TX
x if x ∈ TX (7)
where R(x) = {(w1, . . . , wn)} (a single trivial decomposition including all subparts). With a single decomposition, the repeated operation reduces to a single term; and here γ is the product (it will be clear subsequently, when we apply the Convolution Conjecture to these models, why we are assuming different decomposition sets for Additive and Multiplicative).
Definition 1 (Convolution Conjecture) For every CDSM f along with its R(x) set, there exist functions K, Ki and a function g such that:
K( f (x), f (y)) = ∑
x∈R(x) y∈R(y)
g(K1( f (x1), f (y1)), K2( f (x2), f (y2)), . . . , Kn( f (xn), f (yn))) (8)
The Convolution Conjecture postulates that the similarity K( f (x), f (y)) between the tensors f (x) and f (y) is computed by combining operations on the subparts, that is, Ki( f (xi), f (yi)), using the function g. This is exactly what happens in convolution kernels (Haussler 1999). K is usually the dot product, but this is not necessary: We will show that for the dual-space model of Turney (2012) K turns out to be the fourth root of the Frobenius tensor.","4 comparing composed phrases :We illustrate now how the Convolution Conjecture (CC) applies to the considered CDSMs, exemplifying with adjective–noun and subject–verb–object phrases. Without loss of generality we use tall boy and red cat for adjective–noun phrases and goats eat grass and cows drink water for subject–verb–object phrases.
Additive Model. K and Ki are dot products, g is the identity function, and f is as in Equation (6). The structure of the input is a word sequence (i.e., x = (w1 w2)) and the relevant decompositions consist of these single words, R(x) = {(w1), (w2)}. Then
K( f (tall boy), f (red cat)) = 〈 tall + boy, red + cats〉 = = 〈 tall, red〉+ 〈 tall, cat〉+ 〈 boy, red〉+ 〈 boy, cat〉 = =
∑ x∈{tall,boy} y∈{red,cat} 〈 f (x), f (y)〉 = ∑ x∈{tall,boy} y∈{red,cat} K( f (x), f (y)) (9)
The CC form of Additive shows that the overall dot product can be decomposed into dot products of the vectors of the single words. Composition does not add any further information. These results can be easily extended to longer phrases and to phrases of different length.
Multiplicative Model. K, g are dot products, Ki the component-wise product, and f is as in Equation (7). The structure of the input is x = (w1 w2), and we use the trivial single decomposition consisting of all subparts (thus summation reduces to a single term):
K( f (tall boy), f (red cat)) = 〈 tall boy, red cat〉 = 〈 tall red boy cat, 1〉 = = 〈 tall red, boy cat〉 = g(K1( tall, red), K2( boy, cat)) (10)
This is the dot product between an indistinct chain of element-wise products and a vector 1 of all ones or the product of two separate element-wise products, one on adjectives tall red, and one on nouns boy cat. In this latter CC form, the final dot product is obtained in two steps: first separately operating on the adjectives and on the nouns; then taking the dot product of the resulting vectors. The comparison operations are thus reflecting the input syntactic structure. The results can be easily extended to longer phrases and to phrases of different lengths.
Full Additive Model. The input consists of a sequence of (label,word) pairs x = ((L1 w1), . . . , (Ln wn)) and the relevant decomposition set includes the single tuples, that is, R(x) = {(L1 w1), . . . , (Ln wn)}. The CDSM f is defined as
f (x) = ⎧⎪⎪⎨ ⎪⎪⎩ ∑ (L w)∈R(x) f (L)f (w) if x /∈ TX
X if x ∈ TX is a label L w if x ∈ TX is a word w
(11)
The repeated operation ⊙
here is summation, and γ the matrix-by-vector product. In the CC form, K is the dot product, g the Frobenius product, K1( f (x), f (y)) = f (x)Tf (y), and K2( f (x), f (y)) = f (x)f (y)T. We have then for adjective–noun composition (by using the property in Equation (3)):
K( f ((A tall) (N boy)), f ((A red) (N cat))) = 〈A tall + N boy, A red + N cat〉 = = 〈A tall, A red〉+ 〈A tall, N cat〉+ 〈N boy, A red〉+ 〈N boy, N cat〉 = = 〈ATA, tall redT〉F + 〈NTA, boy red
T〉F + 〈ATN, tall catT〉F + 〈NTN, boy catT〉F = = ∑ (lx wx)∈{(A tall),(N boy)} (ly wy )∈{(A red),(N cat)} g(K1( f (lx), f (ly)), K2( f (wx), f (wy)) (12)
The CC form shows how Full Additive factorizes into a more structural and a more lexical part: Each element of the sum is the Frobenius product between the product of two matrices representing syntactic labels and the tensor product between
two vectors representing the corresponding words. For subject–verb–object phrases ((S w1) (V w2) (O w3)) we have
K( f (((S goats) (V eat) (O grass))), f (((S cows) (V drink) (O water)))) = = 〈S goats + V eat + O grass, S cows + V drink + O water〉 = = 〈STS, goats cowsT〉F + 〈STV, goats drink
T〉F + 〈STO, goats waterT〉F +〈VTS, eat cowsT〉F + 〈VTV, eat drink
T〉F + 〈VTO, eat waterT〉F +〈OTS, grass cowsT〉F + 〈OTV, grass drink
T〉F + 〈OTO, grass waterT〉F = ∑ (lx wx)∈{(S goats),(V eat),(O grass)}
(ly wy )∈{(S cows),(V drink),(O water)}
g(K1( f (lx), f (ly)), K2( f (wx), f (wy))
(13)
Again, we observe the factoring into products of syntactic and lexical representations. By looking at Full Additive in the CC form, we observe that when XTY ≈ I for all matrix pairs, it degenerates to Additive. Interestingly, Full Additive can also approximate a semantic convolution kernel (Mehdad, Moschitti, and Zanzotto 2010), which combines dot products of elements in the same slot. In the adjective–noun case, we obtain this approximation by choosing two nearly orthonormal matrices A and N such that AAT = NNT ≈ I and ANT ≈ 0 and applying Equation (2): 〈A tall + N boy, A red + N cat〉 ≈ 〈tall, red〉+ 〈boy, cat〉.
This approximation is valid also for three-word phrases. When the matrices S, V, and O are such that XXT ≈ I with X one of the three matrices and YXT ≈ 0 with X and Y two different matrices, Full Additive approximates a semantic convolution kernel comparing two sentences by summing the dot products of the words in the same role, that is,
〈S goats + V eat + O grass, S cows + V drink + O water〉 ≈ ≈ 〈goats, cows〉+ 〈eat, drink〉+ 〈grass, water〉 (14)
Results can again be easily extended to longer and different-length phrases.
Lexical Function Model. We distinguish composition with one- vs. two argument predicates. We illustrate the first through adjective–noun composition, where the adjective acts as the predicate, and the second with transitive verb constructions. Although we use the relevant syntactic labels, the formulas generalize to any construction with the same argument count. For adjective–noun phrases, the input is a sequence of (label, word) pairs (x = ((A, w1), (N, w2))) and the relevant decomposition set again includes only the single trivial decomposition into all the subparts: R(x) = {((A, w1), (N, w2))}. The method itself is recursively defined as
f (x) = ⎧⎪⎨ ⎪⎩ f ((A, w1))f ((N, w2)) if x /∈ TX = ((A, w1), (N, w2)) W1 if x ∈ Tx = (A, w1) w2 if x ∈ Tx = (N, w2)
(15)
Here, K and g are, respectively, the dot and Frobenius product, K1( f (x), f (y)) = f (x)Tf (y), and K2( f (x), f (y)) = f (x)f (y)T. Using Equation (3), we have then
K( f (tall boy)), f (red cat)) = 〈TALL boy, RED cat〉 = = 〈TALLTRED, boy catT〉F = g(K1( f (tall), f (red)), K2( f (boy), f (cat)))
(16)
The role of predicate and argument words in the final dot product is clearly separated, showing again the structure-sensitive nature of the decomposition of the comparison operations. In the two-place predicate case, again, the input is a set of (label, word) tuples, and the relevant decomposition set only includes the single trivial decomposition into all subparts. The CDSM f is defined as
f (x) = ⎧⎪⎨ ⎪⎩ f ((S w1)) ⊗ f ((V w2)) ⊗ f ((O w3)) if x /∈ TX = ((S w1) (V w2) (O w3)) w if x ∈ TX = (l w) and l is S or O W if x ∈ TX = (V w) (17)
K is the dot product and g(x, y, z) = 〈x, y ⊗ z〉F, K1( f (x), f (y)) = ∑
j( f (x) ⊗ f (y))j— that is, the tensor contraction2 along the second index of the tensor product between f (x) and f (y)—and K2( f (x), f (y)) = K3( f (x), f (y)) = f (x) ⊗ f (y) are tensor products. The dot product of goats EAT grass and cows DRINK water is (by using Equation (4))
K( f (((S goats) (V eat) (O grass))), f (((S cows) (V drink) (O water)))) = = 〈 goats EAT grass, cows DRINK water〉 = = 〈∑j(EAT ⊗ DRINK)j, goats ⊗ grass ⊗ cows ⊗ water〉F = = g(K1( f ((V eat)), f ((V drink))),
K2( f ((S goats)), f ((S cows))) ⊗ K3( f ((O grass)), f ((O water))))
(18)
We rewrote the equation as a Frobenius product between two fourth-order tensors. The first combines the two third-order tensors of the verbs ∑ j(EAT ⊗ DRINK)j and the second combines the vectors representing the arguments of the verb, that is: goats ⊗ grass ⊗ cows ⊗ water. In this case as well we can separate the role of predicate and argument types in the comparison computation. Extension of the Lexical Function to structured objects of different lengths is treated by using the identity element for missing parts. As an example, we show here the comparison between tall boy and cat where the identity element is the identity matrix I:
K( f (tall boy)), f (cat)) = 〈TALL boy, cat〉 = 〈TALL boy, I cat〉 = = 〈TALLTI, boy catT〉F = g(K1( f (tall), f ( )), K2( f (boy), f (cat)))
(19)
Dual Space Model. We have until now applied the CC to linear CDSMs with the dot product as the final comparison operator (what we called K). The CC also holds for the effective Dual Space model of Turney (2012), which assumes that each word has two distributional representations, wd in “domain” space and wf in “function” space. The similarity of two phrases is directly computed as the geometric average of the separate similarities between the first and second words in both spaces. Even though
2 Grefenstette et al. (2013) first framed the Lexical Function in terms of tensor contraction.
there is no explicit composition step, it is still possible to put the model in CC form. Take x = (x1, x2) and its trivial decomposition. Define, for a word w with vector representations wd and wf : f (w) = wd wf T. Define also K1( f (x1), f (y1)) = √〈 f (x1), f (y1)〉F,
K2( f (x2), f (y2)) = √〈 f (x2), f (y2)〉F and g(a, b) to be √ab. Then
g(K1( f (x1), f (y1)), K2( f (x2), f (y2))) =
= √√ 〈 xd1 xf 1T, yd1 yf 1T〉F · √ 〈 xd2 xf 2T, yd2 yf 2T〉F =
= 4 √ 〈 xd1, yd1〉 · 〈 xf 1, yf 1〉 · 〈 xd2, yd2〉 · 〈 xf 2, yf 2〉 =
= geo(sim(xd1, yd1), sim(xd2, yd2), sim(xf 1, yf 1), sim(xf 2, yf 2))
(20)
A Neural-network-like Model. Consider the phrase (w1, w2, . . . , wn) and the model defined by f (x) = σ( w1 + w2 + . . .+ wn), where σ(·) is a component-wise logistic function. Here we have a single trivial decomposition that includes all the subparts, and γ(x1, . . . , xn) is defined as σ(x1 + . . .+ xn). To see that for this model the CC cannot hold, consider two two-word phrases (a b) and (c d)
K( f ((a, b)), f ((c, d))) = 〈 f ((a, b)), f ((c, d))〉 = ∑i [ σ( a + b) ] i · [ σ( c + d) ] i
= ∑
i
( 1 + e−ai−bi + e−ci−di + e−ai−bi−ci−di )−1 (21)
We need to rewrite this as
g(K1( a, c), K2( b, d)) (22)
But there is no possible choice of g, K1, and K2 that allows Equation (21) to be written as Equation (22). This example can be regarded as a simplified version of the neuralnetwork model of Socher et al. (2011). The fact that the CC does not apply to it suggests that it will not apply to other models in this family.","5 conclusion :The Convolution Conjecture offers a general way to rewrite the phrase similarity computations of CDSMs by highlighting the role played by the subparts of a composed representation. This perspective allows for a better understanding of the exact operations that a composition model applies to its input. The Convolution Conjecture also suggests a strong connection between CDSMs and semantic convolution kernels. This link suggests that insights from the CDSM literature could be directly integrated in the development of convolution kernels, with all the benefits offered by this wellunderstood general machine-learning framework.",,,,"Distributional semantics has been extended to phrases and sentences by means of composition operations. We look at how these operations affect similarity measurements, showing that similarity equations of an important class of composition methods can be decomposed into operations performed on the subparts of the input phrases. This establishes a strong link between these models and convolution kernels.","[{""affiliations"": [], ""name"": ""Fabio Massimo Zanzotto""}, {""affiliations"": [], ""name"": ""Lorenzo Ferrone""}, {""affiliations"": [], ""name"": ""Marco Baroni""}]",SP:1111fb9daa320f08257d60408e9f6a49aad5c0c6,"[{""authors"": [""Clark"", ""Stephen.""], ""title"": ""Vector space models of lexical meaning"", ""venue"": ""Shalom Lappin and Chris Fox, editors, Handbook of Contemporary Semantics, 2nd ed. Blackwell, Malden, MA. In press."", ""year"": 2015}, {""authors"": [""Coecke"", ""Bob"", ""Mehrnoosh Sadrzadeh"", ""Stephen Clark.""], ""title"": ""Mathematical foundations for a compositional distributional model of meaning"", ""venue"": ""Linguistic Analysis, 36:345\u2013384."", ""year"": 2010}, {""authors"": [""Ganesalingam"", ""Mohan"", ""Aur\u00e9lie Herbelot.""], ""title"": ""Composing distributions: Mathematical structures and their linguistic interpretation"", ""venue"": ""Working paper, Computer Laboratory, University"", ""year"": 2013}, {""authors"": [""Grefenstette"", ""Edward"", ""Georgiana Dinu"", ""Yao-Zhong Zhang"", ""Mehrnoosh Sadrzadeh"", ""Marco Baroni.""], ""title"": ""Multi-step regression learning for compositional distributional semantics"", ""venue"": ""Proceedings"", ""year"": 2013}, {""authors"": [""Guevara"", ""Emiliano.""], ""title"": ""A regression model of adjective-noun compositionality in distributional semantics"", ""venue"": ""Proceedings of GEMS, pages 33\u201337, Uppsala."", ""year"": 2010}, {""authors"": [""Haussler"", ""David.""], ""title"": ""Convolution kernels on discrete structures"", ""venue"": ""Technical report USCS-CL-99-10, University of California at Santa Cruz."", ""year"": 1999}, {""authors"": [""Mitchell"", ""Jeff"", ""Mirella Lapata.""], ""title"": ""Vector-based models of semantic composition"", ""venue"": ""Proceedings of ACL, pages 236\u2013244, Columbus, OH."", ""year"": 2008}, {""authors"": [""Socher"", ""Richard"", ""Eric Huang"", ""Jeffrey Pennin"", ""Andrew Ng"", ""Christopher Manning.""], ""title"": ""Dynamic pooling and unfolding recursive autoencoders for paraphrase detection"", ""venue"": ""Proceedings of NIPS,"", ""year"": 2011}, {""authors"": [""Turney"", ""Peter.""], ""title"": ""Domain and function: A dual-space model of semantic relations and compositions"", ""venue"": ""Journal of Artificial Intelligence Research, 44:533\u2013585."", ""year"": 2012}, {""authors"": [""Turney"", ""Peter"", ""Patrick Pantel.""], ""title"": ""From frequency to meaning: Vector space models of semantics"", ""venue"": ""Journal of Artificial Intelligence Research, 37:141\u2013188."", ""year"": 2010}, {""authors"": [""Zanzotto"", ""Fabio Massimo"", ""Ioannis Korkontzelos"", ""Francesca Falucchi"", ""Suresh Manandhar.""], ""title"": ""Estimating linear models for compositional distributional semantics"", ""venue"": ""Proceedings of COLING,"", ""year"": 2010}]","acknowledgments  :We thank the reviewers for helpful comments. Marco Baroni acknowledges ERC 2011 Starting Independent Research Grant n. 283554 (COMPOSES).
References Clark, Stephen. 2015. Vector space models
of lexical meaning. In Shalom Lappin and Chris Fox, editors, Handbook of Contemporary Semantics, 2nd ed. Blackwell, Malden, MA. In press.
Coecke, Bob, Mehrnoosh Sadrzadeh, and Stephen Clark. 2010. Mathematical foundations for a compositional distributional model of meaning. Linguistic Analysis, 36:345–384. Ganesalingam, Mohan and Aurélie Herbelot. 2013. Composing distributions: Mathematical structures and their linguistic interpretation. Working paper, Computer Laboratory, University of Cambridge. Available at www.cl.cam.ac.uk/∼ah433/. Grefenstette, Edward, Georgiana Dinu, Yao-Zhong Zhang, Mehrnoosh Sadrzadeh, and Marco Baroni. 2013. Multi-step regression learning for compositional distributional semantics. Proceedings of IWCS, pages 131–142, Potsdam. Guevara, Emiliano. 2010. A regression model of adjective-noun compositionality in distributional semantics. In Proceedings of GEMS, pages 33–37, Uppsala. Haussler, David. 1999. Convolution kernels on discrete structures. Technical report USCS-CL-99-10, University of California at Santa Cruz. Mehdad, Yashar, Alessandro Moschitti, and Fabio Massimo Zanzotto. 2010.
Syntactic/semantic structures for textual entailment recognition. In Proceedings of NAACL, pages 1,020–1,028, Los Angeles, CA.
Mitchell, Jeff and Mirella Lapata. 2008. Vector-based models of semantic composition. In Proceedings of ACL, pages 236–244, Columbus, OH. Socher, Richard, Eric Huang, Jeffrey Pennin, Andrew Ng, and Christopher Manning. 2011. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In Proceedings of NIPS, pages 801–809, Granada. Turney, Peter. 2012. Domain and function: A dual-space model of semantic relations and compositions. Journal of Artificial Intelligence Research, 44:533–585. Turney, Peter and Patrick Pantel. 2010. From frequency to meaning: Vector space models of semantics. Journal of Artificial Intelligence Research, 37:141–188. Zanzotto, Fabio Massimo, Ioannis Korkontzelos, Francesca Falucchi, and Suresh Manandhar. 2010. Estimating linear models for compositional distributional semantics. In Proceedings of COLING, pages 1,263–1,271, Beijing.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,when the whole is not greater :,than the combination of its parts: :,a “decompositional” look at :,"compositional distributional semantics :Fabio Massimo Zanzotto∗ University of Rome “Tor Vergata”
Lorenzo Ferrone University of Rome “Tor Vergata”
Marco Baroni University of Trento
Distributional semantics has been extended to phrases and sentences by means of composition operations. We look at how these operations affect similarity measurements, showing that similarity equations of an important class of composition methods can be decomposed into operations performed on the subparts of the input phrases. This establishes a strong link between these models and convolution kernels.",,,,,,,,,,,,,,,,"1 introduction :Distributional semantics approximates word meanings with vectors tracking cooccurrence in corpora (Turney and Pantel 2010). Recent work has extended this approach to phrases and sentences through vector composition (Clark 2015). Resulting compositional distributional semantic models (CDSMs) estimate degrees of semantic similarity (or, more generally, relatedness) between two phrases: A good CDSM might tell us that green bird is closer to parrot than to pigeon, useful for tasks such as paraphrasing.
We take a mathematical look1 at how the composition operations postulated by CDSMs affect similarity measurements involving the vectors they produce for phrases or sentences. We show that, for an important class of composition methods, encompassing at least those based on linear transformations, the similarity equations can be decomposed into operations performed on the subparts of the input phrases,
∗ Department of Enterprise Engineering, University of Rome “Tor Vergata,” Viale del Politecnico, 1, 00133 Rome, Italy. E-mail: fabio.massimo.zanzotto@uniroma2.it. 1 Ganesalingam and Herbelot (2013) also present a mathematical investigation of CDSMs. However, except for the tensor product (a composition method we do not consider here as it is not empirically effective), they do not look at how composition strategies affect similarity comparisons.
Original submission received: 10 December 2013; revision received: 26 May 2014; accepted for publication: 1 August 2014.
doi:10.1162/COLI a 00215
© 2015 Association for Computational Linguistics
and typically factorized into terms that reflect the linguistic structure of the input. This establishes a strong link between CDSMs and convolution kernels (Haussler 1999), which act in the same way. We thus refer to our claim as the “Convolution Conjecture.”
We focus on the models in Table 1. These CDSMs all apply linear methods, and we suspect that linearity is a sufficient (but not necessary) condition to ensure that the Convolution Conjecture holds. We will first illustrate the conjecture for linear methods, and then briefly consider two nonlinear approaches: the dual space model of Turney (2012), for which it does, and a representative of the recent strand of work on neuralnetwork models of composition, for which it does not. 2 mathematical preliminaries :Vectors are represented as small letters with an arrow a and their elements are ai, matrices as capital letters in bold A and their elements are Aij, and third-order or fourth-order tensors as capital letters in the form A and their elements are Aijk or Aijkh. The symbol represents the element-wise product and ⊗ is the tensor product. The dot product is 〈 a, b〉 and the Frobenius product—that is, the generalization of the dot product to matrices and high-order tensors—is represented as 〈A, B〉F and 〈A,B〉F. The Frobenius product acts on vectors, matrices, and third-order tensors as follows:
〈 a, b〉F = ∑
i
aibi = 〈 a, b〉 〈A, B〉F = ∑ ij AijBij 〈A,B〉F = ∑
ijk
AijkBijk (1)
A simple property that relates the dot product between two vectors and the Frobenius product between two general tensors is the following:
〈 a, b〉 = 〈I, a bT〉F (2)
where I is the identity matrix. The dot product of A x and B y can be rewritten as:
〈A x, B y〉 = 〈ATB, x yT〉F (3)
Let A and B be two third-order tensors and x, y, a, c four vectors. It can be shown that:
〈 xA y, aB c 〉 = 〈 ∑
j
(A ⊗ B)j, x ⊗ y ⊗ a ⊗ c 〉F (4)
where C = ∑
j(A ⊗ B)j is a non-standard way to indicate the tensor contraction of the tensor product between two third-order tensors. In this particular tensor contraction, the elements Ciknm of the resulting fourth-order tensor C are Ciknm = ∑ j AijkBnjm. The elements Diknm of the tensor D = x ⊗ y ⊗ a ⊗ c are Diknm = xiykancm. 3 formalizing the convolution conjecture :Structured Objects. In line with Haussler (1999), a structured object x ∈ X is either a terminal object that cannot be furthermore decomposed, or a non-terminal object that can be decomposed into n subparts. We indicate with x = (x1, . . . , xn) one such decomposition, where the subparts xi ∈ X are structured objects themselves. The set X is the set of the structured objects and TX ⊆ X is the set of the terminal objects. A structured object x can be anything according to the representational needs. Here, x is a representation of a text fragment, and so it can be a sequence of words, a sequence of words along with their part of speech, a tree structure, and so on. The set R(x) is the set of decompositions of x relevant to define a specific CDSM. Note that a given decomposition of a structured object x does not need to contain all the subparts of the original object. For example, let us consider the phrase x = tall boy. We can then define R(x) = {(tall, boy), (tall), (boy)}. This set contains the three possible decompositions of the phrase: ( tall︸︷︷︸
x1 , boy︸︷︷︸ x2 ), ( tall︸︷︷︸ x1 ), and (boy︸︷︷︸ x1 ).
Recursive formulation of CDSM. A CDSM can be viewed as a function f that acts recursively on a structured object x. If x is a non-terminal object
f (x) = ⊙
x∈R(x) γ( f (x1), f (x2), . . . , f (xn)) (5)
where R(x) is the set of relevant decompositions, ⊙ is a repeated operation on this set, γ is a function defined on f (xi) where xi are the subparts of a decomposition of x. If x is a terminal object, f (x) is directly mapped to a tensor. The function f may operate differently on different kinds of structured objects, with tensor degree varying accordingly. The set R(x) and the functions f , γ, and ⊙ depend on the specific CDSM, and the same CDSM might be susceptible to alternative analyses satisfying the form in Equation (5). As an example, under Additive, x is a sequence of words and f is
f (x) = ⎧⎨ ⎩ ∑ y∈R(x) f (y) if x /∈ TX
x if x ∈ TX (6)
where R((w1, . . . , wn)) = {(w1), . . . , (wn)}. The repeated operation ⊙
corresponds to summing and γ is identity. For Multiplicative we have
f (x) = ⎧⎨ ⎩ y∈R(x) f (y) if x /∈ TX
x if x ∈ TX (7)
where R(x) = {(w1, . . . , wn)} (a single trivial decomposition including all subparts). With a single decomposition, the repeated operation reduces to a single term; and here γ is the product (it will be clear subsequently, when we apply the Convolution Conjecture to these models, why we are assuming different decomposition sets for Additive and Multiplicative).
Definition 1 (Convolution Conjecture) For every CDSM f along with its R(x) set, there exist functions K, Ki and a function g such that:
K( f (x), f (y)) = ∑
x∈R(x) y∈R(y)
g(K1( f (x1), f (y1)), K2( f (x2), f (y2)), . . . , Kn( f (xn), f (yn))) (8)
The Convolution Conjecture postulates that the similarity K( f (x), f (y)) between the tensors f (x) and f (y) is computed by combining operations on the subparts, that is, Ki( f (xi), f (yi)), using the function g. This is exactly what happens in convolution kernels (Haussler 1999). K is usually the dot product, but this is not necessary: We will show that for the dual-space model of Turney (2012) K turns out to be the fourth root of the Frobenius tensor. 4 comparing composed phrases :We illustrate now how the Convolution Conjecture (CC) applies to the considered CDSMs, exemplifying with adjective–noun and subject–verb–object phrases. Without loss of generality we use tall boy and red cat for adjective–noun phrases and goats eat grass and cows drink water for subject–verb–object phrases.
Additive Model. K and Ki are dot products, g is the identity function, and f is as in Equation (6). The structure of the input is a word sequence (i.e., x = (w1 w2)) and the relevant decompositions consist of these single words, R(x) = {(w1), (w2)}. Then
K( f (tall boy), f (red cat)) = 〈 tall + boy, red + cats〉 = = 〈 tall, red〉+ 〈 tall, cat〉+ 〈 boy, red〉+ 〈 boy, cat〉 = =
∑ x∈{tall,boy} y∈{red,cat} 〈 f (x), f (y)〉 = ∑ x∈{tall,boy} y∈{red,cat} K( f (x), f (y)) (9)
The CC form of Additive shows that the overall dot product can be decomposed into dot products of the vectors of the single words. Composition does not add any further information. These results can be easily extended to longer phrases and to phrases of different length.
Multiplicative Model. K, g are dot products, Ki the component-wise product, and f is as in Equation (7). The structure of the input is x = (w1 w2), and we use the trivial single decomposition consisting of all subparts (thus summation reduces to a single term):
K( f (tall boy), f (red cat)) = 〈 tall boy, red cat〉 = 〈 tall red boy cat, 1〉 = = 〈 tall red, boy cat〉 = g(K1( tall, red), K2( boy, cat)) (10)
This is the dot product between an indistinct chain of element-wise products and a vector 1 of all ones or the product of two separate element-wise products, one on adjectives tall red, and one on nouns boy cat. In this latter CC form, the final dot product is obtained in two steps: first separately operating on the adjectives and on the nouns; then taking the dot product of the resulting vectors. The comparison operations are thus reflecting the input syntactic structure. The results can be easily extended to longer phrases and to phrases of different lengths.
Full Additive Model. The input consists of a sequence of (label,word) pairs x = ((L1 w1), . . . , (Ln wn)) and the relevant decomposition set includes the single tuples, that is, R(x) = {(L1 w1), . . . , (Ln wn)}. The CDSM f is defined as
f (x) = ⎧⎪⎪⎨ ⎪⎪⎩ ∑ (L w)∈R(x) f (L)f (w) if x /∈ TX
X if x ∈ TX is a label L w if x ∈ TX is a word w
(11)
The repeated operation ⊙
here is summation, and γ the matrix-by-vector product. In the CC form, K is the dot product, g the Frobenius product, K1( f (x), f (y)) = f (x)Tf (y), and K2( f (x), f (y)) = f (x)f (y)T. We have then for adjective–noun composition (by using the property in Equation (3)):
K( f ((A tall) (N boy)), f ((A red) (N cat))) = 〈A tall + N boy, A red + N cat〉 = = 〈A tall, A red〉+ 〈A tall, N cat〉+ 〈N boy, A red〉+ 〈N boy, N cat〉 = = 〈ATA, tall redT〉F + 〈NTA, boy red
T〉F + 〈ATN, tall catT〉F + 〈NTN, boy catT〉F = = ∑ (lx wx)∈{(A tall),(N boy)} (ly wy )∈{(A red),(N cat)} g(K1( f (lx), f (ly)), K2( f (wx), f (wy)) (12)
The CC form shows how Full Additive factorizes into a more structural and a more lexical part: Each element of the sum is the Frobenius product between the product of two matrices representing syntactic labels and the tensor product between
two vectors representing the corresponding words. For subject–verb–object phrases ((S w1) (V w2) (O w3)) we have
K( f (((S goats) (V eat) (O grass))), f (((S cows) (V drink) (O water)))) = = 〈S goats + V eat + O grass, S cows + V drink + O water〉 = = 〈STS, goats cowsT〉F + 〈STV, goats drink
T〉F + 〈STO, goats waterT〉F +〈VTS, eat cowsT〉F + 〈VTV, eat drink
T〉F + 〈VTO, eat waterT〉F +〈OTS, grass cowsT〉F + 〈OTV, grass drink
T〉F + 〈OTO, grass waterT〉F = ∑ (lx wx)∈{(S goats),(V eat),(O grass)}
(ly wy )∈{(S cows),(V drink),(O water)}
g(K1( f (lx), f (ly)), K2( f (wx), f (wy))
(13)
Again, we observe the factoring into products of syntactic and lexical representations. By looking at Full Additive in the CC form, we observe that when XTY ≈ I for all matrix pairs, it degenerates to Additive. Interestingly, Full Additive can also approximate a semantic convolution kernel (Mehdad, Moschitti, and Zanzotto 2010), which combines dot products of elements in the same slot. In the adjective–noun case, we obtain this approximation by choosing two nearly orthonormal matrices A and N such that AAT = NNT ≈ I and ANT ≈ 0 and applying Equation (2): 〈A tall + N boy, A red + N cat〉 ≈ 〈tall, red〉+ 〈boy, cat〉.
This approximation is valid also for three-word phrases. When the matrices S, V, and O are such that XXT ≈ I with X one of the three matrices and YXT ≈ 0 with X and Y two different matrices, Full Additive approximates a semantic convolution kernel comparing two sentences by summing the dot products of the words in the same role, that is,
〈S goats + V eat + O grass, S cows + V drink + O water〉 ≈ ≈ 〈goats, cows〉+ 〈eat, drink〉+ 〈grass, water〉 (14)
Results can again be easily extended to longer and different-length phrases.
Lexical Function Model. We distinguish composition with one- vs. two argument predicates. We illustrate the first through adjective–noun composition, where the adjective acts as the predicate, and the second with transitive verb constructions. Although we use the relevant syntactic labels, the formulas generalize to any construction with the same argument count. For adjective–noun phrases, the input is a sequence of (label, word) pairs (x = ((A, w1), (N, w2))) and the relevant decomposition set again includes only the single trivial decomposition into all the subparts: R(x) = {((A, w1), (N, w2))}. The method itself is recursively defined as
f (x) = ⎧⎪⎨ ⎪⎩ f ((A, w1))f ((N, w2)) if x /∈ TX = ((A, w1), (N, w2)) W1 if x ∈ Tx = (A, w1) w2 if x ∈ Tx = (N, w2)
(15)
Here, K and g are, respectively, the dot and Frobenius product, K1( f (x), f (y)) = f (x)Tf (y), and K2( f (x), f (y)) = f (x)f (y)T. Using Equation (3), we have then
K( f (tall boy)), f (red cat)) = 〈TALL boy, RED cat〉 = = 〈TALLTRED, boy catT〉F = g(K1( f (tall), f (red)), K2( f (boy), f (cat)))
(16)
The role of predicate and argument words in the final dot product is clearly separated, showing again the structure-sensitive nature of the decomposition of the comparison operations. In the two-place predicate case, again, the input is a set of (label, word) tuples, and the relevant decomposition set only includes the single trivial decomposition into all subparts. The CDSM f is defined as
f (x) = ⎧⎪⎨ ⎪⎩ f ((S w1)) ⊗ f ((V w2)) ⊗ f ((O w3)) if x /∈ TX = ((S w1) (V w2) (O w3)) w if x ∈ TX = (l w) and l is S or O W if x ∈ TX = (V w) (17)
K is the dot product and g(x, y, z) = 〈x, y ⊗ z〉F, K1( f (x), f (y)) = ∑
j( f (x) ⊗ f (y))j— that is, the tensor contraction2 along the second index of the tensor product between f (x) and f (y)—and K2( f (x), f (y)) = K3( f (x), f (y)) = f (x) ⊗ f (y) are tensor products. The dot product of goats EAT grass and cows DRINK water is (by using Equation (4))
K( f (((S goats) (V eat) (O grass))), f (((S cows) (V drink) (O water)))) = = 〈 goats EAT grass, cows DRINK water〉 = = 〈∑j(EAT ⊗ DRINK)j, goats ⊗ grass ⊗ cows ⊗ water〉F = = g(K1( f ((V eat)), f ((V drink))),
K2( f ((S goats)), f ((S cows))) ⊗ K3( f ((O grass)), f ((O water))))
(18)
We rewrote the equation as a Frobenius product between two fourth-order tensors. The first combines the two third-order tensors of the verbs ∑ j(EAT ⊗ DRINK)j and the second combines the vectors representing the arguments of the verb, that is: goats ⊗ grass ⊗ cows ⊗ water. In this case as well we can separate the role of predicate and argument types in the comparison computation. Extension of the Lexical Function to structured objects of different lengths is treated by using the identity element for missing parts. As an example, we show here the comparison between tall boy and cat where the identity element is the identity matrix I:
K( f (tall boy)), f (cat)) = 〈TALL boy, cat〉 = 〈TALL boy, I cat〉 = = 〈TALLTI, boy catT〉F = g(K1( f (tall), f ( )), K2( f (boy), f (cat)))
(19)
Dual Space Model. We have until now applied the CC to linear CDSMs with the dot product as the final comparison operator (what we called K). The CC also holds for the effective Dual Space model of Turney (2012), which assumes that each word has two distributional representations, wd in “domain” space and wf in “function” space. The similarity of two phrases is directly computed as the geometric average of the separate similarities between the first and second words in both spaces. Even though
2 Grefenstette et al. (2013) first framed the Lexical Function in terms of tensor contraction.
there is no explicit composition step, it is still possible to put the model in CC form. Take x = (x1, x2) and its trivial decomposition. Define, for a word w with vector representations wd and wf : f (w) = wd wf T. Define also K1( f (x1), f (y1)) = √〈 f (x1), f (y1)〉F,
K2( f (x2), f (y2)) = √〈 f (x2), f (y2)〉F and g(a, b) to be √ab. Then
g(K1( f (x1), f (y1)), K2( f (x2), f (y2))) =
= √√ 〈 xd1 xf 1T, yd1 yf 1T〉F · √ 〈 xd2 xf 2T, yd2 yf 2T〉F =
= 4 √ 〈 xd1, yd1〉 · 〈 xf 1, yf 1〉 · 〈 xd2, yd2〉 · 〈 xf 2, yf 2〉 =
= geo(sim(xd1, yd1), sim(xd2, yd2), sim(xf 1, yf 1), sim(xf 2, yf 2))
(20)
A Neural-network-like Model. Consider the phrase (w1, w2, . . . , wn) and the model defined by f (x) = σ( w1 + w2 + . . .+ wn), where σ(·) is a component-wise logistic function. Here we have a single trivial decomposition that includes all the subparts, and γ(x1, . . . , xn) is defined as σ(x1 + . . .+ xn). To see that for this model the CC cannot hold, consider two two-word phrases (a b) and (c d)
K( f ((a, b)), f ((c, d))) = 〈 f ((a, b)), f ((c, d))〉 = ∑i [ σ( a + b) ] i · [ σ( c + d) ] i
= ∑
i
( 1 + e−ai−bi + e−ci−di + e−ai−bi−ci−di )−1 (21)
We need to rewrite this as
g(K1( a, c), K2( b, d)) (22)
But there is no possible choice of g, K1, and K2 that allows Equation (21) to be written as Equation (22). This example can be regarded as a simplified version of the neuralnetwork model of Socher et al. (2011). The fact that the CC does not apply to it suggests that it will not apply to other models in this family. 5 conclusion :The Convolution Conjecture offers a general way to rewrite the phrase similarity computations of CDSMs by highlighting the role played by the subparts of a composed representation. This perspective allows for a better understanding of the exact operations that a composition model applies to its input. The Convolution Conjecture also suggests a strong connection between CDSMs and semantic convolution kernels. This link suggests that insights from the CDSM literature could be directly integrated in the development of convolution kernels, with all the benefits offered by this wellunderstood general machine-learning framework. Distributional semantics has been extended to phrases and sentences by means of composition operations. We look at how these operations affect similarity measurements, showing that similarity equations of an important class of composition methods can be decomposed into operations performed on the subparts of the input phrases. This establishes a strong link between these models and convolution kernels. [{""affiliations"": [], ""name"": ""Fabio Massimo Zanzotto""}, {""affiliations"": [], ""name"": ""Lorenzo Ferrone""}, {""affiliations"": [], ""name"": ""Marco Baroni""}] SP:1111fb9daa320f08257d60408e9f6a49aad5c0c6 [{""authors"": [""Clark"", ""Stephen.""], ""title"": ""Vector space models of lexical meaning"", ""venue"": ""Shalom Lappin and Chris Fox, editors, Handbook of Contemporary Semantics, 2nd ed. Blackwell, Malden, MA. In press."", ""year"": 2015}, {""authors"": [""Coecke"", ""Bob"", ""Mehrnoosh Sadrzadeh"", ""Stephen Clark.""], ""title"": ""Mathematical foundations for a compositional distributional model of meaning"", ""venue"": ""Linguistic Analysis, 36:345\u2013384."", ""year"": 2010}, {""authors"": [""Ganesalingam"", ""Mohan"", ""Aur\u00e9lie Herbelot.""], ""title"": ""Composing distributions: Mathematical structures and their linguistic interpretation"", ""venue"": ""Working paper, Computer Laboratory, University"", ""year"": 2013}, {""authors"": [""Grefenstette"", ""Edward"", ""Georgiana Dinu"", ""Yao-Zhong Zhang"", ""Mehrnoosh Sadrzadeh"", ""Marco Baroni.""], ""title"": ""Multi-step regression learning for compositional distributional semantics"", ""venue"": ""Proceedings"", ""year"": 2013}, {""authors"": [""Guevara"", ""Emiliano.""], ""title"": ""A regression model of adjective-noun compositionality in distributional semantics"", ""venue"": ""Proceedings of GEMS, pages 33\u201337, Uppsala."", ""year"": 2010}, {""authors"": [""Haussler"", ""David.""], ""title"": ""Convolution kernels on discrete structures"", ""venue"": ""Technical report USCS-CL-99-10, University of California at Santa Cruz."", ""year"": 1999}, {""authors"": [""Mitchell"", ""Jeff"", ""Mirella Lapata.""], ""title"": ""Vector-based models of semantic composition"", ""venue"": ""Proceedings of ACL, pages 236\u2013244, Columbus, OH."", ""year"": 2008}, {""authors"": [""Socher"", ""Richard"", ""Eric Huang"", ""Jeffrey Pennin"", ""Andrew Ng"", ""Christopher Manning.""], ""title"": ""Dynamic pooling and unfolding recursive autoencoders for paraphrase detection"", ""venue"": ""Proceedings of NIPS,"", ""year"": 2011}, {""authors"": [""Turney"", ""Peter.""], ""title"": ""Domain and function: A dual-space model of semantic relations and compositions"", ""venue"": ""Journal of Artificial Intelligence Research, 44:533\u2013585."", ""year"": 2012}, {""authors"": [""Turney"", ""Peter"", ""Patrick Pantel.""], ""title"": ""From frequency to meaning: Vector space models of semantics"", ""venue"": ""Journal of Artificial Intelligence Research, 37:141\u2013188."", ""year"": 2010}, {""authors"": [""Zanzotto"", ""Fabio Massimo"", ""Ioannis Korkontzelos"", ""Francesca Falucchi"", ""Suresh Manandhar.""], ""title"": ""Estimating linear models for compositional distributional semantics"", ""venue"": ""Proceedings of COLING,"", ""year"": 2010}] acknowledgments  :We thank the reviewers for helpful comments. Marco Baroni acknowledges ERC 2011 Starting Independent Research Grant n. 283554 (COMPOSES).
References Clark, Stephen. 2015. Vector space models
of lexical meaning. In Shalom Lappin and Chris Fox, editors, Handbook of Contemporary Semantics, 2nd ed. Blackwell, Malden, MA. In press.
Coecke, Bob, Mehrnoosh Sadrzadeh, and Stephen Clark. 2010. Mathematical foundations for a compositional distributional model of meaning. Linguistic Analysis, 36:345–384. Ganesalingam, Mohan and Aurélie Herbelot. 2013. Composing distributions: Mathematical structures and their linguistic interpretation. Working paper, Computer Laboratory, University of Cambridge. Available at www.cl.cam.ac.uk/∼ah433/. Grefenstette, Edward, Georgiana Dinu, Yao-Zhong Zhang, Mehrnoosh Sadrzadeh, and Marco Baroni. 2013. Multi-step regression learning for compositional distributional semantics. Proceedings of IWCS, pages 131–142, Potsdam. Guevara, Emiliano. 2010. A regression model of adjective-noun compositionality in distributional semantics. In Proceedings of GEMS, pages 33–37, Uppsala. Haussler, David. 1999. Convolution kernels on discrete structures. Technical report USCS-CL-99-10, University of California at Santa Cruz. Mehdad, Yashar, Alessandro Moschitti, and Fabio Massimo Zanzotto. 2010.
Syntactic/semantic structures for textual entailment recognition. In Proceedings of NAACL, pages 1,020–1,028, Los Angeles, CA.
Mitchell, Jeff and Mirella Lapata. 2008. Vector-based models of semantic composition. In Proceedings of ACL, pages 236–244, Columbus, OH. Socher, Richard, Eric Huang, Jeffrey Pennin, Andrew Ng, and Christopher Manning. 2011. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In Proceedings of NIPS, pages 801–809, Granada. Turney, Peter. 2012. Domain and function: A dual-space model of semantic relations and compositions. Journal of Artificial Intelligence Research, 44:533–585. Turney, Peter and Patrick Pantel. 2010. From frequency to meaning: Vector space models of semantics. Journal of Artificial Intelligence Research, 37:141–188. Zanzotto, Fabio Massimo, Ioannis Korkontzelos, Francesca Falucchi, and Suresh Manandhar. 2010. Estimating linear models for compositional distributional semantics. In Proceedings of COLING, pages 1,263–1,271, Beijing. when the whole is not greater : than the combination of its parts: : a “decompositional” look at : compositional distributional semantics :Fabio Massimo Zanzotto∗ University of Rome “Tor Vergata”
Lorenzo Ferrone University of Rome “Tor Vergata”
Marco Baroni University of Trento
Distributional semantics has been extended to phrases and sentences by means of composition operations. We look at how these operations affect similarity measurements, showing that similarity equations of an important class of composition methods can be decomposed into operations performed on the subparts of the input phrases. This establishes a strong link between these models and convolution kernels.",
,,,,,,,,"The phenomenal success of machine learning in engineering natural language applications has led to a curious situation: Natural language processing practitioners who were trained in the last 15 to 20 years may have established a quite successful career in this area with only a haphazard knowledge of the science of natural languages. The premise of the new volume by Emily M. Bender is that greater awareness of linguistics will enable continued technical progress, particularly as language applications are required to perform more intelligent processing in more languages.","[{""affiliations"": [], ""name"": ""Emily M. Bender""}]",SP:99bddb5469a564b2f6ddaf267bc7dc90601cd3a2,"[{""authors"": [""Chomsky"", ""Noam"", ""Morris Halle.""], ""title"": ""The Sound Pattern of English"", ""venue"": ""Harper & Row."", ""year"": 1968}, {""authors"": [""Prince"", ""Alan"", ""Paul Smolensky.""], ""title"": ""Optimality Theory: Constraint Interaction in Generative Grammar"", ""venue"": ""Wiley\u2013Blackwell."", ""year"": 2004}, {""authors"": [""Stabler"", ""Edward P.""], ""title"": ""Derivational minimalism"", ""venue"": ""Christian Retor\u00e9, editor,"", ""year"": 1997}, {""authors"": [""Steedman"", ""Mark""], ""title"": ""Logical Aspects of Computational Linguistics, volume 1328 of LNCS"", ""year"": 2000}, {""authors"": [""Wintner"", ""Shuly""], ""title"": ""Last words: What science underlies natural language engineering"", ""venue"": ""Computational Linguistics,"", ""year"": 2009}]",,,,,,,,,,,,,,,,,,,"reviewed by chris dyer carnegie mellon university :The phenomenal success of machine learning in engineering natural language applications has led to a curious situation: Natural language processing practitioners who were trained in the last 15 to 20 years may have established a quite successful career in this area with only a haphazard knowledge of the science of natural languages. The premise of the new volume by Emily M. Bender is that greater awareness of linguistics will enable continued technical progress, particularly as language applications are required to perform more intelligent processing in more languages.
© 2015 Association for Computational Linguistics
This book is not beholden to any particular theoretical program. Rather, it is a survey of the morphological and syntactic means by which different languages express meaning, anchored by clear and effective examples from typologically diverse languages. Eschewing theorizing to stay close to data permits a remarkably wide range of linguistic phenomena to be covered, and it is this that is the book’s greatest strength. However, in a few places, a seemingly arbitrary theoretical perspective is assumed rather more tacitly than one might hope, with few hints as to alternative analyses (e.g., see the following remarks about parts of speech in Chapter 6). Furthermore, a bit more theoretical scaffolding could have made the presentation more succinct in places (e.g., Chapter 7’s excellent discussion of heads, arguments, and adjuncts could have been more precise with a basic logical calculus). Finally, although theoretical squabbles can be off-putting to outsiders, theoretical diversity can have practical benefits, particularly in a field as omnivorous as NLP. For example, while theorists might disagree about whether morphophonology is best modeled with systems of rewrite rules (e.g., SPE) or constraint satisfaction (e.g., Optimality Theory) (Chomsky and Halle 1968; Prince and Smolensky 2004), each suggests a distinct computational instantiation with different challenges and opportunities. For such reasons, more discussion of theory would not have been unwelcome. This slight objection aside, the book is an excellent introduction to the diversity of linguistic representations that NLP must eventually contend with.
The book is organized into 10 chapters, in roughly two parts (the first part, morphology; the second, syntax), spread over 100 numbered topics.
Chapter 1 gives an overview of the scope of the book, distinguishing morphology and syntax from bag-of-words models. It lays out the premise that knowledge of linguistic structure can guide engineers in profitable directions by facilitating error
doi:10.1162/COLI r 00212
Computational Linguistics Volume 41, Number 1
analysis and feature engineering. The notion of bounded variation is introduced: the idea that while languages exhibit diversity in how they pair sound and meaning, this variation is subject to limits, and that different languages can have similarities due to areal, genetic, and typological relatedness. A brief survey of the genetic taxonomy of the world’s languages is given and the number of speakers they have—as well as the striking difference in distributions of the languages in the NLP literature.
Chapters 2 and 3 introduce morphology and morphophonology, focusing on the internal structure of words and how they are realized in text and speech. Simple English examples motivate the discussion, but more exotic nonconcatenative processes in Semitic languages and infixation examples from Lakhota emphasize phenomena that may be unfamiliar to those with experience only with Indo-European languages. The conventional tripartite distinction of roots and derivational and inflectional affixes is presented to organize the kinds of meaning/function changes characteristic of morphological processes, although compounding and cliticization—which fit less neatly into this taxonomy—are also discussed. Because syntax and semantics were only briefly mentioned in the Introduction, the extensive forward references to the related material in the later chapters were quite helpful for clarifying terms, making the e-book version particularly convenient.
Chapter 4 discusses morphosyntax, reviewing the diverse grammatical functions that different languages encode with morphology. Phenomena covered include functions applying to the verbal domain, including tense, mood, and aspect, negation, evidentiality; the nominal domain, including person, number, and gender, case, definiteness, and possession; and various common agreement processes.
Having established how words are constructed from morphemes, Chapters 5–9 focus on how syntax is used to combine words to form an unbounded number of sentences whose meaning is determined compositionally. Chapter 5 introduces the distinction between grammaticality of sentences and how syntactic structure determines their meanings, and Chapter 6 introduces parts of speech as clusters of distributional regularities of words and phrases in grammatical sentences. The fact that discussion of grammaticality proceeds almost exclusively in terms of POS—a familiar construct to anyone working in NLP, but one that looks quite different in many theories of syntax (Steedman 2000; Stabler 1997)—is a shortcoming.
Chapter 7, perhaps the strongest in the book, discusses syntax in terms of headed phrases that relate to each other either as arguments (which semantically complete the meaning of a predicate) or adjunction (which introduces additional predicates). Diagnostics for distinguishing heads and dependents as well as arguments and adjuncts are given, together with clear examples of their application, and common mistakes (e.g., using optionality as a test of argumenthood or assuming that only verbs can select arguments) are covered. A particularly useful part of this chapter is a discussion of lexical resources (FrameNet, ProbBank) and how they relate to the concepts being discussed.
Chapter 8 discusses argument types and grammatical functions, reviewing not entirely successful attempts to create universal inventories of thematic roles, ultimately demonstrating that syntactic roles are less idiosyncratic (at least within single languages), and capture many generalizations useful for semantic analysis. A discussion of cross-linguistic properties of subjects and the distinction between core and oblique arguments follows. Three important sections discuss the subtle and often confusing distinctions between syntactic and semantic arguments with effective examples. Although most of this chapter focuses on English examples, various morphological strategies for marking grammatical functions is discussed.
154
Book Reviews
Chapter 9 concludes the syntax portion of the book, focusing on the processes that can introduce divergences between syntactic and semantic relationships. Because such divergences underlie many constructions with considerable value in NLP (e.g., wh-questions in English) and directly challenge the simplifying assumption of transparency between syntax and semantics, it is fortunate that this section goes into considerable detail, covering phenomena including passivization, dative shift, expletives, raising, control, and various kinds of long distance movement, as well as a good discussion of phenomena found in other languages, such as causative morphology and discontinuous constituents.
Chapter 10 provides a brief appendix summarizing various large-scale computational resources (morphological analyzers, parsers, typological atlases) that encode linguistic knowledge.
This book serves as a useful introduction to linguistic phenomena that will help NLP researchers orient themselves with respect to phenomena they will encounter as their applications push into new languages and strive for deeper automated understanding of language. The tension between the science of linguistics and natural language engineering and the resulting missed opportunities has been remarked upon in these pages recently (Wintner 2009), and we should applaud this successful effort to find common ground.",,,,,,,,,,,,,,,,linguistic fundamentals for natural language processing: :,"100 essentials from morphology and syntax :Emily M. Bender (University of Washington)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 20), 2013, xvii+166 pp; paperbound, ISBN 978-1-62705-011-1, $40.00; e-book, ISBN 978-1-62705-012-8, $30.00 or by subscription",,,,,,,,,,,,,,"The phenomenal success of machine learning in engineering natural language applications has led to a curious situation: Natural language processing practitioners who were trained in the last 15 to 20 years may have established a quite successful career in this area with only a haphazard knowledge of the science of natural languages. The premise of the new volume by Emily M. Bender is that greater awareness of linguistics will enable continued technical progress, particularly as language applications are required to perform more intelligent processing in more languages. [{""affiliations"": [], ""name"": ""Emily M. Bender""}] SP:99bddb5469a564b2f6ddaf267bc7dc90601cd3a2 [{""authors"": [""Chomsky"", ""Noam"", ""Morris Halle.""], ""title"": ""The Sound Pattern of English"", ""venue"": ""Harper & Row."", ""year"": 1968}, {""authors"": [""Prince"", ""Alan"", ""Paul Smolensky.""], ""title"": ""Optimality Theory: Constraint Interaction in Generative Grammar"", ""venue"": ""Wiley\u2013Blackwell."", ""year"": 2004}, {""authors"": [""Stabler"", ""Edward P.""], ""title"": ""Derivational minimalism"", ""venue"": ""Christian Retor\u00e9, editor,"", ""year"": 1997}, {""authors"": [""Steedman"", ""Mark""], ""title"": ""Logical Aspects of Computational Linguistics, volume 1328 of LNCS"", ""year"": 2000}, {""authors"": [""Wintner"", ""Shuly""], ""title"": ""Last words: What science underlies natural language engineering"", ""venue"": ""Computational Linguistics,"", ""year"": 2009}] reviewed by chris dyer carnegie mellon university :The phenomenal success of machine learning in engineering natural language applications has led to a curious situation: Natural language processing practitioners who were trained in the last 15 to 20 years may have established a quite successful career in this area with only a haphazard knowledge of the science of natural languages. The premise of the new volume by Emily M. Bender is that greater awareness of linguistics will enable continued technical progress, particularly as language applications are required to perform more intelligent processing in more languages.
© 2015 Association for Computational Linguistics
This book is not beholden to any particular theoretical program. Rather, it is a survey of the morphological and syntactic means by which different languages express meaning, anchored by clear and effective examples from typologically diverse languages. Eschewing theorizing to stay close to data permits a remarkably wide range of linguistic phenomena to be covered, and it is this that is the book’s greatest strength. However, in a few places, a seemingly arbitrary theoretical perspective is assumed rather more tacitly than one might hope, with few hints as to alternative analyses (e.g., see the following remarks about parts of speech in Chapter 6). Furthermore, a bit more theoretical scaffolding could have made the presentation more succinct in places (e.g., Chapter 7’s excellent discussion of heads, arguments, and adjuncts could have been more precise with a basic logical calculus). Finally, although theoretical squabbles can be off-putting to outsiders, theoretical diversity can have practical benefits, particularly in a field as omnivorous as NLP. For example, while theorists might disagree about whether morphophonology is best modeled with systems of rewrite rules (e.g., SPE) or constraint satisfaction (e.g., Optimality Theory) (Chomsky and Halle 1968; Prince and Smolensky 2004), each suggests a distinct computational instantiation with different challenges and opportunities. For such reasons, more discussion of theory would not have been unwelcome. This slight objection aside, the book is an excellent introduction to the diversity of linguistic representations that NLP must eventually contend with.
The book is organized into 10 chapters, in roughly two parts (the first part, morphology; the second, syntax), spread over 100 numbered topics.
Chapter 1 gives an overview of the scope of the book, distinguishing morphology and syntax from bag-of-words models. It lays out the premise that knowledge of linguistic structure can guide engineers in profitable directions by facilitating error
doi:10.1162/COLI r 00212
Computational Linguistics Volume 41, Number 1
analysis and feature engineering. The notion of bounded variation is introduced: the idea that while languages exhibit diversity in how they pair sound and meaning, this variation is subject to limits, and that different languages can have similarities due to areal, genetic, and typological relatedness. A brief survey of the genetic taxonomy of the world’s languages is given and the number of speakers they have—as well as the striking difference in distributions of the languages in the NLP literature.
Chapters 2 and 3 introduce morphology and morphophonology, focusing on the internal structure of words and how they are realized in text and speech. Simple English examples motivate the discussion, but more exotic nonconcatenative processes in Semitic languages and infixation examples from Lakhota emphasize phenomena that may be unfamiliar to those with experience only with Indo-European languages. The conventional tripartite distinction of roots and derivational and inflectional affixes is presented to organize the kinds of meaning/function changes characteristic of morphological processes, although compounding and cliticization—which fit less neatly into this taxonomy—are also discussed. Because syntax and semantics were only briefly mentioned in the Introduction, the extensive forward references to the related material in the later chapters were quite helpful for clarifying terms, making the e-book version particularly convenient.
Chapter 4 discusses morphosyntax, reviewing the diverse grammatical functions that different languages encode with morphology. Phenomena covered include functions applying to the verbal domain, including tense, mood, and aspect, negation, evidentiality; the nominal domain, including person, number, and gender, case, definiteness, and possession; and various common agreement processes.
Having established how words are constructed from morphemes, Chapters 5–9 focus on how syntax is used to combine words to form an unbounded number of sentences whose meaning is determined compositionally. Chapter 5 introduces the distinction between grammaticality of sentences and how syntactic structure determines their meanings, and Chapter 6 introduces parts of speech as clusters of distributional regularities of words and phrases in grammatical sentences. The fact that discussion of grammaticality proceeds almost exclusively in terms of POS—a familiar construct to anyone working in NLP, but one that looks quite different in many theories of syntax (Steedman 2000; Stabler 1997)—is a shortcoming.
Chapter 7, perhaps the strongest in the book, discusses syntax in terms of headed phrases that relate to each other either as arguments (which semantically complete the meaning of a predicate) or adjunction (which introduces additional predicates). Diagnostics for distinguishing heads and dependents as well as arguments and adjuncts are given, together with clear examples of their application, and common mistakes (e.g., using optionality as a test of argumenthood or assuming that only verbs can select arguments) are covered. A particularly useful part of this chapter is a discussion of lexical resources (FrameNet, ProbBank) and how they relate to the concepts being discussed.
Chapter 8 discusses argument types and grammatical functions, reviewing not entirely successful attempts to create universal inventories of thematic roles, ultimately demonstrating that syntactic roles are less idiosyncratic (at least within single languages), and capture many generalizations useful for semantic analysis. A discussion of cross-linguistic properties of subjects and the distinction between core and oblique arguments follows. Three important sections discuss the subtle and often confusing distinctions between syntactic and semantic arguments with effective examples. Although most of this chapter focuses on English examples, various morphological strategies for marking grammatical functions is discussed.
154
Book Reviews
Chapter 9 concludes the syntax portion of the book, focusing on the processes that can introduce divergences between syntactic and semantic relationships. Because such divergences underlie many constructions with considerable value in NLP (e.g., wh-questions in English) and directly challenge the simplifying assumption of transparency between syntax and semantics, it is fortunate that this section goes into considerable detail, covering phenomena including passivization, dative shift, expletives, raising, control, and various kinds of long distance movement, as well as a good discussion of phenomena found in other languages, such as causative morphology and discontinuous constituents.
Chapter 10 provides a brief appendix summarizing various large-scale computational resources (morphological analyzers, parsers, typological atlases) that encode linguistic knowledge.
This book serves as a useful introduction to linguistic phenomena that will help NLP researchers orient themselves with respect to phenomena they will encounter as their applications push into new languages and strive for deeper automated understanding of language. The tension between the science of linguistics and natural language engineering and the resulting missed opportunities has been remarked upon in these pages recently (Wintner 2009), and we should applaud this successful effort to find common ground. linguistic fundamentals for natural language processing: : 100 essentials from morphology and syntax :Emily M. Bender (University of Washington)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 20), 2013, xvii+166 pp; paperbound, ISBN 978-1-62705-011-1, $40.00; e-book, ISBN 978-1-62705-012-8, $30.00 or by subscription",
"1 introduction :Since the late 1970s, several grammar formalisms have been proposed that extend the power of context-free grammars in restricted ways. The two most prominent members of this class of “mildly context-sensitive” formalisms (a term coined by Joshi 1985) are Tree-Adjoining Grammar (TAG; Joshi and Schabes 1997) and Combinatory Categorial Grammar (CCG; Steedman 2000; Steedman and Baldridge 2011). Both formalisms have been applied to a broad range of linguistic phenomena, and are being widely used in computational linguistics and natural language processing.
In a seminal paper, Vijay-Shanker and Weir (1994) showed that TAG, CCG, and two other mildly context-sensitive formalisms—Head Grammar (Pollard 1984) and Linear Indexed Grammar (Gazdar 1987)—all characterize the same class of string languages. However, when citing this result it is sometimes overlooked that the result applies to a version of CCG that is quite different from the versions that are in practical use today.
∗ Department of Computer and Information Science, Linköping University, 581 83 Linköping, Sweden. E-mail: marco.kuhlmann@liu.se.
∗∗ Department of Linguistics, Karl-Liebknecht-Str. 24–25, University of Potsdam, 14476 Potsdam, Germany. E-mail: koller@ling.uni-potsdam.de.
† Department of Information Engineering, University of Padua, via Gradenigo 6/A, 35131 Padova, Italy. E-mail: satta@dei.unipd.it.
Submission received: 4 December 2013; revised submission received: 26 July 2014; accepted for publication: 25 November 2014.
doi:10.1162/COLI a 00219
© 2015 Association for Computational Linguistics
The goal of this article is to contribute to a better understanding of the significance of this difference.
The difference between “classical” CCG as formalized by Vijay-Shanker and Weir (1994) and the modern perspective may be illustrated with the combinatory rule of backward-crossed composition. The general form of this rule looks as follows:
Backward-crossed composition, general form:
Y/Z X /Y ⇒ X/Z (<B×)
This rule is frequently used in the analysis of heavy NP shift, as in the sentence Kahn blocked skillfully a powerful shot by Rivaldo (example from Baldridge 2002). However, backward crossed composition cannot be universally active in English as this would cause the grammar to also accept strings such as *a powerful by Rivaldo shot, which is witnessed by the derivation in Figure 1. To solve this problem, in the CCG formalism of Vijay-Shanker and Weir (1994), one may restrict backward-crossed composition to instances where X and Y are both verbal categories—that is, functions into the category S of sentences (cf. Steedman 2000, Section 4.2.2). With this restriction the unwanted derivation in Figure 1 can be blocked, and a powerful shot by Rivaldo is still accepted as grammatical. Other syntactic phenomena require other grammar-specific restrictions, including the complete ban of certain combinatory rules (cf. Steedman 2000, Section 4.2.1).
Over the past 20 years, CCG has evolved to put more emphasis on supporting fully lexicalized grammars (Baldridge and Kruijff 2003; Steedman and Baldridge 2011), in which as much grammatical information as possible is pushed into the lexicon. This follows the tradition of other frameworks such as Lexicalized Tree-Adjoining Grammar (LTAG) and Head-Driven Phrase Structure Grammar (HPSG). Grammar-specific rule restrictions are not connected to individual lexicon entries, and are therefore avoided. Instead, recent versions of CCG have introduced a new, lexicalized control mechanism in the form of modalities or slash types. The basic idea here is that combinatory rules only apply if the slashes in their input categories have the right types. For instance, the modern version of backward-crossed composition (Steedman and Baldridge 2011) takes the following form, where × is a slash type that allows crossed but not harmonic composition:
Backward-crossed composition, modalized form:
Y/×Z X /×Y ⇒ X/×Z (<B×)
The derivation in Figure 1 can now be blocked by assigning the slash in the category of powerful a type that is incompatible with ×, and thus cannot feed into type-aware backward-crossed composition as an input category. This “modalization” of grammatical composition, which imports some of the central ideas from the type-logical tradition of categorial grammar (Moortgat 2011) into the CCG framework, is attractive because it isolates the control over the applicability of combinatory rules in the lexicon. Quoting Steedman and Baldridge (2011, p. 186):
Without typed slashes, language-specific restrictions or even bans on some combinatory rules are necessary in order to block certain ungrammatical word orders. With them, the combinatory rules are truly universal: the grammar of every language utilizes exactly the same set of rules, without modification, thereby leaving all cross-linguistic variation in the lexicon. As such, CCG is a fully lexicalized grammar formalism.
The stated goal is thus to express all linguistically relevant restrictions on the use of combinatory rules in terms of lexically specified, typed slashes. But to what extent can this goal actually be achieved? So far, there have been only partial answers—even when, as in the formalism of Vijay-Shanker and Weir (1994), we restrict ourselves to the rules of composition, but exclude other rules such as type-raising and substitution. Baldridge and Kruijff (2003) note that the machinery of their multi-modal formalism can be simulated by rule restrictions, which shows that this version of lexicalized CCG is at most as expressive as the classical formalism. At the other end of the spectrum, we have shown in previous work that a “pure” form of lexicalized CCG with neither rule restrictions nor slash types is strictly less expressive (Kuhlmann, Koller, and Satta 2010). The general question of whether “classical” CCG (rule restrictions) and “modern” CCG (slash types instead of rule restrictions) are weakly equivalent has remained open.
In this article, we answer this question. Our results are summarized in Figure 2. After setting the stage (Section 2), we first pinpoint the exact type of rule restrictions that make the classical CCG formalism weakly equivalent to TAG (Section 3). We do so by focusing on a class of grammars that we call prefix-closed. Unlike the “pure” grammars that we studied in earlier work, prefix-closed grammars do permit rule restrictions (under certain conditions that seem to be satisfied by linguistic grammars). We show that the generative power of prefix-closed CCG depends on its ability to express target restrictions, which are exactly the functions-into type of restrictions that are needed to block the derivation in Figure 1. We prove that the full class of prefix-closed CCGs is weakly equivalent to TAG, and the subclass that cannot express target restrictions is strictly less powerful. This result significantly sharpens the picture of the generative capacity of CCG.
In a further step (Section 4), we then prove that for at least one popular incarnation of modern, lexicalized CCG, slash types are strictly less expressive than rule
Weakly equivalent to TAG Strictly less powerful than TAG","2 background :In this section we provide the technical background of the article. We introduce the basic architecture of CCG, present the formalism of Vijay-Shanker and Weir (1994), and set the points of reference for our results about generative capacity. The two central components of CCG are a lexicon that associates words with categories, and rules that specify how categories can be combined. Taken together, these components give rise to derivations, such as the one shown in Figure 3.
Lexicon. A category is a syntactic type that identifies a constituent as either “complete” or “incomplete.” The categories of complete constituents are taken to be primitives; the categories of incomplete constituents are modeled as (curried) functions that specify the type and the direction of their arguments. Every category is projected from the lexicon, which is a finite collection of word–category pairs. To give examples (taken from Steedman 2012), an English intransitive verb such as walks has a category that identifies it as a function seeking a (subject) noun phrase to its left (indicated by a backward slash) and returning a complete sentence; and a transitive verb such as admires has a category that identifies it as a function seeking an (object) noun phrase to its right (indicated by a forward slash) and returning a constituent that acts as an intransitive verb. We denote lexical assignment using the “colon equals” operator:
walks := S /NP admires := (S /NP)/NP
Formally, the set C(A) of categories is built over a finite set A of atomic categories, the primitive syntactic types. It is the smallest set such that
2. if X, Y ∈ C(A) then X/Y ∈ C(A) and X /Y ∈ C(A).
We use the letters A, B, C to denote atomic categories, the letters X, Y, Z to denote arbitrary categories, and the symbol | to denote slashes (forward or backward). We treat slashes as left-associative operators and omit unnecessary parentheses. By this convention, every category X can be written in the form
X = A|mXm · · · |1X1
where m ≥ 0, A is an atomic category that we call the target of X, and the |iXi are slash–category pairs that we call the arguments of X. Intuitively, the target of X is the atomic category that is obtained after X has been applied to all of its arguments. We use the Greek letters α, β, γ to denote (potentially empty) sequences of arguments. The number m is called the arity of the category X.
Rules. Categories can combine under a number of rules; this gives rise to derivations such as the one shown in Figure 3. Each rule specifies an operation that assembles two input categories into an output category. The two most basic rules of CCG are the directed versions of function application:
X/Y Y ⇒ X (forward application) (>) Y X /Y ⇒ X (backward application) (<)
Formally, a rule is a syntactic object in which the letters X, Y, Z act as variables for categories. A rule instance is obtained by substituting concrete categories for all variables in the rule. For example, the derivation in Figure 3 contains the following instances of function application. We denote rule instances by using a triple arrow instead of the double arrow in our notation for rules.
(S /NP)/NP NP S /NP and NP S /NP S
Application rules give rise to derivations equivalent to those of context-free grammar. Indeed, versions of categorial grammar where application is the only mode of combination, such as AB-grammar (Ajdukiewicz 1935; Bar-Hillel, Gaifman, and Shamir 1960), can only generate context-free languages. CCG can be more powerful because it also includes other rules, derived from the combinators of combinatory logic (Curry, Feys, and Craig 1958). In this article, as in most of the formal work on CCG, we restrict our attention to the rules of (generalized) composition, which are based on the B combinator.1
The general form of composition rules is shown in Figure 4. In each rule, the two input categories are distinguished into one primary (shaded) and one secondary input category. The number n of outermost arguments of the secondary input category is called the degree of the rule.2 In particular, for n = 0 we obtain the rules of function
1 This means that we ignore other rules required for linguistic analysis, in particular type-raising (from the T combinator), substitution (from the S combinator), and coordination. 2 The literature on CCG assumes a bound on n; for English, Steedman (2000, p. 42) puts n ≤ 3. Adding rules of unbounded degree increases the generative capacity of the formalism (Weir and Joshi 1988).
application. In contexts where we refer to both application and composition, we use the latter term for composition rules with degree n > 0.
Derivation Trees. Derivation trees can now be schematically defined as in Figure 5. They contain two types of branchings: unary branchings correspond to lexicon entries; binary branchings correspond to rule instances. The yield of a derivation tree is the left-to-right concatenation of its leaves. We now define the classical CCG formalism that was studied by Vijay-Shanker and Weir (1994) and originally introduced by Weir and Joshi (1988). As mentioned in Section 1, the central feature of this formalism is its ability to impose restrictions on the applicability of combinatory rules. Specifically, a restricted rule is a rule annotated with constraints that
(a) restrict the target of the primary input category; and/or
(b) restrict the secondary input category, either in parts or in its entirety.
Every grammar lists a finite number of restricted rules (where one and the same base rule may occur with several different restrictions). A valid rule instance is an instance that is compatible with at least one of the restricted rules.
Example 1 Linguistic grammars make frequent use of rule restrictions. To exclude the undesired derivation in Figure 1 we restricted backward crossed composition to instances where both the primary and the secondary input category are functions into the category of
sentences, S. Writing target for the function that returns the target of a category, the restricted rule can be written as
Y/Z X /Y ⇒ X/Z (backward crossed composition) (<B×) where target(X) = S and target(Y) = S
In the following definition we write ε for the empty string.
Definition 1 A combinatory categorial grammar in the sense of Vijay-Shanker and Weir (1994), or VW-CCG for short, is a structure
G = (Σ,A, :=, R, S)
where Σ is a finite vocabulary, A is a finite set of atomic categories, the lexicon := is a finite relation between the sets Σε = Σ ∪ {ε} and C(A), R is a finite set of (possibly) restricted rules, and S ∈ A is a distinguished atomic category.
A derivation tree of G consists of lexicon entries and valid instances of rules from R. The grammar G generates a string w if there exists a derivation tree whose yield is w and whose root node is labeled with the distinguished atomic category S. The language L(G) generated by G is the set of all generated strings.
Example 2 We specify a VW-CCG G1 = (Σ1,A1, :=1, R1, S1) that generates the language
L1 = {anbncn | n ≥ 1} .
Let Σ1 = {a, b, c}, A1 = {A, B, C, S}, S1 = S, and let :=1 consist of the entries
a :=1 A , b :=1 S/C , b :=1 S/C/B , b :=1 B/C , b :=1 B/C/B , c :=1 C /A
Finally, let R1 consist of the following rules:
1. Forward and backward application.
2. Forward and backward composition of degree 1.
3. Forward and backward composition of degree 2.
Each of these rules is restricted to instances where the target of the primary input category is S. A sample derivation of G1 is shown in Figure 6. By contrast, Figure 7 shows a derivation that is not valid in G1 because it uses application rules in which the target of the primary input category is B or C (rather than S).
Example 3 AB-grammar is a categorial grammar formalism in which forward and backward application are the only combinatory rules that are allowed. Furthermore, it does not support
rule restrictions.3 Every AB-grammar can therefore be written as a VW-CCG that allows all instances of application, but no other rules.
The following lemmas establish two fundamental properties of VW-CCG that we shall use at several places in this article. Both of them are minor variants of lemmas proved by Vijay-Shanker and Weir (1994) (Lemma 3.1 and Lemma 3.2) and were previously stated by Weir and Joshi (1988).
Lemma 1 The set of arguments that occur in the derivations of a VW-CCG is finite. In this article we are interested in the generative power of CCG, in particular in its relation to that of Tree-Adjoining Grammar. We therefore provide a compact introduction to TAG. For more details, we refer to Joshi and Schabes (1997).
Elementary Trees. Tree-Adjoining Grammar is a formalism for generating trees. These trees can be characterized as rooted, ordered trees in which internal nodes are labeled with nonterminal symbols—including a distinguished start symbol S—and leaf nodes are labeled with nonterminals, terminals, or the empty string. Every grammar specifies a finite set of such trees; these are called elementary trees. There are two types: initial trees and auxiliary trees. They differ in that auxiliary trees have a distinguished nonterminal-labeled leaf node, called foot node; this node is conventionally marked with an asterisk. An elementary tree whose root node is labeled with a nonterminal A is called an A-tree.
Substitution and Adjunction. New trees may be derived by combining other trees using two operations called substitution and adjunction. Substitution replaces some leaf node of a given tree with an initial tree (or a tree derived from an initial tree). Adjunction replaces some internal node u of a given tree with an auxiliary tree (or a tree derived from an auxiliary tree); the subtree with root u replaces the foot node of the auxiliary tree. All replacements are subject to the condition that the node being replaced and the root of the tree that replaces it are labeled with the same nonterminal.
Generated Languages. The tree language generated by a TAG is the set of all trees that can be derived from its initial S-trees. Derivations are considered complete only if they satisfy additional, node-specific constraints. In particular, substitution is obligatory at every node where it is possible, and adjunction may be specified as either obligatory (OA, Obligatory Adjunction) or forbidden (NA, Null Adjunction) at a given node. In derived trees corresponding to complete derivations, all leaf nodes are labeled with terminal symbols. The left-to-right concatenation of these symbols forms the yield of the tree, and the yields of all trees in the tree language form the string language generated by the TAG.
Example 5 Figure 8 shows a TAG that generates the language L1 = {anbncn | n ≥ 1} defined in Example 2. Derivations start with adjoining the auxiliary tree t2 at the root of the initial tree t1. New trees can be derived by repeatedly adjoining t2 at an S-node. Vijay-Shanker and Weir (1994) proved the following:
Theorem 1 VW-CCG and TAG are weakly equivalent.
The inclusion of the VW-CCG languages in the TAG languages follows from a chain of inclusions that connects VW-CCG and TAG via Linear Indexed Grammar (LIG; Gazdar 1987) and Head Grammar (HG; Pollard 1984). All of these inclusions were proved by Vijay-Shanker and Weir (1994). Here we sketch a proof of the inclusion of the TAG languages in the VW-CCG languages. Our proof closely follows that of Weir (1988, Section 5.2.2), whose construction we shall return to when establishing our own results.
Lemma 3 The TAG languages are included in the VW-CCG languages.","3 relevance of target restrictions in prefix-closed ccg :In this section we present the central technical results of this article. We study a class of VW-CCGs that we call prefix-closed and show that for this class, weak equivalence with TAG stands and falls with the ability to specify target restrictions. Rule restrictions are important tools in grammars for natural language; but not all of their potential uses have obvious linguistic motivations. For instance, one could write a grammar that permits all compositions with a functional category A/B as the secondary input category, but rules out application with the “shorter” category A. Such constraints do not seem to be used in linguistically motivated grammars; for example, none of the grammars developed by Steedman (2000) needs them. In prefix-closed grammars, this use of rule restrictions is explicitly barred.
Definition 2 A VW-CCG is prefix-closed if it satisfies the following implication:
if X/Y Y|nYn · · · |1Y1 X|nYn · · · |1Y1 is a valid rule instance then so is X/Y Y|nYn · · · |kYk X|nYn · · · |kYk for any k ≥ 1
and similarly for backward rules.
Note that prefix-closed grammars still allow some types of rule restrictions. The crucial property is that, if a certain combinatory rule applies at all, then it also applies to combinations where the secondary input category has already been applied to some (k ≤ n) or even all (k > n) of its arguments.
Example 6 We illustrate prefix-closedness using some examples:
1. Every AB-grammar (when seen as a VW-CCG) is trivially prefix-closed; in these grammars, n = 0.
2. The “pure” grammars that we considered in our earlier work (Kuhlmann, Koller, and Satta 2010) are trivially prefix-closed.
3. The grammar G1 from Example 2 is prefix-closed.
4. The grammars constructed in the proof of Lemma 3 are not prefix-closed; they do not allow the following instances of application, where the secondary input category is of the form Bc (rather than Ba):
Aa /Bc Bc Aa where A, B ∈ V Bc Aa /Bc Aa where A, B ∈ V
Example 7 The linguistic intuition underlying prefix-closed grammars is that if such a grammar allows us to delay the combination of a functor and its argument (via composition), then it also allows us to combine the functor and its argument immediately (via application). To illustrate this intuition, consider Figure 11, which shows two derivations related to the discussion of word order in Swiss German subordinate clauses (Shieber 1985):
. . . mer em Hans es huus hälfed aastriche . . . we Hansdat the houseacc helped paint
“. . . we helped Hans paint the house”
Derivation (5) (simplified from Steedman and Baldridge 2011, p. 201) starts by composing the tensed verb hälfed into the infinitive aastriche and then applies the resulting category to the accusative argument of the infinitive, es huus. Prefix-closedness implies that, if the combination of hälfed and aastriche is allowed when the latter is still waiting for es huus, then it must also be allowed if es huus has already been found.
Thus prefix-closedness predicts derivation (6), and along with it the alternative word order
. . . mer em Hans hälfed es huus aastriche . . . we Hansdat helped the houseacc paint
This word order is in fact grammatical (Shieber 1985, pp. 338–339). We now show that the restriction to prefix-closed grammars does not change the generative capacity of VW-CCG.
Theorem 2 Prefix-closed VW-CCG and TAG are weakly equivalent. In this section we shall see that the weak equivalence between prefix-closed VW-CCG and TAG depends on the ability to restrict the target of the primary input category in a combinatory rule. These are the restrictions that we referred to as constraints of type (a) in Section 2.2. We say that a grammar that does not make use of these constraints is without target restrictions. This property can be formally defined as follows.
Definition 3 A VW-CCG is without target restrictions if it satisfies the following implication:
if X/Y Yβ Xβ is a valid rule instance then so is X̄/Y Yβ X̄β for any category X̄ of the grammar
and similarly for backward rules.
Example 8
1. Every AB-grammar is without target restrictions; it allows forward and backward application for every primary input category.
2. The grammar G1 from Example (2) is not without target restrictions, because its rules are restricted to primary input categories with target S.
Target restrictions on the primary input category are useful in CCGs for natural languages; recall our discussion of backward-crossed composition in Section 1. As we shall see, target restrictions are also relevant from a formal point of view: If we require VW-CCGs to be without target restrictions, then we lose some of their weak generative capacity. This is the main technical result of this article. For its proof we need the following standard concept from formal language theory:
Definition 4 Two languages L and L′ are Parikh-equivalent if for every string w ∈ L there exists a permuted version w′ of w such that w′ ∈ L′, and vice versa.
Theorem 3 The languages generated by prefix-closed VW-CCG without target restrictions are properly included in the TAG languages. We shall now prove the central lemma that we used in the proof of Theorem 3.
Lemma 6 (Main Lemma for VW-CCG) For every language L that is generated by some prefix-closed VW-CCG without target restrictions, there is a sublanguage L′ ⊆ L such that
1. L′ and L are Parikh-equivalent, and
2. L′ is context-free.
Throughout this section, we let G be some arbitrary prefix-closed VW-CCG without target restrictions. The basic idea is to transform the derivations of G into a certain special form, and to prove that the transformed derivations yield a context-free language. The transformation is formalized by the rewriting system in Figure 12.4 To see how the rules of this system work, consider rule R1; the other rules are symmetric. Rule R1 rewrites an entire derivation into another derivation. It states that, whenever we have a situation where a category of the form X/Y is combined with a category of the form Yβ/Z by means of composition, and the resulting category is combined with a category Z by means of application, then we may just as well first combine Yβ/Z with Z, and then use the resulting category as a secondary input category together with X/Y.
4 Recall that we use the Greek letter β to denote a (possibly empty) sequence of arguments.
Note that R1 and R2 produce a new derivation for the original sentence, whereas R3 and R4 produce a derivation that yields a permutation of that sentence: The order of the substrings corresponding to the categories Z and X/Y (in the case of rule R3) or X /Y (in the case of rule R4) is reversed. In particular, R3 captures the relation between the two derivations of Swiss German word orders shown in Figure 11: Applying R3 to derivation (5) gives derivation (6). Importantly though, while the transformation may reorder the yield of a derivation, every transformed derivation still is a derivation of G.
Example 9 If we take the derivation in Figure 6 and exhaustively apply the rewriting rules from Figure 12, then the derivation that we obtain is the one in Figure 7. Note that although the latter derivation is not grammatical with respect to the grammar G1 from Example 2, it is grammatical with respect to the grammar G2 from the proof of Lemma 4, which is without target restrictions.
It is instructive to compare the rewriting rules in Figure 12 to the rules that establish the normal form of Eisner (1996). This normal form is used in practical CCG parsers to solve the problem of “spurious ambiguity,” where one and the same semantic interpretation (which in CCG takes the form of a lambda term) has multiple syntactic derivation trees. It is established by rewriting rules such as the following:
X/Y Yβ/Z Xβ/Z † Zγ
Xβγ
−→ X/Y Yβ/Z Zγ
Yβγ Xβγ
†† (5)
The rules in Figure 12 have much in common with the Eisner rules; yet there are two important differences. First, as already mentioned, our rules (in particular, rules R3 and R4) may reorder the yield of a derivation, whereas Eisner’s normal form
preserves yields. Second, our rules decrease the degrees of the involved composition operations, whereas Eisner’s rules may in fact increase them. To see this, note that the left-hand side of derivation (7) involves a composition of degree |β|+ 1 (†), whereas the right-hand side involves a composition of degree |β|+ |γ| (††). This means that rewriting will increase the degree in situations where |γ| > 1. In contrast, our rules only fire in the case where the combination with Z happens by means of an application, that is, if |γ| = 0. Under this condition, each rewrite step is guaranteed to decrease the degree of the composition. We will use this observation in the proof of Lemma 7.
3.4.1 Properties of the Transformation. The next two lemmas show that the rewriting system in Figure 12 implements a total function on the derivations of G.
Lemma 7 The rewriting system is terminating and confluent: Rewriting a derivation ends after a finite number of steps, and different rewriting orders all result in the same output. Lemma 9 The yields of the transformed derivations are a subset of and Parikh-equivalent to L(G). Theorem 3 pinpoints the exact mechanism that VW-CCG uses to achieve weak equivalence to TAG: At least for the class of prefix-closed grammars, TAG equivalence is achieved if and only if we allow target restrictions. Although target restrictions are frequently used in linguistically motivated grammars, it is important and perhaps surprising to realize that they are indeed necessary to achieve the full generative capacity of VW-CCG.
In the grammar formalisms folklore, the generative capacity of CCG is often attributed to generalized composition, and indeed we have seen (in Lemma 4) that even grammars without target restrictions can generate non-context-free languages such as L(G2). However, our results show that composition by itself is not enough to achieve weak equivalence with TAG: The yields of the transformed derivations from Section 3.4 form a context-free language despite the fact that these derivations may still contain compositions, including compositions of degree n > 2. In addition to composition, VWCCG also needs target restrictions to exert enough control on word order to block unwanted permutations. One way to think about this is that target restrictions can enforce alternations of composition and application (as in the derivation shown in
Figure 6), while transformed derivations are characterized by projection paths without such alternations (Lemma 10).
We can sharpen the picture even more by observing that the target restrictions that are crucial for the generative capacity of VW-CCG are not those on generalized composition, but those on function application. To see this we can note that the proof of Lemma 8 goes through also only if application rules such as (9) and (10) are without target restrictions. This means that we have the following qualification of Theorem 1.
Lemma 13 Prefix-closed VW-CCG is weakly equivalent to TAG only because it supports target restrictions on forward and backward application.
This finding is unexpected indeed—for instance, no grammar in Steedman (2000) uses target restrictions on the application rules.","4 generative capacity of multimodal ccg :After clarifying the mechanisms that “classical” CCG uses to achieve weak equivalence with TAG, we now turn our attention to “modern,” multimodal versions of CCG (Baldridge and Kruijff 2003; Steedman and Baldridge 2011). These versions emphasize the use of fully lexicalized grammars in which no rule restrictions are allowed, and instead equip slashes with types in order to control the use of the combinatory rules. Our central question is whether the use of slash types is sufficient to recover the expressiveness that we lose by giving up rule restrictions.
We need to fix a specific variant of multimodal CCG to study this question formally. Published works on multimodal CCG differ with respect to the specific inventories of slash types they assume. Some important details, such as a precise definition of generalized composition with slash types, are typically not discussed at all. In this article we define a variant of multimodal CCG which we call O-CCG. This formalism extends our definition of VW-CCG (Definition 1) with the slash inventory and the composition rules of the popular OpenCCG grammar development system (White 2013). Our technical result is that the main Lemma (Lemma 6) also holds for O-CCG. With this we can conclude that the answer to our question is negative: Slash types are not sufficient to replace rule restrictions; O-CCG is strictly less powerful than TAG. Although this is primarily a theoretical result, at the end of this section we also discuss its implications for practical grammar development. We define O-CCG as a formalism that extends VW-CCG with the slash types of OpenCCG, but abandons rule restrictions. Note that OpenCCG has a number of additional features that affect the generative capacity; we discuss these in Section 4.4.
Slash Types. Like other incarnations of multimodal CCG, O-CCG uses an enriched notion of categories where every slash has a type. There are eight such types:5
core types: ∗ × · left types: × right types: ×
5 The type system of OpenCCG is an extension of the system used by Baldridge (2002).
The basic idea behind these types is as follows. Slashes with type ∗ can only be used to instantiate application rules. Type also licenses harmonic composition rules, and type × also licenses crossed composition rules. Type · is the least restrictive type and can be used to instantiate all rules. The remaining types refine the system by incorporating a dimension of directionality. The exact type–rule compatibilities are specified in Figure 14.
Inertness. O-CCG is distinguished from other versions of multimodal CCG, such as that of Baldridge and Kruijff (2003), in that every slash not only has a type but also an inertness status. Inertness was introduced by Baldridge (2002, Section 8.2.2) as an implementation of the “antecedent government” (ANT) feature of Steedman (1996), which is used to control the word order in certain English relative clauses. It is a two-valued feature. Arguments whose slash type has inertness status + are called active; arguments whose slash type has inertness status − are called inert. Only active arguments can be eliminated by means of combinatory rules; however, an inert argument can still be consumed as part of a secondary input category. For example, the following instance of application is valid because the outermost slash of the primary input category has inertness status +:
X/+(Y/−Z) Y/−Z X
We use the notations /st and / s t to denote the forward and backward slashes with
slash type t and inertness status s.
Rules. All O-CCG grammars share a fixed set of combinatory rules, shown in Figure 15. Every grammar uses all rules, up to some grammar-specific bound on the degree of generalized composition. As mentioned earlier, a combinatory rule can only be instantiated if the slashes of the input categories have compatible types. Additionally, all composition rules require the slashes of the secondary input category to have a uniform direction. This is a somewhat peculiar feature of OpenCCG, and is in contrast to VW-CCG and other versions of CCG, which also allow composition rules with mixed directions.
Composition rules are classified into harmonic and crossed forms. This distinction is based on the direction of the slashes in the secondary input category. If these have the same direction as the outermost slash of the primary input category, then the rule is called harmonic; otherwise it is called crossed.6
6 In versions of CCG that allow rules with mixed slash directions, the distinction between harmonic and crossed is made based on the direction of the innermost slash of the secondary input category, |i.
When a rule is applied, in most cases the arguments of the secondary input category are simply copied into the output category, as in VW-CCG. The one exception happens for crossed composition rules if not all slash directions match the direction of their slash type (left or right). In this case, the arguments of the secondary input category become inert. Thus the inertness status of an argument may change over the course of a derivation—but only from active to inert, not back again.
Definition 5 A multimodal combinatory categorial grammar in the sense of OpenCCG, or O-CCG for short, is a structure
G = (Σ,A, :=, d, S)
where Σ is a finite vocabulary, A is a finite set of atomic categories, := is a finite relation between Σ and the set of (multimodal) categories over A, d ≥ 0 is the maximal degree of generalized composition, and S ∈ A is a distinguished atomic category.
We generalize the notions of rule instances, derivation trees, and generated language to categories over slashes with types and inertness statuses in the obvious way: Instead of two slashes, we now have one slash for every combination of a direction, type, and inertness status. Similarly, we generalize the concepts of a grammar being prefix-closed (Definition 2) and without target restrictions (Definition 3) to O-CCG. We now investigate the generative capacity of O-CCG. We start with the (unsurprising) observation that O-CCG can describe non-context-free languages.
Lemma 14 The languages generated by O-CCG properly include the context-free languages. The proof of Lemma 15 adapts the rewriting system from Figure 12. We simply let each rewriting step copy the type and inertness status of each slash from the left-hand side to the right-hand side of the rewriting rule. With this change, it is easy to verify that the proofs of Lemma 7 (termination and confluence), Lemma 10 (projection paths in transformed derivations are split), Lemma 11 (transformed derivations contain a finite number of categories), and Lemma 12 (transformed derivations yield a contextfree language) go through without problems. The proof of Lemma 8, however, is not straightforward, because of the dynamic nature of the inertness statuses. We therefore restate the lemma for O-CCG:
Lemma 16 The rewriting system transforms O-CCG derivations into O-CCG derivations. In this section we have shown that the languages generated by O-CCG are properly included in the languages generated by TAG, and equivalently, in the languages generated by VW-CCG. This means that the multimodal machinery of OpenCCG is not powerful enough to express the rule restrictions of VW-CCG in a fully lexicalized way. The result is easy to obtain for O-CCG without inertness, which is prefix-closed and without target restrictions; but it is remarkably robust in that it also applies to O-CCG with inertness, which is not prefix-closed. As we have already mentioned, the result carries over also to other multimodal versions of CCG, such as the formalism of Baldridge and Kruijff (2003).
Our result has implications for practical grammar development with OpenCCG. To illustrate this, recall Example 7, which showed that every VW-CCG without target restrictions for Swiss German that allows cross–serial word orders as in derivation (5) also permits alternative word orders, as in derivation (6). By Lemma 15, this remains true for O-CCG or weaker multimodal formalisms. This is not a problem in the case of Swiss German, where the alternative word orders are grammatical. However, there is at least one language, Dutch, where dependencies in subordinate clauses must cross. For this case, our result shows that the modalized composition rules of OpenCCG are not powerful enough to write adequate grammars. Consider the following classical example:
. . . ik Cecilia de paarden zag voeren . . . I Cecilia the horses saw feed
“. . . I saw Cecilia feed the horses”
The straightforward derivation of the cross–serial dependencies in this sentence (adapted from Steedman 2000, p. 141) is exemplified in Figure 16. It takes the same form as derivation (5) for Swiss German: The verbs and their NP arguments lie on a single, right-branching path projected from the tensed verb zag. This projection path is not split; specifically, it starts with a composition that produces a category which acts as the primary input category of an application. As a consequence, the derivation can be transformed (by our rewriting rule R3) in exactly the same way as instance (5) could be transformed into derivation (6). The crucial difference is that the yield of the transformed derivation, *ik Cecilia zag de paarden voeren, is not a grammatical clause of Dutch.
To address the problem of ungrammatical word orders in Dutch subordinate clauses, the VW-CCG grammar of Steedman (2000) and the multimodal CCG grammar of Baldridge (2002, Section 5.3.1) resort to combinatorial rules other than composition. In particular, they assume that all complement noun phrases undergo obligatory type-raising, and become primary input categories of application rules. This gives rise to derivations such as the one shown in Figure 17, which cannot be transformed using our rewriting rules because the result of the forward crossed composition >1
now is a secondary rather than a primary input category. As a consequence, this grammar is capable of enforcing the obligatory cross-serial dependencies of Dutch. However, it is important to note that it requires type-raising over arbitary categories with target S (observe the increasingly complex type-raised categories for the NPs). This kind of type-raising is allowed in many variants of CCG, including the full formalism underlying OpenCCG. VW-CCG and O-CCG, however, are limited to generalized composition, and can only support derivations like the one in Figure 17 if all the type-raised categories for the noun phrases are available in the lexicon. The unbounded type-raising required by the Steedman–Baldridge analysis of Dutch would translate into an infinite lexicon, and so this analysis is not possible in VW-CCG and O-CCG.
We conclude by discussing the impact of several other constructs of OpenCCG that we have not captured in O-CCG. First, OpenCCG allows us to use generalized composition rules of arbitrary degree; there is no upper bound d on the composition degree as in an O-CCG grammar. It is known that this extends the generative capacity of CCG beyond that of TAG (Weir 1988). Second, OpenCCG allows categories to be annotated with feature structures. This has no impact on the generative capacity, as the features must take values from finite domains and can therefore be compiled into the atomic categories of the grammar. Finally, OpenCCG includes the combinatory rules of substitution and coordination, as well as multiset slashes, another extension frequently used in linguistic grammars. We have deliberately left these constructs out of O-CCG to establish the most direct comparison to the literature on VW-CCG. It is conceivable that their inclusion could restore the weak equivalence to TAG, but a proof of this result would require a non-trivial extension of the work of Vijay-Shanker and Weir (1994). Regarding multiset slashes, it is also worth noting that these were introduced with the expressed goal of allowing more flexible word
order, whereas restoration of weak equivalence would require more controlled word order.","5 conclusion :In this article we have contributed two technical results to the literature on CCG. First, we have refined the weak equivalence result for CCG and TAG (Vijay-Shanker and Weir 1994) by showing that prefix-closed grammars are weakly equivalent to TAG only if target restrictions are allowed. Second, we have shown that O-CCG, the formal, composition-only core of OpenCCG, is not weakly equivalent to TAG. These results point to a tension in CCG between lexicalization and generative capacity: Lexicalized versions of the framework are less powerful than classical versions, which allow rule restrictions.
What conclusions one draws from these technical results depends on the perspective. One way to look at CCG is as a system for defining formal languages. Under this view, one is primarily interested in results on generative capacity and parsing complexity such as those obtained by Vijay-Shanker and Weir (1993, 1994). Here, our results clarify the precise mechanisms that make CCG weakly equivalent to TAG. Perhaps surprisingly, it is not the availability of generalized composition rules by itself that explains the generative power of CCG, but the ability to constrain the interaction between generalized composition and function application by means of target restrictions.
On the other hand, one may be interested in CCG primarily as a formalism for developing grammars for natural languages (Steedman 2000; Baldridge 2002; Steedman 2012). From this point of view, the suitability of CCG for the development of lexicalized grammars has been amply demonstrated. However, our technical results still serve as important reminders that extra care must be taken to avoid overgeneration when designing a grammar. In particular, it is worth double-checking that an OpenCCG grammar does not generate word orders that the grammar developer did not intend. Here the rewriting system that we presented in Figure 12 can serve as a useful tool: A grammar developer can take any derivation for a grammatical sentence, transform the derivation according to our rewriting rules, and check whether the transformed derivation still yields a grammatical sentence.
It remains an open question how the conflicting desires for generative capacity and lexicalization might be reconciled. A simple answer is to add some lexicalized method for enforcing target restrictions to CCG, specifically on the application rules. However, we are not aware that this idea has seen widespread use in the CCG literature, so it may not be called for empirically. Alternatively, one might modify the rules of O-CCG in such a way that they are no longer prefix-closed—for example, by introducing some new slash type. Finally, it is possible that the constructs of OpenCCG that we set aside in O-CCG (such as type-raising, substitution, and multiset slashes) might be sufficient to achieve the generative capacity of classical CCG and TAG. A detailed study of the expressive power of these constructs would make an interesting avenue for future research.",,,,"The weak equivalence of Combinatory Categorial Grammar (CCG) and Tree-Adjoining Grammar (TAG) is a central result of the literature on mildly context-sensitive grammar formalisms. However, the categorial formalism for which this equivalence has been established differs significantly from the versions of CCG that are in use today. In particular, it allows restriction of combinatory rules on a per grammar basis, whereas modern CCG assumes a universal set of rules, isolating all cross-linguistic variation in the lexicon. In this article we investigate the formal significance of this difference. Our main result is that lexicalized versions of the classical CCG formalism are strictly less powerful than TAG.","[{""affiliations"": [], ""name"": ""Marco Kuhlmann""}, {""affiliations"": [], ""name"": ""Alexander Koller""}, {""affiliations"": [], ""name"": ""Giorgio Satta""}]",SP:52fb58337ad0d4ec8da3bb73ae06c06bd09b356a,"[{""authors"": [""Ajdukiewicz"", ""Kazimierz.""], ""title"": ""Die syntaktische Konnexit\u00e4t"", ""venue"": ""Studia Philosophica, 1:1\u201327."", ""year"": 1935}, {""authors"": [""Baldridge"", ""Jason"", ""Geert-Jan M. Kruijff.""], ""title"": ""Multi-modal combinatory categorial grammar"", ""venue"": ""Tenth Conference of the European Chapter of the Association for Computational Linguistics (EACL),"", ""year"": 2003}, {""authors"": [""Bar-Hillel"", ""Yehoshua"", ""Haim Gaifman"", ""Eli Shamir.""], ""title"": ""On categorial and phrase structure grammars"", ""venue"": ""Bulletin of the Research Council of Israel, 9F(1):1\u201316. Reprinted in Yehoshua Bar-Hillel. Language and"", ""year"": 1960}, {""authors"": [""Curry"", ""Haskell B."", ""Robert Feys"", ""William Craig.""], ""title"": ""Combinatory Logic"", ""venue"": ""Volume 1. Studies in Logic and the Foundations of Mathematics. North-Holland."", ""year"": 1958}, {""authors"": [""Eisner"", ""Jason.""], ""title"": ""Efficient normal-form parsing for Combinatory Categorial Grammar"", ""venue"": ""Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics (ACL), pages 79\u201386,"", ""year"": 1996}, {""authors"": [""Gazdar"", ""Gerald.""], ""title"": ""Applicability of indexed grammars to natural language"", ""venue"": ""Uwe Reyle and Christian Rohrer, editors, Natural Language Parsing and Linguistic Theories. D. Reidel, pages 69\u201394."", ""year"": 1987}, {""authors"": [""Joshi"", ""Aravind K.""], ""title"": ""Tree Adjoining Grammars: How much context-sensitivity is required to provide reasonable structural descriptions? In David R"", ""venue"": ""Dowty, Lauri Karttunen, and Arnold M."", ""year"": 1985}, {""authors"": [""Joshi"", ""Aravind K."", ""Yves Schabes.""], ""title"": ""Tree-Adjoining Grammars"", ""venue"": ""Grzegorz Rozenberg and Arto Salomaa, editors, Handbook of Formal Languages, volume 3. Springer, pages 69\u2013123."", ""year"": 1997}, {""authors"": [""Kuhlmann"", ""Marco"", ""Alexander Koller"", ""Giorgio Satta.""], ""title"": ""The importance of rule restrictions in CCG"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (ACL),"", ""year"": 2010}, {""authors"": [""Moortgat"", ""Michael.""], ""title"": ""Categorial type logics"", ""venue"": ""Johan van Benthem and Alice ter"", ""year"": 2011}, {""authors"": [""Pollard"", ""Carl J.""], ""title"": ""Generalized Phrase Structure Grammars, Head Grammars, and Natural Language"", ""venue"": ""Ph.D. thesis, Stanford University."", ""year"": 1984}, {""authors"": [""Shieber"", ""Stuart M.""], ""title"": ""Evidence against the context-freeness of natural language"", ""venue"": ""Linguistics and Philosophy, 8(3):333\u2013343."", ""year"": 1985}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""Surface Structure and Interpretation, volume 30 of Linguistic Inquiry Monographs"", ""venue"": ""MIT Press."", ""year"": 1996}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""The Syntactic Process"", ""venue"": ""MIT Press."", ""year"": 2000}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""Taking Scope"", ""venue"": ""MIT Press."", ""year"": 2012}, {""authors"": [""Steedman"", ""Mark"", ""Jason Baldridge.""], ""title"": ""Combinatory Categorial Grammar"", ""venue"": ""Robert D. Borsley and Kersti B\u00f6rjars, editors, Non-Transformational Syntax: Formal and Explicit Models of Grammar."", ""year"": 2011}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""Parsing some constrained grammar formalisms"", ""venue"": ""Computational Linguistics, 19(4):591\u2013636."", ""year"": 1993}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""The equivalence of four extensions of context-free grammars"", ""venue"": ""Mathematical Systems Theory, 27(6):511\u2013546."", ""year"": 1994}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir"", ""Aravind K. Joshi.""], ""title"": ""Tree adjoining and head wrapping"", ""venue"": ""Proceedings of the Eleventh International Conference on Computational Linguistics (COLING),"", ""year"": 1986}, {""authors"": [""Weir"", ""David J.""], ""title"": ""Characterizing Mildly Context-Sensitive Grammar Formalisms"", ""venue"": ""Ph.D. thesis, University of Pennsylvania."", ""year"": 1988}, {""authors"": [""Weir"", ""David J."", ""Aravind K. Joshi.""], ""title"": ""Combinatory categorial grammars: Generative power and relationship to linear context-free rewriting systems"", ""venue"": ""Proceedings of the 26th Annual Meeting"", ""year"": 1988}, {""authors"": [""White"", ""Michael.""], ""title"": ""OpenCCG: The OpenNLP CCG Library"", ""venue"": ""http://openccg.sourceforge.net/ Accessed November 13, 2013. 219"", ""year"": 2013}]","acknowledgments  :We are grateful to Mark Steedman and Jason Baldridge for enlightening discussions of the material presented in this article, and to the four anonymous reviewers of the article for their detailed and constructive comments.
References Ajdukiewicz, Kazimierz. 1935. Die
syntaktische Konnexität. Studia Philosophica, 1:1–27.
Baldridge, Jason. 2002. Lexically Specified Derivational Control in Combinatory
Categorial Grammar. Ph.D. thesis, University of Edinburgh, Edinburgh, UK.
Baldridge, Jason and Geert-Jan M. Kruijff. 2003. Multi-modal combinatory categorial grammar. In Tenth Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 211–218, Budapest. Bar-Hillel, Yehoshua, Haim Gaifman, and Eli Shamir. 1960. On categorial and phrase structure grammars. Bulletin of the Research Council of Israel, 9F(1):1–16. Reprinted in Yehoshua Bar-Hillel. Language and Information: Selected Essays on Their Theory and Application, pages 99–115. Addison-Wesley, 1964. Curry, Haskell B., Robert Feys, and William Craig. 1958. Combinatory Logic. Volume 1. Studies in Logic and the Foundations of Mathematics. North-Holland. Eisner, Jason. 1996. Efficient normal-form parsing for Combinatory Categorial Grammar. In Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics (ACL), pages 79–86, Santa Cruz, CA. Gazdar, Gerald. 1987. Applicability of indexed grammars to natural language. In Uwe Reyle and Christian Rohrer, editors, Natural Language Parsing and Linguistic Theories. D. Reidel, pages 69–94. Joshi, Aravind K. 1985. Tree Adjoining Grammars: How much context-sensitivity is required to provide reasonable structural descriptions? In David R. Dowty, Lauri Karttunen, and Arnold M. Zwicky, editors, Natural Language Parsing. Cambridge University Press, pages 206–250. Joshi, Aravind K. and Yves Schabes. 1997. Tree-Adjoining Grammars. In Grzegorz Rozenberg and Arto Salomaa, editors, Handbook of Formal Languages, volume 3. Springer, pages 69–123. Kuhlmann, Marco, Alexander Koller, and Giorgio Satta. 2010. The importance of rule restrictions in CCG. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (ACL), pages 534–543, Uppsala. Moortgat, Michael. 2011. Categorial type logics. In Johan van Benthem and Alice ter
Meulen, editors, Handbook of Logic and Language. Elsevier, second edition, chapter 2, pages 95–179.
Pollard, Carl J. 1984. Generalized Phrase Structure Grammars, Head Grammars, and Natural Language. Ph.D. thesis, Stanford University. Shieber, Stuart M. 1985. Evidence against the context-freeness of natural language. Linguistics and Philosophy, 8(3):333–343. Steedman, Mark. 1996. Surface Structure and Interpretation, volume 30 of Linguistic Inquiry Monographs. MIT Press. Steedman, Mark. 2000. The Syntactic Process. MIT Press. Steedman, Mark. 2012. Taking Scope. MIT Press. Steedman, Mark and Jason Baldridge. 2011. Combinatory Categorial Grammar. In Robert D. Borsley and Kersti Börjars, editors, Non-Transformational Syntax: Formal and Explicit Models of Grammar. Blackwell, chapter 5, pages 181–224. Vijay-Shanker, K. and David J. Weir. 1993. Parsing some constrained grammar formalisms. Computational Linguistics, 19(4):591–636. Vijay-Shanker, K. and David J. Weir. 1994. The equivalence of four extensions of context-free grammars. Mathematical Systems Theory, 27(6):511–546. Vijay-Shanker, K., David J. Weir, and Aravind K. Joshi. 1986. Tree adjoining and head wrapping. In Proceedings of the Eleventh International Conference on Computational Linguistics (COLING), pages 202–207, Bonn. Weir, David J. 1988. Characterizing Mildly Context-Sensitive Grammar Formalisms. Ph.D. thesis, University of Pennsylvania. Weir, David J. and Aravind K. Joshi. 1988. Combinatory categorial grammars: Generative power and relationship to linear context-free rewriting systems. In Proceedings of the 26th Annual Meeting of the Association for Computational Linguistics (ACL), pages 278–285, Buffalo, NY. White, Michael. 2013. OpenCCG: The OpenNLP CCG Library. http://openccg.sourceforge.net/ Accessed November 13, 2013.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"lexicalization and generative power in ccg :Marco Kuhlmann∗ Linköping University
Alexander Koller∗∗ University of Potsdam
Giorgio Satta† University of Padua
The weak equivalence of Combinatory Categorial Grammar (CCG) and Tree-Adjoining Grammar (TAG) is a central result of the literature on mildly context-sensitive grammar formalisms. However, the categorial formalism for which this equivalence has been established differs significantly from the versions of CCG that are in use today. In particular, it allows restriction of combinatory rules on a per grammar basis, whereas modern CCG assumes a universal set of rules, isolating all cross-linguistic variation in the lexicon. In this article we investigate the formal significance of this difference. Our main result is that lexicalized versions of the classical CCG formalism are strictly less powerful than TAG.","vw-ccg  :restrictions. More specifically, we look at a variant of CCG consisting of the composition rules implemented in OpenCCG (White 2013), the most widely used development platform for CCG grammars. We show that this formalism is (almost) prefix-closed and cannot express target restrictions, which enables us to apply our generative capacity result from the first step. The same result holds for (the composition-only fragment of) the formalism of Baldridge and Kruijff (2003). Thus we find that, at least with existing means, the weak equivalence result of Vijay-Shanker and Weir cannot be obtained for lexicalized CCG. We conclude the article by discussing the implications of our results (Section 5).","proof  :No composition rule creates new arguments: Every argument that occurs in an output category already occurs in one of the input categories. Therefore, every argument must come from some word–category pair in the lexicon, of which there are only finitely many.
Lemma 2 The set of secondary input categories that occur in the derivations of a VW-CCG is finite.
3 Also, AB-grammar does not support lexicon entries for the empty string. Every secondary input category is obtained by substituting concrete categories for the variables that occur in the non-shaded component of one of the rules specified in Figure 4. After the substitution, all of these categories occur as part of arguments. Then, with Lemma 1, we deduce that the substituted categories come from a finite set. At the same time, each grammar specifies a finite set of rules. This means that there are only finitely many ways to obtain a secondary input category.
When specifying VW-CCGs, we find it convenient sometimes to provide an explicit list of valid rule instances, rather than a textual description of rule restrictions. For this we use a special type of restricted rule that we call templates. A template is a restricted rule that simultaneously fixes both
(a) the target of the primary input category of the rule, and
(b) the entire secondary input category.
We illustrate the idea with an example.
Example 4 We list the templates that correspond to the rule instances in the derivation from Figure 6. (The grammar allows other instances that are not listed here.) We use the symbol as a placeholder for that part of a primary input category that is unconstrained by rule restrictions, and therefore may consist of an arbitrary sequence of arguments.
A S /A S (1) S /C C /A S /A (2) S /B B/C S /C (3)
S /B B/C/B S /C/B (4)
For example, template (1) characterizes backward application (<0) where the target of the primary input category is S and the secondary input category is A, and template (4) characterizes forward composition of degree 2 (>2) where the target of the primary input category is S and the secondary input category is B/C/B.
Note that every VW-CCG can be specified using a finite set of templates: It has a finite set of combinatory rules; the set of possible targets of the primary input category of each rule is finite because each target is an atomic category; and the set of possible secondary input categories is finite because of Lemma 2. We are given a TAG G and construct a weakly equivalent VW-CCG G′. The basic idea is to make the lexical categories of G′ correspond to the elementary trees of G, and to set up the combinatory rules and their restrictions in such a way that the derivations of G′ correspond to derivations of G.
Vocabulary, Atomic Categories. The vocabulary of G′ is the set of all terminal symbols of G; the set of atomic categories consists of all symbols of the form At, where either A is a nonterminal symbol of G and t ∈ {a, c}, or A is a terminal symbol of G and t = a. The distinguished atomic category of G′ is Sa, where S is the start symbol of G.
Lexicon. One may assume (cf. Vijay-Shanker, Weir, and Joshi 1986) that G is in the normal form shown in Figure 9. In this normal form there is a single initial S-tree, and all remaining elementary trees are auxiliary trees of one of five possible types. For each such tree, one constructs two lexicon entries for the empty string ε as specified in Figure 9. Additionally, for each terminal symbol x of G, one constructs a lexicon entry x := xa.
Rules. The rules of G′ are forward and backward application and forward and backward composition of degree at most 2. They are used to simulate adjunction operations in derivations of G: Application simulates adjunction into nodes to the left or right of the foot node; composition simulates adjunction into nodes above the foot node. Without restrictions, these rules would allow derivations that do not correspond to derivations of G. Therefore, rules are restricted such that an argument of the form |At can be eliminated by means of an application rule only if t = a, and by means of a composition rule only if t = c. This enforces two properties that are central for the correctness of the construction (Weir 1988, p. 119): First, the secondary input category in every instance of composition is a category that has just been introduced from the
lexicon. Second, categories cannot be combined in arbitrary orders. The rule restrictions are:
1. Forward and backward application are restricted to instances where both the target of the primary input category and the entire secondary input category take the form Aa.
2. Forward and backward composition are restricted to instances where the target of the primary input category takes the form Aa and the target of the secondary input category takes the form Ac.
Using our template notation, the restricted rules can be written as in Figure 10.
As an aside we note that the proof of Lemma 3 makes heavy use of the ability of VW-CCG to assign lexicon entries to the empty string. Such lexicon entries violate one of the central linguistic principles of CCG, the Principle of Adjacency, according to which combinatory rules may only apply to phonologically realized entities (Steedman 2000, p. 54). It is an interesting question for future research whether a version of VW-CCG without lexicon entries for the empty string remains weakly equivalent to TAG. Every prefix-closed VW-CCG is a VW-CCG, therefore the inclusion follows from Theorem 1. To show that every TAG language can be generated by a prefix-closed VWCCG, we recall the construction of a weakly equivalent VW-CCG for a given TAG that we sketched in the proof of Lemma 3. As already mentioned in Example 6, the grammar G′ constructed there is not prefix-closed. However, we can make it prefixclosed by explicitly allowing the “missing” rule instances:
Aa /Bc Bc Aa where A, B ∈ V Bc Aa /Bc Aa where A, B ∈ V
We shall now argue that this modification does not actually change the language generated by G′. The only categories that qualify as secondary input categories of the new instances are atomic categories of the form Bc where B is a nonterminal of the TAG G. Now the lexical categories of G′ either are of the form xa (where x is a terminal symbol) or are non-atomic. Categories of the form Bc are not among the derived categories of G′ either, as the combinatory rules only yield output categories whose targets have the form Ba. This means that the new rule instances can never be used in a complete derivation of G′, and therefore do not change the generated language. Thus we have a construction that turns a TAG into a weakly equivalent prefix-closed VW-CCG. Every prefix-closed VW-CCG without target restrictions is a VW-CCG, so the inclusion follows from Theorem 1. To see that the inclusion is proper, consider the TAG language L1 = {anbncn | n ≥ 1} from Example 5. We are interested in sublanguages L′ ⊆ L1 that are Parikh-equivalent to the full language L1. This property is trivially satisfied by L1 itself. Moreover, it is not hard to see that L1 is in fact the only sublanguage of L1 that has this property. Now in Section 3.4 we shall prove a central lemma (Lemma 6), which asserts that, if we assume that L1 is generated by a prefix-closed VW-CCG without target restrictions, then at least one of the Parikh-equivalent sublanguages of L1 must be context-free. Because L1 is the only such sublanguage, this would give us proof that L1 is context-free; but we know it is not. Therefore we conclude that L1 is not generated by a prefix-closed VW-CCG without target restrictions.
Before turning to the proof of the central lemma (Lemma 6), we establish two other results about the languages generated by grammars without target restrictions.
Lemma 4 The languages generated by prefix-closed VW-CCG without target restrictions properly include the context-free languages. Inclusion follows from the fact that AB-grammars (which generate all context-free languages) are prefix-closed VW-CCGs without target restrictions. To see that the inclusion is proper, consider a grammar G2 that is like G1 but does not have any rule restrictions. This grammar is trivially prefix-closed and without target restrictions; it is actually “pure” in the sense of Kuhlmann, Koller, and Satta (2010). The language L2 = L(G2) contains all the strings in L1 = {anbncn | n ≥ 1}, together with other strings, including the string bbbacacac, whose derivation we showed in Figure 7. It is not hard to see that all of these additional strings have an equal number of as, bs, and cs. We can therefore write L1 as an intersection of L2 and a regular language: L1 = L2 ∩ a∗b∗c∗. To obtain a contradiction, suppose that L2 is context-free; then because of the fact that context-free languages are closed under intersection with regular languages, the language L1 would be context-free as well—but we know it is not. Therefore we conclude that L2 is not context-free either.
Lemma 5 The class of languages generated by prefix-closed VW-CCG without target restrictions is not closed under intersection with regular languages. If the class of languages generated by prefix-closed VW-CCG without target restrictions was closed under intersection with regular languages, then with L2 (the language mentioned in the previous proof) it would also include the language L1 = L2 ∩ a∗b∗c∗. However, from the proof of Theorem 3 we know that L1 is not generated by any prefixclosed VW-CCG without target restrictions. To argue that the system is terminating, we note that each rewriting step decreases the arity of one secondary input category in the derivation by one unit, while all other secondary input categories are left unchanged. As an example, consider rewriting under R1. The secondary input categories in the scope of that rule are Yβ/Z and Z on the left-hand side and Yβ and Z on the right-hand side. Here the arity of Yβ equals the arity of Yβ/Z, minus one. Because the system is terminating, to see that it is also confluent, it suffices to note that the left-hand sides of the rewrite rules do not overlap.
Lemma 8 The rewriting system transforms derivations of G into derivations of G. We prove the stronger result that every rewriting step transforms derivations of G into derivations of G. We only consider rewriting under R1; the arguments for the other rules are similar. Assume that R1 is applied to a derivation of G. The rule instances in the scope of the left-hand side of R1 take the following form:
X/Y Yβ/Z Xβ/Z (8) Xβ/Z Z Xβ (9)
Turning to the right-hand side, the rule instances in the rewritten derivation are
Yβ/Z Z Yβ (10) X/Y Yβ Xβ (11)
The relation between instances (8) and (11) is the characteristic relation of prefixclosed grammars (Definition 2): If instance (8) is valid, then because G is prefix-closed,
instance (11) is valid as well. Similarly, the relation between instances (9) and (10) is the characteristic relation of grammars without target restrictions (Definition 3): If instance (9) is valid, then because G is without target restrictions, instance (10) is valid as well. We conclude that if R1 is applied to a derivation of G, then the result is another derivation of G.
Combining Lemma 7 and Lemma 8, we see that for every derivation d of G, exhaustive application of the rewriting rules produces another uniquely determined derivation of G. We shall refer to this derivation as R(d). A transformed derivation is any derivation d′ such that d′ = R(d) for some derivation d. Let Y be the set of yields of the transformed derivations. Every string w′ ∈ Y is obtained from a string w ∈ L(G) by choosing some derivation d of w, rewriting this derivation into the transformed derivation R(d), and taking the yield. Inclusion then follows from Lemma 8. Because of the permuting rules R3 and R4, the strings w and w′ will in general be different. What we can say, however, is that w and w′ will be equal up to permutation. Thus we have established that Y and L(G) are Parikh-equivalent.
What remains in order to prove Lemma 6 is to show that the yields of the transformed derivations form a context-free language.
3.4.3 Context-Freeness of the Sublanguage. In a derivation tree, every node except the root node is labeled with either the primary or the secondary input category of a combinatory rule. We refer to these two types of nodes as primary nodes and secondary nodes, respectively. To simplify our presentation, we shall treat the root node as a secondary node. We restrict our attention to derivation trees for strings in L(G); in these trees, the root node is labeled with the distinguished atomic category S. For a leaf node u, the projection path of u is the path that starts at the parent of u and ends at the first secondary node that is encountered on the way towards the root node. We denote a projection path as a sequence X1, . . . , Xn (n ≥ 1), where X1 is the category at the parent of u and Xn is the category at the secondary node. Note that the category X1 is taken from the lexicon, while every other category is derived by combining the preceding category on the path with some secondary input category (not on the path) by means of some combinatory rule.
Example 10 In the derivation in Figure 6, the projection path of the first b goes all the way to the root, while all other projection paths have length 1, starting and ending with a lexical category. In Figure 7, the projection path of the first b ends at the root, while the projection paths of the remaining bs end at the nodes with category B, and the projection paths of the cs end at the nodes with category C.
A projection path X1, . . . , Xn is split if it can be segmented into two parts
X1, . . . , Xs and Xs, . . . , Xn (1 ≤ s ≤ n)
such that the first part only uses application rules and the second part only uses composition rules. Note that any part may consist of a single category only, in which case no combinatory rule is used in that part. If n = 1, then the path is trivially split. All projection paths in Figures 6 and 7 are split, except for the path of the first b in Figure 6, which alternates between composition (with C /A) and application (with A).
Lemma 10 In transformed derivations, every projection path is split. We show that as long as a derivation d contains a projection path that is not split, it can be rewritten. A projection path that is not split contains three adjacent categories U, V, W, such that V is derived by means of a composition with primary input U, and W is derived by means of an application with primary input V. Suppose that both the composition and the application are forward. (The arguments for the other three cases are similar.) Then U can be written as X/Y for some category X and argument /Y, V can be written as Xβ/Z for some argument /Z and some (possibly empty) sequence of arguments β, and W can be written as Xβ. We can then convince ourselves that d contains the following configuration, which matches the left-hand side of rewriting rule R1:
X/Y Yβ/Z Xβ/Z Z
Xβ
Lemma 11 The set of all categories that occur in transformed derivations is finite. Every category that occurs in transformed derivations occurs on some of its projection paths. Consider any such path. By Lemma 10 we know that this path is split; its two parts, here called P1 and P2, are visualized in Figure 13. We now reason about the arities of the categories in these two parts.
1. Because P1 only uses application, the arities in this part get smaller and smaller until they reach their minimum at Xs. This means that the arities of P1 are bounded by the arity of the first category on the path, which is a category from the lexicon.
2. Because P2 only uses composition, the arities in this part either get larger or stay the same until they reach a maximum at Xn. This means that the arities of P2 are bounded by the arity of the last category on the path, which is either the distinguished atomic category S or a secondary input category.
Thus the arities of our chosen path are bounded by the maximum of three grammarspecific constants: the maximal arity of a lexical category, the arity of S (which is 0), and the maximal arity of a secondary input category. The latter value is well-defined because there are only finitely many such categories (by Lemma 2). Let k be the maximum among the three constants, and let K be the set of all categories of the form A|mXm · · · |1X1 where A is an atomic category of G, m ≤ k, and each |iXi is an argument that may occur in derivations of G. The set K contains all categories that occur on some projection path, and therefore all categories that occur in transformed derivations, but it may also include other categories. As there are only finitely many atomic categories and finitely many arguments (Lemma 1), we conclude that the set K, and hence the set of categories that occur in transformed derivations, are finite as well.
Lemma 12 The transformed derivations yield a context-free language. We construct a context-free grammar H that generates the set Y of yields of the transformed derivations. To simplify the presentation, we first construct a grammar H′ that generates a superset of Y .
Construction of H′. The construction of the grammar H′ is the same as the construction in the classical proof that showed the context-freeness of AB-grammars, by Bar-Hillel, Gaifman, and Shamir (1960): The production rules of H′ are set up to correspond to the valid rule instances of G. The reason that this construction is not useful for VW-CCGs in general is that these may admit infinitely many rule instances, whereas a context-free grammar can only have finitely many productions. The set of rule instances may be infinite because VW-CCG has access to composition rules (specifically, rules of degrees greater than 1); in contrast, AB-grammars are restricted to application. Crucially though, by Lemma 11 we know that as long as we are interested only in transformed derivations it is sufficient to use a finite number of rule instances—more specifically those whose input and output categories are included in the set K of arity-bounded categories. Thus for every instance X/Y Yβ Xβ where all three categories are in K, we construct a production
[Xβ] → [X/Y] [Yβ]
and similarly for backward rules. (We enclose categories in square brackets for clarity.) In addition, for every lexicon entry σ := X in G we add to H′ a production [X] → σ. As the terminal alphabet of H′ we choose the vocabulary of G; as the nonterminal alphabet we choose the set K; and as the start symbol we choose the distinguished atomic category S. Every transformed derivation of G corresponds (in an obvious way) to some derivation in H′, which proves that Y ⊆ L(H′). Conversely, every derivation of H′ represents a derivation of G (though not necessarily a transformed derivation), thus L(H′) ⊆ L(G).
Construction of H. The chain of inclusions Y ⊆ L(H′) ⊆ L(G) is sufficient to prove Lemma 6: Because Y and L(G) are Parikh-equivalent (which we observed at the beginning of Section 3.4.2), so are L(H′) and L(G), which means that L(H′) satisfies all of the properties claimed in Lemma 6, even though this does not suffice to prove our current lemma. However, once H′ is given, it is not hard to also obtain a grammar H that generates exactly Y . For this, we need to filter out derivations whose projection paths do not have the characteristic property of transformed derivations that we established in Lemma 10. (It is not hard to see that every derivation that does have this property is a transformed derivation.) We annotate the left-hand side nonterminals in the productions of H′ with a flag t ∈ {a, c} to reflect whether the corresponding category has been derived by means of application (t = a) or composition (t = c); the value of this flag is simply the type of combinatory rule that gave rise to the production. The nonterminals in the right-hand sides are annotated in all possible ways, except that the following combinations are ruled out:
[X]a → [X/Y]c[Y]t and [X]a → [Y]t[X /Y]c t ∈ {a, c}
These combinations represent exactly the cases where the output category of a composition rule is used as the primary input category of an application rule, which are the cases that violate the “split” property that we established in Lemma 10.
This concludes the proof of Lemma 6, and therefore the proof of Theorem 3. Inclusion follows from the fact that every AB-grammar can be written as an O-CCG with only application (d = 0). To show that the inclusion is proper, we use the same argument as in the proof of Lemma 4. The grammar G2 that we constructed there can be turned into an equivalent O-CCG by decorating each slash with ·, the least restrictive type, and setting its inertness status to +.
What is less obvious is whether O-CCG generates the same class of languages as VW-CCG and TAG. Our main result is that this is not the case.
Theorem 4 The languages generated by O-CCG are properly included in the TAG languages.
O-CCG without Inertness. To approach Theorem 4, we set inertness aside for a moment and focus on the use of the slash types as a mechanism for imposing rule restrictions. Each of the rules in Figure 15 requires all of the slash types of the n outermost arguments of its secondary input category to be compatible with the rule, in the sense specified in Figure 14. If we now remove one or more of these arguments from a valid rule instance, then the new instance is clearly still valid, as we have reduced the number of potential violations of the type–rule compatibility. This shows that the rule system is prefix-closed. As none of the rules is conditioned on the target of the primary input category, the rule system is even without target restrictions. With these two properties established, Theorem 4 can be proved by literally the same arguments as those that we gave in Section 3. Thus we see directly that the theorem holds for versions of multimodal CCG without inertness, such as the formalism of Baldridge and Kruijff (2003).
O-CCG with Inertness. In the general case, the situation is complicated by the fact that the crossed composition rules change the inertness status of some argument categories if the slash types have conflicting directions. This means that the crossed composition rules in O-CCG are not entirely prefix-closed, as illustrated by the following example.
Example 11 Consider the following two rule instances:
X/+× Y Y / + ×Z2 / + × Z1 X / − ×Z2 / − × Z1 (12)
X/+× Y Y / + ×Z2 X / − ×Z2 (13)
Instance (12) is a valid instance of forward crossed composition. Prefix-closedness would require instance (13) to be valid as well; but it is not. In instance (12) the inertness status of /+ ×Z2 is changed for the only reason that the slash type of / + × Z1 does not match the required direction. In instance (13) the argument /+× Z1 is not present, and
therefore the inertness status of /+ ×Z2 is not changed, but is carried over to the output category.
We therefore have to prove that the following analogue of Lemma 6 holds for O-CCG:
Lemma 15 (Main Lemma for O-CCG) For every language L generated by some O-CCG there is a sublanguage L′ ⊆ L such that
1. L′ and L are Parikh-equivalent, and
2. L′ is context-free.
This lemma implies that the language L1 = {anbncn | n ≥ 1} from Example 2 cannot be generated by O-CCG (but by prefix-closed VW-CCG with target restrictions, and by TAG). The argument is the same as in the proof of Theorem 3. As in the proof of Lemma 8 we establish the stronger result that the claimed property holds for every single rewriting step. We only give the argument for rewriting under R3, which involves instances of forward crossed composition. The argument for R4 is analogous, and R1 and R2 are simpler cases because they involve harmonic composition, where the inertness status does not change.
Suppose that R3 is applied to a derivation of some O-CCG. In their most general form, the rule instances in the scope of the left-hand side of R3 may be written as follows, where the function σ is defined as specified in Figure 15:
X/+t0 Y Y / sn tn Zn · · · / s2 t2 Z2 / s1 t1 Z1 X / σ(sn) tn Zn · · · / σ(s2) t2 Z2 / σ(s1) t1 Z1 (14)
Z1 X / σ(sn) tn Zn · · · / σ(s2) t2 Z2 / + t1 Z1 X / σ(sn) tn Zn · · · / σ(s2) t1 Z2 (15)
Here instance (14) is an instance of forward-crossed composition, so each of the types ti is compatible with that rule. Because the two marked arguments are identical, we have σ(s1) = +. This is only possible if the inertness statuses of the slashes / si ti do not change in the context of derivation (14), that is, if σ(si) = si for all 1 ≤ i ≤ n. Note that in this case, t0 is either a right type or one of the four undirected core types, and each t1, . . . , tn is either a left type or a core type. We can now alternatively write instances (14) and (15) as
X/+t0 Y Y / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 X / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 (14 ′)
Z1 X / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 X / sn tn Zn · · · / s2 t1 Z2 (15 ′)
Then the rule instances in the rewritten derivation can be written as follows:
Z1 Y / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 Y / sn tn Zn · · · / s2 t2 Z2 (16)
X/+t0 Y Y / sn tn Zn · · · / s2 t2 Z2 X / sn tn Zn · · · / s2 t2 Z2 (17)
Here instance (16) is clearly a valid instance of backward application. Based on our earlier observations about the ti and their compatibility with crossed composition, we also see that instance (17) is a valid instance of forward crossed composition (if n > 1), or of forward application (if n = 1).
This completes the proof of Lemma 15. To finish the proof of Theorem 4 we have to also establish the inclusion of the O-CCG languages in the TAG languages. This is a known result for other dialects of multimodal CCG (Baldridge and Kruijff 2003), but O-CCG once again requires some extra work because of inertness.
Lemma 17 The O-CCG languages are included in the TAG languages. It suffices to show that the O-CCG languages are included in the class of languages generated by LIG (Gazdar 1987); the claim then follows from the weak equivalence of LIG and TAG. Vijay-Shanker and Weir (1994, Section 3.1) present a construction that transforms an arbitrary VW-CCG into a weakly equivalent LIG. It is straightforward to adapt their construction to O-CCG. As we do not have the space here to define LIG, we only provide a sketch of the adapted construction.
As in the case of VW-CCG, the valid instances of an O-CCG rule can be written down using our template notation. The adapted construction converts each such template into a production rule of a weakly equivalent LIG. Consider for instance the following instance of forward crossed composition from Example 11:
A /+× Y Y / + ×Z2 / + × Z1 A / − ×Z2 / − × Z1
This template is converted into the following LIG rule. We adopt the notation of VijayShanker and Weir (1994) and write ◦◦ for the tail of a stack of unbounded size.
A[◦◦ /− ×Z2 /−× Z1] → A[◦◦/+× Y] Y[ /+ ×Z2 /+× Z1]
In this way, every O-CCG can be written as a weakly equivalent LIG.",,,,,,,,,,,"1 introduction :Since the late 1970s, several grammar formalisms have been proposed that extend the power of context-free grammars in restricted ways. The two most prominent members of this class of “mildly context-sensitive” formalisms (a term coined by Joshi 1985) are Tree-Adjoining Grammar (TAG; Joshi and Schabes 1997) and Combinatory Categorial Grammar (CCG; Steedman 2000; Steedman and Baldridge 2011). Both formalisms have been applied to a broad range of linguistic phenomena, and are being widely used in computational linguistics and natural language processing.
In a seminal paper, Vijay-Shanker and Weir (1994) showed that TAG, CCG, and two other mildly context-sensitive formalisms—Head Grammar (Pollard 1984) and Linear Indexed Grammar (Gazdar 1987)—all characterize the same class of string languages. However, when citing this result it is sometimes overlooked that the result applies to a version of CCG that is quite different from the versions that are in practical use today.
∗ Department of Computer and Information Science, Linköping University, 581 83 Linköping, Sweden. E-mail: marco.kuhlmann@liu.se.
∗∗ Department of Linguistics, Karl-Liebknecht-Str. 24–25, University of Potsdam, 14476 Potsdam, Germany. E-mail: koller@ling.uni-potsdam.de.
† Department of Information Engineering, University of Padua, via Gradenigo 6/A, 35131 Padova, Italy. E-mail: satta@dei.unipd.it.
Submission received: 4 December 2013; revised submission received: 26 July 2014; accepted for publication: 25 November 2014.
doi:10.1162/COLI a 00219
© 2015 Association for Computational Linguistics
The goal of this article is to contribute to a better understanding of the significance of this difference.
The difference between “classical” CCG as formalized by Vijay-Shanker and Weir (1994) and the modern perspective may be illustrated with the combinatory rule of backward-crossed composition. The general form of this rule looks as follows:
Backward-crossed composition, general form:
Y/Z X /Y ⇒ X/Z (<B×)
This rule is frequently used in the analysis of heavy NP shift, as in the sentence Kahn blocked skillfully a powerful shot by Rivaldo (example from Baldridge 2002). However, backward crossed composition cannot be universally active in English as this would cause the grammar to also accept strings such as *a powerful by Rivaldo shot, which is witnessed by the derivation in Figure 1. To solve this problem, in the CCG formalism of Vijay-Shanker and Weir (1994), one may restrict backward-crossed composition to instances where X and Y are both verbal categories—that is, functions into the category S of sentences (cf. Steedman 2000, Section 4.2.2). With this restriction the unwanted derivation in Figure 1 can be blocked, and a powerful shot by Rivaldo is still accepted as grammatical. Other syntactic phenomena require other grammar-specific restrictions, including the complete ban of certain combinatory rules (cf. Steedman 2000, Section 4.2.1).
Over the past 20 years, CCG has evolved to put more emphasis on supporting fully lexicalized grammars (Baldridge and Kruijff 2003; Steedman and Baldridge 2011), in which as much grammatical information as possible is pushed into the lexicon. This follows the tradition of other frameworks such as Lexicalized Tree-Adjoining Grammar (LTAG) and Head-Driven Phrase Structure Grammar (HPSG). Grammar-specific rule restrictions are not connected to individual lexicon entries, and are therefore avoided. Instead, recent versions of CCG have introduced a new, lexicalized control mechanism in the form of modalities or slash types. The basic idea here is that combinatory rules only apply if the slashes in their input categories have the right types. For instance, the modern version of backward-crossed composition (Steedman and Baldridge 2011) takes the following form, where × is a slash type that allows crossed but not harmonic composition:
Backward-crossed composition, modalized form:
Y/×Z X /×Y ⇒ X/×Z (<B×)
The derivation in Figure 1 can now be blocked by assigning the slash in the category of powerful a type that is incompatible with ×, and thus cannot feed into type-aware backward-crossed composition as an input category. This “modalization” of grammatical composition, which imports some of the central ideas from the type-logical tradition of categorial grammar (Moortgat 2011) into the CCG framework, is attractive because it isolates the control over the applicability of combinatory rules in the lexicon. Quoting Steedman and Baldridge (2011, p. 186):
Without typed slashes, language-specific restrictions or even bans on some combinatory rules are necessary in order to block certain ungrammatical word orders. With them, the combinatory rules are truly universal: the grammar of every language utilizes exactly the same set of rules, without modification, thereby leaving all cross-linguistic variation in the lexicon. As such, CCG is a fully lexicalized grammar formalism.
The stated goal is thus to express all linguistically relevant restrictions on the use of combinatory rules in terms of lexically specified, typed slashes. But to what extent can this goal actually be achieved? So far, there have been only partial answers—even when, as in the formalism of Vijay-Shanker and Weir (1994), we restrict ourselves to the rules of composition, but exclude other rules such as type-raising and substitution. Baldridge and Kruijff (2003) note that the machinery of their multi-modal formalism can be simulated by rule restrictions, which shows that this version of lexicalized CCG is at most as expressive as the classical formalism. At the other end of the spectrum, we have shown in previous work that a “pure” form of lexicalized CCG with neither rule restrictions nor slash types is strictly less expressive (Kuhlmann, Koller, and Satta 2010). The general question of whether “classical” CCG (rule restrictions) and “modern” CCG (slash types instead of rule restrictions) are weakly equivalent has remained open.
In this article, we answer this question. Our results are summarized in Figure 2. After setting the stage (Section 2), we first pinpoint the exact type of rule restrictions that make the classical CCG formalism weakly equivalent to TAG (Section 3). We do so by focusing on a class of grammars that we call prefix-closed. Unlike the “pure” grammars that we studied in earlier work, prefix-closed grammars do permit rule restrictions (under certain conditions that seem to be satisfied by linguistic grammars). We show that the generative power of prefix-closed CCG depends on its ability to express target restrictions, which are exactly the functions-into type of restrictions that are needed to block the derivation in Figure 1. We prove that the full class of prefix-closed CCGs is weakly equivalent to TAG, and the subclass that cannot express target restrictions is strictly less powerful. This result significantly sharpens the picture of the generative capacity of CCG.
In a further step (Section 4), we then prove that for at least one popular incarnation of modern, lexicalized CCG, slash types are strictly less expressive than rule
Weakly equivalent to TAG Strictly less powerful than TAG 2 background :In this section we provide the technical background of the article. We introduce the basic architecture of CCG, present the formalism of Vijay-Shanker and Weir (1994), and set the points of reference for our results about generative capacity. The two central components of CCG are a lexicon that associates words with categories, and rules that specify how categories can be combined. Taken together, these components give rise to derivations, such as the one shown in Figure 3.
Lexicon. A category is a syntactic type that identifies a constituent as either “complete” or “incomplete.” The categories of complete constituents are taken to be primitives; the categories of incomplete constituents are modeled as (curried) functions that specify the type and the direction of their arguments. Every category is projected from the lexicon, which is a finite collection of word–category pairs. To give examples (taken from Steedman 2012), an English intransitive verb such as walks has a category that identifies it as a function seeking a (subject) noun phrase to its left (indicated by a backward slash) and returning a complete sentence; and a transitive verb such as admires has a category that identifies it as a function seeking an (object) noun phrase to its right (indicated by a forward slash) and returning a constituent that acts as an intransitive verb. We denote lexical assignment using the “colon equals” operator:
walks := S /NP admires := (S /NP)/NP
Formally, the set C(A) of categories is built over a finite set A of atomic categories, the primitive syntactic types. It is the smallest set such that
2. if X, Y ∈ C(A) then X/Y ∈ C(A) and X /Y ∈ C(A).
We use the letters A, B, C to denote atomic categories, the letters X, Y, Z to denote arbitrary categories, and the symbol | to denote slashes (forward or backward). We treat slashes as left-associative operators and omit unnecessary parentheses. By this convention, every category X can be written in the form
X = A|mXm · · · |1X1
where m ≥ 0, A is an atomic category that we call the target of X, and the |iXi are slash–category pairs that we call the arguments of X. Intuitively, the target of X is the atomic category that is obtained after X has been applied to all of its arguments. We use the Greek letters α, β, γ to denote (potentially empty) sequences of arguments. The number m is called the arity of the category X.
Rules. Categories can combine under a number of rules; this gives rise to derivations such as the one shown in Figure 3. Each rule specifies an operation that assembles two input categories into an output category. The two most basic rules of CCG are the directed versions of function application:
X/Y Y ⇒ X (forward application) (>) Y X /Y ⇒ X (backward application) (<)
Formally, a rule is a syntactic object in which the letters X, Y, Z act as variables for categories. A rule instance is obtained by substituting concrete categories for all variables in the rule. For example, the derivation in Figure 3 contains the following instances of function application. We denote rule instances by using a triple arrow instead of the double arrow in our notation for rules.
(S /NP)/NP NP S /NP and NP S /NP S
Application rules give rise to derivations equivalent to those of context-free grammar. Indeed, versions of categorial grammar where application is the only mode of combination, such as AB-grammar (Ajdukiewicz 1935; Bar-Hillel, Gaifman, and Shamir 1960), can only generate context-free languages. CCG can be more powerful because it also includes other rules, derived from the combinators of combinatory logic (Curry, Feys, and Craig 1958). In this article, as in most of the formal work on CCG, we restrict our attention to the rules of (generalized) composition, which are based on the B combinator.1
The general form of composition rules is shown in Figure 4. In each rule, the two input categories are distinguished into one primary (shaded) and one secondary input category. The number n of outermost arguments of the secondary input category is called the degree of the rule.2 In particular, for n = 0 we obtain the rules of function
1 This means that we ignore other rules required for linguistic analysis, in particular type-raising (from the T combinator), substitution (from the S combinator), and coordination. 2 The literature on CCG assumes a bound on n; for English, Steedman (2000, p. 42) puts n ≤ 3. Adding rules of unbounded degree increases the generative capacity of the formalism (Weir and Joshi 1988).
application. In contexts where we refer to both application and composition, we use the latter term for composition rules with degree n > 0.
Derivation Trees. Derivation trees can now be schematically defined as in Figure 5. They contain two types of branchings: unary branchings correspond to lexicon entries; binary branchings correspond to rule instances. The yield of a derivation tree is the left-to-right concatenation of its leaves. We now define the classical CCG formalism that was studied by Vijay-Shanker and Weir (1994) and originally introduced by Weir and Joshi (1988). As mentioned in Section 1, the central feature of this formalism is its ability to impose restrictions on the applicability of combinatory rules. Specifically, a restricted rule is a rule annotated with constraints that
(a) restrict the target of the primary input category; and/or
(b) restrict the secondary input category, either in parts or in its entirety.
Every grammar lists a finite number of restricted rules (where one and the same base rule may occur with several different restrictions). A valid rule instance is an instance that is compatible with at least one of the restricted rules.
Example 1 Linguistic grammars make frequent use of rule restrictions. To exclude the undesired derivation in Figure 1 we restricted backward crossed composition to instances where both the primary and the secondary input category are functions into the category of
sentences, S. Writing target for the function that returns the target of a category, the restricted rule can be written as
Y/Z X /Y ⇒ X/Z (backward crossed composition) (<B×) where target(X) = S and target(Y) = S
In the following definition we write ε for the empty string.
Definition 1 A combinatory categorial grammar in the sense of Vijay-Shanker and Weir (1994), or VW-CCG for short, is a structure
G = (Σ,A, :=, R, S)
where Σ is a finite vocabulary, A is a finite set of atomic categories, the lexicon := is a finite relation between the sets Σε = Σ ∪ {ε} and C(A), R is a finite set of (possibly) restricted rules, and S ∈ A is a distinguished atomic category.
A derivation tree of G consists of lexicon entries and valid instances of rules from R. The grammar G generates a string w if there exists a derivation tree whose yield is w and whose root node is labeled with the distinguished atomic category S. The language L(G) generated by G is the set of all generated strings.
Example 2 We specify a VW-CCG G1 = (Σ1,A1, :=1, R1, S1) that generates the language
L1 = {anbncn | n ≥ 1} .
Let Σ1 = {a, b, c}, A1 = {A, B, C, S}, S1 = S, and let :=1 consist of the entries
a :=1 A , b :=1 S/C , b :=1 S/C/B , b :=1 B/C , b :=1 B/C/B , c :=1 C /A
Finally, let R1 consist of the following rules:
1. Forward and backward application.
2. Forward and backward composition of degree 1.
3. Forward and backward composition of degree 2.
Each of these rules is restricted to instances where the target of the primary input category is S. A sample derivation of G1 is shown in Figure 6. By contrast, Figure 7 shows a derivation that is not valid in G1 because it uses application rules in which the target of the primary input category is B or C (rather than S).
Example 3 AB-grammar is a categorial grammar formalism in which forward and backward application are the only combinatory rules that are allowed. Furthermore, it does not support
rule restrictions.3 Every AB-grammar can therefore be written as a VW-CCG that allows all instances of application, but no other rules.
The following lemmas establish two fundamental properties of VW-CCG that we shall use at several places in this article. Both of them are minor variants of lemmas proved by Vijay-Shanker and Weir (1994) (Lemma 3.1 and Lemma 3.2) and were previously stated by Weir and Joshi (1988).
Lemma 1 The set of arguments that occur in the derivations of a VW-CCG is finite. In this article we are interested in the generative power of CCG, in particular in its relation to that of Tree-Adjoining Grammar. We therefore provide a compact introduction to TAG. For more details, we refer to Joshi and Schabes (1997).
Elementary Trees. Tree-Adjoining Grammar is a formalism for generating trees. These trees can be characterized as rooted, ordered trees in which internal nodes are labeled with nonterminal symbols—including a distinguished start symbol S—and leaf nodes are labeled with nonterminals, terminals, or the empty string. Every grammar specifies a finite set of such trees; these are called elementary trees. There are two types: initial trees and auxiliary trees. They differ in that auxiliary trees have a distinguished nonterminal-labeled leaf node, called foot node; this node is conventionally marked with an asterisk. An elementary tree whose root node is labeled with a nonterminal A is called an A-tree.
Substitution and Adjunction. New trees may be derived by combining other trees using two operations called substitution and adjunction. Substitution replaces some leaf node of a given tree with an initial tree (or a tree derived from an initial tree). Adjunction replaces some internal node u of a given tree with an auxiliary tree (or a tree derived from an auxiliary tree); the subtree with root u replaces the foot node of the auxiliary tree. All replacements are subject to the condition that the node being replaced and the root of the tree that replaces it are labeled with the same nonterminal.
Generated Languages. The tree language generated by a TAG is the set of all trees that can be derived from its initial S-trees. Derivations are considered complete only if they satisfy additional, node-specific constraints. In particular, substitution is obligatory at every node where it is possible, and adjunction may be specified as either obligatory (OA, Obligatory Adjunction) or forbidden (NA, Null Adjunction) at a given node. In derived trees corresponding to complete derivations, all leaf nodes are labeled with terminal symbols. The left-to-right concatenation of these symbols forms the yield of the tree, and the yields of all trees in the tree language form the string language generated by the TAG.
Example 5 Figure 8 shows a TAG that generates the language L1 = {anbncn | n ≥ 1} defined in Example 2. Derivations start with adjoining the auxiliary tree t2 at the root of the initial tree t1. New trees can be derived by repeatedly adjoining t2 at an S-node. Vijay-Shanker and Weir (1994) proved the following:
Theorem 1 VW-CCG and TAG are weakly equivalent.
The inclusion of the VW-CCG languages in the TAG languages follows from a chain of inclusions that connects VW-CCG and TAG via Linear Indexed Grammar (LIG; Gazdar 1987) and Head Grammar (HG; Pollard 1984). All of these inclusions were proved by Vijay-Shanker and Weir (1994). Here we sketch a proof of the inclusion of the TAG languages in the VW-CCG languages. Our proof closely follows that of Weir (1988, Section 5.2.2), whose construction we shall return to when establishing our own results.
Lemma 3 The TAG languages are included in the VW-CCG languages. 3 relevance of target restrictions in prefix-closed ccg :In this section we present the central technical results of this article. We study a class of VW-CCGs that we call prefix-closed and show that for this class, weak equivalence with TAG stands and falls with the ability to specify target restrictions. Rule restrictions are important tools in grammars for natural language; but not all of their potential uses have obvious linguistic motivations. For instance, one could write a grammar that permits all compositions with a functional category A/B as the secondary input category, but rules out application with the “shorter” category A. Such constraints do not seem to be used in linguistically motivated grammars; for example, none of the grammars developed by Steedman (2000) needs them. In prefix-closed grammars, this use of rule restrictions is explicitly barred.
Definition 2 A VW-CCG is prefix-closed if it satisfies the following implication:
if X/Y Y|nYn · · · |1Y1 X|nYn · · · |1Y1 is a valid rule instance then so is X/Y Y|nYn · · · |kYk X|nYn · · · |kYk for any k ≥ 1
and similarly for backward rules.
Note that prefix-closed grammars still allow some types of rule restrictions. The crucial property is that, if a certain combinatory rule applies at all, then it also applies to combinations where the secondary input category has already been applied to some (k ≤ n) or even all (k > n) of its arguments.
Example 6 We illustrate prefix-closedness using some examples:
1. Every AB-grammar (when seen as a VW-CCG) is trivially prefix-closed; in these grammars, n = 0.
2. The “pure” grammars that we considered in our earlier work (Kuhlmann, Koller, and Satta 2010) are trivially prefix-closed.
3. The grammar G1 from Example 2 is prefix-closed.
4. The grammars constructed in the proof of Lemma 3 are not prefix-closed; they do not allow the following instances of application, where the secondary input category is of the form Bc (rather than Ba):
Aa /Bc Bc Aa where A, B ∈ V Bc Aa /Bc Aa where A, B ∈ V
Example 7 The linguistic intuition underlying prefix-closed grammars is that if such a grammar allows us to delay the combination of a functor and its argument (via composition), then it also allows us to combine the functor and its argument immediately (via application). To illustrate this intuition, consider Figure 11, which shows two derivations related to the discussion of word order in Swiss German subordinate clauses (Shieber 1985):
. . . mer em Hans es huus hälfed aastriche . . . we Hansdat the houseacc helped paint
“. . . we helped Hans paint the house”
Derivation (5) (simplified from Steedman and Baldridge 2011, p. 201) starts by composing the tensed verb hälfed into the infinitive aastriche and then applies the resulting category to the accusative argument of the infinitive, es huus. Prefix-closedness implies that, if the combination of hälfed and aastriche is allowed when the latter is still waiting for es huus, then it must also be allowed if es huus has already been found.
Thus prefix-closedness predicts derivation (6), and along with it the alternative word order
. . . mer em Hans hälfed es huus aastriche . . . we Hansdat helped the houseacc paint
This word order is in fact grammatical (Shieber 1985, pp. 338–339). We now show that the restriction to prefix-closed grammars does not change the generative capacity of VW-CCG.
Theorem 2 Prefix-closed VW-CCG and TAG are weakly equivalent. In this section we shall see that the weak equivalence between prefix-closed VW-CCG and TAG depends on the ability to restrict the target of the primary input category in a combinatory rule. These are the restrictions that we referred to as constraints of type (a) in Section 2.2. We say that a grammar that does not make use of these constraints is without target restrictions. This property can be formally defined as follows.
Definition 3 A VW-CCG is without target restrictions if it satisfies the following implication:
if X/Y Yβ Xβ is a valid rule instance then so is X̄/Y Yβ X̄β for any category X̄ of the grammar
and similarly for backward rules.
Example 8
1. Every AB-grammar is without target restrictions; it allows forward and backward application for every primary input category.
2. The grammar G1 from Example (2) is not without target restrictions, because its rules are restricted to primary input categories with target S.
Target restrictions on the primary input category are useful in CCGs for natural languages; recall our discussion of backward-crossed composition in Section 1. As we shall see, target restrictions are also relevant from a formal point of view: If we require VW-CCGs to be without target restrictions, then we lose some of their weak generative capacity. This is the main technical result of this article. For its proof we need the following standard concept from formal language theory:
Definition 4 Two languages L and L′ are Parikh-equivalent if for every string w ∈ L there exists a permuted version w′ of w such that w′ ∈ L′, and vice versa.
Theorem 3 The languages generated by prefix-closed VW-CCG without target restrictions are properly included in the TAG languages. We shall now prove the central lemma that we used in the proof of Theorem 3.
Lemma 6 (Main Lemma for VW-CCG) For every language L that is generated by some prefix-closed VW-CCG without target restrictions, there is a sublanguage L′ ⊆ L such that
1. L′ and L are Parikh-equivalent, and
2. L′ is context-free.
Throughout this section, we let G be some arbitrary prefix-closed VW-CCG without target restrictions. The basic idea is to transform the derivations of G into a certain special form, and to prove that the transformed derivations yield a context-free language. The transformation is formalized by the rewriting system in Figure 12.4 To see how the rules of this system work, consider rule R1; the other rules are symmetric. Rule R1 rewrites an entire derivation into another derivation. It states that, whenever we have a situation where a category of the form X/Y is combined with a category of the form Yβ/Z by means of composition, and the resulting category is combined with a category Z by means of application, then we may just as well first combine Yβ/Z with Z, and then use the resulting category as a secondary input category together with X/Y.
4 Recall that we use the Greek letter β to denote a (possibly empty) sequence of arguments.
Note that R1 and R2 produce a new derivation for the original sentence, whereas R3 and R4 produce a derivation that yields a permutation of that sentence: The order of the substrings corresponding to the categories Z and X/Y (in the case of rule R3) or X /Y (in the case of rule R4) is reversed. In particular, R3 captures the relation between the two derivations of Swiss German word orders shown in Figure 11: Applying R3 to derivation (5) gives derivation (6). Importantly though, while the transformation may reorder the yield of a derivation, every transformed derivation still is a derivation of G.
Example 9 If we take the derivation in Figure 6 and exhaustively apply the rewriting rules from Figure 12, then the derivation that we obtain is the one in Figure 7. Note that although the latter derivation is not grammatical with respect to the grammar G1 from Example 2, it is grammatical with respect to the grammar G2 from the proof of Lemma 4, which is without target restrictions.
It is instructive to compare the rewriting rules in Figure 12 to the rules that establish the normal form of Eisner (1996). This normal form is used in practical CCG parsers to solve the problem of “spurious ambiguity,” where one and the same semantic interpretation (which in CCG takes the form of a lambda term) has multiple syntactic derivation trees. It is established by rewriting rules such as the following:
X/Y Yβ/Z Xβ/Z † Zγ
Xβγ
−→ X/Y Yβ/Z Zγ
Yβγ Xβγ
†† (5)
The rules in Figure 12 have much in common with the Eisner rules; yet there are two important differences. First, as already mentioned, our rules (in particular, rules R3 and R4) may reorder the yield of a derivation, whereas Eisner’s normal form
preserves yields. Second, our rules decrease the degrees of the involved composition operations, whereas Eisner’s rules may in fact increase them. To see this, note that the left-hand side of derivation (7) involves a composition of degree |β|+ 1 (†), whereas the right-hand side involves a composition of degree |β|+ |γ| (††). This means that rewriting will increase the degree in situations where |γ| > 1. In contrast, our rules only fire in the case where the combination with Z happens by means of an application, that is, if |γ| = 0. Under this condition, each rewrite step is guaranteed to decrease the degree of the composition. We will use this observation in the proof of Lemma 7.
3.4.1 Properties of the Transformation. The next two lemmas show that the rewriting system in Figure 12 implements a total function on the derivations of G.
Lemma 7 The rewriting system is terminating and confluent: Rewriting a derivation ends after a finite number of steps, and different rewriting orders all result in the same output. Lemma 9 The yields of the transformed derivations are a subset of and Parikh-equivalent to L(G). Theorem 3 pinpoints the exact mechanism that VW-CCG uses to achieve weak equivalence to TAG: At least for the class of prefix-closed grammars, TAG equivalence is achieved if and only if we allow target restrictions. Although target restrictions are frequently used in linguistically motivated grammars, it is important and perhaps surprising to realize that they are indeed necessary to achieve the full generative capacity of VW-CCG.
In the grammar formalisms folklore, the generative capacity of CCG is often attributed to generalized composition, and indeed we have seen (in Lemma 4) that even grammars without target restrictions can generate non-context-free languages such as L(G2). However, our results show that composition by itself is not enough to achieve weak equivalence with TAG: The yields of the transformed derivations from Section 3.4 form a context-free language despite the fact that these derivations may still contain compositions, including compositions of degree n > 2. In addition to composition, VWCCG also needs target restrictions to exert enough control on word order to block unwanted permutations. One way to think about this is that target restrictions can enforce alternations of composition and application (as in the derivation shown in
Figure 6), while transformed derivations are characterized by projection paths without such alternations (Lemma 10).
We can sharpen the picture even more by observing that the target restrictions that are crucial for the generative capacity of VW-CCG are not those on generalized composition, but those on function application. To see this we can note that the proof of Lemma 8 goes through also only if application rules such as (9) and (10) are without target restrictions. This means that we have the following qualification of Theorem 1.
Lemma 13 Prefix-closed VW-CCG is weakly equivalent to TAG only because it supports target restrictions on forward and backward application.
This finding is unexpected indeed—for instance, no grammar in Steedman (2000) uses target restrictions on the application rules. 4 generative capacity of multimodal ccg :After clarifying the mechanisms that “classical” CCG uses to achieve weak equivalence with TAG, we now turn our attention to “modern,” multimodal versions of CCG (Baldridge and Kruijff 2003; Steedman and Baldridge 2011). These versions emphasize the use of fully lexicalized grammars in which no rule restrictions are allowed, and instead equip slashes with types in order to control the use of the combinatory rules. Our central question is whether the use of slash types is sufficient to recover the expressiveness that we lose by giving up rule restrictions.
We need to fix a specific variant of multimodal CCG to study this question formally. Published works on multimodal CCG differ with respect to the specific inventories of slash types they assume. Some important details, such as a precise definition of generalized composition with slash types, are typically not discussed at all. In this article we define a variant of multimodal CCG which we call O-CCG. This formalism extends our definition of VW-CCG (Definition 1) with the slash inventory and the composition rules of the popular OpenCCG grammar development system (White 2013). Our technical result is that the main Lemma (Lemma 6) also holds for O-CCG. With this we can conclude that the answer to our question is negative: Slash types are not sufficient to replace rule restrictions; O-CCG is strictly less powerful than TAG. Although this is primarily a theoretical result, at the end of this section we also discuss its implications for practical grammar development. We define O-CCG as a formalism that extends VW-CCG with the slash types of OpenCCG, but abandons rule restrictions. Note that OpenCCG has a number of additional features that affect the generative capacity; we discuss these in Section 4.4.
Slash Types. Like other incarnations of multimodal CCG, O-CCG uses an enriched notion of categories where every slash has a type. There are eight such types:5
core types: ∗ × · left types: × right types: ×
5 The type system of OpenCCG is an extension of the system used by Baldridge (2002).
The basic idea behind these types is as follows. Slashes with type ∗ can only be used to instantiate application rules. Type also licenses harmonic composition rules, and type × also licenses crossed composition rules. Type · is the least restrictive type and can be used to instantiate all rules. The remaining types refine the system by incorporating a dimension of directionality. The exact type–rule compatibilities are specified in Figure 14.
Inertness. O-CCG is distinguished from other versions of multimodal CCG, such as that of Baldridge and Kruijff (2003), in that every slash not only has a type but also an inertness status. Inertness was introduced by Baldridge (2002, Section 8.2.2) as an implementation of the “antecedent government” (ANT) feature of Steedman (1996), which is used to control the word order in certain English relative clauses. It is a two-valued feature. Arguments whose slash type has inertness status + are called active; arguments whose slash type has inertness status − are called inert. Only active arguments can be eliminated by means of combinatory rules; however, an inert argument can still be consumed as part of a secondary input category. For example, the following instance of application is valid because the outermost slash of the primary input category has inertness status +:
X/+(Y/−Z) Y/−Z X
We use the notations /st and / s t to denote the forward and backward slashes with
slash type t and inertness status s.
Rules. All O-CCG grammars share a fixed set of combinatory rules, shown in Figure 15. Every grammar uses all rules, up to some grammar-specific bound on the degree of generalized composition. As mentioned earlier, a combinatory rule can only be instantiated if the slashes of the input categories have compatible types. Additionally, all composition rules require the slashes of the secondary input category to have a uniform direction. This is a somewhat peculiar feature of OpenCCG, and is in contrast to VW-CCG and other versions of CCG, which also allow composition rules with mixed directions.
Composition rules are classified into harmonic and crossed forms. This distinction is based on the direction of the slashes in the secondary input category. If these have the same direction as the outermost slash of the primary input category, then the rule is called harmonic; otherwise it is called crossed.6
6 In versions of CCG that allow rules with mixed slash directions, the distinction between harmonic and crossed is made based on the direction of the innermost slash of the secondary input category, |i.
When a rule is applied, in most cases the arguments of the secondary input category are simply copied into the output category, as in VW-CCG. The one exception happens for crossed composition rules if not all slash directions match the direction of their slash type (left or right). In this case, the arguments of the secondary input category become inert. Thus the inertness status of an argument may change over the course of a derivation—but only from active to inert, not back again.
Definition 5 A multimodal combinatory categorial grammar in the sense of OpenCCG, or O-CCG for short, is a structure
G = (Σ,A, :=, d, S)
where Σ is a finite vocabulary, A is a finite set of atomic categories, := is a finite relation between Σ and the set of (multimodal) categories over A, d ≥ 0 is the maximal degree of generalized composition, and S ∈ A is a distinguished atomic category.
We generalize the notions of rule instances, derivation trees, and generated language to categories over slashes with types and inertness statuses in the obvious way: Instead of two slashes, we now have one slash for every combination of a direction, type, and inertness status. Similarly, we generalize the concepts of a grammar being prefix-closed (Definition 2) and without target restrictions (Definition 3) to O-CCG. We now investigate the generative capacity of O-CCG. We start with the (unsurprising) observation that O-CCG can describe non-context-free languages.
Lemma 14 The languages generated by O-CCG properly include the context-free languages. The proof of Lemma 15 adapts the rewriting system from Figure 12. We simply let each rewriting step copy the type and inertness status of each slash from the left-hand side to the right-hand side of the rewriting rule. With this change, it is easy to verify that the proofs of Lemma 7 (termination and confluence), Lemma 10 (projection paths in transformed derivations are split), Lemma 11 (transformed derivations contain a finite number of categories), and Lemma 12 (transformed derivations yield a contextfree language) go through without problems. The proof of Lemma 8, however, is not straightforward, because of the dynamic nature of the inertness statuses. We therefore restate the lemma for O-CCG:
Lemma 16 The rewriting system transforms O-CCG derivations into O-CCG derivations. In this section we have shown that the languages generated by O-CCG are properly included in the languages generated by TAG, and equivalently, in the languages generated by VW-CCG. This means that the multimodal machinery of OpenCCG is not powerful enough to express the rule restrictions of VW-CCG in a fully lexicalized way. The result is easy to obtain for O-CCG without inertness, which is prefix-closed and without target restrictions; but it is remarkably robust in that it also applies to O-CCG with inertness, which is not prefix-closed. As we have already mentioned, the result carries over also to other multimodal versions of CCG, such as the formalism of Baldridge and Kruijff (2003).
Our result has implications for practical grammar development with OpenCCG. To illustrate this, recall Example 7, which showed that every VW-CCG without target restrictions for Swiss German that allows cross–serial word orders as in derivation (5) also permits alternative word orders, as in derivation (6). By Lemma 15, this remains true for O-CCG or weaker multimodal formalisms. This is not a problem in the case of Swiss German, where the alternative word orders are grammatical. However, there is at least one language, Dutch, where dependencies in subordinate clauses must cross. For this case, our result shows that the modalized composition rules of OpenCCG are not powerful enough to write adequate grammars. Consider the following classical example:
. . . ik Cecilia de paarden zag voeren . . . I Cecilia the horses saw feed
“. . . I saw Cecilia feed the horses”
The straightforward derivation of the cross–serial dependencies in this sentence (adapted from Steedman 2000, p. 141) is exemplified in Figure 16. It takes the same form as derivation (5) for Swiss German: The verbs and their NP arguments lie on a single, right-branching path projected from the tensed verb zag. This projection path is not split; specifically, it starts with a composition that produces a category which acts as the primary input category of an application. As a consequence, the derivation can be transformed (by our rewriting rule R3) in exactly the same way as instance (5) could be transformed into derivation (6). The crucial difference is that the yield of the transformed derivation, *ik Cecilia zag de paarden voeren, is not a grammatical clause of Dutch.
To address the problem of ungrammatical word orders in Dutch subordinate clauses, the VW-CCG grammar of Steedman (2000) and the multimodal CCG grammar of Baldridge (2002, Section 5.3.1) resort to combinatorial rules other than composition. In particular, they assume that all complement noun phrases undergo obligatory type-raising, and become primary input categories of application rules. This gives rise to derivations such as the one shown in Figure 17, which cannot be transformed using our rewriting rules because the result of the forward crossed composition >1
now is a secondary rather than a primary input category. As a consequence, this grammar is capable of enforcing the obligatory cross-serial dependencies of Dutch. However, it is important to note that it requires type-raising over arbitary categories with target S (observe the increasingly complex type-raised categories for the NPs). This kind of type-raising is allowed in many variants of CCG, including the full formalism underlying OpenCCG. VW-CCG and O-CCG, however, are limited to generalized composition, and can only support derivations like the one in Figure 17 if all the type-raised categories for the noun phrases are available in the lexicon. The unbounded type-raising required by the Steedman–Baldridge analysis of Dutch would translate into an infinite lexicon, and so this analysis is not possible in VW-CCG and O-CCG.
We conclude by discussing the impact of several other constructs of OpenCCG that we have not captured in O-CCG. First, OpenCCG allows us to use generalized composition rules of arbitrary degree; there is no upper bound d on the composition degree as in an O-CCG grammar. It is known that this extends the generative capacity of CCG beyond that of TAG (Weir 1988). Second, OpenCCG allows categories to be annotated with feature structures. This has no impact on the generative capacity, as the features must take values from finite domains and can therefore be compiled into the atomic categories of the grammar. Finally, OpenCCG includes the combinatory rules of substitution and coordination, as well as multiset slashes, another extension frequently used in linguistic grammars. We have deliberately left these constructs out of O-CCG to establish the most direct comparison to the literature on VW-CCG. It is conceivable that their inclusion could restore the weak equivalence to TAG, but a proof of this result would require a non-trivial extension of the work of Vijay-Shanker and Weir (1994). Regarding multiset slashes, it is also worth noting that these were introduced with the expressed goal of allowing more flexible word
order, whereas restoration of weak equivalence would require more controlled word order. 5 conclusion :In this article we have contributed two technical results to the literature on CCG. First, we have refined the weak equivalence result for CCG and TAG (Vijay-Shanker and Weir 1994) by showing that prefix-closed grammars are weakly equivalent to TAG only if target restrictions are allowed. Second, we have shown that O-CCG, the formal, composition-only core of OpenCCG, is not weakly equivalent to TAG. These results point to a tension in CCG between lexicalization and generative capacity: Lexicalized versions of the framework are less powerful than classical versions, which allow rule restrictions.
What conclusions one draws from these technical results depends on the perspective. One way to look at CCG is as a system for defining formal languages. Under this view, one is primarily interested in results on generative capacity and parsing complexity such as those obtained by Vijay-Shanker and Weir (1993, 1994). Here, our results clarify the precise mechanisms that make CCG weakly equivalent to TAG. Perhaps surprisingly, it is not the availability of generalized composition rules by itself that explains the generative power of CCG, but the ability to constrain the interaction between generalized composition and function application by means of target restrictions.
On the other hand, one may be interested in CCG primarily as a formalism for developing grammars for natural languages (Steedman 2000; Baldridge 2002; Steedman 2012). From this point of view, the suitability of CCG for the development of lexicalized grammars has been amply demonstrated. However, our technical results still serve as important reminders that extra care must be taken to avoid overgeneration when designing a grammar. In particular, it is worth double-checking that an OpenCCG grammar does not generate word orders that the grammar developer did not intend. Here the rewriting system that we presented in Figure 12 can serve as a useful tool: A grammar developer can take any derivation for a grammatical sentence, transform the derivation according to our rewriting rules, and check whether the transformed derivation still yields a grammatical sentence.
It remains an open question how the conflicting desires for generative capacity and lexicalization might be reconciled. A simple answer is to add some lexicalized method for enforcing target restrictions to CCG, specifically on the application rules. However, we are not aware that this idea has seen widespread use in the CCG literature, so it may not be called for empirically. Alternatively, one might modify the rules of O-CCG in such a way that they are no longer prefix-closed—for example, by introducing some new slash type. Finally, it is possible that the constructs of OpenCCG that we set aside in O-CCG (such as type-raising, substitution, and multiset slashes) might be sufficient to achieve the generative capacity of classical CCG and TAG. A detailed study of the expressive power of these constructs would make an interesting avenue for future research. The weak equivalence of Combinatory Categorial Grammar (CCG) and Tree-Adjoining Grammar (TAG) is a central result of the literature on mildly context-sensitive grammar formalisms. However, the categorial formalism for which this equivalence has been established differs significantly from the versions of CCG that are in use today. In particular, it allows restriction of combinatory rules on a per grammar basis, whereas modern CCG assumes a universal set of rules, isolating all cross-linguistic variation in the lexicon. In this article we investigate the formal significance of this difference. Our main result is that lexicalized versions of the classical CCG formalism are strictly less powerful than TAG. [{""affiliations"": [], ""name"": ""Marco Kuhlmann""}, {""affiliations"": [], ""name"": ""Alexander Koller""}, {""affiliations"": [], ""name"": ""Giorgio Satta""}] SP:52fb58337ad0d4ec8da3bb73ae06c06bd09b356a [{""authors"": [""Ajdukiewicz"", ""Kazimierz.""], ""title"": ""Die syntaktische Konnexit\u00e4t"", ""venue"": ""Studia Philosophica, 1:1\u201327."", ""year"": 1935}, {""authors"": [""Baldridge"", ""Jason"", ""Geert-Jan M. Kruijff.""], ""title"": ""Multi-modal combinatory categorial grammar"", ""venue"": ""Tenth Conference of the European Chapter of the Association for Computational Linguistics (EACL),"", ""year"": 2003}, {""authors"": [""Bar-Hillel"", ""Yehoshua"", ""Haim Gaifman"", ""Eli Shamir.""], ""title"": ""On categorial and phrase structure grammars"", ""venue"": ""Bulletin of the Research Council of Israel, 9F(1):1\u201316. Reprinted in Yehoshua Bar-Hillel. Language and"", ""year"": 1960}, {""authors"": [""Curry"", ""Haskell B."", ""Robert Feys"", ""William Craig.""], ""title"": ""Combinatory Logic"", ""venue"": ""Volume 1. Studies in Logic and the Foundations of Mathematics. North-Holland."", ""year"": 1958}, {""authors"": [""Eisner"", ""Jason.""], ""title"": ""Efficient normal-form parsing for Combinatory Categorial Grammar"", ""venue"": ""Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics (ACL), pages 79\u201386,"", ""year"": 1996}, {""authors"": [""Gazdar"", ""Gerald.""], ""title"": ""Applicability of indexed grammars to natural language"", ""venue"": ""Uwe Reyle and Christian Rohrer, editors, Natural Language Parsing and Linguistic Theories. D. Reidel, pages 69\u201394."", ""year"": 1987}, {""authors"": [""Joshi"", ""Aravind K.""], ""title"": ""Tree Adjoining Grammars: How much context-sensitivity is required to provide reasonable structural descriptions? In David R"", ""venue"": ""Dowty, Lauri Karttunen, and Arnold M."", ""year"": 1985}, {""authors"": [""Joshi"", ""Aravind K."", ""Yves Schabes.""], ""title"": ""Tree-Adjoining Grammars"", ""venue"": ""Grzegorz Rozenberg and Arto Salomaa, editors, Handbook of Formal Languages, volume 3. Springer, pages 69\u2013123."", ""year"": 1997}, {""authors"": [""Kuhlmann"", ""Marco"", ""Alexander Koller"", ""Giorgio Satta.""], ""title"": ""The importance of rule restrictions in CCG"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (ACL),"", ""year"": 2010}, {""authors"": [""Moortgat"", ""Michael.""], ""title"": ""Categorial type logics"", ""venue"": ""Johan van Benthem and Alice ter"", ""year"": 2011}, {""authors"": [""Pollard"", ""Carl J.""], ""title"": ""Generalized Phrase Structure Grammars, Head Grammars, and Natural Language"", ""venue"": ""Ph.D. thesis, Stanford University."", ""year"": 1984}, {""authors"": [""Shieber"", ""Stuart M.""], ""title"": ""Evidence against the context-freeness of natural language"", ""venue"": ""Linguistics and Philosophy, 8(3):333\u2013343."", ""year"": 1985}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""Surface Structure and Interpretation, volume 30 of Linguistic Inquiry Monographs"", ""venue"": ""MIT Press."", ""year"": 1996}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""The Syntactic Process"", ""venue"": ""MIT Press."", ""year"": 2000}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""Taking Scope"", ""venue"": ""MIT Press."", ""year"": 2012}, {""authors"": [""Steedman"", ""Mark"", ""Jason Baldridge.""], ""title"": ""Combinatory Categorial Grammar"", ""venue"": ""Robert D. Borsley and Kersti B\u00f6rjars, editors, Non-Transformational Syntax: Formal and Explicit Models of Grammar."", ""year"": 2011}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""Parsing some constrained grammar formalisms"", ""venue"": ""Computational Linguistics, 19(4):591\u2013636."", ""year"": 1993}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""The equivalence of four extensions of context-free grammars"", ""venue"": ""Mathematical Systems Theory, 27(6):511\u2013546."", ""year"": 1994}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir"", ""Aravind K. Joshi.""], ""title"": ""Tree adjoining and head wrapping"", ""venue"": ""Proceedings of the Eleventh International Conference on Computational Linguistics (COLING),"", ""year"": 1986}, {""authors"": [""Weir"", ""David J.""], ""title"": ""Characterizing Mildly Context-Sensitive Grammar Formalisms"", ""venue"": ""Ph.D. thesis, University of Pennsylvania."", ""year"": 1988}, {""authors"": [""Weir"", ""David J."", ""Aravind K. Joshi.""], ""title"": ""Combinatory categorial grammars: Generative power and relationship to linear context-free rewriting systems"", ""venue"": ""Proceedings of the 26th Annual Meeting"", ""year"": 1988}, {""authors"": [""White"", ""Michael.""], ""title"": ""OpenCCG: The OpenNLP CCG Library"", ""venue"": ""http://openccg.sourceforge.net/ Accessed November 13, 2013. 219"", ""year"": 2013}] acknowledgments  :We are grateful to Mark Steedman and Jason Baldridge for enlightening discussions of the material presented in this article, and to the four anonymous reviewers of the article for their detailed and constructive comments.
References Ajdukiewicz, Kazimierz. 1935. Die
syntaktische Konnexität. Studia Philosophica, 1:1–27.
Baldridge, Jason. 2002. Lexically Specified Derivational Control in Combinatory
Categorial Grammar. Ph.D. thesis, University of Edinburgh, Edinburgh, UK.
Baldridge, Jason and Geert-Jan M. Kruijff. 2003. Multi-modal combinatory categorial grammar. In Tenth Conference of the European Chapter of the Association for Computational Linguistics (EACL), pages 211–218, Budapest. Bar-Hillel, Yehoshua, Haim Gaifman, and Eli Shamir. 1960. On categorial and phrase structure grammars. Bulletin of the Research Council of Israel, 9F(1):1–16. Reprinted in Yehoshua Bar-Hillel. Language and Information: Selected Essays on Their Theory and Application, pages 99–115. Addison-Wesley, 1964. Curry, Haskell B., Robert Feys, and William Craig. 1958. Combinatory Logic. Volume 1. Studies in Logic and the Foundations of Mathematics. North-Holland. Eisner, Jason. 1996. Efficient normal-form parsing for Combinatory Categorial Grammar. In Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics (ACL), pages 79–86, Santa Cruz, CA. Gazdar, Gerald. 1987. Applicability of indexed grammars to natural language. In Uwe Reyle and Christian Rohrer, editors, Natural Language Parsing and Linguistic Theories. D. Reidel, pages 69–94. Joshi, Aravind K. 1985. Tree Adjoining Grammars: How much context-sensitivity is required to provide reasonable structural descriptions? In David R. Dowty, Lauri Karttunen, and Arnold M. Zwicky, editors, Natural Language Parsing. Cambridge University Press, pages 206–250. Joshi, Aravind K. and Yves Schabes. 1997. Tree-Adjoining Grammars. In Grzegorz Rozenberg and Arto Salomaa, editors, Handbook of Formal Languages, volume 3. Springer, pages 69–123. Kuhlmann, Marco, Alexander Koller, and Giorgio Satta. 2010. The importance of rule restrictions in CCG. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (ACL), pages 534–543, Uppsala. Moortgat, Michael. 2011. Categorial type logics. In Johan van Benthem and Alice ter
Meulen, editors, Handbook of Logic and Language. Elsevier, second edition, chapter 2, pages 95–179.
Pollard, Carl J. 1984. Generalized Phrase Structure Grammars, Head Grammars, and Natural Language. Ph.D. thesis, Stanford University. Shieber, Stuart M. 1985. Evidence against the context-freeness of natural language. Linguistics and Philosophy, 8(3):333–343. Steedman, Mark. 1996. Surface Structure and Interpretation, volume 30 of Linguistic Inquiry Monographs. MIT Press. Steedman, Mark. 2000. The Syntactic Process. MIT Press. Steedman, Mark. 2012. Taking Scope. MIT Press. Steedman, Mark and Jason Baldridge. 2011. Combinatory Categorial Grammar. In Robert D. Borsley and Kersti Börjars, editors, Non-Transformational Syntax: Formal and Explicit Models of Grammar. Blackwell, chapter 5, pages 181–224. Vijay-Shanker, K. and David J. Weir. 1993. Parsing some constrained grammar formalisms. Computational Linguistics, 19(4):591–636. Vijay-Shanker, K. and David J. Weir. 1994. The equivalence of four extensions of context-free grammars. Mathematical Systems Theory, 27(6):511–546. Vijay-Shanker, K., David J. Weir, and Aravind K. Joshi. 1986. Tree adjoining and head wrapping. In Proceedings of the Eleventh International Conference on Computational Linguistics (COLING), pages 202–207, Bonn. Weir, David J. 1988. Characterizing Mildly Context-Sensitive Grammar Formalisms. Ph.D. thesis, University of Pennsylvania. Weir, David J. and Aravind K. Joshi. 1988. Combinatory categorial grammars: Generative power and relationship to linear context-free rewriting systems. In Proceedings of the 26th Annual Meeting of the Association for Computational Linguistics (ACL), pages 278–285, Buffalo, NY. White, Michael. 2013. OpenCCG: The OpenNLP CCG Library. http://openccg.sourceforge.net/ Accessed November 13, 2013. lexicalization and generative power in ccg :Marco Kuhlmann∗ Linköping University
Alexander Koller∗∗ University of Potsdam
Giorgio Satta† University of Padua
The weak equivalence of Combinatory Categorial Grammar (CCG) and Tree-Adjoining Grammar (TAG) is a central result of the literature on mildly context-sensitive grammar formalisms. However, the categorial formalism for which this equivalence has been established differs significantly from the versions of CCG that are in use today. In particular, it allows restriction of combinatory rules on a per grammar basis, whereas modern CCG assumes a universal set of rules, isolating all cross-linguistic variation in the lexicon. In this article we investigate the formal significance of this difference. Our main result is that lexicalized versions of the classical CCG formalism are strictly less powerful than TAG. vw-ccg  :restrictions. More specifically, we look at a variant of CCG consisting of the composition rules implemented in OpenCCG (White 2013), the most widely used development platform for CCG grammars. We show that this formalism is (almost) prefix-closed and cannot express target restrictions, which enables us to apply our generative capacity result from the first step. The same result holds for (the composition-only fragment of) the formalism of Baldridge and Kruijff (2003). Thus we find that, at least with existing means, the weak equivalence result of Vijay-Shanker and Weir cannot be obtained for lexicalized CCG. We conclude the article by discussing the implications of our results (Section 5). proof  :No composition rule creates new arguments: Every argument that occurs in an output category already occurs in one of the input categories. Therefore, every argument must come from some word–category pair in the lexicon, of which there are only finitely many.
Lemma 2 The set of secondary input categories that occur in the derivations of a VW-CCG is finite.
3 Also, AB-grammar does not support lexicon entries for the empty string. Every secondary input category is obtained by substituting concrete categories for the variables that occur in the non-shaded component of one of the rules specified in Figure 4. After the substitution, all of these categories occur as part of arguments. Then, with Lemma 1, we deduce that the substituted categories come from a finite set. At the same time, each grammar specifies a finite set of rules. This means that there are only finitely many ways to obtain a secondary input category.
When specifying VW-CCGs, we find it convenient sometimes to provide an explicit list of valid rule instances, rather than a textual description of rule restrictions. For this we use a special type of restricted rule that we call templates. A template is a restricted rule that simultaneously fixes both
(a) the target of the primary input category of the rule, and
(b) the entire secondary input category.
We illustrate the idea with an example.
Example 4 We list the templates that correspond to the rule instances in the derivation from Figure 6. (The grammar allows other instances that are not listed here.) We use the symbol as a placeholder for that part of a primary input category that is unconstrained by rule restrictions, and therefore may consist of an arbitrary sequence of arguments.
A S /A S (1) S /C C /A S /A (2) S /B B/C S /C (3)
S /B B/C/B S /C/B (4)
For example, template (1) characterizes backward application (<0) where the target of the primary input category is S and the secondary input category is A, and template (4) characterizes forward composition of degree 2 (>2) where the target of the primary input category is S and the secondary input category is B/C/B.
Note that every VW-CCG can be specified using a finite set of templates: It has a finite set of combinatory rules; the set of possible targets of the primary input category of each rule is finite because each target is an atomic category; and the set of possible secondary input categories is finite because of Lemma 2. We are given a TAG G and construct a weakly equivalent VW-CCG G′. The basic idea is to make the lexical categories of G′ correspond to the elementary trees of G, and to set up the combinatory rules and their restrictions in such a way that the derivations of G′ correspond to derivations of G.
Vocabulary, Atomic Categories. The vocabulary of G′ is the set of all terminal symbols of G; the set of atomic categories consists of all symbols of the form At, where either A is a nonterminal symbol of G and t ∈ {a, c}, or A is a terminal symbol of G and t = a. The distinguished atomic category of G′ is Sa, where S is the start symbol of G.
Lexicon. One may assume (cf. Vijay-Shanker, Weir, and Joshi 1986) that G is in the normal form shown in Figure 9. In this normal form there is a single initial S-tree, and all remaining elementary trees are auxiliary trees of one of five possible types. For each such tree, one constructs two lexicon entries for the empty string ε as specified in Figure 9. Additionally, for each terminal symbol x of G, one constructs a lexicon entry x := xa.
Rules. The rules of G′ are forward and backward application and forward and backward composition of degree at most 2. They are used to simulate adjunction operations in derivations of G: Application simulates adjunction into nodes to the left or right of the foot node; composition simulates adjunction into nodes above the foot node. Without restrictions, these rules would allow derivations that do not correspond to derivations of G. Therefore, rules are restricted such that an argument of the form |At can be eliminated by means of an application rule only if t = a, and by means of a composition rule only if t = c. This enforces two properties that are central for the correctness of the construction (Weir 1988, p. 119): First, the secondary input category in every instance of composition is a category that has just been introduced from the
lexicon. Second, categories cannot be combined in arbitrary orders. The rule restrictions are:
1. Forward and backward application are restricted to instances where both the target of the primary input category and the entire secondary input category take the form Aa.
2. Forward and backward composition are restricted to instances where the target of the primary input category takes the form Aa and the target of the secondary input category takes the form Ac.
Using our template notation, the restricted rules can be written as in Figure 10.
As an aside we note that the proof of Lemma 3 makes heavy use of the ability of VW-CCG to assign lexicon entries to the empty string. Such lexicon entries violate one of the central linguistic principles of CCG, the Principle of Adjacency, according to which combinatory rules may only apply to phonologically realized entities (Steedman 2000, p. 54). It is an interesting question for future research whether a version of VW-CCG without lexicon entries for the empty string remains weakly equivalent to TAG. Every prefix-closed VW-CCG is a VW-CCG, therefore the inclusion follows from Theorem 1. To show that every TAG language can be generated by a prefix-closed VWCCG, we recall the construction of a weakly equivalent VW-CCG for a given TAG that we sketched in the proof of Lemma 3. As already mentioned in Example 6, the grammar G′ constructed there is not prefix-closed. However, we can make it prefixclosed by explicitly allowing the “missing” rule instances:
Aa /Bc Bc Aa where A, B ∈ V Bc Aa /Bc Aa where A, B ∈ V
We shall now argue that this modification does not actually change the language generated by G′. The only categories that qualify as secondary input categories of the new instances are atomic categories of the form Bc where B is a nonterminal of the TAG G. Now the lexical categories of G′ either are of the form xa (where x is a terminal symbol) or are non-atomic. Categories of the form Bc are not among the derived categories of G′ either, as the combinatory rules only yield output categories whose targets have the form Ba. This means that the new rule instances can never be used in a complete derivation of G′, and therefore do not change the generated language. Thus we have a construction that turns a TAG into a weakly equivalent prefix-closed VW-CCG. Every prefix-closed VW-CCG without target restrictions is a VW-CCG, so the inclusion follows from Theorem 1. To see that the inclusion is proper, consider the TAG language L1 = {anbncn | n ≥ 1} from Example 5. We are interested in sublanguages L′ ⊆ L1 that are Parikh-equivalent to the full language L1. This property is trivially satisfied by L1 itself. Moreover, it is not hard to see that L1 is in fact the only sublanguage of L1 that has this property. Now in Section 3.4 we shall prove a central lemma (Lemma 6), which asserts that, if we assume that L1 is generated by a prefix-closed VW-CCG without target restrictions, then at least one of the Parikh-equivalent sublanguages of L1 must be context-free. Because L1 is the only such sublanguage, this would give us proof that L1 is context-free; but we know it is not. Therefore we conclude that L1 is not generated by a prefix-closed VW-CCG without target restrictions.
Before turning to the proof of the central lemma (Lemma 6), we establish two other results about the languages generated by grammars without target restrictions.
Lemma 4 The languages generated by prefix-closed VW-CCG without target restrictions properly include the context-free languages. Inclusion follows from the fact that AB-grammars (which generate all context-free languages) are prefix-closed VW-CCGs without target restrictions. To see that the inclusion is proper, consider a grammar G2 that is like G1 but does not have any rule restrictions. This grammar is trivially prefix-closed and without target restrictions; it is actually “pure” in the sense of Kuhlmann, Koller, and Satta (2010). The language L2 = L(G2) contains all the strings in L1 = {anbncn | n ≥ 1}, together with other strings, including the string bbbacacac, whose derivation we showed in Figure 7. It is not hard to see that all of these additional strings have an equal number of as, bs, and cs. We can therefore write L1 as an intersection of L2 and a regular language: L1 = L2 ∩ a∗b∗c∗. To obtain a contradiction, suppose that L2 is context-free; then because of the fact that context-free languages are closed under intersection with regular languages, the language L1 would be context-free as well—but we know it is not. Therefore we conclude that L2 is not context-free either.
Lemma 5 The class of languages generated by prefix-closed VW-CCG without target restrictions is not closed under intersection with regular languages. If the class of languages generated by prefix-closed VW-CCG without target restrictions was closed under intersection with regular languages, then with L2 (the language mentioned in the previous proof) it would also include the language L1 = L2 ∩ a∗b∗c∗. However, from the proof of Theorem 3 we know that L1 is not generated by any prefixclosed VW-CCG without target restrictions. To argue that the system is terminating, we note that each rewriting step decreases the arity of one secondary input category in the derivation by one unit, while all other secondary input categories are left unchanged. As an example, consider rewriting under R1. The secondary input categories in the scope of that rule are Yβ/Z and Z on the left-hand side and Yβ and Z on the right-hand side. Here the arity of Yβ equals the arity of Yβ/Z, minus one. Because the system is terminating, to see that it is also confluent, it suffices to note that the left-hand sides of the rewrite rules do not overlap.
Lemma 8 The rewriting system transforms derivations of G into derivations of G. We prove the stronger result that every rewriting step transforms derivations of G into derivations of G. We only consider rewriting under R1; the arguments for the other rules are similar. Assume that R1 is applied to a derivation of G. The rule instances in the scope of the left-hand side of R1 take the following form:
X/Y Yβ/Z Xβ/Z (8) Xβ/Z Z Xβ (9)
Turning to the right-hand side, the rule instances in the rewritten derivation are
Yβ/Z Z Yβ (10) X/Y Yβ Xβ (11)
The relation between instances (8) and (11) is the characteristic relation of prefixclosed grammars (Definition 2): If instance (8) is valid, then because G is prefix-closed,
instance (11) is valid as well. Similarly, the relation between instances (9) and (10) is the characteristic relation of grammars without target restrictions (Definition 3): If instance (9) is valid, then because G is without target restrictions, instance (10) is valid as well. We conclude that if R1 is applied to a derivation of G, then the result is another derivation of G.
Combining Lemma 7 and Lemma 8, we see that for every derivation d of G, exhaustive application of the rewriting rules produces another uniquely determined derivation of G. We shall refer to this derivation as R(d). A transformed derivation is any derivation d′ such that d′ = R(d) for some derivation d. Let Y be the set of yields of the transformed derivations. Every string w′ ∈ Y is obtained from a string w ∈ L(G) by choosing some derivation d of w, rewriting this derivation into the transformed derivation R(d), and taking the yield. Inclusion then follows from Lemma 8. Because of the permuting rules R3 and R4, the strings w and w′ will in general be different. What we can say, however, is that w and w′ will be equal up to permutation. Thus we have established that Y and L(G) are Parikh-equivalent.
What remains in order to prove Lemma 6 is to show that the yields of the transformed derivations form a context-free language.
3.4.3 Context-Freeness of the Sublanguage. In a derivation tree, every node except the root node is labeled with either the primary or the secondary input category of a combinatory rule. We refer to these two types of nodes as primary nodes and secondary nodes, respectively. To simplify our presentation, we shall treat the root node as a secondary node. We restrict our attention to derivation trees for strings in L(G); in these trees, the root node is labeled with the distinguished atomic category S. For a leaf node u, the projection path of u is the path that starts at the parent of u and ends at the first secondary node that is encountered on the way towards the root node. We denote a projection path as a sequence X1, . . . , Xn (n ≥ 1), where X1 is the category at the parent of u and Xn is the category at the secondary node. Note that the category X1 is taken from the lexicon, while every other category is derived by combining the preceding category on the path with some secondary input category (not on the path) by means of some combinatory rule.
Example 10 In the derivation in Figure 6, the projection path of the first b goes all the way to the root, while all other projection paths have length 1, starting and ending with a lexical category. In Figure 7, the projection path of the first b ends at the root, while the projection paths of the remaining bs end at the nodes with category B, and the projection paths of the cs end at the nodes with category C.
A projection path X1, . . . , Xn is split if it can be segmented into two parts
X1, . . . , Xs and Xs, . . . , Xn (1 ≤ s ≤ n)
such that the first part only uses application rules and the second part only uses composition rules. Note that any part may consist of a single category only, in which case no combinatory rule is used in that part. If n = 1, then the path is trivially split. All projection paths in Figures 6 and 7 are split, except for the path of the first b in Figure 6, which alternates between composition (with C /A) and application (with A).
Lemma 10 In transformed derivations, every projection path is split. We show that as long as a derivation d contains a projection path that is not split, it can be rewritten. A projection path that is not split contains three adjacent categories U, V, W, such that V is derived by means of a composition with primary input U, and W is derived by means of an application with primary input V. Suppose that both the composition and the application are forward. (The arguments for the other three cases are similar.) Then U can be written as X/Y for some category X and argument /Y, V can be written as Xβ/Z for some argument /Z and some (possibly empty) sequence of arguments β, and W can be written as Xβ. We can then convince ourselves that d contains the following configuration, which matches the left-hand side of rewriting rule R1:
X/Y Yβ/Z Xβ/Z Z
Xβ
Lemma 11 The set of all categories that occur in transformed derivations is finite. Every category that occurs in transformed derivations occurs on some of its projection paths. Consider any such path. By Lemma 10 we know that this path is split; its two parts, here called P1 and P2, are visualized in Figure 13. We now reason about the arities of the categories in these two parts.
1. Because P1 only uses application, the arities in this part get smaller and smaller until they reach their minimum at Xs. This means that the arities of P1 are bounded by the arity of the first category on the path, which is a category from the lexicon.
2. Because P2 only uses composition, the arities in this part either get larger or stay the same until they reach a maximum at Xn. This means that the arities of P2 are bounded by the arity of the last category on the path, which is either the distinguished atomic category S or a secondary input category.
Thus the arities of our chosen path are bounded by the maximum of three grammarspecific constants: the maximal arity of a lexical category, the arity of S (which is 0), and the maximal arity of a secondary input category. The latter value is well-defined because there are only finitely many such categories (by Lemma 2). Let k be the maximum among the three constants, and let K be the set of all categories of the form A|mXm · · · |1X1 where A is an atomic category of G, m ≤ k, and each |iXi is an argument that may occur in derivations of G. The set K contains all categories that occur on some projection path, and therefore all categories that occur in transformed derivations, but it may also include other categories. As there are only finitely many atomic categories and finitely many arguments (Lemma 1), we conclude that the set K, and hence the set of categories that occur in transformed derivations, are finite as well.
Lemma 12 The transformed derivations yield a context-free language. We construct a context-free grammar H that generates the set Y of yields of the transformed derivations. To simplify the presentation, we first construct a grammar H′ that generates a superset of Y .
Construction of H′. The construction of the grammar H′ is the same as the construction in the classical proof that showed the context-freeness of AB-grammars, by Bar-Hillel, Gaifman, and Shamir (1960): The production rules of H′ are set up to correspond to the valid rule instances of G. The reason that this construction is not useful for VW-CCGs in general is that these may admit infinitely many rule instances, whereas a context-free grammar can only have finitely many productions. The set of rule instances may be infinite because VW-CCG has access to composition rules (specifically, rules of degrees greater than 1); in contrast, AB-grammars are restricted to application. Crucially though, by Lemma 11 we know that as long as we are interested only in transformed derivations it is sufficient to use a finite number of rule instances—more specifically those whose input and output categories are included in the set K of arity-bounded categories. Thus for every instance X/Y Yβ Xβ where all three categories are in K, we construct a production
[Xβ] → [X/Y] [Yβ]
and similarly for backward rules. (We enclose categories in square brackets for clarity.) In addition, for every lexicon entry σ := X in G we add to H′ a production [X] → σ. As the terminal alphabet of H′ we choose the vocabulary of G; as the nonterminal alphabet we choose the set K; and as the start symbol we choose the distinguished atomic category S. Every transformed derivation of G corresponds (in an obvious way) to some derivation in H′, which proves that Y ⊆ L(H′). Conversely, every derivation of H′ represents a derivation of G (though not necessarily a transformed derivation), thus L(H′) ⊆ L(G).
Construction of H. The chain of inclusions Y ⊆ L(H′) ⊆ L(G) is sufficient to prove Lemma 6: Because Y and L(G) are Parikh-equivalent (which we observed at the beginning of Section 3.4.2), so are L(H′) and L(G), which means that L(H′) satisfies all of the properties claimed in Lemma 6, even though this does not suffice to prove our current lemma. However, once H′ is given, it is not hard to also obtain a grammar H that generates exactly Y . For this, we need to filter out derivations whose projection paths do not have the characteristic property of transformed derivations that we established in Lemma 10. (It is not hard to see that every derivation that does have this property is a transformed derivation.) We annotate the left-hand side nonterminals in the productions of H′ with a flag t ∈ {a, c} to reflect whether the corresponding category has been derived by means of application (t = a) or composition (t = c); the value of this flag is simply the type of combinatory rule that gave rise to the production. The nonterminals in the right-hand sides are annotated in all possible ways, except that the following combinations are ruled out:
[X]a → [X/Y]c[Y]t and [X]a → [Y]t[X /Y]c t ∈ {a, c}
These combinations represent exactly the cases where the output category of a composition rule is used as the primary input category of an application rule, which are the cases that violate the “split” property that we established in Lemma 10.
This concludes the proof of Lemma 6, and therefore the proof of Theorem 3. Inclusion follows from the fact that every AB-grammar can be written as an O-CCG with only application (d = 0). To show that the inclusion is proper, we use the same argument as in the proof of Lemma 4. The grammar G2 that we constructed there can be turned into an equivalent O-CCG by decorating each slash with ·, the least restrictive type, and setting its inertness status to +.
What is less obvious is whether O-CCG generates the same class of languages as VW-CCG and TAG. Our main result is that this is not the case.
Theorem 4 The languages generated by O-CCG are properly included in the TAG languages.
O-CCG without Inertness. To approach Theorem 4, we set inertness aside for a moment and focus on the use of the slash types as a mechanism for imposing rule restrictions. Each of the rules in Figure 15 requires all of the slash types of the n outermost arguments of its secondary input category to be compatible with the rule, in the sense specified in Figure 14. If we now remove one or more of these arguments from a valid rule instance, then the new instance is clearly still valid, as we have reduced the number of potential violations of the type–rule compatibility. This shows that the rule system is prefix-closed. As none of the rules is conditioned on the target of the primary input category, the rule system is even without target restrictions. With these two properties established, Theorem 4 can be proved by literally the same arguments as those that we gave in Section 3. Thus we see directly that the theorem holds for versions of multimodal CCG without inertness, such as the formalism of Baldridge and Kruijff (2003).
O-CCG with Inertness. In the general case, the situation is complicated by the fact that the crossed composition rules change the inertness status of some argument categories if the slash types have conflicting directions. This means that the crossed composition rules in O-CCG are not entirely prefix-closed, as illustrated by the following example.
Example 11 Consider the following two rule instances:
X/+× Y Y / + ×Z2 / + × Z1 X / − ×Z2 / − × Z1 (12)
X/+× Y Y / + ×Z2 X / − ×Z2 (13)
Instance (12) is a valid instance of forward crossed composition. Prefix-closedness would require instance (13) to be valid as well; but it is not. In instance (12) the inertness status of /+ ×Z2 is changed for the only reason that the slash type of / + × Z1 does not match the required direction. In instance (13) the argument /+× Z1 is not present, and
therefore the inertness status of /+ ×Z2 is not changed, but is carried over to the output category.
We therefore have to prove that the following analogue of Lemma 6 holds for O-CCG:
Lemma 15 (Main Lemma for O-CCG) For every language L generated by some O-CCG there is a sublanguage L′ ⊆ L such that
1. L′ and L are Parikh-equivalent, and
2. L′ is context-free.
This lemma implies that the language L1 = {anbncn | n ≥ 1} from Example 2 cannot be generated by O-CCG (but by prefix-closed VW-CCG with target restrictions, and by TAG). The argument is the same as in the proof of Theorem 3. As in the proof of Lemma 8 we establish the stronger result that the claimed property holds for every single rewriting step. We only give the argument for rewriting under R3, which involves instances of forward crossed composition. The argument for R4 is analogous, and R1 and R2 are simpler cases because they involve harmonic composition, where the inertness status does not change.
Suppose that R3 is applied to a derivation of some O-CCG. In their most general form, the rule instances in the scope of the left-hand side of R3 may be written as follows, where the function σ is defined as specified in Figure 15:
X/+t0 Y Y / sn tn Zn · · · / s2 t2 Z2 / s1 t1 Z1 X / σ(sn) tn Zn · · · / σ(s2) t2 Z2 / σ(s1) t1 Z1 (14)
Z1 X / σ(sn) tn Zn · · · / σ(s2) t2 Z2 / + t1 Z1 X / σ(sn) tn Zn · · · / σ(s2) t1 Z2 (15)
Here instance (14) is an instance of forward-crossed composition, so each of the types ti is compatible with that rule. Because the two marked arguments are identical, we have σ(s1) = +. This is only possible if the inertness statuses of the slashes / si ti do not change in the context of derivation (14), that is, if σ(si) = si for all 1 ≤ i ≤ n. Note that in this case, t0 is either a right type or one of the four undirected core types, and each t1, . . . , tn is either a left type or a core type. We can now alternatively write instances (14) and (15) as
X/+t0 Y Y / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 X / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 (14 ′)
Z1 X / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 X / sn tn Zn · · · / s2 t1 Z2 (15 ′)
Then the rule instances in the rewritten derivation can be written as follows:
Z1 Y / sn tn Zn · · · / s2 t2 Z2 / + t1 Z1 Y / sn tn Zn · · · / s2 t2 Z2 (16)
X/+t0 Y Y / sn tn Zn · · · / s2 t2 Z2 X / sn tn Zn · · · / s2 t2 Z2 (17)
Here instance (16) is clearly a valid instance of backward application. Based on our earlier observations about the ti and their compatibility with crossed composition, we also see that instance (17) is a valid instance of forward crossed composition (if n > 1), or of forward application (if n = 1).
This completes the proof of Lemma 15. To finish the proof of Theorem 4 we have to also establish the inclusion of the O-CCG languages in the TAG languages. This is a known result for other dialects of multimodal CCG (Baldridge and Kruijff 2003), but O-CCG once again requires some extra work because of inertness.
Lemma 17 The O-CCG languages are included in the TAG languages. It suffices to show that the O-CCG languages are included in the class of languages generated by LIG (Gazdar 1987); the claim then follows from the weak equivalence of LIG and TAG. Vijay-Shanker and Weir (1994, Section 3.1) present a construction that transforms an arbitrary VW-CCG into a weakly equivalent LIG. It is straightforward to adapt their construction to O-CCG. As we do not have the space here to define LIG, we only provide a sketch of the adapted construction.
As in the case of VW-CCG, the valid instances of an O-CCG rule can be written down using our template notation. The adapted construction converts each such template into a production rule of a weakly equivalent LIG. Consider for instance the following instance of forward crossed composition from Example 11:
A /+× Y Y / + ×Z2 / + × Z1 A / − ×Z2 / − × Z1
This template is converted into the following LIG rule. We adopt the notation of VijayShanker and Weir (1994) and write ◦◦ for the tail of a stack of unbounded size.
A[◦◦ /− ×Z2 /−× Z1] → A[◦◦/+× Y] Y[ /+ ×Z2 /+× Z1]
In this way, every O-CCG can be written as a weakly equivalent LIG.","5 conclusion :In this article we have contributed two technical results to the literature on CCG. First, we have refined the weak equivalence result for CCG and TAG (Vijay-Shanker and Weir 1994) by showing that prefix-closed grammars are weakly equivalent to TAG only if target restrictions are allowed. Second, we have shown that O-CCG, the formal, composition-only core of OpenCCG, is not weakly equivalent to TAG. These results point to a tension in CCG between lexicalization and generative capacity: Lexicalized versions of the framework are less powerful than classical versions, which allow rule restrictions.
What conclusions one draws from these technical results depends on the perspective. One way to look at CCG is as a system for defining formal languages. Under this view, one is primarily interested in results on generative capacity and parsing complexity such as those obtained by Vijay-Shanker and Weir (1993, 1994). Here, our results clarify the precise mechanisms that make CCG weakly equivalent to TAG. Perhaps surprisingly, it is not the availability of generalized composition rules by itself that explains the generative power of CCG, but the ability to constrain the interaction between generalized composition and function application by means of target restrictions.
On the other hand, one may be interested in CCG primarily as a formalism for developing grammars for natural languages (Steedman 2000; Baldridge 2002; Steedman 2012). From this point of view, the suitability of CCG for the development of lexicalized grammars has been amply demonstrated. However, our technical results still serve as important reminders that extra care must be taken to avoid overgeneration when designing a grammar. In particular, it is worth double-checking that an OpenCCG grammar does not generate word orders that the grammar developer did not intend. Here the rewriting system that we presented in Figure 12 can serve as a useful tool: A grammar developer can take any derivation for a grammatical sentence, transform the derivation according to our rewriting rules, and check whether the transformed derivation still yields a grammatical sentence.
It remains an open question how the conflicting desires for generative capacity and lexicalization might be reconciled. A simple answer is to add some lexicalized method for enforcing target restrictions to CCG, specifically on the application rules. However, we are not aware that this idea has seen widespread use in the CCG literature, so it may not be called for empirically. Alternatively, one might modify the rules of O-CCG in such a way that they are no longer prefix-closed—for example, by introducing some new slash type. Finally, it is possible that the constructs of OpenCCG that we set aside in O-CCG (such as type-raising, substitution, and multiset slashes) might be sufficient to achieve the generative capacity of classical CCG and TAG. A detailed study of the expressive power of these constructs would make an interesting avenue for future research."
"1 introduction :A growing body of work in computational linguistics (CL hereafter) or natural language processing manifests an interest in corpus studies, and requires reference annotations for system evaluation or machine learning purposes. The question is how to ensure that an annotation can be considered, if not as the “truth,” than at least as a suitable
∗ Normandie University, France; UNICAEN, GREYC, F-14032 Caen, France; CNRS, UMR 6072, F-14032 Caen, France. E-mail: {yann.mathet, antoine.widlocher, jean-philippe.metivier}@unicaen.fr
Submission received: 15 July 2013; revised version received: 5 October 2014; accepted for publication: 12 January 2015.
doi:10.1162/COLI a 00227
© 2015 Association for Computational Linguistics
reference. For some simple and systematic tasks, domain experts may be able to annotate texts with almost total confidence, but this is generally not the case when no expert is available, or when the tasks become harder. The very notion of “truth” may even be utopian when the annotation process includes a certain degree of interpretation, and we should in such cases look for a consensus, also called the “gold standard,” rather than for the “truth.”
For these reasons, a classic strategy for building annotated corpora with sufficient confidence is to give the same annotation task to several annotators, and to analyze to what extent they agree in order to assess the reliability of their annotations. This is the aim of inter-annotator agreement measures. It is important to point out that most of these measures do not evaluate the distance from annotations to the “truth,” but rather the distance across annotators. Of course, the hope is that the annotators will agree as far as possible, and it is usually considered that a good inter-annotator agreement ensures the constancy and the reproducibility of the annotations: When agreement is high, then the task is consistent and correctly defined, and the annotators can be expected to agree on another part of the corpus, or at another time, and their annotations therefore constitute a consensual reference (even if, as shown for example by Reidsma and Carletta [2008], such an agreement is not necessarily informative for machine learning purposes). Moreover, once several annotators reach good agreement on a given part of a corpus, then each of them can annotate alone other parts of the corpus with great confidence in the reproducibility (see the preface to Gwet [2012, page 6] for illuminating considerations). Consequently, inter-annotator agreement measurement is an important point for all annotation efforts because it is often considered that a given agreement value provided by a given method validates or invalidates the consistency of an annotation effort.
How to measure agreement, and how we define a good measure, is another part of the problem. There is no universal answer, because how to measure depends on the nature of the task, hence on the kind of annotations.
Admittedly, much work has already been done for some kinds of annotation efforts, namely, when annotators have to choose a category for previously identified entities. This approach, which we will call pure categorization, has led to several well-known and widely discussed coefficients such as κ, π, or α, since the 1950s. Some more recent efforts have been made in the domain of unitizing, following Krippendorff’s terminology (Krippendorff 2013), where annotators have to identify by themselves what the elements to be annotated in a text are, and where they are located. Studies are scarce, however, as Krippendorff pointed out: “Measuring the reliability of unitizing has been largely ignored in favor of coding predefined units” (Krippendorff 2013, page 310). This scarcity concerns either segmentation, where annotators simply have to mark boundaries in texts to separate contiguous segments, or more generally unitizing, where gaps may exist between units. Moreover, some even more complex configurations may occur (overlapping or embedding units), which are more rarely taken into account.
And when categorization meets unitizing, as is the case in CL in such fields as, for example, NAMED ENTITY RECOGNITION1 or DISCOURSE FRAMING, very few methods are proposed and discussed. That is the main problem we focus on in this article and to which γ provides solutions.
1 Small caps are used to refer to the examples of annotation tasks introduced in Section 2.2.
The new coefficient γ that is introduced in this article is an agreement measure concerning the joint tasks of unit locating (unitizing) and unit labeling (categorization). It relies on an alignment of units between different annotators, with penalties associated with each positional and categorial discrepancy. The alignment itself is chosen to minimize the overall discrepancy in a holistic way, considering the full continuum to make choices, rather than making local choices. The proposed method is unified because the computation of γ and the selection of the best alignment are interdependent: The computed measure depends on the chosen alignment, whose selection depends on the measure.
This method and the principles proposed in this article have been built up since 2010, and were first presented to the French community in a very early version in Mathet and Widlöcher (2011). The initial motivation for their development was the lack of dedicated agreement measures for annotations at the discourse level, and more specifically for annotation tasks related to TOPIC TRANSITION phenomena.
The article is organized as follows. First, we fix the scope of this work by defining the important notions that are necessary to characterize annotation tasks and by introducing the examples of linguistic objects and annotation tasks used in this article to compare available metrics. Second, we analyze the state of the art and identify the weaknesses of current methods. Then, we introduce our method, called γ. As this method is new, we compare it to the ones already in use, even in their specialized fields (pure categorization, or pure segmentation), and show that it has better properties overall for CL purposes.","2 motivations, scope, and illustrations :We focus in the present work on both categorizing and unitizing, and consider therefore annotation tasks where annotators are not provided with preselected units, but have to locate them and to categorize them at the same time. An example of a multi-annotated continuum (this continuum may be a text or, for example, an audio or a video recording) is provided in Figure 1, where each line represents the annotations of a given annotator, from left to right, respecting the continuum order. In order to characterize the annotation efforts focusing on specific linguistic objects, we consider the following properties, illustrated in Figure 1.
Categorization occurs when the annotator is required to label (predefined or not) units. Unitizing occurs when the annotator is asked to identify the units in the continuum:
She has to determine each of them (and the number of units that she wants) and to locate them by positioning their boundaries.
Embedding (hierarchical overlap) may occur if units may be embedded in larger ones (of the same type, or not). Free overlap may occur when guidelines tolerate the partial overlap of elements (mainly of different types). Embedding is a special case of overlapping. A segmentation without overlap (hierarchical or free) is said to be strictly linear. Full-covering (vs. sporadicity) applies when all parts of the continuum are to be annotated. For other tasks, parts of the continuum are selected. Aggregatable types or instances correspond to the fact that several adjacent elements having the same type may aggregate, without shift in meaning, in a larger span having the same type. This larger span is said to be atomizable: Labeling the whole span or labeling all of its atoms are considered as equivalent, as illustrated by Figure 2.
Two specific cases. We call hereafter pure segmentation (illustrated by Figure 3) the special case of unitizing with full-covering and without categorization, and we call pure categorization categorization without unitizing. To present the state of the art as well as our own propositions, and to make all of them more concrete, it is useful to mention examples of linguistic objects and annotation tasks for which agreement measures may be required. The following sections will then refer to these examples as often as possible, in order to illustrate discussions on abstract problems or configurations. Small caps are used to refer to the names of these tasks.
A
Aggregatable atoms
Atomizable unit
Table 1 summarizes the properties of the linguistic objects and annotation tasks mentioned in this article to illustrate and compare methods and metrics. These objects and tasks are briefly described for convenience in Appendix A. This table shows that annotation of TOPIC TRANSITIONS is the most demanding of the tasks regarding the number of necessary criteria a suitable agreement metric should assess, but most of the tasks listed here require assessment of both unitizing and categorization.","3 state of the art :As we saw in the previous section, different studies in linguistics or CL involve quite different structures, which may lead to annotation guidelines having very different properties. They require suitable metrics in order to assess agreement among annotators. As we will see, some of the needs for which γ is suitable are not satisfied by other available metrics.
Note that this description of the state of the art mainly focuses on the questions which are of most importance for this work, in particular, chance correction and unitizing. For a thorough introduction to the most popular measures that concern categorizing, we refer the reader to the excellent survey by Artstein and Poesio (2008).
In this section, we first address the question of chance correction in agreement measures, then we give an overview of available measures in three domains: pure categorization, pure segmentation, and unitizing. We begin the state of the art with the question of chance correction, because it is a crosscutting issue in all agreement measure domains, and because it influences the final value provided by most agreement measures.
It is important to distinguish between (1) measures to evaluate systems, where the output of an annotating system is compared to a valid reference, and (2) inter-annotator agreement measures, which try to quantify the degree of similarity between what different
annotators say about the same data, and which are the ones we are really concerned with in this article.
In case (1), the result is straightforward, providing for instance the percentage of valid answers of a system: We know exactly how far the evaluated system is from the gold standard, and we can compare this system to others just by comparing their results.
However, case (2) is more difficult. Here, measures do not compare annotations from one annotator to a valid reference (and, most of the time, no reference already exists), but they compare annotations from different annotators. As such, they are clearly not direct distances to the “truth.” So, the question is: Above what amount of agreement can we reasonably trust the annotators? The answer is not straightforward, and this is where chance correction is involved.
For instance, consider a task where two annotators have to label items with 10 categories. If they annotate at random (with the 10 categories having equal prevalence), they will have an agreement of 10%. If we consider another task involving two categories only, still at random, the agreement expected by chance rises to 50%. Based on this observation, most agreement measures try to remove chance from the observed measure, that is to say, to provide the amount of agreement that is above chance. More precisely, most agreement measures (for about 60 years, with well-known measures κ, S, π) rely on the same formula: If we note Ao the observed agreement (i.e., the agreement directly observed between annotators) and Ae the so-called expected agreement (i.e. the agreement which should be obtained by chance), the final agreement A is defined by Equation (1).
To illustrate this formula, assume that the observed agreement is seemingly high, say Ao = 0.9. If Ae = 0.5, A = 0.4/0.5 = 0.8, which is still considered as good, but if Ae = 0.7, A = 0.2/0.3 = 0.67, which is not that good, and if Ae = 0.9, which means annotators did not perform better than chance, then A = 0.
Some other measures, namely, all α from Krippendorff, and the new γ introduced in this article, are computed from observed and expected disagreements (instead of agreements), denoted here respectively Do and De, and they define the final agreement by Equation (2).
A = Ao − Ae1− Ae (1)
A = 1− DoDe (2)
However, the way the expected value is computed is the only difference between many coefficients (κ, S, π, and their generalizations), and is a controversial question. As precisely described in Artstein and Poesio (2008), there are three main ways to model chance in an annotation effort:
1. By considering a uniform distribution. For instance, in a categorization task, considering that each category (for each coder) has the same probability. The limitation of this approach is that it provides a poor model for chance annotation. Moreover, for a given task, the greater the number of categories, the lesser the expected value, hence the higher the final agreement.
2. By considering the mean distribution of the different annotators hence regarded as interchangeable. For instance, in a categorization task with two categories A and B, where the prevalences are respectively 90% for category A and 10% for category B, the expected value is computed as 0.9× 0.9 + 0.1× 0.1 = 0.82, which is much higher than the 0.5 obtained by considering a uniform distribution.
3. By considering the individual distributions of annotators. Here, annotators are considered as not interchangeable; each of them is considered to have her own probability for each category (for a categorization task) based on her own observed distribution. It leads to the same results as with the mean distribution if annotators all have the same distribution, or to a lesser value (hence a higher final agreement) if not.
In the two cases of mean and individual distributions, expected agreement may be very high, depending on the prevalence of categories. In some annotation tasks, expected agreement becomes critically high, and any disagreements on the minor category have huge consequences on the chance-corrected agreement, as hotly debated by Berry (1992) and Goldman (1992), and criticized in CL by Di Eugenio and Glass (2004). However, we follow Krippendorff (2013, page 320), who argues that disagreements on rare categories are more serious than on frequent ones. For instance, let us consider the reliability of medical diagnostics concerning a rare disease that affects one person out of 1,000. There are 5,000 patients, 4,995 being healthy, 5 being affected. If doctors fail to agree on the 5 affected patients, their diagnostics cannot be trusted, even if they agree on the 4,995 healthy ones.
These principles have been mainly introduced and used for categorization tasks, because most coefficients address these tasks, but they are more general and may also concern segmentation and, as we will see further, unitizing. The simplest measure of agreement for categorization is percentage of agreement (see for example Scott 1955, page 323). Because it does not feature chance correction, it should be used carefully for the reasons we have just seen.
Consequently, most popular measures are chance-corrected: S (Bennett, Alpert, and Goldstein 1954) relies on a uniform distribution model of chance, π (Scott 1955) and α (Krippendorff 1980) on the mean distribution, and κ (Cohen 1960) on individual distributions. Generalizations to three or more annotators have been provided, such as κ (Fleiss 1971), also known as K (Siegel and Castellan 1988). Moreover, weighted coefficients such as α and κw (Cohen 1968) are designed to take into account the fact that disagreements between two categories are not necessarily all of the same importance. For instance, for scaled categories from 1 to 10 (as opposed to so-called nominal categories), a mistake between categories 3 and 4 should be less penalized than a mistake between categories 1 and 10.
These metrics, widely used in CL, are suitable to assess agreement for pure categorization tasks—for example, in the domains of PART-OF-SPEECH TAGGING, GENE RENAMING, or WORD SENSE ANNOTATION.
From Carletta (1996) to Artstein and Poesio (2008), most of these methods have already been discussed and compared in the perspective of CL and we will not do so here. In the domain of TOPIC SEGMENTATION, several measures have been proposed, especially to evaluate the quality of automatic segmentation systems. In most cases, this evaluation consists in comparing the output of these systems with a reference annotation. We mention them here because their use tends to be extended to interannotator agreement because of the lack of dedicated agreement measures, as illustrated by Artstein and Poesio (2008), who mention these metrics in a survey related to interannotator agreement, or by Kazantseva and Szpakowicz (2012).
In this domain, annotations consist of boundaries (between topic segments), and the penalty must depend on the distance from a true boundary. Thus, dedicated measures have been proposed, such as WindowDiff (WD hereafter; Pevzner and Hearst 2002), based on Pk (Beeferman, Berger, and Lafferty 1997). WD relies on the following principle: A fixed-sized window slides over the text and the numbers of boundaries in the system output and reference are compared. Several limitations of this method have been demonstrated and adjustments proposed, for example, by Lamprier et al. (2007) or by Bestgen (2009), who recommends the use of the Generalized Hamming Distance (GHD hereafter; Bookstein, Kulyukin, and Raita 2002), in order to improve the stability of the measure, especially when the variance of segment size increases.
Because these metrics are dedicated to the evaluation of automatic segmentation systems, their most serious weakness for assessing agreement is that they are not chance-corrected, but they present another limitation: They are dedicated to segmentation and assume a full-covering and linear tiling of the continuum and only one category of objects (topic segments). This strong constraint makes them unsuitable for unitizing tasks using several categories (ARGUMENTATIVE ZONING), targeting more sporadic phenomena (ANNOTATION OF COMMUNICATIVE BEHAVIOR), or involving more complex structures (NAMED ENTITY RECOGNITION, HIERARCHICAL TOPIC SEGMENTATION, TOPIC TRANSITION, DISCOURSE FRAMING, ENUMERATIVE STRUCTURES). 3.4.1 Using Measures for Categorization to Measure Agreement on Unitizing. Because of the lack of dedicated measures, some attempts have been made to transform the task of unitizing into a task of categorizing in order to use well-known coefficients such as κ.
They consist of atomizing the continuum by considering each segment as a sequence of atoms, thereby reducing a unitizing problem to a categorization problem. This is illustrated by Figure 4, where real unitizing annotations are on the left (with two annotators), and the transformed annotations are on the right. To do so, an atom granularity is chosen—for instance, in the case of texts, it may be character, word, sentence, or paragraph atoms. Then, each unit is transformed into a set of items labeled with the category of this unit, and a new “blank” category is added in order to emulate gaps between units.
In most cases, this method has severe limitations:
1. Two contiguous units seen as one. In zone (1) of the left part of Figure 4, one annotator has created two units (of the same category), and the other annotator has created only one unit covering the same space. However, once the continua are discretized, the two annotators seem to agree on this zone (with the four same atoms), as we can see in the right part of the figure.
2. False positive/negative disagreement and slight positional disagreement considered as the same. Zone (2) of Figure 4 shows a case where annotators disagree on whether there is a unit or not, which is quite a severe disagreement, whereas zone (3) shows a case of a slight positional disagreement. Surprisingly, these two discrepancies are counted with the same severity, as we can see in the right side of the figure, because in each case, there is a difference of category for one item only (respectively, “blank” with “blue” in case (2), and “blue” with “blank” in case (3)).
3. Agreement on gaps. Because of the discretization, artificial blank items are created, with the result that annotators may agree on “blanks.” The more gaps in real annotations, the more artificial “blank” agreement, and hence the greater the artificial increase in global agreement. Indeed, the expected agreement is less impacted by artificial “blanks,” and it may even decrease.
4. Overlapping and embedding units are not possible. This results because of the discretizing process, which requires a given position to be assigned a single category (or it would require creating as many artificial categories as possible combinations of categories).
This kind of reduction is used to evaluate the annotation of COMMUNICATIVE BEHAVIOR in video recordings by Reidsma (2008), where unitizing is, on the contrary, clearly required: The time-line is discretized (atomized), then κ and α are computed using discretized time spans as units. It should be noted that Reidsman, Heylen, and Ordelman (2006, page 1119) and Reidsma (2008) claim that this “fairly standard” method (which we call discretizing measure henceforth) has certain drawbacks, such as the fact that it “does not compensate for differences in length of segments,” whereas “short segments are as important as long segments” in their corpus (which is an additional limitation to the ones we have just mentioned). They propose a second approach relying on an alignment, as we mention in Section 4.2.1.
This reduction is also unacceptable for other annotation tasks. For instance, in the perspective of DISCOURSE FRAMING, two adjacent temporal frames should not be aggregated in a larger one. In the same manner, for TOPIC SEGMENTATION, it clearly makes no sense to aggregate two consecutive segments.
3.4.2 A Measure for Unitizing Without Chance Correction. Another approach, derived from Slot Error Rate (Makhoul et al. 1999), presented in Galibert et al. (2010), and called SER below, was more specifically used in the context of evaluation of NAMED ENTITY recognition systems.
Comparing a “hypothesis” to a reference, this metric counts the costs of different error types: error “T” on type (i.e., category) with cost 0.5, error “B” on boundaries
(i.e., position) with cost 0.5, error “TB” on both type and boundaries with cost 1, error “I” of insertion (i.e., false positive) with cost 1, and error “D” of deletion (false negative) with cost 1. The overall cost relies on an alignment of objects from reference and hypothesis, which is chosen to minimize this cost. The final value provided by SER is the average cost of the aligned pairs of units—0 meaning perfect agreement, 1 roughly meaning systematic disagreement. An example is given in Figure 5.
This attempt to extend Slot Error Rate to unitizing suffers from severe limitations. In particular, all positioning and categorizing errors have the same penalty, which may be a serious drawback for annotation tasks where some fuzziness in boundary positions is allowed, such as TOPIC SEGMENTATION, TOPIC TRANSITION, or DISCOURSE FRAMING. Moreover, it is difficult to interpret because its output is surprisingly not upper bounded by 1 (in the case where there are many false positives). Additionally, it was initially designed to compare an output to a reference, and so requires some adjustments to cope with more than 2 annotators. Last but not least, it is not chance corrected.
3.4.3 Specific Measures for Unitizing. To our knowledge, the family of α measures proposed by Krippendorff is by far the broadest attempt to provide suitable metrics for various annotation tasks, involving both categorization and unitizing.
In the survey by Artstein and Poesio (2008, page 581), some hope of finding an answer to unitizing is formulated as follows: “We suspect that the methods proposed by Krippendorff (1995) for measuring agreement on unitizing may be appropriate for the purpose of measuring agreement on discourse segmentation.” Unfortunately, as far as we know, its usage in CL is rare, despite the fact that it is the first coefficient that copes both with unitizing and categorizing at the same time, while taking chance into account. The family of α measures would then be suitable for annotation tasks related, for example, to COMMUNICATIVE BEHAVIOR or DIALOG ACTS.
We will therefore pay special attention to Krippendorff’s work in this article, because it constitutes a very interesting reference to compare with, both in terms of theoretical choices and of results. Let us briefly recap Krippendorff’s studies on unitizing from 1995 to 2013 and introduce some of the α measures, which will be discussed in this article. The α coefficient (Krippendorff 1980, 2004, 2013), dedicated to agreement measures on categorization tasks, generalizes several other broadly used statistics and allows various categorization values (nominal, ordinal, ratio, etc.). Besides this wellknown α measure, which copes with categorizing, a new coefficient called αU has been proposed since 1995 in Krippendorff (1995) and then Krippendorff (2004), which can apply to unitizing. Recently, Krippendorff (2013, pages 310, 315) proposed a new version of this coefficient, called uα, “with major simplifications and improvements over
previous proposals,” and which is meant to “assess the reliability of distinctions within a continuum—how well units and gaps coincide and whether units are of the same or of a different kind.” To supplement uα, which mainly focuses on positioning, Krippendorff has proposed c|uα (Krippendorff 2013), which ignores positioning disagreement and focuses mainly on categories.
These measures will be discussed in the following sections. For now, it must be noted that uα and c|uα are not currently designed to cope with embedding or free overlapping between the units of the same annotator. These metrics are then unsuitable for annotation tasks such as, for instance, TOPIC TRANSITION, HIERARCHICAL TOPIC SEGMENTATION, DISCOURSE FRAMING, or ENUMERATIVE STRUCTURES. To conclude the state of the art, we draw up a final overview of the coverage of the requirements by the different measures in Table 2. The γ measure, introduced in the next section, aims at satisfying all these needs.","4 the proposed method: introducing γ : The basic idea of this new coefficient is as follows: All local disagreements (called disorders) between units from different annotators are averaged to compute an overall disorder. However, these local disorders can be computed only if we know for each unit of a given annotator, which units, if any, from the other annotators it should be compared with (via what is called unitary alignment)—that is to say, if we can rely on a suitable alignment of the whole (called alignment). Because it is not possible to get a reliable preconceived alignment (as explained in Section 4.2.1), γ considers all
possible ones, and computes for each of them the associated overall disorder. Then, γ retains as the best alignment the one that minimizes the overall disorder, and the latter value is retained as the correct disorder. To obtain the final agreement, as with the familiar kappa and alpha coefficients, this disorder is then chance-corrected by a so-called expected disorder, which is calculated by randomly resampling existing annotations.
First of all, we introduce three main principles of γ in Section 4.2. We introduce in Section 4.3 the basic definitions. The comparison of two units (depending on their relative positions and categories) relies on the concept of dissimilarity (Section 4.4). A unitary alignment groups at most one unit of each annotator, and a set of unitary alignments covering all units of all annotators is called an alignment (Section 4.5). The disorder associated with a unitary alignment results from dissimilarities between all its pairs of units, and the disorder associated with an alignment depends on those of its unitary alignments (Section 4.6). The alignment having the minimal disorder (Section 4.7) is used to compute the agreement value, taking chance correction into account (Section 4.8). 4.2.1 Measuring and Aligning at the Same Time: γ is Unified. For a given phenomenon identified by several annotators, it is necessary to provide an agreement measure permissive enough to cope with a double discrepancy concerning its position in the continuum, and the category attributed to the phenomenon.
Because of discrepancy in positioning, it is necessary to provide an agreement measure with an inter-annotator alignment, which shows which unit of a given annotator corresponds, if any, to which unit of another annotator. If such an alignment is provided, it becomes possible, for each phenomenon identified by annotators, to determine to what extent the annotators agree both on its categorization and its positioning. This quantification relies on a certain measure (called dissimilarity hereafter) between annotated units: The more the units are considered as similar, the lesser the dissimilarity.
But how can such an alignment be achieved? For instance, in Figure 6, aligning unit A1 of annotator A with unit B1 of annotator B consists in considering that their properties are similar enough to be associated: annotator A and annotator B have accounted for the same phenomenon, even if in a slightly different manner. Consequently, to operate, the alignment method should rely on a measure of distance (in location, in category assignment, or both) between units.
Therefore, agreement measure and aligning are interdependent: It is not possible to correctly measure without aligning, and it is not possible to align units without measuring their distances. In that respect, measuring and aligning cannot constitute two successive stages, but must be considered as a whole process. This interdependence
reflects the unity of the objective: Establishing to what extent some elements, possibly different, may be considered as similar enough either to quantify their differences (when measuring agreement), or to associate them (when aligning).
Interestingly, Reidsma, Heylen, and Ordelman (2006, page 1119), not really satisfied by the use of the discretizing measure as already mentioned, “have developed an extra method of comparison in which [they] try to align the various segments.” This attempt highlights the necessity to rely on an alignment. Unfortunately, the way the alignment is computed, adapted from Kuper et al. (2003), is disconnected from the measure itself, being an ad hoc procedure to which other measures are applied.
4.2.2 Aligning Globally: γ is Holistic. Let us consider two annotators A and B having respectively produced unit A5, and units B4 and B5, as shown in Figure 7. When considering this configuration at a local level, we may consider, based on the overlapping area for instance, that A5 fits B5 slightly better than B4. However, this local consideration may be misleading. Indeed, Figure 8 shows two larger configurations, where A5, B4, and B5 are unchanged from Figure 7. With a larger view, the choice of alignment of A5 may be driven by the whole configuration, possibly leading to an alignment with B4 in Figure 8a, and with B5 in Figure 8b: Alignment choices depend on the whole system and the method should consequently be holistic.
4.2.3 Accounting for Different Severity Rates of Errors: Positional and Categorial Permissiveness of γ. As far as positional discrepancies between annotators are concerned, it is important for a measure to rely on a progressive error count, not on a binary one: Two positions from two annotators may be more or less close to each other but still concern the same phenomenon (partial agreement), or may be too far to be considered as related to the same phenomenon (no possible alignment). For instance, for segmentation, specific measures such as GHD or WD rely on a progressive error count for positions, with an upper limit being half the average size of the segments. For unitizing, Krippendorff
considers with uα that units can be compared as long as they overlap. However, γ considers that in some cases, units by different annotators may correspond to the same phenomenon though they do not intersect. We base this claim on two grounds. First, if we observe the configuration given in Figure 9, annotators 2 and 3 have both annotated part of the NAMED ENTITY that has been annotated by annotator 1. Consequently, though they do not overlap, their units refer to the same phenomenon. In addition, we find a direct echo of this assumption in Reidsma (2008, pages 16–17) where, in a video corpus concerning COMMUNICATIVE BEHAVIOR, “different timing (non-overlapping) [of the same episode] was assigned by [...] two annotators.” Regarding categorization, some available measures consider all disagreements between all pairs of categories as equal. Other coefficients, called weighted coefficients (see Artstein and Poesio 2008), as well as γ, consider on the contrary that mismatches may not all have the same weight, some pairs of categories being closer than others. This closeness is often referred to as overlap.
In our terminology, we call category-overlapping this closeness between categories, and overlap means positional overlap. For example, within annotation efforts related to WORD SENSE or DIALOG ACTS, it is clear that disagreements on labels are not all alike. Given a multi-annotated continuum t:r let A = {a1, ..., an} be the set of annotatorsr let n = |A| be the number of annotatorsr let U be the set of units from all annotatorsr ∀i ∈ J1, nK, let xi be the number of units by annotator ai for t r let x = n∑i=1 xin be the average number of annotations per annotatorr for annotator a = ai, ∀j ∈ J1, xiK, we note uaj unit from a of rank j Annotation set: An annotation set s is a set of units attached to the same continuum and produced by a given set of annotators. Corpus: A corpus c is defined with respect to a given annotation effort, and is composed of a set of continua, and of the set of annotations related to these continua.
Unit: A unit u bears a category denoted cat(u), and a location given by its two boundaries, each of them corresponding to a position in the continuum, respectively denoted start(u) and end(u), start and end being functions from U to N+.
Equality between units is defined as follows:
∀(u, v) ∈ U2, u = v⇔  cat(u) = cat(v)start(u) = start(v)end(u) = end(v) We introduce here the first brick to build the notion of disorder, which works at a very local level, between two units. A dissimilarity tells to what degree two units should be considered as different, taking into account such features as their positions, their categories, or a combination of the two.
A dissimilarity is a function d : U2 → R+, so that :
∀(u, v) ∈ U2, {
d(u, v) = d(v, u) (d is symmetric) u = v⇒ d(u, v) = 0
A dissimilarity is not necessarily a distance in the mathematical sense of the term, in particular because triangular inequality is not mandatory (for instance, in Figure 10, d(A1, B2) > d(A1, C1) + d(C1, B2)).
4.4.1 Empty Unit u∅, Empty Dissimilarity ∆∅. As we will see, γ relies on an alignment of units by different annotators. In particular, this alignment indicates for unit ua1i of annotator a1, to which unit u a2 j of annotator a2 it corresponds, in order to compute the associated dissimilarity. In some cases, though, the method will choose not to align ua1i with any unit of annotator a2 (none corresponds sufficiently). We define the empty pseudo unit, denoted u∅, which corresponds to the realization of this phenomenon: ultimately, a pseudo unit u∅ is added to the annotations of a2, and u a1 i is aligned with it.
We also define the associated cost ∆∅:
∀u ∈ U , d(u, u∅) = d(u∅, u) = ∆∅ and d(u∅, u∅) = ∆∅
Dissimilarities should be calibrated so that ∆∅ is the value beyond which two compared units are considered critically different. Consequently, it constitutes a reference, and dissimilarities will be expressed in this article as multiples of ∆∅ for better clarity. It is not a parameter of gamma, but a constant (which is set to 1 in our implementation).
4.4.2 Positional Dissimilarity dpos. Different positional dissimilarities may be created, in order to deal with different annotation tasks. In this article, we use the dissimilarity shown in Equation 3, which is very versatile.
dpos−sporadic(u, v) = ( |start(u)− start(v)|+ |end(u)− end(v)| (end(u)− start(u)) + (end(v)− start(v)) )2 ·∆∅ (3)
Equation (3) sums the differences between the right and left boundaries of both units in its numerator. Its denominator sums the lengths of both units, so that this dissimilarity is not scale-dependent. Squaring the value is an option used here to accelerate dissimilarity when differences of positions increase. It is illustrated in Figure 10 with different configurations and their associated values, from 0 for the perfectly aligned pair of units (A1, B1) to 22.2 ·∆∅ for the worst pair (A1, C2).
4.4.3 Categorial Dissimilarity dcat. Let K be the set of categories. For a given annotation effort, |K| different categories are defined. For more convenience, we first define categorial distance between categories distcat via a square matrix of size |K|, with each category appearing both in row titles and column titles. Each cell gives the distance between two categories through a value in [0, 1]. Value 0 means perfect equality, whereas the maximum value 1 means that the categories are considered as totally different. As distcat is symmetric, such a matrix is necessarily symmetric, and bears 0 in each diagonal cell. Table 3 gives an example for three categories, and shows that an association between a unit in category cat1 with one in category cat3 is the worst possible (distance = 1), whereas it is half as much between cat1 with cat2 (distance = 0.5). This makes it possible to take into account so-called category-overlapping (in our example, cat1 and cat2 are said to overlap, which means they are not completely different), as weighted coefficients such as κw or α already do. Note that in the case of so-called “nominal categories,” the matrix will be full of 1 outside the diagonal, and full of 0 in the diagonal (different categories are considered as not matching at all).
This categorial distance matrix is then used to build the categorial dissimilarities, taking into account the ∆∅ value. We define categorial dissimilarity between two units by:
dcat(u, v) = fcat(distcat (cat(u), cat(v))) ·∆∅ (4)
Function fcat can be used to adjust the way the dissimilarity grows with respect to the categorial distance values. The standard option2 (used in this article) is to simply consider fcat(x) = x, with which dcat naturally increases gradually from zero when categories match, to ∆∅ when categories are totally different (distcat(cat(u), cat(v)) = 1 =⇒ dcat(u, v) = ∆∅).
4.4.4 Combined Dissimilarity dcombi. Because in some annotation tasks units may differ both in position and in category, it is necessary to combine the associated dissimilarities so that all costs are cumulated. This is provided by a combined dissimilarity.
Let d1 and d2 be two dissimilarities. We define:
dα,βcombi(d1,d2 )(u, v) = α.d1(u, v) + β.d2(u, v) (5)
It is easy to demonstrate that this linear combination of dissimilarities is itself a dissimilarity (if (α,β) 6= (0, 0)). It enables the same weight to be assigned to positions and categories using d1,1combi(dpos,dcat ), which is currently used for γ.
Then, we can note that it is the same cost ∆∅ for a unit either not to be aligned with any other one, or to be aligned with a unit in the same configuration as (A1, C1) of Figure 10 (if they have the same category), or to be aligned with a unit having an incompatible category (if they occupy the same position). Unitary alignment ă. A unitary alignment ă is an i-tuple, i belonging to J1, nK (n the number of annotators), containing at most one unit by each annotator: It represents the hypothesis that i annotators agree to some extent on a given phenomenon to be unitized. In order to make all unitary alignments homogenous, we eventually complete any unitary alignment that is an i-tuple with n− i empty units u∅, so that all unitary alignments are ultimately n-tuples. Figure 11 illustrates unitary alignments with some u∅ units.
Alignment ā. For a given annotation set, an alignment ā is defined as a set of unitary alignments such that each unit of each annotator belongs to one and only one of its
2 Another option is, for example, to use fcat(x) = −ln(1− x)x30 + x, which is a function almost similar to fcat(x) = x on the [0, 0.9] range, and reaches∞ near 1. Then, when the categorial distance is equal to 1, the categorial dissimilarity reaches infinity, which guarantees that the units cannot be aligned.
unitary alignments. Mathematically, it constitutes a partition of the set of units (if we do not take u∅ into account). 4.6.1 Disorder of a Unitary Alignment. The disorder of a unitary alignment ă, denoted δ̆(ă), is defined for a given dissimilarity d as the average of the one-to-one dissimilarities of its units:
δ̆(ă) = 1 C2n · ∑ (u,v)∈ă2 d(u, v) (6)
Averaging dissimilarities rather than summing them makes the result independent of the number of annotators.
4.6.2 Disorder of an Alignment. The disorder of an alignment ā, denoted δ̄(ā), is the sum of the disorders of all its unitary alignments divided by the mean number of units per annotator:
δ̄(ā) = 1x · |ā|∑
i=1
δ̆(ăi) (7)
We chose to consider the average value rather than the sum so that the disorder does not depend on the size of the continuum. Best alignment â. An alignment ā of the annotation set s is considered as the best (with respect to a dissimilarity) if it minimizes its disorder among all possible alignments of s. It is denoted â. The proposed method is holistic in that it is necessary to take into account the whole set of annotations in order to determine each unitary alignment.
Disorder of an annotation set δ(s). The disorder of the annotation set s, denoted δ(s), is defined as the disorder of its best alignment(s) δ̄(â). Note that it may happen that several alignments produce the lowest disorder.
We have just presented the two crucial definitions of our new method, which make it “unified.” Indeed, the best alignment is chosen with respect to the disorder,
therefore with respect to what computes the agreement measure; and, conversely and simultaneously, the resulting agreement value (see below) is given by the best alignment: agreement computation and alignment are fully intertwined, whereas in most agreement metrics, the alignment is fixed a priori or no alignment is used. 4.8.1 The Model of Chance of γ. As we have already mentioned in the state of the art, it is necessary for an inter-annotator agreement measure to provide chance correction. We have also seen that there are several chance correction models, and that it is a controversial question. However, for γ, we follow Krippendorff, who claims that annotators should be interchangeable, because, as stressed by Krippendorff (2011) and Zwick (1988), Cohen’s definition of expected agreement (using individual distributions) numerically rewards annotators for not agreeing on their use of values, that is to say when they have different prevalences of categories, and punishes those that do agree. Therefore, expected values of γ are computed on the basis of the average distribution of observed annotations of the several annotators.
More precisely, we define the expected (chance) disorder as the average disorder of a multi randomly annotated continuum where:r The random annotations fulfill the observed annotation distributions for the
following features:
– the distribution of the number of units per annotator
– the distribution of categories
– the distribution of unit length per category
– the distribution of gaps’ length
– the distribution of overlapping and/or covering between each pair of categories (for instance, units of categories A and B may never intersect, 7% of the units of category A may cover one unit of category C, and so on).r The number of random annotators is the same as the number of annotators
in the observed data
4.8.2 Two Possible Sources to Build Chance: Local Data versus Corpus Data. In addition, whereas other studies systematically compute the expected value on the data also used to compute the observed value (see Section 3.1), we consider that it should be computed, when possible (that is to say, when several continua have been annotated with the same set of categories and the same instructions), from the distribution observed in all continua of the annotation effort the evaluated continuum comes from: If distribution changes from one continuum to another one, it is more because of the content of each continuum than because of chance. Let us illustrate this by a simple example, where two annotators have to annotate several texts from a sentiment analysis point of view, using three available categories: positive, negative, and neutral. On average, on the whole corpus, we assume that the prevalence is 13 for each category. The expected agreement on the whole corpus is thus 0.33. We also assume that for one particular text, there are only positive and neutral annotations, 50% of each, and no negative one. The expected agreement for this particular text is 0.5, which means that
this particular text is considered to facilitate agreement by chance, and which has the consequence that the final agreement will be more conservative than for the rest of the corpus. Why does the third category, “negative,” not appear in this expected agreement computation? This conception of chance considers that when an annotator begins to annotate this particular text, which she does not already know, the third category no longer exists in her mind, and that it is the case for every other annotator, whereas they are not supposed to cooperate. It cannot be by chance that all annotators use one category in some texts, and not in another one, but because of the content, and of the interpretation, of a given text. For this reason, from our point of view, it is better to take into account the data observed on a whole annotation effort rather than on each individual continuum. The complete data tell more about the mean behavior of annotators, whereas data of a given continuum may depend more on the particularities of its content.
As a consequence, γ provides two ways to compute the expected values: one which considers only the data of the continuum being evaluated, as does every other coefficient; and a second one, which considers the data from all continua of the annotation effort the evaluated continuum comes from. When available, we recommend using the second one, for the reasons already expressed.
4.8.3 Using Sampling to Compute the Expected Value. Expected agreement (or disagreement) is the expected value of a random variable. But which random variable? For coefficients like kappa and alpha, observed agreement (or disagreement) is the mean agreement (or disagreement) on all pairs of instances, so the random variable can be as simple as a random pair of instances (however we interpret “random”). This value can be readily computed. For gamma, however, observed disagreement is determined on a whole annotation, so the random variable needs to be a whole random annotation. The expected value of such a complicated variable is much more difficult to determine analytically. Instead, gamma uses sampling, as introduced in Section 5. Now that the disorder and the expected disorder have been introduced, we can define the agreement measure (of annotation set s belonging to corpus c, with c = {s} if s is a sole annotation set) with Equation 8, which is derived from Equation 2:
∀s ∈ c,γ = 1− δ(s) δe(c)
(8)
If all annotators perfectly agree (Figure 11a), γ = 1. Figure 11c corresponds to the worst case, where the annotators are worse than annotating at random, with γ < 0. Figure 11b shows an intermediate situation.","5 implementation :In this section, we first propose an efficient solution to compute the disorder of an annotated continuum, which relies on linear programming. Second, we propose two ways to generate random annotated continua (with respect to the observed distributions) to compute the expected disorder, one relying on a single continuum, the other one relying on a corpus (i.e., several continua). Third, we determine the number of random data sets
that we must generate (and compute the disorder of) to obtain an accurate value of the expected disorder. In order to simplify the discussion and the demonstrations, we consider in this section that n annotators all made the same number of annotations p.
The proposed method has now been fully described on a theoretical level, but, being holistic, its software implementation leads to a major problem of complexity. One can demonstrate that there are theoretically (p!)n−1 possible alignments. However, we will (1) show how to reduce the initial complexity, and (2) provide an efficient linear programming solution.
5.1.1 Reducing the Initial Complexity. The initial number of possible unitary alignments (which are used to build a possible alignment) is pn. Fortunately, theorem provided as Equation (9) states that any unitary alignment with a cost beyond the value n ·∆∅ cannot belong to the best alignment, and so can be discarded. Indeed, any unitary alignment with a cost above ∆∅ can be replaced by creating a separate unitary alignment for each unit (of cost ∆∅ per unitary alignment, so of total cost n ·∆∅).
Demonstration. Consider the best alignment â, of cardinality m. Let ă be any of its unitary alignments. For convenience, we attribute to it the index 1 (ă = ă1), while the others are indexed from 2 to m. This unitary alignment ă contains n units (either real or u∅). For each of these units ui (1 ≤ i ≤ n), we create the unitary alignment ăm+i = (ui, u∅, ..., u∅) of cardinality n. It is possible to create an alignment ā made up of the set of unitary alignments of â \ {ă}, to which we add the unitary alignments ăm+1 to ăm+n that we have just created.3 It is of cardinality m + n− 1. Because â minimizes the disorder, we obtain:
δ̄(â) ≤ δ̄(ā)⇒ 1x m∑ i=1 δ̆(ăi) ≤ 1x m+n∑ i=2 δ̆(ăi)
⇒ m∑
i=1 δ̆(ăi) ≤ m+n∑ i=2 δ̆(ăi)
⇒ δ̆(ă1) ≤ m+n∑
i=m+1
δ̆(ăi)
since ∀i > m, δ̆(ăi) = 1C2n (C 2 n∆∅) = ∆∅, and since we have denoted ă = ă1,
⇒ δ̆(ă) ≤ n ·∆∅ (9)
Experiments have shown that this theorem allows us to discard about 90% of the unitary alignments.
3 ā is indeed an alignment, because each of its units appears in one and only one unitary alignment.
5.1.2 Finding the Best Alignment: A Linear Programming Solution. Finding the best alignment consists of minimizing the global disorder. Such a problem may be described as a linear programming problem, so that the solution can be computed by a linear programming solver. For convenience, we introduce two new definitions:r Let UA be the set of all unitary alignments.r Let UAu be the set of the unitary alignments which contain unit u.
The description of the problem in linear programming terms is threefold. First, for a given alignment ā, for each possible unitary alignment ăi, we define the Boolean variable Xăi , which indicates if this unitary alignment belongs or not to the alignment:
∀ăi ∈ UA, Xāăi = { 0 iff ăi 6∈ ā 1 iff ăi ∈ ā
Second, we have to express the fact that, by definition, each unit u (of each annotator) should belong to one and only one unitary alignment of the alignment ā, that is to say that among all unitary alignments containing u, exactly one Xăi equals 1, and all the others equal 0:
∀u ∈ U , ∑
ăi ∈ UAu
Xāăi = 1
Third, the goal is to minimize the global disorder δ̄(ā) associated with ā, among all possible alignments ā:
Minimize δ̄(ā) = ∑
ăi ∈ UA δ̆(ăi) · Xāăi
The LPSolve solver4 finds the best solution in less than one second with n = 3 annotators and p = 100 annotations per annotator on a current laptop (once the initial complexity has been reduced thanks to the previous theorem), which is fast enough to be practical. The next two subsections detail two strategies to generate randomly annotated continua with respect to the definition of the expected disorder of γ, and the third subsection explains how to choose the number of expected disorder samples to generate so that their average is an accurate enough value of the theoretical expected value. The two strategies correspond to the need expressed in Section 4.8.1 to compute the expected value on the largest set of available data, either a single continuum, or, when available, several continua from the same corpus.
4 http://lpsolve.sourceforge.net.
5.2.1 A Strategy to Compute the Expected Disorder Using a Single Continuum. When the annotation effort is limited to a single continuum, we can only rely on the annotated continuum itself to compute the expected value. To create random annotations that fulfill the observed distributions, the implemented strategy is as follows: We take the real annotated continuum of an annotator (such as the example shown on the left in Figure 12), choose at random a position on this continuum, split the continuum at this position, and permute the two parts of the split continuum. Three examples of split and permutation are shown in the right part of the figure, for split positions of, respectively, 15, 24, and 38, all coming from the same real continuum, with units that are no longer aligned (except by chance). However, we have to address the fact that some units may intersect with themselves, generating some part of agreement beyond chance. For instance, in Figure 12, unit 3 intersects with itself between #15 and #24, because the length of the unit, 12, is higher than the difference of shifts 24− 15 = 9. To limit this phenomenon, we do not allow the distance between two shifts to be less than the average length of units.
5.2.2 A Strategy to Compute the Expected Disorder Using Several Continua (from the Same Corpus). This strategy consists of mixing annotations coming from different continua, so that their units may align only by chance. To create a random annotation of n annotators, we randomly choose n different continua of the corpus, and pick the annotations of one annotator (randomly chosen) of each of these continua. When different texts are of different lengths, each of them is adjusted to the longest one by duplicating as many times as necessary (like a mosaic).
This is shown in Figure 13 for n = 3 annotators. We assume the corpus contains eight continua, each annotated by three annotators. To generate a random set of three annotations, we have randomly selected a combination of three values between 1 and 8, here (2, 4, 7), to select three different continua among the eight available ones of the corpus. Then, for each of these selected continua, we choose one annotator, here annotator 2 for continuum 2, annotator 3 for continuum 4, and annotator 1 for continuum 7. We combine the associated annotations as shown in the right part of the figure, and obtain a set of random annotations that fulfill (on average) the observed distributions. The (very limited) extent of the resulting agreement we can see in this example (only two units have matching categories, but with discrepancies in position) is only by chance because the compared annotations come from different continua.
In addition, it is possible to create a great number of random sets of annotations with this strategy: With n annotators and m continua (m ≥ n), it is possible to generate up to Cnm · nn different combinations. For instance, in our example, which assumes n = 3 and m = 8, there are 56× 33 = 1512 combinations to create random annotations. Because the expected disorder is by definition randomly obtained on average, and because there is virtually an infinite number of possible random annotations (with a discrete and finite continuum, it is not really infinite, but still too big to be computed), we can only compute a reduced but sufficient number of experiments and obtain an approximate value of the expected disorder. This is a sampling problem as described, for example, in Israel (1992). What statistics provide is a way to determine the minimal number n0 of experiments to do (and to average) so that we get an approximate result of a given precision with a given confidence level. It consists in: First, taking a small sample to estimate the mean and standard deviation; then, using these estimates to determine the sample size n0 that is needed.
We follow the strategy provided in Olivero (2001) to compute a disorder value that differs less than e = 2% from the real value with a (1− α) = 95% confidence (the software distribution we provide is set by default with these values).
First, we consider a sample of chance disorder values of size n = 30. Let µ be the sample mean, and σ′ be its standard deviation. µ is directly an unbiased estimator of the population mean, and σ = √ ( nn−1 ) · σ′ is an unbiased estimator of the real standard deviation. Let Cv = σµ be the coefficient of variation (i.e., the relative standard deviation). Let U1−α2 be the abscissa of the normal curve that cuts off an area α at the tails. This value is provided in statistical tables. We get n0 by the following equation:
n0 = (Cv ·U1−α2
e )2 Let us consider a real example. We generate a sample of random disorders of size n = 30. We compute its mean µ = 3.49, its standard deviation σ′ = 0.1379, hence σ = 0.1403, and Cv = 0.040188. We get U1− 0.052 = 1.96 from the corresponding available table, hence we obtain n0 = 15.5. This means that a sample of 16 disorder values gives 2% of precision with 95% confidence. The mean we have already computed with 30 values fulfills this condition, and is a good approximation of the real expected disorder. If we wish to obtain a high precision of 1%, we need n0 = 62. It is beyond the
initial size of our sample (which is 30), and we will have to generate an additional set of 32 values in order to obtain the required number.","6 comparing and benchmarking γ :As γ is an entirely new agreement measure method, it is necessary to analyze how it compares with some well known and much studied methods. First, we carry out a thorough comparison between γ and the two dedicated alphas, uα and c|uα, which are the most specific measures in the domain. Second, we benchmark γ by comparing it with other main measures, thanks to a special tool that is briefly introduced. As already mentioned, Krippendorff’s uα and c|uα are clearly the most suitable coefficients for combined unitizing and categorizing. To better understand the pros and cons as well as the behavior of these measures compared with γ, we first explain how they are designed in Section 6.1.1, and then make thorough comparisons with γ from Section 6.1.2 to 6.1.6 including: (1) how they react to slight categorial disagreements, (2) interlacement of positional and categorial disagreements, (3) the impact of the size of the units on positional disagreement, (4) split disagreements, and (5) the impact of scale (e.g., if the size of all units is multiplied by 2). We finish by showing a paradox of uα in Section 6.1.7.
6.1.1 Introducing uα and c|uα. To introduce how these two coefficients work, let us consider the example taken from Krippendorff (2013), shown in Figure 14. The length of the continuum is 76, there are two annotators, and there are four possible categories, numbered 1 to 4.
The uα coefficient basically relies on the comparison of all pairs of sections among annotators, a section being either a categorized unit or a gap. To get the observed disagreement value uDo, squared lengths of the unmatching intersections are summed, and this sum is then divided by the product of the length of the continuum and m(m− 1), m being the number of annotators. In the example, mismatches occur around the second and third units of the two annotators. From left to right, there are the following intersections: cat 1 with gap (l = 10), cat 1 with cat 3 (l = 5), gap with cat 3 (l = 8), cat 2 with cat 1 (l = 5), and cat 2 with gap (l = 5). This leads twice (by symmetry) to the sum 102 + 52 + 82 + 52 + 52, and so the observed disagreement uDo = 2(102+52+82+52+52 )
76·2(2−1) = 3.145. The expected value uDe is obtained by considering all the possible positional combinations of each pair, and not only the observed ones. This means that for a given pair, one of the two units is virtually slid in front of the other in all
possible ways, and the corresponding values are averaged. In this example, uDe = 5.286. Therefore, uα = 1− 3.1455.286 = 0.405.
Coefficient c|uα relies on a coincidence matrix between categories, filled with the sums of the lengths of all intersections of units for each given pair of categories. For instance, in the example, the observed coincidence between category 1 and category 3 is 5, and so on. A metric matrix is chosen for these categories, for instance, an interval metric (for numerical categories), which says that the distance between category i and category j is (i− j)2. Hence, the cost for a unitary intersection between categories 1 and 2 is (1− 2)2 = 1, but is 22 = 4 between categories 1 and 3, and so on. Then, the observed disagreement is computed according to these two matrices. To finish, an expected matrix is filled (in a way which cannot be detailed here due to space constraints), and the expected value is computed the same way. In the example, c|uα = 1− 0.8333.145 = 0.744.
Hence, Krippendorff’s alphas provide two clues to analyze the agreement between annotators. In the example, uα = 0.405 indicates that the unitizing is not so good, but also that the categorizing is much better, with c|uα = 0.744 (even though of course, these two values are not independent, since unitizing and categorizing coexist here by nature).
Now that these coefficients have been introduced in detail, let us analyze to what extent they differ from γ.
6.1.2 Slight Categorial Disagreements: Alphas Versus γ. When annotators have slight categorial disagreements (with overlapping categories), c|uα is slightly lowered. However, uα does not take categorial overlapping into account, but has a binary response to such disagreements, and is lowered as much as if they were severe categorial disagreements. A consequence of this approach is illustrated in Figure 15, where two annotators perfectly agree both on positions and categories in the experiment on the left, and still perfectly agree on position but slightly diverge concerning categories in the experiment on the right (1/2, 6/7, and 8/9 are assumed to be close categories). However, uα drops from 1 in the left experiment to –0.34 (a negative value means worse than random) in the right experiment, despite, in the latter, the positions being all correct, and the categories being quite good, since c|uα = 0.85. On such data, γ considers that there is no positional disagreement, and c|uα and γ both consider that there are slight categorial disagreements.
6.1.3 Positional Disagreements Impacting Categorial Agreement: c|uα. Two different conceptions of how to account for categorial disagreement have, respectively, led to c|uα and γ: c|uα relies on intersections between the units of different annotators, which is basically equivalent to an observation at the atom level, whereas γ relies on alignments between units (any unit being finally attached and compared to, at most, only one other) based both on positional and categorial observation. Hence, in a configuration such as the one given in Figure 16, where two annotators annotated three units with the same categories
1
1
9
9
7
7
Observer 1
Observer 2
1
2
9
8
7
6
Observer 1
Observer 2
Figure 15 Consequences of no categorial disagreement (left) compared with slight categorial disagreements (right).
1, 4, and 2, but not exactly at the same locations; c|uα considers a certain account of categorial disagreements, whereas γ does not. According to the principles of c|uα, any part of the continuum (even at the atom level) with an intersection between different categories means some confusion between them, whereas γ considers here that the annotators fully agree on categories (they both observed a “1” then a “4” then a “2” with no confusion), and disagree only on where phenomena exactly start and finish. The crucial difference between the two methods is probably whether we consider units to be non-atomizable (and therefore consider alignments, as γ does), or atomizable (in which case two different parts of a given unit may be simultaneously and respectively compared to two different units from another annotator).
6.1.4 Disagreements on Long versus Short Spans of Texts. Here again, the way disagreements are accounted for may differ markedly between uα and γ: when a unit does not match with any other, uα takes into account the length of the corresponding span of text to assess a disagreement. As shown in Figure 17, an orphan unit of size 10 will cost 100 times as much as an orphan unit of size 1, whereas for γ, they will have the same cost. In the whole example in Figure 17, to compute the observed disagreements, uα says the first case is 50 times worse than the second, whereas γ says on the contrary that the second case is twice as bad as the first. Here, γ fulfills the need (already mentioned) expressed by Reidsma, Heylen, and Ordelman (2006, page 3) to consider that “short segments are as important as long segments.” This phenomenon is the same for categories between c|uα and γ, the size of the units having consequences only for c|uα.
6.1.5 Split Disagreements. Sometimes, an annotator may divide a given span of text into several contiguous units of the same type, or may annotate the same span with one whole unit. In these cases, c|uα computes its observed disagreement the same in both
configurations, and uα assigns decreasing disagreement when splitting increases, as shown in Figure 18, whereas γ assigns increasing disagreements.
Moreover, in Figure 19, the observed uα is not responsive to splits at all, whereas γ is still responsive.
6.1.6 Scale Effects. The way uα computes dissimilarities is directly proportional to squared lengths, as shown in Figure 20. On the other hand, γ may use any positional dissimilarity, and usually uses ones that are not scale-dependent for CL applications, such as dpos−sporadic (Equation (3)). For instance, if a text is annotated with two categories, one at word level, the other one at paragraph level, we may prefer to account for relative disagreements so that a missing word will be more heavily penalized in the first case than in the second. In Figure 20, the observed disagreement of uα is 32 = 9 times greater for B units than for A units, but would be the same for γwith dpos−sporadic since:
( 0 + 3( 7+10
2 ))2 = ( 0 + 9( 21+30 2 ))2 . Figure 21a, the annotators disagree on categorization, and have a moderate agreement on unitizing. This configuration leads to uα = 0.107. In Figure 21b, the configuration is quite similar, but now annotators fully agree on unitizing: Each of them puts units in the same positions. Paradoxically, uα drops to −0.287, which is less than in the first configuration. In brief, the reason for this behavior is that in the first case, the computed disagreement regarding a given pair of units is virtually distributed into shorter parts of the whole (an intersection of length 80 between them, and an intersection of length
20 with a gap for each of them, which leads to 802 + 2× 202 = 7, 200) whereas the disagreement is maximum in the second case (an intersection of length 100 with a unit of another category, which leads to 1002 = 10, 000). Contrarily, with similar data, γ provides a better agreement in the second case than in the first one. With its design, it considers that there is the same categorial agreement in both cases, but better positional agreement in the second case, which seems to better correspond to the CL tasks we have considered.
6.1.8 Overlapping Units (Embedding or Free Overlap). Both alpha coefficients are currently designed to cope only with non-overlapping units (the term overlapping also stands here for embedding), which is a limitation for several fields in CL. It is debatable whether they could be generalized to handle overlapping units. It seems that it would involve a major change in the strategy, which currently necessitates comparing the intersections of all pairs of units. In the example shown in Figure 22, even though annotators fully agree on their two units, the alphas will inherently compare A1 with B2 and A2 with B1 (in addition to the normal comparisons between A1 with B1 and A2 with B2), and will count the resulting intersections as disagreements. It is necessary here to choose once and for all what unit to compare to what other, rather than to perform all the comparisons. But making such a choice precisely consists in making an alignment, which is a fundamental feature of γ. Consequently, it seems that the alphas would need a structural modification to cope with overlapping. As explained by Reidsma, Heylen, and Ordelman (2006), because of the lack of specialized coefficients coping with unitizing, a fairly standard practice is to use categorization coefficients on a discretized (i.e., atomized) version of the continuum: For instance, each character (or each word, or each paragraph) of a text is considered as an item; and a standard categorization coefficient such as κ is used to compute agreement. Such a measure is called κd (for discretized κ) hereafter. Several weaknesses of this approach have been already mentioned in the state-of-the-art section. It is interesting to compare such a measure to the specialized one c|uα: even if they both bear the aggregatable hypothesis, they have, however, significant differences (as confirmed by the experiments presented in the next section). The main one is that c|uα does not use an artificial atomization of the continuum, and only compares units with units. In doing so, it is not prone to agreement on blanks, in contrast to κd. Another difference is that, for the same reason, c|uα is not inherently limited to non-overlapping units: Even if it is not currently designed to cope with them, as we have already seen, it is possible to submit overlapping units to this measure (some results are shown in the next section). In this section on benchmarking, we use the Corpus Shuffling Tool (CST) introduced by Mathet et al. (2012) to compare γ concretely and accurately to the other measures.
We first introduce the possible error types that it will provide: category (category mistakes may occur), position (the boundaries may be shifted), false positives (the annotators add units to the reference units), false negatives (the annotators miss some of the reference units), and splits (the annotators put two or more contiguous units instead of a reference unit, which occupy the same span of text).
This tool is used to simulate varying degrees of disagreement among different error types, and the metrics are compared with each other according to how they react to these disagreements. For a given error type, for each magnitude between 0 and 1 (with a step of 0.05), the tool creates 40 artificial, multi-annotator shuffled annotation sets, and computes the different measures for them. Hence, we obtain a full graph showing the behavior of each measure for this error type, with the magnitude on the x-axis, and the average agreement (over the 40 annotation sets) on the y-axis. This provides a sort of “X-ray” of the capabilities of the measures with respect to this error type, which should be evaluated against the following desiderata:r A measure should provide a full response to the whole range of
magnitudes, which means in particular that the curve should ideally start from 1 (at m = 0) and reach 0 (at m = 1), but never go below 0 (indeed, negative agreement values require a part of systematic disagreement, which is not simulated by the current version of the CST).r The response should be strictly decreasing: A flat part would mean the measure does not differentiate between different magnitudes, and, even worse, an increasing part would mean that the measure is counter-effective at some magnitudes, where a worse error is penalized less severely.
We emphasize the fact that the whole graph is important, up to magnitude 1. Indeed, in most real annotated corpora, even when the overall agreement is high, errors
corresponding to all magnitudes may occur. For instance, an agreement of 0.8 does not necessarily correspond to the fact that all annotations are affected by slight errors (which correspond to magnitudes close to 0), but may for instance correspond to the fact that a few units are affected by severe errors (which may correspond to magnitudes close or equal to 1).
It is important to note that this tool was designed by the authors of γ, for tasks where units cannot be considered as atomizable. In particular, it was conceived so that disagreements concerning small units are as important as those concerning large ones. However, it is provided as open-source (see Conclusions section) so that anyone can test and modify it, and propose new experiments to test γ and other measures in the future.
6.3.1 Introducing the CST. The main principle of this tool is as follows. A reference corpus is built, with respect to a statistical model, which defines the number of categories, their prevalence, the minimum and maximum length for each category, and so forth. Then, this reference is used by the shuffling tool to generate a multi-annotator corpus, simulating the fact that each annotator makes mistakes of a certain type, and of a certain magnitude. It is important to remark that the generated corpus does not include the reference it is built from.
The magnitude m is the strength of the shuffling, that is to say the severity of mistakes annotators make compared to the reference. It can be set from 0, which means no damage is applied (and the annotators are perfect) to the extreme value 1, which means annotators are assumed to behave in the worst possible way (but still being independent of each other)—namely, at random.
Figure 23 illustrates the way such a corpus is built: From the reference containing some categorized units, three new sets of annotations are built, simulating three annotators who are assumed to have the same annotating skill level, which is set in this example at magnitude 0.1. The applied error type is position only, that is to say that each annotator makes mistakes only when positioning boundaries, but does not make any other mistake (the units are reproduced in the same order, with the correct category, and in the same number). At this low magnitude, the positions are still close to those of the reference, but often vary a little. Hence, we obtain here a slightly shuffled multiannotator corpus. Let us sum up the way error types are currently designed in the CST.
Position. At magnitude m, for a given unit, we define a value shiftmax that is proportional to m and to the length of the unit, and each boundary of the unit is shifted by a
value randomly chosen between −shiftmax and shiftmax (note: at magnitude 0, because shiftmax = 0, units are not shifted).
Category. This shuffling cannot be described in a few words (see Mathet et al. [2012] for details). It uses special matrices to simulate, using conditional probabilities, progressive confusion between categories, and can be configured to take into account overlapping of categories. The higher the magnitude, the more frequent and severe the confusion.
False negatives. At magnitude m, each unit has the probability m to be forgotten. For instance, at magnitude m = 0.5, each annotator misses (on average) half of the units from the reference (but not necessarily the same units as the other annotators).
False positives. At magnitude m, each annotator adds a certain number of units (proportional to m) to the ones of the reference.
Splits. At magnitude m, each annotator splits a certain number of units (proportional to m). A split unit may be re-split, and so on.
6.3.2 Pure Segmentation: γ, WD, GHD. Even if γ was created to cope with error types that are poorly or not at all dealt with by other methods, and, moreover, to cope simultaneously with all of them (unitizing of categorized and overlapping categories), it is illuminating to observe how it behaves in more specific error types, to which specialized and well known methods are dedicated. We start with pure segmentation. Figure 24 shows the behavior of WD, GHD, and γ for two error types.
For false negatives, WD and GHD are quite close, with an almost linear response until magnitude 0.6. Their drawback is that their responses are limited by an asymptote, because of the absence of chance correction, while γ shows a full range of agreements; for shifts, WD and GHD show an asymptote at about agreement = 0.4, while γ shows values from 1 to 0. This experiment confirms the advantage of using γ instead of these distances for inter-annotator agreement measure.
6.3.3 Pure Categorization. In this experiment, the CST is set to three annotators, four categories with given prevalences. The units are all of the same size, positioned in fixed, predefined positions, so that the focus is on categorizing only. It should be noted that, with such a configuration, α and κ behave exactly in the same way as c|uα. It is particularly striking in Figure 25 that γ behaves in almost the same way as c|uα. In fact, the observed values of these measures are exactly the same, the only difference coming from a slight difference in the expected values, due to sampling. Other tests carried out with the pure categorizing coefficient κ yielded the same results on this particular error type, which means that γ performs as well as recognized measures as far as categorizing is concerned, with two or more annotators. The uα curve goes below zero at magnitude 0.5 (probably for the reasons seen in Section 6.1.7). Moreover, its behavior depends on the size of the gaps: Indeed, with other settings of the shuffling, the curve may, on the contrary, be stuck over zero. κd fails to reach 0 because of the virtual agreement on gaps (but it would if there were no gaps). Lastly, SER (averaging the results of each pair of annotators) is bounded below by 0.6, which results from not taking chance into account.
6.3.4 Almost General Case: Unitizing + Categorizing. This section concerns the more general uses of γ, combining both unitizing and categorizing. However, in order to be compliant with uα, c|uα, and κd, we limit the configurations here so that the units do not overlap at all. In particular, the reference was built with no overlapping units, and we have used a modified version of the shifting shuffling procedure so that the nonoverlapping constraint is fully satisfied, even at high magnitudes.
Positional errors (Figure 26a). An important point is that this shuffling error type, which is based only on moving positions, has a paradoxical consequence on category agreement, since units of different categories align when sufficient shifting is applied. Consequently, c|uα is not blocked at 1, even though it is designed to focus on categories. Additionally, it starts to decrease from the very first shifts, as soon as units from different annotators start overlapping. This is a concrete consequence of what has been formally studied in Section 6.1.3. γ has a most progressive response, reaches 0.1 at magnitude 1,
and is the only measure to be strictly decreasing. SER immediately drops to agreement 0.5 at magnitude 0.05. As it relies on a binary positional distance, it fails to distinguish between small and large errors. This is a serious drawback of such a measure for most CL tasks. Then it goes below zero and is not strictly decreasing. uα is mostly strictly decreasing, but has some increasing parts, and, even more problematic, negative values from 0.6 to 0.9, probably because of the reason explained in Section 6.1.7. κd is too responsive at the very first magnitudes, and is not strictly decreasing, probably because it “does not compensate for differences in length of segments” (Reidsma, Heylen, and Ordelman 2006, page 3).
Positional and categorial errors (Figure 26b). γ is strictly decreasing and reaches 0. The alphas are not strictly decreasing, and once again uα drops below 0 from magnitude 0.6 onwards. κd is not strictly decreasing (again, probably because it “does not compensate for differences in length of segments”), but its general shape is not that far from γ.
Split errors (Figure 27). The split error type would need to create an infinite number of splits to mean pure chaos at magnitude 1. As this is computationally not possible, we restricted the number of splits to five times the number of units of the reference. We should therefore not expect measures to reach 0. In this context, γ shows a good range of responses, from 1 to 0.2, in an almost linear curve. SER is also quite linear, but gives very confusing values for this error type because it reaches negative values above magnitude 0.6. Finally, uα, c|uα, and κd are not responsive at all to this error type, as expected, and remain blocked at 1 (which is normal for c|uα, which focuses on categorizing).
False positives and false negatives (Figure 28). In the current version of the CST, the false positive error type creates some overlapping (new units may overlap), and it is the reason why uα and κd were discarded from this experiment. However, we have kept c|uα because it behaves quite well despite overlapping units.
All the measures have overall a good response to the false positives error type, as shown in Figure 28a, even if the shape of c|uα is delayed compared with the others, but it should be pointed out that SER has a curious and unfortunate final increasing section (not visible in the figure because this section is below 0).
On the other hand, bigger differences appear with false negatives (Figure 28b). γ is still strictly decreasing and almost reaches 0 (0.025), but uα is not strictly decreasing, and is at 0 or below from m = 0.3; SER quickly drops below 0 from m = 0.4, κd is not strictly decreasing, and c|uα, as for splits, does not react at all but remains stuck at 1, which is desired for this coefficient focused on categories (values of c|uα over m = 0.7
are missing since there are not enough intersections between units for this measure to work).
Overview of each measure for the almost general case. In order to summarize the behavior of each measure in response to the different error types for the almost general case (without overlap), we pick all curves relative to a given measure out of the previous plots and draw them in the same graph, as shown in Figure 29. Briefly, γ shows a steady behavior for all error types, almost strictly decreasing from 1 to 0. uα has some increasing parts and negative values and is sometimes not responsive. c|uα is very responsive for some error types, is less responsive for some other types, and is sometimes not responsive at all (which is desired, as already said). SER has unreliable responses, being either too responsive (reaching negative values) or not responsive enough. Finally, κd is not always responsive, is most of the time not strictly decreasing, but is sometimes quite progressive.
6.3.5 Fully General Case: Unitizing + Categorizing. This last section considers the fully general case, where overlapping of units within an annotator is allowed. In this experiment, we took a reference corpus with no overlap, but the errors applied (combination of positioning and false positives) progressively lead to overlapping units. The results are shown in Figure 30. As expected, γ behaves quite the same as it does with non-overlapping configurations. Admittedly, c|uα was not designed to handle these configurations (and so should not be included in this experiment), but surprisingly it
seems to perform in rather the same way as it does with no overlapping; this must be investigated further, but judging from this preliminary observation, it seems this coefficient could still be operational and useful in such cases. On the contrary, uα does not handle correctly this experiment and so was not included in the graph.","7 conclusion :The present work addresses an aspect of inter-annotator agreement that is rarely studied in other approaches: the combination of unitizing and categorizing. Nevertheless, the use of methods that have been transposed from other domains (such as κ, which was originally dedicated to pure categorizing) in CL, for example at the discourse level, leads to severe biases, and manifests the need for specialized coefficients, fair and meaningful, suitable for annotation tasks focusing on complex objects.
In the end, only Krippendorff’s coefficients uα and c|uα come close to the needs we expressed in the introduction, with the restriction that they are natively limited to non-overlapping units.
The main reason why research on this topic is sparse, and why it may be difficult to enlarge Krippendorff’s coefficients to overlapping units, probably results from the fact that we are facing here a major difficulty: the simultaneous double discrepancy between annotators, with annotations possibly differing both in positioning relevant units anywhere on a continuum, and in categorizing each of these free units. Consequently, it is difficult for a method to choose precisely which features to compare between different annotators (unlike pure categorizing, where we know exactly what each annotator says for each predefined element to be categorized); and this problem is exacerbated when overlapping units (within an annotator) occur.
To cope with this critical point, we advocate the use of an alignment that ultimately expresses which unit from one annotator should be compared to which unit, if any, from another one, and consequently makes it natural and easier to compute the agreement. Moreover, we have shown that this alignment cannot be done in an independent way, but is part of the measure method itself. This is the “unified” aspect of our approach.
We have also shown that in order to be relevant, this alignment cannot be done at a local level (unit by unit), but should consider the whole set of annotations at the same time, which is the “holistic” aspect.
This is how the new method γ presented here was designed. Moreover, this method is highly configurable to cope with different annotation tasks (in particular, boundary errors should not necessarily be considered the same for all kinds of annotations), and it provides the alignment that emerges from an agreement measurement. Not only is this alignment a result in itself, which can be used to build a gold standard very quickly from a multi-annotated corpus (by listing all unitary alignments, and for each of them showing the corresponding frontiers and category proposed by each annotator), but it also behaves as a kind of a “flight recorder” of the measure: Observing these alignments gives crucial information on the choices the measure makes and whether it needs to be adjusted, unlike other methods which only provide a sole “out of the box” value.
Finally, we have compared γ to several other popular coefficients, even in their specific domains (pure categorization, pure segmentation), through a specific benchmark tool (namely, CST) which scans the responsivity of the measures to different kinds of errors and at all degrees of severity. Overall, γ provides broader and more progressive responsivity than the others in the experiments shown here. Concerning pure categorizing, γ does not have an edge over the well-known coefficients, such as α, but it is interesting to see that it behaves in much the same way as others in this specific field. Concerning segmentation, γ outperforms WD and GHD, by taking chance into account, but also by not depending on the heterogeneity of the segment sizes. Concerning unitizing with categorizing, as theoretically expected and confirmed by the benchmarking, SER shows severe limitations, such as a binary response to various (small or severe) positional or categorial errors, the fact that it does not make chance correction, or its limitation to two annotators only. Krippendorff’s coefficients uα and c|uα present very interesting properties, such as chance correction. However, as we have shown with thorough comparisons, they rely on quite different hypotheses to ours, since they consider intersections between units whereas we advocate considering aligned units. We have identified several situations in CL where considering alignments is advantageous, for instance, when contiguous segments of the same type may occur, or when errors on several short units should be considered as more serious than one error on a long unit, but we do not posit these situtations as a universal rule. In conclusion, when unitizing and categorizing involve internal overlapping of units, only γ is currently available, and, even if it cannot be compared to any other method at the moment for this reason, benchmarking reveals very similar responses to overlapping configurations and to non-overlapping ones, which already demonstrates its consistency and its relevance.
We can summarize the features of γ as follows: It takes into account all varieties of unitizing, combines unitizing and categorizing simultaneously, enables any number of annotators, provides chance correction, processes an alignment while it measures agreement, and provides progressive responsivity to errors both for unitizing and for categorizing. This makes γ suitable for annotation tasks such as relative to NAMED ENTITY, DISCOURSE FRAMING, TOPIC TRANSITION, or ENUMERATIVE STRUCTURES.
The full implementation of γ is provided as Java open-source packages on the http://gamma.greyc.fr Web site. It is already compatible with annotations created with the Glozz Annotation Platform (Widlöcher and Mathet 2012), and with annotations generated by the Corpus Shuffling Tool.",,"Agreement measures have been widely used in computational linguistics for more than 15 years to check the reliability of annotation processes. Although considerable effort has been made concerning categorization, fewer studies address unitizing, and when both paradigms are combined even fewer methods are available and discussed. The aim of this article is threefold. First, we advocate that to deal with unitizing, alignment and agreement measures should be considered as a unified process, because a relevant measure should rely on an alignment of the units from different annotators, and this alignment should be computed according to the principles of the measure. Second, we propose the new versatile measure γ, which fulfills this requirement and copes with both paradigms, and we introduce its implementation. Third, we show that this new method performs as well as, or even better than, other more specialized methods devoted to categorization or segmentation, while combining the two paradigms at the same time.","[{""affiliations"": [], ""name"": ""Yann Mathet""}, {""affiliations"": [], ""name"": ""Antoine Widl\u00f6cher""}, {""affiliations"": [], ""name"": ""Jean-Philippe M\u00e9tivier""}]",SP:48c2046e4054182647ac9058902fdedb3869a97e,"[{""authors"": [""Ludovic Tanguy"", ""Marianne Vergez-Couret"", ""Laure Vieu.""], ""title"": ""An empirical resource for discovering cognitive principles of discourse organisation: The annodis corpus"", ""venue"": ""Proceedings of the Eighth"", ""year"": 2012}, {""authors"": [""Artstein"", ""Ron"", ""Massimo Poesio.""], ""title"": ""Inter-coder agreement for computational linguistics"", ""venue"": ""Computational Linguistics, 34(4):555\u2013596."", ""year"": 2008}, {""authors"": [""Beeferman"", ""Douglas"", ""Adam Berger"", ""John Lafferty.""], ""title"": ""Text segmentation using exponential models"", ""venue"": ""Proceedings of the 2nd Conference on Empirical Methods in Natural Language Processing, pages 35\u201346,"", ""year"": 1997}, {""authors"": [""E.M. Bennett"", ""R. Alpert"", ""A.C. Goldstein.""], ""title"": ""Communications through limited questioning"", ""venue"": ""Public Opinion Quarterly, 18(3):303\u2013308."", ""year"": 1954}, {""authors"": [""Berry"", ""Charles C.""], ""title"": ""The K statistic [letter"", ""venue"": ""Journal of the American Medical Association, 268(18):2513\u20132514."", ""year"": 1992}, {""authors"": [""Bestgen"", ""Yves.""], ""title"": ""Improving text segmentation using latent semantic analysis: A reanalysis of Choi, Wiemer-Hastings, and Moore (2001)"", ""venue"": ""Computational Linguistics, 32(1):5\u201312."", ""year"": 2006}, {""authors"": [""Bestgen"", ""Yves""], ""title"": ""Quel indice pour mesurer l\u2019efficacit\u00e9 en segmentation de textes"", ""venue"": ""Actes de TALN"", ""year"": 2009}, {""authors"": [""A. Bookstein"", ""V.A. Kulyukin"", ""T. Raita.""], ""title"": ""Generalized Hamming Distance"", ""venue"": ""Information Retrieval, (5):353\u2013375."", ""year"": 2002}, {""authors"": [""J. Carletta""], ""title"": ""Assessing agreement on classification tasks: The kappa statistic"", ""venue"": ""Computational Linguistics, 22(2):249\u2013254."", ""year"": 1996}, {""authors"": [""Carletta"", ""Jean.""], ""title"": ""Unleashing the killer corpus: Experiences in creating the multi-everything ami meeting corpus"", ""venue"": ""Language Resources and Evaluation, 41(2):181\u2013190."", ""year"": 2007}, {""authors"": [""Charolles"", ""Michel"", ""Anne Le Draoulec"", ""Marie-Paule Pery-Woodley"", ""Laure Sarda.""], ""title"": ""Temporal and spatial dimensions of discourse organisation"", ""venue"": ""Journal of French Language Studies,"", ""year"": 2005}, {""authors"": [""J. Cohen""], ""title"": ""A coefficient of agreement for nominal scales"", ""venue"": ""Educational and Psychological Measurement, 20(1):37\u201346."", ""year"": 1960}, {""authors"": [""J. Cohen""], ""title"": ""Weighted kappa: Nominal scale agreement with provision for scaled disagreement or partial credit"", ""venue"": ""Psychological Bulletin, 70(4):213\u2013220."", ""year"": 1968}, {""authors"": [""Di Eugenio"", ""Barbara"", ""Michael Glass.""], ""title"": ""The kappa statistic: A second look"", ""venue"": ""Computational Linguistics, 30(1):95\u2013101."", ""year"": 2004}, {""authors"": [""Eisenstein"", ""Jacob.""], ""title"": ""Hierarchical text segmentation from multi-scale lexical cohesion"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association"", ""year"": 2009}, {""authors"": [""Fleiss"", ""Joseph.""], ""title"": ""Measuring nominal scale agreement among many raters"", ""venue"": ""Psychological Bulletin, (5):378\u2013382."", ""year"": 1971}, {""authors"": [""Fort"", ""Kar\u00ebn"", ""Claire Fran\u00e7ois"", ""Olivier Galibert"", ""Maha Ghribi.""], ""title"": ""Analyzing the impact of prevalence on the evaluation of a manual annotation campaign"", ""venue"": ""Eighth International Conference on Language"", ""year"": 2012}, {""authors"": [""Laurent.""], ""title"": ""Named and specific entity detection in varied data: The quaero named entity baseline evaluation"", ""venue"": ""Seventh International Conference on Language 477"", ""year"": 2010}, {""authors"": [""Hills"", ""CA. Krippendorff"", ""Klaus""], ""title"": ""On the reliability"", ""year"": 1995}, {""authors"": [""Labadi\u00e9"", ""Alexandre"", ""Patrice Enjalbert"", ""Yann Mathet"", ""Antoine Widl\u00f6cher.""], ""title"": ""Discourse structure annotation: Creating reference corpora"", ""venue"": ""Workshop on Language Resource and Language"", ""year"": 2010}, {""authors"": [""S. Lamprier"", ""T. Amghar"", ""B. Levrat"", ""F. Saubion.""], ""title"": ""On evaluation methodologies for text segmentation algorithms"", ""venue"": ""Proceedings of ICTAI 2007, pages 19\u201326. Patras."", ""year"": 2007}, {""authors"": [""Makhoul"", ""John"", ""Francis Kubala"", ""Richard Schwartz"", ""Ralph Weischedel.""], ""title"": ""Performance measures for information extraction"", ""venue"": ""Proceedings of DARPA Broadcast News Workshop, pages 249\u2013252,"", ""year"": 1999}, {""authors"": [""Mathet"", ""Yann"", ""Antoine Widl\u00f6cher.""], ""title"": ""Une approche holiste et unifi\u00e9e de l\u2019alignement et de la mesure d\u2019accord inter-annotateurs"", ""venue"": ""Traitement Automatique des Langues Naturelles 2011"", ""year"": 2011}, {""authors"": [""Nadeau"", ""David"", ""Satoshi Sekine.""], ""title"": ""A survey of named entity recognition and classification"", ""venue"": ""Linguisticae Investigationes, 30(1):3\u201326."", ""year"": 2007}, {""authors"": [""Olivero"", ""Patrick.""], ""title"": ""Calcul de la taille des \u00c9chantillons"", ""venue"": ""CETE du Sud-Ouest / DAT / ZELT. Technical report."", ""year"": 2001}, {""authors"": [""Passonneau"", ""Rebecca J."", ""Vikas Bhardwaj"", ""Ansaf Salleb-Aouissi"", ""Nancy Ide.""], ""title"": ""Multiplicity and word sense: Evaluating and learning from multiply labeled word sense annotations"", ""venue"": ""Language Resources and"", ""year"": 2012}, {""authors"": [""L. Pevzner"", ""M. Hearst.""], ""title"": ""A critique and improvement of an evaluation metric for text segmentation"", ""venue"": ""Computational Linguistics, 28(1):19\u201336."", ""year"": 2002}, {""authors"": [""D. Reidsma""], ""title"": ""Annotations and Subjective Machines of Annotators, Embodied Agents, Users, and Other Humans"", ""venue"": ""Ph.D. thesis, University of Twente."", ""year"": 2008}, {""authors"": [""D. Reidsma"", ""D.K.J. Heylen"", ""R.J.F. Ordelman.""], ""title"": ""Annotating emotions in meetings"", ""venue"": ""Proceedings of the Fifth International Conference on Language Resources and Evaluation, LREC 2006,"", ""year"": 2006}, {""authors"": [""Reidsma"", ""Denis"", ""Jean Carletta.""], ""title"": ""Reliability measurement without limits"", ""venue"": ""Computational Linguistics, 34(3):319\u2013326."", ""year"": 2008}, {""authors"": [""Scott"", ""William.""], ""title"": ""Reliability of content analysis: The case of nominal scale coding"", ""venue"": ""Public Opinion Quarterly, 19(3):321\u2013325."", ""year"": 1955}, {""authors"": [""Siegel"", ""Sidney"", ""N. John Castellan.""], ""title"": ""Nonparametric Statistics for the Behavioral Sciences"", ""venue"": ""McGraw-Hill, New York, 2nd edition."", ""year"": 1988}, {""authors"": [""Teufel"", ""Simone.""], ""title"": ""Argumentative Zoning: Information Extraction from Scientific Articles"", ""venue"": ""Ph.D. thesis, University of Edinburgh."", ""year"": 1999}, {""authors"": [""Teufel"", ""Simone"", ""Jean Carletta"", ""Marc Moens""], ""title"": ""An annotation scheme for discourse-level argumentation in research"", ""year"": 1999}, {""authors"": [""Teufel"", ""Simone"", ""Marc Moens.""], ""title"": ""Summarizing scientific articles: Experiments with relevance and rhetorical status"", ""venue"": ""Computational Linguistics, 28(4):409\u2013445."", ""year"": 2002}, {""authors"": [""Widl\u00f6cher"", ""Antoine"", ""Yann Mathet.""], ""title"": ""The glozz platform: A corpus annotation and mining tool"", ""venue"": ""ACM Symposium on Document Engineering (DocEng\u201912), pages 171\u2013180, Paris."", ""year"": 2012}, {""authors"": [""Zwick"", ""Rebecca.""], ""title"": ""Another look at interrater agreement"", ""venue"": ""Psychological Bulletin, (103):347\u2013387. 479"", ""year"": 1988}]","acknowledgments  :We wish to thank three anonymous reviewers for helpful comments and discussion. The authors would also like to warmly thank Klaus Krippendorff for his support when they implemented his coefficients in order to test them. This work was carried out in the GREYC Laboratory, Caen, France, with the strong support of the French Contrat de Projet État-Région (CPER) and the Région Basse-Normandie, which have provided this research with two engineers, Jérôme Chauveau and Stéphane Bouvry.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"appendix a. examples of linguistic objects and possible annotation tasks :Terms emphasized hereafter refer to the terminology defined in Section 2.
PART-OF-SPEECH. Part-of-speech (POS) tagging (see, for example, Güngör [2010] for a recent state of the art) gives a well-known illustration of a pure categorization without unitizing task: for all the words in a text (predefined units, full-covering, no overlap), annotators have to select a category, belonging to quite a small set of exclusive elements. POS units (words) having the same label are obviously not aggregatable.
GENE RENAMING. In a study on gene renaming presented in Fort et al. (2012), all the tokens (predefined units, no overlap) are markable (categorization) with “Nothing” (the default value), “Former” (the original name of a gene) or “New” (its new name). This work at word level considers sparser (sporadic) phenomena than POS tagging. However, the annotation is defined as full-covering, with “Nothing” as a default tag. Note that the presence of the “Nothing” category also reveals here the reduction of a unitizing problem (detection of renaming) to a pure coding system (categorization). These units are not aggregatable.
WORD SENSE. For the annotation task described in Passonneau et al. (2012), annotators were asked to assign sense labels (categorization without unitizing) to preselected moderately polysemous words (sporadicity, predefined units, no overlap) in preselected sentences where they occur. Adjacent words are not aggregatable with sense preservation.
NAMED ENTITIY. Well-established named entity (NE) recognition tasks (see, for example, Nadeau and Sekine [2007]) led to many annotation efforts. In such tasks, the annotator is often asked to identify the units in the text’s continuum (unitizing, sporadicity) and to select a NE type from an inventory (categorization). It is well known that some difficulties of NE annotation relate to the delimitation of NE boundaries. For example, for a phrase such as “Mr X, the President of Y,” it makes sense to annotate subparts (“X,” “Mr X,” “the President of Y”) and/or the whole. “Y” is also a NE of another type. This may result in hierarchical or free overlapping structures. Adjacent NE are not aggregatable.
ARGUMENTATIVE ZONING. Studies concerned by argumentative zoning (Teufel 1999; Teufel, Carletta, and Moens 1999; Teufel and Moens 2002) consider the argumenative structure of texts, and identify text spans having specific roles. For each sentence (fullcovering, predefined units, no overlap), a category (categorization) is selected. Adjacent sentences of the same type are aggregated into larger spans (argumentative zones). This reveals an underlying question of unitizing. However, it has to be noted that the categorization mainly concerns predefined sentences: argumentative types are aggregatable.
DISCOURSE FRAMING. In Charolles et al.’s discourse framing hypothesis (Charolles et al. 2005, page 121), “a discourse frame is described as the grouping together of a number of propositions which are linked by the fact that they must be interpreted with reference to a specific criterion, realized in a frame-initial introducing expression.” Thus, temporal or spatial introducing expressions lead, for example, to temporal or spatial discourse frames in the text continuum (unitizing, sporadicity, categorization). Discourse frames are not aggregatable. Subordination is possible, leading to possibly hierarchical overlap, where frames (of the same type or of different types) are embedded.
COMMUNICATIVE BEHAVIOR. The multimodal AMI Meeting corpus (Carletta 2007) covers a wide range of phenomena, and contains many different layers of annotation describing the communicative behavior of the participants of meetings. For example, in Reidsma (2008), annotators are required to identify fragments in a video recording (unitizing, sporadicity) and to categorize them (categorization). For such an annotation task, one can easily imagine instruction manuals allowing annotators to use multiple labels and to identify embedded (hierarchical overlap) or free overlapping units, even if the example provided by Reidsma (2008) does not.
DIALOG ACT. Annotating dialog act conforming to a standard as defined, for example, in Bunt et al. (2010), leads annotators to assign communicative function labels and types of semantic content (categorization) to stretches of dialogue called functional segments. The possible mulitfunctionality of segments (one functional segment is related to one or more dialog acts), and the fact that annotations may be attached directly to the primary data such as stretches of speech defined by begin and end points, or attached to structures at other levels of analysis, seems to allow different kinds of configurations and annotation instructions: unitizing or pure categorization of pre-existing structures, sporadicity or full-covering, hierarchical, overlapping or linear segmentation.
TOPIC SEGMENTATION. Topic segmentation (see, for example, the seminal work by Hearst [1997] or Bestgen [2006] for a more recent state of the art), which aims at detecting the most important thematic breaks in the text’s continuum, gives an illuminating example of pure segmentation. This unitizing problem of linear segmentation is full-covering and restricted to the detection of breaks (the right boundary of a unit corresponds to the left boundary of the following segment) (no overlap). If we consider the resulting segments, there is just one category (topic segment) and then no categorization. Adjacent topic segments are obviously not aggregatable without a shift in meaning.
HIERARCHICAL TOPIC SEGMENTATION. In order to take better into account the fact that lexical cohesion is a multiscale phenomenon and that discourse displays a hierarchical structure, hierarchical topic segmentation proposed, for example, by Eisenstein (2009) preserves the main goal and properties of text-tiling (unitizing, no categorization, fullcovering, not aggregatable segments), but allows a topic segment to be subsegmented into sub-topic segments (hierarchical [but not free] overlap).
TOPIC TRANSITION. The topic zoning annotation model presented in Labadié et al. (2010) is based on the hypothesis that, in a well constructed text, abrupt topic boundaries are more the exception than the rule. This model introduces transition zones (unitizing) between topics, zones that help the reader to move from one topic to another. The annotator is asked to identify and categorize (categorization) topic segments, introduction, conclusion, and transition zones. Hierarchical overlap is possible (embedded elements of the same type or of different types are allowed). Free overlapping structures are frequent, by virtue of the nature of transitions. Adjacent topic zones and adjacent transition zones are not aggregatable.
ENUMERATIVE STRUCTURES. A study on complex discourse objects such as enumerative structures (Afantenos et al. 2012) illustrates both the need for sporadic unitizing and the need for categorization. The enumerative structures have a complex internal organization, which is composed of various types of subelements (hierarchical overlap) (a trigger of the enumeration, items composing its body, etc.) which are not aggregatable.",,,,,,,,,,"1 introduction :A growing body of work in computational linguistics (CL hereafter) or natural language processing manifests an interest in corpus studies, and requires reference annotations for system evaluation or machine learning purposes. The question is how to ensure that an annotation can be considered, if not as the “truth,” than at least as a suitable
∗ Normandie University, France; UNICAEN, GREYC, F-14032 Caen, France; CNRS, UMR 6072, F-14032 Caen, France. E-mail: {yann.mathet, antoine.widlocher, jean-philippe.metivier}@unicaen.fr
Submission received: 15 July 2013; revised version received: 5 October 2014; accepted for publication: 12 January 2015.
doi:10.1162/COLI a 00227
© 2015 Association for Computational Linguistics
reference. For some simple and systematic tasks, domain experts may be able to annotate texts with almost total confidence, but this is generally not the case when no expert is available, or when the tasks become harder. The very notion of “truth” may even be utopian when the annotation process includes a certain degree of interpretation, and we should in such cases look for a consensus, also called the “gold standard,” rather than for the “truth.”
For these reasons, a classic strategy for building annotated corpora with sufficient confidence is to give the same annotation task to several annotators, and to analyze to what extent they agree in order to assess the reliability of their annotations. This is the aim of inter-annotator agreement measures. It is important to point out that most of these measures do not evaluate the distance from annotations to the “truth,” but rather the distance across annotators. Of course, the hope is that the annotators will agree as far as possible, and it is usually considered that a good inter-annotator agreement ensures the constancy and the reproducibility of the annotations: When agreement is high, then the task is consistent and correctly defined, and the annotators can be expected to agree on another part of the corpus, or at another time, and their annotations therefore constitute a consensual reference (even if, as shown for example by Reidsma and Carletta [2008], such an agreement is not necessarily informative for machine learning purposes). Moreover, once several annotators reach good agreement on a given part of a corpus, then each of them can annotate alone other parts of the corpus with great confidence in the reproducibility (see the preface to Gwet [2012, page 6] for illuminating considerations). Consequently, inter-annotator agreement measurement is an important point for all annotation efforts because it is often considered that a given agreement value provided by a given method validates or invalidates the consistency of an annotation effort.
How to measure agreement, and how we define a good measure, is another part of the problem. There is no universal answer, because how to measure depends on the nature of the task, hence on the kind of annotations.
Admittedly, much work has already been done for some kinds of annotation efforts, namely, when annotators have to choose a category for previously identified entities. This approach, which we will call pure categorization, has led to several well-known and widely discussed coefficients such as κ, π, or α, since the 1950s. Some more recent efforts have been made in the domain of unitizing, following Krippendorff’s terminology (Krippendorff 2013), where annotators have to identify by themselves what the elements to be annotated in a text are, and where they are located. Studies are scarce, however, as Krippendorff pointed out: “Measuring the reliability of unitizing has been largely ignored in favor of coding predefined units” (Krippendorff 2013, page 310). This scarcity concerns either segmentation, where annotators simply have to mark boundaries in texts to separate contiguous segments, or more generally unitizing, where gaps may exist between units. Moreover, some even more complex configurations may occur (overlapping or embedding units), which are more rarely taken into account.
And when categorization meets unitizing, as is the case in CL in such fields as, for example, NAMED ENTITY RECOGNITION1 or DISCOURSE FRAMING, very few methods are proposed and discussed. That is the main problem we focus on in this article and to which γ provides solutions.
1 Small caps are used to refer to the examples of annotation tasks introduced in Section 2.2.
The new coefficient γ that is introduced in this article is an agreement measure concerning the joint tasks of unit locating (unitizing) and unit labeling (categorization). It relies on an alignment of units between different annotators, with penalties associated with each positional and categorial discrepancy. The alignment itself is chosen to minimize the overall discrepancy in a holistic way, considering the full continuum to make choices, rather than making local choices. The proposed method is unified because the computation of γ and the selection of the best alignment are interdependent: The computed measure depends on the chosen alignment, whose selection depends on the measure.
This method and the principles proposed in this article have been built up since 2010, and were first presented to the French community in a very early version in Mathet and Widlöcher (2011). The initial motivation for their development was the lack of dedicated agreement measures for annotations at the discourse level, and more specifically for annotation tasks related to TOPIC TRANSITION phenomena.
The article is organized as follows. First, we fix the scope of this work by defining the important notions that are necessary to characterize annotation tasks and by introducing the examples of linguistic objects and annotation tasks used in this article to compare available metrics. Second, we analyze the state of the art and identify the weaknesses of current methods. Then, we introduce our method, called γ. As this method is new, we compare it to the ones already in use, even in their specialized fields (pure categorization, or pure segmentation), and show that it has better properties overall for CL purposes. 2 motivations, scope, and illustrations :We focus in the present work on both categorizing and unitizing, and consider therefore annotation tasks where annotators are not provided with preselected units, but have to locate them and to categorize them at the same time. An example of a multi-annotated continuum (this continuum may be a text or, for example, an audio or a video recording) is provided in Figure 1, where each line represents the annotations of a given annotator, from left to right, respecting the continuum order. In order to characterize the annotation efforts focusing on specific linguistic objects, we consider the following properties, illustrated in Figure 1.
Categorization occurs when the annotator is required to label (predefined or not) units. Unitizing occurs when the annotator is asked to identify the units in the continuum:
She has to determine each of them (and the number of units that she wants) and to locate them by positioning their boundaries.
Embedding (hierarchical overlap) may occur if units may be embedded in larger ones (of the same type, or not). Free overlap may occur when guidelines tolerate the partial overlap of elements (mainly of different types). Embedding is a special case of overlapping. A segmentation without overlap (hierarchical or free) is said to be strictly linear. Full-covering (vs. sporadicity) applies when all parts of the continuum are to be annotated. For other tasks, parts of the continuum are selected. Aggregatable types or instances correspond to the fact that several adjacent elements having the same type may aggregate, without shift in meaning, in a larger span having the same type. This larger span is said to be atomizable: Labeling the whole span or labeling all of its atoms are considered as equivalent, as illustrated by Figure 2.
Two specific cases. We call hereafter pure segmentation (illustrated by Figure 3) the special case of unitizing with full-covering and without categorization, and we call pure categorization categorization without unitizing. To present the state of the art as well as our own propositions, and to make all of them more concrete, it is useful to mention examples of linguistic objects and annotation tasks for which agreement measures may be required. The following sections will then refer to these examples as often as possible, in order to illustrate discussions on abstract problems or configurations. Small caps are used to refer to the names of these tasks.
A
Aggregatable atoms
Atomizable unit
Table 1 summarizes the properties of the linguistic objects and annotation tasks mentioned in this article to illustrate and compare methods and metrics. These objects and tasks are briefly described for convenience in Appendix A. This table shows that annotation of TOPIC TRANSITIONS is the most demanding of the tasks regarding the number of necessary criteria a suitable agreement metric should assess, but most of the tasks listed here require assessment of both unitizing and categorization. 3 state of the art :As we saw in the previous section, different studies in linguistics or CL involve quite different structures, which may lead to annotation guidelines having very different properties. They require suitable metrics in order to assess agreement among annotators. As we will see, some of the needs for which γ is suitable are not satisfied by other available metrics.
Note that this description of the state of the art mainly focuses on the questions which are of most importance for this work, in particular, chance correction and unitizing. For a thorough introduction to the most popular measures that concern categorizing, we refer the reader to the excellent survey by Artstein and Poesio (2008).
In this section, we first address the question of chance correction in agreement measures, then we give an overview of available measures in three domains: pure categorization, pure segmentation, and unitizing. We begin the state of the art with the question of chance correction, because it is a crosscutting issue in all agreement measure domains, and because it influences the final value provided by most agreement measures.
It is important to distinguish between (1) measures to evaluate systems, where the output of an annotating system is compared to a valid reference, and (2) inter-annotator agreement measures, which try to quantify the degree of similarity between what different
annotators say about the same data, and which are the ones we are really concerned with in this article.
In case (1), the result is straightforward, providing for instance the percentage of valid answers of a system: We know exactly how far the evaluated system is from the gold standard, and we can compare this system to others just by comparing their results.
However, case (2) is more difficult. Here, measures do not compare annotations from one annotator to a valid reference (and, most of the time, no reference already exists), but they compare annotations from different annotators. As such, they are clearly not direct distances to the “truth.” So, the question is: Above what amount of agreement can we reasonably trust the annotators? The answer is not straightforward, and this is where chance correction is involved.
For instance, consider a task where two annotators have to label items with 10 categories. If they annotate at random (with the 10 categories having equal prevalence), they will have an agreement of 10%. If we consider another task involving two categories only, still at random, the agreement expected by chance rises to 50%. Based on this observation, most agreement measures try to remove chance from the observed measure, that is to say, to provide the amount of agreement that is above chance. More precisely, most agreement measures (for about 60 years, with well-known measures κ, S, π) rely on the same formula: If we note Ao the observed agreement (i.e., the agreement directly observed between annotators) and Ae the so-called expected agreement (i.e. the agreement which should be obtained by chance), the final agreement A is defined by Equation (1).
To illustrate this formula, assume that the observed agreement is seemingly high, say Ao = 0.9. If Ae = 0.5, A = 0.4/0.5 = 0.8, which is still considered as good, but if Ae = 0.7, A = 0.2/0.3 = 0.67, which is not that good, and if Ae = 0.9, which means annotators did not perform better than chance, then A = 0.
Some other measures, namely, all α from Krippendorff, and the new γ introduced in this article, are computed from observed and expected disagreements (instead of agreements), denoted here respectively Do and De, and they define the final agreement by Equation (2).
A = Ao − Ae1− Ae (1)
A = 1− DoDe (2)
However, the way the expected value is computed is the only difference between many coefficients (κ, S, π, and their generalizations), and is a controversial question. As precisely described in Artstein and Poesio (2008), there are three main ways to model chance in an annotation effort:
1. By considering a uniform distribution. For instance, in a categorization task, considering that each category (for each coder) has the same probability. The limitation of this approach is that it provides a poor model for chance annotation. Moreover, for a given task, the greater the number of categories, the lesser the expected value, hence the higher the final agreement.
2. By considering the mean distribution of the different annotators hence regarded as interchangeable. For instance, in a categorization task with two categories A and B, where the prevalences are respectively 90% for category A and 10% for category B, the expected value is computed as 0.9× 0.9 + 0.1× 0.1 = 0.82, which is much higher than the 0.5 obtained by considering a uniform distribution.
3. By considering the individual distributions of annotators. Here, annotators are considered as not interchangeable; each of them is considered to have her own probability for each category (for a categorization task) based on her own observed distribution. It leads to the same results as with the mean distribution if annotators all have the same distribution, or to a lesser value (hence a higher final agreement) if not.
In the two cases of mean and individual distributions, expected agreement may be very high, depending on the prevalence of categories. In some annotation tasks, expected agreement becomes critically high, and any disagreements on the minor category have huge consequences on the chance-corrected agreement, as hotly debated by Berry (1992) and Goldman (1992), and criticized in CL by Di Eugenio and Glass (2004). However, we follow Krippendorff (2013, page 320), who argues that disagreements on rare categories are more serious than on frequent ones. For instance, let us consider the reliability of medical diagnostics concerning a rare disease that affects one person out of 1,000. There are 5,000 patients, 4,995 being healthy, 5 being affected. If doctors fail to agree on the 5 affected patients, their diagnostics cannot be trusted, even if they agree on the 4,995 healthy ones.
These principles have been mainly introduced and used for categorization tasks, because most coefficients address these tasks, but they are more general and may also concern segmentation and, as we will see further, unitizing. The simplest measure of agreement for categorization is percentage of agreement (see for example Scott 1955, page 323). Because it does not feature chance correction, it should be used carefully for the reasons we have just seen.
Consequently, most popular measures are chance-corrected: S (Bennett, Alpert, and Goldstein 1954) relies on a uniform distribution model of chance, π (Scott 1955) and α (Krippendorff 1980) on the mean distribution, and κ (Cohen 1960) on individual distributions. Generalizations to three or more annotators have been provided, such as κ (Fleiss 1971), also known as K (Siegel and Castellan 1988). Moreover, weighted coefficients such as α and κw (Cohen 1968) are designed to take into account the fact that disagreements between two categories are not necessarily all of the same importance. For instance, for scaled categories from 1 to 10 (as opposed to so-called nominal categories), a mistake between categories 3 and 4 should be less penalized than a mistake between categories 1 and 10.
These metrics, widely used in CL, are suitable to assess agreement for pure categorization tasks—for example, in the domains of PART-OF-SPEECH TAGGING, GENE RENAMING, or WORD SENSE ANNOTATION.
From Carletta (1996) to Artstein and Poesio (2008), most of these methods have already been discussed and compared in the perspective of CL and we will not do so here. In the domain of TOPIC SEGMENTATION, several measures have been proposed, especially to evaluate the quality of automatic segmentation systems. In most cases, this evaluation consists in comparing the output of these systems with a reference annotation. We mention them here because their use tends to be extended to interannotator agreement because of the lack of dedicated agreement measures, as illustrated by Artstein and Poesio (2008), who mention these metrics in a survey related to interannotator agreement, or by Kazantseva and Szpakowicz (2012).
In this domain, annotations consist of boundaries (between topic segments), and the penalty must depend on the distance from a true boundary. Thus, dedicated measures have been proposed, such as WindowDiff (WD hereafter; Pevzner and Hearst 2002), based on Pk (Beeferman, Berger, and Lafferty 1997). WD relies on the following principle: A fixed-sized window slides over the text and the numbers of boundaries in the system output and reference are compared. Several limitations of this method have been demonstrated and adjustments proposed, for example, by Lamprier et al. (2007) or by Bestgen (2009), who recommends the use of the Generalized Hamming Distance (GHD hereafter; Bookstein, Kulyukin, and Raita 2002), in order to improve the stability of the measure, especially when the variance of segment size increases.
Because these metrics are dedicated to the evaluation of automatic segmentation systems, their most serious weakness for assessing agreement is that they are not chance-corrected, but they present another limitation: They are dedicated to segmentation and assume a full-covering and linear tiling of the continuum and only one category of objects (topic segments). This strong constraint makes them unsuitable for unitizing tasks using several categories (ARGUMENTATIVE ZONING), targeting more sporadic phenomena (ANNOTATION OF COMMUNICATIVE BEHAVIOR), or involving more complex structures (NAMED ENTITY RECOGNITION, HIERARCHICAL TOPIC SEGMENTATION, TOPIC TRANSITION, DISCOURSE FRAMING, ENUMERATIVE STRUCTURES). 3.4.1 Using Measures for Categorization to Measure Agreement on Unitizing. Because of the lack of dedicated measures, some attempts have been made to transform the task of unitizing into a task of categorizing in order to use well-known coefficients such as κ.
They consist of atomizing the continuum by considering each segment as a sequence of atoms, thereby reducing a unitizing problem to a categorization problem. This is illustrated by Figure 4, where real unitizing annotations are on the left (with two annotators), and the transformed annotations are on the right. To do so, an atom granularity is chosen—for instance, in the case of texts, it may be character, word, sentence, or paragraph atoms. Then, each unit is transformed into a set of items labeled with the category of this unit, and a new “blank” category is added in order to emulate gaps between units.
In most cases, this method has severe limitations:
1. Two contiguous units seen as one. In zone (1) of the left part of Figure 4, one annotator has created two units (of the same category), and the other annotator has created only one unit covering the same space. However, once the continua are discretized, the two annotators seem to agree on this zone (with the four same atoms), as we can see in the right part of the figure.
2. False positive/negative disagreement and slight positional disagreement considered as the same. Zone (2) of Figure 4 shows a case where annotators disagree on whether there is a unit or not, which is quite a severe disagreement, whereas zone (3) shows a case of a slight positional disagreement. Surprisingly, these two discrepancies are counted with the same severity, as we can see in the right side of the figure, because in each case, there is a difference of category for one item only (respectively, “blank” with “blue” in case (2), and “blue” with “blank” in case (3)).
3. Agreement on gaps. Because of the discretization, artificial blank items are created, with the result that annotators may agree on “blanks.” The more gaps in real annotations, the more artificial “blank” agreement, and hence the greater the artificial increase in global agreement. Indeed, the expected agreement is less impacted by artificial “blanks,” and it may even decrease.
4. Overlapping and embedding units are not possible. This results because of the discretizing process, which requires a given position to be assigned a single category (or it would require creating as many artificial categories as possible combinations of categories).
This kind of reduction is used to evaluate the annotation of COMMUNICATIVE BEHAVIOR in video recordings by Reidsma (2008), where unitizing is, on the contrary, clearly required: The time-line is discretized (atomized), then κ and α are computed using discretized time spans as units. It should be noted that Reidsman, Heylen, and Ordelman (2006, page 1119) and Reidsma (2008) claim that this “fairly standard” method (which we call discretizing measure henceforth) has certain drawbacks, such as the fact that it “does not compensate for differences in length of segments,” whereas “short segments are as important as long segments” in their corpus (which is an additional limitation to the ones we have just mentioned). They propose a second approach relying on an alignment, as we mention in Section 4.2.1.
This reduction is also unacceptable for other annotation tasks. For instance, in the perspective of DISCOURSE FRAMING, two adjacent temporal frames should not be aggregated in a larger one. In the same manner, for TOPIC SEGMENTATION, it clearly makes no sense to aggregate two consecutive segments.
3.4.2 A Measure for Unitizing Without Chance Correction. Another approach, derived from Slot Error Rate (Makhoul et al. 1999), presented in Galibert et al. (2010), and called SER below, was more specifically used in the context of evaluation of NAMED ENTITY recognition systems.
Comparing a “hypothesis” to a reference, this metric counts the costs of different error types: error “T” on type (i.e., category) with cost 0.5, error “B” on boundaries
(i.e., position) with cost 0.5, error “TB” on both type and boundaries with cost 1, error “I” of insertion (i.e., false positive) with cost 1, and error “D” of deletion (false negative) with cost 1. The overall cost relies on an alignment of objects from reference and hypothesis, which is chosen to minimize this cost. The final value provided by SER is the average cost of the aligned pairs of units—0 meaning perfect agreement, 1 roughly meaning systematic disagreement. An example is given in Figure 5.
This attempt to extend Slot Error Rate to unitizing suffers from severe limitations. In particular, all positioning and categorizing errors have the same penalty, which may be a serious drawback for annotation tasks where some fuzziness in boundary positions is allowed, such as TOPIC SEGMENTATION, TOPIC TRANSITION, or DISCOURSE FRAMING. Moreover, it is difficult to interpret because its output is surprisingly not upper bounded by 1 (in the case where there are many false positives). Additionally, it was initially designed to compare an output to a reference, and so requires some adjustments to cope with more than 2 annotators. Last but not least, it is not chance corrected.
3.4.3 Specific Measures for Unitizing. To our knowledge, the family of α measures proposed by Krippendorff is by far the broadest attempt to provide suitable metrics for various annotation tasks, involving both categorization and unitizing.
In the survey by Artstein and Poesio (2008, page 581), some hope of finding an answer to unitizing is formulated as follows: “We suspect that the methods proposed by Krippendorff (1995) for measuring agreement on unitizing may be appropriate for the purpose of measuring agreement on discourse segmentation.” Unfortunately, as far as we know, its usage in CL is rare, despite the fact that it is the first coefficient that copes both with unitizing and categorizing at the same time, while taking chance into account. The family of α measures would then be suitable for annotation tasks related, for example, to COMMUNICATIVE BEHAVIOR or DIALOG ACTS.
We will therefore pay special attention to Krippendorff’s work in this article, because it constitutes a very interesting reference to compare with, both in terms of theoretical choices and of results. Let us briefly recap Krippendorff’s studies on unitizing from 1995 to 2013 and introduce some of the α measures, which will be discussed in this article. The α coefficient (Krippendorff 1980, 2004, 2013), dedicated to agreement measures on categorization tasks, generalizes several other broadly used statistics and allows various categorization values (nominal, ordinal, ratio, etc.). Besides this wellknown α measure, which copes with categorizing, a new coefficient called αU has been proposed since 1995 in Krippendorff (1995) and then Krippendorff (2004), which can apply to unitizing. Recently, Krippendorff (2013, pages 310, 315) proposed a new version of this coefficient, called uα, “with major simplifications and improvements over
previous proposals,” and which is meant to “assess the reliability of distinctions within a continuum—how well units and gaps coincide and whether units are of the same or of a different kind.” To supplement uα, which mainly focuses on positioning, Krippendorff has proposed c|uα (Krippendorff 2013), which ignores positioning disagreement and focuses mainly on categories.
These measures will be discussed in the following sections. For now, it must be noted that uα and c|uα are not currently designed to cope with embedding or free overlapping between the units of the same annotator. These metrics are then unsuitable for annotation tasks such as, for instance, TOPIC TRANSITION, HIERARCHICAL TOPIC SEGMENTATION, DISCOURSE FRAMING, or ENUMERATIVE STRUCTURES. To conclude the state of the art, we draw up a final overview of the coverage of the requirements by the different measures in Table 2. The γ measure, introduced in the next section, aims at satisfying all these needs. 4 the proposed method: introducing γ : The basic idea of this new coefficient is as follows: All local disagreements (called disorders) between units from different annotators are averaged to compute an overall disorder. However, these local disorders can be computed only if we know for each unit of a given annotator, which units, if any, from the other annotators it should be compared with (via what is called unitary alignment)—that is to say, if we can rely on a suitable alignment of the whole (called alignment). Because it is not possible to get a reliable preconceived alignment (as explained in Section 4.2.1), γ considers all
possible ones, and computes for each of them the associated overall disorder. Then, γ retains as the best alignment the one that minimizes the overall disorder, and the latter value is retained as the correct disorder. To obtain the final agreement, as with the familiar kappa and alpha coefficients, this disorder is then chance-corrected by a so-called expected disorder, which is calculated by randomly resampling existing annotations.
First of all, we introduce three main principles of γ in Section 4.2. We introduce in Section 4.3 the basic definitions. The comparison of two units (depending on their relative positions and categories) relies on the concept of dissimilarity (Section 4.4). A unitary alignment groups at most one unit of each annotator, and a set of unitary alignments covering all units of all annotators is called an alignment (Section 4.5). The disorder associated with a unitary alignment results from dissimilarities between all its pairs of units, and the disorder associated with an alignment depends on those of its unitary alignments (Section 4.6). The alignment having the minimal disorder (Section 4.7) is used to compute the agreement value, taking chance correction into account (Section 4.8). 4.2.1 Measuring and Aligning at the Same Time: γ is Unified. For a given phenomenon identified by several annotators, it is necessary to provide an agreement measure permissive enough to cope with a double discrepancy concerning its position in the continuum, and the category attributed to the phenomenon.
Because of discrepancy in positioning, it is necessary to provide an agreement measure with an inter-annotator alignment, which shows which unit of a given annotator corresponds, if any, to which unit of another annotator. If such an alignment is provided, it becomes possible, for each phenomenon identified by annotators, to determine to what extent the annotators agree both on its categorization and its positioning. This quantification relies on a certain measure (called dissimilarity hereafter) between annotated units: The more the units are considered as similar, the lesser the dissimilarity.
But how can such an alignment be achieved? For instance, in Figure 6, aligning unit A1 of annotator A with unit B1 of annotator B consists in considering that their properties are similar enough to be associated: annotator A and annotator B have accounted for the same phenomenon, even if in a slightly different manner. Consequently, to operate, the alignment method should rely on a measure of distance (in location, in category assignment, or both) between units.
Therefore, agreement measure and aligning are interdependent: It is not possible to correctly measure without aligning, and it is not possible to align units without measuring their distances. In that respect, measuring and aligning cannot constitute two successive stages, but must be considered as a whole process. This interdependence
reflects the unity of the objective: Establishing to what extent some elements, possibly different, may be considered as similar enough either to quantify their differences (when measuring agreement), or to associate them (when aligning).
Interestingly, Reidsma, Heylen, and Ordelman (2006, page 1119), not really satisfied by the use of the discretizing measure as already mentioned, “have developed an extra method of comparison in which [they] try to align the various segments.” This attempt highlights the necessity to rely on an alignment. Unfortunately, the way the alignment is computed, adapted from Kuper et al. (2003), is disconnected from the measure itself, being an ad hoc procedure to which other measures are applied.
4.2.2 Aligning Globally: γ is Holistic. Let us consider two annotators A and B having respectively produced unit A5, and units B4 and B5, as shown in Figure 7. When considering this configuration at a local level, we may consider, based on the overlapping area for instance, that A5 fits B5 slightly better than B4. However, this local consideration may be misleading. Indeed, Figure 8 shows two larger configurations, where A5, B4, and B5 are unchanged from Figure 7. With a larger view, the choice of alignment of A5 may be driven by the whole configuration, possibly leading to an alignment with B4 in Figure 8a, and with B5 in Figure 8b: Alignment choices depend on the whole system and the method should consequently be holistic.
4.2.3 Accounting for Different Severity Rates of Errors: Positional and Categorial Permissiveness of γ. As far as positional discrepancies between annotators are concerned, it is important for a measure to rely on a progressive error count, not on a binary one: Two positions from two annotators may be more or less close to each other but still concern the same phenomenon (partial agreement), or may be too far to be considered as related to the same phenomenon (no possible alignment). For instance, for segmentation, specific measures such as GHD or WD rely on a progressive error count for positions, with an upper limit being half the average size of the segments. For unitizing, Krippendorff
considers with uα that units can be compared as long as they overlap. However, γ considers that in some cases, units by different annotators may correspond to the same phenomenon though they do not intersect. We base this claim on two grounds. First, if we observe the configuration given in Figure 9, annotators 2 and 3 have both annotated part of the NAMED ENTITY that has been annotated by annotator 1. Consequently, though they do not overlap, their units refer to the same phenomenon. In addition, we find a direct echo of this assumption in Reidsma (2008, pages 16–17) where, in a video corpus concerning COMMUNICATIVE BEHAVIOR, “different timing (non-overlapping) [of the same episode] was assigned by [...] two annotators.” Regarding categorization, some available measures consider all disagreements between all pairs of categories as equal. Other coefficients, called weighted coefficients (see Artstein and Poesio 2008), as well as γ, consider on the contrary that mismatches may not all have the same weight, some pairs of categories being closer than others. This closeness is often referred to as overlap.
In our terminology, we call category-overlapping this closeness between categories, and overlap means positional overlap. For example, within annotation efforts related to WORD SENSE or DIALOG ACTS, it is clear that disagreements on labels are not all alike. Given a multi-annotated continuum t:r let A = {a1, ..., an} be the set of annotatorsr let n = |A| be the number of annotatorsr let U be the set of units from all annotatorsr ∀i ∈ J1, nK, let xi be the number of units by annotator ai for t r let x = n∑i=1 xin be the average number of annotations per annotatorr for annotator a = ai, ∀j ∈ J1, xiK, we note uaj unit from a of rank j Annotation set: An annotation set s is a set of units attached to the same continuum and produced by a given set of annotators. Corpus: A corpus c is defined with respect to a given annotation effort, and is composed of a set of continua, and of the set of annotations related to these continua.
Unit: A unit u bears a category denoted cat(u), and a location given by its two boundaries, each of them corresponding to a position in the continuum, respectively denoted start(u) and end(u), start and end being functions from U to N+.
Equality between units is defined as follows:
∀(u, v) ∈ U2, u = v⇔  cat(u) = cat(v)start(u) = start(v)end(u) = end(v) We introduce here the first brick to build the notion of disorder, which works at a very local level, between two units. A dissimilarity tells to what degree two units should be considered as different, taking into account such features as their positions, their categories, or a combination of the two.
A dissimilarity is a function d : U2 → R+, so that :
∀(u, v) ∈ U2, {
d(u, v) = d(v, u) (d is symmetric) u = v⇒ d(u, v) = 0
A dissimilarity is not necessarily a distance in the mathematical sense of the term, in particular because triangular inequality is not mandatory (for instance, in Figure 10, d(A1, B2) > d(A1, C1) + d(C1, B2)).
4.4.1 Empty Unit u∅, Empty Dissimilarity ∆∅. As we will see, γ relies on an alignment of units by different annotators. In particular, this alignment indicates for unit ua1i of annotator a1, to which unit u a2 j of annotator a2 it corresponds, in order to compute the associated dissimilarity. In some cases, though, the method will choose not to align ua1i with any unit of annotator a2 (none corresponds sufficiently). We define the empty pseudo unit, denoted u∅, which corresponds to the realization of this phenomenon: ultimately, a pseudo unit u∅ is added to the annotations of a2, and u a1 i is aligned with it.
We also define the associated cost ∆∅:
∀u ∈ U , d(u, u∅) = d(u∅, u) = ∆∅ and d(u∅, u∅) = ∆∅
Dissimilarities should be calibrated so that ∆∅ is the value beyond which two compared units are considered critically different. Consequently, it constitutes a reference, and dissimilarities will be expressed in this article as multiples of ∆∅ for better clarity. It is not a parameter of gamma, but a constant (which is set to 1 in our implementation).
4.4.2 Positional Dissimilarity dpos. Different positional dissimilarities may be created, in order to deal with different annotation tasks. In this article, we use the dissimilarity shown in Equation 3, which is very versatile.
dpos−sporadic(u, v) = ( |start(u)− start(v)|+ |end(u)− end(v)| (end(u)− start(u)) + (end(v)− start(v)) )2 ·∆∅ (3)
Equation (3) sums the differences between the right and left boundaries of both units in its numerator. Its denominator sums the lengths of both units, so that this dissimilarity is not scale-dependent. Squaring the value is an option used here to accelerate dissimilarity when differences of positions increase. It is illustrated in Figure 10 with different configurations and their associated values, from 0 for the perfectly aligned pair of units (A1, B1) to 22.2 ·∆∅ for the worst pair (A1, C2).
4.4.3 Categorial Dissimilarity dcat. Let K be the set of categories. For a given annotation effort, |K| different categories are defined. For more convenience, we first define categorial distance between categories distcat via a square matrix of size |K|, with each category appearing both in row titles and column titles. Each cell gives the distance between two categories through a value in [0, 1]. Value 0 means perfect equality, whereas the maximum value 1 means that the categories are considered as totally different. As distcat is symmetric, such a matrix is necessarily symmetric, and bears 0 in each diagonal cell. Table 3 gives an example for three categories, and shows that an association between a unit in category cat1 with one in category cat3 is the worst possible (distance = 1), whereas it is half as much between cat1 with cat2 (distance = 0.5). This makes it possible to take into account so-called category-overlapping (in our example, cat1 and cat2 are said to overlap, which means they are not completely different), as weighted coefficients such as κw or α already do. Note that in the case of so-called “nominal categories,” the matrix will be full of 1 outside the diagonal, and full of 0 in the diagonal (different categories are considered as not matching at all).
This categorial distance matrix is then used to build the categorial dissimilarities, taking into account the ∆∅ value. We define categorial dissimilarity between two units by:
dcat(u, v) = fcat(distcat (cat(u), cat(v))) ·∆∅ (4)
Function fcat can be used to adjust the way the dissimilarity grows with respect to the categorial distance values. The standard option2 (used in this article) is to simply consider fcat(x) = x, with which dcat naturally increases gradually from zero when categories match, to ∆∅ when categories are totally different (distcat(cat(u), cat(v)) = 1 =⇒ dcat(u, v) = ∆∅).
4.4.4 Combined Dissimilarity dcombi. Because in some annotation tasks units may differ both in position and in category, it is necessary to combine the associated dissimilarities so that all costs are cumulated. This is provided by a combined dissimilarity.
Let d1 and d2 be two dissimilarities. We define:
dα,βcombi(d1,d2 )(u, v) = α.d1(u, v) + β.d2(u, v) (5)
It is easy to demonstrate that this linear combination of dissimilarities is itself a dissimilarity (if (α,β) 6= (0, 0)). It enables the same weight to be assigned to positions and categories using d1,1combi(dpos,dcat ), which is currently used for γ.
Then, we can note that it is the same cost ∆∅ for a unit either not to be aligned with any other one, or to be aligned with a unit in the same configuration as (A1, C1) of Figure 10 (if they have the same category), or to be aligned with a unit having an incompatible category (if they occupy the same position). Unitary alignment ă. A unitary alignment ă is an i-tuple, i belonging to J1, nK (n the number of annotators), containing at most one unit by each annotator: It represents the hypothesis that i annotators agree to some extent on a given phenomenon to be unitized. In order to make all unitary alignments homogenous, we eventually complete any unitary alignment that is an i-tuple with n− i empty units u∅, so that all unitary alignments are ultimately n-tuples. Figure 11 illustrates unitary alignments with some u∅ units.
Alignment ā. For a given annotation set, an alignment ā is defined as a set of unitary alignments such that each unit of each annotator belongs to one and only one of its
2 Another option is, for example, to use fcat(x) = −ln(1− x)x30 + x, which is a function almost similar to fcat(x) = x on the [0, 0.9] range, and reaches∞ near 1. Then, when the categorial distance is equal to 1, the categorial dissimilarity reaches infinity, which guarantees that the units cannot be aligned.
unitary alignments. Mathematically, it constitutes a partition of the set of units (if we do not take u∅ into account). 4.6.1 Disorder of a Unitary Alignment. The disorder of a unitary alignment ă, denoted δ̆(ă), is defined for a given dissimilarity d as the average of the one-to-one dissimilarities of its units:
δ̆(ă) = 1 C2n · ∑ (u,v)∈ă2 d(u, v) (6)
Averaging dissimilarities rather than summing them makes the result independent of the number of annotators.
4.6.2 Disorder of an Alignment. The disorder of an alignment ā, denoted δ̄(ā), is the sum of the disorders of all its unitary alignments divided by the mean number of units per annotator:
δ̄(ā) = 1x · |ā|∑
i=1
δ̆(ăi) (7)
We chose to consider the average value rather than the sum so that the disorder does not depend on the size of the continuum. Best alignment â. An alignment ā of the annotation set s is considered as the best (with respect to a dissimilarity) if it minimizes its disorder among all possible alignments of s. It is denoted â. The proposed method is holistic in that it is necessary to take into account the whole set of annotations in order to determine each unitary alignment.
Disorder of an annotation set δ(s). The disorder of the annotation set s, denoted δ(s), is defined as the disorder of its best alignment(s) δ̄(â). Note that it may happen that several alignments produce the lowest disorder.
We have just presented the two crucial definitions of our new method, which make it “unified.” Indeed, the best alignment is chosen with respect to the disorder,
therefore with respect to what computes the agreement measure; and, conversely and simultaneously, the resulting agreement value (see below) is given by the best alignment: agreement computation and alignment are fully intertwined, whereas in most agreement metrics, the alignment is fixed a priori or no alignment is used. 4.8.1 The Model of Chance of γ. As we have already mentioned in the state of the art, it is necessary for an inter-annotator agreement measure to provide chance correction. We have also seen that there are several chance correction models, and that it is a controversial question. However, for γ, we follow Krippendorff, who claims that annotators should be interchangeable, because, as stressed by Krippendorff (2011) and Zwick (1988), Cohen’s definition of expected agreement (using individual distributions) numerically rewards annotators for not agreeing on their use of values, that is to say when they have different prevalences of categories, and punishes those that do agree. Therefore, expected values of γ are computed on the basis of the average distribution of observed annotations of the several annotators.
More precisely, we define the expected (chance) disorder as the average disorder of a multi randomly annotated continuum where:r The random annotations fulfill the observed annotation distributions for the
following features:
– the distribution of the number of units per annotator
– the distribution of categories
– the distribution of unit length per category
– the distribution of gaps’ length
– the distribution of overlapping and/or covering between each pair of categories (for instance, units of categories A and B may never intersect, 7% of the units of category A may cover one unit of category C, and so on).r The number of random annotators is the same as the number of annotators
in the observed data
4.8.2 Two Possible Sources to Build Chance: Local Data versus Corpus Data. In addition, whereas other studies systematically compute the expected value on the data also used to compute the observed value (see Section 3.1), we consider that it should be computed, when possible (that is to say, when several continua have been annotated with the same set of categories and the same instructions), from the distribution observed in all continua of the annotation effort the evaluated continuum comes from: If distribution changes from one continuum to another one, it is more because of the content of each continuum than because of chance. Let us illustrate this by a simple example, where two annotators have to annotate several texts from a sentiment analysis point of view, using three available categories: positive, negative, and neutral. On average, on the whole corpus, we assume that the prevalence is 13 for each category. The expected agreement on the whole corpus is thus 0.33. We also assume that for one particular text, there are only positive and neutral annotations, 50% of each, and no negative one. The expected agreement for this particular text is 0.5, which means that
this particular text is considered to facilitate agreement by chance, and which has the consequence that the final agreement will be more conservative than for the rest of the corpus. Why does the third category, “negative,” not appear in this expected agreement computation? This conception of chance considers that when an annotator begins to annotate this particular text, which she does not already know, the third category no longer exists in her mind, and that it is the case for every other annotator, whereas they are not supposed to cooperate. It cannot be by chance that all annotators use one category in some texts, and not in another one, but because of the content, and of the interpretation, of a given text. For this reason, from our point of view, it is better to take into account the data observed on a whole annotation effort rather than on each individual continuum. The complete data tell more about the mean behavior of annotators, whereas data of a given continuum may depend more on the particularities of its content.
As a consequence, γ provides two ways to compute the expected values: one which considers only the data of the continuum being evaluated, as does every other coefficient; and a second one, which considers the data from all continua of the annotation effort the evaluated continuum comes from. When available, we recommend using the second one, for the reasons already expressed.
4.8.3 Using Sampling to Compute the Expected Value. Expected agreement (or disagreement) is the expected value of a random variable. But which random variable? For coefficients like kappa and alpha, observed agreement (or disagreement) is the mean agreement (or disagreement) on all pairs of instances, so the random variable can be as simple as a random pair of instances (however we interpret “random”). This value can be readily computed. For gamma, however, observed disagreement is determined on a whole annotation, so the random variable needs to be a whole random annotation. The expected value of such a complicated variable is much more difficult to determine analytically. Instead, gamma uses sampling, as introduced in Section 5. Now that the disorder and the expected disorder have been introduced, we can define the agreement measure (of annotation set s belonging to corpus c, with c = {s} if s is a sole annotation set) with Equation 8, which is derived from Equation 2:
∀s ∈ c,γ = 1− δ(s) δe(c)
(8)
If all annotators perfectly agree (Figure 11a), γ = 1. Figure 11c corresponds to the worst case, where the annotators are worse than annotating at random, with γ < 0. Figure 11b shows an intermediate situation. 5 implementation :In this section, we first propose an efficient solution to compute the disorder of an annotated continuum, which relies on linear programming. Second, we propose two ways to generate random annotated continua (with respect to the observed distributions) to compute the expected disorder, one relying on a single continuum, the other one relying on a corpus (i.e., several continua). Third, we determine the number of random data sets
that we must generate (and compute the disorder of) to obtain an accurate value of the expected disorder. In order to simplify the discussion and the demonstrations, we consider in this section that n annotators all made the same number of annotations p.
The proposed method has now been fully described on a theoretical level, but, being holistic, its software implementation leads to a major problem of complexity. One can demonstrate that there are theoretically (p!)n−1 possible alignments. However, we will (1) show how to reduce the initial complexity, and (2) provide an efficient linear programming solution.
5.1.1 Reducing the Initial Complexity. The initial number of possible unitary alignments (which are used to build a possible alignment) is pn. Fortunately, theorem provided as Equation (9) states that any unitary alignment with a cost beyond the value n ·∆∅ cannot belong to the best alignment, and so can be discarded. Indeed, any unitary alignment with a cost above ∆∅ can be replaced by creating a separate unitary alignment for each unit (of cost ∆∅ per unitary alignment, so of total cost n ·∆∅).
Demonstration. Consider the best alignment â, of cardinality m. Let ă be any of its unitary alignments. For convenience, we attribute to it the index 1 (ă = ă1), while the others are indexed from 2 to m. This unitary alignment ă contains n units (either real or u∅). For each of these units ui (1 ≤ i ≤ n), we create the unitary alignment ăm+i = (ui, u∅, ..., u∅) of cardinality n. It is possible to create an alignment ā made up of the set of unitary alignments of â \ {ă}, to which we add the unitary alignments ăm+1 to ăm+n that we have just created.3 It is of cardinality m + n− 1. Because â minimizes the disorder, we obtain:
δ̄(â) ≤ δ̄(ā)⇒ 1x m∑ i=1 δ̆(ăi) ≤ 1x m+n∑ i=2 δ̆(ăi)
⇒ m∑
i=1 δ̆(ăi) ≤ m+n∑ i=2 δ̆(ăi)
⇒ δ̆(ă1) ≤ m+n∑
i=m+1
δ̆(ăi)
since ∀i > m, δ̆(ăi) = 1C2n (C 2 n∆∅) = ∆∅, and since we have denoted ă = ă1,
⇒ δ̆(ă) ≤ n ·∆∅ (9)
Experiments have shown that this theorem allows us to discard about 90% of the unitary alignments.
3 ā is indeed an alignment, because each of its units appears in one and only one unitary alignment.
5.1.2 Finding the Best Alignment: A Linear Programming Solution. Finding the best alignment consists of minimizing the global disorder. Such a problem may be described as a linear programming problem, so that the solution can be computed by a linear programming solver. For convenience, we introduce two new definitions:r Let UA be the set of all unitary alignments.r Let UAu be the set of the unitary alignments which contain unit u.
The description of the problem in linear programming terms is threefold. First, for a given alignment ā, for each possible unitary alignment ăi, we define the Boolean variable Xăi , which indicates if this unitary alignment belongs or not to the alignment:
∀ăi ∈ UA, Xāăi = { 0 iff ăi 6∈ ā 1 iff ăi ∈ ā
Second, we have to express the fact that, by definition, each unit u (of each annotator) should belong to one and only one unitary alignment of the alignment ā, that is to say that among all unitary alignments containing u, exactly one Xăi equals 1, and all the others equal 0:
∀u ∈ U , ∑
ăi ∈ UAu
Xāăi = 1
Third, the goal is to minimize the global disorder δ̄(ā) associated with ā, among all possible alignments ā:
Minimize δ̄(ā) = ∑
ăi ∈ UA δ̆(ăi) · Xāăi
The LPSolve solver4 finds the best solution in less than one second with n = 3 annotators and p = 100 annotations per annotator on a current laptop (once the initial complexity has been reduced thanks to the previous theorem), which is fast enough to be practical. The next two subsections detail two strategies to generate randomly annotated continua with respect to the definition of the expected disorder of γ, and the third subsection explains how to choose the number of expected disorder samples to generate so that their average is an accurate enough value of the theoretical expected value. The two strategies correspond to the need expressed in Section 4.8.1 to compute the expected value on the largest set of available data, either a single continuum, or, when available, several continua from the same corpus.
4 http://lpsolve.sourceforge.net.
5.2.1 A Strategy to Compute the Expected Disorder Using a Single Continuum. When the annotation effort is limited to a single continuum, we can only rely on the annotated continuum itself to compute the expected value. To create random annotations that fulfill the observed distributions, the implemented strategy is as follows: We take the real annotated continuum of an annotator (such as the example shown on the left in Figure 12), choose at random a position on this continuum, split the continuum at this position, and permute the two parts of the split continuum. Three examples of split and permutation are shown in the right part of the figure, for split positions of, respectively, 15, 24, and 38, all coming from the same real continuum, with units that are no longer aligned (except by chance). However, we have to address the fact that some units may intersect with themselves, generating some part of agreement beyond chance. For instance, in Figure 12, unit 3 intersects with itself between #15 and #24, because the length of the unit, 12, is higher than the difference of shifts 24− 15 = 9. To limit this phenomenon, we do not allow the distance between two shifts to be less than the average length of units.
5.2.2 A Strategy to Compute the Expected Disorder Using Several Continua (from the Same Corpus). This strategy consists of mixing annotations coming from different continua, so that their units may align only by chance. To create a random annotation of n annotators, we randomly choose n different continua of the corpus, and pick the annotations of one annotator (randomly chosen) of each of these continua. When different texts are of different lengths, each of them is adjusted to the longest one by duplicating as many times as necessary (like a mosaic).
This is shown in Figure 13 for n = 3 annotators. We assume the corpus contains eight continua, each annotated by three annotators. To generate a random set of three annotations, we have randomly selected a combination of three values between 1 and 8, here (2, 4, 7), to select three different continua among the eight available ones of the corpus. Then, for each of these selected continua, we choose one annotator, here annotator 2 for continuum 2, annotator 3 for continuum 4, and annotator 1 for continuum 7. We combine the associated annotations as shown in the right part of the figure, and obtain a set of random annotations that fulfill (on average) the observed distributions. The (very limited) extent of the resulting agreement we can see in this example (only two units have matching categories, but with discrepancies in position) is only by chance because the compared annotations come from different continua.
In addition, it is possible to create a great number of random sets of annotations with this strategy: With n annotators and m continua (m ≥ n), it is possible to generate up to Cnm · nn different combinations. For instance, in our example, which assumes n = 3 and m = 8, there are 56× 33 = 1512 combinations to create random annotations. Because the expected disorder is by definition randomly obtained on average, and because there is virtually an infinite number of possible random annotations (with a discrete and finite continuum, it is not really infinite, but still too big to be computed), we can only compute a reduced but sufficient number of experiments and obtain an approximate value of the expected disorder. This is a sampling problem as described, for example, in Israel (1992). What statistics provide is a way to determine the minimal number n0 of experiments to do (and to average) so that we get an approximate result of a given precision with a given confidence level. It consists in: First, taking a small sample to estimate the mean and standard deviation; then, using these estimates to determine the sample size n0 that is needed.
We follow the strategy provided in Olivero (2001) to compute a disorder value that differs less than e = 2% from the real value with a (1− α) = 95% confidence (the software distribution we provide is set by default with these values).
First, we consider a sample of chance disorder values of size n = 30. Let µ be the sample mean, and σ′ be its standard deviation. µ is directly an unbiased estimator of the population mean, and σ = √ ( nn−1 ) · σ′ is an unbiased estimator of the real standard deviation. Let Cv = σµ be the coefficient of variation (i.e., the relative standard deviation). Let U1−α2 be the abscissa of the normal curve that cuts off an area α at the tails. This value is provided in statistical tables. We get n0 by the following equation:
n0 = (Cv ·U1−α2
e )2 Let us consider a real example. We generate a sample of random disorders of size n = 30. We compute its mean µ = 3.49, its standard deviation σ′ = 0.1379, hence σ = 0.1403, and Cv = 0.040188. We get U1− 0.052 = 1.96 from the corresponding available table, hence we obtain n0 = 15.5. This means that a sample of 16 disorder values gives 2% of precision with 95% confidence. The mean we have already computed with 30 values fulfills this condition, and is a good approximation of the real expected disorder. If we wish to obtain a high precision of 1%, we need n0 = 62. It is beyond the
initial size of our sample (which is 30), and we will have to generate an additional set of 32 values in order to obtain the required number. 6 comparing and benchmarking γ :As γ is an entirely new agreement measure method, it is necessary to analyze how it compares with some well known and much studied methods. First, we carry out a thorough comparison between γ and the two dedicated alphas, uα and c|uα, which are the most specific measures in the domain. Second, we benchmark γ by comparing it with other main measures, thanks to a special tool that is briefly introduced. As already mentioned, Krippendorff’s uα and c|uα are clearly the most suitable coefficients for combined unitizing and categorizing. To better understand the pros and cons as well as the behavior of these measures compared with γ, we first explain how they are designed in Section 6.1.1, and then make thorough comparisons with γ from Section 6.1.2 to 6.1.6 including: (1) how they react to slight categorial disagreements, (2) interlacement of positional and categorial disagreements, (3) the impact of the size of the units on positional disagreement, (4) split disagreements, and (5) the impact of scale (e.g., if the size of all units is multiplied by 2). We finish by showing a paradox of uα in Section 6.1.7.
6.1.1 Introducing uα and c|uα. To introduce how these two coefficients work, let us consider the example taken from Krippendorff (2013), shown in Figure 14. The length of the continuum is 76, there are two annotators, and there are four possible categories, numbered 1 to 4.
The uα coefficient basically relies on the comparison of all pairs of sections among annotators, a section being either a categorized unit or a gap. To get the observed disagreement value uDo, squared lengths of the unmatching intersections are summed, and this sum is then divided by the product of the length of the continuum and m(m− 1), m being the number of annotators. In the example, mismatches occur around the second and third units of the two annotators. From left to right, there are the following intersections: cat 1 with gap (l = 10), cat 1 with cat 3 (l = 5), gap with cat 3 (l = 8), cat 2 with cat 1 (l = 5), and cat 2 with gap (l = 5). This leads twice (by symmetry) to the sum 102 + 52 + 82 + 52 + 52, and so the observed disagreement uDo = 2(102+52+82+52+52 )
76·2(2−1) = 3.145. The expected value uDe is obtained by considering all the possible positional combinations of each pair, and not only the observed ones. This means that for a given pair, one of the two units is virtually slid in front of the other in all
possible ways, and the corresponding values are averaged. In this example, uDe = 5.286. Therefore, uα = 1− 3.1455.286 = 0.405.
Coefficient c|uα relies on a coincidence matrix between categories, filled with the sums of the lengths of all intersections of units for each given pair of categories. For instance, in the example, the observed coincidence between category 1 and category 3 is 5, and so on. A metric matrix is chosen for these categories, for instance, an interval metric (for numerical categories), which says that the distance between category i and category j is (i− j)2. Hence, the cost for a unitary intersection between categories 1 and 2 is (1− 2)2 = 1, but is 22 = 4 between categories 1 and 3, and so on. Then, the observed disagreement is computed according to these two matrices. To finish, an expected matrix is filled (in a way which cannot be detailed here due to space constraints), and the expected value is computed the same way. In the example, c|uα = 1− 0.8333.145 = 0.744.
Hence, Krippendorff’s alphas provide two clues to analyze the agreement between annotators. In the example, uα = 0.405 indicates that the unitizing is not so good, but also that the categorizing is much better, with c|uα = 0.744 (even though of course, these two values are not independent, since unitizing and categorizing coexist here by nature).
Now that these coefficients have been introduced in detail, let us analyze to what extent they differ from γ.
6.1.2 Slight Categorial Disagreements: Alphas Versus γ. When annotators have slight categorial disagreements (with overlapping categories), c|uα is slightly lowered. However, uα does not take categorial overlapping into account, but has a binary response to such disagreements, and is lowered as much as if they were severe categorial disagreements. A consequence of this approach is illustrated in Figure 15, where two annotators perfectly agree both on positions and categories in the experiment on the left, and still perfectly agree on position but slightly diverge concerning categories in the experiment on the right (1/2, 6/7, and 8/9 are assumed to be close categories). However, uα drops from 1 in the left experiment to –0.34 (a negative value means worse than random) in the right experiment, despite, in the latter, the positions being all correct, and the categories being quite good, since c|uα = 0.85. On such data, γ considers that there is no positional disagreement, and c|uα and γ both consider that there are slight categorial disagreements.
6.1.3 Positional Disagreements Impacting Categorial Agreement: c|uα. Two different conceptions of how to account for categorial disagreement have, respectively, led to c|uα and γ: c|uα relies on intersections between the units of different annotators, which is basically equivalent to an observation at the atom level, whereas γ relies on alignments between units (any unit being finally attached and compared to, at most, only one other) based both on positional and categorial observation. Hence, in a configuration such as the one given in Figure 16, where two annotators annotated three units with the same categories
1
1
9
9
7
7
Observer 1
Observer 2
1
2
9
8
7
6
Observer 1
Observer 2
Figure 15 Consequences of no categorial disagreement (left) compared with slight categorial disagreements (right).
1, 4, and 2, but not exactly at the same locations; c|uα considers a certain account of categorial disagreements, whereas γ does not. According to the principles of c|uα, any part of the continuum (even at the atom level) with an intersection between different categories means some confusion between them, whereas γ considers here that the annotators fully agree on categories (they both observed a “1” then a “4” then a “2” with no confusion), and disagree only on where phenomena exactly start and finish. The crucial difference between the two methods is probably whether we consider units to be non-atomizable (and therefore consider alignments, as γ does), or atomizable (in which case two different parts of a given unit may be simultaneously and respectively compared to two different units from another annotator).
6.1.4 Disagreements on Long versus Short Spans of Texts. Here again, the way disagreements are accounted for may differ markedly between uα and γ: when a unit does not match with any other, uα takes into account the length of the corresponding span of text to assess a disagreement. As shown in Figure 17, an orphan unit of size 10 will cost 100 times as much as an orphan unit of size 1, whereas for γ, they will have the same cost. In the whole example in Figure 17, to compute the observed disagreements, uα says the first case is 50 times worse than the second, whereas γ says on the contrary that the second case is twice as bad as the first. Here, γ fulfills the need (already mentioned) expressed by Reidsma, Heylen, and Ordelman (2006, page 3) to consider that “short segments are as important as long segments.” This phenomenon is the same for categories between c|uα and γ, the size of the units having consequences only for c|uα.
6.1.5 Split Disagreements. Sometimes, an annotator may divide a given span of text into several contiguous units of the same type, or may annotate the same span with one whole unit. In these cases, c|uα computes its observed disagreement the same in both
configurations, and uα assigns decreasing disagreement when splitting increases, as shown in Figure 18, whereas γ assigns increasing disagreements.
Moreover, in Figure 19, the observed uα is not responsive to splits at all, whereas γ is still responsive.
6.1.6 Scale Effects. The way uα computes dissimilarities is directly proportional to squared lengths, as shown in Figure 20. On the other hand, γ may use any positional dissimilarity, and usually uses ones that are not scale-dependent for CL applications, such as dpos−sporadic (Equation (3)). For instance, if a text is annotated with two categories, one at word level, the other one at paragraph level, we may prefer to account for relative disagreements so that a missing word will be more heavily penalized in the first case than in the second. In Figure 20, the observed disagreement of uα is 32 = 9 times greater for B units than for A units, but would be the same for γwith dpos−sporadic since:
( 0 + 3( 7+10
2 ))2 = ( 0 + 9( 21+30 2 ))2 . Figure 21a, the annotators disagree on categorization, and have a moderate agreement on unitizing. This configuration leads to uα = 0.107. In Figure 21b, the configuration is quite similar, but now annotators fully agree on unitizing: Each of them puts units in the same positions. Paradoxically, uα drops to −0.287, which is less than in the first configuration. In brief, the reason for this behavior is that in the first case, the computed disagreement regarding a given pair of units is virtually distributed into shorter parts of the whole (an intersection of length 80 between them, and an intersection of length
20 with a gap for each of them, which leads to 802 + 2× 202 = 7, 200) whereas the disagreement is maximum in the second case (an intersection of length 100 with a unit of another category, which leads to 1002 = 10, 000). Contrarily, with similar data, γ provides a better agreement in the second case than in the first one. With its design, it considers that there is the same categorial agreement in both cases, but better positional agreement in the second case, which seems to better correspond to the CL tasks we have considered.
6.1.8 Overlapping Units (Embedding or Free Overlap). Both alpha coefficients are currently designed to cope only with non-overlapping units (the term overlapping also stands here for embedding), which is a limitation for several fields in CL. It is debatable whether they could be generalized to handle overlapping units. It seems that it would involve a major change in the strategy, which currently necessitates comparing the intersections of all pairs of units. In the example shown in Figure 22, even though annotators fully agree on their two units, the alphas will inherently compare A1 with B2 and A2 with B1 (in addition to the normal comparisons between A1 with B1 and A2 with B2), and will count the resulting intersections as disagreements. It is necessary here to choose once and for all what unit to compare to what other, rather than to perform all the comparisons. But making such a choice precisely consists in making an alignment, which is a fundamental feature of γ. Consequently, it seems that the alphas would need a structural modification to cope with overlapping. As explained by Reidsma, Heylen, and Ordelman (2006), because of the lack of specialized coefficients coping with unitizing, a fairly standard practice is to use categorization coefficients on a discretized (i.e., atomized) version of the continuum: For instance, each character (or each word, or each paragraph) of a text is considered as an item; and a standard categorization coefficient such as κ is used to compute agreement. Such a measure is called κd (for discretized κ) hereafter. Several weaknesses of this approach have been already mentioned in the state-of-the-art section. It is interesting to compare such a measure to the specialized one c|uα: even if they both bear the aggregatable hypothesis, they have, however, significant differences (as confirmed by the experiments presented in the next section). The main one is that c|uα does not use an artificial atomization of the continuum, and only compares units with units. In doing so, it is not prone to agreement on blanks, in contrast to κd. Another difference is that, for the same reason, c|uα is not inherently limited to non-overlapping units: Even if it is not currently designed to cope with them, as we have already seen, it is possible to submit overlapping units to this measure (some results are shown in the next section). In this section on benchmarking, we use the Corpus Shuffling Tool (CST) introduced by Mathet et al. (2012) to compare γ concretely and accurately to the other measures.
We first introduce the possible error types that it will provide: category (category mistakes may occur), position (the boundaries may be shifted), false positives (the annotators add units to the reference units), false negatives (the annotators miss some of the reference units), and splits (the annotators put two or more contiguous units instead of a reference unit, which occupy the same span of text).
This tool is used to simulate varying degrees of disagreement among different error types, and the metrics are compared with each other according to how they react to these disagreements. For a given error type, for each magnitude between 0 and 1 (with a step of 0.05), the tool creates 40 artificial, multi-annotator shuffled annotation sets, and computes the different measures for them. Hence, we obtain a full graph showing the behavior of each measure for this error type, with the magnitude on the x-axis, and the average agreement (over the 40 annotation sets) on the y-axis. This provides a sort of “X-ray” of the capabilities of the measures with respect to this error type, which should be evaluated against the following desiderata:r A measure should provide a full response to the whole range of
magnitudes, which means in particular that the curve should ideally start from 1 (at m = 0) and reach 0 (at m = 1), but never go below 0 (indeed, negative agreement values require a part of systematic disagreement, which is not simulated by the current version of the CST).r The response should be strictly decreasing: A flat part would mean the measure does not differentiate between different magnitudes, and, even worse, an increasing part would mean that the measure is counter-effective at some magnitudes, where a worse error is penalized less severely.
We emphasize the fact that the whole graph is important, up to magnitude 1. Indeed, in most real annotated corpora, even when the overall agreement is high, errors
corresponding to all magnitudes may occur. For instance, an agreement of 0.8 does not necessarily correspond to the fact that all annotations are affected by slight errors (which correspond to magnitudes close to 0), but may for instance correspond to the fact that a few units are affected by severe errors (which may correspond to magnitudes close or equal to 1).
It is important to note that this tool was designed by the authors of γ, for tasks where units cannot be considered as atomizable. In particular, it was conceived so that disagreements concerning small units are as important as those concerning large ones. However, it is provided as open-source (see Conclusions section) so that anyone can test and modify it, and propose new experiments to test γ and other measures in the future.
6.3.1 Introducing the CST. The main principle of this tool is as follows. A reference corpus is built, with respect to a statistical model, which defines the number of categories, their prevalence, the minimum and maximum length for each category, and so forth. Then, this reference is used by the shuffling tool to generate a multi-annotator corpus, simulating the fact that each annotator makes mistakes of a certain type, and of a certain magnitude. It is important to remark that the generated corpus does not include the reference it is built from.
The magnitude m is the strength of the shuffling, that is to say the severity of mistakes annotators make compared to the reference. It can be set from 0, which means no damage is applied (and the annotators are perfect) to the extreme value 1, which means annotators are assumed to behave in the worst possible way (but still being independent of each other)—namely, at random.
Figure 23 illustrates the way such a corpus is built: From the reference containing some categorized units, three new sets of annotations are built, simulating three annotators who are assumed to have the same annotating skill level, which is set in this example at magnitude 0.1. The applied error type is position only, that is to say that each annotator makes mistakes only when positioning boundaries, but does not make any other mistake (the units are reproduced in the same order, with the correct category, and in the same number). At this low magnitude, the positions are still close to those of the reference, but often vary a little. Hence, we obtain here a slightly shuffled multiannotator corpus. Let us sum up the way error types are currently designed in the CST.
Position. At magnitude m, for a given unit, we define a value shiftmax that is proportional to m and to the length of the unit, and each boundary of the unit is shifted by a
value randomly chosen between −shiftmax and shiftmax (note: at magnitude 0, because shiftmax = 0, units are not shifted).
Category. This shuffling cannot be described in a few words (see Mathet et al. [2012] for details). It uses special matrices to simulate, using conditional probabilities, progressive confusion between categories, and can be configured to take into account overlapping of categories. The higher the magnitude, the more frequent and severe the confusion.
False negatives. At magnitude m, each unit has the probability m to be forgotten. For instance, at magnitude m = 0.5, each annotator misses (on average) half of the units from the reference (but not necessarily the same units as the other annotators).
False positives. At magnitude m, each annotator adds a certain number of units (proportional to m) to the ones of the reference.
Splits. At magnitude m, each annotator splits a certain number of units (proportional to m). A split unit may be re-split, and so on.
6.3.2 Pure Segmentation: γ, WD, GHD. Even if γ was created to cope with error types that are poorly or not at all dealt with by other methods, and, moreover, to cope simultaneously with all of them (unitizing of categorized and overlapping categories), it is illuminating to observe how it behaves in more specific error types, to which specialized and well known methods are dedicated. We start with pure segmentation. Figure 24 shows the behavior of WD, GHD, and γ for two error types.
For false negatives, WD and GHD are quite close, with an almost linear response until magnitude 0.6. Their drawback is that their responses are limited by an asymptote, because of the absence of chance correction, while γ shows a full range of agreements; for shifts, WD and GHD show an asymptote at about agreement = 0.4, while γ shows values from 1 to 0. This experiment confirms the advantage of using γ instead of these distances for inter-annotator agreement measure.
6.3.3 Pure Categorization. In this experiment, the CST is set to three annotators, four categories with given prevalences. The units are all of the same size, positioned in fixed, predefined positions, so that the focus is on categorizing only. It should be noted that, with such a configuration, α and κ behave exactly in the same way as c|uα. It is particularly striking in Figure 25 that γ behaves in almost the same way as c|uα. In fact, the observed values of these measures are exactly the same, the only difference coming from a slight difference in the expected values, due to sampling. Other tests carried out with the pure categorizing coefficient κ yielded the same results on this particular error type, which means that γ performs as well as recognized measures as far as categorizing is concerned, with two or more annotators. The uα curve goes below zero at magnitude 0.5 (probably for the reasons seen in Section 6.1.7). Moreover, its behavior depends on the size of the gaps: Indeed, with other settings of the shuffling, the curve may, on the contrary, be stuck over zero. κd fails to reach 0 because of the virtual agreement on gaps (but it would if there were no gaps). Lastly, SER (averaging the results of each pair of annotators) is bounded below by 0.6, which results from not taking chance into account.
6.3.4 Almost General Case: Unitizing + Categorizing. This section concerns the more general uses of γ, combining both unitizing and categorizing. However, in order to be compliant with uα, c|uα, and κd, we limit the configurations here so that the units do not overlap at all. In particular, the reference was built with no overlapping units, and we have used a modified version of the shifting shuffling procedure so that the nonoverlapping constraint is fully satisfied, even at high magnitudes.
Positional errors (Figure 26a). An important point is that this shuffling error type, which is based only on moving positions, has a paradoxical consequence on category agreement, since units of different categories align when sufficient shifting is applied. Consequently, c|uα is not blocked at 1, even though it is designed to focus on categories. Additionally, it starts to decrease from the very first shifts, as soon as units from different annotators start overlapping. This is a concrete consequence of what has been formally studied in Section 6.1.3. γ has a most progressive response, reaches 0.1 at magnitude 1,
and is the only measure to be strictly decreasing. SER immediately drops to agreement 0.5 at magnitude 0.05. As it relies on a binary positional distance, it fails to distinguish between small and large errors. This is a serious drawback of such a measure for most CL tasks. Then it goes below zero and is not strictly decreasing. uα is mostly strictly decreasing, but has some increasing parts, and, even more problematic, negative values from 0.6 to 0.9, probably because of the reason explained in Section 6.1.7. κd is too responsive at the very first magnitudes, and is not strictly decreasing, probably because it “does not compensate for differences in length of segments” (Reidsma, Heylen, and Ordelman 2006, page 3).
Positional and categorial errors (Figure 26b). γ is strictly decreasing and reaches 0. The alphas are not strictly decreasing, and once again uα drops below 0 from magnitude 0.6 onwards. κd is not strictly decreasing (again, probably because it “does not compensate for differences in length of segments”), but its general shape is not that far from γ.
Split errors (Figure 27). The split error type would need to create an infinite number of splits to mean pure chaos at magnitude 1. As this is computationally not possible, we restricted the number of splits to five times the number of units of the reference. We should therefore not expect measures to reach 0. In this context, γ shows a good range of responses, from 1 to 0.2, in an almost linear curve. SER is also quite linear, but gives very confusing values for this error type because it reaches negative values above magnitude 0.6. Finally, uα, c|uα, and κd are not responsive at all to this error type, as expected, and remain blocked at 1 (which is normal for c|uα, which focuses on categorizing).
False positives and false negatives (Figure 28). In the current version of the CST, the false positive error type creates some overlapping (new units may overlap), and it is the reason why uα and κd were discarded from this experiment. However, we have kept c|uα because it behaves quite well despite overlapping units.
All the measures have overall a good response to the false positives error type, as shown in Figure 28a, even if the shape of c|uα is delayed compared with the others, but it should be pointed out that SER has a curious and unfortunate final increasing section (not visible in the figure because this section is below 0).
On the other hand, bigger differences appear with false negatives (Figure 28b). γ is still strictly decreasing and almost reaches 0 (0.025), but uα is not strictly decreasing, and is at 0 or below from m = 0.3; SER quickly drops below 0 from m = 0.4, κd is not strictly decreasing, and c|uα, as for splits, does not react at all but remains stuck at 1, which is desired for this coefficient focused on categories (values of c|uα over m = 0.7
are missing since there are not enough intersections between units for this measure to work).
Overview of each measure for the almost general case. In order to summarize the behavior of each measure in response to the different error types for the almost general case (without overlap), we pick all curves relative to a given measure out of the previous plots and draw them in the same graph, as shown in Figure 29. Briefly, γ shows a steady behavior for all error types, almost strictly decreasing from 1 to 0. uα has some increasing parts and negative values and is sometimes not responsive. c|uα is very responsive for some error types, is less responsive for some other types, and is sometimes not responsive at all (which is desired, as already said). SER has unreliable responses, being either too responsive (reaching negative values) or not responsive enough. Finally, κd is not always responsive, is most of the time not strictly decreasing, but is sometimes quite progressive.
6.3.5 Fully General Case: Unitizing + Categorizing. This last section considers the fully general case, where overlapping of units within an annotator is allowed. In this experiment, we took a reference corpus with no overlap, but the errors applied (combination of positioning and false positives) progressively lead to overlapping units. The results are shown in Figure 30. As expected, γ behaves quite the same as it does with non-overlapping configurations. Admittedly, c|uα was not designed to handle these configurations (and so should not be included in this experiment), but surprisingly it
seems to perform in rather the same way as it does with no overlapping; this must be investigated further, but judging from this preliminary observation, it seems this coefficient could still be operational and useful in such cases. On the contrary, uα does not handle correctly this experiment and so was not included in the graph. 7 conclusion :The present work addresses an aspect of inter-annotator agreement that is rarely studied in other approaches: the combination of unitizing and categorizing. Nevertheless, the use of methods that have been transposed from other domains (such as κ, which was originally dedicated to pure categorizing) in CL, for example at the discourse level, leads to severe biases, and manifests the need for specialized coefficients, fair and meaningful, suitable for annotation tasks focusing on complex objects.
In the end, only Krippendorff’s coefficients uα and c|uα come close to the needs we expressed in the introduction, with the restriction that they are natively limited to non-overlapping units.
The main reason why research on this topic is sparse, and why it may be difficult to enlarge Krippendorff’s coefficients to overlapping units, probably results from the fact that we are facing here a major difficulty: the simultaneous double discrepancy between annotators, with annotations possibly differing both in positioning relevant units anywhere on a continuum, and in categorizing each of these free units. Consequently, it is difficult for a method to choose precisely which features to compare between different annotators (unlike pure categorizing, where we know exactly what each annotator says for each predefined element to be categorized); and this problem is exacerbated when overlapping units (within an annotator) occur.
To cope with this critical point, we advocate the use of an alignment that ultimately expresses which unit from one annotator should be compared to which unit, if any, from another one, and consequently makes it natural and easier to compute the agreement. Moreover, we have shown that this alignment cannot be done in an independent way, but is part of the measure method itself. This is the “unified” aspect of our approach.
We have also shown that in order to be relevant, this alignment cannot be done at a local level (unit by unit), but should consider the whole set of annotations at the same time, which is the “holistic” aspect.
This is how the new method γ presented here was designed. Moreover, this method is highly configurable to cope with different annotation tasks (in particular, boundary errors should not necessarily be considered the same for all kinds of annotations), and it provides the alignment that emerges from an agreement measurement. Not only is this alignment a result in itself, which can be used to build a gold standard very quickly from a multi-annotated corpus (by listing all unitary alignments, and for each of them showing the corresponding frontiers and category proposed by each annotator), but it also behaves as a kind of a “flight recorder” of the measure: Observing these alignments gives crucial information on the choices the measure makes and whether it needs to be adjusted, unlike other methods which only provide a sole “out of the box” value.
Finally, we have compared γ to several other popular coefficients, even in their specific domains (pure categorization, pure segmentation), through a specific benchmark tool (namely, CST) which scans the responsivity of the measures to different kinds of errors and at all degrees of severity. Overall, γ provides broader and more progressive responsivity than the others in the experiments shown here. Concerning pure categorizing, γ does not have an edge over the well-known coefficients, such as α, but it is interesting to see that it behaves in much the same way as others in this specific field. Concerning segmentation, γ outperforms WD and GHD, by taking chance into account, but also by not depending on the heterogeneity of the segment sizes. Concerning unitizing with categorizing, as theoretically expected and confirmed by the benchmarking, SER shows severe limitations, such as a binary response to various (small or severe) positional or categorial errors, the fact that it does not make chance correction, or its limitation to two annotators only. Krippendorff’s coefficients uα and c|uα present very interesting properties, such as chance correction. However, as we have shown with thorough comparisons, they rely on quite different hypotheses to ours, since they consider intersections between units whereas we advocate considering aligned units. We have identified several situations in CL where considering alignments is advantageous, for instance, when contiguous segments of the same type may occur, or when errors on several short units should be considered as more serious than one error on a long unit, but we do not posit these situtations as a universal rule. In conclusion, when unitizing and categorizing involve internal overlapping of units, only γ is currently available, and, even if it cannot be compared to any other method at the moment for this reason, benchmarking reveals very similar responses to overlapping configurations and to non-overlapping ones, which already demonstrates its consistency and its relevance.
We can summarize the features of γ as follows: It takes into account all varieties of unitizing, combines unitizing and categorizing simultaneously, enables any number of annotators, provides chance correction, processes an alignment while it measures agreement, and provides progressive responsivity to errors both for unitizing and for categorizing. This makes γ suitable for annotation tasks such as relative to NAMED ENTITY, DISCOURSE FRAMING, TOPIC TRANSITION, or ENUMERATIVE STRUCTURES.
The full implementation of γ is provided as Java open-source packages on the http://gamma.greyc.fr Web site. It is already compatible with annotations created with the Glozz Annotation Platform (Widlöcher and Mathet 2012), and with annotations generated by the Corpus Shuffling Tool. Agreement measures have been widely used in computational linguistics for more than 15 years to check the reliability of annotation processes. Although considerable effort has been made concerning categorization, fewer studies address unitizing, and when both paradigms are combined even fewer methods are available and discussed. The aim of this article is threefold. First, we advocate that to deal with unitizing, alignment and agreement measures should be considered as a unified process, because a relevant measure should rely on an alignment of the units from different annotators, and this alignment should be computed according to the principles of the measure. Second, we propose the new versatile measure γ, which fulfills this requirement and copes with both paradigms, and we introduce its implementation. Third, we show that this new method performs as well as, or even better than, other more specialized methods devoted to categorization or segmentation, while combining the two paradigms at the same time. [{""affiliations"": [], ""name"": ""Yann Mathet""}, {""affiliations"": [], ""name"": ""Antoine Widl\u00f6cher""}, {""affiliations"": [], ""name"": ""Jean-Philippe M\u00e9tivier""}] SP:48c2046e4054182647ac9058902fdedb3869a97e [{""authors"": [""Ludovic Tanguy"", ""Marianne Vergez-Couret"", ""Laure Vieu.""], ""title"": ""An empirical resource for discovering cognitive principles of discourse organisation: The annodis corpus"", ""venue"": ""Proceedings of the Eighth"", ""year"": 2012}, {""authors"": [""Artstein"", ""Ron"", ""Massimo Poesio.""], ""title"": ""Inter-coder agreement for computational linguistics"", ""venue"": ""Computational Linguistics, 34(4):555\u2013596."", ""year"": 2008}, {""authors"": [""Beeferman"", ""Douglas"", ""Adam Berger"", ""John Lafferty.""], ""title"": ""Text segmentation using exponential models"", ""venue"": ""Proceedings of the 2nd Conference on Empirical Methods in Natural Language Processing, pages 35\u201346,"", ""year"": 1997}, {""authors"": [""E.M. Bennett"", ""R. Alpert"", ""A.C. Goldstein.""], ""title"": ""Communications through limited questioning"", ""venue"": ""Public Opinion Quarterly, 18(3):303\u2013308."", ""year"": 1954}, {""authors"": [""Berry"", ""Charles C.""], ""title"": ""The K statistic [letter"", ""venue"": ""Journal of the American Medical Association, 268(18):2513\u20132514."", ""year"": 1992}, {""authors"": [""Bestgen"", ""Yves.""], ""title"": ""Improving text segmentation using latent semantic analysis: A reanalysis of Choi, Wiemer-Hastings, and Moore (2001)"", ""venue"": ""Computational Linguistics, 32(1):5\u201312."", ""year"": 2006}, {""authors"": [""Bestgen"", ""Yves""], ""title"": ""Quel indice pour mesurer l\u2019efficacit\u00e9 en segmentation de textes"", ""venue"": ""Actes de TALN"", ""year"": 2009}, {""authors"": [""A. Bookstein"", ""V.A. Kulyukin"", ""T. Raita.""], ""title"": ""Generalized Hamming Distance"", ""venue"": ""Information Retrieval, (5):353\u2013375."", ""year"": 2002}, {""authors"": [""J. Carletta""], ""title"": ""Assessing agreement on classification tasks: The kappa statistic"", ""venue"": ""Computational Linguistics, 22(2):249\u2013254."", ""year"": 1996}, {""authors"": [""Carletta"", ""Jean.""], ""title"": ""Unleashing the killer corpus: Experiences in creating the multi-everything ami meeting corpus"", ""venue"": ""Language Resources and Evaluation, 41(2):181\u2013190."", ""year"": 2007}, {""authors"": [""Charolles"", ""Michel"", ""Anne Le Draoulec"", ""Marie-Paule Pery-Woodley"", ""Laure Sarda.""], ""title"": ""Temporal and spatial dimensions of discourse organisation"", ""venue"": ""Journal of French Language Studies,"", ""year"": 2005}, {""authors"": [""J. Cohen""], ""title"": ""A coefficient of agreement for nominal scales"", ""venue"": ""Educational and Psychological Measurement, 20(1):37\u201346."", ""year"": 1960}, {""authors"": [""J. Cohen""], ""title"": ""Weighted kappa: Nominal scale agreement with provision for scaled disagreement or partial credit"", ""venue"": ""Psychological Bulletin, 70(4):213\u2013220."", ""year"": 1968}, {""authors"": [""Di Eugenio"", ""Barbara"", ""Michael Glass.""], ""title"": ""The kappa statistic: A second look"", ""venue"": ""Computational Linguistics, 30(1):95\u2013101."", ""year"": 2004}, {""authors"": [""Eisenstein"", ""Jacob.""], ""title"": ""Hierarchical text segmentation from multi-scale lexical cohesion"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association"", ""year"": 2009}, {""authors"": [""Fleiss"", ""Joseph.""], ""title"": ""Measuring nominal scale agreement among many raters"", ""venue"": ""Psychological Bulletin, (5):378\u2013382."", ""year"": 1971}, {""authors"": [""Fort"", ""Kar\u00ebn"", ""Claire Fran\u00e7ois"", ""Olivier Galibert"", ""Maha Ghribi.""], ""title"": ""Analyzing the impact of prevalence on the evaluation of a manual annotation campaign"", ""venue"": ""Eighth International Conference on Language"", ""year"": 2012}, {""authors"": [""Laurent.""], ""title"": ""Named and specific entity detection in varied data: The quaero named entity baseline evaluation"", ""venue"": ""Seventh International Conference on Language 477"", ""year"": 2010}, {""authors"": [""Hills"", ""CA. Krippendorff"", ""Klaus""], ""title"": ""On the reliability"", ""year"": 1995}, {""authors"": [""Labadi\u00e9"", ""Alexandre"", ""Patrice Enjalbert"", ""Yann Mathet"", ""Antoine Widl\u00f6cher.""], ""title"": ""Discourse structure annotation: Creating reference corpora"", ""venue"": ""Workshop on Language Resource and Language"", ""year"": 2010}, {""authors"": [""S. Lamprier"", ""T. Amghar"", ""B. Levrat"", ""F. Saubion.""], ""title"": ""On evaluation methodologies for text segmentation algorithms"", ""venue"": ""Proceedings of ICTAI 2007, pages 19\u201326. Patras."", ""year"": 2007}, {""authors"": [""Makhoul"", ""John"", ""Francis Kubala"", ""Richard Schwartz"", ""Ralph Weischedel.""], ""title"": ""Performance measures for information extraction"", ""venue"": ""Proceedings of DARPA Broadcast News Workshop, pages 249\u2013252,"", ""year"": 1999}, {""authors"": [""Mathet"", ""Yann"", ""Antoine Widl\u00f6cher.""], ""title"": ""Une approche holiste et unifi\u00e9e de l\u2019alignement et de la mesure d\u2019accord inter-annotateurs"", ""venue"": ""Traitement Automatique des Langues Naturelles 2011"", ""year"": 2011}, {""authors"": [""Nadeau"", ""David"", ""Satoshi Sekine.""], ""title"": ""A survey of named entity recognition and classification"", ""venue"": ""Linguisticae Investigationes, 30(1):3\u201326."", ""year"": 2007}, {""authors"": [""Olivero"", ""Patrick.""], ""title"": ""Calcul de la taille des \u00c9chantillons"", ""venue"": ""CETE du Sud-Ouest / DAT / ZELT. Technical report."", ""year"": 2001}, {""authors"": [""Passonneau"", ""Rebecca J."", ""Vikas Bhardwaj"", ""Ansaf Salleb-Aouissi"", ""Nancy Ide.""], ""title"": ""Multiplicity and word sense: Evaluating and learning from multiply labeled word sense annotations"", ""venue"": ""Language Resources and"", ""year"": 2012}, {""authors"": [""L. Pevzner"", ""M. Hearst.""], ""title"": ""A critique and improvement of an evaluation metric for text segmentation"", ""venue"": ""Computational Linguistics, 28(1):19\u201336."", ""year"": 2002}, {""authors"": [""D. Reidsma""], ""title"": ""Annotations and Subjective Machines of Annotators, Embodied Agents, Users, and Other Humans"", ""venue"": ""Ph.D. thesis, University of Twente."", ""year"": 2008}, {""authors"": [""D. Reidsma"", ""D.K.J. Heylen"", ""R.J.F. Ordelman.""], ""title"": ""Annotating emotions in meetings"", ""venue"": ""Proceedings of the Fifth International Conference on Language Resources and Evaluation, LREC 2006,"", ""year"": 2006}, {""authors"": [""Reidsma"", ""Denis"", ""Jean Carletta.""], ""title"": ""Reliability measurement without limits"", ""venue"": ""Computational Linguistics, 34(3):319\u2013326."", ""year"": 2008}, {""authors"": [""Scott"", ""William.""], ""title"": ""Reliability of content analysis: The case of nominal scale coding"", ""venue"": ""Public Opinion Quarterly, 19(3):321\u2013325."", ""year"": 1955}, {""authors"": [""Siegel"", ""Sidney"", ""N. John Castellan.""], ""title"": ""Nonparametric Statistics for the Behavioral Sciences"", ""venue"": ""McGraw-Hill, New York, 2nd edition."", ""year"": 1988}, {""authors"": [""Teufel"", ""Simone.""], ""title"": ""Argumentative Zoning: Information Extraction from Scientific Articles"", ""venue"": ""Ph.D. thesis, University of Edinburgh."", ""year"": 1999}, {""authors"": [""Teufel"", ""Simone"", ""Jean Carletta"", ""Marc Moens""], ""title"": ""An annotation scheme for discourse-level argumentation in research"", ""year"": 1999}, {""authors"": [""Teufel"", ""Simone"", ""Marc Moens.""], ""title"": ""Summarizing scientific articles: Experiments with relevance and rhetorical status"", ""venue"": ""Computational Linguistics, 28(4):409\u2013445."", ""year"": 2002}, {""authors"": [""Widl\u00f6cher"", ""Antoine"", ""Yann Mathet.""], ""title"": ""The glozz platform: A corpus annotation and mining tool"", ""venue"": ""ACM Symposium on Document Engineering (DocEng\u201912), pages 171\u2013180, Paris."", ""year"": 2012}, {""authors"": [""Zwick"", ""Rebecca.""], ""title"": ""Another look at interrater agreement"", ""venue"": ""Psychological Bulletin, (103):347\u2013387. 479"", ""year"": 1988}] acknowledgments  :We wish to thank three anonymous reviewers for helpful comments and discussion. The authors would also like to warmly thank Klaus Krippendorff for his support when they implemented his coefficients in order to test them. This work was carried out in the GREYC Laboratory, Caen, France, with the strong support of the French Contrat de Projet État-Région (CPER) and the Région Basse-Normandie, which have provided this research with two engineers, Jérôme Chauveau and Stéphane Bouvry. appendix a. examples of linguistic objects and possible annotation tasks :Terms emphasized hereafter refer to the terminology defined in Section 2.
PART-OF-SPEECH. Part-of-speech (POS) tagging (see, for example, Güngör [2010] for a recent state of the art) gives a well-known illustration of a pure categorization without unitizing task: for all the words in a text (predefined units, full-covering, no overlap), annotators have to select a category, belonging to quite a small set of exclusive elements. POS units (words) having the same label are obviously not aggregatable.
GENE RENAMING. In a study on gene renaming presented in Fort et al. (2012), all the tokens (predefined units, no overlap) are markable (categorization) with “Nothing” (the default value), “Former” (the original name of a gene) or “New” (its new name). This work at word level considers sparser (sporadic) phenomena than POS tagging. However, the annotation is defined as full-covering, with “Nothing” as a default tag. Note that the presence of the “Nothing” category also reveals here the reduction of a unitizing problem (detection of renaming) to a pure coding system (categorization). These units are not aggregatable.
WORD SENSE. For the annotation task described in Passonneau et al. (2012), annotators were asked to assign sense labels (categorization without unitizing) to preselected moderately polysemous words (sporadicity, predefined units, no overlap) in preselected sentences where they occur. Adjacent words are not aggregatable with sense preservation.
NAMED ENTITIY. Well-established named entity (NE) recognition tasks (see, for example, Nadeau and Sekine [2007]) led to many annotation efforts. In such tasks, the annotator is often asked to identify the units in the text’s continuum (unitizing, sporadicity) and to select a NE type from an inventory (categorization). It is well known that some difficulties of NE annotation relate to the delimitation of NE boundaries. For example, for a phrase such as “Mr X, the President of Y,” it makes sense to annotate subparts (“X,” “Mr X,” “the President of Y”) and/or the whole. “Y” is also a NE of another type. This may result in hierarchical or free overlapping structures. Adjacent NE are not aggregatable.
ARGUMENTATIVE ZONING. Studies concerned by argumentative zoning (Teufel 1999; Teufel, Carletta, and Moens 1999; Teufel and Moens 2002) consider the argumenative structure of texts, and identify text spans having specific roles. For each sentence (fullcovering, predefined units, no overlap), a category (categorization) is selected. Adjacent sentences of the same type are aggregated into larger spans (argumentative zones). This reveals an underlying question of unitizing. However, it has to be noted that the categorization mainly concerns predefined sentences: argumentative types are aggregatable.
DISCOURSE FRAMING. In Charolles et al.’s discourse framing hypothesis (Charolles et al. 2005, page 121), “a discourse frame is described as the grouping together of a number of propositions which are linked by the fact that they must be interpreted with reference to a specific criterion, realized in a frame-initial introducing expression.” Thus, temporal or spatial introducing expressions lead, for example, to temporal or spatial discourse frames in the text continuum (unitizing, sporadicity, categorization). Discourse frames are not aggregatable. Subordination is possible, leading to possibly hierarchical overlap, where frames (of the same type or of different types) are embedded.
COMMUNICATIVE BEHAVIOR. The multimodal AMI Meeting corpus (Carletta 2007) covers a wide range of phenomena, and contains many different layers of annotation describing the communicative behavior of the participants of meetings. For example, in Reidsma (2008), annotators are required to identify fragments in a video recording (unitizing, sporadicity) and to categorize them (categorization). For such an annotation task, one can easily imagine instruction manuals allowing annotators to use multiple labels and to identify embedded (hierarchical overlap) or free overlapping units, even if the example provided by Reidsma (2008) does not.
DIALOG ACT. Annotating dialog act conforming to a standard as defined, for example, in Bunt et al. (2010), leads annotators to assign communicative function labels and types of semantic content (categorization) to stretches of dialogue called functional segments. The possible mulitfunctionality of segments (one functional segment is related to one or more dialog acts), and the fact that annotations may be attached directly to the primary data such as stretches of speech defined by begin and end points, or attached to structures at other levels of analysis, seems to allow different kinds of configurations and annotation instructions: unitizing or pure categorization of pre-existing structures, sporadicity or full-covering, hierarchical, overlapping or linear segmentation.
TOPIC SEGMENTATION. Topic segmentation (see, for example, the seminal work by Hearst [1997] or Bestgen [2006] for a more recent state of the art), which aims at detecting the most important thematic breaks in the text’s continuum, gives an illuminating example of pure segmentation. This unitizing problem of linear segmentation is full-covering and restricted to the detection of breaks (the right boundary of a unit corresponds to the left boundary of the following segment) (no overlap). If we consider the resulting segments, there is just one category (topic segment) and then no categorization. Adjacent topic segments are obviously not aggregatable without a shift in meaning.
HIERARCHICAL TOPIC SEGMENTATION. In order to take better into account the fact that lexical cohesion is a multiscale phenomenon and that discourse displays a hierarchical structure, hierarchical topic segmentation proposed, for example, by Eisenstein (2009) preserves the main goal and properties of text-tiling (unitizing, no categorization, fullcovering, not aggregatable segments), but allows a topic segment to be subsegmented into sub-topic segments (hierarchical [but not free] overlap).
TOPIC TRANSITION. The topic zoning annotation model presented in Labadié et al. (2010) is based on the hypothesis that, in a well constructed text, abrupt topic boundaries are more the exception than the rule. This model introduces transition zones (unitizing) between topics, zones that help the reader to move from one topic to another. The annotator is asked to identify and categorize (categorization) topic segments, introduction, conclusion, and transition zones. Hierarchical overlap is possible (embedded elements of the same type or of different types are allowed). Free overlapping structures are frequent, by virtue of the nature of transitions. Adjacent topic zones and adjacent transition zones are not aggregatable.
ENUMERATIVE STRUCTURES. A study on complex discourse objects such as enumerative structures (Afantenos et al. 2012) illustrates both the need for sporadic unitizing and the need for categorization. The enumerative structures have a complex internal organization, which is composed of various types of subelements (hierarchical overlap) (a trigger of the enumeration, items composing its body, etc.) which are not aggregatable.","6 comparing and benchmarking γ :As γ is an entirely new agreement measure method, it is necessary to analyze how it compares with some well known and much studied methods. First, we carry out a thorough comparison between γ and the two dedicated alphas, uα and c|uα, which are the most specific measures in the domain. Second, we benchmark γ by comparing it with other main measures, thanks to a special tool that is briefly introduced. As already mentioned, Krippendorff’s uα and c|uα are clearly the most suitable coefficients for combined unitizing and categorizing. To better understand the pros and cons as well as the behavior of these measures compared with γ, we first explain how they are designed in Section 6.1.1, and then make thorough comparisons with γ from Section 6.1.2 to 6.1.6 including: (1) how they react to slight categorial disagreements, (2) interlacement of positional and categorial disagreements, (3) the impact of the size of the units on positional disagreement, (4) split disagreements, and (5) the impact of scale (e.g., if the size of all units is multiplied by 2). We finish by showing a paradox of uα in Section 6.1.7.
6.1.1 Introducing uα and c|uα. To introduce how these two coefficients work, let us consider the example taken from Krippendorff (2013), shown in Figure 14. The length of the continuum is 76, there are two annotators, and there are four possible categories, numbered 1 to 4.
The uα coefficient basically relies on the comparison of all pairs of sections among annotators, a section being either a categorized unit or a gap. To get the observed disagreement value uDo, squared lengths of the unmatching intersections are summed, and this sum is then divided by the product of the length of the continuum and m(m− 1), m being the number of annotators. In the example, mismatches occur around the second and third units of the two annotators. From left to right, there are the following intersections: cat 1 with gap (l = 10), cat 1 with cat 3 (l = 5), gap with cat 3 (l = 8), cat 2 with cat 1 (l = 5), and cat 2 with gap (l = 5). This leads twice (by symmetry) to the sum 102 + 52 + 82 + 52 + 52, and so the observed disagreement uDo = 2(102+52+82+52+52 )
76·2(2−1) = 3.145. The expected value uDe is obtained by considering all the possible positional combinations of each pair, and not only the observed ones. This means that for a given pair, one of the two units is virtually slid in front of the other in all
possible ways, and the corresponding values are averaged. In this example, uDe = 5.286. Therefore, uα = 1− 3.1455.286 = 0.405.
Coefficient c|uα relies on a coincidence matrix between categories, filled with the sums of the lengths of all intersections of units for each given pair of categories. For instance, in the example, the observed coincidence between category 1 and category 3 is 5, and so on. A metric matrix is chosen for these categories, for instance, an interval metric (for numerical categories), which says that the distance between category i and category j is (i− j)2. Hence, the cost for a unitary intersection between categories 1 and 2 is (1− 2)2 = 1, but is 22 = 4 between categories 1 and 3, and so on. Then, the observed disagreement is computed according to these two matrices. To finish, an expected matrix is filled (in a way which cannot be detailed here due to space constraints), and the expected value is computed the same way. In the example, c|uα = 1− 0.8333.145 = 0.744.
Hence, Krippendorff’s alphas provide two clues to analyze the agreement between annotators. In the example, uα = 0.405 indicates that the unitizing is not so good, but also that the categorizing is much better, with c|uα = 0.744 (even though of course, these two values are not independent, since unitizing and categorizing coexist here by nature).
Now that these coefficients have been introduced in detail, let us analyze to what extent they differ from γ.
6.1.2 Slight Categorial Disagreements: Alphas Versus γ. When annotators have slight categorial disagreements (with overlapping categories), c|uα is slightly lowered. However, uα does not take categorial overlapping into account, but has a binary response to such disagreements, and is lowered as much as if they were severe categorial disagreements. A consequence of this approach is illustrated in Figure 15, where two annotators perfectly agree both on positions and categories in the experiment on the left, and still perfectly agree on position but slightly diverge concerning categories in the experiment on the right (1/2, 6/7, and 8/9 are assumed to be close categories). However, uα drops from 1 in the left experiment to –0.34 (a negative value means worse than random) in the right experiment, despite, in the latter, the positions being all correct, and the categories being quite good, since c|uα = 0.85. On such data, γ considers that there is no positional disagreement, and c|uα and γ both consider that there are slight categorial disagreements.
6.1.3 Positional Disagreements Impacting Categorial Agreement: c|uα. Two different conceptions of how to account for categorial disagreement have, respectively, led to c|uα and γ: c|uα relies on intersections between the units of different annotators, which is basically equivalent to an observation at the atom level, whereas γ relies on alignments between units (any unit being finally attached and compared to, at most, only one other) based both on positional and categorial observation. Hence, in a configuration such as the one given in Figure 16, where two annotators annotated three units with the same categories
1
1
9
9
7
7
Observer 1
Observer 2
1
2
9
8
7
6
Observer 1
Observer 2
Figure 15 Consequences of no categorial disagreement (left) compared with slight categorial disagreements (right).
1, 4, and 2, but not exactly at the same locations; c|uα considers a certain account of categorial disagreements, whereas γ does not. According to the principles of c|uα, any part of the continuum (even at the atom level) with an intersection between different categories means some confusion between them, whereas γ considers here that the annotators fully agree on categories (they both observed a “1” then a “4” then a “2” with no confusion), and disagree only on where phenomena exactly start and finish. The crucial difference between the two methods is probably whether we consider units to be non-atomizable (and therefore consider alignments, as γ does), or atomizable (in which case two different parts of a given unit may be simultaneously and respectively compared to two different units from another annotator).
6.1.4 Disagreements on Long versus Short Spans of Texts. Here again, the way disagreements are accounted for may differ markedly between uα and γ: when a unit does not match with any other, uα takes into account the length of the corresponding span of text to assess a disagreement. As shown in Figure 17, an orphan unit of size 10 will cost 100 times as much as an orphan unit of size 1, whereas for γ, they will have the same cost. In the whole example in Figure 17, to compute the observed disagreements, uα says the first case is 50 times worse than the second, whereas γ says on the contrary that the second case is twice as bad as the first. Here, γ fulfills the need (already mentioned) expressed by Reidsma, Heylen, and Ordelman (2006, page 3) to consider that “short segments are as important as long segments.” This phenomenon is the same for categories between c|uα and γ, the size of the units having consequences only for c|uα.
6.1.5 Split Disagreements. Sometimes, an annotator may divide a given span of text into several contiguous units of the same type, or may annotate the same span with one whole unit. In these cases, c|uα computes its observed disagreement the same in both
configurations, and uα assigns decreasing disagreement when splitting increases, as shown in Figure 18, whereas γ assigns increasing disagreements.
Moreover, in Figure 19, the observed uα is not responsive to splits at all, whereas γ is still responsive.
6.1.6 Scale Effects. The way uα computes dissimilarities is directly proportional to squared lengths, as shown in Figure 20. On the other hand, γ may use any positional dissimilarity, and usually uses ones that are not scale-dependent for CL applications, such as dpos−sporadic (Equation (3)). For instance, if a text is annotated with two categories, one at word level, the other one at paragraph level, we may prefer to account for relative disagreements so that a missing word will be more heavily penalized in the first case than in the second. In Figure 20, the observed disagreement of uα is 32 = 9 times greater for B units than for A units, but would be the same for γwith dpos−sporadic since:
( 0 + 3( 7+10
2 ))2 = ( 0 + 9( 21+30 2 ))2 . Figure 21a, the annotators disagree on categorization, and have a moderate agreement on unitizing. This configuration leads to uα = 0.107. In Figure 21b, the configuration is quite similar, but now annotators fully agree on unitizing: Each of them puts units in the same positions. Paradoxically, uα drops to −0.287, which is less than in the first configuration. In brief, the reason for this behavior is that in the first case, the computed disagreement regarding a given pair of units is virtually distributed into shorter parts of the whole (an intersection of length 80 between them, and an intersection of length
20 with a gap for each of them, which leads to 802 + 2× 202 = 7, 200) whereas the disagreement is maximum in the second case (an intersection of length 100 with a unit of another category, which leads to 1002 = 10, 000). Contrarily, with similar data, γ provides a better agreement in the second case than in the first one. With its design, it considers that there is the same categorial agreement in both cases, but better positional agreement in the second case, which seems to better correspond to the CL tasks we have considered.
6.1.8 Overlapping Units (Embedding or Free Overlap). Both alpha coefficients are currently designed to cope only with non-overlapping units (the term overlapping also stands here for embedding), which is a limitation for several fields in CL. It is debatable whether they could be generalized to handle overlapping units. It seems that it would involve a major change in the strategy, which currently necessitates comparing the intersections of all pairs of units. In the example shown in Figure 22, even though annotators fully agree on their two units, the alphas will inherently compare A1 with B2 and A2 with B1 (in addition to the normal comparisons between A1 with B1 and A2 with B2), and will count the resulting intersections as disagreements. It is necessary here to choose once and for all what unit to compare to what other, rather than to perform all the comparisons. But making such a choice precisely consists in making an alignment, which is a fundamental feature of γ. Consequently, it seems that the alphas would need a structural modification to cope with overlapping. As explained by Reidsma, Heylen, and Ordelman (2006), because of the lack of specialized coefficients coping with unitizing, a fairly standard practice is to use categorization coefficients on a discretized (i.e., atomized) version of the continuum: For instance, each character (or each word, or each paragraph) of a text is considered as an item; and a standard categorization coefficient such as κ is used to compute agreement. Such a measure is called κd (for discretized κ) hereafter. Several weaknesses of this approach have been already mentioned in the state-of-the-art section. It is interesting to compare such a measure to the specialized one c|uα: even if they both bear the aggregatable hypothesis, they have, however, significant differences (as confirmed by the experiments presented in the next section). The main one is that c|uα does not use an artificial atomization of the continuum, and only compares units with units. In doing so, it is not prone to agreement on blanks, in contrast to κd. Another difference is that, for the same reason, c|uα is not inherently limited to non-overlapping units: Even if it is not currently designed to cope with them, as we have already seen, it is possible to submit overlapping units to this measure (some results are shown in the next section). In this section on benchmarking, we use the Corpus Shuffling Tool (CST) introduced by Mathet et al. (2012) to compare γ concretely and accurately to the other measures.
We first introduce the possible error types that it will provide: category (category mistakes may occur), position (the boundaries may be shifted), false positives (the annotators add units to the reference units), false negatives (the annotators miss some of the reference units), and splits (the annotators put two or more contiguous units instead of a reference unit, which occupy the same span of text).
This tool is used to simulate varying degrees of disagreement among different error types, and the metrics are compared with each other according to how they react to these disagreements. For a given error type, for each magnitude between 0 and 1 (with a step of 0.05), the tool creates 40 artificial, multi-annotator shuffled annotation sets, and computes the different measures for them. Hence, we obtain a full graph showing the behavior of each measure for this error type, with the magnitude on the x-axis, and the average agreement (over the 40 annotation sets) on the y-axis. This provides a sort of “X-ray” of the capabilities of the measures with respect to this error type, which should be evaluated against the following desiderata:r A measure should provide a full response to the whole range of
magnitudes, which means in particular that the curve should ideally start from 1 (at m = 0) and reach 0 (at m = 1), but never go below 0 (indeed, negative agreement values require a part of systematic disagreement, which is not simulated by the current version of the CST).r The response should be strictly decreasing: A flat part would mean the measure does not differentiate between different magnitudes, and, even worse, an increasing part would mean that the measure is counter-effective at some magnitudes, where a worse error is penalized less severely.
We emphasize the fact that the whole graph is important, up to magnitude 1. Indeed, in most real annotated corpora, even when the overall agreement is high, errors
corresponding to all magnitudes may occur. For instance, an agreement of 0.8 does not necessarily correspond to the fact that all annotations are affected by slight errors (which correspond to magnitudes close to 0), but may for instance correspond to the fact that a few units are affected by severe errors (which may correspond to magnitudes close or equal to 1).
It is important to note that this tool was designed by the authors of γ, for tasks where units cannot be considered as atomizable. In particular, it was conceived so that disagreements concerning small units are as important as those concerning large ones. However, it is provided as open-source (see Conclusions section) so that anyone can test and modify it, and propose new experiments to test γ and other measures in the future.
6.3.1 Introducing the CST. The main principle of this tool is as follows. A reference corpus is built, with respect to a statistical model, which defines the number of categories, their prevalence, the minimum and maximum length for each category, and so forth. Then, this reference is used by the shuffling tool to generate a multi-annotator corpus, simulating the fact that each annotator makes mistakes of a certain type, and of a certain magnitude. It is important to remark that the generated corpus does not include the reference it is built from.
The magnitude m is the strength of the shuffling, that is to say the severity of mistakes annotators make compared to the reference. It can be set from 0, which means no damage is applied (and the annotators are perfect) to the extreme value 1, which means annotators are assumed to behave in the worst possible way (but still being independent of each other)—namely, at random.
Figure 23 illustrates the way such a corpus is built: From the reference containing some categorized units, three new sets of annotations are built, simulating three annotators who are assumed to have the same annotating skill level, which is set in this example at magnitude 0.1. The applied error type is position only, that is to say that each annotator makes mistakes only when positioning boundaries, but does not make any other mistake (the units are reproduced in the same order, with the correct category, and in the same number). At this low magnitude, the positions are still close to those of the reference, but often vary a little. Hence, we obtain here a slightly shuffled multiannotator corpus. Let us sum up the way error types are currently designed in the CST.
Position. At magnitude m, for a given unit, we define a value shiftmax that is proportional to m and to the length of the unit, and each boundary of the unit is shifted by a
value randomly chosen between −shiftmax and shiftmax (note: at magnitude 0, because shiftmax = 0, units are not shifted).
Category. This shuffling cannot be described in a few words (see Mathet et al. [2012] for details). It uses special matrices to simulate, using conditional probabilities, progressive confusion between categories, and can be configured to take into account overlapping of categories. The higher the magnitude, the more frequent and severe the confusion.
False negatives. At magnitude m, each unit has the probability m to be forgotten. For instance, at magnitude m = 0.5, each annotator misses (on average) half of the units from the reference (but not necessarily the same units as the other annotators).
False positives. At magnitude m, each annotator adds a certain number of units (proportional to m) to the ones of the reference.
Splits. At magnitude m, each annotator splits a certain number of units (proportional to m). A split unit may be re-split, and so on.
6.3.2 Pure Segmentation: γ, WD, GHD. Even if γ was created to cope with error types that are poorly or not at all dealt with by other methods, and, moreover, to cope simultaneously with all of them (unitizing of categorized and overlapping categories), it is illuminating to observe how it behaves in more specific error types, to which specialized and well known methods are dedicated. We start with pure segmentation. Figure 24 shows the behavior of WD, GHD, and γ for two error types.
For false negatives, WD and GHD are quite close, with an almost linear response until magnitude 0.6. Their drawback is that their responses are limited by an asymptote, because of the absence of chance correction, while γ shows a full range of agreements; for shifts, WD and GHD show an asymptote at about agreement = 0.4, while γ shows values from 1 to 0. This experiment confirms the advantage of using γ instead of these distances for inter-annotator agreement measure.
6.3.3 Pure Categorization. In this experiment, the CST is set to three annotators, four categories with given prevalences. The units are all of the same size, positioned in fixed, predefined positions, so that the focus is on categorizing only. It should be noted that, with such a configuration, α and κ behave exactly in the same way as c|uα. It is particularly striking in Figure 25 that γ behaves in almost the same way as c|uα. In fact, the observed values of these measures are exactly the same, the only difference coming from a slight difference in the expected values, due to sampling. Other tests carried out with the pure categorizing coefficient κ yielded the same results on this particular error type, which means that γ performs as well as recognized measures as far as categorizing is concerned, with two or more annotators. The uα curve goes below zero at magnitude 0.5 (probably for the reasons seen in Section 6.1.7). Moreover, its behavior depends on the size of the gaps: Indeed, with other settings of the shuffling, the curve may, on the contrary, be stuck over zero. κd fails to reach 0 because of the virtual agreement on gaps (but it would if there were no gaps). Lastly, SER (averaging the results of each pair of annotators) is bounded below by 0.6, which results from not taking chance into account.
6.3.4 Almost General Case: Unitizing + Categorizing. This section concerns the more general uses of γ, combining both unitizing and categorizing. However, in order to be compliant with uα, c|uα, and κd, we limit the configurations here so that the units do not overlap at all. In particular, the reference was built with no overlapping units, and we have used a modified version of the shifting shuffling procedure so that the nonoverlapping constraint is fully satisfied, even at high magnitudes.
Positional errors (Figure 26a). An important point is that this shuffling error type, which is based only on moving positions, has a paradoxical consequence on category agreement, since units of different categories align when sufficient shifting is applied. Consequently, c|uα is not blocked at 1, even though it is designed to focus on categories. Additionally, it starts to decrease from the very first shifts, as soon as units from different annotators start overlapping. This is a concrete consequence of what has been formally studied in Section 6.1.3. γ has a most progressive response, reaches 0.1 at magnitude 1,
and is the only measure to be strictly decreasing. SER immediately drops to agreement 0.5 at magnitude 0.05. As it relies on a binary positional distance, it fails to distinguish between small and large errors. This is a serious drawback of such a measure for most CL tasks. Then it goes below zero and is not strictly decreasing. uα is mostly strictly decreasing, but has some increasing parts, and, even more problematic, negative values from 0.6 to 0.9, probably because of the reason explained in Section 6.1.7. κd is too responsive at the very first magnitudes, and is not strictly decreasing, probably because it “does not compensate for differences in length of segments” (Reidsma, Heylen, and Ordelman 2006, page 3).
Positional and categorial errors (Figure 26b). γ is strictly decreasing and reaches 0. The alphas are not strictly decreasing, and once again uα drops below 0 from magnitude 0.6 onwards. κd is not strictly decreasing (again, probably because it “does not compensate for differences in length of segments”), but its general shape is not that far from γ.
Split errors (Figure 27). The split error type would need to create an infinite number of splits to mean pure chaos at magnitude 1. As this is computationally not possible, we restricted the number of splits to five times the number of units of the reference. We should therefore not expect measures to reach 0. In this context, γ shows a good range of responses, from 1 to 0.2, in an almost linear curve. SER is also quite linear, but gives very confusing values for this error type because it reaches negative values above magnitude 0.6. Finally, uα, c|uα, and κd are not responsive at all to this error type, as expected, and remain blocked at 1 (which is normal for c|uα, which focuses on categorizing).
False positives and false negatives (Figure 28). In the current version of the CST, the false positive error type creates some overlapping (new units may overlap), and it is the reason why uα and κd were discarded from this experiment. However, we have kept c|uα because it behaves quite well despite overlapping units.
All the measures have overall a good response to the false positives error type, as shown in Figure 28a, even if the shape of c|uα is delayed compared with the others, but it should be pointed out that SER has a curious and unfortunate final increasing section (not visible in the figure because this section is below 0).
On the other hand, bigger differences appear with false negatives (Figure 28b). γ is still strictly decreasing and almost reaches 0 (0.025), but uα is not strictly decreasing, and is at 0 or below from m = 0.3; SER quickly drops below 0 from m = 0.4, κd is not strictly decreasing, and c|uα, as for splits, does not react at all but remains stuck at 1, which is desired for this coefficient focused on categories (values of c|uα over m = 0.7
are missing since there are not enough intersections between units for this measure to work).
Overview of each measure for the almost general case. In order to summarize the behavior of each measure in response to the different error types for the almost general case (without overlap), we pick all curves relative to a given measure out of the previous plots and draw them in the same graph, as shown in Figure 29. Briefly, γ shows a steady behavior for all error types, almost strictly decreasing from 1 to 0. uα has some increasing parts and negative values and is sometimes not responsive. c|uα is very responsive for some error types, is less responsive for some other types, and is sometimes not responsive at all (which is desired, as already said). SER has unreliable responses, being either too responsive (reaching negative values) or not responsive enough. Finally, κd is not always responsive, is most of the time not strictly decreasing, but is sometimes quite progressive.
6.3.5 Fully General Case: Unitizing + Categorizing. This last section considers the fully general case, where overlapping of units within an annotator is allowed. In this experiment, we took a reference corpus with no overlap, but the errors applied (combination of positioning and false positives) progressively lead to overlapping units. The results are shown in Figure 30. As expected, γ behaves quite the same as it does with non-overlapping configurations. Admittedly, c|uα was not designed to handle these configurations (and so should not be included in this experiment), but surprisingly it
seems to perform in rather the same way as it does with no overlapping; this must be investigated further, but judging from this preliminary observation, it seems this coefficient could still be operational and useful in such cases. On the contrary, uα does not handle correctly this experiment and so was not included in the graph."
"1 introduction :A well-written text is not merely a sequence of independent and isolated sentences, but instead a sequence of structured and related sentences, where the meaning of a sentence relates to the previous and the following ones. In other words, a well-written
∗ Arabic Language Technologies, Qatar Computing Research Institute, Qatar Foundation, Doha, Qatar. E-mail: sjoty@qf.org.qa.
∗∗ Computer Science Department, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4. E-mail: carenini@cs.ubc.ca.
† Computer Science Department, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4. E-mail: rng@cs.ubc.ca.
Submission received: 11 May 2014; revised version received: 29 January 2015; accepted for publication: 18 March 2015.
doi:10.1162/COLI a 00226
No rights reserved. This work was authored as part of the Contributor’s official duties as an Employee of the United States Government and is therefore a work of the United States Government. In accordance with 17 U.S.C. 105, no copyright protection is available for such works under U.S. Law.
text has a coherence structure (Halliday and Hasan 1976; Hobbs 1979), which logically binds its clauses and sentences together to express a meaning as a whole. Rhetorical analysis seeks to uncover this coherence structure underneath the text; this has been shown to be beneficial for many Natural Language Processing (NLP) applications, including text summarization and compression (Marcu 2000b; Daumé and Marcu 2002; Sporleder and Lapata 2005; Louis, Joshi, and Nenkova 2010), text generation (Prasad et al. 2005), machine translation evaluation (Guzmán et al. 2014a, 2014b; Joty et al. 2014), sentiment analysis (Somasundaran 2010; Lazaridou, Titov, and Sporleder 2013), information extraction (Teufel and Moens 2002; Maslennikov and Chua 2007), and question answering (Verberne et al. 2007). Furthermore, rhetorical structures can be useful for other discourse analysis tasks, including co-reference resolution using Veins theory (Cristea, Ide, and Romary 1998).
Different formal theories of discourse have been proposed from different viewpoints to describe the coherence structure of a text. For example, Martin (1992) and Knott and Dale (1994) propose discourse relations based on the usage of discourse connectives (e.g., because, but) in the text. Asher and Lascarides (2003) propose Segmented Discourse Representation Theory, which is driven by sentence semantics. Webber (2004) and Danlos (2009) extend sentence grammar to formalize discourse structure. Rhetorical Structure Theory (RST), proposed by Mann and Thompson (1988), is perhaps the most influential theory of discourse in computational linguistics. Although it was initially intended to be used in text generation, later it became popular as a framework for parsing the structure of a text (Taboada and Mann 2006). RST represents texts by labeled hierarchical structures, called Discourse Trees (DTs). For example, consider the DT shown in Figure 1 for the following text:
But he added: “Some people use the purchasers’ index as a leading indicator, some use it as a coincident indicator. But the thing it’s supposed to measure—manufacturing strength—it missed altogether last month.”
The leaves of a DT correspond to contiguous atomic text spans, called elementary discourse units (EDUs; six in the example). EDUs are clause-like units that serve as building blocks. Adjacent EDUs are connected by coherence relations (e.g., Elaboration, Contrast), forming larger discourse units (represented by internal nodes), which in turn are also subject to this relation linking. Discourse units linked by a rhetorical relation are
further distinguished based on their relative importance in the text: nuclei are the core parts of the relation and satellites are peripheral or supportive ones. For example, in Figure 1, Elaboration is a relation between a nucleus (EDU 4) and a satellite (EDU 5), and Contrast is a relation between two nuclei (EDUs 2 and 3). Carlson, Marcu, and Okurowski (2002) constructed the first large RST-annotated corpus (RST–DT) on Wall Street Journal articles from the Penn Treebank. Whereas Mann and Thompson (1988) had suggested about 25 relations, the RST–DT uses 53 mono-nuclear and 25 multi-nuclear relations. The relations are grouped into 16 coarse-grained categories; see Carlson and Marcu (2001) for a detailed description of the relations. Conventionally, rhetorical analysis in RST involves two subtasks: discourse segmentation is the task of breaking the text into a sequence of EDUs, and discourse parsing is the task of linking the discourse units (EDUs and larger units) into a labeled tree. In this article, we use the terms discourse parsing and rhetorical parsing interchangeably.
While recent advances in automatic discourse segmentation have attained high accuracies (an F-score of 90.5% reported by Fisher and Roark [2007]), discourse parsing still poses significant challenges (Feng and Hirst 2012) and the performance of the existing discourse parsers (Soricut and Marcu 2003; Subba and Di-Eugenio 2009; Hernault et al. 2010) is still considerably inferior compared with the human gold standard. Thus, the impact of rhetorical structure in downstream NLP applications is still very limited. The work we present in this article aims to reduce this performance gap and take discourse parsing one step further. To this end, we address three key limitations of existing discourse parsers.
First, existing discourse parsers typically model the structure and the labels of a DT separately, and also do not take into account the sequential dependencies between the DT constituents. However, for several NLP tasks, it has recently been shown that joint models typically outperform independent or pipeline models (Murphy 2012, page 687). This is also supported in a recent study by Feng and Hirst (2012), in which the performance of a greedy bottom–up discourse parser improved when sequential dependencies were considered by using gold annotations for the neighboring (i.e., previous and next) discourse units as contextual features in the parsing model. To address this limitation of existing parsers, as the first contribution, we propose a novel discourse parser based on probabilistic discriminative parsing models, expressed as Conditional Random Fields (CRFs) (Sutton, McCallum, and Rohanimanesh 2007), to infer the probability of all possible DT constituents. The CRF models effectively represent the structure and the label of a DT constituent jointly, and, whenever possible, capture the sequential dependencies.
Second, existing discourse parsers typically apply greedy and sub-optimal parsing algorithms to build a DT. To cope with this limitation, we use the inferred (posterior) probabilities from our CRF parsing models in a probabilistic CKY-like bottom–up parsing algorithm (Jurafsky and Martin 2008), which is non-greedy and optimal. Furthermore, a simple modification of this parsing algorithm allows us to generate k-best (i.e., the k highest probability) parse hypotheses for each input text that could then be used in a reranker to improve over the initial ranking using additional (global) features of the discourse tree as evidence, a strategy that has been successfully explored in syntactic parsing (Charniak and Johnson 2005; Collins and Koo 2005).
Third, most of the existing discourse parsers do not discriminate between intrasentential parsing (i.e., building the DTs for the individual sentences) and multisentential parsing (i.e., building the DT for the whole document). However, we argue that distinguishing between these two parsing conditions can result in more effective parsing. Two separate parsing models could exploit the fact that rhetorical relations
are distributed differently intra-sententially versus multi-sententially. Also, they could independently choose their own informative feature sets. As another key contribution of our work, we devise two different parsing components: one for intra-sentential parsing, the other for multi-sentential parsing. This provides for scalable, modular, and flexible solutions that can exploit the strong correlation observed between the text structure (i.e., sentence boundaries) and the structure of the discourse tree.
In order to develop a complete and robust discourse parser, we combine our intrasentential and multi-sentential parsing components in two different ways. Because most sentences have a well-formed discourse sub-tree in the full DT (e.g., the second sentence in Figure 1), our first approach constructs a DT for every sentence using our intrasentential parser, and then runs the multi-sentential parser on the resulting sentencelevel DTs to build a complete DT for the whole document. However, this approach would fail in those cases where discourse structures violate sentence boundaries, also called “leaky” boundaries (Vliet and Redeker 2011). For example, consider the first sentence in Figure 1. It does not have a well-formed discourse sub-tree because the unit containing EDUs 2 and 3 merges with the next sentence and only then is the resulting unit merged with EDU 1. Our second approach, in order to deal with these leaky cases, builds sentence-level sub-trees by applying the intra-sentential parser on a sliding window covering two adjacent sentences and by then consolidating the results produced by overlapping windows. After that, the multi-sentential parser takes all these sentence-level sub-trees and builds a full DT for the whole document.
Our discourse parser assumes that the input text has already been segmented into elementary discourse units. As an additional contribution, we propose a novel discriminative approach to discourse segmentation that not only achieves state-of-theart performance, but also reduces time and space complexities by using fewer features. Notice that the combination of our segmenter with our parser forms a COmplete probabilistic Discriminative framework for Rhetorical Analysis (CODRA).
Whereas previous systems have been tested on only one corpus, we evaluate our framework on texts from two very different genres: news articles and instructional howto manuals. The results demonstrate that our approach to discourse parsing provides consistent and statistically significant improvements over previous methods both at the sentence level and at the document level. The performance of our final system compares very favorably to the performance of state-of-the-art discourse parsers. Finally, the oracle accuracy computed based on the k-best parse hypotheses generated by our parser demonstrates that a reranker could potentially improve the accuracy further.
After discussing related work in Section 2, we present our rhetorical analysis framework in Section 3. In Section 4, we describe our discourse parser. Then, in Section 5 we present our discourse segmenter. The experiments and analysis of results are presented in Section 6. Finally, we summarize our contributions with future directions in Section 7.","2 related work :Rhetorical analysis has a long history—dating back to Mann and Thompson (1988), when RST was initially proposed as a useful linguistic method for describing natural texts, to more recent attempts to automatically extract the rhetorical structure of a given text (Hernault et al. 2010). In this section, we provide a brief overview of the computational approaches that follow RST as the theory of discourse, and that are related to our work; see the survey by Stede (2011) for a broader overview that also includes other theories of discourse. Although the most effective approaches to rhetorical analysis to date rely on supervised machine learning methods trained on human-annotated data, unsupervised methods have also been proposed, as they do not require human-annotated data and can be more easily applied to new domains.
Often, discourse connectives like but, because, and although convey clear information on the kind of relation linking the two text segments. In his early work, Marcu (2000a) presented a shallow rule-based approach relying on discourse connectives (or cues) and surface patterns. He used hand-coded rules, derived from an extensive corpus study, to break the text into EDUs and to build DTs for sentences first, then for paragraphs, and so on. Despite the fact that this work pioneered the field of rhetorical analysis, it has many limitations. First, identifying discourse connectives is a difficult task on its own, because (depending on the usage), the same phrase may or may not signal a discourse relation (Pitler and Nenkova 2009). For example, but can either signal a Contrast discourse relation or can simply perform non-discourse acts. Second, discourse segmentation using only discourse connectives fails to attain high accuracy (Soricut and Marcu 2003). Third, DT structures do not always correspond to paragraph structures; for example, Sporleder and Lapata (2004) report that more than 20% of the paragraphs in the RST–DT corpus (Carlson, Marcu, and Okurowski 2002) do not correspond to a discourse unit in the DT. Fourth, discourse cues are sometimes ambiguous; for example, but can signal Contrast, Antithesis and Concession, and so on.
Finally, a more serious problem with the rule-based approach is that often rhetorical relations are not explicitly signaled by discourse cues. For example, in RST–DT, Marcu and Echihabi (2002) found that only 61 out of 238 Contrast relations and 79 out of 307 Cause–Explanation relations were explicitly signaled by cue phrases. In the British National Corpus, Sporleder and Lascarides (2008) report that half of the sentences lack a discourse cue. Other studies (Schauer and Hahn 2001; Stede 2004; Taboada 2006; Subba and Di-Eugenio 2009) report even higher figures: About 60% of discourse relations are not explicitly signaled. Therefore, rather than relying on hand-coded rules based on discourse cues and surface patterns, recent approaches use machine learning techniques with a large set of informative features.
While some rhetorical relations need to be explicitly signaled by discourse cues (e.g., Concession) and some do not (e.g., Background), there is a large middle ground of relations that may be signaled or not. For these “middle ground” relations, can we exploit features present in the signaled cases to automatically identify relations when they are not explicitly signaled? The idea is to use unambiguous discourse cues (e.g., although for Contrast, for example for Elaboration) to automatically label a large corpus with rhetorical relations that could then be used to train a supervised model.1
A series of previous studies have explored this idea. Marcu and Echihabi (2002) first attempted to identify four broad classes of relations: Contrast, Elaboration, Condition, and Cause–Explanation–Evidence. They used a naive Bayes classifier based on word pairs (w1, w2), where w1 occurs in the left segment, and w2 occurs in the right segment. Sporleder and Lascarides (2005) included other features (e.g., words and their stems, Part-of-Speech [POS] tags, positions, segment lengths) in a boosting-based classifier (i.e., BoosTexter [Schapire and Singer 2000]) to further improve relation classification accuracy. However, these studies evaluated classification performance on the instances
1 We categorize this approach as unsupervised because it does not rely on human-annotated data.
where rhetorical relations were originally signaled (i.e., the discourse cues were artificially removed), and did not verify how well this approach performs on the instances that are not originally signaled. Subsequent studies (Blair-Goldensohn, McKeown, and Rambow 2007; Sporleder 2007; Sporleder and Lascarides 2008) confirm that classifiers trained on instances stripped of their original discourse cues do not generalize well to implicit cases because they are linguistically quite different.
Note that this approach to identifying discourse relations in the absence of manually labeled data does not fully solve the parsing problem (i.e., building DTs); rather, it only attempts to identify a small subset of coarser relations between two (flat) text segments (i.e., a tagging problem). Arguably, to perform a complete rhetorical analysis, one needs to use supervised machine learning techniques based on human-annotated data. Marcu (1999) applies supervised machine learning techniques to build a discourse segmenter and a shift–reduce discourse parser. Both the segmenter and the parser rely on C4.5 decision tree classifiers (Poole and Mackworth 2010) to learn the rules automatically from the data. The discourse segmenter mainly uses discourse cues, shallowsyntactic (i.e., POS tags) and contextual features (i.e., neighboring words and their POS tags). To learn the shift–reduce actions, the discourse parser encodes five types of features: lexical (e.g., discourse cues), shallow-syntactic, textual similarity, operational (previous n shift–reduce operations), and rhetorical sub-structural features. Despite the fact that this work has pioneered many of today’s machine learning approaches to discourse parsing, it has all the limitations mentioned in Section 1.
The work of Marcu (1999) is considerably improved by Soricut and Marcu (2003). They present the publicly available SPADE system,2 which comes with probabilistic models for discourse segmentation and sentence-level discourse parsing. Their segmentation and parsing models are based on lexico-syntactic patterns (or features) extracted from the lexicalized syntactic tree of a sentence. The discourse parser uses an optimal parsing algorithm to find the most probable DT structure for a sentence. SPADE was trained and tested on the RST–DT corpus. This work, by showing empirically the connection between syntax and discourse structure at the sentence level, has greatly influenced all major contributions in this area ever since. However, it is limited in several ways. First, SPADE does not produce a full-text (i.e., document-level) parse. Second, it applies a generative parsing model based on only lexico-syntactic features, whereas discriminative models are generally considered to be more accurate, and can incorporate arbitrary features more effectively (Murphy 2012). Third, the parsing model makes an independence assumption between the label and the structure of a DT constituent, and it ignores the sequential and the hierarchical dependencies between the DT constituents.
Subsequent research addresses the question of how much syntax one really needs in rhetorical analysis. Sporleder and Lapata (2005) focus on the discourse chunking problem, comprising two subtasks: discourse segmentation and (flat) nuclearity assignment. They formulate discourse chunking in two alternative ways. First, one-step classification, where the discourse chunker, a multi-class classifier, assigns to each token one of the four labels: (1) B–NUC (beginning of a nucleus), (2) I–NUC (inside a nucleus), (3) B– SAT (beginning of a satellite), and (4) I–SAT (inside a satellite). Therefore, this approach performs discourse segmentation and nuclearity assignment simultaneously. Second,
2 http://www.isi.edu/licensed-sw/spade/.
two-step classification, where in the first step, the discourse segmenter (a binary classifier) labels each token as either B (beginning of an EDU) or I (inside an EDU). Then, in the second step, a nuclearity labeler (another binary classifier) assigns a nuclearity status to each segment. The two-step approach avoids illegal chunk sequences like a B–NUC followed by an I–SAT or a B–SAT followed by an I–NUC, and in this approach, it is easier to incorporate sentence-level properties like the constraint that a sentence must contain at least one nucleus. They examine whether shallow-syntactic features (e.g., POS and phrase tags) would be sufficient for these purposes. The evaluation on the RST–DT shows that the two-step approach outperforms the one-step approach, and its performance is comparable to that of SPADE, which requires relatively expensive full syntactic parses.
In follow–up work, Fisher and Roark (2007) demonstrate over 4% absolute performance gain in discourse segmentation, by combining the features extracted from the syntactic tree with the ones derived via POS tagging and shallow syntactic parsing (i.e., chunking). Using quite a large number of features in a binary log-linear model, they achieve state-of-the-art performance in discourse segmentation on the RST–DT test set.
In a different approach, Regneri, Egg, and Koller (2008) propose to use Underspecified Discourse Representation (UDR) as an intermediate representation for discourse parsing. Underspecified representations offer a single compact representation to express possible ambiguities in a linguistic structure, and have been primarily used to deal with scope ambiguity in semantic structures (Reyle 1993; Egg, Koller, and Niehren 2001; Althaus et al. 2003; Koller, Regneri, and Thater 2008). Assuming that a UDR of a DT is already given in the form of a dominance graph (Althaus et al. 2003), Regneri, Egg, and Koller (2008) convert it into a more expressive and complete UDR representation called regular tree grammar (Koller, Regneri, and Thater 2008), for which efficient algorithms (Knight and Graehl 2005) already exist to derive the best configuration (i.e., the best discourse tree).
Hernault et al. (2010) present the publicly available HILDA system,3 which comes with a discourse segmenter and a parser based on Support Vector Machines (SVMs). The discourse segmenter is a binary SVM classifier that uses the same lexico-syntactic features used in SPADE, but with more context (i.e., the lexico-syntactic features for the previous two words and the following two words). The discourse parser iteratively uses two SVM classifiers in a pipeline to build a DT. In each iteration, a binary classifier first decides which of the adjacent units to merge, then a multi-class classifier connects the selected units with an appropriate relation label. Using this simple method, they report promising results in document-level discourse parsing on the RST–DT.
For a different genre, instructional texts, Subba and Di-Eugenio (2009) propose a shift–reduce discourse parser that relies on a classifier for relation labeling. Their classifier uses Inductive Logic Programming (ILP) to learn first-order logic rules from a large set of features including the linguistically rich compositional semantics coming from a semantic parser. They demonstrate that including compositional semantics with other features improves the performance of the classifier, thus, also improves the performance of the parser.
Both HILDA and the ILP-based approach of Subba and Di-Eugenio (2009) are limited in several ways. First, they do not differentiate between intra- and multi-sentential
3 http://nlp.prendingerlab.net/hilda/.
parsing, and both scenarios use a single uniform parsing model. Second, they take a greedy (i.e., sub-optimal) approach to construct a DT. Third, they disregard sequential dependencies between DT constituents. Furthermore, HILDA considers the structure and the labels of a DT separately. Our discourse parser CODRA, as described in the next section, addresses all these limitations.
More recent work than ours also attempts to address some of the above-mentioned limitations of the existing discourse parsers. Similar to us, Feng and Hirst (2014) generate a document-level DT in two stages, where a multi-sentential parsing follows an intra-sentential one. At each stage, they iteratively use two separate linear-chain CRFs (Lafferty, McCallum, and Pereira 2001) in a cascade: one for predicting the presence of rhetorical relations between adjacent discourse units in a sequence, and the other to predict the relation label between the two most probable adjacent units to be merged as selected by the previous CRF. While they use CRFs to take into account the sequential dependencies between DT constituents, they use them greedily during parsing to achieve efficiency. They also propose a greedy post-editing step based on an additional feature (i.e., depth of a discourse unit) to modify the initial DT, which gives them a significant gain in performance. In a different approach, Li et al. (2014) propose a discourse-level dependency structure to capture direct relationships between EDUs rather than deep hierarchical relationships. They first create a discourse dependency treebank by converting the deep annotations in RST–DT to shallow head-dependent annotations between EDUs. To find the dependency parse (i.e., an optimal spanning tree) for a given text, they apply Eisner (1996) and Maximum Spanning Tree (McDonald et al. 2005) dependency parsing algorithms with the Margin Infused Relaxed Algorithm online learning framework (McDonald, Crammer, and Pereira 2005).
With the successful application of deep learning to numerous NLP problems including syntactic parsing (Socher et al. 2013a), sentiment analysis (Socher et al. 2013b), and various tagging tasks (Collobert et al. 2011), a couple of recent studies in discourse parsing also use deep neural networks (DNNs) and related feature representation methods. Inspired by the work of Socher et al. (2013a, 2013b), Li, Li, and Hovy (2014) propose a recursive DNN for discourse parsing. However, as in Socher et al. (2013a, 2013b), word vectors (i.e., embeddings) are not learned explicitly for the task, rather they are taken from Collobert et al. (2011). Given the vectors of the words in an EDU, their model first composes them hierarchically based on a syntactic parse tree to get the vector representation for the EDU. Adjacent discourse units are then merged hierarchically to get the vector representations for the higher order discourse units. In every step, the merging is done using one binary (structure) and one multi-class (relation) classifier, each having a three-layer neural network architecture. The cost function for training the model is given by these two cascaded classifiers applied at different levels of the DT. Similar to our method, they use the classifier probabilities in a CKY-like parsing algorithm to find the global optimal DT. Finally, Ji and Eisenstein (2014) present a feature representation learning method in a shift–reduce discourse parser (Marcu 1999). Unlike DNNs, which learn non-linear feature transformations in a maximum likelihood model, they learn linear transformations of features in a max margin classification model.","3 overview of our rhetorical analysis framework :CODRA takes as input a raw text and produces a discourse tree that describes the text in terms of coherence relations that hold between adjacent discourse units (i.e., clauses, sentences) in the text. An example DT generated by an online demo of CODRA
is shown in Appendix A.4 The color of a node represents its nuclearity status: blue denoting nucleus and yellow denoting satellite. The demo also allows some useful user interactions—for example, collapsing or expanding a node, highlighting an EDU, and so on.5
CODRA follows a pipeline architecture, shown in Figure 2. Given a raw text, the first task in the rhetorical analysis pipeline is to break the text into a sequence of EDUs (i.e., discourse segmentation). Because it is taken for granted that sentence boundaries are also EDU boundaries (i.e., EDUs do not span across multiple sentences), the discourse segmentation task boils down to finding EDU boundaries inside sentences. CODRA uses a maximum entropy model for discourse segmentation (see Section 5).
Once the EDUs are identified, the discourse parsing problem is determining which discourse units (EDUs or larger units) to relate (i.e., the structure), and what relations (i.e., the labels) to use in the process of building the DT. Specifically, discourse parsing requires: (1) a parsing model to explore the search space of possible structures and labels for their nodes, and (2) a parsing algorithm for selecting the best parse tree(s) among the candidates. A probabilistic parsing model like ours assigns a probability to every possible DT. The parsing algorithm then picks the most probable DTs.
The existing discourse parsers (Marcu 1999; Soricut and Marcu 2003; Subba and Di-Eugenio 2009; Hernault et al. 2010) described in Section 2 use parsing models that disregard the structural interdependencies between the DT constituents. However, we hypothesize that, like syntactic parsing, discourse parsing is also a structured prediction problem, which involves predicting multiple variables (i.e., the structure and the relation labels) that depend on each other (Smith 2011). Recently, Feng and Hirst (2012) also found these interdependencies to be critical for parsing performance. To capture the structural dependencies between the DT constituents, CODRA uses undirected conditional graphical models (i.e., CRFs) as its parsing models.
To find the most probable DT, unlike most previous studies (Marcu 1999; Subba and Di-Eugenio 2009; Hernault et al. 2010), which adopt a greedy solution, CODRA applies an optimal CKY parsing algorithm to the inferred posterior probabilities (obtained from the CRFs) of all possible DT constituents. Furthermore, the parsing algorithm allows CODRA to generate a list of k-best parse hypotheses for a given text.
4 The demo of CODRA is available at http://109.228.0.153/Discourse Parser Demo/. The source code of CODRA is available from http://alt.qcri.org/tools/. 5 The input text in the demo in Appendix A is taken from www.bbc.co.uk/news/world-asia-26106490.
Note that the way CRFs and CKY are used in CODRA is quite different from the way they are used in syntactic parsing. For example, in the CRF-based constituency parsing proposed by Finkel, Kleeman, and Manning (2008), the conditional probability distribution of a parse tree given a sentence decomposes across factors defined over productions, and the standard inside–outside algorithm is used for inference on possible trees. In contrast, CODRA first uses the standard forward–backward algorithm in a “fat” chain structured6 CRF (to be discussed in Section 4.1.1) to compute the posterior probabilities of all possible DT constituents for a given text (i.e., EDUs); then it uses a CKY parsing algorithm to combine those probabilities and find the most probable DT.
Another crucial question related to parsing models is whether to use a single model or two different models for parsing at the sentence-level (i.e., intra-sentential) and at the document-level (i.e., multi-sentential). A simple and straightforward strategy would be to use a single unified parsing model for both intra- and multi-sentential parsing without distinguishing the two cases, as was previously done (Marcu 1999; Subba and Di-Eugenio 2009; Hernault et al. 2010). That approach has the advantages of making the parsing process easier, and the model gets more data to learn from. However, for a solution like ours, which tries to capture the interdependencies between constituents, this would be problematic with respect to scalability and inappropriate because of two modeling issues.
More specifically, for scalability note that the number of valid trees grows exponentially with the number of EDUs in a document.7 Therefore, an exhaustive search over all the valid DTs is often infeasible, even for relatively small documents.
For modeling, a single unified approach is inappropriate for two reasons. On the one hand, it appears that discourse relations are distributed differently intra- versus multi-sententially. For example, Figure 3 shows a comparison between the two distributions of the eight most frequent relations in the RST–DT training set. Notice that Same–Unit is more frequent than Joint in the intra-sentential case, whereas Joint is more frequent than Same–Unit in the multi-sentential case. Similarly, the relative distributions
6 By the term “fat” we refer to CRFs with multiple (interconnected) chains of output variables. 7 For n + 1 EDUs, the number of valid discourse tree structures (i.e., not counting possible variations in the
nuclearity and relation labels) is the Catalan number Cn.
of Background, Contrast, Cause, and Explanation are different in the two parsing scenarios. On the other hand, different kinds of features are applicable and informative for intraversus multi-sentential parsing. For example, syntactic features like dominance sets (Soricut and Marcu 2003) are extremely useful for parsing at the sentence-level, but are not even applicable in the multi-sentential case. Likewise, lexical chain features (Sporleder and Lapata 2004), which are useful for multi-sentential parsing, are not applicable at the sentence level.
Based on these above observations, CODRA comprises two separate modules: an intra-sentential parser and a multi-sentential parser, as shown in Figure 2. First, the intra-sentential parser produces one or more discourse sub-trees for each sentence. Then, the multi-sentential parser generates a full DT for the document from these sub-trees. Both of our parsers have the same two components: a parsing model and a parsing algorithm. Whereas the two parsing models are rather different, the same parsing algorithm is shared by the two modules. Staging multi-sentential parsing on top of intra-sentential parsing in this way allows CODRA to explicitly exploit the strong correlation observed between the text structure and the DT structure, as explained in detail in Section 4.3.","4 the discourse parser :Before describing the parsing models and the parsing algorithm of CODRA in detail, we introduce some terminology that we will use throughout this article.
A DT can be formally represented as a set of constituents of the form R[i, m, j], where i ≤ m < j. This refers to a rhetorical relation R between the discourse unit containing EDUs i through m and the discourse unit containing EDUs m+1 through j. For example, the DT for the second sentence in Figure 1 can be represented as {Elaboration–NS[4,4,5], Same–Unit–NN[4,5,6]}. Notice that in this representation, a relation R also specifies the nuclearity status of the discourse units involved, which can be one of Nucleus–Satellite (NS), Satellite–Nucleus (SN), or Nucleus–Nucleus (NN). Attaching nuclearity status to the relations allows us to perform the two subtasks of discourse parsing, relation identification and nuclearity assignment, simultaneously.
A common assumption made for generating DTs effectively is that they are binary trees (Soricut and Marcu 2003; Hernault et al. 2010). That is, multi-nuclear relations (e.g., Joint, Same–Unit) involving more than two discourse units are mapped to a hierarchical right-branching binary tree. For example, a flat Joint(e1, e2, e3, e4) (Figure 4a) is mapped to a right-branching binary tree Joint(e1, Joint(e2, Joint(e3, e4))) (Figure 4b). As mentioned before, the job of the intra- and multi-sentential parsing models of CODRA is to assign a probability to each of the constituents of all possible DTs at the sentence level and at the document level, respectively. Formally, given the model parameters Θ at a particular parsing scenario (i.e., sentence-level or document-level), for each possible constituent R[i, m, j] in a candidate DT at that parsing scenario, the parsing model estimates P(R[i, m, j]|Θ), which specifies a joint distribution over the label R and the structure [i, m, j] of the constituent. For example, when applied to the sentences in Figure 1 separately, the intra-sentential parsing model (with learned parameters Θs) estimates P(R[1, 1, 2]|Θs), P(R[2, 2, 3]|Θs), P(R[1, 2, 3]|Θs), and P(R[1, 1, 3]|Θs) for the first sentence, and P(R[4, 4, 5]|Θs), P(R[5, 5, 6]|Θs), P(R[4, 5, 6]|Θs), and P(R[4, 4, 6]|Θs) for the second sentence, respectively, for all R ranging over the set of relations.
4.1.1 Intra-Sentential Parsing Model. Figure 5 shows the parsing model of CODRA for intra-sentential parsing. The observed nodes Uj (at the bottom) in a sequence represent the discourse units (EDUs or larger units). The first layer of hidden nodes are the structure nodes, where Sj ∈ {0, 1} denotes whether two adjacent discourse units Uj−1 and Uj should be connected or not. The second layer of hidden nodes are the relation nodes, with Rj ∈ {1 . . .M} denoting the relation between two adjacent units Uj−1 and Uj, where M is the total number of relations in the relation set. The connections between adjacent nodes in a hidden layer encode sequential dependencies between the respective hidden nodes, and can enforce constraints such as the fact that a node must have a unique mother, namely, a Sj= 1 must not follow a Sj−1= 1. The connections between the two hidden layers model the structure and the relation of DT constituents jointly.
Notice that the probabilistic graphical model shown in Figure 5 is a chain-structured undirected graphical model (also known as Markov Random Field or MRF [Murphy 2012]) with two hidden layers, i.e., structure chain and relation chain. It becomes a Dynamic Conditional Random Field (DCRF) (Sutton, McCallum, and Rohanimanesh 2007) when we directly model the hidden (output) variables by conditioning the clique potentials (i.e., factors) on the observed (input) variables:
P(R2:t, S2:t|x,Θs) = 1Z(x,Θs) t−1∏ i=2 ϕ(Ri, Ri+1|x,Θs,r)ψ(Si, Si+1|x,Θs,s)ω(Ri, Si|x,Θs,c) (1)
where {ϕ} and {ψ} are the factors over the edges of the relation and structure chains, respectively, and {ω} are the factors over the edges connecting the relation and structure nodes (i.e., between-chain edges). Here, x represents input features extracted from the observed variables, Θs = [Θs,r,Θs,s,Θs,c] are model parameters, and Z(x,Θs) is the partition function. We use the standard log-linear representation of the factors:
ϕ(Ri, Ri+1|x,Θs,r) = exp(ΘTs,r f (Ri, Ri+1, x)) (2)
ψ(Si, Si+1|x,Θs,s) = exp(ΘTs,s f (Si, Si+1, x)) (3)
ω(Ri, Si|x,Θs,c) = exp(ΘTs,c f (Ri, Si, x)) (4)
where f (Y, Z, x) is a feature vector derived from the input features x and the local labels Y and Z, and Θs,y is the corresponding weight vector—that is, Θs,r and Θs,s are the weight vectors for the factors over the relation edges and the structure edges, respectively, and Θs,c is the weight vector for the factors over the between-chain edges.
A DCRF is a generalization of linear-chain CRFs (Lafferty, McCallum, and Pereira 2001) to represent complex interactions between output variables (i.e., labels), such as when performing multiple labeling tasks on the same sequence. Recently, there has been an explosion of interest in CRFs for solving structured output classification problems, with many successful applications in NLP including syntactic parsing (Finkel, Kleeman, and Manning 2008), syntactic chunking (Sha and Pereira 2003), and discourse chunking (Ghosh et al. 2011) in accordance with the Penn Discourse Treebank (Prasad et al. 2008).
DCRFs, being a discriminative approach to sequence modeling, have several advantages over their generative counterparts such as Hidden Markov Models (HMMs) and MRFs, which first model the joint distribution p(y, x|Θ), and then infer the conditional distribution p(y|x,Θ). It has been advocated that discriminative models are generally more accurate than generative ones because they do not “waste resources” modeling complex distributions that are observed (i.e., p(x)); instead, they focus directly on modeling what we care about, namely, the distribution of labels given the data (Murphy 2012).
Other key advantages include the ability to incorporate arbitrary overlapping local and global features, and the ability to relax strong independence assumptions. Furthermore, CRFs surmount the label bias problem (Lafferty, McCallum, and Pereira 2001) of the Maximum Entropy Markov Model (McCallum, Freitag, and Pereira 2000), which is considered to be a discriminative version of the HMM.
4.1.2 Training and Applying the Intra-Sentential Parsing Model. In order to obtain the probability of the constituents of all candidate DTs for a sentence, CODRA applies the intra-sentential parsing model (with learned parameters Θs) recursively to sequences at different levels of the DT, and computes the posterior marginals over the relation– structure pairs. It uses the standard forward–backward algorithm to compute the posterior marginals. To illustrate the process, let us assume that the sentence contains four EDUs, e1, · · · , e4 (see Figure 6). At the first (i.e., bottom) level of the DT, when all the discourse units are EDUs, there is only one unit sequence (e1, e2, e3, e4) to which CODRA applies the DCRF model. Figure 6a at the top left shows the corresponding DCRF model. For this sequence it computes the posterior marginals P(R2, S2=1|e1, e2, e3, e4,Θs),
P(R3, S3=1|e1, e2, e3, e4,Θs), and P(R4, S4=1| e1, e2, e3, e4,Θs) to obtain the probability of the DT constituents R[1, 1, 2], R[2, 2, 3], and R[3, 3, 4], respectively.
At the second level, there are three unit sequences: (e1:2, e3, e4), (e1,e2:3, e4), and (e1,e2,e3:4). Figure 6b shows their corresponding DCRF models. Notice that each of these sequences has a discourse unit that connects two EDUs, and the probability of this connection has already been computed at the previous level. CODRA computes the posterior marginals P(R3, S3=1|e1:2,e3,e4,Θs), P(R2:3S2:3=1|e1,e2:3,e4,Θs), P(R4, S4=1|e1,e2:3,e4,Θs), and P(R3:4, S3:4=1|e1,e2,e3:4,Θs) from these three sequences, which correspond to the probability of the constituents R[1, 2, 3], R[1, 1, 3], R[2, 3, 4], and R[2, 2, 4], respectively. Similarly, it attains the probability of the constituents R[1, 1, 4], R[1, 2, 4], and R[1, 3, 4] by computing their respective posterior marginals from the three sequences at the third (i.e., top) level of the candidate DTs (see Figure 6c).
Algorithm 1 describes how CODRA generates the unit sequences at different levels of the candidate DTs for a given number of EDUs in a sentence. Specifically, to compute the probability of a DT constituent R[i, k, j], CODRA generates sequences like (e1, · · · , ei−1, ei:k, ek+1:j, ej+1, · · · , en) for 1 ≤ i ≤ k < j ≤ n. However, in doing so, it may generate some duplicate sequences. Clearly, the sequence (e1, · · · , ei−1, ei:i, ei+1:j, ej+1, · · · , en) for 1 ≤ i ≤ k < j < n is already considered for computing the probability of the constituent R[i + 1, j, j + 1]. Therefore, it is a duplicate sequence that CODRA excludes from the list of sequences. The algorithm has a complexity of O(n3), where n is the number of EDUs in the sentence.
Once CODRA acquires the probability of all possible intra-sentential DT constituents, the discourse sub-trees for the sentences are built by applying an optimal parsing algorithm (Section 4.2) using one of the methods described in Section 4.3.
Algorithm 1 is also used to generate sequences for training the model (i.e., learning Θs). For example, Figure 7 demonstrates how we generate the training instances (right) from a gold DT with four EDUs (left). To find the relevant labels for the sequences
Algorithm 1 Generating unit sequences for a sentence with n EDUs. Input: Sequence of EDUs: (e1, e2, · · · , en) Output: List of sequences: L for i = 1 → n − 1 do // all possible starting positions for the subsequence
for j = i + 1 → n do // all possible ending positions for the subsequence if j == n then // sequences at top and bottom levels
for k = i → j − 1 do // all possible cut points within the subsequence L.append ((e1, · · · , ei−1, ei:k, ek+1:j, ej+1, · · · , en))
end else // sequences at intermediate levels
for k = i + 1 → j − 1 do // cut points excluding duplicate sequences L.append ((e1, · · · , ei−1, ei:k, ek+1:j, ej+1, · · · , en))
end end
end end
generated by the algorithm, we consult the gold DT and see if two discourse units are connected by a relation r (i.e., the corresponding labels are S = 1, R = r) or not (i.e., the corresponding labels are S = 0, R =NR). We train the model by maximizing the conditional likelihood of the labels in each of these training examples (see Equation (1)).
4.1.3 Multi-Sentential Parsing Model. Given the discourse units (sub-trees) for all the individual sentences in a document, a simple approach to build the DT of the document would be to apply a new DCRF model, similar to the one in Figure 5 (with different parameters), to all the possible sequences generated from these units by Algorithm 1 to infer the probability of all possible higher-order (multi-sentential) constituents. However, the number of possible sequences and their length increase with the number of sentences in a document. For example, assuming that each sentence has a well-formed DT, for a document with n sentences, Algorithm 1 generates O(n3) sequences, where
the sequence at the bottom level has n units, each of the sequences at the second level has n-1 units, and so on. Because the DCRF model in Figure 5 has a “fat” chain structure, one could use the forward–backward algorithm for exact inference in this model (Murphy 2012). Forward–backward on a sequence containing T units costs O(TM2) time, where M is the number of relations in our relation set. This makes the chainstructured DCRF model impractical for multi-sentential parsing of long documents, since learning requires running inference on every training sequence with an overall time complexity of O(TM2n3) = O(M2n4) per document (Sutton and McCallum 2012).
To address this problem, we have developed a simplified parsing model for multisentential parsing. Our model is shown in Figure 8. The two observed nodes Ut−1 and Ut are two adjacent (multi-sentential) discourse units. The (hidden) structure node S ∈ {0, 1} denotes whether the two discourse units should be linked or not. The other hidden node R ∈ {1 . . .M} represents the relation between the two units. Notice that similar to the model in Figure 5, this is also an undirected graphical model and becomes a CRF model if we directly model the labels by conditioning the clique potential ϕ on the input features x, derived from the observed variables:
P(Rt, St|x,Θd) = 1Z(x,Θd) ϕ(Rt, St|x,Θd) (5)
ϕ(Rt, St|x,Θd) = exp(ΘTd f (Rt, St, x)) (6)
where f (Rt, St, x) is a feature vector derived from the input features x and the labels Rt and St, and Θd is the corresponding weight vector. Although this model is similar in spirit to the parsing model in Figure 5, it now breaks the chain structure, which makes the inference much faster (i.e., a complexity of O(M2)). Breaking the chain structure also allows CODRA to balance the data for training (an equal number of instances with S=1 and S=0), which dramatically reduces the learning time of the model.
CODRA applies this parsing model to all possible adjacent units at all levels in the multi-sentential case, and computes the posterior marginals of the relation– structure pairs P(Rt, St=1|Ut−1, Ut,Θd) using the forward–backward algorithm to obtain the probability of all possible DT constituents. Given the sentence-level discourse units, Algorithm 2, which is a simplified variation of Algorithm 1, extracts all possible adjacent discourse units for multi-sentential parsing. Similar to Algorithm 1, Algorithm 2 also has a complexity of O(n3), where n is the number of sentence-level discourse units.
Algorithm 2 Generating all possible adjacent discourse units at all levels of a documentlevel discourse tree. Input: Sequence of units: (U1, U2, · · ·Un), where Ux[0]:= start EDU ID of unit x, and Ux[1]:= end EDU ID of unit x. Output: List of adjacent units: L for i = 1 → n − 1 do // all possible starting positions for the subsequence
for j = i + 1 → n do // all possible ending positions for the subsequence for k = i → j − 1 do // all possible cut points within the subsequence
Left = Ui[0] : Uk[1] Right = Uk+1[0] : Uj[1] L.append ((Left, Right))
end end
end
Both our intra- and multi-sentential parsing models are designed using MALLET’s graphical model toolkit GRMM (McCallum 2002). In order to avoid overfitting, we regularize the CRF models with l2 regularization and learn the model parameters using the limited-memory BFGS (L-BFGS) fitting algorithm.
4.1.4 Features Used in the Parsing Models. Crucial to parsing performance is the set of features used in the parsing models, as summarized in Table 1. We categorize the features into seven groups and specify which groups are used in what parsing model. Notice that some of the features are used in both models. Most of the features have been explored in previous studies (e.g., Soricut and Marcu 2003; Sporleder and Lapata 2005; Hernault et al. 2010). However, we improve some of these as explained subsequently.
The features are extracted from two adjacent discourse units Ut−1 and Ut. Organizational features encode useful information about text organization as shown by duVerle and Prendinger (2009). We measure the length of the discourse units as the number of EDUs and tokens in it. However, in order to better adjust to the length variations, rather than computing their absolute numbers in a unit, we choose to measure their relative numbers with respect to their total numbers in the two units. For example, if the two discourse units under consideration contain three EDUs in total, a unit containing two of the EDUs will have a relative EDU number of 0.67. We also measure the distances of the units in terms of the number of EDUs from the beginning and end of the sentence (or text in the multi-sentential case). Text structural features capture the correlation between text structure and rhetorical structure by counting the number of sentence and paragraph boundaries in the discourse units.
Discourse cues (e.g., because, but), when present, signal rhetorical relations between two text segments, and have been used as a primary source of information in earlier studies (Knott and Dale 1994; Marcu 2000a). However, recent studies (Hernault et al. 2010; Biran and Rambow 2011) suggest that an empirically acquired lexical N-gram dictionary is more effective than a fixed list of cue phrases, since this approach is domain independent and capable of capturing non-lexical cues such as punctuation.
In order to build a lexical N-gram dictionary empirically from the training corpus, we extract the first and last N tokens (N∈{1, 2, 3}) of each discourse unit and rank them according to their mutual information with the two labels, Structure (S) and Relation (R).
8 N-gram features N∈{1, 2, 3} Intra & Multi-Sentential
Beginning (or end) lexical N-grams in unit 1. Beginning (or end) lexical N-grams in unit 2. Beginning (or end) POS N-grams in unit 1. Beginning (or end) POS N-grams in unit 2.
5 Dominance set features Intra-Sentential
Syntactic labels of the head node and the attachment node. Lexical heads of the head node and the attachment node. Dominance relationship between the two units.
9 Lexical chain features Multi-Sentential
Number of chains spanning unit 1 and unit 2. Number of chains start in unit 1 and end in unit 2. Number of chains start (or end) in unit 1 (or in unit 2). Number of chains skipping both unit 1 and unit 2. Number of chains skipping unit 1 (or unit 2).
2 Contextual features Intra & Multi-Sentential
Previous and next feature vectors.
2 Sub-structural features Intra & Multi-Sentential
Root nodes of the left and right rhetorical sub-trees.
More specifically, given an N-gram x, we compute its conditional entropy H with respect to S and R as follows:8
H(S, R|x) = − ∑
s∈S r∈R log c(x, s, r)c(x) (7)
where c(x) is the empirical count of N-gram x, and c(x, s, r) is the joint empirical count of N-gram x with the labels s and r. This is in contrast to HILDA (Hernault et al. 2010), which ranks the N-grams by their frequencies in the training corpus. However, Blitzer
8 The higher the conditional entropy, the lower the mutual information, and vice versa.
(2008) found mutual information to be more effective than frequency as a method for feature selection. Intuitively, the most informative discourse cues are not only the most frequent, but also the ones that are indicative of the labels in the training data. In addition to the lexical N-grams we also encode the POS tags of the first and last N tokens (N∈{1, 2, 3}) in a discourse unit as shallow-syntactic features in our models.
Lexico-syntactic features dominance sets extracted from the Discourse Segmented Lexicalized Syntactic Tree (DS-LST) of a sentence have been shown to be extremely effective for intra-sentential discourse parsing in SPADE (Soricut and Marcu 2003). Figure 9a shows the DS-LST (i.e., lexicalized syntactic tree with EDUs identified) for a sentence with three EDUs from the RST–DT corpus, and Figure 9b shows the corresponding discourse tree. In a DS-LST, each EDU except the one with the root node must have a head node NH that is attached to an attachment node NA residing in a separate EDU. A dominance set D (shown at the bottom of Figure 9a) contains these attachment points (shown in boxes) of the EDUs in a DS-LST. In addition to the syntactic and lexical information of the head and attachment nodes, each element in the dominance set also includes a dominance relationship between the EDUs involved; the EDU with the attachment node dominates (represented by “>”) the EDU with the head node.
Soricut and Marcu (2003) hypothesize that the dominance set (i.e., lexical heads, syntactic labels, and dominance relationships) carries the most informative clues for intra-sentential parsing. For instance, the dominance relationship between the EDUs in our example sentence is 3 > 1 > 2, which favors the DT structure [1, 1, 2] over [2, 2, 3]. In order to extract dominance set features for two adjacent discourse units Ut−1 and Ut, containing EDUs ei:j and ej+1:k, respectively, we first compute the dominance set from the DS-LST of the sentence. We then extract the element from the set that holds across the EDUs j and j + 1. In our example, for the two units, containing EDUs e1 and e2, respectively, the relevant dominance set element is (1, efforts/NP)>(2, to/S). We encode the syntactic labels and lexical heads of NH and NA, and the dominance relationship as features in our intra-sentential parsing model.
Lexical chains (Morris and Hirst 1991) are sequences of semantically related words that can indicate topical boundaries in a text (Galley et al. 2003; Joty, Carenini, and Ng 2013). Features extracted from lexical chains are also shown to be useful for finding paragraph-level discourse structure (Sporleder and Lapata 2004). For example, consider the text with four paragraphs (P1 to P4) in Figure 10a. Now, let us assume that there is a lexical chain that spans the whole text, skipping paragraphs P2 and P3, while a second chain only spans P2 and P3. This situation makes it more likely that P2 and P3 should be linked in the DT before either of them is linked with another paragraph. Therefore, the DT structure in Figure 10b should be more likely than the structure in Figure 10c.
One challenge in computing lexical chains is that words can have multiple senses, and semantic relationships depend on the sense rather than the word itself. Several methods have been proposed to compute lexical chains (Barzilay and Elhadad 1997; Hirst and St. Onge 1997; Silber and McCoy 2002; Galley and McKeown 2003). We follow the state-of-the-art approach proposed by Galley and McKeown (2003), which extracts lexical chains after performing Word Sense Disambiguation (WSD).
In the preprocessing step, we extract the nouns from the document and lemmatize them using WordNet’s built-in morphy function (Fellbaum 1998). Then, by looking up in WordNet we expand each noun to all of its senses, and build a Lexical Semantic Relatedness Graph (LSRG) (Galley and McKeown 2003; Chali and Joty 2007). In an LSRG, the nodes represent noun-tokens with their candidate senses, and the weighted edges between senses of two different tokens represent one of the three semantic relations: repetition, synonym, and hypernym. For example, Figure 11a shows a partial LSRG, where the token bank has two possible senses, namely, money bank and river bank. Using the money bank sense, bank is connected with institution and company by hypernymy relations (edges marked with H), and with another bank by a repetition relation (edges marked with R). Similarly, using the river bank sense, it is connected with riverside by a hypernymy relation and with bank by a repetition relation. Nouns that are not found in WordNet are considered as proper nouns having only one sense, and are connected by only repetition relations.
We use this LSRG first to perform WSD, then to construct lexical chains. For WSD, the weights of all edges leaving the nodes under their different senses are summed up and the one with the highest score is considered to be the right sense for the wordtoken. For example, if repetition and synonymy are weighted equally, and hypernymy is given half as much weight as either of them, the score of bank’s two senses are: 1 + 0.5 + 0.5 = 2 for the sense money bank and 1 + 0.5 = 1.5 for the sense river bank. Therefore, the selected sense for bank in this context is river bank. In case of a tie, we select the sense that is most frequent (i.e., the first sense in WordNet). Note that this approach to WSD is different from that of Sporleder and Lapata (2004), which takes a greedy approach.
Finally, we prune the graph by only keeping the links that connect words with the selected senses. At the end of the process, we are left with the edges that form the actual lexical chains. For example, Figure 11b shows the result of pruning the graph in Figure 11a. The lexical chains extracted from the pruned graph are shown in the box at the bottom. Following Sporleder and Lapata (2004), for each chain element, we keep track of the location (i.e., sentence ID) in the text where that element was found, and exclude chains containing only one element. Given two discourse units, we count the number of chains that: hit the two units, exclusively hit the two units, skip both units, skip one of the units, start in a unit, and end in a unit.
We also consider more contextual information by including the above features computed for the neighboring adjacent discourse unit pairs in the current feature vector. For example, the contextual features for units Ut−1 and Ut include the feature vector computed from Ut−2 and Ut−1 and the feature vector computed from Ut and Ut+1.
We incorporate hierarchical dependencies between the constituents in a DT by rhetorical sub-structural features. For two adjacent units Ut−1 and Ut, we extract the roots of the two rhetorical sub-trees. For example, the root of the rhetorical sub-tree spanning over EDUs e1:2 in Figure 9b is Elaboration–NS. However, extraction of these features assumes the presence of labels for the sub-trees, which is not the case when we apply the parser to a new text (sentence or document) in order to build its DT in a nongreedy fashion. One way to deal with this is to loop twice through the parsing process using two different parsing models—one trained with the complete feature set, and the other trained without the sub-structural features. We first build an initial, sub-optimal DT using the parsing model that is trained without the sub-structural features. This intermediate DT will now provide labels for the sub-structures. Next we can build a final, more accurate DT by using the complete parsing model. This idea of twopass discourse parsing, where the second pass performs post-editing using additional features, has recently been adopted by Feng and Hirst (2014) in their greedy parser.
One could even continue doing post-editing multiple times until the DT converges. However, this could be very time consuming as each post-editing pass requires: (1) applying the parsing model to every possible unit sequence and computing the posterior marginals for all possible DT constituents, and (2) using the parsing algorithm to find the most probable DT. Recall from our earlier discussion in Section 4.1.3 that for n discourse units and M rhetorical relations, the first step requires O(M2n4) and O(M2n3) for intra- and multi-sentential parsing, respectively; we will see in the next section that the second step requires O(Mn3). In spite of the computational cost, the gain we attained in the subsequent passes was not significant for our development set. Therefore, we restrict our parser to only one-pass post-editing.
Note that in parsing models where the score (i.e., likelihood) of a parse tree decomposes across local factors (e.g., the CRF-based syntactic parser of Finkel, Kleeman, and Manning [2008]), it is possible to define a semiring using the factors and the local scores (e.g., given by the inside algorithm). The CKY algorithm could then give the optimal parse tree in a single post-editing pass (Smith 2011). However, because our intra-sentential parsing model is designed to capture sequential dependencies between DT constituents, the score of a DT does not directly decompose across factors over discourse productions. Therefore, designing such a semiring was not possible in our case.
In addition to these features, we also experimented with other features including WordNet-based lexical semantics, subjectivity, and TF.IDF-based cosine similarity. However, because such features did not improve parsing performance on our development set, they were excluded from our final set of features. The intra- and multi-sentential parsing models of CODRA assign a probability to every possible DT constituent in their respective parsing scenarios. The job of the parsing algorithm is then to find the k most probable DTs for a given text. We implement a probabilistic CKY-like bottom–up parsing algorithm that uses dynamic programming to compute the most likely parses (Jurafsky and Martin 2008). For simplicity, we first
describe the specific case of generating the single most probable DT, then we describe how to generalize this algorithm to produce the k most probable DTs for a given text.
Formally, the search problem for finding the most probable DT can be written as
DT∗ = argmax DT P(DT|Θ) (8)
where Θ specifies the parameters of the parsing model (intra- or multi-sentential). Given n discourse units, our parsing algorithm uses the upper-triangular portion of the n×n dynamic programming table D, where cell D[i, j] (for i < j) stores:
D[i, j] = P(r∗[Ui(0), Um∗ (1), Uj(1)]) (9)
where Ux(0) and Ux(1) are the start and end EDU Ids of discourse unit Ux, and
(m∗, r∗) = argmax i≤m<j ; R∈{1···M} P(R[Ui(0), Um(1), Uj(1)]) × D[i, m] × D[m + 1, j] (10)
Recall that the notation R[Ui(0), Um(1), Uj(1)] in this expression refers to a rhetorical relation R that holds between the discourse unit containing EDUs Ui(0) through Um(1) and the unit containing EDUs Um(1) + 1 through Uj(1).
In addition to D, which stores the probability of the most probable constituents of a DT, the algorithm also simultaneously maintains two other n×n dynamic programming tables S and R for storing the structure (i.e., Um∗ (1)) and the relations (i.e., r∗) of the corresponding DT constituents, respectively. For example, given four EDUs e1 · · · e4, the S and R dynamic programming tables at the left side in Figure 12 together represent the DT shown at the right. More specifically, to find the DT, we first look at the top-right entries in the two tables, and find S[1, 4] = 2 and R[1, 4] = r2, which specify that the two discourse units e1:2 and e3:4 should be connected by the relation r2 (the root in the DT). Then, we see how EDUs e1 and e2 should be connected by looking at the entries S[1, 2] and R[1, 2], and find S[1, 2] = 1 and R[1, 2] = r1, which indicates that these two units should be connected by the relation r1 (the left pre-terminal in the DT). Finally, to see how EDUs e3 and e4 should be linked, we look at the entries S[3, 4] and R[3, 4], which tell us that they should be linked by the relation r4 (the right pre-terminal). The algorithm
works in polynomial time. Specifically, for n discourse units and M number of relations, the time and space complexities are O(n3M) and O(n2), respectively.
A key advantage of using a probabilistic parsing algorithm like the one we use is that it allows us to generate a list of k most probable parse trees. It is straightforward to generalize the above algorithm to produce k most probable DTs. Specifically, when filling up the dynamic programming tables, rather than storing a single best parse for each sub-tree, we store and keep track (i.e., using back-pointers) of k-best candidates simultaneously. One can show that the time and space complexities of the k-best version of the algorithm are O(n3Mk2 log k) and O(k2n), respectively (Huang and Chiang 2005).
Note that, in contrast to other document-level discourse parsers (Marcu 2000b; Subba and Di-Eugenio 2009; Hernault et al. 2010; Feng and Hirst 2012, 2014), which use a greedy algorithm, CODRA finds a discourse tree that is globally optimal.9 This approach of CODRA is also different from the sentence-level discourse parser SPADE (Soricut and Marcu 2003). SPADE first finds the tree structure that is globally optimal, then it assigns the most probable relations to the internal nodes. More specifically, the cell D[i, j] in SPADE’s dynamic programming table stores
D[i, j] = P([Ui(0), Um∗ (1), Uj(1)]) (11)
where m∗ = argmax i≤m<j P([Ui(0), Um(1), Uj(1)]). Disregarding the relation label R while populating D, this approach may find a discourse tree that is not globally optimal. Now that we have presented our intra-sentential and multi-sentential parsing components, we are ready to describe how they can be effectively combined in a unified framework (Figure 2) to perform document-level rhetorical analysis. Recall that a key motivation for a two-stage10 parsing is that it allows us to capture the strong correlation between text structure and discourse structure in a scalable, modular, and flexible way. In the following, we describe two different approaches to model this correlation.
4.3.1 1S–1S (1 Sentence–1 Sub-tree). A key finding from previous studies on sentencelevel discourse analysis is that most sentences have a well-formed discourse sub-tree in the full document-level DT (Soricut and Marcu 2003; Fisher and Roark 2007). For example, Figure 13a shows 10 EDUs in three sentences (see boxes), where the DTs for the sentences obey their respective sentence boundaries.
Our first approach, called 1S–1S (1 Sentence–1 Sub-tree), aims to maximally exploit this finding. It first constructs a DT for every sentence using our intra-sentential parser, and then it provides our multi-sentential parser with the sentence-level DTs to build the rhetorical parse for the whole document.
4.3.2 Sliding Window. Although the assumption made by 1S–1S clearly simplifies the parsing process, it completely ignores the cases where rhetorical structures violate
9 We agree that with potentially sub-optimal, sub-structural features in the parsing model, CKY may end up finding a sub-optimal DT. But that is a separate issue.
10 Do not confuse the term two-stage with the term two-pass.
sentence boundaries. For example, in the DT shown in Figure 13b, sentence S2 does not have a well-formed sub-tree because some of its units attach to the left (i.e., 4–5 and 6) and some to the right (i.e., 7). Vliet and Redeker (2011) call these cases “leaky” boundaries.
Although we find fewer than 5% of the sentences in the RST–DT have leaky boundaries, in other corpora this can be true for a larger portion of the sentences. For example, we observe that over 12% of the sentences in the instructional corpus of Subba and Di-Eugenio (2009) have leaky boundaries. However, we notice that in most cases where DT structures violate sentence boundaries, its units are merged with the units of its adjacent sentences, as in Figure 13b. For example, this is true for 75% of the leaky cases in our development set containing 20 news articles from the RST–DT and for 79% of the leaky cases in our development set containing 20 how-to manuals from the instructional corpus. Based on this observation, we propose a sliding window approach.
In this approach, our intra-sentential parser works with a window of two consecutive sentences, and builds a DT for the two sentences. For example, given the three sentences in Figure 13, our intra-sentential parser constructs a DT for S1–S2 and a DT for S2–S3. In this process, each sentence in a document except the boundary sentences (i.e., the first and the last) will be associated with two DTs: one with the previous sentence (say, DTp) and one with the next (say, DTn). In other words, for each nonboundary sentence, we will have two decisions: one from DTp and one from DTn. Our parser consolidates the two decisions and generates one or more sub-trees for each sentence by checking the following three mutually exclusive conditions one after another:r Same in both: If the sentence under consideration has the same (in both
structure and labels) well-formed sub-tree in both DTp and DTn, we take this sub-tree. For example, in Figure 14a, S2 has the same sub-tree in the two DTs (one for S1–S2 and one for S2–S3). The two decisions agree on the DT for the sentence.r Different but no cross: If the sentence under consideration has a wellformed sub-tree in both DTp and DTn, but the two sub-trees vary either in structure or in labels, we pick the most probable one. For example, consider the DT for S1–S2 (at the left) in Figure 14a and the DT for S2–S3
in Figure 14b. In both cases S2 has a well-formed sub-tree, but they differ in structure. We pick the sub-tree which has the higher probability in the two dynamic programming tables.r Cross: If either or both of DTp and DTn segment the sentence into multiple sub-trees, we pick the one having more sub-trees. For example, consider the two DTs in Figure 14c. In the DT for S1–S2 on the left, S2 has three sub-trees (4–5, 6, 7), whereas in the DT for S2–S3 on the right, it has two (4–6, 7). So, we extract the three sub-trees for S2 from the first DT. If the sentence has the same number of sub-trees in both DTp and DTn, we pick the one with higher probability in the dynamic programming tables. Note that our choice of picking the DT with more sub-trees is intended to allow the parser to find more leaky cases. However, other heuristics are also possible. For example, another simple heuristic that one could try is: When both DTs segment the sentence into multiple sub-trees, pick the one with fewer sub-trees, and when only one of the DTs segment the sentence into multiple sub-trees, pick that one.
At the end, the multi-sentential parser takes all these sentence-level sub-trees for a document, and builds a full rhetorical parse for the whole document.","5 the discourse segmenter :Our discourse parser assumes that the input text has been already segmented into a sequence of EDUs. However, discourse segmentation is also a challenging problem, and previous studies (Soricut and Marcu 2003; Fisher and Roark 2007) have identified
it as a primary source of inaccuracy for discourse parsing. Regardless of its importance in discourse parsing, discourse segmentation itself can be useful in several NLP applications, including sentence compression (Sporleder and Lapata 2005) and textual alignment in statistical machine translation (Stede 2011). Therefore, in CODRA, we have developed our own discourse segmenter, which not only achieves state-of-theart performance as shown later, but also reduces the time complexity by using fewer features. The discourse segmenter in CODRA implements a binary classifier to decide for each word–token (except the last) in a sentence, whether to place an EDU boundary after that token. We use a maximum entropy model to build a discriminative classifier. More specifically, we use a Logistic Regression classifier with parameter Θ:
P(y|w,Θ) = Ber(y| Sigm(ΘTx)) (12)
where the output y ∈ {0, 1} denotes whether to put an EDU boundary (i.e., y = 1) or not (i.e., y = 0) after the word–token w, which is represented by a feature vector x. In the equation, Ber(η) and Sigm(η) refer to the Bernoulli distribution and the Sigmoid (also known as logistic) function, respectively. The negative log-likelihood (NLL) of the model with l2 regularization for N data points (i.e., word–tokens) is given by
NLL(θ) = − N∑
i=1
yi log Sigm(ΘTxi) + (1 − yi) log (1 − Sigm(ΘTxi)) + λΘTΘ (13)
where yi is the gold label for word–token wi (represented by feature vector xi). We learn the model parameters Θ using the L-BFGS fitting algorithm, which is time- and spaceefficient. To avoid overfitting, we use 5-fold cross validation to learn the regularization strength parameter λ from the training data. We also use a simple bagging technique (Breiman 1996) to deal with the sparsity of boundary (i.e., y = 1) tags.
Note that our first attempt at the discourse segmentation task implemented a linearchain CRF model (Lafferty, McCallum, and Pereira 2001) to capture the sequential dependencies between the tags in a discriminative way. However, the binary Logistic Regression classifier, using the same set of features, not only outperforms the CRF model, but also reduces time and space complexity. One possible explanation for the low performance of the CRF model is that Markov dependencies between tags cannot be effectively captured due to the sparsity of boundary tags. Also, because we could not balance the data by using techniques like bagging in the CRF model, this further degrades the performance. Our set of features for discourse segmentation are mostly inspired from previous studies but used in a novel way, as we describe here.
Our first subset of features, which we call SPADE features, includes the lexicosyntactic patterns extracted from the lexicalized syntactic tree of the given sentence.
These features replicate the features used in SPADE’s segmenter, but used in a discriminative way. In order to decide on an EDU boundary after a word–token wk, we search for the lowest constituent in the lexicalized syntactic tree that spans over tokens wi . . .wj such that i ≤ k < j. The production that expands this constituent in the tree, with the potential EDU boundary marked, forms the primary feature. For instance, to determine the existence of an EDU boundary after the word efforts in our sample sentence shown in Figure 9, the production NP(efforts) → PRP$(its) NNS(efforts) ↑ S(to) extracted from the lexicalized syntactic tree in Figure 9a constitutes the primary feature, where ↑ denotes the potential EDU boundary.
SPADE predicts an EDU boundary if the relative frequency (i.e., maximum likelihood estimate) of a potential boundary given the production in the training data is greater than 0.5. If the production has not been observed frequently enough, the unlexicalized version of the production (e.g., NP → PRP$ NNS ↑ S) is used for prediction. If the unlexicalized version is also found to be rare, other variations of the production, depending on whether they include the lexical heads and how many non-terminals (one or two) they consider before and after the potential boundary, are examined one after another (see Fisher and Roark [2007] for details). In contrast, we compute the maximum likelihood estimates for a primary production (feature) and its other variations, and use those directly as features with/without binarizing the values.
Shallow syntactic features like Chunk and POS tags have been shown to possess valuable clues for discourse segmentation (Fisher and Roark 2007; Sporleder and Lapata 2005). For example, it is less likely that an EDU boundary occurs within a chunk. We annotate the tokens of a sentence with chunk and POS tags using the state-of-the-art Illinois tagger11 and encode these as features in our model. Note that the chunker assigns each token a tag using the BIO notation, where B stands for beginning of a particular phrase (e.g., noun or verb phrase), I stands for inside of a particular phrase, and O stands for outside of a particular phrase. The rationale for using the Illinois chunker is that it uses a larger set of tags (23 in total); thus it is more informative than most of the other existing taggers, which typically use only five tags (B–NP, I–NP, B–VP, I–VP, and O).
EDUs are normally multi-word strings. Thus, a token near the beginning or end of a sentence is unlikely to be the end of a segment. Therefore, for each token we include its relative position (i.e., absolute position/total number of tokens) in the sentence and distances to the beginning and end of the sentence as features.
It is unlikely that two consecutive tokens are tagged with EDU boundaries. Therefore, we incorporate contextual information for a token into our model by including the above features computed for its neighboring tokens.
We also experimented with different N-gram (N ∈ {1, 2, 3}) features extracted from the token sequence, POS sequence, and chunk sequence. However, because such features did not improve segmentation accuracy on the development set, they were excluded from our final set of features.","6 experiments :In this section we present our experimental results. First, we describe the corpora on which the experiments were performed and the evaluation metrics used to measure the
11 Available at http://cogcomp.cs.illinois.edu/page/software.
performance of the discourse segmenter and the parser. Then we show the performance of our discourse segmenter, followed by the performance of our discourse parser. Whereas previous studies on discourse analysis only report their results on a particular corpus, to demonstrate the generality of our method, we experiment with texts from two very different genres: news articles and instructional how-to manuals.
Our first corpus is the standard RST–DT (Carlson, Marcu, and Okurowski 2002), which contains discourse annotations for 385 Wall Street Journal news articles taken from the Penn Treebank corpus (Marcus, Marcinkiewicz, and Santorini 1994). The corpus is partitioned into a training set of 347 documents and a test set of 38 documents. A total of 53 documents selected from both training and test sets were annotated by two human annotators. We measure human agreements based on this doubly annotated data set. We used 25 documents from the training set as our development set. In RST–DT, the original 25 rhetorical relations defined by Mann and Thompson (1988) are further divided into a set of 18 coarser relation classes with 78 finer-grained relations (see Carlson and Marcu [2001] for details). Our second corpus is the instructional corpus prepared by Subba and Di-Eugenio (2009), which contains discourse annotations for 176 how-to manuals on home repair. The corpus was annotated with 26 informational relations (e.g., Preparation–Act, Act–Goal).
For our experiments with the intra-sentential discourse parser, we extracted a sentence-level DT from a document-level DT by finding the sub-tree that exactly spans over the sentence. In RST–DT, by our count, 7, 321 out of 7, 673 sentences in the training set, 951 out of 991 sentences in the test set, and 1, 114 out of 1, 208 sentences in the doubly-annotated set have a well-formed DT. On the other hand, 3, 032 out of 3, 430 sentences in the instructional corpus have a well-formed DT. This forms the corpora for our experiments with intra-sentential discourse parsing. However, the existence of a well-formed DT is not a necessity for discourse segmentation; therefore, we do not exclude any sentence in our discourse segmentation experiments. In this subsection we describe the metrics used to measure both how much the annotators agree with each other, and how well the systems perform when their outputs are compared with human annotations for the discourse analysis tasks.
6.2.1 Metrics for Discourse Segmentation. Because sentence boundaries are considered to also be the EDU boundaries, we measure segmentation accuracy with respect to the intra-sentential segment boundaries, which is a standard method (Soricut and Marcu 2003; Fisher and Roark 2007). Specifically, if a sentence contains n EDUs, which corresponds to n − 1 intra-sentential segment boundaries, we measure the segmenter’s ability to correctly identify these n − 1 boundaries. Let h be the total number of intrasentential segment boundaries in the human annotation, m be the total number of intrasentential segment boundaries in the model output, and c be the total number of correct segment boundaries in the model output. Then, we measure Precision (P), Recall (R), and F-score for segmentation performance as follows:
P = cm, R = c h , and F−score = 2PR P + R = 2c h + m (14)
6.2.2 Metrics for Discourse Parsing. To evaluate parsing performance, we use the standard unlabeled and labeled precision, recall, and F-score as proposed by Marcu (2000b). The unlabeled metric measures how accurate the discourse parser is in finding the right structure (i.e., the skeleton) of the DT, while the labeled metrics measure the parser’s ability to find the right labels (i.e., nuclearity statuses or relation labels) in addition to the right structure. Assume, for example, that given the two sentences of Figure 1, our system generates the DT shown in Figure 15. In Figure 16, we show the same gold DT shown in Figure 1 (on the left), and the same system-generated DT shown in Figure 15 (on the right), when the two trees are aligned. For the sake of illustration, instead of showing the real EDUs, we only show their IDs. Notice that the automatic segmenter made two mistakes: (1) it broke the EDU marked 2–3 (Some people use the purchasers’ index as a leading indicator) in the human annotation into two separate EDUs, and (2) it could not identify EDU 5 (But the thing it’s supposed to measure) and EDU 6 (— manufacturing strength —) as two separate EDUs. Therefore, when we align the two annotations, we obtain seven EDUs in total.
In Table 2, we list all constituents of the two DTs and their associated labels at the span, nuclei, and relation levels. The recall (R) and precision (P) figures are shown at the bottom of the table. Notice that, following (Marcu 2000b), the relation labels are assigned to the children nodes rather than to the parent nodes in the evaluation process to deal with non-binary trees in human annotations. To our knowledge, no implementation of
the evaluation metrics was made publicly available. Therefore, to help other researchers, we have made our source code of the evaluation metrics publicly available.12
Given this evaluation setup, it is easy to understand that if the number of EDUs is the same in the human and system annotations (e.g., when the discourse parser uses gold discourse segmentation), and the discourse trees are binary, then we get the same figures for precision, recall, and F-score. In this section we present our experiments on discourse segmentation.
6.3.1 Experimental Set-up for Discourse Segmentation. We compare the performance of our discourse segmenter with the performance of the two publicly available discourse segmenters, namely, the discourse segmenters of the HILDA (Hernault et al. 2010) and SPADE (Soricut and Marcu 2003) systems. We also compare our results with the stateof-the-art results reported by Fisher and Roark (2007) on the RST–DT test set. In all our experiments when comparing two systems, we use paired t-test on the F-scores to measure statistical significance and report the p-value.
We ran HILDA with its default settings. For SPADE, we applied the same modifications to its default settings as described in Fisher and Roark (2007), which delivers significantly improved performance over its original version. Specifically, in our experiments on the RST–DT corpus, we trained SPADE using the human-annotated syntactic trees extracted from the Penn Treebank (Marcus, Marcinkiewicz, and Santorini 1994), and, during testing, we replaced the Charniak parser (Charniak 2000) with a more
12 Available from alt.qcri.org/tools/.
accurate reranking parser (Charniak and Johnson 2005). However, because of the lack of gold syntactic trees in the instructional corpus, we trained SPADE in this corpus using the syntactic trees produced by the reranking parser. To avoid using the gold syntactic trees, we used the reranking parser in our system for both training and testing purposes. This syntactic parser was trained on the sections of the Penn Treebank not included in our test set. We applied the same canonical lexical head projection rules (Magerman 1995; Collins 2003) to lexicalize the syntactic trees as done in HILDA and SPADE.
Note that previous studies (Fisher and Roark 2007; Soricut and Marcu 2003; Hernault et al. 2010) on discourse segmentation only report their performance on the RST–DT test set. To compare our results with them, we evaluated our model on the RST–DT test set. In addition, we showed a more general performance of SPADE and our system on the two corpora based on 10-fold cross validation.13 However, SPADE does not come with a training module for its segmenter. We reimplemented this module and verified its correctness by reproducing the results on the RST–DT test set.
6.3.2 Results for Discourse Segmentation. Table 3 shows the discourse segmentation results of different systems in Precision, Recall, and F-score on the two corpora. On the RST– DT corpus, HILDA’s segmenter delivers the weakest performance, having an F-score of only 74.1. Note that the high segmentation accuracy reported by Hernault et al. (2010) is due to a less stringent evaluation metric. SPADE performs much better than HILDA with an absolute F-score improvement of 11.1%. Our segmenter DS outperforms SPADE with an absolute F-score improvement of 4.9% (p-value < 2.4e-06), and also achieves comparable results to the ones of Fisher and Roark (2007) (F&R), even though we use fewer features.14 Notice that human agreement for this task is quite high—namely, an F-score of 98.3 computed on the doubly-annotated portion of the RST–DT corpus mentioned in Section 6.1.
Because Fisher and Roark (2007) only report their results on the RST–DT test set and we did not have access to their system, we compare our approach only with SPADE when evaluating on a whole corpus based on 10-fold cross validation. On the RST–DT corpus, our segmenter delivers an absolute F-score improvement of 3.8 percentage points, which represents a more than 25% relative error rate reduction.
13 Because the two tasks—discourse segmentation and intra-sentential parsing—operate at the sentence level, the cross validation was performed over sentences for their evaluation. 14 Because we did not have access to the system or to the complete output/results of Fisher and Roark (2007), we were not able to perform a statistical significance test.
The improvement is higher on the instructional corpus with an absolute F-score improvement of 8.1 percentage points, which corresponds to a relative error reduction of 30%. The improvements for both corpora are statistically significant (p-value < 3.0e-06). When we compare our results on the two corpora, we observe a substantial decrease in performance on the instructional corpus. This could be because of a smaller amount of data in this corpus and/or to the inaccuracies in the syntactic parser and taggers, which are trained on news articles. A promising future direction would be to apply effective domain adaptation methods (e.g., easyadapt [Daumé 2007]) to improve discourse segmentation performance in the instructional domain by leveraging the rich data in the news domain (i.e., RST–DT). In this section we present our experiments on discourse parsing. First, we describe the experimental set-up. Then, we present the results of the parsers. While presenting the performance of our discourse parser, we show a breakdown of intra-sentential versus inter-sentential results, in addition to the aggregated results at the document level.
6.4.1 Experimental Set-up for Discourse Parsing. In our experiments on sentence-level (i.e., intra-sentential) discourse parsing, we compare our approach with SPADE (Soricut and Marcu 2003) on the RST–DT corpus, and with the ILP-based approach of Subba and Di-Eugenio (2009) on the instructional corpus, because they are the state of the art in their respective genres. For SPADE, we applied the same modifications to its default settings as described in Section 6.3.1, which leads to improved performance. Similarly, in our experiments on document-level (i.e., multi-sentential) parsing, we compare our approach with HILDA (Hernault et al. 2010) on the RST–DT corpus, and with the ILPbased approach (Subba and Di-Eugenio 2009) on the instructional corpus. The results for HILDA were obtained by running the system with default settings on the same inputs we provided to our system. Because we could not run the ILP-based system (not publicly available), we report the performance presented in their paper.
Our experiments on the RST–DT corpus use the same 18 coarser coherence relations (see Figure 18 later in this article), defined by Carlson and Marcu (2001) and also used in SPADE and HILDA systems. More specifically, the relation set consists of 16 relation categories and two pseudo-relations, namely, Textual–Organization and Same–Unit. After attaching the nuclearity statuses (NS, SN, NN) to these relations, we obtain 41 distinct relations.15 Our experiments on the instructional corpus consider the same 26 primary relations (e.g., Goal:Act, Cause:Effect) used by Subba and Di-Eugenio (2009) and also treat the reversals of non-commutative relations as separate relations. That is, Goal–Act and Act–Goal are considered to be two different coherence relations. Attaching the nuclearity statuses to these relations provides 76 distinct relations.
Based on our experiments on the development set, the size of the automatically built bi-gram and tri-gram dictionaries was set to 95% of their total number of items, and the size of the unigram dictionary was set to 100%. Note that the unigram dictionary contains only special tags denoting EDU, sentence, and paragraph boundaries.
15 Not all relations take all the possible nuclearity statuses. For example, Elaboration and Attribution are mono-nuclear relations, and Same–Unit and Joint are multi-nuclear relations.
6.4.2 Evaluation of the Intra-Sentential Discourse Parser. This section presents our experimental evaluation on intra-sentential discourse parsing. First, we show the performance of the sentence-level parsers when they are provided with manual (or gold) discourse segmentations. This allows us to judge the parsing performance independently of the segmentation task. Then, we show the end-to-end performance of our intra-sentential framework, that is, the intra-sentential parsing performance based on automatic discourse segmentation.
Intra-sentential parsing results based on manual segmentation Table 4 presents the intra-sentential discourse parsing results when manual discourse segmentation is used. Recall from our discussion on evaluation metrics in Section 6.2.2 that precision, recall, and F-score are the same when manual segmentation is used. Therefore, we report only one of them. Notice that our sentence-level discourse parser PAR-S consistently outperforms SPADE on the RST–DT test set in all three metrics, and the improvements are statistically significant (p-value< 0.01). Especially, on the relation labeling task, which is the hardest among the three tasks, we achieve an absolute F-score improvement of 12.2 percentage points, which represents a relative error rate reduction of 37.7%.
To verify our claim that capturing the sequential dependencies between DT constituents using a DCRF model actually contributes to the performance gain, we also compare our results with an intra-sentential parser (see CRF-NC in Table 4) that uses a simplified CRF model similar to the one shown in Figure 8. Although the simplified model has two hidden variables to model the structure and the relation of a DT constituent jointly, it does not have a chain structure, thus it ignores the sequential dependencies between DT constituents. The comparison in all three measures demonstrates that the improvements are indeed partly due to the DCRF model (p–value < 0.01).16 A comparison between CRF-NC and SPADE shows that CRF-NC significantly outperforms SPADE in all three measures (p-value < 0.01). This could be due to the fact that CRF-NC is trained discriminatively with a large number of features, whereas SPADE is trained generatively with only lexico-syntactic features.
Notice that the scores of our parser (PAR-S) are close to the human agreement on the doubly-annotated data, and these results on the RST–DT test set are also consistent with the mean scores over 10-folds on the whole RST–DT corpus.17
The improvements are higher on the instructional corpus, where we compare our mean results over 10-folds with the reported results of the ILP-based system of Subba and Di-Eugenio (2009), giving absolute F-score improvements of 5.4 percentage points, 17.6 percentage points, and 12.8 percentage points in span, nuclearity, and relations, respectively.18 Our parser PAR-S reduces the errors by 76.1%, 62.4%, and 34.6% in span, nuclearity and relations, respectively.
If we compare the performance of our intra-sentential discourse parser on the two corpora, we notice that our parser PAR-S is more accurate in finding the right tree
16 The parsing performance reported in Table 4 for CRF-NC is when the CRF parsing model is trained on a balanced data set (an equal number of instances with S=1 and S=0); Training on full but imbalanced data set gives slightly lower results. 17 Our EMNLP and ACL publications (Joty, Carenini, and Ng 2012; Joty et al. 2013) reported slightly lower parsing accuracies. Fixing a bug in the parsing algorithm accounts for the difference. 18 Subba and Di-Eugenio (2009) report their results based on an arbitrary split between training and test sets. Because we did not have access to their particular split, we compare our model’s performance based on 10-fold cross validation with their reported results. Also, because we did not have access to their system/output, we could not perform a significance test on the instructional corpus.
structure (see Span row in the table) on the instructional corpus. This may be due to the fact that sentences in the instructional domain are relatively short and contain fewer EDUs than sentences in the news domain, thus making it easier to find the right tree structure. However, when we compare the performance on the relation labeling task, we observe a decrease on the instructional corpus. This may be due to the small amount of data available for training and the imbalanced distribution of a large number of discourse relations (i.e., 76 with nuclearity attached) in this corpus.
Intra-sentential parsing results based on automatic segmentation In order to evaluate the performance of the fully automatic sentence-level discourse analysis systems, we feed the intra-sentential discourse parsers the output of their respective discourse segmenters. Table 5 shows the (P)recision, (R)ecall, and (F)–score results for different evaluation metrics. We compare our intra-sentential parser PAR-S with SPADE on the RST–DT test set. We achieve absolute F-score improvements of 5.7 percentage points, 6.4 percentage points, and 9.5 percentage points in span, nuclearity, and relation, respectively. These improvements are statistically significant (pvalue<0.001). Our system, therefore, reduces the errors by 24.5%, 21.4%, and 22.6% in span, nuclearity, and relations, respectively. These results are also consistent with the mean results over 10-folds on the whole RST–DT corpus.
The rightmost column in the table shows our mean results over 10-folds on the instructional corpus. We could not compare our system with the ILP-based approach of Subba and Di-Eugenio (2009) because no results were reported using an automatic segmenter. It is interesting to observe how much our parser is affected by an automatic segmenter on the two corpora (see Tables 4 and 5). Nevertheless, taking into account the
segmentation results in Table 3, this is not surprising because previous studies (Soricut and Marcu 2003) have already shown that automatic segmentation is the primary impediment to high accuracy discourse parsing. This demonstrates the need for a more accurate discourse segmentation model in the instructional genre.
6.4.3 Evaluation of the Complete Parser. We experiment with our full document-level discourse parser on the two corpora using the two parsing approaches described in Section 4.3, namely, 1S-1S and the sliding window. On RST–DT, the standard split was used for training and testing. On the instructional corpus, Subba and Di-Eugenio (2009) used 151 documents for training and 25 documents for testing. Because we did not have access to their particular split, we took five random samples of 151 documents for training and 25 documents for testing, and report the average performance over the five test sets.
Table 6 presents results for our two-stage discourse parser (TSP) using approaches 1S-1S (TSP 1-1) and the sliding window (TSP SW) on manually segmented texts. Recall that precision, recall, and F-score are the same when manual segmentation is used. We compare our parser with the state-of-the-art on the two corpora: HILDA (Hernault et al. 2010) on RST–DT, and the ILP-based approach (Subba and Di-Eugenio 2009) on the instructional domain. On both corpora, our systems outperform existing systems by a wide margin (p-value <7.1e-05 on RST–DT).19 On RST–DT, our parser TSP 1-1 achieves absolute improvements of 7.9 percentage points, 9.3 percentage points, and 11.5 percentage points in span, nuclearity, and relation, respectively, over HILDA. This represents relative error reductions of 31.2%, 22.7%, and 20.7% in span, nuclearity, and relation, respectively.
Beside HILDA, we also compare our results with two baseline parsers on RST–DT: (1) CRF-O, which uses a single unified CRF-based parsing model shown in Figure 8 (the one used for multi-sentential parsing) without distinguishing between intra- and multisentential parsing, and (2) CRF-T, which uses two different CRF-based parsing models for intra- and multi-sentential parsing in the two-stage approach 1S-1S, both models having the same structure as in Figure 8. Thus, CRF-T is a variation of TSP 1-1, where the DCRF-based (chain-structured) intra-sentential parsing model is replaced with a simpler CRF-based parsing model.20 Note that although CRF-O does not explicitly
19 Because we did not have access to the output or to the system of Subba and Di-Eugenio (2009), we were not able to perform a significance test on the instructional corpus. 20 The performance of this model for intra-sentential parsing is reported in Table 4 under the name CRF-NC.
discriminate between intra- and multi-sentential parsing, it uses N-gram features that include sentence and EDU boundaries to encode this information into the model.
Table 6 shows that both CRF-O and CRF-T outperform HILDA by a good margin (p-value <0.0001). This improvement can be attributed to the optimal parsing algorithm and better feature selection strategy. When we compare CRF-T with CRF-O, we notice significant performance gains for CRF-T (p-value <0.001). The absolute gains are 4.32 percentage points, 2.68 percentage points, and 4.55 percentage points in span, nuclearity, and relation, respectively. This comparison clearly demonstrates the benefit of using a two-stage approach with two different parsing models over a framework with one single unified parsing model. Finally, when we compare our best results with the human agreements, we still observe room for further improvement in all three measures.
On the instructional genre, our parser TSP 1-1 delivers absolute F-score improvements of 10.3 percentage points, 13.6 percentage points, and 8.1 percentage points in span, nuclearity, and relations, respectively, over the ILP-based approach of Subba and Di-Eugenio (2009). Our parser, therefore, reduces errors by 34.7%, 26.9%, and 12.5% in span, nuclearity, and relations, respectively.
If we compare the performance of our discourse parsers on the two corpora, we observe lower results on the instructional corpus. There could be two reasons for this. First, the instructional corpus has a smaller amount of data with a larger set of relations (76 with nuclearity attached). Second, some of the frequent relations are semantically very similar (e.g., Preparation-Act, Step1-Step2), which makes it difficult even for the human annotators to distinguish them (Subba and Di-Eugenio 2009).
Comparison between our two document-level parsing approaches reveals that TSP SW significantly outperforms TSP 1-1 only in finding the right structure on both corpora (p-value <0.01). Not surprisingly, the improvement is higher on the instructional corpus. A likely explanation is that the instructional corpus contains more leaky boundaries (12%), allowing the sliding window approach to be more effective in finding those, without inducing much noise for the labels. This demonstrates the potential of TSP SW for data sets with even more leaky boundaries, e.g., the Dutch (Vliet and Redeker 2011) and the German Potsdam (Stede 2004) corpora. However, it would be interesting to see how other heuristics to do consolidation in the cross condition (Section 4.3.2) perform.
To analyze errors made by TSP SW, we looked at some poorly parsed examples and found that although TSP SW finds more correct structures, a corresponding improvement in labeling relations is not present because in some cases, it tends to induce noise from the neighboring sentences for the labels. For example, when parsing is performed on the first sentence in Figure 1 in isolation using 1S-1S, our parser rightly identifies the Contrast relation between EDUs 2 and 3. But, when it is considered with its neighboring sentences by the sliding window, the parser labels it as Elaboration. A promising strategy to deal with this and similar problems would be to apply both approaches to each sentence and combine them by consolidating three probabilistic decisions, namely, the one from 1S-1S and the two from the sliding window.
6.4.4 k-best Parsing Results Based on Manual Segmentation. As described in Section 4.2, a straight-forward modification of our probabilistic parsing algorithm allows us to generate a list of k-best parse hypotheses for a given text. We adapt our parsing algorithm accordingly to produce k most probable DTs for each text, and measure the oracle accuracy based on the F-scores of the Relation metric which gives aggregated evaluation
on structure and relation labels (see Table 2). Specifically, the oracle accuracy O−score for k-best discourse parsing is measured as follows:
O−score = ∑N i=1 max k j=1 F−scorer(gi, h j i)
N (15)
where N is the total number of texts (sentences or documents) evaluated, gi is the gold DT annotation for text i, hji is the j
th parse hypothesis generated by the parser for text i, and F-scorer(gi, h j i) is the F-score accuracy of hypothesis h j i on the Relation metric, which essentially measures how similar hji is to gi in terms of its structure and labels. Table 7 presents the oracle scores of our intra-sentential parser PAR-S on the RST– DT test set as a function of k of k-best parsing. The 1-best result tells that the parser has the base accuracy of 79.8%. The 2-best shows dramatic oracle-rate improvements (i.e., 4.65% absolute), meaning that often our parser generates the best tree as its top two outputs. 3-best and 4-best also show moderate improvements (about 2%). Things start to slow down afterwards, and we achieve oracle rates of 90.37% and 92.57% at 10−best and 20-best, respectively. The 30-best parsing gives an oracle score of 93.2%.
The results of our k-best intra-sentential discourse parser demonstrate that a k-best reranking approach like that of Collins and Koo (2005) and Charniak and Johnson (2005) used for syntactic parsing can potentially improve the parsing accuracy even further by exploiting additional global features of the candidate discourse trees as evidence.
The scenario is quite different at the document-level; Table 8 shows the k-best parsing results of TSP 1S-1S on the RST–DT test set. The improvements in oracle-rate are small at the document-level when compared with the sentence-level parsing. For example, the 2-best and the 5-best improve over the base accuracy by only 0.7 percentage points and 1.0 percentage points, respectively. The improvements get even slower after that. However, this is not surprising because generally document-level DTs are big with many constituents, and only a very few of these constituents change from k-best to k + 1-best parsing. These small changes among the candidate DTs do not contribute much to the overall F-score accuracy (for further clarification see how F-score is calculated in Section 6.2.2).
The results of our k-best document-level parsing suggest that often the best tree is missing in the top k parses. Thus, a reranking of k-best document-level parses may not be a suitable option for further improvement at the document-level. An alternative to k-best reranking is to use a sampling-based parsing strategy (Wick et al. 2011) to explore the space of possible trees, as recently used for dependency parsing (Zhang et al. 2014). However, note that the potential gain we may obtain by using a reranker at the sentence level will also improve the (combined) accuracy of the document-level parser.
6.4.5 Analysis of Features. To analyze the relative importance of different features used in our parsing models, Table 9 presents the sentence- and document-level parsing results on a manually segmented RST–DT test set using different subsets of features. The feature subsets were defined in Section 4.1.4. In each parsing condition, the subsets of features are added incrementally, based on their availability and historical importance. The columns in Table 9 represent the inclusion order of the feature subsets.
Because SPADE (Soricut and Marcu 2003) achieved the previous best results on intra-sentential parsing using Dominance set features, these are included as the initial set of features in our intra-sentential parsing model. In HILDA, Hernault et al. (2010) demonstrate the importance of Organizational and N-gram features for full text parsing. We add these two feature subsets one after another in our intra- and multi-sentential parsing models.21 Contextual features require other features to be computed; thus they were added after those features. Because computation of Sub-structural features requires an initial parse tree (i.e., when the parser is applied), they are added at the very end.
Notice that inclusion of every new subset of features appears to improve the performance over the previous set. Specifically, for sentence-level parsing, when we add the Organizational features with the Dominance set features, we achieve about 2 percentage points absolute improvements in nuclearity and relations. With N-gram features, the gain is even higher: 6 percentage points in relations and 3.5 percentage points in nuclearity for sentence-level parsing, and 3.8 percentage points in relations and 3.1 percentage points in nuclearity for document-level parsing. This demonstrates the utility of the N-gram features, which is also consistent with the previous findings of duVerle and Prendinger (2009) and Schilder (2002).
The features extracted from Lexical chains (L-ch) have also proved to be useful for document-level parsing. They deliver absolute improvements of 2.7 percentage points, 2.9 percentage points, and 2.3 percentage points in span, nuclearity, and relations,
21 Text structural features are included in the Organizational features for multi-sentential parsing.
respectively. Including the Contextual features further gives improvements of 3 percentage points in nuclearity and 2.2 percentage points in relation for sentence-level parsing, and 1.3 percentage points in nuclearity, and 1.2 percentage points in relation for document-level parsing. Notice that Sub-structural features are more beneficial for document-level parsing than they are for sentence-level parsing, that is, an improvement of 2.2 percentage points versus an improvement of 0.9 percentage points. This is not surprising because document-level DTs are generally much larger than sentencelevel DTs, making the sub-structural features more effective for document-level parsing.
6.4.6 Error Analysis. We further analyze the errors made by our discourse parser. As described in previous sections, the parser could be wrong in finding the right structure as well as the right nuclearity and relation labels. Figure 17 presents an example where our parser makes mistakes in finding the right structure (notice the units connected by Attribution and Cause in the two example DTs) and the right relation label (TopicComment vs. Background). The comparison between intra- and multi-sentential parsing results presented in Sections 6.4.2 and 6.4.3 tells us that the errors in structure occur more frequently when the DT is large (e.g., at the document level) and the parsing model fails to capture the long-range structural dependencies between the DT constituents.
To further analyze the errors made by our parser on the hardest task of relation labeling, in Figure 18 we present the confusion matrix for our document-level parser TSP 1-1 on the RST–DT test set. In order to judge independently the ability of our parser to assign the correct relation labels, the confusion matrix is computed based on the constituents (see Table 2), where our parser found the right span (i.e., structure).22 The relations in the matrix are ordered according to their frequency in the training set.
In general, the errors can be explained by two different causes acting together: (1) imbalanced distribution of the relations in the corpus, and (2) semantic similarity between the relations. The most frequent relation Elaboration tends to overshadow others, especially the ones that are semantically similar (e.g., Explanation, Background)
22 Therefore, the counts of the relations shown in the table may not match the ones in the test set.
T-C T-O T-CM M-M CMP EV SU CND EN CA TE EX BA CO JO S-U AT EL
and less frequent (e.g., Summary, Evaluation). Furthermore, our models sometimes fail to distinguish relations that are semantically similar (e.g., Temporal vs. Background, Cause vs. Explanation).
Now, let us look more closely at a few of these errors. Figure 19 presents an example where our parser mistakenly labels a Summary as Elaboration. Clearly, in this example the text in parentheses (i.e., (CFD)) is an acronym or summary of the text to the left. However, parenthesized texts are also used to provide additional information (i.e., to elaborate), as exemplified in Figure 20 by two text snippets from the RST-DT. Notice that although the structure of the text (widow of the ..) in the first example is quite distinguishable from the structure of (CFD), the text (D., Maine) in the second example is similar to (CFD) in structure, thus it confuses our model.23
Figure 21 presents two examples where our parser mistakenly labels Background and Cause as Elaboration. However, notice that the two discourse relations (i.e., Background
23 D., Maine in this example refers to Democrat from state Maine.
vs. Elaboration and Cause vs. Elaboration) in these examples are semantically very close, and arguably both can be applicable.
Given these observations, we see two possible ways to improve our system. First, we would like to use a more robust method (e.g., ensemble methods with bagging) to deal with the imbalanced distribution of relations, along with taking advantage of richer semantic knowledge (e.g., compositional semantics) to cope with the errors caused by semantic similarity between the relations. Second, to capture long-range dependencies between DT constituents, we would like to explore the idea of k-best discriminative reranking using tree kernels (Dinarelli, Moschitti, and Riccardi 2011). Because our parser already produces k most probable DTs, developing a reranker based on discourse tree kernels is very much within our reach.","7 conclusions and future directions :In this article we have presented CODRA, a complete probabilistic discriminative framework for performing rhetorical analysis in the RST framework. CODRA comprises components for performing both discourse segmentation and discourse parsing. The discourse segmenter is a binary classifier based on a maximum entropy model, and the discourse parser applies an optimal parsing algorithm to probabilities inferred from two CRF models: one for intra-sentential parsing and the other for multi-sentential parsing. The CRF models effectively represent the structure and the label of discourse tree constituents jointly. Furthermore, the DCRF model for intra-sentential parsing captures the sequential dependencies between the constituents. The two separate parsing models use their own informative feature sets and the distributional variations of the relation labels in their respective parsing conditions.
We have also presented two approaches to effectively combine the intra-sentential and the multi-sentential parsing modules, which can exploit the strong correlation observed between the text structure and the structure of the discourse tree. The first approach (1S–1S) builds a DT for every sentence using the intra-sentential parser, and then
runs the multi-sentential parser on the resulting sentence-level DTs. To deal with leaky boundaries, our second approach (sliding window) builds sentence-level discourse sub-trees by applying the intra-sentential parser on a sliding window, covering two adjacent sentences and then consolidating the results produced by overlapping windows. After that, the multi-sentential parser takes all these sentence-level sub-trees and builds a full rhetorical parse for the whole document.
Finally, we have extended the parsing algorithm to generate k most probable parse hypotheses for each input text, which could be used in a reranker to improve over the initial ranking using global features like long-range structural dependencies.
Empirical evaluations on two different genres demonstrate that our approach to discourse segmentation achieves state-of-the-art performance more efficiently using fewer features. A series of experiments on the discourse parsing task shows that both our intra- and multi-sentential parsers significantly outperform the state of the art, often by a wide margin. A comparison between our combination strategies reveals that the sliding window approach is more robust across domains. Furthermore, the oracle accuracy computed based on the k-best parse hypotheses generated by our parser demonstrates that a reranker could potentially improve the accuracy even further.
Our error analysis reveals that although the sliding window approach finds more correct tree structures, in some cases it induces noise for the relation labels from the neighboring sentences. With respect to the performance of our discourse parser on the relation labeling task we also found that the most frequent relations tend to mislead the identification of the less frequent ones, and the models sometimes fail to distinguish relations that are semantically similar.
The work presented in this article leads us to several interesting future directions. Our short-term goal is to develop a k-best discriminative reranking discourse parser using tree kernels applied to discourse trees. We also plan to investigate to what extent discourse segmentation and discourse parsing can be performed jointly.
We would also like to explore how our system performs on other genres like conversational (e.g., blogs, e-mails) and evaluative (e.g., customer reviews) texts. To address the problem of limited annotated data in various genres, we are planning to develop an interactive version of our system that will allow users to fix the output of the system with minimal effort and let the system learn from that feedback.
Another interesting future direction is to perform extrinsic evaluations of our system in downstream applications. One important application of rhetorical structure is text summarization, where a significant challenge is producing not only informative but also coherent summaries. A number of researchers have already investigated the utility of rhetorical structure for measuring text importance (i.e., informativeness) in summarization (Marcu 2000b; Daumé and Marcu 2002; Louis, Joshi, and Nenkova 2010). Recently, Christensen et al. (2013, 2014) propose to perform sentence selection and ordering at the same time, and use constraints on discourse structure to make the summaries coherent. However, they represent the discourse as an unweighted directed graph, which is shallow and not sufficiently informative in most cases. Furthermore, their approach does not allow compression at the sentence level, which is often beneficial in summarization. In the future, we would like to investigate the utility of our rhetorical structure for performing sentence compression, selection, and ordering in a joint summarization process.
Discourse structure can also play important roles in sentiment analysis. A key research problem in sentiment analysis is extracting fine-grained opinions about different aspects of a product. Several recent papers (Somasundaran 2010; Lazaridou, Titov, and
Sporleder 2013) exploited the rhetorical structure for this task. Another challenging problem is assessing the overall opinion expressed in a review because not all sentences in a review contribute equally to the overall sentiment. For example, some sentences are subjective, whereas others are objective (Pang and Lee 2004); some express the main claims, whereas others support them (Taboada et al. 2011); some express opinions about the main entity, whereas others are about the peripherals. Discourse structure could be useful to capture the relative weights of the discourse units towards the overall sentiment. For example, the nucleus and satellite distinction along with the rhetorical relations could be useful to infer the relative weights of the connecting discourse units.
Among other applications of discourse structure, Machine Translation (MT) and its evaluation have received a resurgence of interest recently. A workshop dedicated to discourse in machine translation was arranged recently at the ACL 2013 conference (Webber et al. 2013). Researchers believe that MT systems should consider discourse phenomena that go beyond the current sentence to ensure consistency in the choice of lexical items or referring expressions, and the fact that source-language coherence relations are also realized in the target language (i.e., translating at the documentlevel [Hardmeier, Nivre, and Tiedemann 2012]). Guzmán et al. (2014a, 2014b) and Joty et al. (2014) propose new discourse-aware automatic evaluation metrics for MT systems using our discourse analysis tool. They demonstrate that sentence-level discourse information is complementary to the state-of-the-art evaluation metrics, and by combining the discourse-based metrics with the metrics from the ASIYA MT evaluation toolkit (Giménez and Màrquez 2010), they won the WMT 2014 metrics shared task challenge (Macháček and Bojar 2014) both at the segment- and at the system-level. These results suggest that discourse structure helps to distinguish better translations from worse ones. Thus, it would be interesting to explore whether discourse information can be used to rerank alternative MT hypotheses as a post-processing step for the MT output.
A longer-term goal is to extend our framework to also work with graph structures of discourse, as recommended by several recent discourse theories (Wolf and Gibson 2005). Once we achieve similar performance on graph structures, we will perform extrinsic evaluations to determine their relative utility for various NLP tasks.
Finally, we hope that the online demo, the source code of CODRA, and the evaluation metrics that we made publicly available in this work will facilitate other researchers in extending our work and in applying discourse parsing to their NLP tasks.
Bibliographic Note
Portions of this work were previously published in two conference proceedings (Joty, Carenini, and Ng 2012; Joty et al. 2013). This article significantly extends our previous work in several ways, most notably: (1) we extend the parsing algorithm to generate k-most probable parse hypotheses for each input text (Section 4.2); (2) we show the oracle accuracies for k-best discourse parsing both at the sentence level and at the document level (Section 6.4.4); (3) to support our claim, we compare our best results with several variations of our approach (see CRF-NC in Section 6.4.2, and CRF-O and CRF-T in Section 6.4.3); (4) we analyze the relative importance of different features for intra- and multi-sentential discourse parsing (Section 6.4.5); and (5) we perform indepth error analysis of our complete rhetorical analysis framework (Section 6.4.6).
Appendix Sample Output Generated by an Online Demo of CODRA
A.",,"Clauses and sentences rarely stand on their own in an actual discourse; rather, the relationship between them carries important information that allows the discourse to express a meaning as a whole beyond the sum of its individual parts. Rhetorical analysis seeks to uncover this coherence structure. In this article, we present CODRA— a COmplete probabilistic Discriminative framework for performing Rhetorical Analysis in accordance with Rhetorical Structure Theory, which posits a tree representation of a discourse. CODRA comprises a discourse segmenter and a discourse parser. First, the discourse segmenter, which is based on a binary classifier, identifies the elementary discourse units in a given text. Then the discourse parser builds a discourse tree by applying an optimal parsing algorithm to probabilities inferred from two Conditional Random Fields: one for intra-sentential parsing and the other for multi-sentential parsing. We present two approaches to combine these two stages of parsing effectively. By conducting a series of empirical evaluations over two different data sets, we demonstrate that CODRA significantly outperforms the state-of-the-art, often by a wide margin. We also show that a reranking of the k-best parse hypotheses generated by CODRA can potentially improve the accuracy even further.","[{""affiliations"": [], ""name"": ""Shafiq Joty""}, {""affiliations"": [], ""name"": ""Giuseppe Carenini""}, {""affiliations"": [], ""name"": ""Raymond T. Ng""}]",SP:6e0be3b1258a50030cf084705cb3a08259f2f268,"[{""authors"": [""Althaus"", ""Ernst"", ""Denys Duchier"", ""Alexander Koller"", ""Kurt Mehlhorn"", ""Joachim Niehren"", ""Sven Thiel.""], ""title"": ""An efficient graph algorithm for dominance constraints"", ""venue"": ""Journal of Algorithms, 48(1):194\u2013219."", ""year"": 2003}, {""authors"": [""Asher"", ""Nicholas"", ""Alex Lascarides""], ""title"": ""Logics of Conversation"", ""year"": 2003}, {""authors"": [""Barzilay"", ""Regina"", ""Michael Elhadad.""], ""title"": ""Using lexical chains for text summarization"", ""venue"": ""Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and the 8th"", ""year"": 1997}, {""authors"": [""Biran"", ""Or"", ""Owen Rambow.""], ""title"": ""Identifying justifications in written dialogs by classifying text as argumentative"", ""venue"": ""International Journal of Semantic Computing, 5(4):363\u2013381."", ""year"": 2011}, {""authors"": [""Blair-Goldensohn"", ""Sasha"", ""Kathleen McKeown"", ""Owen Rambow.""], ""title"": ""Building and refining rhetorical-semantic relation models"", ""venue"": ""Proceedings of the Human Language Technologies: The Annual"", ""year"": 2007}, {""authors"": [""Blitzer"", ""John.""], ""title"": ""Domain Adaptation of Natural Language Processing Systems"", ""venue"": ""Ph.D. thesis, University of Pennsylvania."", ""year"": 2008}, {""authors"": [""Breiman"", ""Leo.""], ""title"": ""Bagging predictors"", ""venue"": ""Machine Learning, 24(2):123\u2013140."", ""year"": 1996}, {""authors"": [""Carlson"", ""Lynn"", ""Daniel Marcu.""], ""title"": ""Discourse tagging reference manual"", ""venue"": ""Technical Report ISI-TR-545, University of Southern California Information Sciences Institute."", ""year"": 2001}, {""authors"": [""Carlson"", ""Lynn"", ""Daniel Marcu"", ""Mary Ellen Okurowski.""], ""title"": ""RST Discourse Treebank (RST\u2013DT) LDC2002T07"", ""venue"": ""Linguistic Data Consortium, Philadelphia."", ""year"": 2002}, {""authors"": [""Chali"", ""Yllias"", ""Shafiq Joty""], ""title"": ""Word sense disambiguation using lexical"", ""year"": 2007}, {""authors"": [""Charniak"", ""Eugene.""], ""title"": ""A maximum-entropy-inspired parser"", ""venue"": ""Proceedings of the 1st North American Chapter of the Association for Computational Linguistics Conference, NAACL\u201900,"", ""year"": 2000}, {""authors"": [""Charniak"", ""Eugene"", ""Mark Johnson.""], ""title"": ""Coarse-to-fine n-best parsing and MaxEnt discriminative reranking"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL\u201905,"", ""year"": 2005}, {""authors"": [""Christensen"", ""Janara"", ""Mausam"", ""Stephen Soderland"", ""Oren Etzioni.""], ""title"": ""Towards coherent multi-document summarization"", ""venue"": ""Proceedings of the 2013 Conference of the North American Chapter of"", ""year"": 2013}, {""authors"": [""Christensen"", ""Janara"", ""Stephen Soderland"", ""Gagan Bansal"", ""Mausam.""], ""title"": ""Hierarchical summarization: Scaling up multi-document summarization"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the"", ""year"": 2014}, {""authors"": [""Collins"", ""Michael.""], ""title"": ""Head-driven statistical models for natural language Parsing"", ""venue"": ""Computational Linguistics, 29(4):589\u2013637."", ""year"": 2003}, {""authors"": [""Collins"", ""Michael"", ""Terry Koo.""], ""title"": ""Discriminative reranking for natural language parsing"", ""venue"": ""Computational Linguistics, 31(1):25\u201370."", ""year"": 2005}, {""authors"": [""Collobert"", ""Ronan"", ""Jason Weston"", ""L\u00e9on Bottou"", ""Michael Karlen"", ""Koray Kavukcuoglu"", ""Pavel Kuksa.""], ""title"": ""Natural language processing (almost) from scratch"", ""venue"": ""Journal of Machine Learning Research, 12:2493\u20132537."", ""year"": 2011}, {""authors"": [""Cristea"", ""Dan"", ""Nancy Ide"", ""Laurent Romary.""], ""title"": ""Veins theory: A model of global discourse cohesion and coherence"", ""venue"": ""Proceedings of the 36th Annual Meeting of the Association for Computational"", ""year"": 1998}, {""authors"": [""Danlos"", ""Laurence.""], ""title"": ""D-STAG: A discourse analysis formalism based on synchronous TAGs"", ""venue"": ""TAL, 50(1):111\u2013143."", ""year"": 2009}, {""authors"": [""III Daum\u00e9"", ""Hal.""], ""title"": ""Frustratingly easy domain adaptation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, ACL\u201907, pages 256\u2013263, Prague."", ""year"": 2007}, {""authors"": [""III Daum\u00e9"", ""Hal"", ""Daniel Marcu.""], ""title"": ""A noisy-channel model for document compression"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, ACL \u201902,"", ""year"": 2002}, {""authors"": [""Dinarelli"", ""Marco"", ""Alessandro Moschitti"", ""Giuseppe Riccardi.""], ""title"": ""Discriminative reranking for spoken language understanding"", ""venue"": ""IEEE Transactions on Audio, Speech and Language Processing (TASLP),"", ""year"": 2011}, {""authors"": [""duVerle"", ""David"", ""Helmut Prendinger""], ""title"": ""A novel discourse parser based on support vector machine classification"", ""venue"": ""In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the"", ""year"": 2009}, {""authors"": [""Egg"", ""Markus"", ""Alexander Koller"", ""Joachim Niehren.""], ""title"": ""The constraint language for lambda structures"", ""venue"": ""Journal of Logic, Language and Information, 10(4):457\u2013485."", ""year"": 2001}, {""authors"": [""Eisner"", ""Jason.""], ""title"": ""Three new probabilistic models for dependency parsing: An exploration"", ""venue"": ""Proceedings of the 16th Conference on Computational Linguistics Volume 1, COLING \u201996, pages 340\u2013345,"", ""year"": 1996}, {""authors"": [""Fellbaum"", ""Christiane.""], ""title"": ""WordNet\u2014An Electronic Lexical Database"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 1998}, {""authors"": [""Feng"", ""Vanessa"", ""Graeme Hirst.""], ""title"": ""Text-level discourse parsing with rich linguistic features"", ""venue"": ""Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, ACL \u201912,"", ""year"": 2012}, {""authors"": [""Feng"", ""Vanessa"", ""Graeme Hirst.""], ""title"": ""A linear-time bottom-up discourse parser with constraints and post-editing"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2014}, {""authors"": [""Finkel"", ""Jenny Rose"", ""Alex Kleeman"", ""Christopher Manning.""], ""title"": ""Efficient, feature-based, conditional random field parsing"", ""venue"": ""Proceedings of the 46th Annual Meeting of the Association for Computational"", ""year"": 2008}, {""authors"": [""Fisher"", ""Seeger"", ""Brian Roark.""], ""title"": ""The utility of parse-derived features for automatic discourse segmentation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2007}, {""authors"": [""Galley"", ""Michel"", ""Kathleen McKeown.""], ""title"": ""Improving word sense disambiguation in lexical chaining"", ""venue"": ""Proceedings of the 18th International Joint Conference on Artificial Intelligence, IJCAI\u201903, pages 1486\u20131488,"", ""year"": 2003}, {""authors"": [""Galley"", ""Michel"", ""Kathleen McKeown"", ""Eric Fosler-Lussier"", ""Hongyan Jing.""], ""title"": ""Discourse segmentation of multi-party conversation"", ""venue"": ""Proceedings of the 41st Annual Meeting of the Association for"", ""year"": 2003}, {""authors"": [""Ghosh"", ""Sucheta"", ""Richard Johansson"", ""Giuseppe Riccardi"", ""Sara Tonelli.""], ""title"": ""Shallow discourse parsing with conditional random fields"", ""venue"": ""Proceedings of the 5th International Joint Conference on"", ""year"": 2011}, {""authors"": [""Gim\u00e9nez"", ""Jes\u00fas"", ""Llu\u0131\u0301s M\u00e0rquez""], ""title"": ""Linguistic measures for automatic machine translation evaluation"", ""venue"": ""Machine Translation,"", ""year"": 2010}, {""authors"": [""Guzm\u00e1n"", ""Francisco"", ""Shafiq Joty"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Alessandro Moschitti"", ""Preslav Nakov"", ""Massimo Nicosia""], ""title"": ""Learning to differentiate better from worse translations"", ""year"": 2014}, {""authors"": [""Guzm\u00e1n"", ""Francisco"", ""Shafiq Joty"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Preslav Nakov""], ""title"": ""Using discourse structure improves machine translation evaluation"", ""venue"": ""In Proceedings of the 52nd Annual Meeting of the Association"", ""year"": 2014}, {""authors"": [""Halliday"", ""Michael"", ""Ruqaiya Hasan.""], ""title"": ""Cohesion in English"", ""venue"": ""Longman, London."", ""year"": 1976}, {""authors"": [""Hardmeier"", ""Christian"", ""Joakim Nivre"", ""J\u00f6rg Tiedemann.""], ""title"": ""Document-wide decoding for phrase-based statistical machine translation"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods"", ""year"": 2012}, {""authors"": [""Hernault"", ""Hugo"", ""Helmut Prendinger"", ""David duVerle"", ""Mitsuru Ishizuka.""], ""title"": ""HILDA: A discourse parser using support vector machine classification"", ""venue"": ""Dialogue and Discourse, 1(3):1\u201333."", ""year"": 2010}, {""authors"": [""Hirst"", ""Graeme"", ""David St-Onge.""], ""title"": ""Lexical Chains as Representation of Context for the Detection and Correction of Malapropisms"", ""venue"": ""Christiane Fellbaum, 431"", ""year"": 1997}, {""authors"": [""Hobbs"", ""Jerry.""], ""title"": ""Coherence and coreference"", ""venue"": ""Cognitive Science, 3:67\u201390."", ""year"": 1979}, {""authors"": [""Huang"", ""Liang"", ""David Chiang.""], ""title"": ""Better k-best parsing"", ""venue"": ""Proceedings of the Ninth International Workshop on Parsing Technology, Parsing \u201905, pages 53\u201364, Stroudsburg, PA."", ""year"": 2005}, {""authors"": [""Ji"", ""Yangfeng"", ""Jacob Eisenstein.""], ""title"": ""Representation learning for text-level discourse parsing"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long"", ""year"": 2014}, {""authors"": [""Joty"", ""Shafiq"", ""Giuseppe Carenini"", ""Raymond T. Ng.""], ""title"": ""A novel discriminative framework for sentence-level discourse analysis"", ""venue"": ""Proceedings of the 2012 Joint Conference on"", ""year"": 2012}, {""authors"": [""Joty"", ""Shafiq"", ""Giuseppe Carenini"", ""Raymond T. Ng.""], ""title"": ""Topic segmentation and labeling in asynchronous Conversations"", ""venue"": ""Journal of Artificial Intelligence Research (JAIR), 47:521\u2013573."", ""year"": 2013}, {""authors"": [""Joty"", ""Shafiq"", ""Giuseppe Carenini"", ""Raymond T. Ng"", ""Yashar Mehdad.""], ""title"": ""Combining intra- and multi-sentential rhetorical parsing for document-level discourse analysis"", ""venue"": ""Proceedings of the 51st Annual"", ""year"": 2013}, {""authors"": [""Joty"", ""Shafiq"", ""Francisco Guzm\u00e1n"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Preslav Nakov""], ""title"": ""DiscoTK: Using discourse structure for machine translation evaluation"", ""venue"": ""In Proceedings of the Ninth Workshop"", ""year"": 2014}, {""authors"": [""Jurafsky"", ""Daniel"", ""James Martin.""], ""title"": ""Statistical parsing"", ""venue"": ""Speech and Language Processing, chapter 14. Prentice Hall."", ""year"": 2008}, {""authors"": [""Knight"", ""Kevin"", ""Jonathan Graehl.""], ""title"": ""An overview of probabilistic tree transducers for natural language processing"", ""venue"": ""Computational Linguistics and Intelligent Text Processing, volume 3406 of Lecture Notes in"", ""year"": 2005}, {""authors"": [""Knott"", ""Alistair"", ""Robert Dale.""], ""title"": ""Using linguistic phenomena to motivate a set of coherence relations"", ""venue"": ""Discourse Processes, 18:35\u201362."", ""year"": 1994}, {""authors"": [""Koller"", ""Alexander"", ""Michaela Regneri"", ""Stefan Thater.""], ""title"": ""Regular tree grammars as a formalism for scope underspecification"", ""venue"": ""Proceedings of the 46th Annual Meeting of the Association for"", ""year"": 2008}, {""authors"": [""Lafferty"", ""John"", ""Andrew McCallum"", ""Fernando Pereira.""], ""title"": ""Conditional random fields: Probabilistic models for segmenting and labeling sequence data"", ""venue"": ""Proceedings of the Eighteenth International"", ""year"": 2001}, {""authors"": [""Lazaridou"", ""Angeliki"", ""Ivan Titov"", ""Caroline Sporleder.""], ""title"": ""A Bayesian model for joint unsupervised induction of sentiment, aspect and discourse representations"", ""venue"": ""Proceedings of the 51st Annual Meeting of the"", ""year"": 2013}, {""authors"": [""Li"", ""Jiwei"", ""Rumeng Li"", ""Eduard Hovy.""], ""title"": ""Recursive deep models for discourse parsing"", ""venue"": ""Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP),"", ""year"": 2014}, {""authors"": [""Li"", ""Sujian"", ""Liang Wang"", ""Ziqiang Cao"", ""Wenjie Li.""], ""title"": ""Text-level discourse dependency parsing"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long"", ""year"": 2014}, {""authors"": [""Louis"", ""Annie"", ""Aravind Joshi"", ""Ani Nenkova.""], ""title"": ""Discourse indicators for content selection in summarization"", ""venue"": ""Proceedings of the 11th Annual Meeting of the Special Interest Group on Discourse and"", ""year"": 2010}, {""authors"": [""Mach\u00e1\u010dek"", ""Matou\u0161"", ""Ond\u0159ej Bojar.""], ""title"": ""Results of the WMT14 metrics shared task"", ""venue"": ""Proceedings of the Ninth Workshop on Statistical Machine Translation, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Magerman"", ""David.""], ""title"": ""Statistical decision-tree models for parsing"", ""venue"": ""Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics, ACL\u201995, pages 276\u2013283, Cambridge, MA."", ""year"": 1995}, {""authors"": [""Mann"", ""William"", ""Sandra Thompson.""], ""title"": ""Rhetorical structure theory: Toward a functional theory of text organization"", ""venue"": ""Text, 8(3):243\u2013281."", ""year"": 1988}, {""authors"": [""Marcu"", ""Daniel.""], ""title"": ""A decision-based approach to rhetorical parsing"", ""venue"": ""Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics on 432"", ""year"": 1999}, {""authors"": [""Marcu"", ""Daniel.""], ""title"": ""The rhetorical parsing of unrestricted texts: A surface-based approach"", ""venue"": ""Computational Linguistics, 26:395\u2013448."", ""year"": 2000}, {""authors"": [""Marcu"", ""Daniel.""], ""title"": ""The Theory and Practice of Discourse Parsing and Summarization"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 2000}, {""authors"": [""Marcu"", ""Daniel"", ""Abdessamad Echihabi.""], ""title"": ""An unsupervised approach to recognizing discourse relations"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for Computational"", ""year"": 2002}, {""authors"": [""Marcus"", ""Mitchell"", ""Mary Marcinkiewicz"", ""Beatrice Santorini.""], ""title"": ""Building a large annotated corpus of English: The Penn Treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1994}, {""authors"": [""Maslennikov"", ""Mstislav"", ""Tat-Seng Chua.""], ""title"": ""A multi-resolution framework for information extraction from free text"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2007}, {""authors"": [""McCallum"", ""Andrew.""], ""title"": ""MALLET: A machine learning for language toolkit"", ""venue"": ""http://mallet.cs.umass.edu."", ""year"": 2002}, {""authors"": [""McCallum"", ""Andrew"", ""Dayne Freitag"", ""Fernando C.N. Pereira.""], ""title"": ""Maximum entropy Markov models for information extraction and segmentation"", ""venue"": ""Proceedings of the Seventeenth International"", ""year"": 2000}, {""authors"": [""McDonald"", ""Ryan"", ""Koby Crammer"", ""Fernando Pereira.""], ""title"": ""Online large-margin training of dependency parsers"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Fernando Pereira"", ""Kiril Ribarov"", ""Jan Haji\u010d.""], ""title"": ""Non-projective dependency parsing using spanning tree algorithms"", ""venue"": ""Proceedings of the Conference on Human Language"", ""year"": 2005}, {""authors"": [""Morris"", ""Jane"", ""Graeme Hirst.""], ""title"": ""Lexical cohesion computed by thesaural relations as an indicator of structure of text"", ""venue"": ""Computational Linguistics, 17(1):21\u201348."", ""year"": 1991}, {""authors"": [""Murphy"", ""Kevin.""], ""title"": ""Machine Learning: A Probabilistic Perspective"", ""venue"": ""The MIT Press. Cambridge, MA."", ""year"": 2012}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts"", ""venue"": ""Proceedings of the 42nd Annual Meeting of the Association for"", ""year"": 2004}, {""authors"": [""Pitler"", ""Emily"", ""Ani Nenkova.""], ""title"": ""Using syntax to disambiguate explicit discourse connectives in text"", ""venue"": ""Proceedings of the ACL-IJCNLP 2009 Conference Short Papers, ACLShort \u201909, pages 13\u201316, Suntec."", ""year"": 2009}, {""authors"": [""Poole"", ""David"", ""Alan Mackworth""], ""title"": ""Artificial Intelligence: Foundations of Computational Agents"", ""year"": 2010}, {""authors"": [""Prasad"", ""Rashmi"", ""Nikhil Dinesh"", ""Alan Lee"", ""Eleni Miltsakaki"", ""Livio Robaldo"", ""Aravind Joshi"", ""Bonnie Webber""], ""title"": ""The Penn Discourse TreeBank 2.0"", ""venue"": ""In Proceedings of the Sixth International Conference"", ""year"": 2008}, {""authors"": [""Prasad"", ""Rashmi"", ""Aravind Joshi"", ""Nikhil Dinesh"", ""Alan Lee"", ""Eleni Miltsakaki"", ""Bonnie Webber.""], ""title"": ""The Penn Discourse TreeBank as a resource for natural language generation"", ""venue"": ""Proceedings of the"", ""year"": 2005}, {""authors"": [""Regneri"", ""Michaela"", ""Markus Egg"", ""Alexander Koller.""], ""title"": ""Efficient processing of underspecified discourse representations"", ""venue"": ""Proceedings of the 46th Annual Meeting of the Association for"", ""year"": 2008}, {""authors"": [""Reyle"", ""Uwe.""], ""title"": ""Dealing with ambiguities by underspecification: Construction, representation and deduction"", ""venue"": ""Journal of Semantics, 10(2):123\u2013179."", ""year"": 1993}, {""authors"": [""Schapire"", ""Robert E."", ""Yoram Singer.""], ""title"": ""Boostexter: A boosting-based system for text categorization"", ""venue"": ""Machine Learning, 39(2\u20133):135\u2013168."", ""year"": 2000}, {""authors"": [""Schauer"", ""Holger"", ""Udo Hahn.""], ""title"": ""Anaphoric cues for coherence relations"", ""venue"": ""Proceedings of the Conference on Recent Advances in Natural Language Processing, RANLP \u201901, pages 228\u2013234,"", ""year"": 2001}, {""authors"": [""Schilder"", ""Frank""], ""title"": ""Robust discourse parsing via discourse markers, topicality"", ""year"": 2002}, {""authors"": [""Sha"", ""Fei"", ""Fernando Pereira.""], ""title"": ""Shallow parsing with conditional random fields"", ""venue"": ""Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human"", ""year"": 2003}, {""authors"": [""Silber"", ""Gregory"", ""Kathleen McCoy.""], ""title"": ""Efficiently computed lexical chains as an intermediate representation for automatic text summarization"", ""venue"": ""Computational Linguistics, 28(4):487\u2013496."", ""year"": 2002}, {""authors"": [""Smith"", ""Noah A.""], ""title"": ""Linguistic Structure Prediction"", ""venue"": ""Synthesis Lectures on Human Language Technologies. Morgan and Claypool."", ""year"": 2011}, {""authors"": [""Socher"", ""Richard"", ""John Bauer"", ""Christopher D. Manning"", ""Ng Andrew Y.""], ""title"": ""Parsing with compositional vector grammars"", ""venue"": ""Proceedings of the 51st Annual Meeting of the Association for Computational"", ""year"": 2013}, {""authors"": [""Socher"", ""Richard"", ""Alex Perelygin"", ""Jean Wu"", ""Jason Chuang"", ""Christopher D. Manning"", ""Andrew Ng"", ""Christopher Potts""], ""title"": ""Recursive deep models for semantic compositionality over a sentiment"", ""year"": 2013}, {""authors"": [""S. Somasundaran""], ""title"": ""Discourse-Level Relations for Opinion Analysis"", ""venue"": ""Ph.D. thesis, University of Pittsburgh, PA."", ""year"": 2010}, {""authors"": [""Soricut"", ""Radu"", ""Daniel Marcu.""], ""title"": ""Sentence level discourse parsing using syntactic and lexical information"", ""venue"": ""Proceedings of the 2003 Conference of the North American Chapter of the Association for"", ""year"": 2003}, {""authors"": [""Sporleder"", ""Caroline.""], ""title"": ""Manually vs"", ""venue"": ""automatically labelled data in discourse relation classification. Effects of example and feature selection. LDV Forum, 22(1):1\u201320."", ""year"": 2007}, {""authors"": [""Sporleder"", ""Caroline"", ""Mirella Lapata.""], ""title"": ""Automatic paragraph identification: A study across languages and domains"", ""venue"": ""Proceedings of the 2004 Conference on Empirical Methods in Natural Language"", ""year"": 2004}, {""authors"": [""Sporleder"", ""Caroline"", ""Mirella Lapata.""], ""title"": ""Discourse chunking and its application to sentence compression"", ""venue"": ""Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural"", ""year"": 2005}, {""authors"": [""Sporleder"", ""Caroline"", ""Alex Lascarides.""], ""title"": ""Exploiting linguistic cues to classify rhetorical relations"", ""venue"": ""Proceedings of Recent Advances in Natural Language Processing (RANLP), pages 157\u2013166, Bulgaria."", ""year"": 2005}, {""authors"": [""Sporleder"", ""Caroline"", ""Alex Lascarides.""], ""title"": ""Using automatically labelled examples to classify rhetorical relations: An assessment"", ""venue"": ""Natural Language Engineering, 14(3):369\u2013416."", ""year"": 2008}, {""authors"": [""Stede"", ""Manfred.""], ""title"": ""The Potsdam Commentary Corpus"", ""venue"": ""Proceedings of the ACL-04 Workshop on Discourse Annotation, pages 96\u2013102, Barcelona."", ""year"": 2004}, {""authors"": [""Stede"", ""Manfred.""], ""title"": ""Discourse Processing"", ""venue"": ""Synthesis Lectures on Human Language Technologies. Morgan and Claypool Publishers."", ""year"": 2011}, {""authors"": [""Subba"", ""Rajen"", ""Barbara Di-Eugenio.""], ""title"": ""An effective discourse parser that uses rich linguistic information"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North"", ""year"": 2009}, {""authors"": [""Sutton"", ""Charles"", ""Andrew McCallum.""], ""title"": ""An introduction to conditional random fields"", ""venue"": ""Foundations and Trends in Machine Learning, 4(4):267\u2013373."", ""year"": 2012}, {""authors"": [""Sutton"", ""Charles"", ""Andrew McCallum"", ""Khashayar Rohanimanesh.""], ""title"": ""Dynamic conditional random fields: Factorized probabilistic models for labeling and segmenting sequence data"", ""venue"": ""Journal of"", ""year"": 2007}, {""authors"": [""Taboada"", ""Maite.""], ""title"": ""Discourse markers as signals (or not) of rhetorical relations"", ""venue"": ""Journal of Pragmatics, 38(4):567\u2013592."", ""year"": 2006}, {""authors"": [""Taboada"", ""Maite"", ""Julian Brooke"", ""Milan Tofiloski"", ""Kimberly Voll"", ""Manfred Stede.""], ""title"": ""Lexicon-based methods for sentiment analysis"", ""venue"": ""Computational Linguistics, 37(2):267\u2013307."", ""year"": 2011}, {""authors"": [""Taboada"", ""Maite"", ""William C. Mann.""], ""title"": ""Rhetorical structure theory: Looking back and moving ahead"", ""venue"": ""Discourse Studies, 8(3):423\u2013459."", ""year"": 2006}, {""authors"": [""Teufel"", ""Simone"", ""Marc Moens""], ""title"": ""Summarizing scientific articles: Experiments with relevance and rhetorical"", ""year"": 2002}, {""authors"": [""Verberne"", ""Suzan"", ""Lou Boves"", ""Nelleke Oostdijk"", ""Peter-Arno Coppen.""], ""title"": ""Evaluating discourse-based answer extraction for why-question answering"", ""venue"": ""Proceedings of the 30th Annual International"", ""year"": 2007}, {""authors"": [""Vliet"", ""Nynke"", ""Gisela Redeker.""], ""title"": ""Complex sentences as leaky units in discourse parsing"", ""venue"": ""Proceedings of Constraints in Discourse, pages 1\u20139, Agay\u2013Saint Raphael."", ""year"": 2011}, {""authors"": [""B. Webber""], ""title"": ""D-LTAG: Extending lexicalized TAG to discourse"", ""venue"": ""Cognitive Science, 28(5):751\u2013779."", ""year"": 2004}, {""authors"": [""Wick"", ""Michael"", ""Khashayar Rohanimanesh"", ""Kedare Bellare"", ""Aron Culotta"", ""Andrew McCallum.""], ""title"": ""Samplerank: Training factor graphs with atomic gradients"", ""venue"": ""Proceedings of the 28th"", ""year"": 2011}, {""authors"": [""Wolf"", ""Florian"", ""Edward Gibson.""], ""title"": ""Representing discourse coherence: A corpus-based study"", ""venue"": ""Computational Linguistics, 31:249\u2013288."", ""year"": 2005}, {""authors"": [""Zhang"", ""Yuan"", ""Tao Lei"", ""Regina Barzilay"", ""Tommi Jaakkola"", ""Amir Globerson.""], ""title"": ""Steps to excellence: Simple inference with refined scoring of dependency trees"", ""venue"": ""Proceedings of the 52nd Annual Meeting of"", ""year"": 2014}]","acknowledgments  :The authors acknowledge the funding support of NSERC Canada Graduate Scholarship (CGS-D). Many thanks to Bonnie Webber, Amanda Stent, Carolyn Rose, Lluis Marquez, Samantha Wray, and the anonymous reviewers for their insightful comments on an earlier version of this article.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :A well-written text is not merely a sequence of independent and isolated sentences, but instead a sequence of structured and related sentences, where the meaning of a sentence relates to the previous and the following ones. In other words, a well-written
∗ Arabic Language Technologies, Qatar Computing Research Institute, Qatar Foundation, Doha, Qatar. E-mail: sjoty@qf.org.qa.
∗∗ Computer Science Department, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4. E-mail: carenini@cs.ubc.ca.
† Computer Science Department, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4. E-mail: rng@cs.ubc.ca.
Submission received: 11 May 2014; revised version received: 29 January 2015; accepted for publication: 18 March 2015.
doi:10.1162/COLI a 00226
No rights reserved. This work was authored as part of the Contributor’s official duties as an Employee of the United States Government and is therefore a work of the United States Government. In accordance with 17 U.S.C. 105, no copyright protection is available for such works under U.S. Law.
text has a coherence structure (Halliday and Hasan 1976; Hobbs 1979), which logically binds its clauses and sentences together to express a meaning as a whole. Rhetorical analysis seeks to uncover this coherence structure underneath the text; this has been shown to be beneficial for many Natural Language Processing (NLP) applications, including text summarization and compression (Marcu 2000b; Daumé and Marcu 2002; Sporleder and Lapata 2005; Louis, Joshi, and Nenkova 2010), text generation (Prasad et al. 2005), machine translation evaluation (Guzmán et al. 2014a, 2014b; Joty et al. 2014), sentiment analysis (Somasundaran 2010; Lazaridou, Titov, and Sporleder 2013), information extraction (Teufel and Moens 2002; Maslennikov and Chua 2007), and question answering (Verberne et al. 2007). Furthermore, rhetorical structures can be useful for other discourse analysis tasks, including co-reference resolution using Veins theory (Cristea, Ide, and Romary 1998).
Different formal theories of discourse have been proposed from different viewpoints to describe the coherence structure of a text. For example, Martin (1992) and Knott and Dale (1994) propose discourse relations based on the usage of discourse connectives (e.g., because, but) in the text. Asher and Lascarides (2003) propose Segmented Discourse Representation Theory, which is driven by sentence semantics. Webber (2004) and Danlos (2009) extend sentence grammar to formalize discourse structure. Rhetorical Structure Theory (RST), proposed by Mann and Thompson (1988), is perhaps the most influential theory of discourse in computational linguistics. Although it was initially intended to be used in text generation, later it became popular as a framework for parsing the structure of a text (Taboada and Mann 2006). RST represents texts by labeled hierarchical structures, called Discourse Trees (DTs). For example, consider the DT shown in Figure 1 for the following text:
But he added: “Some people use the purchasers’ index as a leading indicator, some use it as a coincident indicator. But the thing it’s supposed to measure—manufacturing strength—it missed altogether last month.”
The leaves of a DT correspond to contiguous atomic text spans, called elementary discourse units (EDUs; six in the example). EDUs are clause-like units that serve as building blocks. Adjacent EDUs are connected by coherence relations (e.g., Elaboration, Contrast), forming larger discourse units (represented by internal nodes), which in turn are also subject to this relation linking. Discourse units linked by a rhetorical relation are
further distinguished based on their relative importance in the text: nuclei are the core parts of the relation and satellites are peripheral or supportive ones. For example, in Figure 1, Elaboration is a relation between a nucleus (EDU 4) and a satellite (EDU 5), and Contrast is a relation between two nuclei (EDUs 2 and 3). Carlson, Marcu, and Okurowski (2002) constructed the first large RST-annotated corpus (RST–DT) on Wall Street Journal articles from the Penn Treebank. Whereas Mann and Thompson (1988) had suggested about 25 relations, the RST–DT uses 53 mono-nuclear and 25 multi-nuclear relations. The relations are grouped into 16 coarse-grained categories; see Carlson and Marcu (2001) for a detailed description of the relations. Conventionally, rhetorical analysis in RST involves two subtasks: discourse segmentation is the task of breaking the text into a sequence of EDUs, and discourse parsing is the task of linking the discourse units (EDUs and larger units) into a labeled tree. In this article, we use the terms discourse parsing and rhetorical parsing interchangeably.
While recent advances in automatic discourse segmentation have attained high accuracies (an F-score of 90.5% reported by Fisher and Roark [2007]), discourse parsing still poses significant challenges (Feng and Hirst 2012) and the performance of the existing discourse parsers (Soricut and Marcu 2003; Subba and Di-Eugenio 2009; Hernault et al. 2010) is still considerably inferior compared with the human gold standard. Thus, the impact of rhetorical structure in downstream NLP applications is still very limited. The work we present in this article aims to reduce this performance gap and take discourse parsing one step further. To this end, we address three key limitations of existing discourse parsers.
First, existing discourse parsers typically model the structure and the labels of a DT separately, and also do not take into account the sequential dependencies between the DT constituents. However, for several NLP tasks, it has recently been shown that joint models typically outperform independent or pipeline models (Murphy 2012, page 687). This is also supported in a recent study by Feng and Hirst (2012), in which the performance of a greedy bottom–up discourse parser improved when sequential dependencies were considered by using gold annotations for the neighboring (i.e., previous and next) discourse units as contextual features in the parsing model. To address this limitation of existing parsers, as the first contribution, we propose a novel discourse parser based on probabilistic discriminative parsing models, expressed as Conditional Random Fields (CRFs) (Sutton, McCallum, and Rohanimanesh 2007), to infer the probability of all possible DT constituents. The CRF models effectively represent the structure and the label of a DT constituent jointly, and, whenever possible, capture the sequential dependencies.
Second, existing discourse parsers typically apply greedy and sub-optimal parsing algorithms to build a DT. To cope with this limitation, we use the inferred (posterior) probabilities from our CRF parsing models in a probabilistic CKY-like bottom–up parsing algorithm (Jurafsky and Martin 2008), which is non-greedy and optimal. Furthermore, a simple modification of this parsing algorithm allows us to generate k-best (i.e., the k highest probability) parse hypotheses for each input text that could then be used in a reranker to improve over the initial ranking using additional (global) features of the discourse tree as evidence, a strategy that has been successfully explored in syntactic parsing (Charniak and Johnson 2005; Collins and Koo 2005).
Third, most of the existing discourse parsers do not discriminate between intrasentential parsing (i.e., building the DTs for the individual sentences) and multisentential parsing (i.e., building the DT for the whole document). However, we argue that distinguishing between these two parsing conditions can result in more effective parsing. Two separate parsing models could exploit the fact that rhetorical relations
are distributed differently intra-sententially versus multi-sententially. Also, they could independently choose their own informative feature sets. As another key contribution of our work, we devise two different parsing components: one for intra-sentential parsing, the other for multi-sentential parsing. This provides for scalable, modular, and flexible solutions that can exploit the strong correlation observed between the text structure (i.e., sentence boundaries) and the structure of the discourse tree.
In order to develop a complete and robust discourse parser, we combine our intrasentential and multi-sentential parsing components in two different ways. Because most sentences have a well-formed discourse sub-tree in the full DT (e.g., the second sentence in Figure 1), our first approach constructs a DT for every sentence using our intrasentential parser, and then runs the multi-sentential parser on the resulting sentencelevel DTs to build a complete DT for the whole document. However, this approach would fail in those cases where discourse structures violate sentence boundaries, also called “leaky” boundaries (Vliet and Redeker 2011). For example, consider the first sentence in Figure 1. It does not have a well-formed discourse sub-tree because the unit containing EDUs 2 and 3 merges with the next sentence and only then is the resulting unit merged with EDU 1. Our second approach, in order to deal with these leaky cases, builds sentence-level sub-trees by applying the intra-sentential parser on a sliding window covering two adjacent sentences and by then consolidating the results produced by overlapping windows. After that, the multi-sentential parser takes all these sentence-level sub-trees and builds a full DT for the whole document.
Our discourse parser assumes that the input text has already been segmented into elementary discourse units. As an additional contribution, we propose a novel discriminative approach to discourse segmentation that not only achieves state-of-theart performance, but also reduces time and space complexities by using fewer features. Notice that the combination of our segmenter with our parser forms a COmplete probabilistic Discriminative framework for Rhetorical Analysis (CODRA).
Whereas previous systems have been tested on only one corpus, we evaluate our framework on texts from two very different genres: news articles and instructional howto manuals. The results demonstrate that our approach to discourse parsing provides consistent and statistically significant improvements over previous methods both at the sentence level and at the document level. The performance of our final system compares very favorably to the performance of state-of-the-art discourse parsers. Finally, the oracle accuracy computed based on the k-best parse hypotheses generated by our parser demonstrates that a reranker could potentially improve the accuracy further.
After discussing related work in Section 2, we present our rhetorical analysis framework in Section 3. In Section 4, we describe our discourse parser. Then, in Section 5 we present our discourse segmenter. The experiments and analysis of results are presented in Section 6. Finally, we summarize our contributions with future directions in Section 7. 2 related work :Rhetorical analysis has a long history—dating back to Mann and Thompson (1988), when RST was initially proposed as a useful linguistic method for describing natural texts, to more recent attempts to automatically extract the rhetorical structure of a given text (Hernault et al. 2010). In this section, we provide a brief overview of the computational approaches that follow RST as the theory of discourse, and that are related to our work; see the survey by Stede (2011) for a broader overview that also includes other theories of discourse. Although the most effective approaches to rhetorical analysis to date rely on supervised machine learning methods trained on human-annotated data, unsupervised methods have also been proposed, as they do not require human-annotated data and can be more easily applied to new domains.
Often, discourse connectives like but, because, and although convey clear information on the kind of relation linking the two text segments. In his early work, Marcu (2000a) presented a shallow rule-based approach relying on discourse connectives (or cues) and surface patterns. He used hand-coded rules, derived from an extensive corpus study, to break the text into EDUs and to build DTs for sentences first, then for paragraphs, and so on. Despite the fact that this work pioneered the field of rhetorical analysis, it has many limitations. First, identifying discourse connectives is a difficult task on its own, because (depending on the usage), the same phrase may or may not signal a discourse relation (Pitler and Nenkova 2009). For example, but can either signal a Contrast discourse relation or can simply perform non-discourse acts. Second, discourse segmentation using only discourse connectives fails to attain high accuracy (Soricut and Marcu 2003). Third, DT structures do not always correspond to paragraph structures; for example, Sporleder and Lapata (2004) report that more than 20% of the paragraphs in the RST–DT corpus (Carlson, Marcu, and Okurowski 2002) do not correspond to a discourse unit in the DT. Fourth, discourse cues are sometimes ambiguous; for example, but can signal Contrast, Antithesis and Concession, and so on.
Finally, a more serious problem with the rule-based approach is that often rhetorical relations are not explicitly signaled by discourse cues. For example, in RST–DT, Marcu and Echihabi (2002) found that only 61 out of 238 Contrast relations and 79 out of 307 Cause–Explanation relations were explicitly signaled by cue phrases. In the British National Corpus, Sporleder and Lascarides (2008) report that half of the sentences lack a discourse cue. Other studies (Schauer and Hahn 2001; Stede 2004; Taboada 2006; Subba and Di-Eugenio 2009) report even higher figures: About 60% of discourse relations are not explicitly signaled. Therefore, rather than relying on hand-coded rules based on discourse cues and surface patterns, recent approaches use machine learning techniques with a large set of informative features.
While some rhetorical relations need to be explicitly signaled by discourse cues (e.g., Concession) and some do not (e.g., Background), there is a large middle ground of relations that may be signaled or not. For these “middle ground” relations, can we exploit features present in the signaled cases to automatically identify relations when they are not explicitly signaled? The idea is to use unambiguous discourse cues (e.g., although for Contrast, for example for Elaboration) to automatically label a large corpus with rhetorical relations that could then be used to train a supervised model.1
A series of previous studies have explored this idea. Marcu and Echihabi (2002) first attempted to identify four broad classes of relations: Contrast, Elaboration, Condition, and Cause–Explanation–Evidence. They used a naive Bayes classifier based on word pairs (w1, w2), where w1 occurs in the left segment, and w2 occurs in the right segment. Sporleder and Lascarides (2005) included other features (e.g., words and their stems, Part-of-Speech [POS] tags, positions, segment lengths) in a boosting-based classifier (i.e., BoosTexter [Schapire and Singer 2000]) to further improve relation classification accuracy. However, these studies evaluated classification performance on the instances
1 We categorize this approach as unsupervised because it does not rely on human-annotated data.
where rhetorical relations were originally signaled (i.e., the discourse cues were artificially removed), and did not verify how well this approach performs on the instances that are not originally signaled. Subsequent studies (Blair-Goldensohn, McKeown, and Rambow 2007; Sporleder 2007; Sporleder and Lascarides 2008) confirm that classifiers trained on instances stripped of their original discourse cues do not generalize well to implicit cases because they are linguistically quite different.
Note that this approach to identifying discourse relations in the absence of manually labeled data does not fully solve the parsing problem (i.e., building DTs); rather, it only attempts to identify a small subset of coarser relations between two (flat) text segments (i.e., a tagging problem). Arguably, to perform a complete rhetorical analysis, one needs to use supervised machine learning techniques based on human-annotated data. Marcu (1999) applies supervised machine learning techniques to build a discourse segmenter and a shift–reduce discourse parser. Both the segmenter and the parser rely on C4.5 decision tree classifiers (Poole and Mackworth 2010) to learn the rules automatically from the data. The discourse segmenter mainly uses discourse cues, shallowsyntactic (i.e., POS tags) and contextual features (i.e., neighboring words and their POS tags). To learn the shift–reduce actions, the discourse parser encodes five types of features: lexical (e.g., discourse cues), shallow-syntactic, textual similarity, operational (previous n shift–reduce operations), and rhetorical sub-structural features. Despite the fact that this work has pioneered many of today’s machine learning approaches to discourse parsing, it has all the limitations mentioned in Section 1.
The work of Marcu (1999) is considerably improved by Soricut and Marcu (2003). They present the publicly available SPADE system,2 which comes with probabilistic models for discourse segmentation and sentence-level discourse parsing. Their segmentation and parsing models are based on lexico-syntactic patterns (or features) extracted from the lexicalized syntactic tree of a sentence. The discourse parser uses an optimal parsing algorithm to find the most probable DT structure for a sentence. SPADE was trained and tested on the RST–DT corpus. This work, by showing empirically the connection between syntax and discourse structure at the sentence level, has greatly influenced all major contributions in this area ever since. However, it is limited in several ways. First, SPADE does not produce a full-text (i.e., document-level) parse. Second, it applies a generative parsing model based on only lexico-syntactic features, whereas discriminative models are generally considered to be more accurate, and can incorporate arbitrary features more effectively (Murphy 2012). Third, the parsing model makes an independence assumption between the label and the structure of a DT constituent, and it ignores the sequential and the hierarchical dependencies between the DT constituents.
Subsequent research addresses the question of how much syntax one really needs in rhetorical analysis. Sporleder and Lapata (2005) focus on the discourse chunking problem, comprising two subtasks: discourse segmentation and (flat) nuclearity assignment. They formulate discourse chunking in two alternative ways. First, one-step classification, where the discourse chunker, a multi-class classifier, assigns to each token one of the four labels: (1) B–NUC (beginning of a nucleus), (2) I–NUC (inside a nucleus), (3) B– SAT (beginning of a satellite), and (4) I–SAT (inside a satellite). Therefore, this approach performs discourse segmentation and nuclearity assignment simultaneously. Second,
2 http://www.isi.edu/licensed-sw/spade/.
two-step classification, where in the first step, the discourse segmenter (a binary classifier) labels each token as either B (beginning of an EDU) or I (inside an EDU). Then, in the second step, a nuclearity labeler (another binary classifier) assigns a nuclearity status to each segment. The two-step approach avoids illegal chunk sequences like a B–NUC followed by an I–SAT or a B–SAT followed by an I–NUC, and in this approach, it is easier to incorporate sentence-level properties like the constraint that a sentence must contain at least one nucleus. They examine whether shallow-syntactic features (e.g., POS and phrase tags) would be sufficient for these purposes. The evaluation on the RST–DT shows that the two-step approach outperforms the one-step approach, and its performance is comparable to that of SPADE, which requires relatively expensive full syntactic parses.
In follow–up work, Fisher and Roark (2007) demonstrate over 4% absolute performance gain in discourse segmentation, by combining the features extracted from the syntactic tree with the ones derived via POS tagging and shallow syntactic parsing (i.e., chunking). Using quite a large number of features in a binary log-linear model, they achieve state-of-the-art performance in discourse segmentation on the RST–DT test set.
In a different approach, Regneri, Egg, and Koller (2008) propose to use Underspecified Discourse Representation (UDR) as an intermediate representation for discourse parsing. Underspecified representations offer a single compact representation to express possible ambiguities in a linguistic structure, and have been primarily used to deal with scope ambiguity in semantic structures (Reyle 1993; Egg, Koller, and Niehren 2001; Althaus et al. 2003; Koller, Regneri, and Thater 2008). Assuming that a UDR of a DT is already given in the form of a dominance graph (Althaus et al. 2003), Regneri, Egg, and Koller (2008) convert it into a more expressive and complete UDR representation called regular tree grammar (Koller, Regneri, and Thater 2008), for which efficient algorithms (Knight and Graehl 2005) already exist to derive the best configuration (i.e., the best discourse tree).
Hernault et al. (2010) present the publicly available HILDA system,3 which comes with a discourse segmenter and a parser based on Support Vector Machines (SVMs). The discourse segmenter is a binary SVM classifier that uses the same lexico-syntactic features used in SPADE, but with more context (i.e., the lexico-syntactic features for the previous two words and the following two words). The discourse parser iteratively uses two SVM classifiers in a pipeline to build a DT. In each iteration, a binary classifier first decides which of the adjacent units to merge, then a multi-class classifier connects the selected units with an appropriate relation label. Using this simple method, they report promising results in document-level discourse parsing on the RST–DT.
For a different genre, instructional texts, Subba and Di-Eugenio (2009) propose a shift–reduce discourse parser that relies on a classifier for relation labeling. Their classifier uses Inductive Logic Programming (ILP) to learn first-order logic rules from a large set of features including the linguistically rich compositional semantics coming from a semantic parser. They demonstrate that including compositional semantics with other features improves the performance of the classifier, thus, also improves the performance of the parser.
Both HILDA and the ILP-based approach of Subba and Di-Eugenio (2009) are limited in several ways. First, they do not differentiate between intra- and multi-sentential
3 http://nlp.prendingerlab.net/hilda/.
parsing, and both scenarios use a single uniform parsing model. Second, they take a greedy (i.e., sub-optimal) approach to construct a DT. Third, they disregard sequential dependencies between DT constituents. Furthermore, HILDA considers the structure and the labels of a DT separately. Our discourse parser CODRA, as described in the next section, addresses all these limitations.
More recent work than ours also attempts to address some of the above-mentioned limitations of the existing discourse parsers. Similar to us, Feng and Hirst (2014) generate a document-level DT in two stages, where a multi-sentential parsing follows an intra-sentential one. At each stage, they iteratively use two separate linear-chain CRFs (Lafferty, McCallum, and Pereira 2001) in a cascade: one for predicting the presence of rhetorical relations between adjacent discourse units in a sequence, and the other to predict the relation label between the two most probable adjacent units to be merged as selected by the previous CRF. While they use CRFs to take into account the sequential dependencies between DT constituents, they use them greedily during parsing to achieve efficiency. They also propose a greedy post-editing step based on an additional feature (i.e., depth of a discourse unit) to modify the initial DT, which gives them a significant gain in performance. In a different approach, Li et al. (2014) propose a discourse-level dependency structure to capture direct relationships between EDUs rather than deep hierarchical relationships. They first create a discourse dependency treebank by converting the deep annotations in RST–DT to shallow head-dependent annotations between EDUs. To find the dependency parse (i.e., an optimal spanning tree) for a given text, they apply Eisner (1996) and Maximum Spanning Tree (McDonald et al. 2005) dependency parsing algorithms with the Margin Infused Relaxed Algorithm online learning framework (McDonald, Crammer, and Pereira 2005).
With the successful application of deep learning to numerous NLP problems including syntactic parsing (Socher et al. 2013a), sentiment analysis (Socher et al. 2013b), and various tagging tasks (Collobert et al. 2011), a couple of recent studies in discourse parsing also use deep neural networks (DNNs) and related feature representation methods. Inspired by the work of Socher et al. (2013a, 2013b), Li, Li, and Hovy (2014) propose a recursive DNN for discourse parsing. However, as in Socher et al. (2013a, 2013b), word vectors (i.e., embeddings) are not learned explicitly for the task, rather they are taken from Collobert et al. (2011). Given the vectors of the words in an EDU, their model first composes them hierarchically based on a syntactic parse tree to get the vector representation for the EDU. Adjacent discourse units are then merged hierarchically to get the vector representations for the higher order discourse units. In every step, the merging is done using one binary (structure) and one multi-class (relation) classifier, each having a three-layer neural network architecture. The cost function for training the model is given by these two cascaded classifiers applied at different levels of the DT. Similar to our method, they use the classifier probabilities in a CKY-like parsing algorithm to find the global optimal DT. Finally, Ji and Eisenstein (2014) present a feature representation learning method in a shift–reduce discourse parser (Marcu 1999). Unlike DNNs, which learn non-linear feature transformations in a maximum likelihood model, they learn linear transformations of features in a max margin classification model. 3 overview of our rhetorical analysis framework :CODRA takes as input a raw text and produces a discourse tree that describes the text in terms of coherence relations that hold between adjacent discourse units (i.e., clauses, sentences) in the text. An example DT generated by an online demo of CODRA
is shown in Appendix A.4 The color of a node represents its nuclearity status: blue denoting nucleus and yellow denoting satellite. The demo also allows some useful user interactions—for example, collapsing or expanding a node, highlighting an EDU, and so on.5
CODRA follows a pipeline architecture, shown in Figure 2. Given a raw text, the first task in the rhetorical analysis pipeline is to break the text into a sequence of EDUs (i.e., discourse segmentation). Because it is taken for granted that sentence boundaries are also EDU boundaries (i.e., EDUs do not span across multiple sentences), the discourse segmentation task boils down to finding EDU boundaries inside sentences. CODRA uses a maximum entropy model for discourse segmentation (see Section 5).
Once the EDUs are identified, the discourse parsing problem is determining which discourse units (EDUs or larger units) to relate (i.e., the structure), and what relations (i.e., the labels) to use in the process of building the DT. Specifically, discourse parsing requires: (1) a parsing model to explore the search space of possible structures and labels for their nodes, and (2) a parsing algorithm for selecting the best parse tree(s) among the candidates. A probabilistic parsing model like ours assigns a probability to every possible DT. The parsing algorithm then picks the most probable DTs.
The existing discourse parsers (Marcu 1999; Soricut and Marcu 2003; Subba and Di-Eugenio 2009; Hernault et al. 2010) described in Section 2 use parsing models that disregard the structural interdependencies between the DT constituents. However, we hypothesize that, like syntactic parsing, discourse parsing is also a structured prediction problem, which involves predicting multiple variables (i.e., the structure and the relation labels) that depend on each other (Smith 2011). Recently, Feng and Hirst (2012) also found these interdependencies to be critical for parsing performance. To capture the structural dependencies between the DT constituents, CODRA uses undirected conditional graphical models (i.e., CRFs) as its parsing models.
To find the most probable DT, unlike most previous studies (Marcu 1999; Subba and Di-Eugenio 2009; Hernault et al. 2010), which adopt a greedy solution, CODRA applies an optimal CKY parsing algorithm to the inferred posterior probabilities (obtained from the CRFs) of all possible DT constituents. Furthermore, the parsing algorithm allows CODRA to generate a list of k-best parse hypotheses for a given text.
4 The demo of CODRA is available at http://109.228.0.153/Discourse Parser Demo/. The source code of CODRA is available from http://alt.qcri.org/tools/. 5 The input text in the demo in Appendix A is taken from www.bbc.co.uk/news/world-asia-26106490.
Note that the way CRFs and CKY are used in CODRA is quite different from the way they are used in syntactic parsing. For example, in the CRF-based constituency parsing proposed by Finkel, Kleeman, and Manning (2008), the conditional probability distribution of a parse tree given a sentence decomposes across factors defined over productions, and the standard inside–outside algorithm is used for inference on possible trees. In contrast, CODRA first uses the standard forward–backward algorithm in a “fat” chain structured6 CRF (to be discussed in Section 4.1.1) to compute the posterior probabilities of all possible DT constituents for a given text (i.e., EDUs); then it uses a CKY parsing algorithm to combine those probabilities and find the most probable DT.
Another crucial question related to parsing models is whether to use a single model or two different models for parsing at the sentence-level (i.e., intra-sentential) and at the document-level (i.e., multi-sentential). A simple and straightforward strategy would be to use a single unified parsing model for both intra- and multi-sentential parsing without distinguishing the two cases, as was previously done (Marcu 1999; Subba and Di-Eugenio 2009; Hernault et al. 2010). That approach has the advantages of making the parsing process easier, and the model gets more data to learn from. However, for a solution like ours, which tries to capture the interdependencies between constituents, this would be problematic with respect to scalability and inappropriate because of two modeling issues.
More specifically, for scalability note that the number of valid trees grows exponentially with the number of EDUs in a document.7 Therefore, an exhaustive search over all the valid DTs is often infeasible, even for relatively small documents.
For modeling, a single unified approach is inappropriate for two reasons. On the one hand, it appears that discourse relations are distributed differently intra- versus multi-sententially. For example, Figure 3 shows a comparison between the two distributions of the eight most frequent relations in the RST–DT training set. Notice that Same–Unit is more frequent than Joint in the intra-sentential case, whereas Joint is more frequent than Same–Unit in the multi-sentential case. Similarly, the relative distributions
6 By the term “fat” we refer to CRFs with multiple (interconnected) chains of output variables. 7 For n + 1 EDUs, the number of valid discourse tree structures (i.e., not counting possible variations in the
nuclearity and relation labels) is the Catalan number Cn.
of Background, Contrast, Cause, and Explanation are different in the two parsing scenarios. On the other hand, different kinds of features are applicable and informative for intraversus multi-sentential parsing. For example, syntactic features like dominance sets (Soricut and Marcu 2003) are extremely useful for parsing at the sentence-level, but are not even applicable in the multi-sentential case. Likewise, lexical chain features (Sporleder and Lapata 2004), which are useful for multi-sentential parsing, are not applicable at the sentence level.
Based on these above observations, CODRA comprises two separate modules: an intra-sentential parser and a multi-sentential parser, as shown in Figure 2. First, the intra-sentential parser produces one or more discourse sub-trees for each sentence. Then, the multi-sentential parser generates a full DT for the document from these sub-trees. Both of our parsers have the same two components: a parsing model and a parsing algorithm. Whereas the two parsing models are rather different, the same parsing algorithm is shared by the two modules. Staging multi-sentential parsing on top of intra-sentential parsing in this way allows CODRA to explicitly exploit the strong correlation observed between the text structure and the DT structure, as explained in detail in Section 4.3. 4 the discourse parser :Before describing the parsing models and the parsing algorithm of CODRA in detail, we introduce some terminology that we will use throughout this article.
A DT can be formally represented as a set of constituents of the form R[i, m, j], where i ≤ m < j. This refers to a rhetorical relation R between the discourse unit containing EDUs i through m and the discourse unit containing EDUs m+1 through j. For example, the DT for the second sentence in Figure 1 can be represented as {Elaboration–NS[4,4,5], Same–Unit–NN[4,5,6]}. Notice that in this representation, a relation R also specifies the nuclearity status of the discourse units involved, which can be one of Nucleus–Satellite (NS), Satellite–Nucleus (SN), or Nucleus–Nucleus (NN). Attaching nuclearity status to the relations allows us to perform the two subtasks of discourse parsing, relation identification and nuclearity assignment, simultaneously.
A common assumption made for generating DTs effectively is that they are binary trees (Soricut and Marcu 2003; Hernault et al. 2010). That is, multi-nuclear relations (e.g., Joint, Same–Unit) involving more than two discourse units are mapped to a hierarchical right-branching binary tree. For example, a flat Joint(e1, e2, e3, e4) (Figure 4a) is mapped to a right-branching binary tree Joint(e1, Joint(e2, Joint(e3, e4))) (Figure 4b). As mentioned before, the job of the intra- and multi-sentential parsing models of CODRA is to assign a probability to each of the constituents of all possible DTs at the sentence level and at the document level, respectively. Formally, given the model parameters Θ at a particular parsing scenario (i.e., sentence-level or document-level), for each possible constituent R[i, m, j] in a candidate DT at that parsing scenario, the parsing model estimates P(R[i, m, j]|Θ), which specifies a joint distribution over the label R and the structure [i, m, j] of the constituent. For example, when applied to the sentences in Figure 1 separately, the intra-sentential parsing model (with learned parameters Θs) estimates P(R[1, 1, 2]|Θs), P(R[2, 2, 3]|Θs), P(R[1, 2, 3]|Θs), and P(R[1, 1, 3]|Θs) for the first sentence, and P(R[4, 4, 5]|Θs), P(R[5, 5, 6]|Θs), P(R[4, 5, 6]|Θs), and P(R[4, 4, 6]|Θs) for the second sentence, respectively, for all R ranging over the set of relations.
4.1.1 Intra-Sentential Parsing Model. Figure 5 shows the parsing model of CODRA for intra-sentential parsing. The observed nodes Uj (at the bottom) in a sequence represent the discourse units (EDUs or larger units). The first layer of hidden nodes are the structure nodes, where Sj ∈ {0, 1} denotes whether two adjacent discourse units Uj−1 and Uj should be connected or not. The second layer of hidden nodes are the relation nodes, with Rj ∈ {1 . . .M} denoting the relation between two adjacent units Uj−1 and Uj, where M is the total number of relations in the relation set. The connections between adjacent nodes in a hidden layer encode sequential dependencies between the respective hidden nodes, and can enforce constraints such as the fact that a node must have a unique mother, namely, a Sj= 1 must not follow a Sj−1= 1. The connections between the two hidden layers model the structure and the relation of DT constituents jointly.
Notice that the probabilistic graphical model shown in Figure 5 is a chain-structured undirected graphical model (also known as Markov Random Field or MRF [Murphy 2012]) with two hidden layers, i.e., structure chain and relation chain. It becomes a Dynamic Conditional Random Field (DCRF) (Sutton, McCallum, and Rohanimanesh 2007) when we directly model the hidden (output) variables by conditioning the clique potentials (i.e., factors) on the observed (input) variables:
P(R2:t, S2:t|x,Θs) = 1Z(x,Θs) t−1∏ i=2 ϕ(Ri, Ri+1|x,Θs,r)ψ(Si, Si+1|x,Θs,s)ω(Ri, Si|x,Θs,c) (1)
where {ϕ} and {ψ} are the factors over the edges of the relation and structure chains, respectively, and {ω} are the factors over the edges connecting the relation and structure nodes (i.e., between-chain edges). Here, x represents input features extracted from the observed variables, Θs = [Θs,r,Θs,s,Θs,c] are model parameters, and Z(x,Θs) is the partition function. We use the standard log-linear representation of the factors:
ϕ(Ri, Ri+1|x,Θs,r) = exp(ΘTs,r f (Ri, Ri+1, x)) (2)
ψ(Si, Si+1|x,Θs,s) = exp(ΘTs,s f (Si, Si+1, x)) (3)
ω(Ri, Si|x,Θs,c) = exp(ΘTs,c f (Ri, Si, x)) (4)
where f (Y, Z, x) is a feature vector derived from the input features x and the local labels Y and Z, and Θs,y is the corresponding weight vector—that is, Θs,r and Θs,s are the weight vectors for the factors over the relation edges and the structure edges, respectively, and Θs,c is the weight vector for the factors over the between-chain edges.
A DCRF is a generalization of linear-chain CRFs (Lafferty, McCallum, and Pereira 2001) to represent complex interactions between output variables (i.e., labels), such as when performing multiple labeling tasks on the same sequence. Recently, there has been an explosion of interest in CRFs for solving structured output classification problems, with many successful applications in NLP including syntactic parsing (Finkel, Kleeman, and Manning 2008), syntactic chunking (Sha and Pereira 2003), and discourse chunking (Ghosh et al. 2011) in accordance with the Penn Discourse Treebank (Prasad et al. 2008).
DCRFs, being a discriminative approach to sequence modeling, have several advantages over their generative counterparts such as Hidden Markov Models (HMMs) and MRFs, which first model the joint distribution p(y, x|Θ), and then infer the conditional distribution p(y|x,Θ). It has been advocated that discriminative models are generally more accurate than generative ones because they do not “waste resources” modeling complex distributions that are observed (i.e., p(x)); instead, they focus directly on modeling what we care about, namely, the distribution of labels given the data (Murphy 2012).
Other key advantages include the ability to incorporate arbitrary overlapping local and global features, and the ability to relax strong independence assumptions. Furthermore, CRFs surmount the label bias problem (Lafferty, McCallum, and Pereira 2001) of the Maximum Entropy Markov Model (McCallum, Freitag, and Pereira 2000), which is considered to be a discriminative version of the HMM.
4.1.2 Training and Applying the Intra-Sentential Parsing Model. In order to obtain the probability of the constituents of all candidate DTs for a sentence, CODRA applies the intra-sentential parsing model (with learned parameters Θs) recursively to sequences at different levels of the DT, and computes the posterior marginals over the relation– structure pairs. It uses the standard forward–backward algorithm to compute the posterior marginals. To illustrate the process, let us assume that the sentence contains four EDUs, e1, · · · , e4 (see Figure 6). At the first (i.e., bottom) level of the DT, when all the discourse units are EDUs, there is only one unit sequence (e1, e2, e3, e4) to which CODRA applies the DCRF model. Figure 6a at the top left shows the corresponding DCRF model. For this sequence it computes the posterior marginals P(R2, S2=1|e1, e2, e3, e4,Θs),
P(R3, S3=1|e1, e2, e3, e4,Θs), and P(R4, S4=1| e1, e2, e3, e4,Θs) to obtain the probability of the DT constituents R[1, 1, 2], R[2, 2, 3], and R[3, 3, 4], respectively.
At the second level, there are three unit sequences: (e1:2, e3, e4), (e1,e2:3, e4), and (e1,e2,e3:4). Figure 6b shows their corresponding DCRF models. Notice that each of these sequences has a discourse unit that connects two EDUs, and the probability of this connection has already been computed at the previous level. CODRA computes the posterior marginals P(R3, S3=1|e1:2,e3,e4,Θs), P(R2:3S2:3=1|e1,e2:3,e4,Θs), P(R4, S4=1|e1,e2:3,e4,Θs), and P(R3:4, S3:4=1|e1,e2,e3:4,Θs) from these three sequences, which correspond to the probability of the constituents R[1, 2, 3], R[1, 1, 3], R[2, 3, 4], and R[2, 2, 4], respectively. Similarly, it attains the probability of the constituents R[1, 1, 4], R[1, 2, 4], and R[1, 3, 4] by computing their respective posterior marginals from the three sequences at the third (i.e., top) level of the candidate DTs (see Figure 6c).
Algorithm 1 describes how CODRA generates the unit sequences at different levels of the candidate DTs for a given number of EDUs in a sentence. Specifically, to compute the probability of a DT constituent R[i, k, j], CODRA generates sequences like (e1, · · · , ei−1, ei:k, ek+1:j, ej+1, · · · , en) for 1 ≤ i ≤ k < j ≤ n. However, in doing so, it may generate some duplicate sequences. Clearly, the sequence (e1, · · · , ei−1, ei:i, ei+1:j, ej+1, · · · , en) for 1 ≤ i ≤ k < j < n is already considered for computing the probability of the constituent R[i + 1, j, j + 1]. Therefore, it is a duplicate sequence that CODRA excludes from the list of sequences. The algorithm has a complexity of O(n3), where n is the number of EDUs in the sentence.
Once CODRA acquires the probability of all possible intra-sentential DT constituents, the discourse sub-trees for the sentences are built by applying an optimal parsing algorithm (Section 4.2) using one of the methods described in Section 4.3.
Algorithm 1 is also used to generate sequences for training the model (i.e., learning Θs). For example, Figure 7 demonstrates how we generate the training instances (right) from a gold DT with four EDUs (left). To find the relevant labels for the sequences
Algorithm 1 Generating unit sequences for a sentence with n EDUs. Input: Sequence of EDUs: (e1, e2, · · · , en) Output: List of sequences: L for i = 1 → n − 1 do // all possible starting positions for the subsequence
for j = i + 1 → n do // all possible ending positions for the subsequence if j == n then // sequences at top and bottom levels
for k = i → j − 1 do // all possible cut points within the subsequence L.append ((e1, · · · , ei−1, ei:k, ek+1:j, ej+1, · · · , en))
end else // sequences at intermediate levels
for k = i + 1 → j − 1 do // cut points excluding duplicate sequences L.append ((e1, · · · , ei−1, ei:k, ek+1:j, ej+1, · · · , en))
end end
end end
generated by the algorithm, we consult the gold DT and see if two discourse units are connected by a relation r (i.e., the corresponding labels are S = 1, R = r) or not (i.e., the corresponding labels are S = 0, R =NR). We train the model by maximizing the conditional likelihood of the labels in each of these training examples (see Equation (1)).
4.1.3 Multi-Sentential Parsing Model. Given the discourse units (sub-trees) for all the individual sentences in a document, a simple approach to build the DT of the document would be to apply a new DCRF model, similar to the one in Figure 5 (with different parameters), to all the possible sequences generated from these units by Algorithm 1 to infer the probability of all possible higher-order (multi-sentential) constituents. However, the number of possible sequences and their length increase with the number of sentences in a document. For example, assuming that each sentence has a well-formed DT, for a document with n sentences, Algorithm 1 generates O(n3) sequences, where
the sequence at the bottom level has n units, each of the sequences at the second level has n-1 units, and so on. Because the DCRF model in Figure 5 has a “fat” chain structure, one could use the forward–backward algorithm for exact inference in this model (Murphy 2012). Forward–backward on a sequence containing T units costs O(TM2) time, where M is the number of relations in our relation set. This makes the chainstructured DCRF model impractical for multi-sentential parsing of long documents, since learning requires running inference on every training sequence with an overall time complexity of O(TM2n3) = O(M2n4) per document (Sutton and McCallum 2012).
To address this problem, we have developed a simplified parsing model for multisentential parsing. Our model is shown in Figure 8. The two observed nodes Ut−1 and Ut are two adjacent (multi-sentential) discourse units. The (hidden) structure node S ∈ {0, 1} denotes whether the two discourse units should be linked or not. The other hidden node R ∈ {1 . . .M} represents the relation between the two units. Notice that similar to the model in Figure 5, this is also an undirected graphical model and becomes a CRF model if we directly model the labels by conditioning the clique potential ϕ on the input features x, derived from the observed variables:
P(Rt, St|x,Θd) = 1Z(x,Θd) ϕ(Rt, St|x,Θd) (5)
ϕ(Rt, St|x,Θd) = exp(ΘTd f (Rt, St, x)) (6)
where f (Rt, St, x) is a feature vector derived from the input features x and the labels Rt and St, and Θd is the corresponding weight vector. Although this model is similar in spirit to the parsing model in Figure 5, it now breaks the chain structure, which makes the inference much faster (i.e., a complexity of O(M2)). Breaking the chain structure also allows CODRA to balance the data for training (an equal number of instances with S=1 and S=0), which dramatically reduces the learning time of the model.
CODRA applies this parsing model to all possible adjacent units at all levels in the multi-sentential case, and computes the posterior marginals of the relation– structure pairs P(Rt, St=1|Ut−1, Ut,Θd) using the forward–backward algorithm to obtain the probability of all possible DT constituents. Given the sentence-level discourse units, Algorithm 2, which is a simplified variation of Algorithm 1, extracts all possible adjacent discourse units for multi-sentential parsing. Similar to Algorithm 1, Algorithm 2 also has a complexity of O(n3), where n is the number of sentence-level discourse units.
Algorithm 2 Generating all possible adjacent discourse units at all levels of a documentlevel discourse tree. Input: Sequence of units: (U1, U2, · · ·Un), where Ux[0]:= start EDU ID of unit x, and Ux[1]:= end EDU ID of unit x. Output: List of adjacent units: L for i = 1 → n − 1 do // all possible starting positions for the subsequence
for j = i + 1 → n do // all possible ending positions for the subsequence for k = i → j − 1 do // all possible cut points within the subsequence
Left = Ui[0] : Uk[1] Right = Uk+1[0] : Uj[1] L.append ((Left, Right))
end end
end
Both our intra- and multi-sentential parsing models are designed using MALLET’s graphical model toolkit GRMM (McCallum 2002). In order to avoid overfitting, we regularize the CRF models with l2 regularization and learn the model parameters using the limited-memory BFGS (L-BFGS) fitting algorithm.
4.1.4 Features Used in the Parsing Models. Crucial to parsing performance is the set of features used in the parsing models, as summarized in Table 1. We categorize the features into seven groups and specify which groups are used in what parsing model. Notice that some of the features are used in both models. Most of the features have been explored in previous studies (e.g., Soricut and Marcu 2003; Sporleder and Lapata 2005; Hernault et al. 2010). However, we improve some of these as explained subsequently.
The features are extracted from two adjacent discourse units Ut−1 and Ut. Organizational features encode useful information about text organization as shown by duVerle and Prendinger (2009). We measure the length of the discourse units as the number of EDUs and tokens in it. However, in order to better adjust to the length variations, rather than computing their absolute numbers in a unit, we choose to measure their relative numbers with respect to their total numbers in the two units. For example, if the two discourse units under consideration contain three EDUs in total, a unit containing two of the EDUs will have a relative EDU number of 0.67. We also measure the distances of the units in terms of the number of EDUs from the beginning and end of the sentence (or text in the multi-sentential case). Text structural features capture the correlation between text structure and rhetorical structure by counting the number of sentence and paragraph boundaries in the discourse units.
Discourse cues (e.g., because, but), when present, signal rhetorical relations between two text segments, and have been used as a primary source of information in earlier studies (Knott and Dale 1994; Marcu 2000a). However, recent studies (Hernault et al. 2010; Biran and Rambow 2011) suggest that an empirically acquired lexical N-gram dictionary is more effective than a fixed list of cue phrases, since this approach is domain independent and capable of capturing non-lexical cues such as punctuation.
In order to build a lexical N-gram dictionary empirically from the training corpus, we extract the first and last N tokens (N∈{1, 2, 3}) of each discourse unit and rank them according to their mutual information with the two labels, Structure (S) and Relation (R).
8 N-gram features N∈{1, 2, 3} Intra & Multi-Sentential
Beginning (or end) lexical N-grams in unit 1. Beginning (or end) lexical N-grams in unit 2. Beginning (or end) POS N-grams in unit 1. Beginning (or end) POS N-grams in unit 2.
5 Dominance set features Intra-Sentential
Syntactic labels of the head node and the attachment node. Lexical heads of the head node and the attachment node. Dominance relationship between the two units.
9 Lexical chain features Multi-Sentential
Number of chains spanning unit 1 and unit 2. Number of chains start in unit 1 and end in unit 2. Number of chains start (or end) in unit 1 (or in unit 2). Number of chains skipping both unit 1 and unit 2. Number of chains skipping unit 1 (or unit 2).
2 Contextual features Intra & Multi-Sentential
Previous and next feature vectors.
2 Sub-structural features Intra & Multi-Sentential
Root nodes of the left and right rhetorical sub-trees.
More specifically, given an N-gram x, we compute its conditional entropy H with respect to S and R as follows:8
H(S, R|x) = − ∑
s∈S r∈R log c(x, s, r)c(x) (7)
where c(x) is the empirical count of N-gram x, and c(x, s, r) is the joint empirical count of N-gram x with the labels s and r. This is in contrast to HILDA (Hernault et al. 2010), which ranks the N-grams by their frequencies in the training corpus. However, Blitzer
8 The higher the conditional entropy, the lower the mutual information, and vice versa.
(2008) found mutual information to be more effective than frequency as a method for feature selection. Intuitively, the most informative discourse cues are not only the most frequent, but also the ones that are indicative of the labels in the training data. In addition to the lexical N-grams we also encode the POS tags of the first and last N tokens (N∈{1, 2, 3}) in a discourse unit as shallow-syntactic features in our models.
Lexico-syntactic features dominance sets extracted from the Discourse Segmented Lexicalized Syntactic Tree (DS-LST) of a sentence have been shown to be extremely effective for intra-sentential discourse parsing in SPADE (Soricut and Marcu 2003). Figure 9a shows the DS-LST (i.e., lexicalized syntactic tree with EDUs identified) for a sentence with three EDUs from the RST–DT corpus, and Figure 9b shows the corresponding discourse tree. In a DS-LST, each EDU except the one with the root node must have a head node NH that is attached to an attachment node NA residing in a separate EDU. A dominance set D (shown at the bottom of Figure 9a) contains these attachment points (shown in boxes) of the EDUs in a DS-LST. In addition to the syntactic and lexical information of the head and attachment nodes, each element in the dominance set also includes a dominance relationship between the EDUs involved; the EDU with the attachment node dominates (represented by “>”) the EDU with the head node.
Soricut and Marcu (2003) hypothesize that the dominance set (i.e., lexical heads, syntactic labels, and dominance relationships) carries the most informative clues for intra-sentential parsing. For instance, the dominance relationship between the EDUs in our example sentence is 3 > 1 > 2, which favors the DT structure [1, 1, 2] over [2, 2, 3]. In order to extract dominance set features for two adjacent discourse units Ut−1 and Ut, containing EDUs ei:j and ej+1:k, respectively, we first compute the dominance set from the DS-LST of the sentence. We then extract the element from the set that holds across the EDUs j and j + 1. In our example, for the two units, containing EDUs e1 and e2, respectively, the relevant dominance set element is (1, efforts/NP)>(2, to/S). We encode the syntactic labels and lexical heads of NH and NA, and the dominance relationship as features in our intra-sentential parsing model.
Lexical chains (Morris and Hirst 1991) are sequences of semantically related words that can indicate topical boundaries in a text (Galley et al. 2003; Joty, Carenini, and Ng 2013). Features extracted from lexical chains are also shown to be useful for finding paragraph-level discourse structure (Sporleder and Lapata 2004). For example, consider the text with four paragraphs (P1 to P4) in Figure 10a. Now, let us assume that there is a lexical chain that spans the whole text, skipping paragraphs P2 and P3, while a second chain only spans P2 and P3. This situation makes it more likely that P2 and P3 should be linked in the DT before either of them is linked with another paragraph. Therefore, the DT structure in Figure 10b should be more likely than the structure in Figure 10c.
One challenge in computing lexical chains is that words can have multiple senses, and semantic relationships depend on the sense rather than the word itself. Several methods have been proposed to compute lexical chains (Barzilay and Elhadad 1997; Hirst and St. Onge 1997; Silber and McCoy 2002; Galley and McKeown 2003). We follow the state-of-the-art approach proposed by Galley and McKeown (2003), which extracts lexical chains after performing Word Sense Disambiguation (WSD).
In the preprocessing step, we extract the nouns from the document and lemmatize them using WordNet’s built-in morphy function (Fellbaum 1998). Then, by looking up in WordNet we expand each noun to all of its senses, and build a Lexical Semantic Relatedness Graph (LSRG) (Galley and McKeown 2003; Chali and Joty 2007). In an LSRG, the nodes represent noun-tokens with their candidate senses, and the weighted edges between senses of two different tokens represent one of the three semantic relations: repetition, synonym, and hypernym. For example, Figure 11a shows a partial LSRG, where the token bank has two possible senses, namely, money bank and river bank. Using the money bank sense, bank is connected with institution and company by hypernymy relations (edges marked with H), and with another bank by a repetition relation (edges marked with R). Similarly, using the river bank sense, it is connected with riverside by a hypernymy relation and with bank by a repetition relation. Nouns that are not found in WordNet are considered as proper nouns having only one sense, and are connected by only repetition relations.
We use this LSRG first to perform WSD, then to construct lexical chains. For WSD, the weights of all edges leaving the nodes under their different senses are summed up and the one with the highest score is considered to be the right sense for the wordtoken. For example, if repetition and synonymy are weighted equally, and hypernymy is given half as much weight as either of them, the score of bank’s two senses are: 1 + 0.5 + 0.5 = 2 for the sense money bank and 1 + 0.5 = 1.5 for the sense river bank. Therefore, the selected sense for bank in this context is river bank. In case of a tie, we select the sense that is most frequent (i.e., the first sense in WordNet). Note that this approach to WSD is different from that of Sporleder and Lapata (2004), which takes a greedy approach.
Finally, we prune the graph by only keeping the links that connect words with the selected senses. At the end of the process, we are left with the edges that form the actual lexical chains. For example, Figure 11b shows the result of pruning the graph in Figure 11a. The lexical chains extracted from the pruned graph are shown in the box at the bottom. Following Sporleder and Lapata (2004), for each chain element, we keep track of the location (i.e., sentence ID) in the text where that element was found, and exclude chains containing only one element. Given two discourse units, we count the number of chains that: hit the two units, exclusively hit the two units, skip both units, skip one of the units, start in a unit, and end in a unit.
We also consider more contextual information by including the above features computed for the neighboring adjacent discourse unit pairs in the current feature vector. For example, the contextual features for units Ut−1 and Ut include the feature vector computed from Ut−2 and Ut−1 and the feature vector computed from Ut and Ut+1.
We incorporate hierarchical dependencies between the constituents in a DT by rhetorical sub-structural features. For two adjacent units Ut−1 and Ut, we extract the roots of the two rhetorical sub-trees. For example, the root of the rhetorical sub-tree spanning over EDUs e1:2 in Figure 9b is Elaboration–NS. However, extraction of these features assumes the presence of labels for the sub-trees, which is not the case when we apply the parser to a new text (sentence or document) in order to build its DT in a nongreedy fashion. One way to deal with this is to loop twice through the parsing process using two different parsing models—one trained with the complete feature set, and the other trained without the sub-structural features. We first build an initial, sub-optimal DT using the parsing model that is trained without the sub-structural features. This intermediate DT will now provide labels for the sub-structures. Next we can build a final, more accurate DT by using the complete parsing model. This idea of twopass discourse parsing, where the second pass performs post-editing using additional features, has recently been adopted by Feng and Hirst (2014) in their greedy parser.
One could even continue doing post-editing multiple times until the DT converges. However, this could be very time consuming as each post-editing pass requires: (1) applying the parsing model to every possible unit sequence and computing the posterior marginals for all possible DT constituents, and (2) using the parsing algorithm to find the most probable DT. Recall from our earlier discussion in Section 4.1.3 that for n discourse units and M rhetorical relations, the first step requires O(M2n4) and O(M2n3) for intra- and multi-sentential parsing, respectively; we will see in the next section that the second step requires O(Mn3). In spite of the computational cost, the gain we attained in the subsequent passes was not significant for our development set. Therefore, we restrict our parser to only one-pass post-editing.
Note that in parsing models where the score (i.e., likelihood) of a parse tree decomposes across local factors (e.g., the CRF-based syntactic parser of Finkel, Kleeman, and Manning [2008]), it is possible to define a semiring using the factors and the local scores (e.g., given by the inside algorithm). The CKY algorithm could then give the optimal parse tree in a single post-editing pass (Smith 2011). However, because our intra-sentential parsing model is designed to capture sequential dependencies between DT constituents, the score of a DT does not directly decompose across factors over discourse productions. Therefore, designing such a semiring was not possible in our case.
In addition to these features, we also experimented with other features including WordNet-based lexical semantics, subjectivity, and TF.IDF-based cosine similarity. However, because such features did not improve parsing performance on our development set, they were excluded from our final set of features. The intra- and multi-sentential parsing models of CODRA assign a probability to every possible DT constituent in their respective parsing scenarios. The job of the parsing algorithm is then to find the k most probable DTs for a given text. We implement a probabilistic CKY-like bottom–up parsing algorithm that uses dynamic programming to compute the most likely parses (Jurafsky and Martin 2008). For simplicity, we first
describe the specific case of generating the single most probable DT, then we describe how to generalize this algorithm to produce the k most probable DTs for a given text.
Formally, the search problem for finding the most probable DT can be written as
DT∗ = argmax DT P(DT|Θ) (8)
where Θ specifies the parameters of the parsing model (intra- or multi-sentential). Given n discourse units, our parsing algorithm uses the upper-triangular portion of the n×n dynamic programming table D, where cell D[i, j] (for i < j) stores:
D[i, j] = P(r∗[Ui(0), Um∗ (1), Uj(1)]) (9)
where Ux(0) and Ux(1) are the start and end EDU Ids of discourse unit Ux, and
(m∗, r∗) = argmax i≤m<j ; R∈{1···M} P(R[Ui(0), Um(1), Uj(1)]) × D[i, m] × D[m + 1, j] (10)
Recall that the notation R[Ui(0), Um(1), Uj(1)] in this expression refers to a rhetorical relation R that holds between the discourse unit containing EDUs Ui(0) through Um(1) and the unit containing EDUs Um(1) + 1 through Uj(1).
In addition to D, which stores the probability of the most probable constituents of a DT, the algorithm also simultaneously maintains two other n×n dynamic programming tables S and R for storing the structure (i.e., Um∗ (1)) and the relations (i.e., r∗) of the corresponding DT constituents, respectively. For example, given four EDUs e1 · · · e4, the S and R dynamic programming tables at the left side in Figure 12 together represent the DT shown at the right. More specifically, to find the DT, we first look at the top-right entries in the two tables, and find S[1, 4] = 2 and R[1, 4] = r2, which specify that the two discourse units e1:2 and e3:4 should be connected by the relation r2 (the root in the DT). Then, we see how EDUs e1 and e2 should be connected by looking at the entries S[1, 2] and R[1, 2], and find S[1, 2] = 1 and R[1, 2] = r1, which indicates that these two units should be connected by the relation r1 (the left pre-terminal in the DT). Finally, to see how EDUs e3 and e4 should be linked, we look at the entries S[3, 4] and R[3, 4], which tell us that they should be linked by the relation r4 (the right pre-terminal). The algorithm
works in polynomial time. Specifically, for n discourse units and M number of relations, the time and space complexities are O(n3M) and O(n2), respectively.
A key advantage of using a probabilistic parsing algorithm like the one we use is that it allows us to generate a list of k most probable parse trees. It is straightforward to generalize the above algorithm to produce k most probable DTs. Specifically, when filling up the dynamic programming tables, rather than storing a single best parse for each sub-tree, we store and keep track (i.e., using back-pointers) of k-best candidates simultaneously. One can show that the time and space complexities of the k-best version of the algorithm are O(n3Mk2 log k) and O(k2n), respectively (Huang and Chiang 2005).
Note that, in contrast to other document-level discourse parsers (Marcu 2000b; Subba and Di-Eugenio 2009; Hernault et al. 2010; Feng and Hirst 2012, 2014), which use a greedy algorithm, CODRA finds a discourse tree that is globally optimal.9 This approach of CODRA is also different from the sentence-level discourse parser SPADE (Soricut and Marcu 2003). SPADE first finds the tree structure that is globally optimal, then it assigns the most probable relations to the internal nodes. More specifically, the cell D[i, j] in SPADE’s dynamic programming table stores
D[i, j] = P([Ui(0), Um∗ (1), Uj(1)]) (11)
where m∗ = argmax i≤m<j P([Ui(0), Um(1), Uj(1)]). Disregarding the relation label R while populating D, this approach may find a discourse tree that is not globally optimal. Now that we have presented our intra-sentential and multi-sentential parsing components, we are ready to describe how they can be effectively combined in a unified framework (Figure 2) to perform document-level rhetorical analysis. Recall that a key motivation for a two-stage10 parsing is that it allows us to capture the strong correlation between text structure and discourse structure in a scalable, modular, and flexible way. In the following, we describe two different approaches to model this correlation.
4.3.1 1S–1S (1 Sentence–1 Sub-tree). A key finding from previous studies on sentencelevel discourse analysis is that most sentences have a well-formed discourse sub-tree in the full document-level DT (Soricut and Marcu 2003; Fisher and Roark 2007). For example, Figure 13a shows 10 EDUs in three sentences (see boxes), where the DTs for the sentences obey their respective sentence boundaries.
Our first approach, called 1S–1S (1 Sentence–1 Sub-tree), aims to maximally exploit this finding. It first constructs a DT for every sentence using our intra-sentential parser, and then it provides our multi-sentential parser with the sentence-level DTs to build the rhetorical parse for the whole document.
4.3.2 Sliding Window. Although the assumption made by 1S–1S clearly simplifies the parsing process, it completely ignores the cases where rhetorical structures violate
9 We agree that with potentially sub-optimal, sub-structural features in the parsing model, CKY may end up finding a sub-optimal DT. But that is a separate issue.
10 Do not confuse the term two-stage with the term two-pass.
sentence boundaries. For example, in the DT shown in Figure 13b, sentence S2 does not have a well-formed sub-tree because some of its units attach to the left (i.e., 4–5 and 6) and some to the right (i.e., 7). Vliet and Redeker (2011) call these cases “leaky” boundaries.
Although we find fewer than 5% of the sentences in the RST–DT have leaky boundaries, in other corpora this can be true for a larger portion of the sentences. For example, we observe that over 12% of the sentences in the instructional corpus of Subba and Di-Eugenio (2009) have leaky boundaries. However, we notice that in most cases where DT structures violate sentence boundaries, its units are merged with the units of its adjacent sentences, as in Figure 13b. For example, this is true for 75% of the leaky cases in our development set containing 20 news articles from the RST–DT and for 79% of the leaky cases in our development set containing 20 how-to manuals from the instructional corpus. Based on this observation, we propose a sliding window approach.
In this approach, our intra-sentential parser works with a window of two consecutive sentences, and builds a DT for the two sentences. For example, given the three sentences in Figure 13, our intra-sentential parser constructs a DT for S1–S2 and a DT for S2–S3. In this process, each sentence in a document except the boundary sentences (i.e., the first and the last) will be associated with two DTs: one with the previous sentence (say, DTp) and one with the next (say, DTn). In other words, for each nonboundary sentence, we will have two decisions: one from DTp and one from DTn. Our parser consolidates the two decisions and generates one or more sub-trees for each sentence by checking the following three mutually exclusive conditions one after another:r Same in both: If the sentence under consideration has the same (in both
structure and labels) well-formed sub-tree in both DTp and DTn, we take this sub-tree. For example, in Figure 14a, S2 has the same sub-tree in the two DTs (one for S1–S2 and one for S2–S3). The two decisions agree on the DT for the sentence.r Different but no cross: If the sentence under consideration has a wellformed sub-tree in both DTp and DTn, but the two sub-trees vary either in structure or in labels, we pick the most probable one. For example, consider the DT for S1–S2 (at the left) in Figure 14a and the DT for S2–S3
in Figure 14b. In both cases S2 has a well-formed sub-tree, but they differ in structure. We pick the sub-tree which has the higher probability in the two dynamic programming tables.r Cross: If either or both of DTp and DTn segment the sentence into multiple sub-trees, we pick the one having more sub-trees. For example, consider the two DTs in Figure 14c. In the DT for S1–S2 on the left, S2 has three sub-trees (4–5, 6, 7), whereas in the DT for S2–S3 on the right, it has two (4–6, 7). So, we extract the three sub-trees for S2 from the first DT. If the sentence has the same number of sub-trees in both DTp and DTn, we pick the one with higher probability in the dynamic programming tables. Note that our choice of picking the DT with more sub-trees is intended to allow the parser to find more leaky cases. However, other heuristics are also possible. For example, another simple heuristic that one could try is: When both DTs segment the sentence into multiple sub-trees, pick the one with fewer sub-trees, and when only one of the DTs segment the sentence into multiple sub-trees, pick that one.
At the end, the multi-sentential parser takes all these sentence-level sub-trees for a document, and builds a full rhetorical parse for the whole document. 5 the discourse segmenter :Our discourse parser assumes that the input text has been already segmented into a sequence of EDUs. However, discourse segmentation is also a challenging problem, and previous studies (Soricut and Marcu 2003; Fisher and Roark 2007) have identified
it as a primary source of inaccuracy for discourse parsing. Regardless of its importance in discourse parsing, discourse segmentation itself can be useful in several NLP applications, including sentence compression (Sporleder and Lapata 2005) and textual alignment in statistical machine translation (Stede 2011). Therefore, in CODRA, we have developed our own discourse segmenter, which not only achieves state-of-theart performance as shown later, but also reduces the time complexity by using fewer features. The discourse segmenter in CODRA implements a binary classifier to decide for each word–token (except the last) in a sentence, whether to place an EDU boundary after that token. We use a maximum entropy model to build a discriminative classifier. More specifically, we use a Logistic Regression classifier with parameter Θ:
P(y|w,Θ) = Ber(y| Sigm(ΘTx)) (12)
where the output y ∈ {0, 1} denotes whether to put an EDU boundary (i.e., y = 1) or not (i.e., y = 0) after the word–token w, which is represented by a feature vector x. In the equation, Ber(η) and Sigm(η) refer to the Bernoulli distribution and the Sigmoid (also known as logistic) function, respectively. The negative log-likelihood (NLL) of the model with l2 regularization for N data points (i.e., word–tokens) is given by
NLL(θ) = − N∑
i=1
yi log Sigm(ΘTxi) + (1 − yi) log (1 − Sigm(ΘTxi)) + λΘTΘ (13)
where yi is the gold label for word–token wi (represented by feature vector xi). We learn the model parameters Θ using the L-BFGS fitting algorithm, which is time- and spaceefficient. To avoid overfitting, we use 5-fold cross validation to learn the regularization strength parameter λ from the training data. We also use a simple bagging technique (Breiman 1996) to deal with the sparsity of boundary (i.e., y = 1) tags.
Note that our first attempt at the discourse segmentation task implemented a linearchain CRF model (Lafferty, McCallum, and Pereira 2001) to capture the sequential dependencies between the tags in a discriminative way. However, the binary Logistic Regression classifier, using the same set of features, not only outperforms the CRF model, but also reduces time and space complexity. One possible explanation for the low performance of the CRF model is that Markov dependencies between tags cannot be effectively captured due to the sparsity of boundary tags. Also, because we could not balance the data by using techniques like bagging in the CRF model, this further degrades the performance. Our set of features for discourse segmentation are mostly inspired from previous studies but used in a novel way, as we describe here.
Our first subset of features, which we call SPADE features, includes the lexicosyntactic patterns extracted from the lexicalized syntactic tree of the given sentence.
These features replicate the features used in SPADE’s segmenter, but used in a discriminative way. In order to decide on an EDU boundary after a word–token wk, we search for the lowest constituent in the lexicalized syntactic tree that spans over tokens wi . . .wj such that i ≤ k < j. The production that expands this constituent in the tree, with the potential EDU boundary marked, forms the primary feature. For instance, to determine the existence of an EDU boundary after the word efforts in our sample sentence shown in Figure 9, the production NP(efforts) → PRP$(its) NNS(efforts) ↑ S(to) extracted from the lexicalized syntactic tree in Figure 9a constitutes the primary feature, where ↑ denotes the potential EDU boundary.
SPADE predicts an EDU boundary if the relative frequency (i.e., maximum likelihood estimate) of a potential boundary given the production in the training data is greater than 0.5. If the production has not been observed frequently enough, the unlexicalized version of the production (e.g., NP → PRP$ NNS ↑ S) is used for prediction. If the unlexicalized version is also found to be rare, other variations of the production, depending on whether they include the lexical heads and how many non-terminals (one or two) they consider before and after the potential boundary, are examined one after another (see Fisher and Roark [2007] for details). In contrast, we compute the maximum likelihood estimates for a primary production (feature) and its other variations, and use those directly as features with/without binarizing the values.
Shallow syntactic features like Chunk and POS tags have been shown to possess valuable clues for discourse segmentation (Fisher and Roark 2007; Sporleder and Lapata 2005). For example, it is less likely that an EDU boundary occurs within a chunk. We annotate the tokens of a sentence with chunk and POS tags using the state-of-the-art Illinois tagger11 and encode these as features in our model. Note that the chunker assigns each token a tag using the BIO notation, where B stands for beginning of a particular phrase (e.g., noun or verb phrase), I stands for inside of a particular phrase, and O stands for outside of a particular phrase. The rationale for using the Illinois chunker is that it uses a larger set of tags (23 in total); thus it is more informative than most of the other existing taggers, which typically use only five tags (B–NP, I–NP, B–VP, I–VP, and O).
EDUs are normally multi-word strings. Thus, a token near the beginning or end of a sentence is unlikely to be the end of a segment. Therefore, for each token we include its relative position (i.e., absolute position/total number of tokens) in the sentence and distances to the beginning and end of the sentence as features.
It is unlikely that two consecutive tokens are tagged with EDU boundaries. Therefore, we incorporate contextual information for a token into our model by including the above features computed for its neighboring tokens.
We also experimented with different N-gram (N ∈ {1, 2, 3}) features extracted from the token sequence, POS sequence, and chunk sequence. However, because such features did not improve segmentation accuracy on the development set, they were excluded from our final set of features. 6 experiments :In this section we present our experimental results. First, we describe the corpora on which the experiments were performed and the evaluation metrics used to measure the
11 Available at http://cogcomp.cs.illinois.edu/page/software.
performance of the discourse segmenter and the parser. Then we show the performance of our discourse segmenter, followed by the performance of our discourse parser. Whereas previous studies on discourse analysis only report their results on a particular corpus, to demonstrate the generality of our method, we experiment with texts from two very different genres: news articles and instructional how-to manuals.
Our first corpus is the standard RST–DT (Carlson, Marcu, and Okurowski 2002), which contains discourse annotations for 385 Wall Street Journal news articles taken from the Penn Treebank corpus (Marcus, Marcinkiewicz, and Santorini 1994). The corpus is partitioned into a training set of 347 documents and a test set of 38 documents. A total of 53 documents selected from both training and test sets were annotated by two human annotators. We measure human agreements based on this doubly annotated data set. We used 25 documents from the training set as our development set. In RST–DT, the original 25 rhetorical relations defined by Mann and Thompson (1988) are further divided into a set of 18 coarser relation classes with 78 finer-grained relations (see Carlson and Marcu [2001] for details). Our second corpus is the instructional corpus prepared by Subba and Di-Eugenio (2009), which contains discourse annotations for 176 how-to manuals on home repair. The corpus was annotated with 26 informational relations (e.g., Preparation–Act, Act–Goal).
For our experiments with the intra-sentential discourse parser, we extracted a sentence-level DT from a document-level DT by finding the sub-tree that exactly spans over the sentence. In RST–DT, by our count, 7, 321 out of 7, 673 sentences in the training set, 951 out of 991 sentences in the test set, and 1, 114 out of 1, 208 sentences in the doubly-annotated set have a well-formed DT. On the other hand, 3, 032 out of 3, 430 sentences in the instructional corpus have a well-formed DT. This forms the corpora for our experiments with intra-sentential discourse parsing. However, the existence of a well-formed DT is not a necessity for discourse segmentation; therefore, we do not exclude any sentence in our discourse segmentation experiments. In this subsection we describe the metrics used to measure both how much the annotators agree with each other, and how well the systems perform when their outputs are compared with human annotations for the discourse analysis tasks.
6.2.1 Metrics for Discourse Segmentation. Because sentence boundaries are considered to also be the EDU boundaries, we measure segmentation accuracy with respect to the intra-sentential segment boundaries, which is a standard method (Soricut and Marcu 2003; Fisher and Roark 2007). Specifically, if a sentence contains n EDUs, which corresponds to n − 1 intra-sentential segment boundaries, we measure the segmenter’s ability to correctly identify these n − 1 boundaries. Let h be the total number of intrasentential segment boundaries in the human annotation, m be the total number of intrasentential segment boundaries in the model output, and c be the total number of correct segment boundaries in the model output. Then, we measure Precision (P), Recall (R), and F-score for segmentation performance as follows:
P = cm, R = c h , and F−score = 2PR P + R = 2c h + m (14)
6.2.2 Metrics for Discourse Parsing. To evaluate parsing performance, we use the standard unlabeled and labeled precision, recall, and F-score as proposed by Marcu (2000b). The unlabeled metric measures how accurate the discourse parser is in finding the right structure (i.e., the skeleton) of the DT, while the labeled metrics measure the parser’s ability to find the right labels (i.e., nuclearity statuses or relation labels) in addition to the right structure. Assume, for example, that given the two sentences of Figure 1, our system generates the DT shown in Figure 15. In Figure 16, we show the same gold DT shown in Figure 1 (on the left), and the same system-generated DT shown in Figure 15 (on the right), when the two trees are aligned. For the sake of illustration, instead of showing the real EDUs, we only show their IDs. Notice that the automatic segmenter made two mistakes: (1) it broke the EDU marked 2–3 (Some people use the purchasers’ index as a leading indicator) in the human annotation into two separate EDUs, and (2) it could not identify EDU 5 (But the thing it’s supposed to measure) and EDU 6 (— manufacturing strength —) as two separate EDUs. Therefore, when we align the two annotations, we obtain seven EDUs in total.
In Table 2, we list all constituents of the two DTs and their associated labels at the span, nuclei, and relation levels. The recall (R) and precision (P) figures are shown at the bottom of the table. Notice that, following (Marcu 2000b), the relation labels are assigned to the children nodes rather than to the parent nodes in the evaluation process to deal with non-binary trees in human annotations. To our knowledge, no implementation of
the evaluation metrics was made publicly available. Therefore, to help other researchers, we have made our source code of the evaluation metrics publicly available.12
Given this evaluation setup, it is easy to understand that if the number of EDUs is the same in the human and system annotations (e.g., when the discourse parser uses gold discourse segmentation), and the discourse trees are binary, then we get the same figures for precision, recall, and F-score. In this section we present our experiments on discourse segmentation.
6.3.1 Experimental Set-up for Discourse Segmentation. We compare the performance of our discourse segmenter with the performance of the two publicly available discourse segmenters, namely, the discourse segmenters of the HILDA (Hernault et al. 2010) and SPADE (Soricut and Marcu 2003) systems. We also compare our results with the stateof-the-art results reported by Fisher and Roark (2007) on the RST–DT test set. In all our experiments when comparing two systems, we use paired t-test on the F-scores to measure statistical significance and report the p-value.
We ran HILDA with its default settings. For SPADE, we applied the same modifications to its default settings as described in Fisher and Roark (2007), which delivers significantly improved performance over its original version. Specifically, in our experiments on the RST–DT corpus, we trained SPADE using the human-annotated syntactic trees extracted from the Penn Treebank (Marcus, Marcinkiewicz, and Santorini 1994), and, during testing, we replaced the Charniak parser (Charniak 2000) with a more
12 Available from alt.qcri.org/tools/.
accurate reranking parser (Charniak and Johnson 2005). However, because of the lack of gold syntactic trees in the instructional corpus, we trained SPADE in this corpus using the syntactic trees produced by the reranking parser. To avoid using the gold syntactic trees, we used the reranking parser in our system for both training and testing purposes. This syntactic parser was trained on the sections of the Penn Treebank not included in our test set. We applied the same canonical lexical head projection rules (Magerman 1995; Collins 2003) to lexicalize the syntactic trees as done in HILDA and SPADE.
Note that previous studies (Fisher and Roark 2007; Soricut and Marcu 2003; Hernault et al. 2010) on discourse segmentation only report their performance on the RST–DT test set. To compare our results with them, we evaluated our model on the RST–DT test set. In addition, we showed a more general performance of SPADE and our system on the two corpora based on 10-fold cross validation.13 However, SPADE does not come with a training module for its segmenter. We reimplemented this module and verified its correctness by reproducing the results on the RST–DT test set.
6.3.2 Results for Discourse Segmentation. Table 3 shows the discourse segmentation results of different systems in Precision, Recall, and F-score on the two corpora. On the RST– DT corpus, HILDA’s segmenter delivers the weakest performance, having an F-score of only 74.1. Note that the high segmentation accuracy reported by Hernault et al. (2010) is due to a less stringent evaluation metric. SPADE performs much better than HILDA with an absolute F-score improvement of 11.1%. Our segmenter DS outperforms SPADE with an absolute F-score improvement of 4.9% (p-value < 2.4e-06), and also achieves comparable results to the ones of Fisher and Roark (2007) (F&R), even though we use fewer features.14 Notice that human agreement for this task is quite high—namely, an F-score of 98.3 computed on the doubly-annotated portion of the RST–DT corpus mentioned in Section 6.1.
Because Fisher and Roark (2007) only report their results on the RST–DT test set and we did not have access to their system, we compare our approach only with SPADE when evaluating on a whole corpus based on 10-fold cross validation. On the RST–DT corpus, our segmenter delivers an absolute F-score improvement of 3.8 percentage points, which represents a more than 25% relative error rate reduction.
13 Because the two tasks—discourse segmentation and intra-sentential parsing—operate at the sentence level, the cross validation was performed over sentences for their evaluation. 14 Because we did not have access to the system or to the complete output/results of Fisher and Roark (2007), we were not able to perform a statistical significance test.
The improvement is higher on the instructional corpus with an absolute F-score improvement of 8.1 percentage points, which corresponds to a relative error reduction of 30%. The improvements for both corpora are statistically significant (p-value < 3.0e-06). When we compare our results on the two corpora, we observe a substantial decrease in performance on the instructional corpus. This could be because of a smaller amount of data in this corpus and/or to the inaccuracies in the syntactic parser and taggers, which are trained on news articles. A promising future direction would be to apply effective domain adaptation methods (e.g., easyadapt [Daumé 2007]) to improve discourse segmentation performance in the instructional domain by leveraging the rich data in the news domain (i.e., RST–DT). In this section we present our experiments on discourse parsing. First, we describe the experimental set-up. Then, we present the results of the parsers. While presenting the performance of our discourse parser, we show a breakdown of intra-sentential versus inter-sentential results, in addition to the aggregated results at the document level.
6.4.1 Experimental Set-up for Discourse Parsing. In our experiments on sentence-level (i.e., intra-sentential) discourse parsing, we compare our approach with SPADE (Soricut and Marcu 2003) on the RST–DT corpus, and with the ILP-based approach of Subba and Di-Eugenio (2009) on the instructional corpus, because they are the state of the art in their respective genres. For SPADE, we applied the same modifications to its default settings as described in Section 6.3.1, which leads to improved performance. Similarly, in our experiments on document-level (i.e., multi-sentential) parsing, we compare our approach with HILDA (Hernault et al. 2010) on the RST–DT corpus, and with the ILPbased approach (Subba and Di-Eugenio 2009) on the instructional corpus. The results for HILDA were obtained by running the system with default settings on the same inputs we provided to our system. Because we could not run the ILP-based system (not publicly available), we report the performance presented in their paper.
Our experiments on the RST–DT corpus use the same 18 coarser coherence relations (see Figure 18 later in this article), defined by Carlson and Marcu (2001) and also used in SPADE and HILDA systems. More specifically, the relation set consists of 16 relation categories and two pseudo-relations, namely, Textual–Organization and Same–Unit. After attaching the nuclearity statuses (NS, SN, NN) to these relations, we obtain 41 distinct relations.15 Our experiments on the instructional corpus consider the same 26 primary relations (e.g., Goal:Act, Cause:Effect) used by Subba and Di-Eugenio (2009) and also treat the reversals of non-commutative relations as separate relations. That is, Goal–Act and Act–Goal are considered to be two different coherence relations. Attaching the nuclearity statuses to these relations provides 76 distinct relations.
Based on our experiments on the development set, the size of the automatically built bi-gram and tri-gram dictionaries was set to 95% of their total number of items, and the size of the unigram dictionary was set to 100%. Note that the unigram dictionary contains only special tags denoting EDU, sentence, and paragraph boundaries.
15 Not all relations take all the possible nuclearity statuses. For example, Elaboration and Attribution are mono-nuclear relations, and Same–Unit and Joint are multi-nuclear relations.
6.4.2 Evaluation of the Intra-Sentential Discourse Parser. This section presents our experimental evaluation on intra-sentential discourse parsing. First, we show the performance of the sentence-level parsers when they are provided with manual (or gold) discourse segmentations. This allows us to judge the parsing performance independently of the segmentation task. Then, we show the end-to-end performance of our intra-sentential framework, that is, the intra-sentential parsing performance based on automatic discourse segmentation.
Intra-sentential parsing results based on manual segmentation Table 4 presents the intra-sentential discourse parsing results when manual discourse segmentation is used. Recall from our discussion on evaluation metrics in Section 6.2.2 that precision, recall, and F-score are the same when manual segmentation is used. Therefore, we report only one of them. Notice that our sentence-level discourse parser PAR-S consistently outperforms SPADE on the RST–DT test set in all three metrics, and the improvements are statistically significant (p-value< 0.01). Especially, on the relation labeling task, which is the hardest among the three tasks, we achieve an absolute F-score improvement of 12.2 percentage points, which represents a relative error rate reduction of 37.7%.
To verify our claim that capturing the sequential dependencies between DT constituents using a DCRF model actually contributes to the performance gain, we also compare our results with an intra-sentential parser (see CRF-NC in Table 4) that uses a simplified CRF model similar to the one shown in Figure 8. Although the simplified model has two hidden variables to model the structure and the relation of a DT constituent jointly, it does not have a chain structure, thus it ignores the sequential dependencies between DT constituents. The comparison in all three measures demonstrates that the improvements are indeed partly due to the DCRF model (p–value < 0.01).16 A comparison between CRF-NC and SPADE shows that CRF-NC significantly outperforms SPADE in all three measures (p-value < 0.01). This could be due to the fact that CRF-NC is trained discriminatively with a large number of features, whereas SPADE is trained generatively with only lexico-syntactic features.
Notice that the scores of our parser (PAR-S) are close to the human agreement on the doubly-annotated data, and these results on the RST–DT test set are also consistent with the mean scores over 10-folds on the whole RST–DT corpus.17
The improvements are higher on the instructional corpus, where we compare our mean results over 10-folds with the reported results of the ILP-based system of Subba and Di-Eugenio (2009), giving absolute F-score improvements of 5.4 percentage points, 17.6 percentage points, and 12.8 percentage points in span, nuclearity, and relations, respectively.18 Our parser PAR-S reduces the errors by 76.1%, 62.4%, and 34.6% in span, nuclearity and relations, respectively.
If we compare the performance of our intra-sentential discourse parser on the two corpora, we notice that our parser PAR-S is more accurate in finding the right tree
16 The parsing performance reported in Table 4 for CRF-NC is when the CRF parsing model is trained on a balanced data set (an equal number of instances with S=1 and S=0); Training on full but imbalanced data set gives slightly lower results. 17 Our EMNLP and ACL publications (Joty, Carenini, and Ng 2012; Joty et al. 2013) reported slightly lower parsing accuracies. Fixing a bug in the parsing algorithm accounts for the difference. 18 Subba and Di-Eugenio (2009) report their results based on an arbitrary split between training and test sets. Because we did not have access to their particular split, we compare our model’s performance based on 10-fold cross validation with their reported results. Also, because we did not have access to their system/output, we could not perform a significance test on the instructional corpus.
structure (see Span row in the table) on the instructional corpus. This may be due to the fact that sentences in the instructional domain are relatively short and contain fewer EDUs than sentences in the news domain, thus making it easier to find the right tree structure. However, when we compare the performance on the relation labeling task, we observe a decrease on the instructional corpus. This may be due to the small amount of data available for training and the imbalanced distribution of a large number of discourse relations (i.e., 76 with nuclearity attached) in this corpus.
Intra-sentential parsing results based on automatic segmentation In order to evaluate the performance of the fully automatic sentence-level discourse analysis systems, we feed the intra-sentential discourse parsers the output of their respective discourse segmenters. Table 5 shows the (P)recision, (R)ecall, and (F)–score results for different evaluation metrics. We compare our intra-sentential parser PAR-S with SPADE on the RST–DT test set. We achieve absolute F-score improvements of 5.7 percentage points, 6.4 percentage points, and 9.5 percentage points in span, nuclearity, and relation, respectively. These improvements are statistically significant (pvalue<0.001). Our system, therefore, reduces the errors by 24.5%, 21.4%, and 22.6% in span, nuclearity, and relations, respectively. These results are also consistent with the mean results over 10-folds on the whole RST–DT corpus.
The rightmost column in the table shows our mean results over 10-folds on the instructional corpus. We could not compare our system with the ILP-based approach of Subba and Di-Eugenio (2009) because no results were reported using an automatic segmenter. It is interesting to observe how much our parser is affected by an automatic segmenter on the two corpora (see Tables 4 and 5). Nevertheless, taking into account the
segmentation results in Table 3, this is not surprising because previous studies (Soricut and Marcu 2003) have already shown that automatic segmentation is the primary impediment to high accuracy discourse parsing. This demonstrates the need for a more accurate discourse segmentation model in the instructional genre.
6.4.3 Evaluation of the Complete Parser. We experiment with our full document-level discourse parser on the two corpora using the two parsing approaches described in Section 4.3, namely, 1S-1S and the sliding window. On RST–DT, the standard split was used for training and testing. On the instructional corpus, Subba and Di-Eugenio (2009) used 151 documents for training and 25 documents for testing. Because we did not have access to their particular split, we took five random samples of 151 documents for training and 25 documents for testing, and report the average performance over the five test sets.
Table 6 presents results for our two-stage discourse parser (TSP) using approaches 1S-1S (TSP 1-1) and the sliding window (TSP SW) on manually segmented texts. Recall that precision, recall, and F-score are the same when manual segmentation is used. We compare our parser with the state-of-the-art on the two corpora: HILDA (Hernault et al. 2010) on RST–DT, and the ILP-based approach (Subba and Di-Eugenio 2009) on the instructional domain. On both corpora, our systems outperform existing systems by a wide margin (p-value <7.1e-05 on RST–DT).19 On RST–DT, our parser TSP 1-1 achieves absolute improvements of 7.9 percentage points, 9.3 percentage points, and 11.5 percentage points in span, nuclearity, and relation, respectively, over HILDA. This represents relative error reductions of 31.2%, 22.7%, and 20.7% in span, nuclearity, and relation, respectively.
Beside HILDA, we also compare our results with two baseline parsers on RST–DT: (1) CRF-O, which uses a single unified CRF-based parsing model shown in Figure 8 (the one used for multi-sentential parsing) without distinguishing between intra- and multisentential parsing, and (2) CRF-T, which uses two different CRF-based parsing models for intra- and multi-sentential parsing in the two-stage approach 1S-1S, both models having the same structure as in Figure 8. Thus, CRF-T is a variation of TSP 1-1, where the DCRF-based (chain-structured) intra-sentential parsing model is replaced with a simpler CRF-based parsing model.20 Note that although CRF-O does not explicitly
19 Because we did not have access to the output or to the system of Subba and Di-Eugenio (2009), we were not able to perform a significance test on the instructional corpus. 20 The performance of this model for intra-sentential parsing is reported in Table 4 under the name CRF-NC.
discriminate between intra- and multi-sentential parsing, it uses N-gram features that include sentence and EDU boundaries to encode this information into the model.
Table 6 shows that both CRF-O and CRF-T outperform HILDA by a good margin (p-value <0.0001). This improvement can be attributed to the optimal parsing algorithm and better feature selection strategy. When we compare CRF-T with CRF-O, we notice significant performance gains for CRF-T (p-value <0.001). The absolute gains are 4.32 percentage points, 2.68 percentage points, and 4.55 percentage points in span, nuclearity, and relation, respectively. This comparison clearly demonstrates the benefit of using a two-stage approach with two different parsing models over a framework with one single unified parsing model. Finally, when we compare our best results with the human agreements, we still observe room for further improvement in all three measures.
On the instructional genre, our parser TSP 1-1 delivers absolute F-score improvements of 10.3 percentage points, 13.6 percentage points, and 8.1 percentage points in span, nuclearity, and relations, respectively, over the ILP-based approach of Subba and Di-Eugenio (2009). Our parser, therefore, reduces errors by 34.7%, 26.9%, and 12.5% in span, nuclearity, and relations, respectively.
If we compare the performance of our discourse parsers on the two corpora, we observe lower results on the instructional corpus. There could be two reasons for this. First, the instructional corpus has a smaller amount of data with a larger set of relations (76 with nuclearity attached). Second, some of the frequent relations are semantically very similar (e.g., Preparation-Act, Step1-Step2), which makes it difficult even for the human annotators to distinguish them (Subba and Di-Eugenio 2009).
Comparison between our two document-level parsing approaches reveals that TSP SW significantly outperforms TSP 1-1 only in finding the right structure on both corpora (p-value <0.01). Not surprisingly, the improvement is higher on the instructional corpus. A likely explanation is that the instructional corpus contains more leaky boundaries (12%), allowing the sliding window approach to be more effective in finding those, without inducing much noise for the labels. This demonstrates the potential of TSP SW for data sets with even more leaky boundaries, e.g., the Dutch (Vliet and Redeker 2011) and the German Potsdam (Stede 2004) corpora. However, it would be interesting to see how other heuristics to do consolidation in the cross condition (Section 4.3.2) perform.
To analyze errors made by TSP SW, we looked at some poorly parsed examples and found that although TSP SW finds more correct structures, a corresponding improvement in labeling relations is not present because in some cases, it tends to induce noise from the neighboring sentences for the labels. For example, when parsing is performed on the first sentence in Figure 1 in isolation using 1S-1S, our parser rightly identifies the Contrast relation between EDUs 2 and 3. But, when it is considered with its neighboring sentences by the sliding window, the parser labels it as Elaboration. A promising strategy to deal with this and similar problems would be to apply both approaches to each sentence and combine them by consolidating three probabilistic decisions, namely, the one from 1S-1S and the two from the sliding window.
6.4.4 k-best Parsing Results Based on Manual Segmentation. As described in Section 4.2, a straight-forward modification of our probabilistic parsing algorithm allows us to generate a list of k-best parse hypotheses for a given text. We adapt our parsing algorithm accordingly to produce k most probable DTs for each text, and measure the oracle accuracy based on the F-scores of the Relation metric which gives aggregated evaluation
on structure and relation labels (see Table 2). Specifically, the oracle accuracy O−score for k-best discourse parsing is measured as follows:
O−score = ∑N i=1 max k j=1 F−scorer(gi, h j i)
N (15)
where N is the total number of texts (sentences or documents) evaluated, gi is the gold DT annotation for text i, hji is the j
th parse hypothesis generated by the parser for text i, and F-scorer(gi, h j i) is the F-score accuracy of hypothesis h j i on the Relation metric, which essentially measures how similar hji is to gi in terms of its structure and labels. Table 7 presents the oracle scores of our intra-sentential parser PAR-S on the RST– DT test set as a function of k of k-best parsing. The 1-best result tells that the parser has the base accuracy of 79.8%. The 2-best shows dramatic oracle-rate improvements (i.e., 4.65% absolute), meaning that often our parser generates the best tree as its top two outputs. 3-best and 4-best also show moderate improvements (about 2%). Things start to slow down afterwards, and we achieve oracle rates of 90.37% and 92.57% at 10−best and 20-best, respectively. The 30-best parsing gives an oracle score of 93.2%.
The results of our k-best intra-sentential discourse parser demonstrate that a k-best reranking approach like that of Collins and Koo (2005) and Charniak and Johnson (2005) used for syntactic parsing can potentially improve the parsing accuracy even further by exploiting additional global features of the candidate discourse trees as evidence.
The scenario is quite different at the document-level; Table 8 shows the k-best parsing results of TSP 1S-1S on the RST–DT test set. The improvements in oracle-rate are small at the document-level when compared with the sentence-level parsing. For example, the 2-best and the 5-best improve over the base accuracy by only 0.7 percentage points and 1.0 percentage points, respectively. The improvements get even slower after that. However, this is not surprising because generally document-level DTs are big with many constituents, and only a very few of these constituents change from k-best to k + 1-best parsing. These small changes among the candidate DTs do not contribute much to the overall F-score accuracy (for further clarification see how F-score is calculated in Section 6.2.2).
The results of our k-best document-level parsing suggest that often the best tree is missing in the top k parses. Thus, a reranking of k-best document-level parses may not be a suitable option for further improvement at the document-level. An alternative to k-best reranking is to use a sampling-based parsing strategy (Wick et al. 2011) to explore the space of possible trees, as recently used for dependency parsing (Zhang et al. 2014). However, note that the potential gain we may obtain by using a reranker at the sentence level will also improve the (combined) accuracy of the document-level parser.
6.4.5 Analysis of Features. To analyze the relative importance of different features used in our parsing models, Table 9 presents the sentence- and document-level parsing results on a manually segmented RST–DT test set using different subsets of features. The feature subsets were defined in Section 4.1.4. In each parsing condition, the subsets of features are added incrementally, based on their availability and historical importance. The columns in Table 9 represent the inclusion order of the feature subsets.
Because SPADE (Soricut and Marcu 2003) achieved the previous best results on intra-sentential parsing using Dominance set features, these are included as the initial set of features in our intra-sentential parsing model. In HILDA, Hernault et al. (2010) demonstrate the importance of Organizational and N-gram features for full text parsing. We add these two feature subsets one after another in our intra- and multi-sentential parsing models.21 Contextual features require other features to be computed; thus they were added after those features. Because computation of Sub-structural features requires an initial parse tree (i.e., when the parser is applied), they are added at the very end.
Notice that inclusion of every new subset of features appears to improve the performance over the previous set. Specifically, for sentence-level parsing, when we add the Organizational features with the Dominance set features, we achieve about 2 percentage points absolute improvements in nuclearity and relations. With N-gram features, the gain is even higher: 6 percentage points in relations and 3.5 percentage points in nuclearity for sentence-level parsing, and 3.8 percentage points in relations and 3.1 percentage points in nuclearity for document-level parsing. This demonstrates the utility of the N-gram features, which is also consistent with the previous findings of duVerle and Prendinger (2009) and Schilder (2002).
The features extracted from Lexical chains (L-ch) have also proved to be useful for document-level parsing. They deliver absolute improvements of 2.7 percentage points, 2.9 percentage points, and 2.3 percentage points in span, nuclearity, and relations,
21 Text structural features are included in the Organizational features for multi-sentential parsing.
respectively. Including the Contextual features further gives improvements of 3 percentage points in nuclearity and 2.2 percentage points in relation for sentence-level parsing, and 1.3 percentage points in nuclearity, and 1.2 percentage points in relation for document-level parsing. Notice that Sub-structural features are more beneficial for document-level parsing than they are for sentence-level parsing, that is, an improvement of 2.2 percentage points versus an improvement of 0.9 percentage points. This is not surprising because document-level DTs are generally much larger than sentencelevel DTs, making the sub-structural features more effective for document-level parsing.
6.4.6 Error Analysis. We further analyze the errors made by our discourse parser. As described in previous sections, the parser could be wrong in finding the right structure as well as the right nuclearity and relation labels. Figure 17 presents an example where our parser makes mistakes in finding the right structure (notice the units connected by Attribution and Cause in the two example DTs) and the right relation label (TopicComment vs. Background). The comparison between intra- and multi-sentential parsing results presented in Sections 6.4.2 and 6.4.3 tells us that the errors in structure occur more frequently when the DT is large (e.g., at the document level) and the parsing model fails to capture the long-range structural dependencies between the DT constituents.
To further analyze the errors made by our parser on the hardest task of relation labeling, in Figure 18 we present the confusion matrix for our document-level parser TSP 1-1 on the RST–DT test set. In order to judge independently the ability of our parser to assign the correct relation labels, the confusion matrix is computed based on the constituents (see Table 2), where our parser found the right span (i.e., structure).22 The relations in the matrix are ordered according to their frequency in the training set.
In general, the errors can be explained by two different causes acting together: (1) imbalanced distribution of the relations in the corpus, and (2) semantic similarity between the relations. The most frequent relation Elaboration tends to overshadow others, especially the ones that are semantically similar (e.g., Explanation, Background)
22 Therefore, the counts of the relations shown in the table may not match the ones in the test set.
T-C T-O T-CM M-M CMP EV SU CND EN CA TE EX BA CO JO S-U AT EL
and less frequent (e.g., Summary, Evaluation). Furthermore, our models sometimes fail to distinguish relations that are semantically similar (e.g., Temporal vs. Background, Cause vs. Explanation).
Now, let us look more closely at a few of these errors. Figure 19 presents an example where our parser mistakenly labels a Summary as Elaboration. Clearly, in this example the text in parentheses (i.e., (CFD)) is an acronym or summary of the text to the left. However, parenthesized texts are also used to provide additional information (i.e., to elaborate), as exemplified in Figure 20 by two text snippets from the RST-DT. Notice that although the structure of the text (widow of the ..) in the first example is quite distinguishable from the structure of (CFD), the text (D., Maine) in the second example is similar to (CFD) in structure, thus it confuses our model.23
Figure 21 presents two examples where our parser mistakenly labels Background and Cause as Elaboration. However, notice that the two discourse relations (i.e., Background
23 D., Maine in this example refers to Democrat from state Maine.
vs. Elaboration and Cause vs. Elaboration) in these examples are semantically very close, and arguably both can be applicable.
Given these observations, we see two possible ways to improve our system. First, we would like to use a more robust method (e.g., ensemble methods with bagging) to deal with the imbalanced distribution of relations, along with taking advantage of richer semantic knowledge (e.g., compositional semantics) to cope with the errors caused by semantic similarity between the relations. Second, to capture long-range dependencies between DT constituents, we would like to explore the idea of k-best discriminative reranking using tree kernels (Dinarelli, Moschitti, and Riccardi 2011). Because our parser already produces k most probable DTs, developing a reranker based on discourse tree kernels is very much within our reach. 7 conclusions and future directions :In this article we have presented CODRA, a complete probabilistic discriminative framework for performing rhetorical analysis in the RST framework. CODRA comprises components for performing both discourse segmentation and discourse parsing. The discourse segmenter is a binary classifier based on a maximum entropy model, and the discourse parser applies an optimal parsing algorithm to probabilities inferred from two CRF models: one for intra-sentential parsing and the other for multi-sentential parsing. The CRF models effectively represent the structure and the label of discourse tree constituents jointly. Furthermore, the DCRF model for intra-sentential parsing captures the sequential dependencies between the constituents. The two separate parsing models use their own informative feature sets and the distributional variations of the relation labels in their respective parsing conditions.
We have also presented two approaches to effectively combine the intra-sentential and the multi-sentential parsing modules, which can exploit the strong correlation observed between the text structure and the structure of the discourse tree. The first approach (1S–1S) builds a DT for every sentence using the intra-sentential parser, and then
runs the multi-sentential parser on the resulting sentence-level DTs. To deal with leaky boundaries, our second approach (sliding window) builds sentence-level discourse sub-trees by applying the intra-sentential parser on a sliding window, covering two adjacent sentences and then consolidating the results produced by overlapping windows. After that, the multi-sentential parser takes all these sentence-level sub-trees and builds a full rhetorical parse for the whole document.
Finally, we have extended the parsing algorithm to generate k most probable parse hypotheses for each input text, which could be used in a reranker to improve over the initial ranking using global features like long-range structural dependencies.
Empirical evaluations on two different genres demonstrate that our approach to discourse segmentation achieves state-of-the-art performance more efficiently using fewer features. A series of experiments on the discourse parsing task shows that both our intra- and multi-sentential parsers significantly outperform the state of the art, often by a wide margin. A comparison between our combination strategies reveals that the sliding window approach is more robust across domains. Furthermore, the oracle accuracy computed based on the k-best parse hypotheses generated by our parser demonstrates that a reranker could potentially improve the accuracy even further.
Our error analysis reveals that although the sliding window approach finds more correct tree structures, in some cases it induces noise for the relation labels from the neighboring sentences. With respect to the performance of our discourse parser on the relation labeling task we also found that the most frequent relations tend to mislead the identification of the less frequent ones, and the models sometimes fail to distinguish relations that are semantically similar.
The work presented in this article leads us to several interesting future directions. Our short-term goal is to develop a k-best discriminative reranking discourse parser using tree kernels applied to discourse trees. We also plan to investigate to what extent discourse segmentation and discourse parsing can be performed jointly.
We would also like to explore how our system performs on other genres like conversational (e.g., blogs, e-mails) and evaluative (e.g., customer reviews) texts. To address the problem of limited annotated data in various genres, we are planning to develop an interactive version of our system that will allow users to fix the output of the system with minimal effort and let the system learn from that feedback.
Another interesting future direction is to perform extrinsic evaluations of our system in downstream applications. One important application of rhetorical structure is text summarization, where a significant challenge is producing not only informative but also coherent summaries. A number of researchers have already investigated the utility of rhetorical structure for measuring text importance (i.e., informativeness) in summarization (Marcu 2000b; Daumé and Marcu 2002; Louis, Joshi, and Nenkova 2010). Recently, Christensen et al. (2013, 2014) propose to perform sentence selection and ordering at the same time, and use constraints on discourse structure to make the summaries coherent. However, they represent the discourse as an unweighted directed graph, which is shallow and not sufficiently informative in most cases. Furthermore, their approach does not allow compression at the sentence level, which is often beneficial in summarization. In the future, we would like to investigate the utility of our rhetorical structure for performing sentence compression, selection, and ordering in a joint summarization process.
Discourse structure can also play important roles in sentiment analysis. A key research problem in sentiment analysis is extracting fine-grained opinions about different aspects of a product. Several recent papers (Somasundaran 2010; Lazaridou, Titov, and
Sporleder 2013) exploited the rhetorical structure for this task. Another challenging problem is assessing the overall opinion expressed in a review because not all sentences in a review contribute equally to the overall sentiment. For example, some sentences are subjective, whereas others are objective (Pang and Lee 2004); some express the main claims, whereas others support them (Taboada et al. 2011); some express opinions about the main entity, whereas others are about the peripherals. Discourse structure could be useful to capture the relative weights of the discourse units towards the overall sentiment. For example, the nucleus and satellite distinction along with the rhetorical relations could be useful to infer the relative weights of the connecting discourse units.
Among other applications of discourse structure, Machine Translation (MT) and its evaluation have received a resurgence of interest recently. A workshop dedicated to discourse in machine translation was arranged recently at the ACL 2013 conference (Webber et al. 2013). Researchers believe that MT systems should consider discourse phenomena that go beyond the current sentence to ensure consistency in the choice of lexical items or referring expressions, and the fact that source-language coherence relations are also realized in the target language (i.e., translating at the documentlevel [Hardmeier, Nivre, and Tiedemann 2012]). Guzmán et al. (2014a, 2014b) and Joty et al. (2014) propose new discourse-aware automatic evaluation metrics for MT systems using our discourse analysis tool. They demonstrate that sentence-level discourse information is complementary to the state-of-the-art evaluation metrics, and by combining the discourse-based metrics with the metrics from the ASIYA MT evaluation toolkit (Giménez and Màrquez 2010), they won the WMT 2014 metrics shared task challenge (Macháček and Bojar 2014) both at the segment- and at the system-level. These results suggest that discourse structure helps to distinguish better translations from worse ones. Thus, it would be interesting to explore whether discourse information can be used to rerank alternative MT hypotheses as a post-processing step for the MT output.
A longer-term goal is to extend our framework to also work with graph structures of discourse, as recommended by several recent discourse theories (Wolf and Gibson 2005). Once we achieve similar performance on graph structures, we will perform extrinsic evaluations to determine their relative utility for various NLP tasks.
Finally, we hope that the online demo, the source code of CODRA, and the evaluation metrics that we made publicly available in this work will facilitate other researchers in extending our work and in applying discourse parsing to their NLP tasks.
Bibliographic Note
Portions of this work were previously published in two conference proceedings (Joty, Carenini, and Ng 2012; Joty et al. 2013). This article significantly extends our previous work in several ways, most notably: (1) we extend the parsing algorithm to generate k-most probable parse hypotheses for each input text (Section 4.2); (2) we show the oracle accuracies for k-best discourse parsing both at the sentence level and at the document level (Section 6.4.4); (3) to support our claim, we compare our best results with several variations of our approach (see CRF-NC in Section 6.4.2, and CRF-O and CRF-T in Section 6.4.3); (4) we analyze the relative importance of different features for intra- and multi-sentential discourse parsing (Section 6.4.5); and (5) we perform indepth error analysis of our complete rhetorical analysis framework (Section 6.4.6).
Appendix Sample Output Generated by an Online Demo of CODRA
A. Clauses and sentences rarely stand on their own in an actual discourse; rather, the relationship between them carries important information that allows the discourse to express a meaning as a whole beyond the sum of its individual parts. Rhetorical analysis seeks to uncover this coherence structure. In this article, we present CODRA— a COmplete probabilistic Discriminative framework for performing Rhetorical Analysis in accordance with Rhetorical Structure Theory, which posits a tree representation of a discourse. CODRA comprises a discourse segmenter and a discourse parser. First, the discourse segmenter, which is based on a binary classifier, identifies the elementary discourse units in a given text. Then the discourse parser builds a discourse tree by applying an optimal parsing algorithm to probabilities inferred from two Conditional Random Fields: one for intra-sentential parsing and the other for multi-sentential parsing. We present two approaches to combine these two stages of parsing effectively. By conducting a series of empirical evaluations over two different data sets, we demonstrate that CODRA significantly outperforms the state-of-the-art, often by a wide margin. We also show that a reranking of the k-best parse hypotheses generated by CODRA can potentially improve the accuracy even further. [{""affiliations"": [], ""name"": ""Shafiq Joty""}, {""affiliations"": [], ""name"": ""Giuseppe Carenini""}, {""affiliations"": [], ""name"": ""Raymond T. Ng""}] SP:6e0be3b1258a50030cf084705cb3a08259f2f268 [{""authors"": [""Althaus"", ""Ernst"", ""Denys Duchier"", ""Alexander Koller"", ""Kurt Mehlhorn"", ""Joachim Niehren"", ""Sven Thiel.""], ""title"": ""An efficient graph algorithm for dominance constraints"", ""venue"": ""Journal of Algorithms, 48(1):194\u2013219."", ""year"": 2003}, {""authors"": [""Asher"", ""Nicholas"", ""Alex Lascarides""], ""title"": ""Logics of Conversation"", ""year"": 2003}, {""authors"": [""Barzilay"", ""Regina"", ""Michael Elhadad.""], ""title"": ""Using lexical chains for text summarization"", ""venue"": ""Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and the 8th"", ""year"": 1997}, {""authors"": [""Biran"", ""Or"", ""Owen Rambow.""], ""title"": ""Identifying justifications in written dialogs by classifying text as argumentative"", ""venue"": ""International Journal of Semantic Computing, 5(4):363\u2013381."", ""year"": 2011}, {""authors"": [""Blair-Goldensohn"", ""Sasha"", ""Kathleen McKeown"", ""Owen Rambow.""], ""title"": ""Building and refining rhetorical-semantic relation models"", ""venue"": ""Proceedings of the Human Language Technologies: The Annual"", ""year"": 2007}, {""authors"": [""Blitzer"", ""John.""], ""title"": ""Domain Adaptation of Natural Language Processing Systems"", ""venue"": ""Ph.D. thesis, University of Pennsylvania."", ""year"": 2008}, {""authors"": [""Breiman"", ""Leo.""], ""title"": ""Bagging predictors"", ""venue"": ""Machine Learning, 24(2):123\u2013140."", ""year"": 1996}, {""authors"": [""Carlson"", ""Lynn"", ""Daniel Marcu.""], ""title"": ""Discourse tagging reference manual"", ""venue"": ""Technical Report ISI-TR-545, University of Southern California Information Sciences Institute."", ""year"": 2001}, {""authors"": [""Carlson"", ""Lynn"", ""Daniel Marcu"", ""Mary Ellen Okurowski.""], ""title"": ""RST Discourse Treebank (RST\u2013DT) LDC2002T07"", ""venue"": ""Linguistic Data Consortium, Philadelphia."", ""year"": 2002}, {""authors"": [""Chali"", ""Yllias"", ""Shafiq Joty""], ""title"": ""Word sense disambiguation using lexical"", ""year"": 2007}, {""authors"": [""Charniak"", ""Eugene.""], ""title"": ""A maximum-entropy-inspired parser"", ""venue"": ""Proceedings of the 1st North American Chapter of the Association for Computational Linguistics Conference, NAACL\u201900,"", ""year"": 2000}, {""authors"": [""Charniak"", ""Eugene"", ""Mark Johnson.""], ""title"": ""Coarse-to-fine n-best parsing and MaxEnt discriminative reranking"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL\u201905,"", ""year"": 2005}, {""authors"": [""Christensen"", ""Janara"", ""Mausam"", ""Stephen Soderland"", ""Oren Etzioni.""], ""title"": ""Towards coherent multi-document summarization"", ""venue"": ""Proceedings of the 2013 Conference of the North American Chapter of"", ""year"": 2013}, {""authors"": [""Christensen"", ""Janara"", ""Stephen Soderland"", ""Gagan Bansal"", ""Mausam.""], ""title"": ""Hierarchical summarization: Scaling up multi-document summarization"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the"", ""year"": 2014}, {""authors"": [""Collins"", ""Michael.""], ""title"": ""Head-driven statistical models for natural language Parsing"", ""venue"": ""Computational Linguistics, 29(4):589\u2013637."", ""year"": 2003}, {""authors"": [""Collins"", ""Michael"", ""Terry Koo.""], ""title"": ""Discriminative reranking for natural language parsing"", ""venue"": ""Computational Linguistics, 31(1):25\u201370."", ""year"": 2005}, {""authors"": [""Collobert"", ""Ronan"", ""Jason Weston"", ""L\u00e9on Bottou"", ""Michael Karlen"", ""Koray Kavukcuoglu"", ""Pavel Kuksa.""], ""title"": ""Natural language processing (almost) from scratch"", ""venue"": ""Journal of Machine Learning Research, 12:2493\u20132537."", ""year"": 2011}, {""authors"": [""Cristea"", ""Dan"", ""Nancy Ide"", ""Laurent Romary.""], ""title"": ""Veins theory: A model of global discourse cohesion and coherence"", ""venue"": ""Proceedings of the 36th Annual Meeting of the Association for Computational"", ""year"": 1998}, {""authors"": [""Danlos"", ""Laurence.""], ""title"": ""D-STAG: A discourse analysis formalism based on synchronous TAGs"", ""venue"": ""TAL, 50(1):111\u2013143."", ""year"": 2009}, {""authors"": [""III Daum\u00e9"", ""Hal.""], ""title"": ""Frustratingly easy domain adaptation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, ACL\u201907, pages 256\u2013263, Prague."", ""year"": 2007}, {""authors"": [""III Daum\u00e9"", ""Hal"", ""Daniel Marcu.""], ""title"": ""A noisy-channel model for document compression"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, ACL \u201902,"", ""year"": 2002}, {""authors"": [""Dinarelli"", ""Marco"", ""Alessandro Moschitti"", ""Giuseppe Riccardi.""], ""title"": ""Discriminative reranking for spoken language understanding"", ""venue"": ""IEEE Transactions on Audio, Speech and Language Processing (TASLP),"", ""year"": 2011}, {""authors"": [""duVerle"", ""David"", ""Helmut Prendinger""], ""title"": ""A novel discourse parser based on support vector machine classification"", ""venue"": ""In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the"", ""year"": 2009}, {""authors"": [""Egg"", ""Markus"", ""Alexander Koller"", ""Joachim Niehren.""], ""title"": ""The constraint language for lambda structures"", ""venue"": ""Journal of Logic, Language and Information, 10(4):457\u2013485."", ""year"": 2001}, {""authors"": [""Eisner"", ""Jason.""], ""title"": ""Three new probabilistic models for dependency parsing: An exploration"", ""venue"": ""Proceedings of the 16th Conference on Computational Linguistics Volume 1, COLING \u201996, pages 340\u2013345,"", ""year"": 1996}, {""authors"": [""Fellbaum"", ""Christiane.""], ""title"": ""WordNet\u2014An Electronic Lexical Database"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 1998}, {""authors"": [""Feng"", ""Vanessa"", ""Graeme Hirst.""], ""title"": ""Text-level discourse parsing with rich linguistic features"", ""venue"": ""Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics, ACL \u201912,"", ""year"": 2012}, {""authors"": [""Feng"", ""Vanessa"", ""Graeme Hirst.""], ""title"": ""A linear-time bottom-up discourse parser with constraints and post-editing"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2014}, {""authors"": [""Finkel"", ""Jenny Rose"", ""Alex Kleeman"", ""Christopher Manning.""], ""title"": ""Efficient, feature-based, conditional random field parsing"", ""venue"": ""Proceedings of the 46th Annual Meeting of the Association for Computational"", ""year"": 2008}, {""authors"": [""Fisher"", ""Seeger"", ""Brian Roark.""], ""title"": ""The utility of parse-derived features for automatic discourse segmentation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2007}, {""authors"": [""Galley"", ""Michel"", ""Kathleen McKeown.""], ""title"": ""Improving word sense disambiguation in lexical chaining"", ""venue"": ""Proceedings of the 18th International Joint Conference on Artificial Intelligence, IJCAI\u201903, pages 1486\u20131488,"", ""year"": 2003}, {""authors"": [""Galley"", ""Michel"", ""Kathleen McKeown"", ""Eric Fosler-Lussier"", ""Hongyan Jing.""], ""title"": ""Discourse segmentation of multi-party conversation"", ""venue"": ""Proceedings of the 41st Annual Meeting of the Association for"", ""year"": 2003}, {""authors"": [""Ghosh"", ""Sucheta"", ""Richard Johansson"", ""Giuseppe Riccardi"", ""Sara Tonelli.""], ""title"": ""Shallow discourse parsing with conditional random fields"", ""venue"": ""Proceedings of the 5th International Joint Conference on"", ""year"": 2011}, {""authors"": [""Gim\u00e9nez"", ""Jes\u00fas"", ""Llu\u0131\u0301s M\u00e0rquez""], ""title"": ""Linguistic measures for automatic machine translation evaluation"", ""venue"": ""Machine Translation,"", ""year"": 2010}, {""authors"": [""Guzm\u00e1n"", ""Francisco"", ""Shafiq Joty"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Alessandro Moschitti"", ""Preslav Nakov"", ""Massimo Nicosia""], ""title"": ""Learning to differentiate better from worse translations"", ""year"": 2014}, {""authors"": [""Guzm\u00e1n"", ""Francisco"", ""Shafiq Joty"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Preslav Nakov""], ""title"": ""Using discourse structure improves machine translation evaluation"", ""venue"": ""In Proceedings of the 52nd Annual Meeting of the Association"", ""year"": 2014}, {""authors"": [""Halliday"", ""Michael"", ""Ruqaiya Hasan.""], ""title"": ""Cohesion in English"", ""venue"": ""Longman, London."", ""year"": 1976}, {""authors"": [""Hardmeier"", ""Christian"", ""Joakim Nivre"", ""J\u00f6rg Tiedemann.""], ""title"": ""Document-wide decoding for phrase-based statistical machine translation"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods"", ""year"": 2012}, {""authors"": [""Hernault"", ""Hugo"", ""Helmut Prendinger"", ""David duVerle"", ""Mitsuru Ishizuka.""], ""title"": ""HILDA: A discourse parser using support vector machine classification"", ""venue"": ""Dialogue and Discourse, 1(3):1\u201333."", ""year"": 2010}, {""authors"": [""Hirst"", ""Graeme"", ""David St-Onge.""], ""title"": ""Lexical Chains as Representation of Context for the Detection and Correction of Malapropisms"", ""venue"": ""Christiane Fellbaum, 431"", ""year"": 1997}, {""authors"": [""Hobbs"", ""Jerry.""], ""title"": ""Coherence and coreference"", ""venue"": ""Cognitive Science, 3:67\u201390."", ""year"": 1979}, {""authors"": [""Huang"", ""Liang"", ""David Chiang.""], ""title"": ""Better k-best parsing"", ""venue"": ""Proceedings of the Ninth International Workshop on Parsing Technology, Parsing \u201905, pages 53\u201364, Stroudsburg, PA."", ""year"": 2005}, {""authors"": [""Ji"", ""Yangfeng"", ""Jacob Eisenstein.""], ""title"": ""Representation learning for text-level discourse parsing"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long"", ""year"": 2014}, {""authors"": [""Joty"", ""Shafiq"", ""Giuseppe Carenini"", ""Raymond T. Ng.""], ""title"": ""A novel discriminative framework for sentence-level discourse analysis"", ""venue"": ""Proceedings of the 2012 Joint Conference on"", ""year"": 2012}, {""authors"": [""Joty"", ""Shafiq"", ""Giuseppe Carenini"", ""Raymond T. Ng.""], ""title"": ""Topic segmentation and labeling in asynchronous Conversations"", ""venue"": ""Journal of Artificial Intelligence Research (JAIR), 47:521\u2013573."", ""year"": 2013}, {""authors"": [""Joty"", ""Shafiq"", ""Giuseppe Carenini"", ""Raymond T. Ng"", ""Yashar Mehdad.""], ""title"": ""Combining intra- and multi-sentential rhetorical parsing for document-level discourse analysis"", ""venue"": ""Proceedings of the 51st Annual"", ""year"": 2013}, {""authors"": [""Joty"", ""Shafiq"", ""Francisco Guzm\u00e1n"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Preslav Nakov""], ""title"": ""DiscoTK: Using discourse structure for machine translation evaluation"", ""venue"": ""In Proceedings of the Ninth Workshop"", ""year"": 2014}, {""authors"": [""Jurafsky"", ""Daniel"", ""James Martin.""], ""title"": ""Statistical parsing"", ""venue"": ""Speech and Language Processing, chapter 14. Prentice Hall."", ""year"": 2008}, {""authors"": [""Knight"", ""Kevin"", ""Jonathan Graehl.""], ""title"": ""An overview of probabilistic tree transducers for natural language processing"", ""venue"": ""Computational Linguistics and Intelligent Text Processing, volume 3406 of Lecture Notes in"", ""year"": 2005}, {""authors"": [""Knott"", ""Alistair"", ""Robert Dale.""], ""title"": ""Using linguistic phenomena to motivate a set of coherence relations"", ""venue"": ""Discourse Processes, 18:35\u201362."", ""year"": 1994}, {""authors"": [""Koller"", ""Alexander"", ""Michaela Regneri"", ""Stefan Thater.""], ""title"": ""Regular tree grammars as a formalism for scope underspecification"", ""venue"": ""Proceedings of the 46th Annual Meeting of the Association for"", ""year"": 2008}, {""authors"": [""Lafferty"", ""John"", ""Andrew McCallum"", ""Fernando Pereira.""], ""title"": ""Conditional random fields: Probabilistic models for segmenting and labeling sequence data"", ""venue"": ""Proceedings of the Eighteenth International"", ""year"": 2001}, {""authors"": [""Lazaridou"", ""Angeliki"", ""Ivan Titov"", ""Caroline Sporleder.""], ""title"": ""A Bayesian model for joint unsupervised induction of sentiment, aspect and discourse representations"", ""venue"": ""Proceedings of the 51st Annual Meeting of the"", ""year"": 2013}, {""authors"": [""Li"", ""Jiwei"", ""Rumeng Li"", ""Eduard Hovy.""], ""title"": ""Recursive deep models for discourse parsing"", ""venue"": ""Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP),"", ""year"": 2014}, {""authors"": [""Li"", ""Sujian"", ""Liang Wang"", ""Ziqiang Cao"", ""Wenjie Li.""], ""title"": ""Text-level discourse dependency parsing"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long"", ""year"": 2014}, {""authors"": [""Louis"", ""Annie"", ""Aravind Joshi"", ""Ani Nenkova.""], ""title"": ""Discourse indicators for content selection in summarization"", ""venue"": ""Proceedings of the 11th Annual Meeting of the Special Interest Group on Discourse and"", ""year"": 2010}, {""authors"": [""Mach\u00e1\u010dek"", ""Matou\u0161"", ""Ond\u0159ej Bojar.""], ""title"": ""Results of the WMT14 metrics shared task"", ""venue"": ""Proceedings of the Ninth Workshop on Statistical Machine Translation, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Magerman"", ""David.""], ""title"": ""Statistical decision-tree models for parsing"", ""venue"": ""Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics, ACL\u201995, pages 276\u2013283, Cambridge, MA."", ""year"": 1995}, {""authors"": [""Mann"", ""William"", ""Sandra Thompson.""], ""title"": ""Rhetorical structure theory: Toward a functional theory of text organization"", ""venue"": ""Text, 8(3):243\u2013281."", ""year"": 1988}, {""authors"": [""Marcu"", ""Daniel.""], ""title"": ""A decision-based approach to rhetorical parsing"", ""venue"": ""Proceedings of the 37th Annual Meeting of the Association for Computational Linguistics on 432"", ""year"": 1999}, {""authors"": [""Marcu"", ""Daniel.""], ""title"": ""The rhetorical parsing of unrestricted texts: A surface-based approach"", ""venue"": ""Computational Linguistics, 26:395\u2013448."", ""year"": 2000}, {""authors"": [""Marcu"", ""Daniel.""], ""title"": ""The Theory and Practice of Discourse Parsing and Summarization"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 2000}, {""authors"": [""Marcu"", ""Daniel"", ""Abdessamad Echihabi.""], ""title"": ""An unsupervised approach to recognizing discourse relations"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for Computational"", ""year"": 2002}, {""authors"": [""Marcus"", ""Mitchell"", ""Mary Marcinkiewicz"", ""Beatrice Santorini.""], ""title"": ""Building a large annotated corpus of English: The Penn Treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1994}, {""authors"": [""Maslennikov"", ""Mstislav"", ""Tat-Seng Chua.""], ""title"": ""A multi-resolution framework for information extraction from free text"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2007}, {""authors"": [""McCallum"", ""Andrew.""], ""title"": ""MALLET: A machine learning for language toolkit"", ""venue"": ""http://mallet.cs.umass.edu."", ""year"": 2002}, {""authors"": [""McCallum"", ""Andrew"", ""Dayne Freitag"", ""Fernando C.N. Pereira.""], ""title"": ""Maximum entropy Markov models for information extraction and segmentation"", ""venue"": ""Proceedings of the Seventeenth International"", ""year"": 2000}, {""authors"": [""McDonald"", ""Ryan"", ""Koby Crammer"", ""Fernando Pereira.""], ""title"": ""Online large-margin training of dependency parsers"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Fernando Pereira"", ""Kiril Ribarov"", ""Jan Haji\u010d.""], ""title"": ""Non-projective dependency parsing using spanning tree algorithms"", ""venue"": ""Proceedings of the Conference on Human Language"", ""year"": 2005}, {""authors"": [""Morris"", ""Jane"", ""Graeme Hirst.""], ""title"": ""Lexical cohesion computed by thesaural relations as an indicator of structure of text"", ""venue"": ""Computational Linguistics, 17(1):21\u201348."", ""year"": 1991}, {""authors"": [""Murphy"", ""Kevin.""], ""title"": ""Machine Learning: A Probabilistic Perspective"", ""venue"": ""The MIT Press. Cambridge, MA."", ""year"": 2012}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts"", ""venue"": ""Proceedings of the 42nd Annual Meeting of the Association for"", ""year"": 2004}, {""authors"": [""Pitler"", ""Emily"", ""Ani Nenkova.""], ""title"": ""Using syntax to disambiguate explicit discourse connectives in text"", ""venue"": ""Proceedings of the ACL-IJCNLP 2009 Conference Short Papers, ACLShort \u201909, pages 13\u201316, Suntec."", ""year"": 2009}, {""authors"": [""Poole"", ""David"", ""Alan Mackworth""], ""title"": ""Artificial Intelligence: Foundations of Computational Agents"", ""year"": 2010}, {""authors"": [""Prasad"", ""Rashmi"", ""Nikhil Dinesh"", ""Alan Lee"", ""Eleni Miltsakaki"", ""Livio Robaldo"", ""Aravind Joshi"", ""Bonnie Webber""], ""title"": ""The Penn Discourse TreeBank 2.0"", ""venue"": ""In Proceedings of the Sixth International Conference"", ""year"": 2008}, {""authors"": [""Prasad"", ""Rashmi"", ""Aravind Joshi"", ""Nikhil Dinesh"", ""Alan Lee"", ""Eleni Miltsakaki"", ""Bonnie Webber.""], ""title"": ""The Penn Discourse TreeBank as a resource for natural language generation"", ""venue"": ""Proceedings of the"", ""year"": 2005}, {""authors"": [""Regneri"", ""Michaela"", ""Markus Egg"", ""Alexander Koller.""], ""title"": ""Efficient processing of underspecified discourse representations"", ""venue"": ""Proceedings of the 46th Annual Meeting of the Association for"", ""year"": 2008}, {""authors"": [""Reyle"", ""Uwe.""], ""title"": ""Dealing with ambiguities by underspecification: Construction, representation and deduction"", ""venue"": ""Journal of Semantics, 10(2):123\u2013179."", ""year"": 1993}, {""authors"": [""Schapire"", ""Robert E."", ""Yoram Singer.""], ""title"": ""Boostexter: A boosting-based system for text categorization"", ""venue"": ""Machine Learning, 39(2\u20133):135\u2013168."", ""year"": 2000}, {""authors"": [""Schauer"", ""Holger"", ""Udo Hahn.""], ""title"": ""Anaphoric cues for coherence relations"", ""venue"": ""Proceedings of the Conference on Recent Advances in Natural Language Processing, RANLP \u201901, pages 228\u2013234,"", ""year"": 2001}, {""authors"": [""Schilder"", ""Frank""], ""title"": ""Robust discourse parsing via discourse markers, topicality"", ""year"": 2002}, {""authors"": [""Sha"", ""Fei"", ""Fernando Pereira.""], ""title"": ""Shallow parsing with conditional random fields"", ""venue"": ""Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human"", ""year"": 2003}, {""authors"": [""Silber"", ""Gregory"", ""Kathleen McCoy.""], ""title"": ""Efficiently computed lexical chains as an intermediate representation for automatic text summarization"", ""venue"": ""Computational Linguistics, 28(4):487\u2013496."", ""year"": 2002}, {""authors"": [""Smith"", ""Noah A.""], ""title"": ""Linguistic Structure Prediction"", ""venue"": ""Synthesis Lectures on Human Language Technologies. Morgan and Claypool."", ""year"": 2011}, {""authors"": [""Socher"", ""Richard"", ""John Bauer"", ""Christopher D. Manning"", ""Ng Andrew Y.""], ""title"": ""Parsing with compositional vector grammars"", ""venue"": ""Proceedings of the 51st Annual Meeting of the Association for Computational"", ""year"": 2013}, {""authors"": [""Socher"", ""Richard"", ""Alex Perelygin"", ""Jean Wu"", ""Jason Chuang"", ""Christopher D. Manning"", ""Andrew Ng"", ""Christopher Potts""], ""title"": ""Recursive deep models for semantic compositionality over a sentiment"", ""year"": 2013}, {""authors"": [""S. Somasundaran""], ""title"": ""Discourse-Level Relations for Opinion Analysis"", ""venue"": ""Ph.D. thesis, University of Pittsburgh, PA."", ""year"": 2010}, {""authors"": [""Soricut"", ""Radu"", ""Daniel Marcu.""], ""title"": ""Sentence level discourse parsing using syntactic and lexical information"", ""venue"": ""Proceedings of the 2003 Conference of the North American Chapter of the Association for"", ""year"": 2003}, {""authors"": [""Sporleder"", ""Caroline.""], ""title"": ""Manually vs"", ""venue"": ""automatically labelled data in discourse relation classification. Effects of example and feature selection. LDV Forum, 22(1):1\u201320."", ""year"": 2007}, {""authors"": [""Sporleder"", ""Caroline"", ""Mirella Lapata.""], ""title"": ""Automatic paragraph identification: A study across languages and domains"", ""venue"": ""Proceedings of the 2004 Conference on Empirical Methods in Natural Language"", ""year"": 2004}, {""authors"": [""Sporleder"", ""Caroline"", ""Mirella Lapata.""], ""title"": ""Discourse chunking and its application to sentence compression"", ""venue"": ""Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural"", ""year"": 2005}, {""authors"": [""Sporleder"", ""Caroline"", ""Alex Lascarides.""], ""title"": ""Exploiting linguistic cues to classify rhetorical relations"", ""venue"": ""Proceedings of Recent Advances in Natural Language Processing (RANLP), pages 157\u2013166, Bulgaria."", ""year"": 2005}, {""authors"": [""Sporleder"", ""Caroline"", ""Alex Lascarides.""], ""title"": ""Using automatically labelled examples to classify rhetorical relations: An assessment"", ""venue"": ""Natural Language Engineering, 14(3):369\u2013416."", ""year"": 2008}, {""authors"": [""Stede"", ""Manfred.""], ""title"": ""The Potsdam Commentary Corpus"", ""venue"": ""Proceedings of the ACL-04 Workshop on Discourse Annotation, pages 96\u2013102, Barcelona."", ""year"": 2004}, {""authors"": [""Stede"", ""Manfred.""], ""title"": ""Discourse Processing"", ""venue"": ""Synthesis Lectures on Human Language Technologies. Morgan and Claypool Publishers."", ""year"": 2011}, {""authors"": [""Subba"", ""Rajen"", ""Barbara Di-Eugenio.""], ""title"": ""An effective discourse parser that uses rich linguistic information"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North"", ""year"": 2009}, {""authors"": [""Sutton"", ""Charles"", ""Andrew McCallum.""], ""title"": ""An introduction to conditional random fields"", ""venue"": ""Foundations and Trends in Machine Learning, 4(4):267\u2013373."", ""year"": 2012}, {""authors"": [""Sutton"", ""Charles"", ""Andrew McCallum"", ""Khashayar Rohanimanesh.""], ""title"": ""Dynamic conditional random fields: Factorized probabilistic models for labeling and segmenting sequence data"", ""venue"": ""Journal of"", ""year"": 2007}, {""authors"": [""Taboada"", ""Maite.""], ""title"": ""Discourse markers as signals (or not) of rhetorical relations"", ""venue"": ""Journal of Pragmatics, 38(4):567\u2013592."", ""year"": 2006}, {""authors"": [""Taboada"", ""Maite"", ""Julian Brooke"", ""Milan Tofiloski"", ""Kimberly Voll"", ""Manfred Stede.""], ""title"": ""Lexicon-based methods for sentiment analysis"", ""venue"": ""Computational Linguistics, 37(2):267\u2013307."", ""year"": 2011}, {""authors"": [""Taboada"", ""Maite"", ""William C. Mann.""], ""title"": ""Rhetorical structure theory: Looking back and moving ahead"", ""venue"": ""Discourse Studies, 8(3):423\u2013459."", ""year"": 2006}, {""authors"": [""Teufel"", ""Simone"", ""Marc Moens""], ""title"": ""Summarizing scientific articles: Experiments with relevance and rhetorical"", ""year"": 2002}, {""authors"": [""Verberne"", ""Suzan"", ""Lou Boves"", ""Nelleke Oostdijk"", ""Peter-Arno Coppen.""], ""title"": ""Evaluating discourse-based answer extraction for why-question answering"", ""venue"": ""Proceedings of the 30th Annual International"", ""year"": 2007}, {""authors"": [""Vliet"", ""Nynke"", ""Gisela Redeker.""], ""title"": ""Complex sentences as leaky units in discourse parsing"", ""venue"": ""Proceedings of Constraints in Discourse, pages 1\u20139, Agay\u2013Saint Raphael."", ""year"": 2011}, {""authors"": [""B. Webber""], ""title"": ""D-LTAG: Extending lexicalized TAG to discourse"", ""venue"": ""Cognitive Science, 28(5):751\u2013779."", ""year"": 2004}, {""authors"": [""Wick"", ""Michael"", ""Khashayar Rohanimanesh"", ""Kedare Bellare"", ""Aron Culotta"", ""Andrew McCallum.""], ""title"": ""Samplerank: Training factor graphs with atomic gradients"", ""venue"": ""Proceedings of the 28th"", ""year"": 2011}, {""authors"": [""Wolf"", ""Florian"", ""Edward Gibson.""], ""title"": ""Representing discourse coherence: A corpus-based study"", ""venue"": ""Computational Linguistics, 31:249\u2013288."", ""year"": 2005}, {""authors"": [""Zhang"", ""Yuan"", ""Tao Lei"", ""Regina Barzilay"", ""Tommi Jaakkola"", ""Amir Globerson.""], ""title"": ""Steps to excellence: Simple inference with refined scoring of dependency trees"", ""venue"": ""Proceedings of the 52nd Annual Meeting of"", ""year"": 2014}] acknowledgments  :The authors acknowledge the funding support of NSERC Canada Graduate Scholarship (CGS-D). Many thanks to Bonnie Webber, Amanda Stent, Carolyn Rose, Lluis Marquez, Samantha Wray, and the anonymous reviewers for their insightful comments on an earlier version of this article.","1 introduction :A well-written text is not merely a sequence of independent and isolated sentences, but instead a sequence of structured and related sentences, where the meaning of a sentence relates to the previous and the following ones. In other words, a well-written
∗ Arabic Language Technologies, Qatar Computing Research Institute, Qatar Foundation, Doha, Qatar. E-mail: sjoty@qf.org.qa.
∗∗ Computer Science Department, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4. E-mail: carenini@cs.ubc.ca.
† Computer Science Department, University of British Columbia, Vancouver, BC, Canada, V6T 1Z4. E-mail: rng@cs.ubc.ca.
Submission received: 11 May 2014; revised version received: 29 January 2015; accepted for publication: 18 March 2015.
doi:10.1162/COLI a 00226
No rights reserved. This work was authored as part of the Contributor’s official duties as an Employee of the United States Government and is therefore a work of the United States Government. In accordance with 17 U.S.C. 105, no copyright protection is available for such works under U.S. Law.
text has a coherence structure (Halliday and Hasan 1976; Hobbs 1979), which logically binds its clauses and sentences together to express a meaning as a whole. Rhetorical analysis seeks to uncover this coherence structure underneath the text; this has been shown to be beneficial for many Natural Language Processing (NLP) applications, including text summarization and compression (Marcu 2000b; Daumé and Marcu 2002; Sporleder and Lapata 2005; Louis, Joshi, and Nenkova 2010), text generation (Prasad et al. 2005), machine translation evaluation (Guzmán et al. 2014a, 2014b; Joty et al. 2014), sentiment analysis (Somasundaran 2010; Lazaridou, Titov, and Sporleder 2013), information extraction (Teufel and Moens 2002; Maslennikov and Chua 2007), and question answering (Verberne et al. 2007). Furthermore, rhetorical structures can be useful for other discourse analysis tasks, including co-reference resolution using Veins theory (Cristea, Ide, and Romary 1998).
Different formal theories of discourse have been proposed from different viewpoints to describe the coherence structure of a text. For example, Martin (1992) and Knott and Dale (1994) propose discourse relations based on the usage of discourse connectives (e.g., because, but) in the text. Asher and Lascarides (2003) propose Segmented Discourse Representation Theory, which is driven by sentence semantics. Webber (2004) and Danlos (2009) extend sentence grammar to formalize discourse structure. Rhetorical Structure Theory (RST), proposed by Mann and Thompson (1988), is perhaps the most influential theory of discourse in computational linguistics. Although it was initially intended to be used in text generation, later it became popular as a framework for parsing the structure of a text (Taboada and Mann 2006). RST represents texts by labeled hierarchical structures, called Discourse Trees (DTs). For example, consider the DT shown in Figure 1 for the following text:
But he added: “Some people use the purchasers’ index as a leading indicator, some use it as a coincident indicator. But the thing it’s supposed to measure—manufacturing strength—it missed altogether last month.”
The leaves of a DT correspond to contiguous atomic text spans, called elementary discourse units (EDUs; six in the example). EDUs are clause-like units that serve as building blocks. Adjacent EDUs are connected by coherence relations (e.g., Elaboration, Contrast), forming larger discourse units (represented by internal nodes), which in turn are also subject to this relation linking. Discourse units linked by a rhetorical relation are
further distinguished based on their relative importance in the text: nuclei are the core parts of the relation and satellites are peripheral or supportive ones. For example, in Figure 1, Elaboration is a relation between a nucleus (EDU 4) and a satellite (EDU 5), and Contrast is a relation between two nuclei (EDUs 2 and 3). Carlson, Marcu, and Okurowski (2002) constructed the first large RST-annotated corpus (RST–DT) on Wall Street Journal articles from the Penn Treebank. Whereas Mann and Thompson (1988) had suggested about 25 relations, the RST–DT uses 53 mono-nuclear and 25 multi-nuclear relations. The relations are grouped into 16 coarse-grained categories; see Carlson and Marcu (2001) for a detailed description of the relations. Conventionally, rhetorical analysis in RST involves two subtasks: discourse segmentation is the task of breaking the text into a sequence of EDUs, and discourse parsing is the task of linking the discourse units (EDUs and larger units) into a labeled tree. In this article, we use the terms discourse parsing and rhetorical parsing interchangeably.
While recent advances in automatic discourse segmentation have attained high accuracies (an F-score of 90.5% reported by Fisher and Roark [2007]), discourse parsing still poses significant challenges (Feng and Hirst 2012) and the performance of the existing discourse parsers (Soricut and Marcu 2003; Subba and Di-Eugenio 2009; Hernault et al. 2010) is still considerably inferior compared with the human gold standard. Thus, the impact of rhetorical structure in downstream NLP applications is still very limited. The work we present in this article aims to reduce this performance gap and take discourse parsing one step further. To this end, we address three key limitations of existing discourse parsers.
First, existing discourse parsers typically model the structure and the labels of a DT separately, and also do not take into account the sequential dependencies between the DT constituents. However, for several NLP tasks, it has recently been shown that joint models typically outperform independent or pipeline models (Murphy 2012, page 687). This is also supported in a recent study by Feng and Hirst (2012), in which the performance of a greedy bottom–up discourse parser improved when sequential dependencies were considered by using gold annotations for the neighboring (i.e., previous and next) discourse units as contextual features in the parsing model. To address this limitation of existing parsers, as the first contribution, we propose a novel discourse parser based on probabilistic discriminative parsing models, expressed as Conditional Random Fields (CRFs) (Sutton, McCallum, and Rohanimanesh 2007), to infer the probability of all possible DT constituents. The CRF models effectively represent the structure and the label of a DT constituent jointly, and, whenever possible, capture the sequential dependencies.
Second, existing discourse parsers typically apply greedy and sub-optimal parsing algorithms to build a DT. To cope with this limitation, we use the inferred (posterior) probabilities from our CRF parsing models in a probabilistic CKY-like bottom–up parsing algorithm (Jurafsky and Martin 2008), which is non-greedy and optimal. Furthermore, a simple modification of this parsing algorithm allows us to generate k-best (i.e., the k highest probability) parse hypotheses for each input text that could then be used in a reranker to improve over the initial ranking using additional (global) features of the discourse tree as evidence, a strategy that has been successfully explored in syntactic parsing (Charniak and Johnson 2005; Collins and Koo 2005).
Third, most of the existing discourse parsers do not discriminate between intrasentential parsing (i.e., building the DTs for the individual sentences) and multisentential parsing (i.e., building the DT for the whole document). However, we argue that distinguishing between these two parsing conditions can result in more effective parsing. Two separate parsing models could exploit the fact that rhetorical relations
are distributed differently intra-sententially versus multi-sententially. Also, they could independently choose their own informative feature sets. As another key contribution of our work, we devise two different parsing components: one for intra-sentential parsing, the other for multi-sentential parsing. This provides for scalable, modular, and flexible solutions that can exploit the strong correlation observed between the text structure (i.e., sentence boundaries) and the structure of the discourse tree.
In order to develop a complete and robust discourse parser, we combine our intrasentential and multi-sentential parsing components in two different ways. Because most sentences have a well-formed discourse sub-tree in the full DT (e.g., the second sentence in Figure 1), our first approach constructs a DT for every sentence using our intrasentential parser, and then runs the multi-sentential parser on the resulting sentencelevel DTs to build a complete DT for the whole document. However, this approach would fail in those cases where discourse structures violate sentence boundaries, also called “leaky” boundaries (Vliet and Redeker 2011). For example, consider the first sentence in Figure 1. It does not have a well-formed discourse sub-tree because the unit containing EDUs 2 and 3 merges with the next sentence and only then is the resulting unit merged with EDU 1. Our second approach, in order to deal with these leaky cases, builds sentence-level sub-trees by applying the intra-sentential parser on a sliding window covering two adjacent sentences and by then consolidating the results produced by overlapping windows. After that, the multi-sentential parser takes all these sentence-level sub-trees and builds a full DT for the whole document.
Our discourse parser assumes that the input text has already been segmented into elementary discourse units. As an additional contribution, we propose a novel discriminative approach to discourse segmentation that not only achieves state-of-theart performance, but also reduces time and space complexities by using fewer features. Notice that the combination of our segmenter with our parser forms a COmplete probabilistic Discriminative framework for Rhetorical Analysis (CODRA).
Whereas previous systems have been tested on only one corpus, we evaluate our framework on texts from two very different genres: news articles and instructional howto manuals. The results demonstrate that our approach to discourse parsing provides consistent and statistically significant improvements over previous methods both at the sentence level and at the document level. The performance of our final system compares very favorably to the performance of state-of-the-art discourse parsers. Finally, the oracle accuracy computed based on the k-best parse hypotheses generated by our parser demonstrates that a reranker could potentially improve the accuracy further.
After discussing related work in Section 2, we present our rhetorical analysis framework in Section 3. In Section 4, we describe our discourse parser. Then, in Section 5 we present our discourse segmenter. The experiments and analysis of results are presented in Section 6. Finally, we summarize our contributions with future directions in Section 7. 6 experiments :In this section we present our experimental results. First, we describe the corpora on which the experiments were performed and the evaluation metrics used to measure the
11 Available at http://cogcomp.cs.illinois.edu/page/software.
performance of the discourse segmenter and the parser. Then we show the performance of our discourse segmenter, followed by the performance of our discourse parser. Whereas previous studies on discourse analysis only report their results on a particular corpus, to demonstrate the generality of our method, we experiment with texts from two very different genres: news articles and instructional how-to manuals.
Our first corpus is the standard RST–DT (Carlson, Marcu, and Okurowski 2002), which contains discourse annotations for 385 Wall Street Journal news articles taken from the Penn Treebank corpus (Marcus, Marcinkiewicz, and Santorini 1994). The corpus is partitioned into a training set of 347 documents and a test set of 38 documents. A total of 53 documents selected from both training and test sets were annotated by two human annotators. We measure human agreements based on this doubly annotated data set. We used 25 documents from the training set as our development set. In RST–DT, the original 25 rhetorical relations defined by Mann and Thompson (1988) are further divided into a set of 18 coarser relation classes with 78 finer-grained relations (see Carlson and Marcu [2001] for details). Our second corpus is the instructional corpus prepared by Subba and Di-Eugenio (2009), which contains discourse annotations for 176 how-to manuals on home repair. The corpus was annotated with 26 informational relations (e.g., Preparation–Act, Act–Goal).
For our experiments with the intra-sentential discourse parser, we extracted a sentence-level DT from a document-level DT by finding the sub-tree that exactly spans over the sentence. In RST–DT, by our count, 7, 321 out of 7, 673 sentences in the training set, 951 out of 991 sentences in the test set, and 1, 114 out of 1, 208 sentences in the doubly-annotated set have a well-formed DT. On the other hand, 3, 032 out of 3, 430 sentences in the instructional corpus have a well-formed DT. This forms the corpora for our experiments with intra-sentential discourse parsing. However, the existence of a well-formed DT is not a necessity for discourse segmentation; therefore, we do not exclude any sentence in our discourse segmentation experiments. In this subsection we describe the metrics used to measure both how much the annotators agree with each other, and how well the systems perform when their outputs are compared with human annotations for the discourse analysis tasks.
6.2.1 Metrics for Discourse Segmentation. Because sentence boundaries are considered to also be the EDU boundaries, we measure segmentation accuracy with respect to the intra-sentential segment boundaries, which is a standard method (Soricut and Marcu 2003; Fisher and Roark 2007). Specifically, if a sentence contains n EDUs, which corresponds to n − 1 intra-sentential segment boundaries, we measure the segmenter’s ability to correctly identify these n − 1 boundaries. Let h be the total number of intrasentential segment boundaries in the human annotation, m be the total number of intrasentential segment boundaries in the model output, and c be the total number of correct segment boundaries in the model output. Then, we measure Precision (P), Recall (R), and F-score for segmentation performance as follows:
P = cm, R = c h , and F−score = 2PR P + R = 2c h + m (14)
6.2.2 Metrics for Discourse Parsing. To evaluate parsing performance, we use the standard unlabeled and labeled precision, recall, and F-score as proposed by Marcu (2000b). The unlabeled metric measures how accurate the discourse parser is in finding the right structure (i.e., the skeleton) of the DT, while the labeled metrics measure the parser’s ability to find the right labels (i.e., nuclearity statuses or relation labels) in addition to the right structure. Assume, for example, that given the two sentences of Figure 1, our system generates the DT shown in Figure 15. In Figure 16, we show the same gold DT shown in Figure 1 (on the left), and the same system-generated DT shown in Figure 15 (on the right), when the two trees are aligned. For the sake of illustration, instead of showing the real EDUs, we only show their IDs. Notice that the automatic segmenter made two mistakes: (1) it broke the EDU marked 2–3 (Some people use the purchasers’ index as a leading indicator) in the human annotation into two separate EDUs, and (2) it could not identify EDU 5 (But the thing it’s supposed to measure) and EDU 6 (— manufacturing strength —) as two separate EDUs. Therefore, when we align the two annotations, we obtain seven EDUs in total.
In Table 2, we list all constituents of the two DTs and their associated labels at the span, nuclei, and relation levels. The recall (R) and precision (P) figures are shown at the bottom of the table. Notice that, following (Marcu 2000b), the relation labels are assigned to the children nodes rather than to the parent nodes in the evaluation process to deal with non-binary trees in human annotations. To our knowledge, no implementation of
the evaluation metrics was made publicly available. Therefore, to help other researchers, we have made our source code of the evaluation metrics publicly available.12
Given this evaluation setup, it is easy to understand that if the number of EDUs is the same in the human and system annotations (e.g., when the discourse parser uses gold discourse segmentation), and the discourse trees are binary, then we get the same figures for precision, recall, and F-score. In this section we present our experiments on discourse segmentation.
6.3.1 Experimental Set-up for Discourse Segmentation. We compare the performance of our discourse segmenter with the performance of the two publicly available discourse segmenters, namely, the discourse segmenters of the HILDA (Hernault et al. 2010) and SPADE (Soricut and Marcu 2003) systems. We also compare our results with the stateof-the-art results reported by Fisher and Roark (2007) on the RST–DT test set. In all our experiments when comparing two systems, we use paired t-test on the F-scores to measure statistical significance and report the p-value.
We ran HILDA with its default settings. For SPADE, we applied the same modifications to its default settings as described in Fisher and Roark (2007), which delivers significantly improved performance over its original version. Specifically, in our experiments on the RST–DT corpus, we trained SPADE using the human-annotated syntactic trees extracted from the Penn Treebank (Marcus, Marcinkiewicz, and Santorini 1994), and, during testing, we replaced the Charniak parser (Charniak 2000) with a more
12 Available from alt.qcri.org/tools/.
accurate reranking parser (Charniak and Johnson 2005). However, because of the lack of gold syntactic trees in the instructional corpus, we trained SPADE in this corpus using the syntactic trees produced by the reranking parser. To avoid using the gold syntactic trees, we used the reranking parser in our system for both training and testing purposes. This syntactic parser was trained on the sections of the Penn Treebank not included in our test set. We applied the same canonical lexical head projection rules (Magerman 1995; Collins 2003) to lexicalize the syntactic trees as done in HILDA and SPADE.
Note that previous studies (Fisher and Roark 2007; Soricut and Marcu 2003; Hernault et al. 2010) on discourse segmentation only report their performance on the RST–DT test set. To compare our results with them, we evaluated our model on the RST–DT test set. In addition, we showed a more general performance of SPADE and our system on the two corpora based on 10-fold cross validation.13 However, SPADE does not come with a training module for its segmenter. We reimplemented this module and verified its correctness by reproducing the results on the RST–DT test set.
6.3.2 Results for Discourse Segmentation. Table 3 shows the discourse segmentation results of different systems in Precision, Recall, and F-score on the two corpora. On the RST– DT corpus, HILDA’s segmenter delivers the weakest performance, having an F-score of only 74.1. Note that the high segmentation accuracy reported by Hernault et al. (2010) is due to a less stringent evaluation metric. SPADE performs much better than HILDA with an absolute F-score improvement of 11.1%. Our segmenter DS outperforms SPADE with an absolute F-score improvement of 4.9% (p-value < 2.4e-06), and also achieves comparable results to the ones of Fisher and Roark (2007) (F&R), even though we use fewer features.14 Notice that human agreement for this task is quite high—namely, an F-score of 98.3 computed on the doubly-annotated portion of the RST–DT corpus mentioned in Section 6.1.
Because Fisher and Roark (2007) only report their results on the RST–DT test set and we did not have access to their system, we compare our approach only with SPADE when evaluating on a whole corpus based on 10-fold cross validation. On the RST–DT corpus, our segmenter delivers an absolute F-score improvement of 3.8 percentage points, which represents a more than 25% relative error rate reduction.
13 Because the two tasks—discourse segmentation and intra-sentential parsing—operate at the sentence level, the cross validation was performed over sentences for their evaluation. 14 Because we did not have access to the system or to the complete output/results of Fisher and Roark (2007), we were not able to perform a statistical significance test.
The improvement is higher on the instructional corpus with an absolute F-score improvement of 8.1 percentage points, which corresponds to a relative error reduction of 30%. The improvements for both corpora are statistically significant (p-value < 3.0e-06). When we compare our results on the two corpora, we observe a substantial decrease in performance on the instructional corpus. This could be because of a smaller amount of data in this corpus and/or to the inaccuracies in the syntactic parser and taggers, which are trained on news articles. A promising future direction would be to apply effective domain adaptation methods (e.g., easyadapt [Daumé 2007]) to improve discourse segmentation performance in the instructional domain by leveraging the rich data in the news domain (i.e., RST–DT). In this section we present our experiments on discourse parsing. First, we describe the experimental set-up. Then, we present the results of the parsers. While presenting the performance of our discourse parser, we show a breakdown of intra-sentential versus inter-sentential results, in addition to the aggregated results at the document level.
6.4.1 Experimental Set-up for Discourse Parsing. In our experiments on sentence-level (i.e., intra-sentential) discourse parsing, we compare our approach with SPADE (Soricut and Marcu 2003) on the RST–DT corpus, and with the ILP-based approach of Subba and Di-Eugenio (2009) on the instructional corpus, because they are the state of the art in their respective genres. For SPADE, we applied the same modifications to its default settings as described in Section 6.3.1, which leads to improved performance. Similarly, in our experiments on document-level (i.e., multi-sentential) parsing, we compare our approach with HILDA (Hernault et al. 2010) on the RST–DT corpus, and with the ILPbased approach (Subba and Di-Eugenio 2009) on the instructional corpus. The results for HILDA were obtained by running the system with default settings on the same inputs we provided to our system. Because we could not run the ILP-based system (not publicly available), we report the performance presented in their paper.
Our experiments on the RST–DT corpus use the same 18 coarser coherence relations (see Figure 18 later in this article), defined by Carlson and Marcu (2001) and also used in SPADE and HILDA systems. More specifically, the relation set consists of 16 relation categories and two pseudo-relations, namely, Textual–Organization and Same–Unit. After attaching the nuclearity statuses (NS, SN, NN) to these relations, we obtain 41 distinct relations.15 Our experiments on the instructional corpus consider the same 26 primary relations (e.g., Goal:Act, Cause:Effect) used by Subba and Di-Eugenio (2009) and also treat the reversals of non-commutative relations as separate relations. That is, Goal–Act and Act–Goal are considered to be two different coherence relations. Attaching the nuclearity statuses to these relations provides 76 distinct relations.
Based on our experiments on the development set, the size of the automatically built bi-gram and tri-gram dictionaries was set to 95% of their total number of items, and the size of the unigram dictionary was set to 100%. Note that the unigram dictionary contains only special tags denoting EDU, sentence, and paragraph boundaries.
15 Not all relations take all the possible nuclearity statuses. For example, Elaboration and Attribution are mono-nuclear relations, and Same–Unit and Joint are multi-nuclear relations.
6.4.2 Evaluation of the Intra-Sentential Discourse Parser. This section presents our experimental evaluation on intra-sentential discourse parsing. First, we show the performance of the sentence-level parsers when they are provided with manual (or gold) discourse segmentations. This allows us to judge the parsing performance independently of the segmentation task. Then, we show the end-to-end performance of our intra-sentential framework, that is, the intra-sentential parsing performance based on automatic discourse segmentation.
Intra-sentential parsing results based on manual segmentation Table 4 presents the intra-sentential discourse parsing results when manual discourse segmentation is used. Recall from our discussion on evaluation metrics in Section 6.2.2 that precision, recall, and F-score are the same when manual segmentation is used. Therefore, we report only one of them. Notice that our sentence-level discourse parser PAR-S consistently outperforms SPADE on the RST–DT test set in all three metrics, and the improvements are statistically significant (p-value< 0.01). Especially, on the relation labeling task, which is the hardest among the three tasks, we achieve an absolute F-score improvement of 12.2 percentage points, which represents a relative error rate reduction of 37.7%.
To verify our claim that capturing the sequential dependencies between DT constituents using a DCRF model actually contributes to the performance gain, we also compare our results with an intra-sentential parser (see CRF-NC in Table 4) that uses a simplified CRF model similar to the one shown in Figure 8. Although the simplified model has two hidden variables to model the structure and the relation of a DT constituent jointly, it does not have a chain structure, thus it ignores the sequential dependencies between DT constituents. The comparison in all three measures demonstrates that the improvements are indeed partly due to the DCRF model (p–value < 0.01).16 A comparison between CRF-NC and SPADE shows that CRF-NC significantly outperforms SPADE in all three measures (p-value < 0.01). This could be due to the fact that CRF-NC is trained discriminatively with a large number of features, whereas SPADE is trained generatively with only lexico-syntactic features.
Notice that the scores of our parser (PAR-S) are close to the human agreement on the doubly-annotated data, and these results on the RST–DT test set are also consistent with the mean scores over 10-folds on the whole RST–DT corpus.17
The improvements are higher on the instructional corpus, where we compare our mean results over 10-folds with the reported results of the ILP-based system of Subba and Di-Eugenio (2009), giving absolute F-score improvements of 5.4 percentage points, 17.6 percentage points, and 12.8 percentage points in span, nuclearity, and relations, respectively.18 Our parser PAR-S reduces the errors by 76.1%, 62.4%, and 34.6% in span, nuclearity and relations, respectively.
If we compare the performance of our intra-sentential discourse parser on the two corpora, we notice that our parser PAR-S is more accurate in finding the right tree
16 The parsing performance reported in Table 4 for CRF-NC is when the CRF parsing model is trained on a balanced data set (an equal number of instances with S=1 and S=0); Training on full but imbalanced data set gives slightly lower results. 17 Our EMNLP and ACL publications (Joty, Carenini, and Ng 2012; Joty et al. 2013) reported slightly lower parsing accuracies. Fixing a bug in the parsing algorithm accounts for the difference. 18 Subba and Di-Eugenio (2009) report their results based on an arbitrary split between training and test sets. Because we did not have access to their particular split, we compare our model’s performance based on 10-fold cross validation with their reported results. Also, because we did not have access to their system/output, we could not perform a significance test on the instructional corpus.
structure (see Span row in the table) on the instructional corpus. This may be due to the fact that sentences in the instructional domain are relatively short and contain fewer EDUs than sentences in the news domain, thus making it easier to find the right tree structure. However, when we compare the performance on the relation labeling task, we observe a decrease on the instructional corpus. This may be due to the small amount of data available for training and the imbalanced distribution of a large number of discourse relations (i.e., 76 with nuclearity attached) in this corpus.
Intra-sentential parsing results based on automatic segmentation In order to evaluate the performance of the fully automatic sentence-level discourse analysis systems, we feed the intra-sentential discourse parsers the output of their respective discourse segmenters. Table 5 shows the (P)recision, (R)ecall, and (F)–score results for different evaluation metrics. We compare our intra-sentential parser PAR-S with SPADE on the RST–DT test set. We achieve absolute F-score improvements of 5.7 percentage points, 6.4 percentage points, and 9.5 percentage points in span, nuclearity, and relation, respectively. These improvements are statistically significant (pvalue<0.001). Our system, therefore, reduces the errors by 24.5%, 21.4%, and 22.6% in span, nuclearity, and relations, respectively. These results are also consistent with the mean results over 10-folds on the whole RST–DT corpus.
The rightmost column in the table shows our mean results over 10-folds on the instructional corpus. We could not compare our system with the ILP-based approach of Subba and Di-Eugenio (2009) because no results were reported using an automatic segmenter. It is interesting to observe how much our parser is affected by an automatic segmenter on the two corpora (see Tables 4 and 5). Nevertheless, taking into account the
segmentation results in Table 3, this is not surprising because previous studies (Soricut and Marcu 2003) have already shown that automatic segmentation is the primary impediment to high accuracy discourse parsing. This demonstrates the need for a more accurate discourse segmentation model in the instructional genre.
6.4.3 Evaluation of the Complete Parser. We experiment with our full document-level discourse parser on the two corpora using the two parsing approaches described in Section 4.3, namely, 1S-1S and the sliding window. On RST–DT, the standard split was used for training and testing. On the instructional corpus, Subba and Di-Eugenio (2009) used 151 documents for training and 25 documents for testing. Because we did not have access to their particular split, we took five random samples of 151 documents for training and 25 documents for testing, and report the average performance over the five test sets.
Table 6 presents results for our two-stage discourse parser (TSP) using approaches 1S-1S (TSP 1-1) and the sliding window (TSP SW) on manually segmented texts. Recall that precision, recall, and F-score are the same when manual segmentation is used. We compare our parser with the state-of-the-art on the two corpora: HILDA (Hernault et al. 2010) on RST–DT, and the ILP-based approach (Subba and Di-Eugenio 2009) on the instructional domain. On both corpora, our systems outperform existing systems by a wide margin (p-value <7.1e-05 on RST–DT).19 On RST–DT, our parser TSP 1-1 achieves absolute improvements of 7.9 percentage points, 9.3 percentage points, and 11.5 percentage points in span, nuclearity, and relation, respectively, over HILDA. This represents relative error reductions of 31.2%, 22.7%, and 20.7% in span, nuclearity, and relation, respectively.
Beside HILDA, we also compare our results with two baseline parsers on RST–DT: (1) CRF-O, which uses a single unified CRF-based parsing model shown in Figure 8 (the one used for multi-sentential parsing) without distinguishing between intra- and multisentential parsing, and (2) CRF-T, which uses two different CRF-based parsing models for intra- and multi-sentential parsing in the two-stage approach 1S-1S, both models having the same structure as in Figure 8. Thus, CRF-T is a variation of TSP 1-1, where the DCRF-based (chain-structured) intra-sentential parsing model is replaced with a simpler CRF-based parsing model.20 Note that although CRF-O does not explicitly
19 Because we did not have access to the output or to the system of Subba and Di-Eugenio (2009), we were not able to perform a significance test on the instructional corpus. 20 The performance of this model for intra-sentential parsing is reported in Table 4 under the name CRF-NC.
discriminate between intra- and multi-sentential parsing, it uses N-gram features that include sentence and EDU boundaries to encode this information into the model.
Table 6 shows that both CRF-O and CRF-T outperform HILDA by a good margin (p-value <0.0001). This improvement can be attributed to the optimal parsing algorithm and better feature selection strategy. When we compare CRF-T with CRF-O, we notice significant performance gains for CRF-T (p-value <0.001). The absolute gains are 4.32 percentage points, 2.68 percentage points, and 4.55 percentage points in span, nuclearity, and relation, respectively. This comparison clearly demonstrates the benefit of using a two-stage approach with two different parsing models over a framework with one single unified parsing model. Finally, when we compare our best results with the human agreements, we still observe room for further improvement in all three measures.
On the instructional genre, our parser TSP 1-1 delivers absolute F-score improvements of 10.3 percentage points, 13.6 percentage points, and 8.1 percentage points in span, nuclearity, and relations, respectively, over the ILP-based approach of Subba and Di-Eugenio (2009). Our parser, therefore, reduces errors by 34.7%, 26.9%, and 12.5% in span, nuclearity, and relations, respectively.
If we compare the performance of our discourse parsers on the two corpora, we observe lower results on the instructional corpus. There could be two reasons for this. First, the instructional corpus has a smaller amount of data with a larger set of relations (76 with nuclearity attached). Second, some of the frequent relations are semantically very similar (e.g., Preparation-Act, Step1-Step2), which makes it difficult even for the human annotators to distinguish them (Subba and Di-Eugenio 2009).
Comparison between our two document-level parsing approaches reveals that TSP SW significantly outperforms TSP 1-1 only in finding the right structure on both corpora (p-value <0.01). Not surprisingly, the improvement is higher on the instructional corpus. A likely explanation is that the instructional corpus contains more leaky boundaries (12%), allowing the sliding window approach to be more effective in finding those, without inducing much noise for the labels. This demonstrates the potential of TSP SW for data sets with even more leaky boundaries, e.g., the Dutch (Vliet and Redeker 2011) and the German Potsdam (Stede 2004) corpora. However, it would be interesting to see how other heuristics to do consolidation in the cross condition (Section 4.3.2) perform.
To analyze errors made by TSP SW, we looked at some poorly parsed examples and found that although TSP SW finds more correct structures, a corresponding improvement in labeling relations is not present because in some cases, it tends to induce noise from the neighboring sentences for the labels. For example, when parsing is performed on the first sentence in Figure 1 in isolation using 1S-1S, our parser rightly identifies the Contrast relation between EDUs 2 and 3. But, when it is considered with its neighboring sentences by the sliding window, the parser labels it as Elaboration. A promising strategy to deal with this and similar problems would be to apply both approaches to each sentence and combine them by consolidating three probabilistic decisions, namely, the one from 1S-1S and the two from the sliding window.
6.4.4 k-best Parsing Results Based on Manual Segmentation. As described in Section 4.2, a straight-forward modification of our probabilistic parsing algorithm allows us to generate a list of k-best parse hypotheses for a given text. We adapt our parsing algorithm accordingly to produce k most probable DTs for each text, and measure the oracle accuracy based on the F-scores of the Relation metric which gives aggregated evaluation
on structure and relation labels (see Table 2). Specifically, the oracle accuracy O−score for k-best discourse parsing is measured as follows:
O−score = ∑N i=1 max k j=1 F−scorer(gi, h j i)
N (15)
where N is the total number of texts (sentences or documents) evaluated, gi is the gold DT annotation for text i, hji is the j
th parse hypothesis generated by the parser for text i, and F-scorer(gi, h j i) is the F-score accuracy of hypothesis h j i on the Relation metric, which essentially measures how similar hji is to gi in terms of its structure and labels. Table 7 presents the oracle scores of our intra-sentential parser PAR-S on the RST– DT test set as a function of k of k-best parsing. The 1-best result tells that the parser has the base accuracy of 79.8%. The 2-best shows dramatic oracle-rate improvements (i.e., 4.65% absolute), meaning that often our parser generates the best tree as its top two outputs. 3-best and 4-best also show moderate improvements (about 2%). Things start to slow down afterwards, and we achieve oracle rates of 90.37% and 92.57% at 10−best and 20-best, respectively. The 30-best parsing gives an oracle score of 93.2%.
The results of our k-best intra-sentential discourse parser demonstrate that a k-best reranking approach like that of Collins and Koo (2005) and Charniak and Johnson (2005) used for syntactic parsing can potentially improve the parsing accuracy even further by exploiting additional global features of the candidate discourse trees as evidence.
The scenario is quite different at the document-level; Table 8 shows the k-best parsing results of TSP 1S-1S on the RST–DT test set. The improvements in oracle-rate are small at the document-level when compared with the sentence-level parsing. For example, the 2-best and the 5-best improve over the base accuracy by only 0.7 percentage points and 1.0 percentage points, respectively. The improvements get even slower after that. However, this is not surprising because generally document-level DTs are big with many constituents, and only a very few of these constituents change from k-best to k + 1-best parsing. These small changes among the candidate DTs do not contribute much to the overall F-score accuracy (for further clarification see how F-score is calculated in Section 6.2.2).
The results of our k-best document-level parsing suggest that often the best tree is missing in the top k parses. Thus, a reranking of k-best document-level parses may not be a suitable option for further improvement at the document-level. An alternative to k-best reranking is to use a sampling-based parsing strategy (Wick et al. 2011) to explore the space of possible trees, as recently used for dependency parsing (Zhang et al. 2014). However, note that the potential gain we may obtain by using a reranker at the sentence level will also improve the (combined) accuracy of the document-level parser.
6.4.5 Analysis of Features. To analyze the relative importance of different features used in our parsing models, Table 9 presents the sentence- and document-level parsing results on a manually segmented RST–DT test set using different subsets of features. The feature subsets were defined in Section 4.1.4. In each parsing condition, the subsets of features are added incrementally, based on their availability and historical importance. The columns in Table 9 represent the inclusion order of the feature subsets.
Because SPADE (Soricut and Marcu 2003) achieved the previous best results on intra-sentential parsing using Dominance set features, these are included as the initial set of features in our intra-sentential parsing model. In HILDA, Hernault et al. (2010) demonstrate the importance of Organizational and N-gram features for full text parsing. We add these two feature subsets one after another in our intra- and multi-sentential parsing models.21 Contextual features require other features to be computed; thus they were added after those features. Because computation of Sub-structural features requires an initial parse tree (i.e., when the parser is applied), they are added at the very end.
Notice that inclusion of every new subset of features appears to improve the performance over the previous set. Specifically, for sentence-level parsing, when we add the Organizational features with the Dominance set features, we achieve about 2 percentage points absolute improvements in nuclearity and relations. With N-gram features, the gain is even higher: 6 percentage points in relations and 3.5 percentage points in nuclearity for sentence-level parsing, and 3.8 percentage points in relations and 3.1 percentage points in nuclearity for document-level parsing. This demonstrates the utility of the N-gram features, which is also consistent with the previous findings of duVerle and Prendinger (2009) and Schilder (2002).
The features extracted from Lexical chains (L-ch) have also proved to be useful for document-level parsing. They deliver absolute improvements of 2.7 percentage points, 2.9 percentage points, and 2.3 percentage points in span, nuclearity, and relations,
21 Text structural features are included in the Organizational features for multi-sentential parsing.
respectively. Including the Contextual features further gives improvements of 3 percentage points in nuclearity and 2.2 percentage points in relation for sentence-level parsing, and 1.3 percentage points in nuclearity, and 1.2 percentage points in relation for document-level parsing. Notice that Sub-structural features are more beneficial for document-level parsing than they are for sentence-level parsing, that is, an improvement of 2.2 percentage points versus an improvement of 0.9 percentage points. This is not surprising because document-level DTs are generally much larger than sentencelevel DTs, making the sub-structural features more effective for document-level parsing.
6.4.6 Error Analysis. We further analyze the errors made by our discourse parser. As described in previous sections, the parser could be wrong in finding the right structure as well as the right nuclearity and relation labels. Figure 17 presents an example where our parser makes mistakes in finding the right structure (notice the units connected by Attribution and Cause in the two example DTs) and the right relation label (TopicComment vs. Background). The comparison between intra- and multi-sentential parsing results presented in Sections 6.4.2 and 6.4.3 tells us that the errors in structure occur more frequently when the DT is large (e.g., at the document level) and the parsing model fails to capture the long-range structural dependencies between the DT constituents.
To further analyze the errors made by our parser on the hardest task of relation labeling, in Figure 18 we present the confusion matrix for our document-level parser TSP 1-1 on the RST–DT test set. In order to judge independently the ability of our parser to assign the correct relation labels, the confusion matrix is computed based on the constituents (see Table 2), where our parser found the right span (i.e., structure).22 The relations in the matrix are ordered according to their frequency in the training set.
In general, the errors can be explained by two different causes acting together: (1) imbalanced distribution of the relations in the corpus, and (2) semantic similarity between the relations. The most frequent relation Elaboration tends to overshadow others, especially the ones that are semantically similar (e.g., Explanation, Background)
22 Therefore, the counts of the relations shown in the table may not match the ones in the test set.
T-C T-O T-CM M-M CMP EV SU CND EN CA TE EX BA CO JO S-U AT EL
and less frequent (e.g., Summary, Evaluation). Furthermore, our models sometimes fail to distinguish relations that are semantically similar (e.g., Temporal vs. Background, Cause vs. Explanation).
Now, let us look more closely at a few of these errors. Figure 19 presents an example where our parser mistakenly labels a Summary as Elaboration. Clearly, in this example the text in parentheses (i.e., (CFD)) is an acronym or summary of the text to the left. However, parenthesized texts are also used to provide additional information (i.e., to elaborate), as exemplified in Figure 20 by two text snippets from the RST-DT. Notice that although the structure of the text (widow of the ..) in the first example is quite distinguishable from the structure of (CFD), the text (D., Maine) in the second example is similar to (CFD) in structure, thus it confuses our model.23
Figure 21 presents two examples where our parser mistakenly labels Background and Cause as Elaboration. However, notice that the two discourse relations (i.e., Background
23 D., Maine in this example refers to Democrat from state Maine.
vs. Elaboration and Cause vs. Elaboration) in these examples are semantically very close, and arguably both can be applicable.
Given these observations, we see two possible ways to improve our system. First, we would like to use a more robust method (e.g., ensemble methods with bagging) to deal with the imbalanced distribution of relations, along with taking advantage of richer semantic knowledge (e.g., compositional semantics) to cope with the errors caused by semantic similarity between the relations. Second, to capture long-range dependencies between DT constituents, we would like to explore the idea of k-best discriminative reranking using tree kernels (Dinarelli, Moschitti, and Riccardi 2011). Because our parser already produces k most probable DTs, developing a reranker based on discourse tree kernels is very much within our reach. 7 conclusions and future directions :In this article we have presented CODRA, a complete probabilistic discriminative framework for performing rhetorical analysis in the RST framework. CODRA comprises components for performing both discourse segmentation and discourse parsing. The discourse segmenter is a binary classifier based on a maximum entropy model, and the discourse parser applies an optimal parsing algorithm to probabilities inferred from two CRF models: one for intra-sentential parsing and the other for multi-sentential parsing. The CRF models effectively represent the structure and the label of discourse tree constituents jointly. Furthermore, the DCRF model for intra-sentential parsing captures the sequential dependencies between the constituents. The two separate parsing models use their own informative feature sets and the distributional variations of the relation labels in their respective parsing conditions.
We have also presented two approaches to effectively combine the intra-sentential and the multi-sentential parsing modules, which can exploit the strong correlation observed between the text structure and the structure of the discourse tree. The first approach (1S–1S) builds a DT for every sentence using the intra-sentential parser, and then
runs the multi-sentential parser on the resulting sentence-level DTs. To deal with leaky boundaries, our second approach (sliding window) builds sentence-level discourse sub-trees by applying the intra-sentential parser on a sliding window, covering two adjacent sentences and then consolidating the results produced by overlapping windows. After that, the multi-sentential parser takes all these sentence-level sub-trees and builds a full rhetorical parse for the whole document.
Finally, we have extended the parsing algorithm to generate k most probable parse hypotheses for each input text, which could be used in a reranker to improve over the initial ranking using global features like long-range structural dependencies.
Empirical evaluations on two different genres demonstrate that our approach to discourse segmentation achieves state-of-the-art performance more efficiently using fewer features. A series of experiments on the discourse parsing task shows that both our intra- and multi-sentential parsers significantly outperform the state of the art, often by a wide margin. A comparison between our combination strategies reveals that the sliding window approach is more robust across domains. Furthermore, the oracle accuracy computed based on the k-best parse hypotheses generated by our parser demonstrates that a reranker could potentially improve the accuracy even further.
Our error analysis reveals that although the sliding window approach finds more correct tree structures, in some cases it induces noise for the relation labels from the neighboring sentences. With respect to the performance of our discourse parser on the relation labeling task we also found that the most frequent relations tend to mislead the identification of the less frequent ones, and the models sometimes fail to distinguish relations that are semantically similar.
The work presented in this article leads us to several interesting future directions. Our short-term goal is to develop a k-best discriminative reranking discourse parser using tree kernels applied to discourse trees. We also plan to investigate to what extent discourse segmentation and discourse parsing can be performed jointly.
We would also like to explore how our system performs on other genres like conversational (e.g., blogs, e-mails) and evaluative (e.g., customer reviews) texts. To address the problem of limited annotated data in various genres, we are planning to develop an interactive version of our system that will allow users to fix the output of the system with minimal effort and let the system learn from that feedback.
Another interesting future direction is to perform extrinsic evaluations of our system in downstream applications. One important application of rhetorical structure is text summarization, where a significant challenge is producing not only informative but also coherent summaries. A number of researchers have already investigated the utility of rhetorical structure for measuring text importance (i.e., informativeness) in summarization (Marcu 2000b; Daumé and Marcu 2002; Louis, Joshi, and Nenkova 2010). Recently, Christensen et al. (2013, 2014) propose to perform sentence selection and ordering at the same time, and use constraints on discourse structure to make the summaries coherent. However, they represent the discourse as an unweighted directed graph, which is shallow and not sufficiently informative in most cases. Furthermore, their approach does not allow compression at the sentence level, which is often beneficial in summarization. In the future, we would like to investigate the utility of our rhetorical structure for performing sentence compression, selection, and ordering in a joint summarization process.
Discourse structure can also play important roles in sentiment analysis. A key research problem in sentiment analysis is extracting fine-grained opinions about different aspects of a product. Several recent papers (Somasundaran 2010; Lazaridou, Titov, and
Sporleder 2013) exploited the rhetorical structure for this task. Another challenging problem is assessing the overall opinion expressed in a review because not all sentences in a review contribute equally to the overall sentiment. For example, some sentences are subjective, whereas others are objective (Pang and Lee 2004); some express the main claims, whereas others support them (Taboada et al. 2011); some express opinions about the main entity, whereas others are about the peripherals. Discourse structure could be useful to capture the relative weights of the discourse units towards the overall sentiment. For example, the nucleus and satellite distinction along with the rhetorical relations could be useful to infer the relative weights of the connecting discourse units.
Among other applications of discourse structure, Machine Translation (MT) and its evaluation have received a resurgence of interest recently. A workshop dedicated to discourse in machine translation was arranged recently at the ACL 2013 conference (Webber et al. 2013). Researchers believe that MT systems should consider discourse phenomena that go beyond the current sentence to ensure consistency in the choice of lexical items or referring expressions, and the fact that source-language coherence relations are also realized in the target language (i.e., translating at the documentlevel [Hardmeier, Nivre, and Tiedemann 2012]). Guzmán et al. (2014a, 2014b) and Joty et al. (2014) propose new discourse-aware automatic evaluation metrics for MT systems using our discourse analysis tool. They demonstrate that sentence-level discourse information is complementary to the state-of-the-art evaluation metrics, and by combining the discourse-based metrics with the metrics from the ASIYA MT evaluation toolkit (Giménez and Màrquez 2010), they won the WMT 2014 metrics shared task challenge (Macháček and Bojar 2014) both at the segment- and at the system-level. These results suggest that discourse structure helps to distinguish better translations from worse ones. Thus, it would be interesting to explore whether discourse information can be used to rerank alternative MT hypotheses as a post-processing step for the MT output.
A longer-term goal is to extend our framework to also work with graph structures of discourse, as recommended by several recent discourse theories (Wolf and Gibson 2005). Once we achieve similar performance on graph structures, we will perform extrinsic evaluations to determine their relative utility for various NLP tasks.
Finally, we hope that the online demo, the source code of CODRA, and the evaluation metrics that we made publicly available in this work will facilitate other researchers in extending our work and in applying discourse parsing to their NLP tasks.
Bibliographic Note
Portions of this work were previously published in two conference proceedings (Joty, Carenini, and Ng 2012; Joty et al. 2013). This article significantly extends our previous work in several ways, most notably: (1) we extend the parsing algorithm to generate k-most probable parse hypotheses for each input text (Section 4.2); (2) we show the oracle accuracies for k-best discourse parsing both at the sentence level and at the document level (Section 6.4.4); (3) to support our claim, we compare our best results with several variations of our approach (see CRF-NC in Section 6.4.2, and CRF-O and CRF-T in Section 6.4.3); (4) we analyze the relative importance of different features for intra- and multi-sentential discourse parsing (Section 6.4.5); and (5) we perform indepth error analysis of our complete rhetorical analysis framework (Section 6.4.6).
Appendix Sample Output Generated by an Online Demo of CODRA
A."
"1 introduction :Performing textual inference is at the heart of many semantic inference applications, such as Question Answering (QA) and Information Extraction (IE). A prominent generic
∗ Computer Science Department, Stanford University, Stanford, CA 94305. Web: http://www.stanford.edu/∼joberant.
∗∗ Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv, 6997801, Israel. Web: http://www.tau.ac.il/∼nogaa. † Computer Science Department, Bar-Ilan University, Ramat-Gan, 52900, Israel.
Web: http://www.cs.biu.ac.il/∼dagan. ‡ Engineering Department, Bar-Ilan University, Ramat-Gan, 52900, Israel.
Web: http://www.eng.biu.ac.il/goldbej.
Submission received: 23 November 2013; revised version received: 26 September 2014; accepted for publication: 25 November 2014.
doi:10.1162/COLI a 00220
© 2015 Association for Computational Linguistics
paradigm for textual inference is Textual Entailment (Dagan et al. 2013). In textual entailment, the goal is to recognize, given two text fragments termed text and hypothesis, whether the hypothesis can be inferred from the text. For example, the text Cyprus was invaded by the Ottoman Empire in 1571 implies the hypothesis The Ottomans attacked Cyprus.
Semantic inference applications such as QA and IE crucially rely on entailment rules (Ravichandran and Hovy 2002; Shinyama and Sekine 2006; Berant and Liang 2014) or equivalently, inference rules—that is, rules that describe a directional inference relation between two fragments of text. An important type of entailment rule specifies the entailment relation between natural language predicates, for example, the entailment rule X invade Y⇒X attack Y can be helpful in inferring the aforementioned hypothesis. Consequently, substantial effort has been made to learn such rules (Lin and Pantel 2001; Sekine 2005; Szpektor and Dagan 2008; Schoenmackers et al. 2010; Melamud et al. 2013).
Textual entailment is inherently a transitive relation, that is, the rules x⇒ y and y⇒ z imply the rule x⇒ z. For example, from the rules X reduce nausea⇒ X help with nausea and X help with nausea⇒ X associated with nausea we can infer the rule X reduce nausea⇒ X associated with nausea (Figure 1). Accordingly, Berant, Dagan, and Goldberger (2012) proposed taking advantage of transitivity to improve learning of entailment rules. They modeled learning entailment rules as a graph optimization problem, where nodes are predicates and edges represent entailment rules that respect transitivity. To solve this optimization problem, they formulated it as an Integer Linear Program (ILP) and used an off-the-shelf ILP solver to find an exact solution. Indeed, they showed that applying global transitivity constraints results in more accurate graphs (known as entailment graphs) than methods that ignore the property of transitivity.
Although using an off-the-shelf ILP solver is straightforward, finding the optimal set of edges respecting transitivity is NP-hard, and practically, transitivity constraints impose substantial restrictions on the scalability of the methods. In fact, in some cases finding the optimal set of edges for entailment graphs with even ∼50 nodes was quite slow.
In this article, we develop algorithms for learning entailment graphs that take advantage of structural properties such as transitivity, but substantially reduce the computational cost of inference, thus allowing us to solve the aforementioned optimization problem on very large graphs.
Our method contains two main steps. The first step is based on a sparsity assumption that even if an entailment graph contains many predicates, most of them do not entail one another, and thus we can decompose the graph into smaller components that can be solved more efficiently. For example, the predicates X parent of Y, X child of Y and X relative of Y are independent from the predicates X works with Y, X boss of Y, and X manages Y, and thus we can consider each one of these two sets separately. We prove that finding the optimal solution for each of the smaller components results in a global optimal solution for our optimization problem.
The second step proposes a polynomial heuristic approximation algorithm for finding a transitive set of edges in each one of the smaller components. It is based on a novel modeling assumption that entailment graphs exhibit a “tree-like” property, which we term forest reducible. For example, the graph in Figure 1 is not a tree, because the predicates X related to nausea and X associated with nausea form a cycle. However, these two predicates are synonymous, and if they were merged into a single node, then the graph would become a tree. We propose a simple iterative approximation algorithm, where in each iteration a single node is deleted from the graph and then inserted back in a way that improves the objective function value. We prove that if we impose a constraint that entailment graphs must be forest reducible, then each iteration can be performed in linear time. This results in an algorithm that can scale to entailment graphs containing tens of thousands of nodes.
We apply our algorithm on two data sets. The first data set includes medium-sized entailment graphs where predicates are typed, that is, the arguments are restricted to belong to a particular semantic type (for instance, Xperson parent of Yperson). We show that using our algorithm, we can substantially improve runtime, suffering from only a slight reduction in the quality of learned entailment graphs. The second data set includes a much larger graph containing 20,000 untyped nodes, where applying state-of-the-art methods that use an ILP solver is completely impossible. We run our algorithm on this data set and demonstrate that we can learn knowledge bases with more than 100,000 entailment rules at a higher precision compared with local learning methods.
The article is organized as follows. In Section 2 we survey prior work on the learning of entailment rules. We first focus on local methods (Section 2.1), that is, methods that handle each pair of predicates independently of other predicates, and then describe global methods (Section 2.2), that is, methods that take into account multiple predicates simultaneously. Section 3 is the core of the article and describes our main algorithmic contributions. After formalizing entailment rule learning as a graph optimization problem, we present the two steps of our algorithm. Section 3.1 describes the first step, in which a large entailment graph is decomposed into smaller components. Section 3.2 describes the second step, in which we develop an efficient heuristic approximation based on the assumption that entailment graphs are forest reducible. Section 4 describes experiments on the first data set containing medium-sized entailment graphs with typed predicates. Section 5 presents an empirical evaluation on a large graph containing 20,000 untyped predicates. We also perform a qualitative analysis in Section 5.4 to further elucidate the behavior of our algorithm. Section 6 concludes the article.
This article is based on previous work (Berant, Dagan, and Goldberger 2011; Berant et al. 2012), but expands over it in multiple directions. Empirically, we present results on a novel data set (Section 5) that is by orders of magnitude larger than in the past. Algorithmically, we present the Tree-Node-And-Component-Fix algorithm, which is an extension of the Tree-Node-Fix algorithm presented in Berant et al. (2012) and achieves best results in our experimental evaluation. Theoretically, we provide an NP-hardness
proof for the Max-Trans-Forest optimization problem presented in Berant et al. (2012) and an ILP formulation for it. Last, we perform in Section 5.4 an extensive qualitative analysis of the graphs learned by our local and global algorithms from which we draw conclusions for future research directions.","2 background :In this section we describe prior work relevant for entailment rule learning. First, we describe local methods that estimate entailment for a pair of predicates, focusing on methods that we employ in this article. Then, we describe global methods that perform inference over a larger set of predicates. Specifically, we provide details on the method and optimization problem developed by Berant, Dagan, and Goldberger (2012) for which we propose scalable inference algorithms in this article. Last, we survey some other related work in NLP that uses global inference. In local learning, given a pair of predicates (i, j) we would like to determine whether i⇒ j. The main sources of information utilized in the past for local learning were (1) lexicographic resources, (2) co-occurrence, and (3) distributional similarity. We briefly describe the first two and then expand more on distributional similarity, which is the most commonly used source of information.
Lexicographic resources are manually built knowledge bases from which semantic information may be extracted. For example, the hyponymy, toponymy, and synonymy relations in WordNet (Fellbaum 1998) can be used to detect entailment between nouns and verbs. Although WordNet is the most popular lexicographic resource, other resources such as CatVar, Nomlex, and FrameNet have also been utilized to extract inference rules (Meyers et al. 2004; Budanitsky and Hirst 2006; Szpektor and Dagan 2009; Coyne and Rambow 2009; Ben Aharon, Szpektor, and Dagan 2010).
Pattern-based methods attempt to identify the semantic relation between a pair of predicates by examining their co-occurrence in a large corpus. For example, the sentence people snore while they sleep provides evidence that snore ⇒ sleep. While most patternbased methods focused on identifying semantic relations between nouns (for example, Hearst patterns [Hearst 1992]), several works (Pekar 2008; Chambers and Jurafsky 2011; Weisman et al. 2012) attempted to extract relations between predicates as well. Chklovsky and Pantel (2004) used pattern-based methods to generate the commonly used VerbOcean resource.
Both lexicographic as well as pattern-based methods suffer from limited coverage. Distributional similarity is therefore used to automatically construct broad-scale resources. Distributional similarity methods are based on the “distributional hypothesis” (Harris 1954) that semantically similar predicates occur with similar arguments. Quite a few methods have been suggested (Lin and Pantel 2001; Szpektor et al. 2004; Bhagat, Pantel, and Hovy 2007; Szpektor and Dagan 2008; Yates and Etzioni 2009; Schoenmackers et al. 2010), which differ in terms of the specifics of the ways in which predicates are represented, the features that are extracted, and the function used to compute feature vector similarity. Next, we elaborate on the methods that we use in this article.
Lin and Pantel (2001) proposed an algorithm for learning paraphrase relations between binary predicates, that is, predicates with two variables such as X treat Y. For each binary predicate, Lin and Pantel compute two sets of features Fx and Fy,
which are the words that instantiate the arguments X and Y, respectively, in a large corpus. Given a predicate u and its feature set for the X variable Fx, every feature fx ∈ Fx is weighted by pointwise mutual information between the predicate and the feature: w( fx) = log Pr( fx|u) Pr( fx )
, where the probabilities are computed using maximum likelihood over the corpus. Given two predicates u and v, the Lin measure (Lin 1998) is computed for the variable X in the following manner:
Linx(u, v) =
∑ f∈Fux∩Fvx [w u x ( f ) + wvx( f )]∑
f∈Fux w u x ( f ) + ∑ f∈Fvx w v x( f )
(1)
The measure is computed analogously for the variable Y and the final distributional similarity score, termed DIRT, is the geometric average of the scores for the two variables: If DIRT(u, v) is high, this means that the templates u and v share many “informative” arguments and so it is possible that u⇒ v.
Szpektor and Dagan (2008) suggested two modifications to DIRT. First, they looked at unary predicates, that is, predicates with a single variable such as X treat. Secondly, they computed a directional score that is more suited for capturing entailment relations compared to the symmetric Lin score. They proposed that if for two unary predicates u⇒ v, then relatively many of the features of u should be covered by the features of v. This is captured by the asymmetric Cover measure (Weeds and Weir 2003):
Cover(u, v) =
∑ f∈Fu∩Fv w
u( f )∑ f∈Fu wu( f )
(2)
The final directional score, termed BInc (Balanced Inclusion), is the geometric average of the Lin measure and the Cover measure.
Both Lin and Pantel as well as Szpektor and Dagan compute a similarity score using a single argument. However, it is clear that although this alleviates sparsity problems, considering pairs of arguments jointly provides more information. Yates and Etzioni (2009), Schoenmackers et al. (2010), and even earlier, Szpektor et al. (2004), presented methods that compute semantic similarity based on pairs of arguments.
A problem common to all local methods presented above is predicate ambiguity— predicates may have different meanings and different entailment relations in different contexts. Some works resolved the problem of ambiguity by representing predicates with argument variables that are typed (Pantel et al. 2007; Schoenmackers et al. 2010). For example, argument variables in the work of Schoenmackers et al. were restricted to belong to one of 156 types, such as country or profession. A different solution that has attracted substantial attention recently is to represent the various contexts in which a predicate can appear in a low-dimensional latent space (for example, using Latent Dirichlet Allocation [Blei, Ng, and Jordan 2003]) and infer entailment relations between predicates based on the contexts in which they appear (Ritter, Mausam, and Etzioni 2010; ÓSéaghdha 2010; Dinu and Lapata 2010; Melamud et al. 2013). In the experiments presented in this article we will use the representation of Schoenmackers et al. in one experiment, and ignore the problem of predicate ambiguity in the other. The idea of global learning is that, by jointly learning semantic relations between a large number of natural language phrases, one can use the dependencies between the relations to improve accuracy. A natural way to model that is with a graph where nodes are phrases and edges represent semantic similarity. Snow, Jurafsky, and Ng (2006) presented one of the early examples of global learning in the context of learning noun taxonomies. In their work, they enforced a transitivity constraint over the taxonomy using a greedy inference procedure and demonstrated that this improves the quality of the taxonomy. Transitivity constraints were also enforced by Yates and Etzioni (2009), who proposed a clustering algorithm for learning undirected synonymy relations. Nakashole, Weikum, and Suchanek (2012) learned a taxonomy of binary predicates and also enforced transitivity with a greedy algorithm.
Berant, Dagan, and Goldberger (2012) developed a global method for learning entailment relations between predicates. In this article we present scalable approximation algorithms for the optimization problem they propose, and so we provide further detail on their method. The input to their algorithm is a large corpus and a lexicographic resource, such as WordNet, and the output is a set of entailment rules that respect the constraint of transitivity. The algorithm is composed of two main steps. In the first step, a set of predicates is extracted from the corpus and a local entailment classifier is trained based on examples automatically generated from the lexicographic resource. At this point, we can derive for each pair of predicates (i, j) a score wij ∈ R that estimates whether i⇒ j.
In the second step of their algorithm, they construct a graph, where predicates are nodes and edges represent entailment rules. Using the local scores they look for the set of edges E that maximizes the objective function ∑ (i,j)∈E wij under the constraint that edges respect transitivity. They show that this optimization problem is NP-hard and find an exact solution using an ILP solver over small graphs. They also avoid problems of predicate ambiguity by partially contextualizing the predicates. In this article, we present efficient and scalable heuristic approximation algorithms for the optimization problem they propose.
In recent years, there has been substantial work on approximation algorithms for global inference problems. Do and Roth (2010) suggested a method for the related task of learning taxonomic relations between terms. Given a pair of terms, they construct a small graph that contains the two terms and a few other related terms, and then impose constraints on the graph structure. They construct these small graphs because their work is geared towards scenarios where relations are determined on-the-fly for a given pair of terms and no global knowledge base is ever explicitly constructed. Because they independently construct a graph for each pair of terms, their method easily produces solutions where global constraints, such as transitivity, are violated.
Another approximation method that violates transitivity constraints is LP relaxation (Martins, Smith, and Xing 2009). In LP relaxation, binary variables are replaced by continuous variables, transforming the problem from an ILP to a Linear Program (LP), which is polynomial. An LP solver is then applied, and variables that are assigned a fractional value are rounded to their nearest integer, so many violations of transitivity may occur. The solution when applying LP relaxation is not a transitive graph; we will show in Section 4 that our approximation method is substantially faster.
Global inference has gained popularity in recent years in NLP, and a common approximation method that has been extensively utilized is dual decomposition (Sontag, Globerson, and Jaakkola 2011). Dual decomposition has been successfully applied in
tasks such as Information Extraction (Reichart and Barzilay 2012), Machine Translation (Chang and Collins 2011), Parsing (Rush et al. 2012), and Named Entity Recognition (Wang, Che, and Manning 2013). To the best of our knowledge, it has not yet been applied for the task of learning entailment relations. The graph decomposition method we present in Section 3.1 can be viewed as an ideal case of dual decomposition, where we can decompose the problem into disjoint components in a way that we do not need to ensure consistency of the results obtained on each component separately.","3 efficient inference :Our goal is to learn a large knowledge base of entailment rules between natural language predicates. Following Berant, Dagan, and Goldberger (2012), we formulate this task as a graph-learning problem, where, given the nodes of an entailment graph, we would like to find the best set of edges that respect a global transitivity constraint.
Our main modeling assumption is that textual entailment is a transitive relation. Berant, Dagan, and Goldberger (2012) have demonstrated that this modeling assumption holds for focused entailment graphs, that is, graphs in which predicates are disambiguated by one of their arguments. For example, a graph that focuses on the concept nausea might contain a en entailment rule between predicates such as X prevents nausea ⇒X affects nausea, where the predicate X prevent Y is disambiguated by the instantiation of the argument nausea. In this work, we will examine this modeling assumption for more general graphs. In Section 4 we show that transitivity holds in typed entailment graphs, that is, graphs where each predicate specifies the semantic type of its arguments (for example, Xdrug prevent Ysymptom). In Section 5.4, we examine the assumption of transitivity when predicates do not carry any typing information (for example, X prevent Y); and we observe that in this setting the assumption of transitivity is often violated because of the predicate ambiguity. Nevertheless, we show that even in this set-up we can empirically take advantage of the transitivity assumption.
Let V be a set of predicates, which are the nodes of the entailment graph, and let w : V × V → R be an entailment weighting function. Given a pair of predicates (i, j), a positive wij indicates local tendency to decide that i entails j, whereas a negative wij indicates local tendency to decide that i does not entail j. We want to find a global entailment transitive graph that is most consistent with the local cues. Formally, our goal is to find the directed graph G = (V, E) that maximizes the sum of edge weights∑
(i,j)∈E wij, under the constraint that the graph is transitive—that is, for every triple of nodes (i, j, k), if (i, j) ∈ E and ( j, k) ∈ E, then (i, k) ∈ E.
Berant, Dagan, and Goldberger (2012) proved that this optimization problem (which we term Max-Trans-Graph) is NP-hard, and provided an ILP formulation for it. Let xij be an indicator for whether i⇒ j, then x = {xij : i 6= j} are the variables of the following ILP:
max x ∑ i6=j wijxij (3)
s.t. ∀i, j, k ∈ V xij + xjk − xik ≤ 1
∀i, j ∈ V xij ∈ {0, 1}
The objective function is the sum of weights over the edges of G, and the constraint xij + xjk − xik ≤ 1 on the binary variables enforces transitivity (i.e., xij = xjk = 1 implies that xik = 1). The weighting function w is trained separately using supervised learning methods and we describe the details of training for each one of our experiments in Sections 4 and 5.
There is a simple probabilistic modeling that motivates the score (3) that we optimize. Assume that for each pair of nodes (i, j), we are given a probability pij(1) = p(xij = 1) that i⇒ j (the probability that i ; j is denoted by pij(0) = 1− pij(1)). Assuming a uniform probability over graphs, the posterior probability (which can also be viewed as the likelihood) of a graph, represented by an edge-set x, is:
p(x) ∝ ∏ i6=j pij(xij) (4)
It can be easily verified that:1
log p(xij) = log pij(1) pij(0) xij + log pij(0) (5)
Hence,
log p(x) = ∑ i 6=j log p(xij) = ∑ i6=j wijxij + const (6)
such that ‘const’ is a scalar that does not depend on x and
wij = log pij(1) pij(0)
(7)
The optimal entailment graph, therefore, is arg maxx p(x) = arg maxx ∑
i6=j wijxij, where the maximization is over all the transitive graphs. Hence the most likely transitive entailment graph is obtained as the solution of the ILP maximization problem (3).
Berant, Dagan, and Goldberger solved this optimization problem with an ILP solver, but because ILP is NP-hard, this does not scale well; the number of variables is O(|V|2) and the number of constraints is O(|V|3). Thus, even a graph with 80 nodes (predicates) has more than half a million constraints. In this section, we describe a heuristic algorithm that empirically provides high-quality solutions for this optimization problem in graphs with tens of thousands of nodes.
Our algorithm contains two main steps. The first step (Section 3.1) is based on a structural assumption that entailment graphs are relatively sparse—that is, most predicates do not entail one another. This allows us to decompose the graph into smaller components in a way that guarantees that an exact solution for each one of the components results in an exact solution for Max-Trans-Graph. However, often even after decomposition, components are too large and finding an exact solution is still intractable.
The second step (Section 3.2) proposes a heuristic algorithm for finding a good solution for each one of the components. This step is based on an observation that
1 We assume that pij(1), pij(0) ∈ (0, 1) and thus log p(xij ) is well defined.
entailment graphs exhibit a “tree-like” property and are very similar to a novel type of graph, which we term forest-reducible graph. We utilize this property to develop an iterative efficient approximation algorithm for learning the graph edges, where each iteration takes linear time. We also prove that finding the optimal forest-reducible graph is NP-hard.
In Sections 4 and 5 we apply our algorithm on two different data sets and show that using transitivity substantially improves performance and that our methods dramatically improve scalability, allowing us to increase the scope of global learning of entailment graphs. The first step of our algorithm takes advantage of graph sparsity: Most predicates in language do not entail one another. Thus, it might be possible to decompose entailment graphs into small components and solve each component separately.
Let V1, V2 be a partitioning of the nodes V into two disjoint non-empty subsets. We term any directed edge, whose two nodes belong to different subsets, a crossing edge.
Proposition 1 If we can partition a set of nodes V into disjoint sets V1, V2 such that no crossing edge has a positive weight, then the optimal set of edges that is the solution of the ILP problem (3) does not contain any crossing edge.
Proof Assume by contradiction that the edge-set of the optimal solution Eopt contains a non-empty set of crossing edges Ecross. We can construct Enew = Eopt \ Ecross. Clearly ∑ (i,j)∈Enew wij ≥ ∑ (i,j)∈Eopt wij, as wij ≤ 0 for any crossing edge.
Next, we show that Enew does not violate transitivity constraints. Assuming it does, then the violation is caused by omitting the edges in Ecross. Thus, there must be, without loss of generality, nodes i ∈ V1 and k ∈ V2 such that for some node j, (i, j) and ( j, k) are in Enew, but (i, k) is not. However, this means either (i, j) or ( j, k) is a crossing edge, which is impossible because we omitted all crossing edges. Thus, Enew is a better solution than Eopt, contradiction. Hence, Ecross is empty.
This proposition suggests a simple algorithm (see Algorithm 1): Construct an undirected graph with the node set V and with an edge connecting i and j if either wij > 0 or wji > 0, then find its connected components, and finally solve each component separately. If an ILP solver is used to find the edges of each component separately,
then we obtain an optimal (not approximate) solution to the optimization problem (3) for the whole graph. Finding the undirected edges (Line 1) and computing connected components (Line 2) can be performed in O(V2). Thus, in this case the efficiency of the algorithm is dominated by the application of an ILP solver (Line 4).
If the entailment graph decomposes into small components, one could obtain an exact solution with an ILP solver, applied on each component separately, without resorting to any approximation. To further extend scalability in this setting we use a cutting-plane method (Kelley 1960). Cutting-plane methods have been used in the past, for example, in dependency parsing (Riedel and Clarke 2006). The idea is that even if we omit all transitivity constraints, we still expect most transitivity constraints to be satisfied, given a good weighting function w. Thus, it makes sense to avoid specifying the constraints ahead of time, but rather add them when they are violated. This is formalized in Algorithm 2.
Line 1 initializes an active set of constraints (ACT). Line 3 applies the ILP solver with the active constraints. Lines 4 and 5 find the violated constraints and add them to the active constraints. The algorithm halts when no constraints are violated. The solution is clearly optimal because we obtain a maximal solution for a less-constrained problem that does not violate any transitivity constraint.
A pre-condition for using cutting-plane methods is that computing the violated constraints (Line 4) is efficient, as it occurs in every iteration. We do that in a straightforward manner: For all edges (i, j) and ( j, k) that are in the current solution E∗, if (i, k) /∈ E∗ we add xij + xjk − xik ≤ 1 to the violated constraints. This is cubic in worst-case and performs very fast in practice.
To conclude, an exact algorithm for Max-Trans-Graph decomposes the graph into components and uses Algorithm 2 to find an exact solution for each component. However, because the problem is NP-hard, this will still fail once components become large. Next, we describe the second step of our method, which replaces Algorithm 2 with an efficient approximation that can scale to much larger graphs. We will show in Section 4 that the solutions obtained by the approximation algorithm are almost as good as the exact solution on a data set where finding an exact solution is feasible. The approximation we present in this section is based on a conjecture that entailment graphs exhibit a “tree-like” property, that is, they can be reduced into a structure similar to a directed forest, which we term forest-reducible graph (FRG). Although FRGs are a more constrained class of directed graphs, we prove that restricting our optimization
problem to FRGs does not make the problem fundamentally easier, that is, the problem remains NP-hard (see Appendix). Then, we present in Section 3.2.2 our iterative approximation algorithm, where in each iteration a node is removed and re-attached back to the graph in a locally optimal way. Combining this scheme with our conjecture about the graph structure yields a linear algorithm for node re-attachment.
Thus, the contribution of this section is twofold: First, we define a novel modeling assumption about the tree-like structure of entailment graphs and empirically demonstrate its validity. Second, we exploit this assumption to develop a polynomial approximation algorithm for learning entailment graphs that can scale to much larger graphs than in the past.
3.2.1 Forest-Reducible Graph. The predicate entailment relation, described by our entailment graphs, is typically from a “semantically specific” predicate to a more “general” one. Thus, intuitively, the topology of an entailment graph is expected to be “tree-like.” In this section we first formalize this intuition and then empirically analyze its validity. This property of entailment graphs is an interesting topological observation on its own, but also enables the efficient approximation algorithm of Section 3.2.2.
For a directed edge i⇒ j in a directed acyclic graph (DAG), we term the node i a child of node j, and j a parent of i.2 A directed forest is a DAG where all nodes have no more than one parent. A strongly connected component in a directed graph is a set of nodes such that there is a path from every node in the set to any other node. In entailment graphs, strongly connected components correspond to semantically equivalent predicates.
The entailment graph in Figure 2a (subgraph from the data set described in Section 4) is clearly not a directed forest—it contains a cycle of size two comprising the nodes X common in Y and X frequent in Y, and in addition the node X be epidemic in Y has three parents. However, we can convert it to a directed forest by applying the following operations. Any directed graph G can be converted into a StronglyConnected-Component (SCC) graph in the following way: Every strongly connected component is contracted into a single node, and an edge is added from SCC S1 to SCC S2 if there is an edge in G from some node in S1 to some node in S2. The SCC graph is always a DAG (Cormen et al. 2002), and if G is transitive, then the SCC graph is also transitive. The graph in Figure 2b is the SCC graph of the one in Figure 2a, but is still not a directed forest because the node X be epidemic in Y has two parents.
The transitive closure of a directed graph G is obtained by adding an edge from node i to node j if there is a path in G from i to j. The transitive reduction of G is obtained by removing all edges whose absence does not affect its transitive closure. In DAGs, the result of transitive reduction is unique (Aho, Garey, and Ullman 1972). We thus define the reduced graph Gred = (Vred, Ered) of a directed graph G as the transitive reduction of its SCC graph. The graph in Figure 2c is the reduced graph of the one in Figure 2a and is a directed forest. We say a graph is a forest-reducible graph if its reduced graph is a directed forest.
We now hypothesize that entailment graphs are approximately FRGs. The intuition is that the predicate on the left-hand-side of an entailment rule has a more specific meaning than the one on the right-hand-side. For instance, in Figure 2a X be epidemic in Y (where X is a type of disease and Y is a country) is more specific than X common in Y and
2 In standard graph terminology an edge is from a parent to a child. We choose the opposite definition to conflate edge direction with the direction of the entailment operator ‘⇒’.
X frequent in Y, which are equivalent, while X occur in Y is even more general. Accordingly, the reduced graph in Figure 2c is an FRG. We note that this is not always the case: For example, the entailment graph in Figure 3 is not an FRG, because X annex Y entails both Y be part of X and X invade Y, while the latter two do not entail one another. However, we hypothesize that this scenario is rather uncommon. Consequently, a natural variant of the Max-Trans-Graph problem is to restrict the required output graph of the optimization problem (3) to an FRG. We term this problem Max-Trans-Forest.
To test whether our hypothesis holds empirically, we performed the following analysis. We sampled seven gold standard typed entailment graphs (that is, graphs where
predicates specify the semantic type of their arguments, such as Xdrug prevents Ysymptom) from the data set described in Section 4, manually transformed them into FRGs by deleting a minimal number of edges, and measured recall over the set of edges in each graph (precision is naturally 1.0, as we only delete gold-standard edges). The lowest recall value obtained was 0.95, illustrating that deleting a very small proportion of edges converts a typed entailment graph into an FRG. Further support for the practical validity of this hypothesis is obtained from our experiments on typed entailment graphs in Section 4. In these experiments we show that exactly solving Max-Trans-Graph and Max-Trans-Forest (with an ILP solver) results in nearly identical performance. In Section 5.4 we qualitatively analyze the validity of the FRG assumption in untyped entailment graphs (where predicates are of the form X prevents Y) and find that this assumption does not always hold, because of predicate ambiguity.
An ILP formulation for Max-Trans-Forest is simple—a transitive graph is an FRG if all nodes in its reduced graph have no more than one parent. It can be verified that this is equivalent to the following statement: For every triplet of nodes i, j, k, if i⇒ j and i⇒ k, then either j⇒ k or k⇒ j (or both). This constraint can be stated in a linear form: xij + xik − xjk − xkj ≤ 1. Therefore, adding this new type of constraint to the ILP given in Equation (3) results in a formulation for Max-Trans-Forest:
max x ∑ i 6=j wijxij (8)
such that ∀i, j, k ∈ V xij + xjk − xik ≤ 1
∀i, j, k ∈ V xij + xik − xjk − xkj ≤ 1
∀i, j ∈ V xij ∈ {0, 1}
A natural question is whether there is a simpler (polynomial) solution for Max-TransForest that avoids the need for an ILP solver. In the Appendix we prove that Max-TransForest is also an NP-hard problem by a polynomial reduction from the X3C problem (Garey and Johnson 1979). Readers who are not interested in the proof can safely skip the Appendix.
3.2.2 Approximation Algorithm. In this section we present Tree-Node-Fix and Tree-NodeAnd-Component-Fix, which are efficient approximation algorithms for Max-TransForest, as well as Graph-Node-Fix, an approximation for Max-Trans-Graph.
Tree-Node-Fix. The scheme of Tree-Node-Fix (TNF) is the following. First, an initial FRG is constructed, using some initialization procedure. Then, at each iteration a single node v is re-attached (see below) to the FRG in a way that improves the objective function. This is repeated until the value of the objective function can no longer be improved by re-attaching a node.
Re-attaching a node v is performed by removing v from the graph (deleting v and all its adjacent edges) and connecting it back with a better set of edges, while maintaining the constraint that it is an FRG. This is done by considering all possible edges from/to the other graph nodes and choosing the optimal subset, while the rest of the graph edges remain fixed. For example, in Figure 2, one way of re-attaching the node X common in Y is to add it as a direct child of X occur in Y that is not a synonym of X frequent in Y. This will result in deletion of the edges X common in Y⇒ X frequent in Y, X frequent in Y⇒ X common in Y, and also X be epidemic in Y⇒ X common in Y, because otherwise the
resulting graph will not be an FRG. We will show that re-attachment can be efficiently performed in linear time using dynamic programming. Formally, let Sv−in = ∑
i 6=v wivxiv be the sum of scores over v’s incoming edges and Sv−out = ∑ k 6=v wvkxvk be the sum of scores over v’s outgoing edges. Re-attachment amounts to optimizing a linear objective:
argmax x(v) (Sv-in + Sv-out) (9)
while maintaining the FRG constraint, where the variables x(v) ⊆ x are indicators for all pairs of nodes involving v. We approximate a solution for Equation (3) by iteratively optimizing the simpler objective (9). Clearly, at each re-attachment the value of the objective function cannot decrease, because the optimization algorithm considers the previous graph as one of its candidate solutions.
We now show that re-attaching a node v is linear. To analyze v’s re-attachment, we consider the structure of the directed forest Gred just before v is re-inserted, and examine the possibilities for v’s insertion relative to that structure. We start by defining some helpful notations. Every node c ∈ Vred is a strongly connected component in G. Let vc ∈ c be an arbitrary representative node in c (we can choose any node vc ∈ c, because the graph G is transitive, and c is a strongly connected component). We denote by Sv-in(c) the sum of weights from all nodes in c and their descendants to v, and by Sv-out(c) the sum of weights from v to all nodes in c and their ancestors:
Sv-in(c) = ∑ i∈c wiv + ∑ k /∈c wkvxkvc (10)
Sv-out(c) = ∑ i∈c wvi + ∑ k /∈c wvkxvck (11)
Note that {xvck, xkvc} are edge indicators in G and not Gred. There are several cases for re-attaching v that we need to consider:
1. Inserting v into an existing component c ∈ Vred (case 1, see Figure 4a).
2. Inserting v as a new component, which breaks into two subcases:
(a) Forming a new component that contains only v, where v is a child of a component c (case 2a, see Figure 4b).
(b) Forming a new component that contains only v, where v is not a child of any component c, and so is a new root in Gred (case 2b, see Figure 4c).
Note that v cannot form a new component that contains other nodes as well, because the rest of the graph is fixed.
To find the optimal way of re-attaching v, we need to compute the score Equation (9), for each of the three cases, and choose the best one. We now describe how this score is computed in each of the three cases.
Case 1: Inserting v into a component c ∈ Vred. In this case we add in G edges from all nodes in c and their descendants to v and from v to all nodes in c and their ancestors. The score (9) in this case is
s1(c) , Sv-in(c) + Sv-out(c) (12)
Case 2a: Inserting v as a child of some c ∈ Vred. Once c is chosen as the parent of v, choosing v’s children in Gred is substantially constrained. A node that is not a descendant of c cannot become a child of v, because this would create a new path from that node to c and would require by transitivity to add a corresponding directed edge to c (but all graph edges not connecting v are fixed). Moreover, only a direct child of c can choose v as a parent instead of c (Figure 4b), because for any other descendant of c, v would become a second parent, and Gred will no longer be a directed forest (Figure 4b’). Thus, this case requires adding in G edges from v to all nodes in c and their ancestors; additionally, for each new child of v, denoted by d ∈ Vred, we add edges from all nodes in d and their descendants to v. Crucially, although the number of possible subsets of c’s children in Gred is exponential, the fact that they are independent trees in Gred allows us to go over them one by one, and decide for each one whether it will be a child of v or not, depending on whether Sv-in(d) is positive. Therefore, the score (9) in this case is:
s2(c) , Sv-out(c)+ ∑
d∈child(c) max(0, Sv-in(d)) (13)
where child(c) are the children of c. Case 2b: Inserting v as a new root in Gred. Similar to case 2a, only roots of Gred can become children of v. In this case for each chosen root r we add in G edges from the nodes in r and their descendants to v. Again, each root can be examined independently. Therefore, the score (9) of re-attaching v is:
s3 , ∑
r
max(0, Sv-in(r)) (14)
where the summation is over the roots of Gred. It can be easily verified that Sv-in(c) and Sv-out(c) satisfy the recursive definitions:
Sv-in(c) = ∑ i∈c wiv + ∑ d∈child(c) Sv-in(d), c ∈ Vred (15)
Sv-out(c) = ∑ i∈c wvi + Sv-out(p), c ∈ Vred (16)
where p is the parent of c in Gred. These recursive definitions allow computing in linear time Sv-in(c) and Sv-out(c) for all c (given Gred) using dynamic programming, before going over the cases for re-attaching v. Sv-in(c) is computed going over Vred leaves-to-root (post-order), and Sv-out(c) is computed going over Vred root-to-leaves (pre-order).
Re-attachment is summarized in Algorithm 3. Computing an SCC graph is linear (Cormen et al. 2002), and it is easy to verify that transitive reduction in FRGs is also linear (Line 1). Computing Sv-in(c) and Sv-out(c) (Lines 2–3) is also linear, as explained. Cases 1 and 2b are trivially linear and in case 2a we go over the children of all nodes in Vred. As the reduced graph is a forest, this simply means going over all nodes of Vred, and so the entire algorithm is linear.
Because re-attachment is linear, re-attaching all nodes is quadratic. Thus if we bound the number of iterations over all nodes, the overall complexity is quadratic. This is dramatically more efficient and scalable than applying an ILP solver. Empirically, we found that TNF converges after 5–10 iterations.
Tree-Node-and-Component-Fix. Assuming that entailment graphs are FRGs allows us to use the node re-attachment operation in linear time. However, this assumption also enables performing other graph operations efficiently. We now suggest a natural extension to the TNF algorithm that better explores the space of FRGs.
A disadvantage of TNF is that it re-attaches a single node in every iteration. Consider, for instance, the reduced graph in Figure 5. In this graph, nodes l, m, and n are direct children of node k, but suppose that in the optimal solution they are all children of node j. Reaching the optimal solution would require three independent re-attachment operations, and it is not clear that each of the three alone would improve the objective function value. However, if we allow re-attachment operations over components in the SCC graph, then we would be able to re-attach the strong connectivity component containing the nodes l, m, and n in a single operation. Thus, the idea of our extended TNF algorithm is to allow re-attachment of both nodes and components. We term this algorithm Tree-Node-and-Component-Fix (TNCF).
There are many ways in which this intuition can be implemented and our TNCF algorithm uses one possible variant. In TNCF we first perform node re-attachment until convergence as in TNF (after initialization), but then compute the SCC graph and perform component re-attachment until convergence. Component re-attachment is identical to node re-attachment, except that we are guaranteed that the reduced graph is a forest. After performing component re-attachment until convergence, we again
perform node re-attachments and then component re-attachments and so on, until the entire process converges.
Graph-Node-Fix. The re-attachment strategy described above can be applied without assuming that the graph is an FRG. Next, we show Graph-Node-Fix (GNF), a similar algorithm that uses the re-attachment strategy to obtain an approximate solution for the ILP problem Max-Trans-Graph (3). In this more general case, finding the optimal reattachment of a node v is done with an ILP solver. Nevertheless, this ILP is simpler than Equation (3), because we consider only candidate edges involving v. Figure 6 illustrates the three types of possible transitivity constraint violations when re-attaching v. The left side depicts a violation when (i, k) /∈ E, expressed by the constraint in Equation (18), and the middle and right depict two violations when the edge (i, k) ∈ E, expressed by the constraints in Equation (19). Thus, the ILP is formulated by adding the following constraints to the objective function (9):
Complexity is exponential because of the ILP solver; however, the ILP size is reduced by an order of magnitude to O(|V|) variables and O(|V|2) constraints. To summarize, the complexity of GNF is higher than the complexity of TNF, but it does not require the assumption that entailment graphs are FRGs. ","4 experimental evaluation: typed entailment graphs :In this and the next section we empirically evaluate the algorithms presented in Section 3 on two different data sets. Our first data set, presented in this section, comprises medium-sized graphs for which obtaining an exact solution is possible, and we demonstrate that our approximation methods substantially improve runtime while causing only a small degradation in performance compared with the optimal solution. The graphs are also particularly suited for global optimization because graph predicates are typed, which substantially reduces their ambiguity. The second data set (Section 5) contains a graph with tens of thousands of untyped nodes, where exact inference is completely impossible. We show that our methods scale to this graph and that transitivity improves performance even when predicates are untyped. The resulting graph contains more than 100,000 entailment rules that can be utilized in downstream semantic applications.
The input to the algorithms presented in Section 3 is a set of nodes V and a weighting function w : V × V → R. We describe how those are constructed before presenting an experimental evaluation. As mentioned in Section 2, one of the challenges in global optimization is that transitivity does not always hold when predicates are ambiguous. Schoenmackers et al. (2010) proposed an algorithm for learning inference rules between typed predicates, a representation that substantially reduces ambiguity. We adopt their representation and learn typed entailment graphs. A typed entailment graph (see Figure 7) is a directed graph where the nodes are typed predicates. A typed predicate is a triple (t1, p, t2), or simply p(t1, t2), representing a predicate in natural language. p is the lexical realization of the predicate and the typed variables t1, t2 indicate that the arguments of the predicate belong to the semantic types t1, t2. Semantic types are taken from a set of types T, where each type t ∈ T is a bag of natural language words or phrases. Examples for typed predicates are: conquer(COUNTRY,CITY) and contain(PRODUCT,MATERIAL). An instance of a typed predicate is a triple (a1, p, a2), or simply p(a1, a2), where a1 ∈ t1 and a2 ∈ t2
be relate to (drug,drug)
be derive from (drug,drug)
be process from (drug,drug)
be convert into (drug,drug)
238
are termed arguments. For example, be common in(asthma,australia) is an instance of be common in(DISEASE,PLACE).
Given a set of typed predicates, entailment rules can only exist between predicates that share the same (unordered) pair of types (such as PLACE and COUNTRY), as otherwise, the rule would contain unbound variables. Hence, given a set of typed predicates we can immediately decompose them into disjoint subsets—all typed predicates sharing the same pair of types define a separate graph that describes the entailment relations between those predicates (see Figure 7).
Edges in typed entailment graphs represent entailment rules in the usual way. If the type t1 is different from the type t2, mapping of arguments is straightforward, as in the rule be find in(MATERIAL,PRODUCT)⇒ contain(PRODUCT,MATERIAL). If t1 and t2 are equal we need to specify how arguments are mapped. This is done by splitting each node into two: For example, the node beat(TEAM,TEAM) is split into two typed predicates beat(Xteam,Yteam) and beat(Yteam,Xteam). This allows us to specify a rule where argument order is reversed such as beat(Xteam,Yteam)⇒ lose to(Yteam,Xteam). This also accommodates rules such as play(Xteam,Yteam)⇒ play(Yteam,Xteam): If team A plays team B, then team B plays team A.
To create typed entailment graphs, we used a data set that was generously provided by Schoenmackers et al. (2010). Schoenmackers et al. produced a mapping of 1.1 million arguments into 156 types (Examples for (argument, TYPE) pairs are (exodus, BOOK), (china, COUNTRY), and (asthma, DISEASE)), and then utilized the types, the mapped arguments, and 1 million TextRunner tuples (Banko et al. 2007) to generate a set of 10,672 typed predicates.3 Because entailment can only occur between predicates that share the same types, we decomposed the 10,672 typed predicates into 2,303 typed entailment graphs. The largest graph contains 188 nodes and the total number of potential rules is 263,756.4 The advantage of typing predicates is that it substantially reduces ambiguity, while still maintaining rules of wide applicability. The weighting function w is derived from an entailment score provided by a local classifier. Given a local classifier that provides an entailment score sij for a pair of predicates (i, j), we define wij = sij − λ, where λ is a prior that controls graph sparseness: As λ increases, wij decreases and becomes negative for more pairs of predicates, rendering the graph more sparse. The probabilistic interpretation of λ is as follows. For each two nodes let q be a prior probability that an edge exists. For large values of q the graph tends to be dense, and vice versa. Defining λ as log 1−qq , the modified weight function is:
wij = log p(xij = 1) p(xij = 0) − log 1− qq = sij − λ (20)
Given the weight function w, the task is to find the maximum a posteriori global graph that satisfies predefined constraints such as transitivity.
3 Readers are referred to their paper for details on mapping of tuples to typed predicates. The mapping of arguments into types can be downloaded from http://www.cs.washington.edu/research/ sherlock-hornclauses/. 4 In more detail, 1,714 graphs have less than 5 nodes, 326 graphs have 5–10 nodes, 145 graphs have 10–20 nodes, 92 graphs have 20–50 nodes, 18 graphs have 50–100 nodes, and 7 graphs have at least 100 nodes.
Training is similar to the method proposed by Berant, Dagan, and Goldberger (2012), and we briefly describe it here. The input for training is a lexicographic resource, for which we use WordNet, and a set of tuples, for which we use the 1 million typed TextRunner tuples provided by Schoenmackers et al. We perform the following steps:
Type Example
direct hypernym beat(TEAM,TEAM)⇒ play(TEAM,TEAM) direct synonym reach(TEAM,GAME)⇒ arrive at(TEAM,GAME) direct cohyponym invade(COUNTRY,CITY) ; bomb(COUNTRY,CITY) hyponym (distance=2) defeat(CITY,CITY) ; eliminate(CITY,CITY) random hold(PLACE,EVENT) ; win(PLACE,EVENT)
predicate. Then, similarity scores are combined by a geometric average. For example, the binary typed predicate invade(COUNTRY, CITY) will be decomposed into two unary predicates, one where the first argument of invade has the type COUNTRY (and the type of the second argument is unspecified), and another where the second argument of invade has the type CITY (and the first argument in unspecified).
The other five scores were provided by Schoenmackers et al. (Schoenmackers et al. 2010) and include SR (Schoenmackers et al. 2010), LIME (McCreath and Sharma 1997), M-estimate (Dzeroski and Brakto 1992), the standard G-test, and a simple implementation of Cover (Weeds and Weir 2003). Overall, the rationale behind this representation is that combining various scores will yield a better classifier than each single measure.
3. Training We sub-sample negative examples and train over an equal number of positive and negative examples. We used SVMperf (Joachims 2005) to train a Gaussian kernel classifier that provides an output score, sij. We tuned the two SVM parameters using 5-fold cross validation on a development set of two typed entailment graphs.
4. Prior knowledge For a small number of pairs of predicates, we might have prior knowledge of whether one entails the other. Berant, Dagan, and Goldberger (2012) integrated prior knowledge by adding hard constraints to the ILP. Because not all of our algorithms use an ILP solver, we integrate prior knowledge by modifying the local classifier score. For pairs of predicates i, j for which we have prior knowledge that i entails j (termed positive local constraints), we set sij =∞. For pairs of predicates i, j for which we have prior knowledge that i does not entail j (termed negative local constraints), we set sij = −∞.
To generate prior knowledge constraints we normalized each predicate by omitting the first word if it is a modal and turned passives into actives. If two normalized predicates are equal, they are a positive local constraint. Negative local constraints were constructed from three sources (1) Predicates differing by a single pair of words that are WordNet antonyms, (2) Predicates differing by a single word of negation, and (3) Predicates p(t1, t2) and p(t2, t1) where p is a transitive verb (for example, beat) in VerbNet (Kipper, Dang, and Palmer 2000).
As reported by Berant, Dagan, and Goldberger (2012), we find that injecting prior knowledge into the graph improves performance, as it provides a good starting point for the inference procedure. To evaluate performance, we manually annotated all edges in 10 typed entailment graphs containing 14, 14, 22, 30, 53, 62, 76, 86, 118, and 118 nodes. This annotation yielded 3,427 edges and 35,585 non-edges, resulting in an empirical edge density of 9%. The data set is publicly available and can be downloaded from http://www-nlp. stanford.edu/joberant/homepage_files/resources/Acl2011Exp.rar.
We implemented the following algorithms for learning graph edges, where in all of them the graph is first decomposed into components as described in Section 3.1.
No-trans Use local scores without transitivity constraints—an edge (i, j) is inserted iff wij > 0, or in other words iff sij > λ.
Exact-graph Find the optimal solution for Max-Trans-Graph in each component by applying an ILP solver in a cutting-plane method, as described in Section 3.1.
Exact-forest Find the exact solution for Max-Trans-Forest (see Equation 8) in each component by applying an ILP solver in a cutting-plane method.
LP-relax Solve Max-Trans-Graph approximately by applying an LP-relaxation (see Section 2) on each graph component. We apply the LP solver within the same cuttingplane method to allow for a direct comparison. As mentioned, our goal is to present a method for learning transitive graphs, whereas LP-relax produces solutions that violate transitivity. However, we run it on our data set to obtain empirical results, and to compare runtimes against TNF.
Graph-Node-Fix (GNF) Initialize each component in the following way: If the graph is very sparse, that is, λ ≥ C for some constant C (set to 1 in our experiments), then solving the graph exactly is not an issue and we use Exact-graph. Otherwise, we initialize by applying Exact-graph in a sparse configuration, that is, λ = C.
Tree-Node-Fix (TNF) Initialize as in GNF, except that if it generates a graph that is not an FRG, it is corrected by a simple heuristic: For every node in the reduced graph Gred that has more than one parent, we choose from its current parents the single one whose SCC is composed of the largest number of nodes in G.
We do not present the results of TNCF in this experiment, because for medium-sized graphs it provides results that are almost identical to TNF.
The No-trans baseline is a state-of-the-art local learning algorithm. It uses state-ofthe-art local scores such as DIRT (Lin and Pantel 2001), BInc (Szpektor and Dagan 2008), Cover (Weeds and Weir 2003), and SR (Schoenmackers et al. 2010) and trains a classifier using these scores. Berant, Dagan, and Goldberger (2010) have demonstrated empiricially that training a classifier over multiple similarity scores improves performance compared with using just a single similarity score.
The Exact-graph algorithm is the state-of-the-art global method presented by Berant, Dagan, and Goldberger (2012). It finds the optimal solution for Max-TransGraph, and our goal is to show that we can obtain good performance more efficiently using our approximation methods. Last, LP-relax is a standard approximation method for solving ILPs (Martins, Smith, and Xing 2009).
We note that a trivial baseline would be to initialize edges based on whether sij > λ and then compute the transitive closure in each component. We empirically found that this adds edges too aggressively and results in very low precision.
We use the Gurobi optimization package5 as our ILP solver in all experiments. The experiments were run on a multi-core 2.5GHz server with 32GB of RAM.
We evaluate algorithms by comparing the set of gold-standard edges with the set of edges learned by each algorithm. We measure recall, precision, and F1 for various values of the sparseness parameter λ, and compute the area under the precision-recall curve (AUC) generated. Efficiency is evaluated by comparing runtimes.
We first focus on runtimes and show that TNF is substantially more efficient than other baselines that use transitivity. Figure 8 compares runtimes of Exact-graph, GNF, TNF, and LP-relax as −λ increases and the graph becomes denser. Note that the y-axis
5 www.gurobi.com
is in logarithmic scale. Clearly, Exact-graph is extremely slow and runtime increases quickly. For λ = 0.3, runtime was already 12 hours and we were unable to obtain results for λ < 0.3, whereas in TNF we easily got a solution for any λ. When λ = 0.6, where both Exact-graph and TNF achieve best F1, TNF is 10 times faster than Exactgraph. When λ = 0.5, TNF is 50 times faster than Exact-graph, and so on. Most importantly, runtime for GNF and TNF increases much more slowly than for Exact-graph. Comparing runtimes for TNF and GNF, we see that the gap between the algorithms decreases as −λ increases. However, for reasonable values of λ, TNF is about four to seven times faster than GNF, and we were unable to run GNF on large graphs, as we report in Section 5.
Runtime of LP-relax is also bad compared with TNF and GNF. Runtime increases more slowly than Exact-graph, but still very fast compared with TNF. When λ = 0.6, LP-relax is almost 10 times slower than TNF, and when λ = −0.1, LP-relax is 200 times slower than TNF. This points to the difficulty of scaling LP-relax to large graphs. Last, Exact-forest is the slowest algorithm and because it is an approximation of Exact-graph we omit if from the figure for clarity.
We now examine the quality of the learned graphs and validity of our modeling assumptions. Figure 9 (left) shows a pair-wise comparison of Exact-graph and Exactforest. As is evident, the two curves are very similar, and the maximal F1 on the curve and AUC are almost identical. This provides further support for our modeling assumption that entailment graphs are roughly forest-reducible. Figure 9 (right) shows a similar comparison for TNF and GNF. We observe again that the curves are similar and performance is almost identical (maximal F1: 0.41, AUC: 0.31), illustrating that using the FRG assumption when using our approximation algorithm is empirically effective.
In Figure 10, we compare the performance of Exact-graph, which exactly solves Max-Trans-Graph, to our most efficient approximation algorithm, TNF, and to No-trans, a baseline that does not use transitivity at all (GNF and LP-relax are omitted from the
recall
figure to improve readability). We observe that both Exact-graph and TNF substantially outperform No-trans and that TNF’s graph quality is only slightly lower than Exact-graph (which is extremely slow). We report in the caption the maximal F1 on the curve and AUC in the recall range 0–0.5 (the widest range for which we have results for all algorithms). Note that compared with Exact-graph, TNF reduces AUC by merely a point and the maximal F1 score by two points only.
As for LP-relax, results are just slightly lower than Exact-graph (maximal F1: 0.43, AUC: 0.32), but its output is not a transitive graph, and as shown above runtime is quite slow.
To summarize, in this section we have empirically evaluated the algorithms presented in Section 3 on medium-sized graphs for which we can find an optimal solution. Our main findings are as follows:
1. Finding the optimal solution for Exact-graph and Exact-forest results in very similar graphs. This supports our assumption that entailment graphs are approximately forest-reducible.
2. Running GNF and TNF also yields very similar graphs, illustrating that using the FRG assumption when using our re-attachment approximation scheme is effective.
3. Our efficient approximation algorithm, TNF, is substantially faster compared with running an ILP solver that exactly solves Max-Trans-Graph.
4. TNF results in only a slight decrease in quality of learned graphs compared with Exact-graph.
5. TNF, which respects the transitivity constraint, learns graphs of higher quality compared with the local baseline, No-trans.
These findings lead us to believe that the algorithms presented in Section 3 can scale to large entailment graphs. In the next section, we empirically evaluate on a much larger data set for which using ILP solvers is impractical. Table 7 demonstrates cases where TNCF correctly omitted an edge and consequently decomposed a weakly connected component into two. These are classical cases where the global constraint of transitivity helps the algorithm avoid errors. In Example 1, although sij = 0.83 (which is higher than λ = 0.2), the edge is not inserted because this would cause the addition of wrong edges from the LHS component that deals with seeing and visiting to the RHS component, which is concerned with help and support. Example 2 is a case of a pair of predicates that are co-hyponyms or even perhaps antonyms. Transitivity helps TNCF avoid this erroneous rule.
To conclude, our analysis reveals that applying structural constraints encourages global algorithms to omit edges. Although this is often desirable, it can also prevent correct rules from being discovered because of problems of ambiguity. Thus, as discussed in Section 5.3, we believe that an important direction for future research is to apply our global graph learning methods over context-sensitive representations (such as the one presented by Melamud et al. [2013]).
Question 2: Node re-attachment. We would like to understand the effect of applying TNF over the entailment graph, or more precisely, check whether TNF substantially alters the set of graph edges compared with an initialization with HTL-FRG. First, we run both
HTL-FRG and TNF with sparseness parameter λ = 0.2 and compute the proportion of edges that is in their intersection. HTL-FRG learns 48,986 edges and TNF learns 61,186 edges. The number of edges in the intersection is 35,636. This means that 27% of the edges learned by HTL-FRG were removed (13,350 edges) and in addition 25,550 edges were added into the graph. Clearly, this indicates that TNF changes the learned graph substantially.
We exemplify this with a few fragments of graphs learned by HTL-FRG and TNF. Figure 14 shows 11 predicates,11 where the black solid edges correspond to HTL-FRG edges and red dashed edges correspond to TNF edges. Nodes that contain more than a single predicate represent an agreement between TNF and HTL-FRG that all predicates in the node entail one another. Clearly, TNF substantially changes the initialization generated by HTL-FRG. HTL-FRG creates an erroneous component in which give a lot of and generate a lot of are synonymous and mean a lot of entails them. TNF disassembles this component and determines that give a lot of is equivalent to offer plenty of, while generate a lot of entails cause of. Moreover, TNF disconnects the predicate mean a lot of completely. TNF also identifies that the predicates offer a wealth of, provide a wealth of, provide a lot of, and provide plenty of entail offer plenty of. Of course, TNF sometimes causes errors—for example, HTL-FRG decided that cause and cause of are equivalent, but TNF deleted the correct entailment rule cause of ⇒ cause.
Figure 15 provides another interesting example for an improvement in the graph due to the application of TNF. In the initialization, HTL-FRG determined that the predicate encrypt entails the predicates convert and convince rather than the predicates code and encode. This is because the local score provided by the classifier for (encrypt, convert) is 0.997—slightly higher than the local score for (encrypt, encode), which is 0.995. Therefore, HTL-FRG inserts the wrong edge encrypt⇒ convert and then it is unable to insert the correct rule encrypt⇒ encode because of the FRG assumption. After HTL-FRG terminates, TNF goes over the nodes one by one and is able to determine that overall the predicate encrypt fits better as a synonym of the predicates code and encode and reattaches it in the correct location. As an aside, notice that both HTL-FRG and TNF are able to identify in this case the correct directionality of the rule compress⇒ encode. The
11 We do not explicitly write the X and Y variables for brevity, and thus we focus only on predicates where the X variable is a subject.
local score given by the classifier for (compress, encode) is 0.65, whereas the local score for (encode, compress) is –0.11.
As a last example, Figure 16 demonstrates again the problem of predicate ambiguity. The predicate depress can occur in two contexts that, though related, instigate different entailment relations. The first context is when the object of the predicate (the Y argument) is a person and then the meaning of depress is similar to discourage. A second context is when the object is some phenomenon, for example, The serious economical situation depresses consumption levels. In initialization, HTL-FRG generates a set of edges that is compatible with the latter context, determining that the predicate depress entails the predicates lower and bring down. TNF re-attaches depress as a synonym of discourage and deter (because this improves the objective function), resulting in a set of edges that corresponds to the first meaning of depress mentioned above. The methods we use in this article do not allow learning the rules for both contexts because this would result in violations of the transitivity and FRG assumptions. This again emphasizes the importance of learning graphs where predicates are marked by their various senses, which will result in a model that can directly benefit from the methods suggested in this article.","5 experimental evaluation: untyped entailment graphs :In this section we evaluate our methods on a graph with 20,000 nodes. Again, we describe how the nodes V and the weighting function w are constructed. The nodes of the entailment graph we learn in this section are not typed. Although ambiguity is a problem in this setting, we will show that nevertheless transitivity constraints can improve results compared with a state-of-the-art local entailment classifier.
The nodes of the graph were generated from a set of two billion tuples of the form (arg1,predicate,arg2) that were extracted by the Reverb open information extraction system (Fader, Soderland, and Etzioni 2011) over the Clueweb096 data set and were generously provided by the authors of Reverb.
We performed some simple preprocessing over the extracted tuples. Each tuple is accompanied by a confidence value and we discarded tuples with confidence smaller than 0.5. Next, we normalized predicates using a procedure that omits unnecessary content such as modals and adverbs. The entire normalization procedure is publicly available as part of the Reverb project.7 Last, we normalized arguments by replacing pronouns and determiners by the tokens PRONOUN and DET. We also ran the BIU
6 http://lemurproject.org/clueweb09.php/. 7 https://github.com/knowitall/reverb-core/blob/master/core/src/main/java/edu/washington/ cs/knowitall/normalization/RelationString.java.
number normalizer8 and replaced numbers larger than one by the token NUM. After these three steps, we are left with a total of 960 million tuples of which 291 million are distinct.
The number of distinct predicates in the set of extracted tuples is 103,315. Because this is still a large number, we restrict the predicate set to the 10,000 predicates that appear with the highest number of pairs of arguments. As previously explained, each predicate is split into two (for example, defeat is split into X defeat Y and Y defeat X), and so the final number of entailment graph nodes is 20,000. As with typed entailment graphs, the weighting function w is obtained by training a classifier that provides a score sij for all pairs of predicates9 and defining wij = sij − λ. This setting, computing the scores sij, involves the following steps:
1. Training set generation. We use the data set released by Zeichner, Berant, and Dagan (2012), which contains 6,567 entailment rule applications annotated for their validity by crowdsourcing. For example, the data set marks that The exercises alleviate pain⇒ The exercises help ease pain is a valid rule application, whereas Obama want to boost the defense budget⇒ Obama increase the defense budget is an invalid rule application. We extract a single rule from each rule application, for example, from the rule application The exercises alleviate pain⇒ The exercises help ease pain we extract the rule X alleviate Y⇒ X help ease Y. We use half of the data set for training, resulting in 1,224 positive examples and 2,060 negative examples. Another two training examples are X unable to pay Y⇒ X owe Y and X own Y ; Y be sold to X.
2. Feature representation. Each pair of predicates (p1, p2) is represented by a feature vector where the first six are distributional similarity features identical to the first six features described in Section 4.2. In addition, for pairs of predicates for which at least one distributional similarity feature is non-zero, we add lexicographic features computed from WordNet (Fellbaum 1998), VerbOcean (Chklovski and Pantel 2004), and CatVar (Habash and Dorr 2003), as well as string-similarity features. Table 2 provides the exact details of these features. A feature is computed for a pair of predicates (p1, p2) where in this context a predicate is a pair (pred,rev): pred is the lexical realization of the predicate, and rev is a Boolean indicating whether arg1 is X and arg2 is Y or vice versa. Overall, each pair of predicates is represented by 27 features.
3. Training. After obtaining a feature representation for every pair of predicates, we train a Gaussian kernel SVM classifier that optimizes F1 (SVMperf implementation [Joachims 2005]), and tune the parameters C and γ by a grid search combined with 5-fold cross validation. We use the trained local classifier to compute a score sij for all pairs of predicates.
8 http://u.cs.biu.ac.il/~nlp/downloads/normalizer.html. 9 We do not need to run the classifier on 10, 0002 pairs of predicates, because for the majority of predicate
pairs the features are all zeros, and for this set of pairs we can run the classifier only once.
(c) passive-active: The passive-active constraint holds for (p1, p2) and (p2, p1) if p2 is the passive form of p1.
Again, local constraints are integrated into our model by changing the score sij =∞ for positive local constraints and sij = −∞ for negative local constraints. To evaluate performance we use the second half of the data set released by Zeichner, Berant, and Dagan (2012) as a test set. This test set contains 3,283 annotated examples, where 1,734 are covered by the 10,000 nodes of our graph (649 positive examples and 1,085 negative examples). As usual, for each algorithm we compute recall and precision with respect to the gold standard at various points by varying the sparseness parameter λ.
Most methods used over typed graphs in Section 4 (Exact-graph, Exact-forest, LPrelax, and GNF) are intractable because they require an ILP solver and our graph does not decompose into components that are small enough. Therefore, we compare the Notrans local baseline, where transitivity is not used, with the efficient global methods TNF and TNCF. Recall that in Section 4 we initialized TNF by applying an ILP solver in a sparse configuration on each graph component. In this experiment the graph is too large to use an ILP solver and so we initialize TNF and TNCF with a simple heuristic we describe next.
We begin with an empty graph and first sort all pairs of predicates (i, j) for which wij > 0, according to their weight w. Then, we go over predicate pairs one-by-one and perform two operations. First, we verify that inserting (i, j) into the graph does not violate the FRG assumption. This is done by going over all edges (i, k) ∈ E and checking that for every k either ( j, k) ∈ E or (k, j) ∈ E. If this is the case, then the edge (i, j) is a valid candidate edge as the resulting reduced graph will remain a directed forest, otherwise adding it will result in i having two parents in the reduced graph, and we do not insert this edge. Second, we compute the transitive closure Tij, which contains all node pairs that must be edges in the graph in case we insert (i, j), due to transitivity. We compute the change in the value of the objective function if we were to insert Tij into the graph. If this change improves the objective function value, we insert Tij into the graph (and so the value of the objective function increases monotonically). We keep going over the candidate edges from highest score to lowest until the objective function can no longer be improved by inserting a candidate edge. We term this initialization HTL-FRG, because we scan edges from high-score to low-score and maintain an FRG. We add this initialization procedure as another baseline.
Figure 11 presents the precision-recall curve of all global algorithms compared with the local classifier for 0.05 ≤ λ ≤ 2.5. One evident property is that No-trans reaches higher recall values than the global algorithms, which means that adding a global transitivity constraint prevents adding correct edges that have positive weight, since this would cause the addition of many other edges that have negative weight. A possible reason for that is predicate ambiguity, where positive weight edges connect connectivity components through an ambiguous predicate. This suggests that a natural direction for future research is to develop algorithms that do not impose transitivity constraints for rules i⇒ j and j⇒ k, if each rule refers to a different meaning of j. One possible direction is to learn latent “sense” or “topic” variables for each rule (as suggested by Melamud et al. Melamud et al. [2013]). Then, we can use the methods presented in this article
to impose transitivity constraints for each sense separately, and easily allow violations of transitivity across different senses. We provide concrete examples for such cases of ambiguity in our qualitative analysis in Section 5.4.
Although global algorithms are limited in their recall, for recall values of 0.1– 0.2 they substantially improve precision over No-trans. To better observe this result, Table 5 presents the results for the recall range 0.1–0.25. Comparing TNCF and Notrans shows that the TNCF algorithm improves over No-trans by 5–10 precision points:
In the recall range of 0.1–.2, the precision of No-trans is 0.68–0.75, whereas the precision of TNCF is 0.74–0.8. The number of rules learned by TNCF in this recall range is about 100,000.
We use HTL-FRG as an initialization method for TNCF, which on its own is not a very good approximation algorithm. Comparing HTL-FRG with No-trans, we observe that it is unable to reach high recall values and it only marginally improves precision compared with No-trans. Applying TNF over HTL-FRG also provides disappointing results—it increases recall compared with HTL-FRG, but precision is not better than No-trans. Indeed, it seems that although performing node re-attachments monotonically improves the objective function value, it is not powerful enough to learn a graph that is better with respect to our gold standard. On the other hand, adding component re-attachments (TNCF) allows the algorithm to better explore the space of graphs and learn a graph with higher recall and precision than HTL-FRG.
Figure 12 presents the runtime for HTL-FRG, TNF, and TNCF (y-axis in logarithmic scale), all of which yield an FRG and can be run on a large untyped graph. Again, we were unable to obtain results with Exact-graph, Exact-forest, LP-relax, and GNF, because using an ILP solver on a graph with 20,000 nodes was impossible. As expected, HTL-FRG is much faster than both TNF and TNCF, and TNCF is somewhat slower than TNF. However, as mentioned earlier, TNCF is able to improve both precision and recall compared with TNF and HTL-FRG. Note that when λ = 1.2, runtime increases suddenly for both TNF and TNCF. This is because at this point a large connected component is formed, as we explain next.
Figure 13 presents the results of graph decomposition as described in Section 3.1. Evidently, when λ = 0 the largest component constitutes almost all of the graph, which contains 20,000 nodes. Note that when λ = 1.2, the size of the largest component increases suddenly to more than half of the graph nodes. Additionally, TNCF obtained recall of 0.1–0.2 when λ ≤ 0.2, and in this case the number of nodes in the largest
component is 90% of the total number of graph nodes. Thus, contrary to the typed graphs, untyped graphs that contain ambiguous predicates do not decompose well into small components.
To summarize, we demonstrated that our efficient global methods scale to a graph with 20,000 nodes and observed that even when predicates are ambiguous we are able to improve precision for moderate values of recall. Nevertheless, there are indications that our modeling assumptions are violated when applied over a large graph with ambiguous predicates. Transitivity and FRG constraints limit the recall we are able to reach, and the graph does not decompose very well into components. Thus, in future work we would like to apply our algorithms over graphs where predicate ambiguity is modeled and so transitivity constraints can be properly applied. In the previous section we quantitatively evaluated global methods for learning entailment graphs over a large set of untyped predicates. Our results revealed interesting issues that we would like to further investigate:
1. Why do methods that assume entailment graphs are transitive and forest-reducible have limited recall?
2. How much does node re-attachment affect graph structure?
In this section, we try to answer these questions by performing a qualitative analysis that allows us to gain better insight on the behavior of our algorithms.
Question 1: Transitivity and FRG assumptions. We would like to understand why using global algorithms such as TNCF results in limited recall. Our hypothesis is that because predicates in the graph are untyped and potentially ambiguous, violations of the
assumptions are possible, which results in recall reduction. To examine this hypothesis, we perform two qualitative analyses:
1. We analyze graphs containing only gold-standard edges, and manually identify cases where the modeling assumptions are violated.
2. We compare the local algorithm, No-trans, to the global algorithm, TNCF. If our hypothesis is correct, then we expect TNCF to remove correct edges that are inserted by No-trans due to ambiguous predicates.
We start by constructing a graph with all edges from our gold-standard data set and search for transitivity and FRG violations caused by ambiguous predicates. Note that the manually annotated gold-standard edges are only a subset of the full set of correct edges, which is actually much larger. Thus, we cannot automatically identify violations, because the graph is missing many true edges. Instead, we manually go over candidate violations and check if indeed this case is a result of a violation of our modeling assumptions, and if so then our global algorithm cannot predict the correct set of edges.
One example is the pair of rules seek ⇒ apply for and apply for ⇒ file for. The first rule was annotated in the rule application Others seek medical care ⇒ Others apply for medical care (Recall that crowdsourcing was used to annotate rule applications, and rules were extracted from these applications.) The second rule was extracted from the rule application Students apply for more than one award⇒ Student file for more than one award. This causes a transitivity violation because Students seek more than one award ; Student file for more than one award. Evidently, the meaning of the predicate apply for is contextdependent. Another example is the pair of rules contain ⇒ supply and supply ⇒ serve annotated in the applications The page contains links⇒ The page supplies links and Users supply information or material⇒Users serve information or material. Clearly, contain ; serve and the transitivity violation is caused by the fact that the predicate supply is contextdependent—in the first context the subject is inanimate, whereas in the second context the subject is human.
A good example for an FRG violation is the predicate come from. The following three entailment rules are annotated by the turkers: come from ⇒ be raised in, come from⇒ be derived from, and come from⇒ come out of. These correspond to the following rule applications: The Messiah comes from Judah ⇒ The Messiah was raised in Judah, The colors come from the sun light ⇒ The colors are derived from the sun light, and The truth comes from the book ⇒ The truth comes out of the book. Clearly, be raised in ; be derived from and be derived from ; be raised in, so this is an FRG violation. Indeed, the three applications correspond to different meanings of the predicate come from, which depend on context. A second example is the pair of rules be dedicated to⇒ be committed to and be dedicated to⇒ contain information on. Again, this is an FRG violation because be committed to ; contain information on and contain information on ; be committed to. This violation occurs because of the ambiguity of the predicate be dedicated to, which can be resolved by knowing whether the subject is human or is an object that carries information (for example, The web site is dedicated to the life of lizards).
This first analysis revealed particular cases where the transitivity and FRG assumptions do not hold. In the next analysis, we would like to directly compare cases where global and local methods disagree with one another. We hypothesize that because predicates are not disambiguated, then global algorithms will delete correct edges due
to an ambiguous predicate. We set the sparseness parameter λ = 0.2 and compare the set of edges learned by No-trans to the set of edges learned by TNCF.
Examining the disagreements, we observe that, as expected, in 90% of the cases TNCF deletes an edge that was inserted by No-trans. We divide these deleted edges into two categories in the following manner: We compute the weakly connected components of the graph learned by TNCF,10 and then for each edge inserted by No-trans and deleted by TNCF we check whether it connects two different weakly connected components or not. There are two reasons for focusing on this property of deleted edges. First, we hypothesize that most deleted edges will connect two different connected components, because such edges result in many violations of transitivity. Second, weakly connected components usually correspond to separate semantic meanings, and thus a correct edge that connects two weakly connected components probably involves an ambiguous predicate.
Indeed, out of all edges where No-trans inserts an edge and TNCF does not, 76.4% connect weakly connected components in the global graph. This proves that deleting such edges is the main cause for disagreement between TNCF and No-trans. Now, we can take a closer look at these edges, and check whether indeed when a correct edge is deleted, it is because of an ambiguous predicate.
We randomly sample five cases where TNCF erroneously deleted an edge connecting weakly connected components—these are cases where ambiguity is likely to occur. For comparison, we also sample five cases where TNCF was right and No-trans erred. Table 6 shows the samples where No-trans is correct. The first column describes the rule and the second column specifies the score sij provided by the local classifier No-trans. The last two columns detail two examples for predicates that are in the same weakly connected component with the rule LHS and RHS in the graph learned by TNCF. These two columns provide a flavor for the overall meaning of predicates that belong to this component. Table 7 is equivalent and shows the samples where TNCF is correct.
Example 1 in Table 6 already demonstrates the problem of ambiguity. The LHS for this rule application in the crowdsourcing experiment was your parents turn off comments and indeed in this case the RHS your parents cut off comments is inferred. However, we can see that the LHS component revolves around the turning off or shutting off of appliances, for example, whereas the RHS component deals with a more explicit and physical meaning of cutting. This is because the predicate X cut off Y is ambiguous and consequently TNCF omits this correct edge. Example 5 also demonstrates well the problem of ambiguity—the LHS of this rule application is This site was put by fans, which in this context entails the RHS This site was run by fans. However, the more common sense of be put by does not entail the predicate be run by; this sense is captured by the predicates in the LHS component Y help put X and X be placed by Y. Example 4 is another good example where the ambiguous predicate is X be open to Y—the RHS component is concerned with the physical state of being opened, whereas the LHS component has a more abstract sense of availability. The LHS of the rule application in this case was The services are available to museums. Examples 2 and 3 are cases where the problem is not in predicate ambiguity but in the connected component itself. Thus, out of five randomly sampled examples where TNCF deleted an edge that connects two weakly connected components, three can be attributed to predicate ambiguity.
10 A weekly connected component in a directed graph is a set of nodes S such that for every pair of nodes u, v ∈ S there is an undirected path from u to v and from v to u, that is, there is a path when edge directions are ignored.","6 conclusions :The problem of language variability is at the heart of many semantic applications such as Information Extraction, Question Answering, Semantic Parsing, and more. Consequently, learning broad coverage knowledge bases of entailment rules and paraphrases (Ganitkevitch, Van Durme, and Callison-Burch 2013) has proved crucial for systems that perform inference over textual representations (Stern and Dagan 2012; Angeli and Manning 2014).
In this article, we have presented algorithms for learning entailment rules between predicates that can scale to large sets of predicates. Our work builds on prior work by Berant, Dagan, and Goldberger (2012), who defined the concept of entailment graphs, and formulated entailment rule learning as a graph optimization problem, where the graph has to obey the structural constraint of transitivity.
Our main contribution is a heuristic polynomial approximation algorithm that can learn entailment graphs containing tens of thousands of nodes. The algorithm is based on two main observations: first, that entailment graphs are sparse, and second, that entailment graphs are forest-reducible. This leads to an efficient algorithm in which at the first step the graph is decomposed into smaller components, and in the second step we find a set of edges for each component. The second step is an iterative approximation algorithm in which at each iteration a node or component of the graph is deleted and then re-attached in a way that improves the overall objective function.
We performed a thorough empirical evaluation, in which we ran our algorithm on two separate data sets. On the first data set we witnessed a substantial improvement in runtime at a relatively small cost in performance. On the second data set, we show that our method scales to an untyped entailment graph containing 20,000 nodes, much larger than was previously possible using state-of-the-art global methods. We see that using a global algorithm improves precision for modest values of recall, but that local methods can achieve higher recall than global methods. We perform an analysis and conclude that the main reason for this limitation is predicate ambiguity, which results in graphs that do not adhere to the structural constraints of transitivity and forest reducibility.
In future work, we hope to address the problem of predicate ambiguity. A natural direction is to combine methods that model the context-sensitivity of predicates with global structural assumptions. In this way, one can apply structural constraints for a set of predicates, given that they appear in similar contexts. A second important direction for future work is to demonstrate that using entailment rules learned by our method improves performance in downstream semantic applications such as QA, IE, and Recognizing Textual Entailment.",,,"Entailment rules between predicates are fundamental to many semantic-inference applications. Consequently, learning such rules has been an active field of research in recent years. Methods for learning entailment rules between predicates that take into account dependencies between different rules (e.g., entailment is a transitive relation) have been shown to improve rule quality, but suffer from scalability issues, that is, the number of predicates handled is often quite small. In this article, we present methods for learning transitive graphs that contain tens of thousands of nodes, where nodes represent predicates and edges correspond to entailment rules (termed entailment graphs). Our methods are able to scale to a large number of predicates by exploiting structural properties of entailment graphs such as the fact that they exhibit a “tree-like” property. We apply our methods on two data sets and demonstrate that our methods find high-quality solutions faster than methods proposed in the past, and moreover our methods for the first time scale to large graphs containing 20,000 nodes and more than 100,000 edges.","[{""affiliations"": [], ""name"": ""Jonathan Berant""}, {""affiliations"": [], ""name"": ""Noga Alon""}, {""affiliations"": [], ""name"": ""Ido Dagan""}, {""affiliations"": [], ""name"": ""Jacob Goldberger""}]",SP:b18e02d0b34892a803b1595d68c0b380680e7437,"[{""authors"": [""Aho"", ""Alfred V."", ""Michael R. Garey"", ""Jeffrey D. Ullman.""], ""title"": ""The transitive reduction of a directed graph"", ""venue"": ""SIAM Journal on Computing, 1(2):131\u2013137."", ""year"": 1972}, {""authors"": [""Angeli"", ""Gabor"", ""Christopher D. Manning.""], ""title"": ""Naturalli: Natural logic inference for common sense reasoning"", ""venue"": ""Proceedings of EMNLP, pages 534\u2013545, Doha."", ""year"": 2014}, {""authors"": [""Oren Etzioni.""], ""title"": ""Open information extraction from the web"", ""venue"": ""Proceedings of IJCAI, pages 68\u201374, Hyderabad."", ""year"": 2007}, {""authors"": [""Ben Aharon"", ""Roni"", ""Idan Szpektor"", ""Ido Dagan.""], ""title"": ""Generating entailment rules from FrameNet"", ""venue"": ""Proceedings of ACL, pages 241\u2013246, Uppsala."", ""year"": 2010}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Meni Adler"", ""Jacob Goldberger""], ""title"": ""Efficient tree-based approximation for entailment"", ""year"": 2012}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Jacob Goldberger.""], ""title"": ""Global learning of focused entailment graphs"", ""venue"": ""Proceedings of ACL, pages 1,220\u20131,229, Uppsala."", ""year"": 2010}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Jacob Goldberger.""], ""title"": ""Global learning of typed entailment rules"", ""venue"": ""Proceedings of ACL, pages 610\u2013619, Portland, OR."", ""year"": 2011}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Jacob Goldberger.""], ""title"": ""Learning entailment relations by global graph structure optimization"", ""venue"": ""Journal of Computational Linguistics,"", ""year"": 2012}, {""authors"": [""Berant"", ""Jonathan"", ""Percy Liang.""], ""title"": ""Semantic parsing via paraphrasing"", ""venue"": ""Proceedings of ACL, pages 1,415\u20131,425, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Bhagat"", ""Rahul"", ""Patrick Pantel"", ""Eduard Hovy.""], ""title"": ""LEDIR: An unsupervised algorithm for learning directionality of inference rules"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 161\u2013170, Prague."", ""year"": 2007}, {""authors"": [""Blei"", ""David M."", ""Andrew Y. Ng"", ""Michael I. Jordan.""], ""title"": ""Latent Dirichlet allocation"", ""venue"": ""Journal of Machine Learning Research, 3:993\u20131022."", ""year"": 2003}, {""authors"": [""Budanitsky"", ""Alexander"", ""Graeme Hirst.""], ""title"": ""Evaluating WordNet-based measures of lexical semantic relatedness"", ""venue"": ""Computational Linguistics, 32(1):13\u201347."", ""year"": 2006}, {""authors"": [""Chambers"", ""Nathanael"", ""Dan Jurafsky.""], ""title"": ""Template-based information extraction without the templates"", ""venue"": ""Proceedings of ACL, pages 976\u2013986, Portland, OR."", ""year"": 2011}, {""authors"": [""Chang"", ""Yin-Wen"", ""Michael Collins.""], ""title"": ""Exact decoding of phrase-based translation models through Lagrangian relaxation"", ""venue"": ""Proceedings of EMNLP, pages 26\u201337, Edinburgh."", ""year"": 2011}, {""authors"": [""Chklovski"", ""Timothy"", ""Patrick Pantel.""], ""title"": ""Verb ocean: Mining the web for fine-grained semantic verb relations"", ""venue"": ""Proceedings of EMNLP, pages 33\u201340, Barcelona."", ""year"": 2004}, {""authors"": [""Cormen"", ""Thomas H."", ""Charles E. Leiserson"", ""Ronald L. Rivest"", ""Clifford Stein.""], ""title"": ""Introduction to Algorithms"", ""venue"": ""The MIT Press."", ""year"": 2002}, {""authors"": [""Coyne"", ""Bob"", ""Owen Rambow.""], ""title"": ""Lexpar: A freely available English paraphrase lexicon automatically extracted from FrameNet"", ""venue"": ""Proceedings of IEEE International Conference on Semantic"", ""year"": 2009}, {""authors"": [""Dinu"", ""Georgiana"", ""Mirella Lapata.""], ""title"": ""Topic models for meaning similarity in context"", ""venue"": ""Proceedings of COLING, pages 250\u2013258, Beijing."", ""year"": 2010}, {""authors"": [""Do"", ""Quang"", ""Dan Roth.""], ""title"": ""Constraints based taxonomic relation classification"", ""venue"": ""Proceedings of EMNLP, pages 1,099\u20131,109, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Dzeroski"", ""Saso"", ""Ivan Brakto.""], ""title"": ""Handling noise in inductive logic programming"", ""venue"": ""Proceedings of the International Workshop on Inductive Logic Programming, pages 48\u201364, Tokyo."", ""year"": 1992}, {""authors"": [""Fader"", ""Anthony"", ""Stephen Soderland"", ""Oren Etzioni.""], ""title"": ""Identifying relations for open information extraction"", ""venue"": ""Proceedings of EMNLP, pages 1,535\u20131,545, Portland, OR."", ""year"": 2011}, {""authors"": [""Fellbaum"", ""Christiane""], ""title"": ""WordNet: An Electronic Lexical Database (Language, Speech, and Communication)"", ""year"": 1998}, {""authors"": [""Ganitkevitch"", ""Juri"", ""Benjamin Van Durme"", ""Chris Callison-Burch.""], ""title"": ""PPDB: The paraphrase database"", ""venue"": ""Proceedings of HLT-NAACL, pages 758\u2013764, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Garey"", ""Michael R."", ""David S. Johnson.""], ""title"": ""Computers and Intractability: A Guide to the Theory of NP-Completeness"", ""venue"": ""W. H. Freeman."", ""year"": 1979}, {""authors"": [""Habash"", ""Nizar"", ""Bonnie Dorr.""], ""title"": ""A categorial variation database for English"", ""venue"": ""Proceedings NAACL-HLT, pages 17\u201323, Edmonton."", ""year"": 2003}, {""authors"": [""Harris"", ""Zellig.""], ""title"": ""Distributional structure"", ""venue"": ""Word, 10(23):146\u2013162."", ""year"": 1954}, {""authors"": [""Hearst"", ""Marti.""], ""title"": ""Automatic acquisition of hyponyms from large text corpora"", ""venue"": ""Proceedings of COLING, pages 23\u201328, Nantes."", ""year"": 1992}, {""authors"": [""Joachims"", ""Thorsten.""], ""title"": ""A support vector method for multivariate performance measures"", ""venue"": ""Proceedings of ICML, pages 377\u2013384, Bonn."", ""year"": 2005}, {""authors"": [""J.E. Kelley""], ""title"": ""The cutting-plane method for solving convex programs"", ""venue"": ""Journal of the Society for Industrial Applied Mathematics, 8:703\u2013712."", ""year"": 1960}, {""authors"": [""Kipper"", ""Karin"", ""Hoa T. Dang"", ""Martha Palmer.""], ""title"": ""Class-based construction of a verb lexicon"", ""venue"": ""Proceedings of AAAI, pages 691\u2013696, Austin, TX."", ""year"": 2000}, {""authors"": [""Lin"", ""Dekang.""], ""title"": ""Automatic retrieval and clustering of similar words"", ""venue"": ""Proceedings of COLING-ACL, pages 768\u2013774, Montreal."", ""year"": 1998}, {""authors"": [""Lin"", ""Dekang"", ""Patrick Pantel""], ""title"": ""Discovery of inference rules for question"", ""year"": 2001}, {""authors"": [""Martins"", ""Andre"", ""Noah Smith"", ""Eric Xing.""], ""title"": ""Concise integer linear programming formulations for dependency parsing"", ""venue"": ""Proceedings of ACL, pages 342\u2013350, Singapore."", ""year"": 2009}, {""authors"": [""McCreath"", ""Eric"", ""Arun Sharma.""], ""title"": ""ILP with noise and fixed example size: a Bayesian approach"", ""venue"": ""Proceedings of IJCAI, pages 1,310\u20131,315, Nagoya."", ""year"": 1997}, {""authors"": [""Melamud"", ""Oren"", ""Jonathan Berant"", ""Ido Dagan"", ""Jacob Goldberger"", ""Idan Szpektor.""], ""title"": ""A two level model for context sensitive inference rules"", ""venue"": ""Proceedings of ACL, pages 1,331\u20131,340, Sofia."", ""year"": 2013}, {""authors"": [""Meyers"", ""Adam"", ""Ruth Reeves"", ""Catherine Macleod"", ""Rachel Szekeley"", ""Veronika Zielinska"", ""Brian Young.""], ""title"": ""The cross-breeding of dictionaries"", ""venue"": ""Proceedings of LREC, pages 1,095\u20131,098,"", ""year"": 2004}, {""authors"": [""Nakashole"", ""Ndapandula"", ""Gerhard Weikum"", ""Fabian M. Suchanek.""], ""title"": ""Patty: A taxonomy of relational patterns with semantic types"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 1,135\u20131,145,"", ""year"": 2012}, {""authors"": [""\u00d3S\u00e9aghdha"", ""Diarmuid.""], ""title"": ""Latent variable models of selectional preference"", ""venue"": ""Proceedings of ACL, pages 435\u2013444, Uppsala."", ""year"": 2010}, {""authors"": [""Pantel"", ""Patrick"", ""Rahul Bhagat"", ""Bonaventura Coppola"", ""Timothy Chklovski"", ""Eduard H. Hovy.""], ""title"": ""ISP: Learning inferential selectional preferences"", ""venue"": ""Proceedings of NAACL-HLT, pages 564\u2013571,"", ""year"": 2007}, {""authors"": [""Pekar"", ""Viktor.""], ""title"": ""Discovery of event entailment knowledge from text corpora"", ""venue"": ""Computer Speech & Language, 22(1):1\u201316."", ""year"": 2008}, {""authors"": [""Ravichandran"", ""Deepak"", ""Eduard H. Hovy.""], ""title"": ""Learning surface text patterns for a question answering system"", ""venue"": ""Proceedings of ACL, pages 41\u201347, Philadelphia, PA."", ""year"": 2002}, {""authors"": [""Reichart"", ""Roi"", ""Regina Barzilay.""], ""title"": ""Multi-event extraction guided by global constraints"", ""venue"": ""Proceedings of NAACL-HLT, pages 70\u201379, Montreal."", ""year"": 2012}, {""authors"": [""Richens"", ""Tom.""], ""title"": ""Anomalies in the WordNet verb hierarchy"", ""venue"": ""Proceedings of COLING, pages 729\u2013736, Manchester."", ""year"": 2008}, {""authors"": [""Riedel"", ""Sebastian"", ""James Clarke.""], ""title"": ""Incremental integer linear programming for non-projective dependency parsing"", ""venue"": ""Proceedings of EMNLP, pages 129\u2013137, Sydney."", ""year"": 2006}, {""authors"": [""Ritter"", ""Alan"", ""Mausam"", ""Oren Etzioni.""], ""title"": ""A latent Dirichlet allocation method for selectional preferences"", ""venue"": ""Proceedings of ACL, pages 424\u2013434, Uppsala."", ""year"": 2010}, {""authors"": [""Roth"", ""Michael"", ""Anette Frank.""], ""title"": ""Aligning predicates across monolingual comparable texts using graph-based clustering"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 171\u2013182, Jeju Island."", ""year"": 2012}, {""authors"": [""Rush"", ""Alexander M."", ""Roi Reichart"", ""Michael Collins"", ""Amir Globerson.""], ""title"": ""Improved parsing and POS tagging using inter-sentence consistency constraints"", ""venue"": ""Proceedings of EMNLP, pages 1434\u20131444,"", ""year"": 2012}, {""authors"": [""Schoenmackers"", ""Stefan"", ""Jesse Davis"", ""Oren Etzioni"", ""Daniel S. Weld.""], ""title"": ""Learning first-order horn clauses from web text"", ""venue"": ""Proceedings of EMNLP, pages 1,088\u20131,098, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Sekine"", ""Satoshi.""], ""title"": ""Automatic paraphrase discovery based on context and keywords between NE pairs"", ""venue"": ""Proceedings of IWP, pages 80\u201387, Jeju Island."", ""year"": 2005}, {""authors"": [""Shinyama"", ""Yusuke"", ""Satoshi Sekine.""], ""title"": ""Preemptive information extraction using unrestricted relation discovery"", ""venue"": ""HLT-NAACL, pages 304\u2013311, New-York, NY."", ""year"": 2006}, {""authors"": [""Snow"", ""Rion"", ""Daniel Jurafsky"", ""Andrew Y. Ng.""], ""title"": ""Semantic taxonomy induction from heterogenous evidence"", ""venue"": ""Proceedings of ACL, pages 801\u2013808, Sydney."", ""year"": 2006}, {""authors"": [""Sontag"", ""David"", ""Amir Globerson"", ""Tommi Jaakkola.""], ""title"": ""Introduction to dual decomposition for inference"", ""venue"": ""Suvrit Sra, Sebastian Nowozin, and Stephen J. Wright, editors, Optimization for Machine"", ""year"": 2011}, {""authors"": [""Stern"", ""Asher"", ""Ido Dagan.""], ""title"": ""BIUTEE: A modular open-source system for recognizing textual entailment"", ""venue"": ""Proceedings of ACL System Demonstrations, pages 73\u201378."", ""year"": 2012}, {""authors"": [""Szpektor"", ""Idan"", ""Ido Dagan.""], ""title"": ""Learning entailment rules for unary templates"", ""venue"": ""Proceedings of COLING, pages 849\u2013856, Manchester."", ""year"": 2008}, {""authors"": [""Szpektor"", ""Idan"", ""Ido Dagan.""], ""title"": ""Augmenting WordNet-based inference with argument mapping"", ""venue"": ""Proceedings of TextInfer, pages 27\u201335, Singapore."", ""year"": 2009}, {""authors"": [""Szpektor"", ""Idan"", ""Hristo Tanev"", ""Ido Dagan"", ""Bonaventura Coppola.""], ""title"": ""Scaling web-based acquisition of entailment relations"", ""venue"": ""Proceedings of EMNLP, pages 41\u201348, Barcelona."", ""year"": 2004}, {""authors"": [""Wang"", ""Mengqiu"", ""Wanxiang Che"", ""Christopher D. Manning.""], ""title"": ""Effective bilingual constraints for semi-supervised learning of named entity recognizers"", ""venue"": ""Proceedings of AAAI, pages 919\u2013925,"", ""year"": 2013}, {""authors"": [""Weeds"", ""Julie"", ""David Weir.""], ""title"": ""A general framework for distributional similarity"", ""venue"": ""Proceedings of EMNLP, pages 81\u201388, Sapporo."", ""year"": 2003}, {""authors"": [""Weisman"", ""Hila"", ""Jonathan Berant"", ""Idan Szpektor"", ""Ido Dagan""], ""title"": ""Learning verb inference rules"", ""year"": 2012}, {""authors"": [""Yates"", ""Alexander"", ""Oren Etzioni.""], ""title"": ""Unsupervised methods for determining object and relation synonyms on the web"", ""venue"": ""Journal of Artificial Intelligence Research, 34:255\u2013296."", ""year"": 2009}, {""authors"": [""Zeichner"", ""Naomi"", ""Jonathan Berant"", ""Ido Dagan.""], ""title"": ""Crowdsourcing inference-rule evaluation"", ""venue"": ""Proceedings of ACL (short papers), pages 156\u2013160, Jeju Island."", ""year"": 2012}]",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"appendix a. max-trans-forest is np-hard :In this appendix we show that the Max-Trans-Forest problem (8) is NP-hard. We prove that the following decision problem (which is clearly easier than Max-TransForest) is NP-hard: Given a set of nodes V, a function w : V × V → R, and an integer number α, is there a transitive forest-reducible (FRG) subgraph G = (V, E) such that∑
e∈E w(e) ≥ α? In this appendix we use the standard terminology that for a directed edge i⇒ j, the node i is termed the parent of node j, and the node j is termed the child of i. We start with a simple polynomial reduction from the following decision problem.
Max-Sub-FRG: Given a directed graph G = (V, E), a function w : E→ Z+, and a positive integer α, is there a transitive FRG subgraph G′ = (V, E′) such that∑
e∈E′ w(e) ≥ α? Max-Sub-FRG ≤p Max-Trans-Forest: Given an instance (G = (V, E), w,α) of MaxSub-FRG, we construct the instance (G′ = (V, E′), w′,α) of Max-Trans-Forest such that w′(u, v) = w(u, v) if (u, v) ∈ E and −∞ otherwise. We need to show that (G = (V, E), w,α) ∈Max-Sub-FRG iff (G′ = (V, E′), w′,α) ∈Max-Trans-Forest. This is trivial: If there is a transitive FRG subgraph of G whose sum of edges≥ α, then choosing the same edges will yield a transitive FRG subgraph of G′ whose sum of edges≥ α. Similarly, any
transitive FRG subgraph of G′ whose sum of edges ≥ α cannot use any −∞ edges, and therefore the edges of this FRG are in E and this defines a subgraph of G whose sum of edges ≥ α.
Note that for Max-Sub-FRG, maximizing the sum of weights of edges in the subgraph is equivalent to minimizing the sum of weights of edges not in the subgraph. We now denote by z the sum of weights of the edges deleted from the graph.
Next we show a polynomial reduction from the NP-hard Exact Cover by 3-sets (X3C) problem (Garey and Johnson 1979) to Max-Sub-FRG. The X3C is defined as follows. Given a set X of size 3n, and m subsets, S1, S2, .., Sm of X, each of size 3, decide if there is a collection of n Si’s whose union covers X.
X3C ≤p Max-Sub-FRG: Given an instance (X, S) of X3C problem, we construct an instance (G = (V, E), w, z) of the Max-Sub-FRG problem as follows (an illustration of the construction is given in Figure A.1). First, we construct the vertices V: We construct x1, .., x3n vertices, corresponding to the points of X, m vertices s1, .., sm corresponding to the subsets S, m additional vertices t1, t2, .., tm, and one more vertex a. Next we construct the edges E and the weight function w : E→ Z+.
1. For each 1 ≤ i ≤ m, we add an edge (ti, si) of weight 2.
2. For each 1 ≤ i ≤ m, we add an edge (a, si) of weight 1.
3. For each 1 ≤ j ≤ 3n, we add an edge (a, xj) of weight M = 4m.
4. For each 1 ≤ i ≤ m, if si = {xp, xq, xr}, we add 3 edges of weight 1: (si, xp), (si, xq), and (si, xr).
To complete the construction of the Max-Sub-FRG problem, we define z = 4m− 2n. We next prove that S has an exact 3-cover of X ⇔ there is a transitive FRG subgraph of G such that the sum of weights deleted is no more than 4m− 2n. ⇒: Assume there is an exact 3-cover of X by S. The FRG subgraph will consist of: n edges (a, si) for the n vertices si that cover X, the 3n edges (si, xj), and m− n edges (tf , sf ), for the m− n vertices sf that do not cover X. The transitive closure contains all edges (a, xi), and the rest of the edges are deleted: for all sf ’s that are not part of the cover, the three edges (sf , xj), and the edge (a, sf ) are deleted. In addition, the weight 2 edges (ti, si) for the si’s that cover X are deleted. The total weight deleted is thus 3(m− n) + m− n + 2n = 4m− 2n. It is easy to verify that the subgraph is a transitive FRG - there are no connected components of size > 1, and in the transitive reduction of G there is no node with more than one parent. ⇐: Assume there is no exact 3-cover of X by S, we will show that any FRG subgraph must delete more than 4m− 2n weight. We cannot omit any edge (a, xi), as the weight of each such edge is too large. Thus all these edges are in the FRG and are either deleted during transitive reduction or not.
Assume first that all these edges are deleted in the transitive reduction, that is, for every xj there exists an si such that (a, si) and (si, xj) are in the forest. Since there is no collection of n subsets Si that cover X, there must be at least k > n such si’s. Consequently, for these si’s, the forest must not contain the edges (ti, si) (otherwise, si would have two parents and violate both the forest and the transitivity properties). For the m− k nodes with no edge (sf , xj) we can either add an edge (a, sf ) or (tf , sf ), but not both (otherwise, sf would have more than one parent). Since w(tf , sf ) > w(a, sf ) it is better to delete the (a, sf ) edges. Hence, the total weight of deleted edges is 3m− 3n for the edges between si’s and xj’s, 2k in the edges (ti, si), for si’s that cover the xj’s, and m− k for the edges (a, sf ). Total weight deleted is 4m− 3n + k > 4m− 2n since k > n.
Assume now that r > 0 edges (a, xj) are not deleted in the transitive reduction. This means that for these xj’s there is no edge (si, xj) for any i (otherwise, xj will have more than one parent after transitive reduction). This means that 3n− r of the xj’s are covered by si’s. To cover xj’s we need at least k ≥ dn− r3e si’s. As before, for these si’s we also have the edges (a, si) and we delete the edges (ti, si), and for the m− k nodes sf that do not cover any xj it is best to add the edges (tf , sf ) and to delete the edges (a, sf ). So the weight deleted is 3m− (3n− r) for edges between si and xj, 2k in the edges (ti, si), and m− k for the edges (a, sf ). Thus, the weight deleted is 4m− 3n + k + r ≥ 4m− 3n + dn− r3e+ r ≥ 4m− 2n + r− b r3c > 4m− 2n.
Because it is known that X3C is NP-hard, this polynomial reduction shows that Max-Trans-Forest is also NP-hard.",,,,,,,,,,"1 introduction :Performing textual inference is at the heart of many semantic inference applications, such as Question Answering (QA) and Information Extraction (IE). A prominent generic
∗ Computer Science Department, Stanford University, Stanford, CA 94305. Web: http://www.stanford.edu/∼joberant.
∗∗ Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv, 6997801, Israel. Web: http://www.tau.ac.il/∼nogaa. † Computer Science Department, Bar-Ilan University, Ramat-Gan, 52900, Israel.
Web: http://www.cs.biu.ac.il/∼dagan. ‡ Engineering Department, Bar-Ilan University, Ramat-Gan, 52900, Israel.
Web: http://www.eng.biu.ac.il/goldbej.
Submission received: 23 November 2013; revised version received: 26 September 2014; accepted for publication: 25 November 2014.
doi:10.1162/COLI a 00220
© 2015 Association for Computational Linguistics
paradigm for textual inference is Textual Entailment (Dagan et al. 2013). In textual entailment, the goal is to recognize, given two text fragments termed text and hypothesis, whether the hypothesis can be inferred from the text. For example, the text Cyprus was invaded by the Ottoman Empire in 1571 implies the hypothesis The Ottomans attacked Cyprus.
Semantic inference applications such as QA and IE crucially rely on entailment rules (Ravichandran and Hovy 2002; Shinyama and Sekine 2006; Berant and Liang 2014) or equivalently, inference rules—that is, rules that describe a directional inference relation between two fragments of text. An important type of entailment rule specifies the entailment relation between natural language predicates, for example, the entailment rule X invade Y⇒X attack Y can be helpful in inferring the aforementioned hypothesis. Consequently, substantial effort has been made to learn such rules (Lin and Pantel 2001; Sekine 2005; Szpektor and Dagan 2008; Schoenmackers et al. 2010; Melamud et al. 2013).
Textual entailment is inherently a transitive relation, that is, the rules x⇒ y and y⇒ z imply the rule x⇒ z. For example, from the rules X reduce nausea⇒ X help with nausea and X help with nausea⇒ X associated with nausea we can infer the rule X reduce nausea⇒ X associated with nausea (Figure 1). Accordingly, Berant, Dagan, and Goldberger (2012) proposed taking advantage of transitivity to improve learning of entailment rules. They modeled learning entailment rules as a graph optimization problem, where nodes are predicates and edges represent entailment rules that respect transitivity. To solve this optimization problem, they formulated it as an Integer Linear Program (ILP) and used an off-the-shelf ILP solver to find an exact solution. Indeed, they showed that applying global transitivity constraints results in more accurate graphs (known as entailment graphs) than methods that ignore the property of transitivity.
Although using an off-the-shelf ILP solver is straightforward, finding the optimal set of edges respecting transitivity is NP-hard, and practically, transitivity constraints impose substantial restrictions on the scalability of the methods. In fact, in some cases finding the optimal set of edges for entailment graphs with even ∼50 nodes was quite slow.
In this article, we develop algorithms for learning entailment graphs that take advantage of structural properties such as transitivity, but substantially reduce the computational cost of inference, thus allowing us to solve the aforementioned optimization problem on very large graphs.
Our method contains two main steps. The first step is based on a sparsity assumption that even if an entailment graph contains many predicates, most of them do not entail one another, and thus we can decompose the graph into smaller components that can be solved more efficiently. For example, the predicates X parent of Y, X child of Y and X relative of Y are independent from the predicates X works with Y, X boss of Y, and X manages Y, and thus we can consider each one of these two sets separately. We prove that finding the optimal solution for each of the smaller components results in a global optimal solution for our optimization problem.
The second step proposes a polynomial heuristic approximation algorithm for finding a transitive set of edges in each one of the smaller components. It is based on a novel modeling assumption that entailment graphs exhibit a “tree-like” property, which we term forest reducible. For example, the graph in Figure 1 is not a tree, because the predicates X related to nausea and X associated with nausea form a cycle. However, these two predicates are synonymous, and if they were merged into a single node, then the graph would become a tree. We propose a simple iterative approximation algorithm, where in each iteration a single node is deleted from the graph and then inserted back in a way that improves the objective function value. We prove that if we impose a constraint that entailment graphs must be forest reducible, then each iteration can be performed in linear time. This results in an algorithm that can scale to entailment graphs containing tens of thousands of nodes.
We apply our algorithm on two data sets. The first data set includes medium-sized entailment graphs where predicates are typed, that is, the arguments are restricted to belong to a particular semantic type (for instance, Xperson parent of Yperson). We show that using our algorithm, we can substantially improve runtime, suffering from only a slight reduction in the quality of learned entailment graphs. The second data set includes a much larger graph containing 20,000 untyped nodes, where applying state-of-the-art methods that use an ILP solver is completely impossible. We run our algorithm on this data set and demonstrate that we can learn knowledge bases with more than 100,000 entailment rules at a higher precision compared with local learning methods.
The article is organized as follows. In Section 2 we survey prior work on the learning of entailment rules. We first focus on local methods (Section 2.1), that is, methods that handle each pair of predicates independently of other predicates, and then describe global methods (Section 2.2), that is, methods that take into account multiple predicates simultaneously. Section 3 is the core of the article and describes our main algorithmic contributions. After formalizing entailment rule learning as a graph optimization problem, we present the two steps of our algorithm. Section 3.1 describes the first step, in which a large entailment graph is decomposed into smaller components. Section 3.2 describes the second step, in which we develop an efficient heuristic approximation based on the assumption that entailment graphs are forest reducible. Section 4 describes experiments on the first data set containing medium-sized entailment graphs with typed predicates. Section 5 presents an empirical evaluation on a large graph containing 20,000 untyped predicates. We also perform a qualitative analysis in Section 5.4 to further elucidate the behavior of our algorithm. Section 6 concludes the article.
This article is based on previous work (Berant, Dagan, and Goldberger 2011; Berant et al. 2012), but expands over it in multiple directions. Empirically, we present results on a novel data set (Section 5) that is by orders of magnitude larger than in the past. Algorithmically, we present the Tree-Node-And-Component-Fix algorithm, which is an extension of the Tree-Node-Fix algorithm presented in Berant et al. (2012) and achieves best results in our experimental evaluation. Theoretically, we provide an NP-hardness
proof for the Max-Trans-Forest optimization problem presented in Berant et al. (2012) and an ILP formulation for it. Last, we perform in Section 5.4 an extensive qualitative analysis of the graphs learned by our local and global algorithms from which we draw conclusions for future research directions. 2 background :In this section we describe prior work relevant for entailment rule learning. First, we describe local methods that estimate entailment for a pair of predicates, focusing on methods that we employ in this article. Then, we describe global methods that perform inference over a larger set of predicates. Specifically, we provide details on the method and optimization problem developed by Berant, Dagan, and Goldberger (2012) for which we propose scalable inference algorithms in this article. Last, we survey some other related work in NLP that uses global inference. In local learning, given a pair of predicates (i, j) we would like to determine whether i⇒ j. The main sources of information utilized in the past for local learning were (1) lexicographic resources, (2) co-occurrence, and (3) distributional similarity. We briefly describe the first two and then expand more on distributional similarity, which is the most commonly used source of information.
Lexicographic resources are manually built knowledge bases from which semantic information may be extracted. For example, the hyponymy, toponymy, and synonymy relations in WordNet (Fellbaum 1998) can be used to detect entailment between nouns and verbs. Although WordNet is the most popular lexicographic resource, other resources such as CatVar, Nomlex, and FrameNet have also been utilized to extract inference rules (Meyers et al. 2004; Budanitsky and Hirst 2006; Szpektor and Dagan 2009; Coyne and Rambow 2009; Ben Aharon, Szpektor, and Dagan 2010).
Pattern-based methods attempt to identify the semantic relation between a pair of predicates by examining their co-occurrence in a large corpus. For example, the sentence people snore while they sleep provides evidence that snore ⇒ sleep. While most patternbased methods focused on identifying semantic relations between nouns (for example, Hearst patterns [Hearst 1992]), several works (Pekar 2008; Chambers and Jurafsky 2011; Weisman et al. 2012) attempted to extract relations between predicates as well. Chklovsky and Pantel (2004) used pattern-based methods to generate the commonly used VerbOcean resource.
Both lexicographic as well as pattern-based methods suffer from limited coverage. Distributional similarity is therefore used to automatically construct broad-scale resources. Distributional similarity methods are based on the “distributional hypothesis” (Harris 1954) that semantically similar predicates occur with similar arguments. Quite a few methods have been suggested (Lin and Pantel 2001; Szpektor et al. 2004; Bhagat, Pantel, and Hovy 2007; Szpektor and Dagan 2008; Yates and Etzioni 2009; Schoenmackers et al. 2010), which differ in terms of the specifics of the ways in which predicates are represented, the features that are extracted, and the function used to compute feature vector similarity. Next, we elaborate on the methods that we use in this article.
Lin and Pantel (2001) proposed an algorithm for learning paraphrase relations between binary predicates, that is, predicates with two variables such as X treat Y. For each binary predicate, Lin and Pantel compute two sets of features Fx and Fy,
which are the words that instantiate the arguments X and Y, respectively, in a large corpus. Given a predicate u and its feature set for the X variable Fx, every feature fx ∈ Fx is weighted by pointwise mutual information between the predicate and the feature: w( fx) = log Pr( fx|u) Pr( fx )
, where the probabilities are computed using maximum likelihood over the corpus. Given two predicates u and v, the Lin measure (Lin 1998) is computed for the variable X in the following manner:
Linx(u, v) =
∑ f∈Fux∩Fvx [w u x ( f ) + wvx( f )]∑
f∈Fux w u x ( f ) + ∑ f∈Fvx w v x( f )
(1)
The measure is computed analogously for the variable Y and the final distributional similarity score, termed DIRT, is the geometric average of the scores for the two variables: If DIRT(u, v) is high, this means that the templates u and v share many “informative” arguments and so it is possible that u⇒ v.
Szpektor and Dagan (2008) suggested two modifications to DIRT. First, they looked at unary predicates, that is, predicates with a single variable such as X treat. Secondly, they computed a directional score that is more suited for capturing entailment relations compared to the symmetric Lin score. They proposed that if for two unary predicates u⇒ v, then relatively many of the features of u should be covered by the features of v. This is captured by the asymmetric Cover measure (Weeds and Weir 2003):
Cover(u, v) =
∑ f∈Fu∩Fv w
u( f )∑ f∈Fu wu( f )
(2)
The final directional score, termed BInc (Balanced Inclusion), is the geometric average of the Lin measure and the Cover measure.
Both Lin and Pantel as well as Szpektor and Dagan compute a similarity score using a single argument. However, it is clear that although this alleviates sparsity problems, considering pairs of arguments jointly provides more information. Yates and Etzioni (2009), Schoenmackers et al. (2010), and even earlier, Szpektor et al. (2004), presented methods that compute semantic similarity based on pairs of arguments.
A problem common to all local methods presented above is predicate ambiguity— predicates may have different meanings and different entailment relations in different contexts. Some works resolved the problem of ambiguity by representing predicates with argument variables that are typed (Pantel et al. 2007; Schoenmackers et al. 2010). For example, argument variables in the work of Schoenmackers et al. were restricted to belong to one of 156 types, such as country or profession. A different solution that has attracted substantial attention recently is to represent the various contexts in which a predicate can appear in a low-dimensional latent space (for example, using Latent Dirichlet Allocation [Blei, Ng, and Jordan 2003]) and infer entailment relations between predicates based on the contexts in which they appear (Ritter, Mausam, and Etzioni 2010; ÓSéaghdha 2010; Dinu and Lapata 2010; Melamud et al. 2013). In the experiments presented in this article we will use the representation of Schoenmackers et al. in one experiment, and ignore the problem of predicate ambiguity in the other. The idea of global learning is that, by jointly learning semantic relations between a large number of natural language phrases, one can use the dependencies between the relations to improve accuracy. A natural way to model that is with a graph where nodes are phrases and edges represent semantic similarity. Snow, Jurafsky, and Ng (2006) presented one of the early examples of global learning in the context of learning noun taxonomies. In their work, they enforced a transitivity constraint over the taxonomy using a greedy inference procedure and demonstrated that this improves the quality of the taxonomy. Transitivity constraints were also enforced by Yates and Etzioni (2009), who proposed a clustering algorithm for learning undirected synonymy relations. Nakashole, Weikum, and Suchanek (2012) learned a taxonomy of binary predicates and also enforced transitivity with a greedy algorithm.
Berant, Dagan, and Goldberger (2012) developed a global method for learning entailment relations between predicates. In this article we present scalable approximation algorithms for the optimization problem they propose, and so we provide further detail on their method. The input to their algorithm is a large corpus and a lexicographic resource, such as WordNet, and the output is a set of entailment rules that respect the constraint of transitivity. The algorithm is composed of two main steps. In the first step, a set of predicates is extracted from the corpus and a local entailment classifier is trained based on examples automatically generated from the lexicographic resource. At this point, we can derive for each pair of predicates (i, j) a score wij ∈ R that estimates whether i⇒ j.
In the second step of their algorithm, they construct a graph, where predicates are nodes and edges represent entailment rules. Using the local scores they look for the set of edges E that maximizes the objective function ∑ (i,j)∈E wij under the constraint that edges respect transitivity. They show that this optimization problem is NP-hard and find an exact solution using an ILP solver over small graphs. They also avoid problems of predicate ambiguity by partially contextualizing the predicates. In this article, we present efficient and scalable heuristic approximation algorithms for the optimization problem they propose.
In recent years, there has been substantial work on approximation algorithms for global inference problems. Do and Roth (2010) suggested a method for the related task of learning taxonomic relations between terms. Given a pair of terms, they construct a small graph that contains the two terms and a few other related terms, and then impose constraints on the graph structure. They construct these small graphs because their work is geared towards scenarios where relations are determined on-the-fly for a given pair of terms and no global knowledge base is ever explicitly constructed. Because they independently construct a graph for each pair of terms, their method easily produces solutions where global constraints, such as transitivity, are violated.
Another approximation method that violates transitivity constraints is LP relaxation (Martins, Smith, and Xing 2009). In LP relaxation, binary variables are replaced by continuous variables, transforming the problem from an ILP to a Linear Program (LP), which is polynomial. An LP solver is then applied, and variables that are assigned a fractional value are rounded to their nearest integer, so many violations of transitivity may occur. The solution when applying LP relaxation is not a transitive graph; we will show in Section 4 that our approximation method is substantially faster.
Global inference has gained popularity in recent years in NLP, and a common approximation method that has been extensively utilized is dual decomposition (Sontag, Globerson, and Jaakkola 2011). Dual decomposition has been successfully applied in
tasks such as Information Extraction (Reichart and Barzilay 2012), Machine Translation (Chang and Collins 2011), Parsing (Rush et al. 2012), and Named Entity Recognition (Wang, Che, and Manning 2013). To the best of our knowledge, it has not yet been applied for the task of learning entailment relations. The graph decomposition method we present in Section 3.1 can be viewed as an ideal case of dual decomposition, where we can decompose the problem into disjoint components in a way that we do not need to ensure consistency of the results obtained on each component separately. 3 efficient inference :Our goal is to learn a large knowledge base of entailment rules between natural language predicates. Following Berant, Dagan, and Goldberger (2012), we formulate this task as a graph-learning problem, where, given the nodes of an entailment graph, we would like to find the best set of edges that respect a global transitivity constraint.
Our main modeling assumption is that textual entailment is a transitive relation. Berant, Dagan, and Goldberger (2012) have demonstrated that this modeling assumption holds for focused entailment graphs, that is, graphs in which predicates are disambiguated by one of their arguments. For example, a graph that focuses on the concept nausea might contain a en entailment rule between predicates such as X prevents nausea ⇒X affects nausea, where the predicate X prevent Y is disambiguated by the instantiation of the argument nausea. In this work, we will examine this modeling assumption for more general graphs. In Section 4 we show that transitivity holds in typed entailment graphs, that is, graphs where each predicate specifies the semantic type of its arguments (for example, Xdrug prevent Ysymptom). In Section 5.4, we examine the assumption of transitivity when predicates do not carry any typing information (for example, X prevent Y); and we observe that in this setting the assumption of transitivity is often violated because of the predicate ambiguity. Nevertheless, we show that even in this set-up we can empirically take advantage of the transitivity assumption.
Let V be a set of predicates, which are the nodes of the entailment graph, and let w : V × V → R be an entailment weighting function. Given a pair of predicates (i, j), a positive wij indicates local tendency to decide that i entails j, whereas a negative wij indicates local tendency to decide that i does not entail j. We want to find a global entailment transitive graph that is most consistent with the local cues. Formally, our goal is to find the directed graph G = (V, E) that maximizes the sum of edge weights∑
(i,j)∈E wij, under the constraint that the graph is transitive—that is, for every triple of nodes (i, j, k), if (i, j) ∈ E and ( j, k) ∈ E, then (i, k) ∈ E.
Berant, Dagan, and Goldberger (2012) proved that this optimization problem (which we term Max-Trans-Graph) is NP-hard, and provided an ILP formulation for it. Let xij be an indicator for whether i⇒ j, then x = {xij : i 6= j} are the variables of the following ILP:
max x ∑ i6=j wijxij (3)
s.t. ∀i, j, k ∈ V xij + xjk − xik ≤ 1
∀i, j ∈ V xij ∈ {0, 1}
The objective function is the sum of weights over the edges of G, and the constraint xij + xjk − xik ≤ 1 on the binary variables enforces transitivity (i.e., xij = xjk = 1 implies that xik = 1). The weighting function w is trained separately using supervised learning methods and we describe the details of training for each one of our experiments in Sections 4 and 5.
There is a simple probabilistic modeling that motivates the score (3) that we optimize. Assume that for each pair of nodes (i, j), we are given a probability pij(1) = p(xij = 1) that i⇒ j (the probability that i ; j is denoted by pij(0) = 1− pij(1)). Assuming a uniform probability over graphs, the posterior probability (which can also be viewed as the likelihood) of a graph, represented by an edge-set x, is:
p(x) ∝ ∏ i6=j pij(xij) (4)
It can be easily verified that:1
log p(xij) = log pij(1) pij(0) xij + log pij(0) (5)
Hence,
log p(x) = ∑ i 6=j log p(xij) = ∑ i6=j wijxij + const (6)
such that ‘const’ is a scalar that does not depend on x and
wij = log pij(1) pij(0)
(7)
The optimal entailment graph, therefore, is arg maxx p(x) = arg maxx ∑
i6=j wijxij, where the maximization is over all the transitive graphs. Hence the most likely transitive entailment graph is obtained as the solution of the ILP maximization problem (3).
Berant, Dagan, and Goldberger solved this optimization problem with an ILP solver, but because ILP is NP-hard, this does not scale well; the number of variables is O(|V|2) and the number of constraints is O(|V|3). Thus, even a graph with 80 nodes (predicates) has more than half a million constraints. In this section, we describe a heuristic algorithm that empirically provides high-quality solutions for this optimization problem in graphs with tens of thousands of nodes.
Our algorithm contains two main steps. The first step (Section 3.1) is based on a structural assumption that entailment graphs are relatively sparse—that is, most predicates do not entail one another. This allows us to decompose the graph into smaller components in a way that guarantees that an exact solution for each one of the components results in an exact solution for Max-Trans-Graph. However, often even after decomposition, components are too large and finding an exact solution is still intractable.
The second step (Section 3.2) proposes a heuristic algorithm for finding a good solution for each one of the components. This step is based on an observation that
1 We assume that pij(1), pij(0) ∈ (0, 1) and thus log p(xij ) is well defined.
entailment graphs exhibit a “tree-like” property and are very similar to a novel type of graph, which we term forest-reducible graph. We utilize this property to develop an iterative efficient approximation algorithm for learning the graph edges, where each iteration takes linear time. We also prove that finding the optimal forest-reducible graph is NP-hard.
In Sections 4 and 5 we apply our algorithm on two different data sets and show that using transitivity substantially improves performance and that our methods dramatically improve scalability, allowing us to increase the scope of global learning of entailment graphs. The first step of our algorithm takes advantage of graph sparsity: Most predicates in language do not entail one another. Thus, it might be possible to decompose entailment graphs into small components and solve each component separately.
Let V1, V2 be a partitioning of the nodes V into two disjoint non-empty subsets. We term any directed edge, whose two nodes belong to different subsets, a crossing edge.
Proposition 1 If we can partition a set of nodes V into disjoint sets V1, V2 such that no crossing edge has a positive weight, then the optimal set of edges that is the solution of the ILP problem (3) does not contain any crossing edge.
Proof Assume by contradiction that the edge-set of the optimal solution Eopt contains a non-empty set of crossing edges Ecross. We can construct Enew = Eopt \ Ecross. Clearly ∑ (i,j)∈Enew wij ≥ ∑ (i,j)∈Eopt wij, as wij ≤ 0 for any crossing edge.
Next, we show that Enew does not violate transitivity constraints. Assuming it does, then the violation is caused by omitting the edges in Ecross. Thus, there must be, without loss of generality, nodes i ∈ V1 and k ∈ V2 such that for some node j, (i, j) and ( j, k) are in Enew, but (i, k) is not. However, this means either (i, j) or ( j, k) is a crossing edge, which is impossible because we omitted all crossing edges. Thus, Enew is a better solution than Eopt, contradiction. Hence, Ecross is empty.
This proposition suggests a simple algorithm (see Algorithm 1): Construct an undirected graph with the node set V and with an edge connecting i and j if either wij > 0 or wji > 0, then find its connected components, and finally solve each component separately. If an ILP solver is used to find the edges of each component separately,
then we obtain an optimal (not approximate) solution to the optimization problem (3) for the whole graph. Finding the undirected edges (Line 1) and computing connected components (Line 2) can be performed in O(V2). Thus, in this case the efficiency of the algorithm is dominated by the application of an ILP solver (Line 4).
If the entailment graph decomposes into small components, one could obtain an exact solution with an ILP solver, applied on each component separately, without resorting to any approximation. To further extend scalability in this setting we use a cutting-plane method (Kelley 1960). Cutting-plane methods have been used in the past, for example, in dependency parsing (Riedel and Clarke 2006). The idea is that even if we omit all transitivity constraints, we still expect most transitivity constraints to be satisfied, given a good weighting function w. Thus, it makes sense to avoid specifying the constraints ahead of time, but rather add them when they are violated. This is formalized in Algorithm 2.
Line 1 initializes an active set of constraints (ACT). Line 3 applies the ILP solver with the active constraints. Lines 4 and 5 find the violated constraints and add them to the active constraints. The algorithm halts when no constraints are violated. The solution is clearly optimal because we obtain a maximal solution for a less-constrained problem that does not violate any transitivity constraint.
A pre-condition for using cutting-plane methods is that computing the violated constraints (Line 4) is efficient, as it occurs in every iteration. We do that in a straightforward manner: For all edges (i, j) and ( j, k) that are in the current solution E∗, if (i, k) /∈ E∗ we add xij + xjk − xik ≤ 1 to the violated constraints. This is cubic in worst-case and performs very fast in practice.
To conclude, an exact algorithm for Max-Trans-Graph decomposes the graph into components and uses Algorithm 2 to find an exact solution for each component. However, because the problem is NP-hard, this will still fail once components become large. Next, we describe the second step of our method, which replaces Algorithm 2 with an efficient approximation that can scale to much larger graphs. We will show in Section 4 that the solutions obtained by the approximation algorithm are almost as good as the exact solution on a data set where finding an exact solution is feasible. The approximation we present in this section is based on a conjecture that entailment graphs exhibit a “tree-like” property, that is, they can be reduced into a structure similar to a directed forest, which we term forest-reducible graph (FRG). Although FRGs are a more constrained class of directed graphs, we prove that restricting our optimization
problem to FRGs does not make the problem fundamentally easier, that is, the problem remains NP-hard (see Appendix). Then, we present in Section 3.2.2 our iterative approximation algorithm, where in each iteration a node is removed and re-attached back to the graph in a locally optimal way. Combining this scheme with our conjecture about the graph structure yields a linear algorithm for node re-attachment.
Thus, the contribution of this section is twofold: First, we define a novel modeling assumption about the tree-like structure of entailment graphs and empirically demonstrate its validity. Second, we exploit this assumption to develop a polynomial approximation algorithm for learning entailment graphs that can scale to much larger graphs than in the past.
3.2.1 Forest-Reducible Graph. The predicate entailment relation, described by our entailment graphs, is typically from a “semantically specific” predicate to a more “general” one. Thus, intuitively, the topology of an entailment graph is expected to be “tree-like.” In this section we first formalize this intuition and then empirically analyze its validity. This property of entailment graphs is an interesting topological observation on its own, but also enables the efficient approximation algorithm of Section 3.2.2.
For a directed edge i⇒ j in a directed acyclic graph (DAG), we term the node i a child of node j, and j a parent of i.2 A directed forest is a DAG where all nodes have no more than one parent. A strongly connected component in a directed graph is a set of nodes such that there is a path from every node in the set to any other node. In entailment graphs, strongly connected components correspond to semantically equivalent predicates.
The entailment graph in Figure 2a (subgraph from the data set described in Section 4) is clearly not a directed forest—it contains a cycle of size two comprising the nodes X common in Y and X frequent in Y, and in addition the node X be epidemic in Y has three parents. However, we can convert it to a directed forest by applying the following operations. Any directed graph G can be converted into a StronglyConnected-Component (SCC) graph in the following way: Every strongly connected component is contracted into a single node, and an edge is added from SCC S1 to SCC S2 if there is an edge in G from some node in S1 to some node in S2. The SCC graph is always a DAG (Cormen et al. 2002), and if G is transitive, then the SCC graph is also transitive. The graph in Figure 2b is the SCC graph of the one in Figure 2a, but is still not a directed forest because the node X be epidemic in Y has two parents.
The transitive closure of a directed graph G is obtained by adding an edge from node i to node j if there is a path in G from i to j. The transitive reduction of G is obtained by removing all edges whose absence does not affect its transitive closure. In DAGs, the result of transitive reduction is unique (Aho, Garey, and Ullman 1972). We thus define the reduced graph Gred = (Vred, Ered) of a directed graph G as the transitive reduction of its SCC graph. The graph in Figure 2c is the reduced graph of the one in Figure 2a and is a directed forest. We say a graph is a forest-reducible graph if its reduced graph is a directed forest.
We now hypothesize that entailment graphs are approximately FRGs. The intuition is that the predicate on the left-hand-side of an entailment rule has a more specific meaning than the one on the right-hand-side. For instance, in Figure 2a X be epidemic in Y (where X is a type of disease and Y is a country) is more specific than X common in Y and
2 In standard graph terminology an edge is from a parent to a child. We choose the opposite definition to conflate edge direction with the direction of the entailment operator ‘⇒’.
X frequent in Y, which are equivalent, while X occur in Y is even more general. Accordingly, the reduced graph in Figure 2c is an FRG. We note that this is not always the case: For example, the entailment graph in Figure 3 is not an FRG, because X annex Y entails both Y be part of X and X invade Y, while the latter two do not entail one another. However, we hypothesize that this scenario is rather uncommon. Consequently, a natural variant of the Max-Trans-Graph problem is to restrict the required output graph of the optimization problem (3) to an FRG. We term this problem Max-Trans-Forest.
To test whether our hypothesis holds empirically, we performed the following analysis. We sampled seven gold standard typed entailment graphs (that is, graphs where
predicates specify the semantic type of their arguments, such as Xdrug prevents Ysymptom) from the data set described in Section 4, manually transformed them into FRGs by deleting a minimal number of edges, and measured recall over the set of edges in each graph (precision is naturally 1.0, as we only delete gold-standard edges). The lowest recall value obtained was 0.95, illustrating that deleting a very small proportion of edges converts a typed entailment graph into an FRG. Further support for the practical validity of this hypothesis is obtained from our experiments on typed entailment graphs in Section 4. In these experiments we show that exactly solving Max-Trans-Graph and Max-Trans-Forest (with an ILP solver) results in nearly identical performance. In Section 5.4 we qualitatively analyze the validity of the FRG assumption in untyped entailment graphs (where predicates are of the form X prevents Y) and find that this assumption does not always hold, because of predicate ambiguity.
An ILP formulation for Max-Trans-Forest is simple—a transitive graph is an FRG if all nodes in its reduced graph have no more than one parent. It can be verified that this is equivalent to the following statement: For every triplet of nodes i, j, k, if i⇒ j and i⇒ k, then either j⇒ k or k⇒ j (or both). This constraint can be stated in a linear form: xij + xik − xjk − xkj ≤ 1. Therefore, adding this new type of constraint to the ILP given in Equation (3) results in a formulation for Max-Trans-Forest:
max x ∑ i 6=j wijxij (8)
such that ∀i, j, k ∈ V xij + xjk − xik ≤ 1
∀i, j, k ∈ V xij + xik − xjk − xkj ≤ 1
∀i, j ∈ V xij ∈ {0, 1}
A natural question is whether there is a simpler (polynomial) solution for Max-TransForest that avoids the need for an ILP solver. In the Appendix we prove that Max-TransForest is also an NP-hard problem by a polynomial reduction from the X3C problem (Garey and Johnson 1979). Readers who are not interested in the proof can safely skip the Appendix.
3.2.2 Approximation Algorithm. In this section we present Tree-Node-Fix and Tree-NodeAnd-Component-Fix, which are efficient approximation algorithms for Max-TransForest, as well as Graph-Node-Fix, an approximation for Max-Trans-Graph.
Tree-Node-Fix. The scheme of Tree-Node-Fix (TNF) is the following. First, an initial FRG is constructed, using some initialization procedure. Then, at each iteration a single node v is re-attached (see below) to the FRG in a way that improves the objective function. This is repeated until the value of the objective function can no longer be improved by re-attaching a node.
Re-attaching a node v is performed by removing v from the graph (deleting v and all its adjacent edges) and connecting it back with a better set of edges, while maintaining the constraint that it is an FRG. This is done by considering all possible edges from/to the other graph nodes and choosing the optimal subset, while the rest of the graph edges remain fixed. For example, in Figure 2, one way of re-attaching the node X common in Y is to add it as a direct child of X occur in Y that is not a synonym of X frequent in Y. This will result in deletion of the edges X common in Y⇒ X frequent in Y, X frequent in Y⇒ X common in Y, and also X be epidemic in Y⇒ X common in Y, because otherwise the
resulting graph will not be an FRG. We will show that re-attachment can be efficiently performed in linear time using dynamic programming. Formally, let Sv−in = ∑
i 6=v wivxiv be the sum of scores over v’s incoming edges and Sv−out = ∑ k 6=v wvkxvk be the sum of scores over v’s outgoing edges. Re-attachment amounts to optimizing a linear objective:
argmax x(v) (Sv-in + Sv-out) (9)
while maintaining the FRG constraint, where the variables x(v) ⊆ x are indicators for all pairs of nodes involving v. We approximate a solution for Equation (3) by iteratively optimizing the simpler objective (9). Clearly, at each re-attachment the value of the objective function cannot decrease, because the optimization algorithm considers the previous graph as one of its candidate solutions.
We now show that re-attaching a node v is linear. To analyze v’s re-attachment, we consider the structure of the directed forest Gred just before v is re-inserted, and examine the possibilities for v’s insertion relative to that structure. We start by defining some helpful notations. Every node c ∈ Vred is a strongly connected component in G. Let vc ∈ c be an arbitrary representative node in c (we can choose any node vc ∈ c, because the graph G is transitive, and c is a strongly connected component). We denote by Sv-in(c) the sum of weights from all nodes in c and their descendants to v, and by Sv-out(c) the sum of weights from v to all nodes in c and their ancestors:
Sv-in(c) = ∑ i∈c wiv + ∑ k /∈c wkvxkvc (10)
Sv-out(c) = ∑ i∈c wvi + ∑ k /∈c wvkxvck (11)
Note that {xvck, xkvc} are edge indicators in G and not Gred. There are several cases for re-attaching v that we need to consider:
1. Inserting v into an existing component c ∈ Vred (case 1, see Figure 4a).
2. Inserting v as a new component, which breaks into two subcases:
(a) Forming a new component that contains only v, where v is a child of a component c (case 2a, see Figure 4b).
(b) Forming a new component that contains only v, where v is not a child of any component c, and so is a new root in Gred (case 2b, see Figure 4c).
Note that v cannot form a new component that contains other nodes as well, because the rest of the graph is fixed.
To find the optimal way of re-attaching v, we need to compute the score Equation (9), for each of the three cases, and choose the best one. We now describe how this score is computed in each of the three cases.
Case 1: Inserting v into a component c ∈ Vred. In this case we add in G edges from all nodes in c and their descendants to v and from v to all nodes in c and their ancestors. The score (9) in this case is
s1(c) , Sv-in(c) + Sv-out(c) (12)
Case 2a: Inserting v as a child of some c ∈ Vred. Once c is chosen as the parent of v, choosing v’s children in Gred is substantially constrained. A node that is not a descendant of c cannot become a child of v, because this would create a new path from that node to c and would require by transitivity to add a corresponding directed edge to c (but all graph edges not connecting v are fixed). Moreover, only a direct child of c can choose v as a parent instead of c (Figure 4b), because for any other descendant of c, v would become a second parent, and Gred will no longer be a directed forest (Figure 4b’). Thus, this case requires adding in G edges from v to all nodes in c and their ancestors; additionally, for each new child of v, denoted by d ∈ Vred, we add edges from all nodes in d and their descendants to v. Crucially, although the number of possible subsets of c’s children in Gred is exponential, the fact that they are independent trees in Gred allows us to go over them one by one, and decide for each one whether it will be a child of v or not, depending on whether Sv-in(d) is positive. Therefore, the score (9) in this case is:
s2(c) , Sv-out(c)+ ∑
d∈child(c) max(0, Sv-in(d)) (13)
where child(c) are the children of c. Case 2b: Inserting v as a new root in Gred. Similar to case 2a, only roots of Gred can become children of v. In this case for each chosen root r we add in G edges from the nodes in r and their descendants to v. Again, each root can be examined independently. Therefore, the score (9) of re-attaching v is:
s3 , ∑
r
max(0, Sv-in(r)) (14)
where the summation is over the roots of Gred. It can be easily verified that Sv-in(c) and Sv-out(c) satisfy the recursive definitions:
Sv-in(c) = ∑ i∈c wiv + ∑ d∈child(c) Sv-in(d), c ∈ Vred (15)
Sv-out(c) = ∑ i∈c wvi + Sv-out(p), c ∈ Vred (16)
where p is the parent of c in Gred. These recursive definitions allow computing in linear time Sv-in(c) and Sv-out(c) for all c (given Gred) using dynamic programming, before going over the cases for re-attaching v. Sv-in(c) is computed going over Vred leaves-to-root (post-order), and Sv-out(c) is computed going over Vred root-to-leaves (pre-order).
Re-attachment is summarized in Algorithm 3. Computing an SCC graph is linear (Cormen et al. 2002), and it is easy to verify that transitive reduction in FRGs is also linear (Line 1). Computing Sv-in(c) and Sv-out(c) (Lines 2–3) is also linear, as explained. Cases 1 and 2b are trivially linear and in case 2a we go over the children of all nodes in Vred. As the reduced graph is a forest, this simply means going over all nodes of Vred, and so the entire algorithm is linear.
Because re-attachment is linear, re-attaching all nodes is quadratic. Thus if we bound the number of iterations over all nodes, the overall complexity is quadratic. This is dramatically more efficient and scalable than applying an ILP solver. Empirically, we found that TNF converges after 5–10 iterations.
Tree-Node-and-Component-Fix. Assuming that entailment graphs are FRGs allows us to use the node re-attachment operation in linear time. However, this assumption also enables performing other graph operations efficiently. We now suggest a natural extension to the TNF algorithm that better explores the space of FRGs.
A disadvantage of TNF is that it re-attaches a single node in every iteration. Consider, for instance, the reduced graph in Figure 5. In this graph, nodes l, m, and n are direct children of node k, but suppose that in the optimal solution they are all children of node j. Reaching the optimal solution would require three independent re-attachment operations, and it is not clear that each of the three alone would improve the objective function value. However, if we allow re-attachment operations over components in the SCC graph, then we would be able to re-attach the strong connectivity component containing the nodes l, m, and n in a single operation. Thus, the idea of our extended TNF algorithm is to allow re-attachment of both nodes and components. We term this algorithm Tree-Node-and-Component-Fix (TNCF).
There are many ways in which this intuition can be implemented and our TNCF algorithm uses one possible variant. In TNCF we first perform node re-attachment until convergence as in TNF (after initialization), but then compute the SCC graph and perform component re-attachment until convergence. Component re-attachment is identical to node re-attachment, except that we are guaranteed that the reduced graph is a forest. After performing component re-attachment until convergence, we again
perform node re-attachments and then component re-attachments and so on, until the entire process converges.
Graph-Node-Fix. The re-attachment strategy described above can be applied without assuming that the graph is an FRG. Next, we show Graph-Node-Fix (GNF), a similar algorithm that uses the re-attachment strategy to obtain an approximate solution for the ILP problem Max-Trans-Graph (3). In this more general case, finding the optimal reattachment of a node v is done with an ILP solver. Nevertheless, this ILP is simpler than Equation (3), because we consider only candidate edges involving v. Figure 6 illustrates the three types of possible transitivity constraint violations when re-attaching v. The left side depicts a violation when (i, k) /∈ E, expressed by the constraint in Equation (18), and the middle and right depict two violations when the edge (i, k) ∈ E, expressed by the constraints in Equation (19). Thus, the ILP is formulated by adding the following constraints to the objective function (9):
Complexity is exponential because of the ILP solver; however, the ILP size is reduced by an order of magnitude to O(|V|) variables and O(|V|2) constraints. To summarize, the complexity of GNF is higher than the complexity of TNF, but it does not require the assumption that entailment graphs are FRGs.  4 experimental evaluation: typed entailment graphs :In this and the next section we empirically evaluate the algorithms presented in Section 3 on two different data sets. Our first data set, presented in this section, comprises medium-sized graphs for which obtaining an exact solution is possible, and we demonstrate that our approximation methods substantially improve runtime while causing only a small degradation in performance compared with the optimal solution. The graphs are also particularly suited for global optimization because graph predicates are typed, which substantially reduces their ambiguity. The second data set (Section 5) contains a graph with tens of thousands of untyped nodes, where exact inference is completely impossible. We show that our methods scale to this graph and that transitivity improves performance even when predicates are untyped. The resulting graph contains more than 100,000 entailment rules that can be utilized in downstream semantic applications.
The input to the algorithms presented in Section 3 is a set of nodes V and a weighting function w : V × V → R. We describe how those are constructed before presenting an experimental evaluation. As mentioned in Section 2, one of the challenges in global optimization is that transitivity does not always hold when predicates are ambiguous. Schoenmackers et al. (2010) proposed an algorithm for learning inference rules between typed predicates, a representation that substantially reduces ambiguity. We adopt their representation and learn typed entailment graphs. A typed entailment graph (see Figure 7) is a directed graph where the nodes are typed predicates. A typed predicate is a triple (t1, p, t2), or simply p(t1, t2), representing a predicate in natural language. p is the lexical realization of the predicate and the typed variables t1, t2 indicate that the arguments of the predicate belong to the semantic types t1, t2. Semantic types are taken from a set of types T, where each type t ∈ T is a bag of natural language words or phrases. Examples for typed predicates are: conquer(COUNTRY,CITY) and contain(PRODUCT,MATERIAL). An instance of a typed predicate is a triple (a1, p, a2), or simply p(a1, a2), where a1 ∈ t1 and a2 ∈ t2
be relate to (drug,drug)
be derive from (drug,drug)
be process from (drug,drug)
be convert into (drug,drug)
238
are termed arguments. For example, be common in(asthma,australia) is an instance of be common in(DISEASE,PLACE).
Given a set of typed predicates, entailment rules can only exist between predicates that share the same (unordered) pair of types (such as PLACE and COUNTRY), as otherwise, the rule would contain unbound variables. Hence, given a set of typed predicates we can immediately decompose them into disjoint subsets—all typed predicates sharing the same pair of types define a separate graph that describes the entailment relations between those predicates (see Figure 7).
Edges in typed entailment graphs represent entailment rules in the usual way. If the type t1 is different from the type t2, mapping of arguments is straightforward, as in the rule be find in(MATERIAL,PRODUCT)⇒ contain(PRODUCT,MATERIAL). If t1 and t2 are equal we need to specify how arguments are mapped. This is done by splitting each node into two: For example, the node beat(TEAM,TEAM) is split into two typed predicates beat(Xteam,Yteam) and beat(Yteam,Xteam). This allows us to specify a rule where argument order is reversed such as beat(Xteam,Yteam)⇒ lose to(Yteam,Xteam). This also accommodates rules such as play(Xteam,Yteam)⇒ play(Yteam,Xteam): If team A plays team B, then team B plays team A.
To create typed entailment graphs, we used a data set that was generously provided by Schoenmackers et al. (2010). Schoenmackers et al. produced a mapping of 1.1 million arguments into 156 types (Examples for (argument, TYPE) pairs are (exodus, BOOK), (china, COUNTRY), and (asthma, DISEASE)), and then utilized the types, the mapped arguments, and 1 million TextRunner tuples (Banko et al. 2007) to generate a set of 10,672 typed predicates.3 Because entailment can only occur between predicates that share the same types, we decomposed the 10,672 typed predicates into 2,303 typed entailment graphs. The largest graph contains 188 nodes and the total number of potential rules is 263,756.4 The advantage of typing predicates is that it substantially reduces ambiguity, while still maintaining rules of wide applicability. The weighting function w is derived from an entailment score provided by a local classifier. Given a local classifier that provides an entailment score sij for a pair of predicates (i, j), we define wij = sij − λ, where λ is a prior that controls graph sparseness: As λ increases, wij decreases and becomes negative for more pairs of predicates, rendering the graph more sparse. The probabilistic interpretation of λ is as follows. For each two nodes let q be a prior probability that an edge exists. For large values of q the graph tends to be dense, and vice versa. Defining λ as log 1−qq , the modified weight function is:
wij = log p(xij = 1) p(xij = 0) − log 1− qq = sij − λ (20)
Given the weight function w, the task is to find the maximum a posteriori global graph that satisfies predefined constraints such as transitivity.
3 Readers are referred to their paper for details on mapping of tuples to typed predicates. The mapping of arguments into types can be downloaded from http://www.cs.washington.edu/research/ sherlock-hornclauses/. 4 In more detail, 1,714 graphs have less than 5 nodes, 326 graphs have 5–10 nodes, 145 graphs have 10–20 nodes, 92 graphs have 20–50 nodes, 18 graphs have 50–100 nodes, and 7 graphs have at least 100 nodes.
Training is similar to the method proposed by Berant, Dagan, and Goldberger (2012), and we briefly describe it here. The input for training is a lexicographic resource, for which we use WordNet, and a set of tuples, for which we use the 1 million typed TextRunner tuples provided by Schoenmackers et al. We perform the following steps:
Type Example
direct hypernym beat(TEAM,TEAM)⇒ play(TEAM,TEAM) direct synonym reach(TEAM,GAME)⇒ arrive at(TEAM,GAME) direct cohyponym invade(COUNTRY,CITY) ; bomb(COUNTRY,CITY) hyponym (distance=2) defeat(CITY,CITY) ; eliminate(CITY,CITY) random hold(PLACE,EVENT) ; win(PLACE,EVENT)
predicate. Then, similarity scores are combined by a geometric average. For example, the binary typed predicate invade(COUNTRY, CITY) will be decomposed into two unary predicates, one where the first argument of invade has the type COUNTRY (and the type of the second argument is unspecified), and another where the second argument of invade has the type CITY (and the first argument in unspecified).
The other five scores were provided by Schoenmackers et al. (Schoenmackers et al. 2010) and include SR (Schoenmackers et al. 2010), LIME (McCreath and Sharma 1997), M-estimate (Dzeroski and Brakto 1992), the standard G-test, and a simple implementation of Cover (Weeds and Weir 2003). Overall, the rationale behind this representation is that combining various scores will yield a better classifier than each single measure.
3. Training We sub-sample negative examples and train over an equal number of positive and negative examples. We used SVMperf (Joachims 2005) to train a Gaussian kernel classifier that provides an output score, sij. We tuned the two SVM parameters using 5-fold cross validation on a development set of two typed entailment graphs.
4. Prior knowledge For a small number of pairs of predicates, we might have prior knowledge of whether one entails the other. Berant, Dagan, and Goldberger (2012) integrated prior knowledge by adding hard constraints to the ILP. Because not all of our algorithms use an ILP solver, we integrate prior knowledge by modifying the local classifier score. For pairs of predicates i, j for which we have prior knowledge that i entails j (termed positive local constraints), we set sij =∞. For pairs of predicates i, j for which we have prior knowledge that i does not entail j (termed negative local constraints), we set sij = −∞.
To generate prior knowledge constraints we normalized each predicate by omitting the first word if it is a modal and turned passives into actives. If two normalized predicates are equal, they are a positive local constraint. Negative local constraints were constructed from three sources (1) Predicates differing by a single pair of words that are WordNet antonyms, (2) Predicates differing by a single word of negation, and (3) Predicates p(t1, t2) and p(t2, t1) where p is a transitive verb (for example, beat) in VerbNet (Kipper, Dang, and Palmer 2000).
As reported by Berant, Dagan, and Goldberger (2012), we find that injecting prior knowledge into the graph improves performance, as it provides a good starting point for the inference procedure. To evaluate performance, we manually annotated all edges in 10 typed entailment graphs containing 14, 14, 22, 30, 53, 62, 76, 86, 118, and 118 nodes. This annotation yielded 3,427 edges and 35,585 non-edges, resulting in an empirical edge density of 9%. The data set is publicly available and can be downloaded from http://www-nlp. stanford.edu/joberant/homepage_files/resources/Acl2011Exp.rar.
We implemented the following algorithms for learning graph edges, where in all of them the graph is first decomposed into components as described in Section 3.1.
No-trans Use local scores without transitivity constraints—an edge (i, j) is inserted iff wij > 0, or in other words iff sij > λ.
Exact-graph Find the optimal solution for Max-Trans-Graph in each component by applying an ILP solver in a cutting-plane method, as described in Section 3.1.
Exact-forest Find the exact solution for Max-Trans-Forest (see Equation 8) in each component by applying an ILP solver in a cutting-plane method.
LP-relax Solve Max-Trans-Graph approximately by applying an LP-relaxation (see Section 2) on each graph component. We apply the LP solver within the same cuttingplane method to allow for a direct comparison. As mentioned, our goal is to present a method for learning transitive graphs, whereas LP-relax produces solutions that violate transitivity. However, we run it on our data set to obtain empirical results, and to compare runtimes against TNF.
Graph-Node-Fix (GNF) Initialize each component in the following way: If the graph is very sparse, that is, λ ≥ C for some constant C (set to 1 in our experiments), then solving the graph exactly is not an issue and we use Exact-graph. Otherwise, we initialize by applying Exact-graph in a sparse configuration, that is, λ = C.
Tree-Node-Fix (TNF) Initialize as in GNF, except that if it generates a graph that is not an FRG, it is corrected by a simple heuristic: For every node in the reduced graph Gred that has more than one parent, we choose from its current parents the single one whose SCC is composed of the largest number of nodes in G.
We do not present the results of TNCF in this experiment, because for medium-sized graphs it provides results that are almost identical to TNF.
The No-trans baseline is a state-of-the-art local learning algorithm. It uses state-ofthe-art local scores such as DIRT (Lin and Pantel 2001), BInc (Szpektor and Dagan 2008), Cover (Weeds and Weir 2003), and SR (Schoenmackers et al. 2010) and trains a classifier using these scores. Berant, Dagan, and Goldberger (2010) have demonstrated empiricially that training a classifier over multiple similarity scores improves performance compared with using just a single similarity score.
The Exact-graph algorithm is the state-of-the-art global method presented by Berant, Dagan, and Goldberger (2012). It finds the optimal solution for Max-TransGraph, and our goal is to show that we can obtain good performance more efficiently using our approximation methods. Last, LP-relax is a standard approximation method for solving ILPs (Martins, Smith, and Xing 2009).
We note that a trivial baseline would be to initialize edges based on whether sij > λ and then compute the transitive closure in each component. We empirically found that this adds edges too aggressively and results in very low precision.
We use the Gurobi optimization package5 as our ILP solver in all experiments. The experiments were run on a multi-core 2.5GHz server with 32GB of RAM.
We evaluate algorithms by comparing the set of gold-standard edges with the set of edges learned by each algorithm. We measure recall, precision, and F1 for various values of the sparseness parameter λ, and compute the area under the precision-recall curve (AUC) generated. Efficiency is evaluated by comparing runtimes.
We first focus on runtimes and show that TNF is substantially more efficient than other baselines that use transitivity. Figure 8 compares runtimes of Exact-graph, GNF, TNF, and LP-relax as −λ increases and the graph becomes denser. Note that the y-axis
5 www.gurobi.com
is in logarithmic scale. Clearly, Exact-graph is extremely slow and runtime increases quickly. For λ = 0.3, runtime was already 12 hours and we were unable to obtain results for λ < 0.3, whereas in TNF we easily got a solution for any λ. When λ = 0.6, where both Exact-graph and TNF achieve best F1, TNF is 10 times faster than Exactgraph. When λ = 0.5, TNF is 50 times faster than Exact-graph, and so on. Most importantly, runtime for GNF and TNF increases much more slowly than for Exact-graph. Comparing runtimes for TNF and GNF, we see that the gap between the algorithms decreases as −λ increases. However, for reasonable values of λ, TNF is about four to seven times faster than GNF, and we were unable to run GNF on large graphs, as we report in Section 5.
Runtime of LP-relax is also bad compared with TNF and GNF. Runtime increases more slowly than Exact-graph, but still very fast compared with TNF. When λ = 0.6, LP-relax is almost 10 times slower than TNF, and when λ = −0.1, LP-relax is 200 times slower than TNF. This points to the difficulty of scaling LP-relax to large graphs. Last, Exact-forest is the slowest algorithm and because it is an approximation of Exact-graph we omit if from the figure for clarity.
We now examine the quality of the learned graphs and validity of our modeling assumptions. Figure 9 (left) shows a pair-wise comparison of Exact-graph and Exactforest. As is evident, the two curves are very similar, and the maximal F1 on the curve and AUC are almost identical. This provides further support for our modeling assumption that entailment graphs are roughly forest-reducible. Figure 9 (right) shows a similar comparison for TNF and GNF. We observe again that the curves are similar and performance is almost identical (maximal F1: 0.41, AUC: 0.31), illustrating that using the FRG assumption when using our approximation algorithm is empirically effective.
In Figure 10, we compare the performance of Exact-graph, which exactly solves Max-Trans-Graph, to our most efficient approximation algorithm, TNF, and to No-trans, a baseline that does not use transitivity at all (GNF and LP-relax are omitted from the
recall
figure to improve readability). We observe that both Exact-graph and TNF substantially outperform No-trans and that TNF’s graph quality is only slightly lower than Exact-graph (which is extremely slow). We report in the caption the maximal F1 on the curve and AUC in the recall range 0–0.5 (the widest range for which we have results for all algorithms). Note that compared with Exact-graph, TNF reduces AUC by merely a point and the maximal F1 score by two points only.
As for LP-relax, results are just slightly lower than Exact-graph (maximal F1: 0.43, AUC: 0.32), but its output is not a transitive graph, and as shown above runtime is quite slow.
To summarize, in this section we have empirically evaluated the algorithms presented in Section 3 on medium-sized graphs for which we can find an optimal solution. Our main findings are as follows:
1. Finding the optimal solution for Exact-graph and Exact-forest results in very similar graphs. This supports our assumption that entailment graphs are approximately forest-reducible.
2. Running GNF and TNF also yields very similar graphs, illustrating that using the FRG assumption when using our re-attachment approximation scheme is effective.
3. Our efficient approximation algorithm, TNF, is substantially faster compared with running an ILP solver that exactly solves Max-Trans-Graph.
4. TNF results in only a slight decrease in quality of learned graphs compared with Exact-graph.
5. TNF, which respects the transitivity constraint, learns graphs of higher quality compared with the local baseline, No-trans.
These findings lead us to believe that the algorithms presented in Section 3 can scale to large entailment graphs. In the next section, we empirically evaluate on a much larger data set for which using ILP solvers is impractical. Table 7 demonstrates cases where TNCF correctly omitted an edge and consequently decomposed a weakly connected component into two. These are classical cases where the global constraint of transitivity helps the algorithm avoid errors. In Example 1, although sij = 0.83 (which is higher than λ = 0.2), the edge is not inserted because this would cause the addition of wrong edges from the LHS component that deals with seeing and visiting to the RHS component, which is concerned with help and support. Example 2 is a case of a pair of predicates that are co-hyponyms or even perhaps antonyms. Transitivity helps TNCF avoid this erroneous rule.
To conclude, our analysis reveals that applying structural constraints encourages global algorithms to omit edges. Although this is often desirable, it can also prevent correct rules from being discovered because of problems of ambiguity. Thus, as discussed in Section 5.3, we believe that an important direction for future research is to apply our global graph learning methods over context-sensitive representations (such as the one presented by Melamud et al. [2013]).
Question 2: Node re-attachment. We would like to understand the effect of applying TNF over the entailment graph, or more precisely, check whether TNF substantially alters the set of graph edges compared with an initialization with HTL-FRG. First, we run both
HTL-FRG and TNF with sparseness parameter λ = 0.2 and compute the proportion of edges that is in their intersection. HTL-FRG learns 48,986 edges and TNF learns 61,186 edges. The number of edges in the intersection is 35,636. This means that 27% of the edges learned by HTL-FRG were removed (13,350 edges) and in addition 25,550 edges were added into the graph. Clearly, this indicates that TNF changes the learned graph substantially.
We exemplify this with a few fragments of graphs learned by HTL-FRG and TNF. Figure 14 shows 11 predicates,11 where the black solid edges correspond to HTL-FRG edges and red dashed edges correspond to TNF edges. Nodes that contain more than a single predicate represent an agreement between TNF and HTL-FRG that all predicates in the node entail one another. Clearly, TNF substantially changes the initialization generated by HTL-FRG. HTL-FRG creates an erroneous component in which give a lot of and generate a lot of are synonymous and mean a lot of entails them. TNF disassembles this component and determines that give a lot of is equivalent to offer plenty of, while generate a lot of entails cause of. Moreover, TNF disconnects the predicate mean a lot of completely. TNF also identifies that the predicates offer a wealth of, provide a wealth of, provide a lot of, and provide plenty of entail offer plenty of. Of course, TNF sometimes causes errors—for example, HTL-FRG decided that cause and cause of are equivalent, but TNF deleted the correct entailment rule cause of ⇒ cause.
Figure 15 provides another interesting example for an improvement in the graph due to the application of TNF. In the initialization, HTL-FRG determined that the predicate encrypt entails the predicates convert and convince rather than the predicates code and encode. This is because the local score provided by the classifier for (encrypt, convert) is 0.997—slightly higher than the local score for (encrypt, encode), which is 0.995. Therefore, HTL-FRG inserts the wrong edge encrypt⇒ convert and then it is unable to insert the correct rule encrypt⇒ encode because of the FRG assumption. After HTL-FRG terminates, TNF goes over the nodes one by one and is able to determine that overall the predicate encrypt fits better as a synonym of the predicates code and encode and reattaches it in the correct location. As an aside, notice that both HTL-FRG and TNF are able to identify in this case the correct directionality of the rule compress⇒ encode. The
11 We do not explicitly write the X and Y variables for brevity, and thus we focus only on predicates where the X variable is a subject.
local score given by the classifier for (compress, encode) is 0.65, whereas the local score for (encode, compress) is –0.11.
As a last example, Figure 16 demonstrates again the problem of predicate ambiguity. The predicate depress can occur in two contexts that, though related, instigate different entailment relations. The first context is when the object of the predicate (the Y argument) is a person and then the meaning of depress is similar to discourage. A second context is when the object is some phenomenon, for example, The serious economical situation depresses consumption levels. In initialization, HTL-FRG generates a set of edges that is compatible with the latter context, determining that the predicate depress entails the predicates lower and bring down. TNF re-attaches depress as a synonym of discourage and deter (because this improves the objective function), resulting in a set of edges that corresponds to the first meaning of depress mentioned above. The methods we use in this article do not allow learning the rules for both contexts because this would result in violations of the transitivity and FRG assumptions. This again emphasizes the importance of learning graphs where predicates are marked by their various senses, which will result in a model that can directly benefit from the methods suggested in this article. 5 experimental evaluation: untyped entailment graphs :In this section we evaluate our methods on a graph with 20,000 nodes. Again, we describe how the nodes V and the weighting function w are constructed. The nodes of the entailment graph we learn in this section are not typed. Although ambiguity is a problem in this setting, we will show that nevertheless transitivity constraints can improve results compared with a state-of-the-art local entailment classifier.
The nodes of the graph were generated from a set of two billion tuples of the form (arg1,predicate,arg2) that were extracted by the Reverb open information extraction system (Fader, Soderland, and Etzioni 2011) over the Clueweb096 data set and were generously provided by the authors of Reverb.
We performed some simple preprocessing over the extracted tuples. Each tuple is accompanied by a confidence value and we discarded tuples with confidence smaller than 0.5. Next, we normalized predicates using a procedure that omits unnecessary content such as modals and adverbs. The entire normalization procedure is publicly available as part of the Reverb project.7 Last, we normalized arguments by replacing pronouns and determiners by the tokens PRONOUN and DET. We also ran the BIU
6 http://lemurproject.org/clueweb09.php/. 7 https://github.com/knowitall/reverb-core/blob/master/core/src/main/java/edu/washington/ cs/knowitall/normalization/RelationString.java.
number normalizer8 and replaced numbers larger than one by the token NUM. After these three steps, we are left with a total of 960 million tuples of which 291 million are distinct.
The number of distinct predicates in the set of extracted tuples is 103,315. Because this is still a large number, we restrict the predicate set to the 10,000 predicates that appear with the highest number of pairs of arguments. As previously explained, each predicate is split into two (for example, defeat is split into X defeat Y and Y defeat X), and so the final number of entailment graph nodes is 20,000. As with typed entailment graphs, the weighting function w is obtained by training a classifier that provides a score sij for all pairs of predicates9 and defining wij = sij − λ. This setting, computing the scores sij, involves the following steps:
1. Training set generation. We use the data set released by Zeichner, Berant, and Dagan (2012), which contains 6,567 entailment rule applications annotated for their validity by crowdsourcing. For example, the data set marks that The exercises alleviate pain⇒ The exercises help ease pain is a valid rule application, whereas Obama want to boost the defense budget⇒ Obama increase the defense budget is an invalid rule application. We extract a single rule from each rule application, for example, from the rule application The exercises alleviate pain⇒ The exercises help ease pain we extract the rule X alleviate Y⇒ X help ease Y. We use half of the data set for training, resulting in 1,224 positive examples and 2,060 negative examples. Another two training examples are X unable to pay Y⇒ X owe Y and X own Y ; Y be sold to X.
2. Feature representation. Each pair of predicates (p1, p2) is represented by a feature vector where the first six are distributional similarity features identical to the first six features described in Section 4.2. In addition, for pairs of predicates for which at least one distributional similarity feature is non-zero, we add lexicographic features computed from WordNet (Fellbaum 1998), VerbOcean (Chklovski and Pantel 2004), and CatVar (Habash and Dorr 2003), as well as string-similarity features. Table 2 provides the exact details of these features. A feature is computed for a pair of predicates (p1, p2) where in this context a predicate is a pair (pred,rev): pred is the lexical realization of the predicate, and rev is a Boolean indicating whether arg1 is X and arg2 is Y or vice versa. Overall, each pair of predicates is represented by 27 features.
3. Training. After obtaining a feature representation for every pair of predicates, we train a Gaussian kernel SVM classifier that optimizes F1 (SVMperf implementation [Joachims 2005]), and tune the parameters C and γ by a grid search combined with 5-fold cross validation. We use the trained local classifier to compute a score sij for all pairs of predicates.
8 http://u.cs.biu.ac.il/~nlp/downloads/normalizer.html. 9 We do not need to run the classifier on 10, 0002 pairs of predicates, because for the majority of predicate
pairs the features are all zeros, and for this set of pairs we can run the classifier only once.
(c) passive-active: The passive-active constraint holds for (p1, p2) and (p2, p1) if p2 is the passive form of p1.
Again, local constraints are integrated into our model by changing the score sij =∞ for positive local constraints and sij = −∞ for negative local constraints. To evaluate performance we use the second half of the data set released by Zeichner, Berant, and Dagan (2012) as a test set. This test set contains 3,283 annotated examples, where 1,734 are covered by the 10,000 nodes of our graph (649 positive examples and 1,085 negative examples). As usual, for each algorithm we compute recall and precision with respect to the gold standard at various points by varying the sparseness parameter λ.
Most methods used over typed graphs in Section 4 (Exact-graph, Exact-forest, LPrelax, and GNF) are intractable because they require an ILP solver and our graph does not decompose into components that are small enough. Therefore, we compare the Notrans local baseline, where transitivity is not used, with the efficient global methods TNF and TNCF. Recall that in Section 4 we initialized TNF by applying an ILP solver in a sparse configuration on each graph component. In this experiment the graph is too large to use an ILP solver and so we initialize TNF and TNCF with a simple heuristic we describe next.
We begin with an empty graph and first sort all pairs of predicates (i, j) for which wij > 0, according to their weight w. Then, we go over predicate pairs one-by-one and perform two operations. First, we verify that inserting (i, j) into the graph does not violate the FRG assumption. This is done by going over all edges (i, k) ∈ E and checking that for every k either ( j, k) ∈ E or (k, j) ∈ E. If this is the case, then the edge (i, j) is a valid candidate edge as the resulting reduced graph will remain a directed forest, otherwise adding it will result in i having two parents in the reduced graph, and we do not insert this edge. Second, we compute the transitive closure Tij, which contains all node pairs that must be edges in the graph in case we insert (i, j), due to transitivity. We compute the change in the value of the objective function if we were to insert Tij into the graph. If this change improves the objective function value, we insert Tij into the graph (and so the value of the objective function increases monotonically). We keep going over the candidate edges from highest score to lowest until the objective function can no longer be improved by inserting a candidate edge. We term this initialization HTL-FRG, because we scan edges from high-score to low-score and maintain an FRG. We add this initialization procedure as another baseline.
Figure 11 presents the precision-recall curve of all global algorithms compared with the local classifier for 0.05 ≤ λ ≤ 2.5. One evident property is that No-trans reaches higher recall values than the global algorithms, which means that adding a global transitivity constraint prevents adding correct edges that have positive weight, since this would cause the addition of many other edges that have negative weight. A possible reason for that is predicate ambiguity, where positive weight edges connect connectivity components through an ambiguous predicate. This suggests that a natural direction for future research is to develop algorithms that do not impose transitivity constraints for rules i⇒ j and j⇒ k, if each rule refers to a different meaning of j. One possible direction is to learn latent “sense” or “topic” variables for each rule (as suggested by Melamud et al. Melamud et al. [2013]). Then, we can use the methods presented in this article
to impose transitivity constraints for each sense separately, and easily allow violations of transitivity across different senses. We provide concrete examples for such cases of ambiguity in our qualitative analysis in Section 5.4.
Although global algorithms are limited in their recall, for recall values of 0.1– 0.2 they substantially improve precision over No-trans. To better observe this result, Table 5 presents the results for the recall range 0.1–0.25. Comparing TNCF and Notrans shows that the TNCF algorithm improves over No-trans by 5–10 precision points:
In the recall range of 0.1–.2, the precision of No-trans is 0.68–0.75, whereas the precision of TNCF is 0.74–0.8. The number of rules learned by TNCF in this recall range is about 100,000.
We use HTL-FRG as an initialization method for TNCF, which on its own is not a very good approximation algorithm. Comparing HTL-FRG with No-trans, we observe that it is unable to reach high recall values and it only marginally improves precision compared with No-trans. Applying TNF over HTL-FRG also provides disappointing results—it increases recall compared with HTL-FRG, but precision is not better than No-trans. Indeed, it seems that although performing node re-attachments monotonically improves the objective function value, it is not powerful enough to learn a graph that is better with respect to our gold standard. On the other hand, adding component re-attachments (TNCF) allows the algorithm to better explore the space of graphs and learn a graph with higher recall and precision than HTL-FRG.
Figure 12 presents the runtime for HTL-FRG, TNF, and TNCF (y-axis in logarithmic scale), all of which yield an FRG and can be run on a large untyped graph. Again, we were unable to obtain results with Exact-graph, Exact-forest, LP-relax, and GNF, because using an ILP solver on a graph with 20,000 nodes was impossible. As expected, HTL-FRG is much faster than both TNF and TNCF, and TNCF is somewhat slower than TNF. However, as mentioned earlier, TNCF is able to improve both precision and recall compared with TNF and HTL-FRG. Note that when λ = 1.2, runtime increases suddenly for both TNF and TNCF. This is because at this point a large connected component is formed, as we explain next.
Figure 13 presents the results of graph decomposition as described in Section 3.1. Evidently, when λ = 0 the largest component constitutes almost all of the graph, which contains 20,000 nodes. Note that when λ = 1.2, the size of the largest component increases suddenly to more than half of the graph nodes. Additionally, TNCF obtained recall of 0.1–0.2 when λ ≤ 0.2, and in this case the number of nodes in the largest
component is 90% of the total number of graph nodes. Thus, contrary to the typed graphs, untyped graphs that contain ambiguous predicates do not decompose well into small components.
To summarize, we demonstrated that our efficient global methods scale to a graph with 20,000 nodes and observed that even when predicates are ambiguous we are able to improve precision for moderate values of recall. Nevertheless, there are indications that our modeling assumptions are violated when applied over a large graph with ambiguous predicates. Transitivity and FRG constraints limit the recall we are able to reach, and the graph does not decompose very well into components. Thus, in future work we would like to apply our algorithms over graphs where predicate ambiguity is modeled and so transitivity constraints can be properly applied. In the previous section we quantitatively evaluated global methods for learning entailment graphs over a large set of untyped predicates. Our results revealed interesting issues that we would like to further investigate:
1. Why do methods that assume entailment graphs are transitive and forest-reducible have limited recall?
2. How much does node re-attachment affect graph structure?
In this section, we try to answer these questions by performing a qualitative analysis that allows us to gain better insight on the behavior of our algorithms.
Question 1: Transitivity and FRG assumptions. We would like to understand why using global algorithms such as TNCF results in limited recall. Our hypothesis is that because predicates in the graph are untyped and potentially ambiguous, violations of the
assumptions are possible, which results in recall reduction. To examine this hypothesis, we perform two qualitative analyses:
1. We analyze graphs containing only gold-standard edges, and manually identify cases where the modeling assumptions are violated.
2. We compare the local algorithm, No-trans, to the global algorithm, TNCF. If our hypothesis is correct, then we expect TNCF to remove correct edges that are inserted by No-trans due to ambiguous predicates.
We start by constructing a graph with all edges from our gold-standard data set and search for transitivity and FRG violations caused by ambiguous predicates. Note that the manually annotated gold-standard edges are only a subset of the full set of correct edges, which is actually much larger. Thus, we cannot automatically identify violations, because the graph is missing many true edges. Instead, we manually go over candidate violations and check if indeed this case is a result of a violation of our modeling assumptions, and if so then our global algorithm cannot predict the correct set of edges.
One example is the pair of rules seek ⇒ apply for and apply for ⇒ file for. The first rule was annotated in the rule application Others seek medical care ⇒ Others apply for medical care (Recall that crowdsourcing was used to annotate rule applications, and rules were extracted from these applications.) The second rule was extracted from the rule application Students apply for more than one award⇒ Student file for more than one award. This causes a transitivity violation because Students seek more than one award ; Student file for more than one award. Evidently, the meaning of the predicate apply for is contextdependent. Another example is the pair of rules contain ⇒ supply and supply ⇒ serve annotated in the applications The page contains links⇒ The page supplies links and Users supply information or material⇒Users serve information or material. Clearly, contain ; serve and the transitivity violation is caused by the fact that the predicate supply is contextdependent—in the first context the subject is inanimate, whereas in the second context the subject is human.
A good example for an FRG violation is the predicate come from. The following three entailment rules are annotated by the turkers: come from ⇒ be raised in, come from⇒ be derived from, and come from⇒ come out of. These correspond to the following rule applications: The Messiah comes from Judah ⇒ The Messiah was raised in Judah, The colors come from the sun light ⇒ The colors are derived from the sun light, and The truth comes from the book ⇒ The truth comes out of the book. Clearly, be raised in ; be derived from and be derived from ; be raised in, so this is an FRG violation. Indeed, the three applications correspond to different meanings of the predicate come from, which depend on context. A second example is the pair of rules be dedicated to⇒ be committed to and be dedicated to⇒ contain information on. Again, this is an FRG violation because be committed to ; contain information on and contain information on ; be committed to. This violation occurs because of the ambiguity of the predicate be dedicated to, which can be resolved by knowing whether the subject is human or is an object that carries information (for example, The web site is dedicated to the life of lizards).
This first analysis revealed particular cases where the transitivity and FRG assumptions do not hold. In the next analysis, we would like to directly compare cases where global and local methods disagree with one another. We hypothesize that because predicates are not disambiguated, then global algorithms will delete correct edges due
to an ambiguous predicate. We set the sparseness parameter λ = 0.2 and compare the set of edges learned by No-trans to the set of edges learned by TNCF.
Examining the disagreements, we observe that, as expected, in 90% of the cases TNCF deletes an edge that was inserted by No-trans. We divide these deleted edges into two categories in the following manner: We compute the weakly connected components of the graph learned by TNCF,10 and then for each edge inserted by No-trans and deleted by TNCF we check whether it connects two different weakly connected components or not. There are two reasons for focusing on this property of deleted edges. First, we hypothesize that most deleted edges will connect two different connected components, because such edges result in many violations of transitivity. Second, weakly connected components usually correspond to separate semantic meanings, and thus a correct edge that connects two weakly connected components probably involves an ambiguous predicate.
Indeed, out of all edges where No-trans inserts an edge and TNCF does not, 76.4% connect weakly connected components in the global graph. This proves that deleting such edges is the main cause for disagreement between TNCF and No-trans. Now, we can take a closer look at these edges, and check whether indeed when a correct edge is deleted, it is because of an ambiguous predicate.
We randomly sample five cases where TNCF erroneously deleted an edge connecting weakly connected components—these are cases where ambiguity is likely to occur. For comparison, we also sample five cases where TNCF was right and No-trans erred. Table 6 shows the samples where No-trans is correct. The first column describes the rule and the second column specifies the score sij provided by the local classifier No-trans. The last two columns detail two examples for predicates that are in the same weakly connected component with the rule LHS and RHS in the graph learned by TNCF. These two columns provide a flavor for the overall meaning of predicates that belong to this component. Table 7 is equivalent and shows the samples where TNCF is correct.
Example 1 in Table 6 already demonstrates the problem of ambiguity. The LHS for this rule application in the crowdsourcing experiment was your parents turn off comments and indeed in this case the RHS your parents cut off comments is inferred. However, we can see that the LHS component revolves around the turning off or shutting off of appliances, for example, whereas the RHS component deals with a more explicit and physical meaning of cutting. This is because the predicate X cut off Y is ambiguous and consequently TNCF omits this correct edge. Example 5 also demonstrates well the problem of ambiguity—the LHS of this rule application is This site was put by fans, which in this context entails the RHS This site was run by fans. However, the more common sense of be put by does not entail the predicate be run by; this sense is captured by the predicates in the LHS component Y help put X and X be placed by Y. Example 4 is another good example where the ambiguous predicate is X be open to Y—the RHS component is concerned with the physical state of being opened, whereas the LHS component has a more abstract sense of availability. The LHS of the rule application in this case was The services are available to museums. Examples 2 and 3 are cases where the problem is not in predicate ambiguity but in the connected component itself. Thus, out of five randomly sampled examples where TNCF deleted an edge that connects two weakly connected components, three can be attributed to predicate ambiguity.
10 A weekly connected component in a directed graph is a set of nodes S such that for every pair of nodes u, v ∈ S there is an undirected path from u to v and from v to u, that is, there is a path when edge directions are ignored. 6 conclusions :The problem of language variability is at the heart of many semantic applications such as Information Extraction, Question Answering, Semantic Parsing, and more. Consequently, learning broad coverage knowledge bases of entailment rules and paraphrases (Ganitkevitch, Van Durme, and Callison-Burch 2013) has proved crucial for systems that perform inference over textual representations (Stern and Dagan 2012; Angeli and Manning 2014).
In this article, we have presented algorithms for learning entailment rules between predicates that can scale to large sets of predicates. Our work builds on prior work by Berant, Dagan, and Goldberger (2012), who defined the concept of entailment graphs, and formulated entailment rule learning as a graph optimization problem, where the graph has to obey the structural constraint of transitivity.
Our main contribution is a heuristic polynomial approximation algorithm that can learn entailment graphs containing tens of thousands of nodes. The algorithm is based on two main observations: first, that entailment graphs are sparse, and second, that entailment graphs are forest-reducible. This leads to an efficient algorithm in which at the first step the graph is decomposed into smaller components, and in the second step we find a set of edges for each component. The second step is an iterative approximation algorithm in which at each iteration a node or component of the graph is deleted and then re-attached in a way that improves the overall objective function.
We performed a thorough empirical evaluation, in which we ran our algorithm on two separate data sets. On the first data set we witnessed a substantial improvement in runtime at a relatively small cost in performance. On the second data set, we show that our method scales to an untyped entailment graph containing 20,000 nodes, much larger than was previously possible using state-of-the-art global methods. We see that using a global algorithm improves precision for modest values of recall, but that local methods can achieve higher recall than global methods. We perform an analysis and conclude that the main reason for this limitation is predicate ambiguity, which results in graphs that do not adhere to the structural constraints of transitivity and forest reducibility.
In future work, we hope to address the problem of predicate ambiguity. A natural direction is to combine methods that model the context-sensitivity of predicates with global structural assumptions. In this way, one can apply structural constraints for a set of predicates, given that they appear in similar contexts. A second important direction for future work is to demonstrate that using entailment rules learned by our method improves performance in downstream semantic applications such as QA, IE, and Recognizing Textual Entailment. Entailment rules between predicates are fundamental to many semantic-inference applications. Consequently, learning such rules has been an active field of research in recent years. Methods for learning entailment rules between predicates that take into account dependencies between different rules (e.g., entailment is a transitive relation) have been shown to improve rule quality, but suffer from scalability issues, that is, the number of predicates handled is often quite small. In this article, we present methods for learning transitive graphs that contain tens of thousands of nodes, where nodes represent predicates and edges correspond to entailment rules (termed entailment graphs). Our methods are able to scale to a large number of predicates by exploiting structural properties of entailment graphs such as the fact that they exhibit a “tree-like” property. We apply our methods on two data sets and demonstrate that our methods find high-quality solutions faster than methods proposed in the past, and moreover our methods for the first time scale to large graphs containing 20,000 nodes and more than 100,000 edges. [{""affiliations"": [], ""name"": ""Jonathan Berant""}, {""affiliations"": [], ""name"": ""Noga Alon""}, {""affiliations"": [], ""name"": ""Ido Dagan""}, {""affiliations"": [], ""name"": ""Jacob Goldberger""}] SP:b18e02d0b34892a803b1595d68c0b380680e7437 [{""authors"": [""Aho"", ""Alfred V."", ""Michael R. Garey"", ""Jeffrey D. Ullman.""], ""title"": ""The transitive reduction of a directed graph"", ""venue"": ""SIAM Journal on Computing, 1(2):131\u2013137."", ""year"": 1972}, {""authors"": [""Angeli"", ""Gabor"", ""Christopher D. Manning.""], ""title"": ""Naturalli: Natural logic inference for common sense reasoning"", ""venue"": ""Proceedings of EMNLP, pages 534\u2013545, Doha."", ""year"": 2014}, {""authors"": [""Oren Etzioni.""], ""title"": ""Open information extraction from the web"", ""venue"": ""Proceedings of IJCAI, pages 68\u201374, Hyderabad."", ""year"": 2007}, {""authors"": [""Ben Aharon"", ""Roni"", ""Idan Szpektor"", ""Ido Dagan.""], ""title"": ""Generating entailment rules from FrameNet"", ""venue"": ""Proceedings of ACL, pages 241\u2013246, Uppsala."", ""year"": 2010}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Meni Adler"", ""Jacob Goldberger""], ""title"": ""Efficient tree-based approximation for entailment"", ""year"": 2012}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Jacob Goldberger.""], ""title"": ""Global learning of focused entailment graphs"", ""venue"": ""Proceedings of ACL, pages 1,220\u20131,229, Uppsala."", ""year"": 2010}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Jacob Goldberger.""], ""title"": ""Global learning of typed entailment rules"", ""venue"": ""Proceedings of ACL, pages 610\u2013619, Portland, OR."", ""year"": 2011}, {""authors"": [""Berant"", ""Jonathan"", ""Ido Dagan"", ""Jacob Goldberger.""], ""title"": ""Learning entailment relations by global graph structure optimization"", ""venue"": ""Journal of Computational Linguistics,"", ""year"": 2012}, {""authors"": [""Berant"", ""Jonathan"", ""Percy Liang.""], ""title"": ""Semantic parsing via paraphrasing"", ""venue"": ""Proceedings of ACL, pages 1,415\u20131,425, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Bhagat"", ""Rahul"", ""Patrick Pantel"", ""Eduard Hovy.""], ""title"": ""LEDIR: An unsupervised algorithm for learning directionality of inference rules"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 161\u2013170, Prague."", ""year"": 2007}, {""authors"": [""Blei"", ""David M."", ""Andrew Y. Ng"", ""Michael I. Jordan.""], ""title"": ""Latent Dirichlet allocation"", ""venue"": ""Journal of Machine Learning Research, 3:993\u20131022."", ""year"": 2003}, {""authors"": [""Budanitsky"", ""Alexander"", ""Graeme Hirst.""], ""title"": ""Evaluating WordNet-based measures of lexical semantic relatedness"", ""venue"": ""Computational Linguistics, 32(1):13\u201347."", ""year"": 2006}, {""authors"": [""Chambers"", ""Nathanael"", ""Dan Jurafsky.""], ""title"": ""Template-based information extraction without the templates"", ""venue"": ""Proceedings of ACL, pages 976\u2013986, Portland, OR."", ""year"": 2011}, {""authors"": [""Chang"", ""Yin-Wen"", ""Michael Collins.""], ""title"": ""Exact decoding of phrase-based translation models through Lagrangian relaxation"", ""venue"": ""Proceedings of EMNLP, pages 26\u201337, Edinburgh."", ""year"": 2011}, {""authors"": [""Chklovski"", ""Timothy"", ""Patrick Pantel.""], ""title"": ""Verb ocean: Mining the web for fine-grained semantic verb relations"", ""venue"": ""Proceedings of EMNLP, pages 33\u201340, Barcelona."", ""year"": 2004}, {""authors"": [""Cormen"", ""Thomas H."", ""Charles E. Leiserson"", ""Ronald L. Rivest"", ""Clifford Stein.""], ""title"": ""Introduction to Algorithms"", ""venue"": ""The MIT Press."", ""year"": 2002}, {""authors"": [""Coyne"", ""Bob"", ""Owen Rambow.""], ""title"": ""Lexpar: A freely available English paraphrase lexicon automatically extracted from FrameNet"", ""venue"": ""Proceedings of IEEE International Conference on Semantic"", ""year"": 2009}, {""authors"": [""Dinu"", ""Georgiana"", ""Mirella Lapata.""], ""title"": ""Topic models for meaning similarity in context"", ""venue"": ""Proceedings of COLING, pages 250\u2013258, Beijing."", ""year"": 2010}, {""authors"": [""Do"", ""Quang"", ""Dan Roth.""], ""title"": ""Constraints based taxonomic relation classification"", ""venue"": ""Proceedings of EMNLP, pages 1,099\u20131,109, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Dzeroski"", ""Saso"", ""Ivan Brakto.""], ""title"": ""Handling noise in inductive logic programming"", ""venue"": ""Proceedings of the International Workshop on Inductive Logic Programming, pages 48\u201364, Tokyo."", ""year"": 1992}, {""authors"": [""Fader"", ""Anthony"", ""Stephen Soderland"", ""Oren Etzioni.""], ""title"": ""Identifying relations for open information extraction"", ""venue"": ""Proceedings of EMNLP, pages 1,535\u20131,545, Portland, OR."", ""year"": 2011}, {""authors"": [""Fellbaum"", ""Christiane""], ""title"": ""WordNet: An Electronic Lexical Database (Language, Speech, and Communication)"", ""year"": 1998}, {""authors"": [""Ganitkevitch"", ""Juri"", ""Benjamin Van Durme"", ""Chris Callison-Burch.""], ""title"": ""PPDB: The paraphrase database"", ""venue"": ""Proceedings of HLT-NAACL, pages 758\u2013764, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Garey"", ""Michael R."", ""David S. Johnson.""], ""title"": ""Computers and Intractability: A Guide to the Theory of NP-Completeness"", ""venue"": ""W. H. Freeman."", ""year"": 1979}, {""authors"": [""Habash"", ""Nizar"", ""Bonnie Dorr.""], ""title"": ""A categorial variation database for English"", ""venue"": ""Proceedings NAACL-HLT, pages 17\u201323, Edmonton."", ""year"": 2003}, {""authors"": [""Harris"", ""Zellig.""], ""title"": ""Distributional structure"", ""venue"": ""Word, 10(23):146\u2013162."", ""year"": 1954}, {""authors"": [""Hearst"", ""Marti.""], ""title"": ""Automatic acquisition of hyponyms from large text corpora"", ""venue"": ""Proceedings of COLING, pages 23\u201328, Nantes."", ""year"": 1992}, {""authors"": [""Joachims"", ""Thorsten.""], ""title"": ""A support vector method for multivariate performance measures"", ""venue"": ""Proceedings of ICML, pages 377\u2013384, Bonn."", ""year"": 2005}, {""authors"": [""J.E. Kelley""], ""title"": ""The cutting-plane method for solving convex programs"", ""venue"": ""Journal of the Society for Industrial Applied Mathematics, 8:703\u2013712."", ""year"": 1960}, {""authors"": [""Kipper"", ""Karin"", ""Hoa T. Dang"", ""Martha Palmer.""], ""title"": ""Class-based construction of a verb lexicon"", ""venue"": ""Proceedings of AAAI, pages 691\u2013696, Austin, TX."", ""year"": 2000}, {""authors"": [""Lin"", ""Dekang.""], ""title"": ""Automatic retrieval and clustering of similar words"", ""venue"": ""Proceedings of COLING-ACL, pages 768\u2013774, Montreal."", ""year"": 1998}, {""authors"": [""Lin"", ""Dekang"", ""Patrick Pantel""], ""title"": ""Discovery of inference rules for question"", ""year"": 2001}, {""authors"": [""Martins"", ""Andre"", ""Noah Smith"", ""Eric Xing.""], ""title"": ""Concise integer linear programming formulations for dependency parsing"", ""venue"": ""Proceedings of ACL, pages 342\u2013350, Singapore."", ""year"": 2009}, {""authors"": [""McCreath"", ""Eric"", ""Arun Sharma.""], ""title"": ""ILP with noise and fixed example size: a Bayesian approach"", ""venue"": ""Proceedings of IJCAI, pages 1,310\u20131,315, Nagoya."", ""year"": 1997}, {""authors"": [""Melamud"", ""Oren"", ""Jonathan Berant"", ""Ido Dagan"", ""Jacob Goldberger"", ""Idan Szpektor.""], ""title"": ""A two level model for context sensitive inference rules"", ""venue"": ""Proceedings of ACL, pages 1,331\u20131,340, Sofia."", ""year"": 2013}, {""authors"": [""Meyers"", ""Adam"", ""Ruth Reeves"", ""Catherine Macleod"", ""Rachel Szekeley"", ""Veronika Zielinska"", ""Brian Young.""], ""title"": ""The cross-breeding of dictionaries"", ""venue"": ""Proceedings of LREC, pages 1,095\u20131,098,"", ""year"": 2004}, {""authors"": [""Nakashole"", ""Ndapandula"", ""Gerhard Weikum"", ""Fabian M. Suchanek.""], ""title"": ""Patty: A taxonomy of relational patterns with semantic types"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 1,135\u20131,145,"", ""year"": 2012}, {""authors"": [""\u00d3S\u00e9aghdha"", ""Diarmuid.""], ""title"": ""Latent variable models of selectional preference"", ""venue"": ""Proceedings of ACL, pages 435\u2013444, Uppsala."", ""year"": 2010}, {""authors"": [""Pantel"", ""Patrick"", ""Rahul Bhagat"", ""Bonaventura Coppola"", ""Timothy Chklovski"", ""Eduard H. Hovy.""], ""title"": ""ISP: Learning inferential selectional preferences"", ""venue"": ""Proceedings of NAACL-HLT, pages 564\u2013571,"", ""year"": 2007}, {""authors"": [""Pekar"", ""Viktor.""], ""title"": ""Discovery of event entailment knowledge from text corpora"", ""venue"": ""Computer Speech & Language, 22(1):1\u201316."", ""year"": 2008}, {""authors"": [""Ravichandran"", ""Deepak"", ""Eduard H. Hovy.""], ""title"": ""Learning surface text patterns for a question answering system"", ""venue"": ""Proceedings of ACL, pages 41\u201347, Philadelphia, PA."", ""year"": 2002}, {""authors"": [""Reichart"", ""Roi"", ""Regina Barzilay.""], ""title"": ""Multi-event extraction guided by global constraints"", ""venue"": ""Proceedings of NAACL-HLT, pages 70\u201379, Montreal."", ""year"": 2012}, {""authors"": [""Richens"", ""Tom.""], ""title"": ""Anomalies in the WordNet verb hierarchy"", ""venue"": ""Proceedings of COLING, pages 729\u2013736, Manchester."", ""year"": 2008}, {""authors"": [""Riedel"", ""Sebastian"", ""James Clarke.""], ""title"": ""Incremental integer linear programming for non-projective dependency parsing"", ""venue"": ""Proceedings of EMNLP, pages 129\u2013137, Sydney."", ""year"": 2006}, {""authors"": [""Ritter"", ""Alan"", ""Mausam"", ""Oren Etzioni.""], ""title"": ""A latent Dirichlet allocation method for selectional preferences"", ""venue"": ""Proceedings of ACL, pages 424\u2013434, Uppsala."", ""year"": 2010}, {""authors"": [""Roth"", ""Michael"", ""Anette Frank.""], ""title"": ""Aligning predicates across monolingual comparable texts using graph-based clustering"", ""venue"": ""Proceedings of EMNLP-CoNLL, pages 171\u2013182, Jeju Island."", ""year"": 2012}, {""authors"": [""Rush"", ""Alexander M."", ""Roi Reichart"", ""Michael Collins"", ""Amir Globerson.""], ""title"": ""Improved parsing and POS tagging using inter-sentence consistency constraints"", ""venue"": ""Proceedings of EMNLP, pages 1434\u20131444,"", ""year"": 2012}, {""authors"": [""Schoenmackers"", ""Stefan"", ""Jesse Davis"", ""Oren Etzioni"", ""Daniel S. Weld.""], ""title"": ""Learning first-order horn clauses from web text"", ""venue"": ""Proceedings of EMNLP, pages 1,088\u20131,098, Cambridge, MA."", ""year"": 2010}, {""authors"": [""Sekine"", ""Satoshi.""], ""title"": ""Automatic paraphrase discovery based on context and keywords between NE pairs"", ""venue"": ""Proceedings of IWP, pages 80\u201387, Jeju Island."", ""year"": 2005}, {""authors"": [""Shinyama"", ""Yusuke"", ""Satoshi Sekine.""], ""title"": ""Preemptive information extraction using unrestricted relation discovery"", ""venue"": ""HLT-NAACL, pages 304\u2013311, New-York, NY."", ""year"": 2006}, {""authors"": [""Snow"", ""Rion"", ""Daniel Jurafsky"", ""Andrew Y. Ng.""], ""title"": ""Semantic taxonomy induction from heterogenous evidence"", ""venue"": ""Proceedings of ACL, pages 801\u2013808, Sydney."", ""year"": 2006}, {""authors"": [""Sontag"", ""David"", ""Amir Globerson"", ""Tommi Jaakkola.""], ""title"": ""Introduction to dual decomposition for inference"", ""venue"": ""Suvrit Sra, Sebastian Nowozin, and Stephen J. Wright, editors, Optimization for Machine"", ""year"": 2011}, {""authors"": [""Stern"", ""Asher"", ""Ido Dagan.""], ""title"": ""BIUTEE: A modular open-source system for recognizing textual entailment"", ""venue"": ""Proceedings of ACL System Demonstrations, pages 73\u201378."", ""year"": 2012}, {""authors"": [""Szpektor"", ""Idan"", ""Ido Dagan.""], ""title"": ""Learning entailment rules for unary templates"", ""venue"": ""Proceedings of COLING, pages 849\u2013856, Manchester."", ""year"": 2008}, {""authors"": [""Szpektor"", ""Idan"", ""Ido Dagan.""], ""title"": ""Augmenting WordNet-based inference with argument mapping"", ""venue"": ""Proceedings of TextInfer, pages 27\u201335, Singapore."", ""year"": 2009}, {""authors"": [""Szpektor"", ""Idan"", ""Hristo Tanev"", ""Ido Dagan"", ""Bonaventura Coppola.""], ""title"": ""Scaling web-based acquisition of entailment relations"", ""venue"": ""Proceedings of EMNLP, pages 41\u201348, Barcelona."", ""year"": 2004}, {""authors"": [""Wang"", ""Mengqiu"", ""Wanxiang Che"", ""Christopher D. Manning.""], ""title"": ""Effective bilingual constraints for semi-supervised learning of named entity recognizers"", ""venue"": ""Proceedings of AAAI, pages 919\u2013925,"", ""year"": 2013}, {""authors"": [""Weeds"", ""Julie"", ""David Weir.""], ""title"": ""A general framework for distributional similarity"", ""venue"": ""Proceedings of EMNLP, pages 81\u201388, Sapporo."", ""year"": 2003}, {""authors"": [""Weisman"", ""Hila"", ""Jonathan Berant"", ""Idan Szpektor"", ""Ido Dagan""], ""title"": ""Learning verb inference rules"", ""year"": 2012}, {""authors"": [""Yates"", ""Alexander"", ""Oren Etzioni.""], ""title"": ""Unsupervised methods for determining object and relation synonyms on the web"", ""venue"": ""Journal of Artificial Intelligence Research, 34:255\u2013296."", ""year"": 2009}, {""authors"": [""Zeichner"", ""Naomi"", ""Jonathan Berant"", ""Ido Dagan.""], ""title"": ""Crowdsourcing inference-rule evaluation"", ""venue"": ""Proceedings of ACL (short papers), pages 156\u2013160, Jeju Island."", ""year"": 2012}] appendix a. max-trans-forest is np-hard :In this appendix we show that the Max-Trans-Forest problem (8) is NP-hard. We prove that the following decision problem (which is clearly easier than Max-TransForest) is NP-hard: Given a set of nodes V, a function w : V × V → R, and an integer number α, is there a transitive forest-reducible (FRG) subgraph G = (V, E) such that∑
e∈E w(e) ≥ α? In this appendix we use the standard terminology that for a directed edge i⇒ j, the node i is termed the parent of node j, and the node j is termed the child of i. We start with a simple polynomial reduction from the following decision problem.
Max-Sub-FRG: Given a directed graph G = (V, E), a function w : E→ Z+, and a positive integer α, is there a transitive FRG subgraph G′ = (V, E′) such that∑
e∈E′ w(e) ≥ α? Max-Sub-FRG ≤p Max-Trans-Forest: Given an instance (G = (V, E), w,α) of MaxSub-FRG, we construct the instance (G′ = (V, E′), w′,α) of Max-Trans-Forest such that w′(u, v) = w(u, v) if (u, v) ∈ E and −∞ otherwise. We need to show that (G = (V, E), w,α) ∈Max-Sub-FRG iff (G′ = (V, E′), w′,α) ∈Max-Trans-Forest. This is trivial: If there is a transitive FRG subgraph of G whose sum of edges≥ α, then choosing the same edges will yield a transitive FRG subgraph of G′ whose sum of edges≥ α. Similarly, any
transitive FRG subgraph of G′ whose sum of edges ≥ α cannot use any −∞ edges, and therefore the edges of this FRG are in E and this defines a subgraph of G whose sum of edges ≥ α.
Note that for Max-Sub-FRG, maximizing the sum of weights of edges in the subgraph is equivalent to minimizing the sum of weights of edges not in the subgraph. We now denote by z the sum of weights of the edges deleted from the graph.
Next we show a polynomial reduction from the NP-hard Exact Cover by 3-sets (X3C) problem (Garey and Johnson 1979) to Max-Sub-FRG. The X3C is defined as follows. Given a set X of size 3n, and m subsets, S1, S2, .., Sm of X, each of size 3, decide if there is a collection of n Si’s whose union covers X.
X3C ≤p Max-Sub-FRG: Given an instance (X, S) of X3C problem, we construct an instance (G = (V, E), w, z) of the Max-Sub-FRG problem as follows (an illustration of the construction is given in Figure A.1). First, we construct the vertices V: We construct x1, .., x3n vertices, corresponding to the points of X, m vertices s1, .., sm corresponding to the subsets S, m additional vertices t1, t2, .., tm, and one more vertex a. Next we construct the edges E and the weight function w : E→ Z+.
1. For each 1 ≤ i ≤ m, we add an edge (ti, si) of weight 2.
2. For each 1 ≤ i ≤ m, we add an edge (a, si) of weight 1.
3. For each 1 ≤ j ≤ 3n, we add an edge (a, xj) of weight M = 4m.
4. For each 1 ≤ i ≤ m, if si = {xp, xq, xr}, we add 3 edges of weight 1: (si, xp), (si, xq), and (si, xr).
To complete the construction of the Max-Sub-FRG problem, we define z = 4m− 2n. We next prove that S has an exact 3-cover of X ⇔ there is a transitive FRG subgraph of G such that the sum of weights deleted is no more than 4m− 2n. ⇒: Assume there is an exact 3-cover of X by S. The FRG subgraph will consist of: n edges (a, si) for the n vertices si that cover X, the 3n edges (si, xj), and m− n edges (tf , sf ), for the m− n vertices sf that do not cover X. The transitive closure contains all edges (a, xi), and the rest of the edges are deleted: for all sf ’s that are not part of the cover, the three edges (sf , xj), and the edge (a, sf ) are deleted. In addition, the weight 2 edges (ti, si) for the si’s that cover X are deleted. The total weight deleted is thus 3(m− n) + m− n + 2n = 4m− 2n. It is easy to verify that the subgraph is a transitive FRG - there are no connected components of size > 1, and in the transitive reduction of G there is no node with more than one parent. ⇐: Assume there is no exact 3-cover of X by S, we will show that any FRG subgraph must delete more than 4m− 2n weight. We cannot omit any edge (a, xi), as the weight of each such edge is too large. Thus all these edges are in the FRG and are either deleted during transitive reduction or not.
Assume first that all these edges are deleted in the transitive reduction, that is, for every xj there exists an si such that (a, si) and (si, xj) are in the forest. Since there is no collection of n subsets Si that cover X, there must be at least k > n such si’s. Consequently, for these si’s, the forest must not contain the edges (ti, si) (otherwise, si would have two parents and violate both the forest and the transitivity properties). For the m− k nodes with no edge (sf , xj) we can either add an edge (a, sf ) or (tf , sf ), but not both (otherwise, sf would have more than one parent). Since w(tf , sf ) > w(a, sf ) it is better to delete the (a, sf ) edges. Hence, the total weight of deleted edges is 3m− 3n for the edges between si’s and xj’s, 2k in the edges (ti, si), for si’s that cover the xj’s, and m− k for the edges (a, sf ). Total weight deleted is 4m− 3n + k > 4m− 2n since k > n.
Assume now that r > 0 edges (a, xj) are not deleted in the transitive reduction. This means that for these xj’s there is no edge (si, xj) for any i (otherwise, xj will have more than one parent after transitive reduction). This means that 3n− r of the xj’s are covered by si’s. To cover xj’s we need at least k ≥ dn− r3e si’s. As before, for these si’s we also have the edges (a, si) and we delete the edges (ti, si), and for the m− k nodes sf that do not cover any xj it is best to add the edges (tf , sf ) and to delete the edges (a, sf ). So the weight deleted is 3m− (3n− r) for edges between si and xj, 2k in the edges (ti, si), and m− k for the edges (a, sf ). Thus, the weight deleted is 4m− 3n + k + r ≥ 4m− 3n + dn− r3e+ r ≥ 4m− 2n + r− b r3c > 4m− 2n.
Because it is known that X3C is NP-hard, this polynomial reduction shows that Max-Trans-Forest is also NP-hard.","1 introduction :Performing textual inference is at the heart of many semantic inference applications, such as Question Answering (QA) and Information Extraction (IE). A prominent generic
∗ Computer Science Department, Stanford University, Stanford, CA 94305. Web: http://www.stanford.edu/∼joberant.
∗∗ Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv, 6997801, Israel. Web: http://www.tau.ac.il/∼nogaa. † Computer Science Department, Bar-Ilan University, Ramat-Gan, 52900, Israel.
Web: http://www.cs.biu.ac.il/∼dagan. ‡ Engineering Department, Bar-Ilan University, Ramat-Gan, 52900, Israel.
Web: http://www.eng.biu.ac.il/goldbej.
Submission received: 23 November 2013; revised version received: 26 September 2014; accepted for publication: 25 November 2014.
doi:10.1162/COLI a 00220
© 2015 Association for Computational Linguistics
paradigm for textual inference is Textual Entailment (Dagan et al. 2013). In textual entailment, the goal is to recognize, given two text fragments termed text and hypothesis, whether the hypothesis can be inferred from the text. For example, the text Cyprus was invaded by the Ottoman Empire in 1571 implies the hypothesis The Ottomans attacked Cyprus.
Semantic inference applications such as QA and IE crucially rely on entailment rules (Ravichandran and Hovy 2002; Shinyama and Sekine 2006; Berant and Liang 2014) or equivalently, inference rules—that is, rules that describe a directional inference relation between two fragments of text. An important type of entailment rule specifies the entailment relation between natural language predicates, for example, the entailment rule X invade Y⇒X attack Y can be helpful in inferring the aforementioned hypothesis. Consequently, substantial effort has been made to learn such rules (Lin and Pantel 2001; Sekine 2005; Szpektor and Dagan 2008; Schoenmackers et al. 2010; Melamud et al. 2013).
Textual entailment is inherently a transitive relation, that is, the rules x⇒ y and y⇒ z imply the rule x⇒ z. For example, from the rules X reduce nausea⇒ X help with nausea and X help with nausea⇒ X associated with nausea we can infer the rule X reduce nausea⇒ X associated with nausea (Figure 1). Accordingly, Berant, Dagan, and Goldberger (2012) proposed taking advantage of transitivity to improve learning of entailment rules. They modeled learning entailment rules as a graph optimization problem, where nodes are predicates and edges represent entailment rules that respect transitivity. To solve this optimization problem, they formulated it as an Integer Linear Program (ILP) and used an off-the-shelf ILP solver to find an exact solution. Indeed, they showed that applying global transitivity constraints results in more accurate graphs (known as entailment graphs) than methods that ignore the property of transitivity.
Although using an off-the-shelf ILP solver is straightforward, finding the optimal set of edges respecting transitivity is NP-hard, and practically, transitivity constraints impose substantial restrictions on the scalability of the methods. In fact, in some cases finding the optimal set of edges for entailment graphs with even ∼50 nodes was quite slow.
In this article, we develop algorithms for learning entailment graphs that take advantage of structural properties such as transitivity, but substantially reduce the computational cost of inference, thus allowing us to solve the aforementioned optimization problem on very large graphs.
Our method contains two main steps. The first step is based on a sparsity assumption that even if an entailment graph contains many predicates, most of them do not entail one another, and thus we can decompose the graph into smaller components that can be solved more efficiently. For example, the predicates X parent of Y, X child of Y and X relative of Y are independent from the predicates X works with Y, X boss of Y, and X manages Y, and thus we can consider each one of these two sets separately. We prove that finding the optimal solution for each of the smaller components results in a global optimal solution for our optimization problem.
The second step proposes a polynomial heuristic approximation algorithm for finding a transitive set of edges in each one of the smaller components. It is based on a novel modeling assumption that entailment graphs exhibit a “tree-like” property, which we term forest reducible. For example, the graph in Figure 1 is not a tree, because the predicates X related to nausea and X associated with nausea form a cycle. However, these two predicates are synonymous, and if they were merged into a single node, then the graph would become a tree. We propose a simple iterative approximation algorithm, where in each iteration a single node is deleted from the graph and then inserted back in a way that improves the objective function value. We prove that if we impose a constraint that entailment graphs must be forest reducible, then each iteration can be performed in linear time. This results in an algorithm that can scale to entailment graphs containing tens of thousands of nodes.
We apply our algorithm on two data sets. The first data set includes medium-sized entailment graphs where predicates are typed, that is, the arguments are restricted to belong to a particular semantic type (for instance, Xperson parent of Yperson). We show that using our algorithm, we can substantially improve runtime, suffering from only a slight reduction in the quality of learned entailment graphs. The second data set includes a much larger graph containing 20,000 untyped nodes, where applying state-of-the-art methods that use an ILP solver is completely impossible. We run our algorithm on this data set and demonstrate that we can learn knowledge bases with more than 100,000 entailment rules at a higher precision compared with local learning methods.
The article is organized as follows. In Section 2 we survey prior work on the learning of entailment rules. We first focus on local methods (Section 2.1), that is, methods that handle each pair of predicates independently of other predicates, and then describe global methods (Section 2.2), that is, methods that take into account multiple predicates simultaneously. Section 3 is the core of the article and describes our main algorithmic contributions. After formalizing entailment rule learning as a graph optimization problem, we present the two steps of our algorithm. Section 3.1 describes the first step, in which a large entailment graph is decomposed into smaller components. Section 3.2 describes the second step, in which we develop an efficient heuristic approximation based on the assumption that entailment graphs are forest reducible. Section 4 describes experiments on the first data set containing medium-sized entailment graphs with typed predicates. Section 5 presents an empirical evaluation on a large graph containing 20,000 untyped predicates. We also perform a qualitative analysis in Section 5.4 to further elucidate the behavior of our algorithm. Section 6 concludes the article.
This article is based on previous work (Berant, Dagan, and Goldberger 2011; Berant et al. 2012), but expands over it in multiple directions. Empirically, we present results on a novel data set (Section 5) that is by orders of magnitude larger than in the past. Algorithmically, we present the Tree-Node-And-Component-Fix algorithm, which is an extension of the Tree-Node-Fix algorithm presented in Berant et al. (2012) and achieves best results in our experimental evaluation. Theoretically, we provide an NP-hardness
proof for the Max-Trans-Forest optimization problem presented in Berant et al. (2012) and an ILP formulation for it. Last, we perform in Section 5.4 an extensive qualitative analysis of the graphs learned by our local and global algorithms from which we draw conclusions for future research directions. 4 experimental evaluation: typed entailment graphs :In this and the next section we empirically evaluate the algorithms presented in Section 3 on two different data sets. Our first data set, presented in this section, comprises medium-sized graphs for which obtaining an exact solution is possible, and we demonstrate that our approximation methods substantially improve runtime while causing only a small degradation in performance compared with the optimal solution. The graphs are also particularly suited for global optimization because graph predicates are typed, which substantially reduces their ambiguity. The second data set (Section 5) contains a graph with tens of thousands of untyped nodes, where exact inference is completely impossible. We show that our methods scale to this graph and that transitivity improves performance even when predicates are untyped. The resulting graph contains more than 100,000 entailment rules that can be utilized in downstream semantic applications.
The input to the algorithms presented in Section 3 is a set of nodes V and a weighting function w : V × V → R. We describe how those are constructed before presenting an experimental evaluation. As mentioned in Section 2, one of the challenges in global optimization is that transitivity does not always hold when predicates are ambiguous. Schoenmackers et al. (2010) proposed an algorithm for learning inference rules between typed predicates, a representation that substantially reduces ambiguity. We adopt their representation and learn typed entailment graphs. A typed entailment graph (see Figure 7) is a directed graph where the nodes are typed predicates. A typed predicate is a triple (t1, p, t2), or simply p(t1, t2), representing a predicate in natural language. p is the lexical realization of the predicate and the typed variables t1, t2 indicate that the arguments of the predicate belong to the semantic types t1, t2. Semantic types are taken from a set of types T, where each type t ∈ T is a bag of natural language words or phrases. Examples for typed predicates are: conquer(COUNTRY,CITY) and contain(PRODUCT,MATERIAL). An instance of a typed predicate is a triple (a1, p, a2), or simply p(a1, a2), where a1 ∈ t1 and a2 ∈ t2
be relate to (drug,drug)
be derive from (drug,drug)
be process from (drug,drug)
be convert into (drug,drug)
238
are termed arguments. For example, be common in(asthma,australia) is an instance of be common in(DISEASE,PLACE).
Given a set of typed predicates, entailment rules can only exist between predicates that share the same (unordered) pair of types (such as PLACE and COUNTRY), as otherwise, the rule would contain unbound variables. Hence, given a set of typed predicates we can immediately decompose them into disjoint subsets—all typed predicates sharing the same pair of types define a separate graph that describes the entailment relations between those predicates (see Figure 7).
Edges in typed entailment graphs represent entailment rules in the usual way. If the type t1 is different from the type t2, mapping of arguments is straightforward, as in the rule be find in(MATERIAL,PRODUCT)⇒ contain(PRODUCT,MATERIAL). If t1 and t2 are equal we need to specify how arguments are mapped. This is done by splitting each node into two: For example, the node beat(TEAM,TEAM) is split into two typed predicates beat(Xteam,Yteam) and beat(Yteam,Xteam). This allows us to specify a rule where argument order is reversed such as beat(Xteam,Yteam)⇒ lose to(Yteam,Xteam). This also accommodates rules such as play(Xteam,Yteam)⇒ play(Yteam,Xteam): If team A plays team B, then team B plays team A.
To create typed entailment graphs, we used a data set that was generously provided by Schoenmackers et al. (2010). Schoenmackers et al. produced a mapping of 1.1 million arguments into 156 types (Examples for (argument, TYPE) pairs are (exodus, BOOK), (china, COUNTRY), and (asthma, DISEASE)), and then utilized the types, the mapped arguments, and 1 million TextRunner tuples (Banko et al. 2007) to generate a set of 10,672 typed predicates.3 Because entailment can only occur between predicates that share the same types, we decomposed the 10,672 typed predicates into 2,303 typed entailment graphs. The largest graph contains 188 nodes and the total number of potential rules is 263,756.4 The advantage of typing predicates is that it substantially reduces ambiguity, while still maintaining rules of wide applicability. The weighting function w is derived from an entailment score provided by a local classifier. Given a local classifier that provides an entailment score sij for a pair of predicates (i, j), we define wij = sij − λ, where λ is a prior that controls graph sparseness: As λ increases, wij decreases and becomes negative for more pairs of predicates, rendering the graph more sparse. The probabilistic interpretation of λ is as follows. For each two nodes let q be a prior probability that an edge exists. For large values of q the graph tends to be dense, and vice versa. Defining λ as log 1−qq , the modified weight function is:
wij = log p(xij = 1) p(xij = 0) − log 1− qq = sij − λ (20)
Given the weight function w, the task is to find the maximum a posteriori global graph that satisfies predefined constraints such as transitivity.
3 Readers are referred to their paper for details on mapping of tuples to typed predicates. The mapping of arguments into types can be downloaded from http://www.cs.washington.edu/research/ sherlock-hornclauses/. 4 In more detail, 1,714 graphs have less than 5 nodes, 326 graphs have 5–10 nodes, 145 graphs have 10–20 nodes, 92 graphs have 20–50 nodes, 18 graphs have 50–100 nodes, and 7 graphs have at least 100 nodes.
Training is similar to the method proposed by Berant, Dagan, and Goldberger (2012), and we briefly describe it here. The input for training is a lexicographic resource, for which we use WordNet, and a set of tuples, for which we use the 1 million typed TextRunner tuples provided by Schoenmackers et al. We perform the following steps:
Type Example
direct hypernym beat(TEAM,TEAM)⇒ play(TEAM,TEAM) direct synonym reach(TEAM,GAME)⇒ arrive at(TEAM,GAME) direct cohyponym invade(COUNTRY,CITY) ; bomb(COUNTRY,CITY) hyponym (distance=2) defeat(CITY,CITY) ; eliminate(CITY,CITY) random hold(PLACE,EVENT) ; win(PLACE,EVENT)
predicate. Then, similarity scores are combined by a geometric average. For example, the binary typed predicate invade(COUNTRY, CITY) will be decomposed into two unary predicates, one where the first argument of invade has the type COUNTRY (and the type of the second argument is unspecified), and another where the second argument of invade has the type CITY (and the first argument in unspecified).
The other five scores were provided by Schoenmackers et al. (Schoenmackers et al. 2010) and include SR (Schoenmackers et al. 2010), LIME (McCreath and Sharma 1997), M-estimate (Dzeroski and Brakto 1992), the standard G-test, and a simple implementation of Cover (Weeds and Weir 2003). Overall, the rationale behind this representation is that combining various scores will yield a better classifier than each single measure.
3. Training We sub-sample negative examples and train over an equal number of positive and negative examples. We used SVMperf (Joachims 2005) to train a Gaussian kernel classifier that provides an output score, sij. We tuned the two SVM parameters using 5-fold cross validation on a development set of two typed entailment graphs.
4. Prior knowledge For a small number of pairs of predicates, we might have prior knowledge of whether one entails the other. Berant, Dagan, and Goldberger (2012) integrated prior knowledge by adding hard constraints to the ILP. Because not all of our algorithms use an ILP solver, we integrate prior knowledge by modifying the local classifier score. For pairs of predicates i, j for which we have prior knowledge that i entails j (termed positive local constraints), we set sij =∞. For pairs of predicates i, j for which we have prior knowledge that i does not entail j (termed negative local constraints), we set sij = −∞.
To generate prior knowledge constraints we normalized each predicate by omitting the first word if it is a modal and turned passives into actives. If two normalized predicates are equal, they are a positive local constraint. Negative local constraints were constructed from three sources (1) Predicates differing by a single pair of words that are WordNet antonyms, (2) Predicates differing by a single word of negation, and (3) Predicates p(t1, t2) and p(t2, t1) where p is a transitive verb (for example, beat) in VerbNet (Kipper, Dang, and Palmer 2000).
As reported by Berant, Dagan, and Goldberger (2012), we find that injecting prior knowledge into the graph improves performance, as it provides a good starting point for the inference procedure. To evaluate performance, we manually annotated all edges in 10 typed entailment graphs containing 14, 14, 22, 30, 53, 62, 76, 86, 118, and 118 nodes. This annotation yielded 3,427 edges and 35,585 non-edges, resulting in an empirical edge density of 9%. The data set is publicly available and can be downloaded from http://www-nlp. stanford.edu/joberant/homepage_files/resources/Acl2011Exp.rar.
We implemented the following algorithms for learning graph edges, where in all of them the graph is first decomposed into components as described in Section 3.1.
No-trans Use local scores without transitivity constraints—an edge (i, j) is inserted iff wij > 0, or in other words iff sij > λ.
Exact-graph Find the optimal solution for Max-Trans-Graph in each component by applying an ILP solver in a cutting-plane method, as described in Section 3.1.
Exact-forest Find the exact solution for Max-Trans-Forest (see Equation 8) in each component by applying an ILP solver in a cutting-plane method.
LP-relax Solve Max-Trans-Graph approximately by applying an LP-relaxation (see Section 2) on each graph component. We apply the LP solver within the same cuttingplane method to allow for a direct comparison. As mentioned, our goal is to present a method for learning transitive graphs, whereas LP-relax produces solutions that violate transitivity. However, we run it on our data set to obtain empirical results, and to compare runtimes against TNF.
Graph-Node-Fix (GNF) Initialize each component in the following way: If the graph is very sparse, that is, λ ≥ C for some constant C (set to 1 in our experiments), then solving the graph exactly is not an issue and we use Exact-graph. Otherwise, we initialize by applying Exact-graph in a sparse configuration, that is, λ = C.
Tree-Node-Fix (TNF) Initialize as in GNF, except that if it generates a graph that is not an FRG, it is corrected by a simple heuristic: For every node in the reduced graph Gred that has more than one parent, we choose from its current parents the single one whose SCC is composed of the largest number of nodes in G.
We do not present the results of TNCF in this experiment, because for medium-sized graphs it provides results that are almost identical to TNF.
The No-trans baseline is a state-of-the-art local learning algorithm. It uses state-ofthe-art local scores such as DIRT (Lin and Pantel 2001), BInc (Szpektor and Dagan 2008), Cover (Weeds and Weir 2003), and SR (Schoenmackers et al. 2010) and trains a classifier using these scores. Berant, Dagan, and Goldberger (2010) have demonstrated empiricially that training a classifier over multiple similarity scores improves performance compared with using just a single similarity score.
The Exact-graph algorithm is the state-of-the-art global method presented by Berant, Dagan, and Goldberger (2012). It finds the optimal solution for Max-TransGraph, and our goal is to show that we can obtain good performance more efficiently using our approximation methods. Last, LP-relax is a standard approximation method for solving ILPs (Martins, Smith, and Xing 2009).
We note that a trivial baseline would be to initialize edges based on whether sij > λ and then compute the transitive closure in each component. We empirically found that this adds edges too aggressively and results in very low precision.
We use the Gurobi optimization package5 as our ILP solver in all experiments. The experiments were run on a multi-core 2.5GHz server with 32GB of RAM.
We evaluate algorithms by comparing the set of gold-standard edges with the set of edges learned by each algorithm. We measure recall, precision, and F1 for various values of the sparseness parameter λ, and compute the area under the precision-recall curve (AUC) generated. Efficiency is evaluated by comparing runtimes.
We first focus on runtimes and show that TNF is substantially more efficient than other baselines that use transitivity. Figure 8 compares runtimes of Exact-graph, GNF, TNF, and LP-relax as −λ increases and the graph becomes denser. Note that the y-axis
5 www.gurobi.com
is in logarithmic scale. Clearly, Exact-graph is extremely slow and runtime increases quickly. For λ = 0.3, runtime was already 12 hours and we were unable to obtain results for λ < 0.3, whereas in TNF we easily got a solution for any λ. When λ = 0.6, where both Exact-graph and TNF achieve best F1, TNF is 10 times faster than Exactgraph. When λ = 0.5, TNF is 50 times faster than Exact-graph, and so on. Most importantly, runtime for GNF and TNF increases much more slowly than for Exact-graph. Comparing runtimes for TNF and GNF, we see that the gap between the algorithms decreases as −λ increases. However, for reasonable values of λ, TNF is about four to seven times faster than GNF, and we were unable to run GNF on large graphs, as we report in Section 5.
Runtime of LP-relax is also bad compared with TNF and GNF. Runtime increases more slowly than Exact-graph, but still very fast compared with TNF. When λ = 0.6, LP-relax is almost 10 times slower than TNF, and when λ = −0.1, LP-relax is 200 times slower than TNF. This points to the difficulty of scaling LP-relax to large graphs. Last, Exact-forest is the slowest algorithm and because it is an approximation of Exact-graph we omit if from the figure for clarity.
We now examine the quality of the learned graphs and validity of our modeling assumptions. Figure 9 (left) shows a pair-wise comparison of Exact-graph and Exactforest. As is evident, the two curves are very similar, and the maximal F1 on the curve and AUC are almost identical. This provides further support for our modeling assumption that entailment graphs are roughly forest-reducible. Figure 9 (right) shows a similar comparison for TNF and GNF. We observe again that the curves are similar and performance is almost identical (maximal F1: 0.41, AUC: 0.31), illustrating that using the FRG assumption when using our approximation algorithm is empirically effective.
In Figure 10, we compare the performance of Exact-graph, which exactly solves Max-Trans-Graph, to our most efficient approximation algorithm, TNF, and to No-trans, a baseline that does not use transitivity at all (GNF and LP-relax are omitted from the
recall
figure to improve readability). We observe that both Exact-graph and TNF substantially outperform No-trans and that TNF’s graph quality is only slightly lower than Exact-graph (which is extremely slow). We report in the caption the maximal F1 on the curve and AUC in the recall range 0–0.5 (the widest range for which we have results for all algorithms). Note that compared with Exact-graph, TNF reduces AUC by merely a point and the maximal F1 score by two points only.
As for LP-relax, results are just slightly lower than Exact-graph (maximal F1: 0.43, AUC: 0.32), but its output is not a transitive graph, and as shown above runtime is quite slow.
To summarize, in this section we have empirically evaluated the algorithms presented in Section 3 on medium-sized graphs for which we can find an optimal solution. Our main findings are as follows:
1. Finding the optimal solution for Exact-graph and Exact-forest results in very similar graphs. This supports our assumption that entailment graphs are approximately forest-reducible.
2. Running GNF and TNF also yields very similar graphs, illustrating that using the FRG assumption when using our re-attachment approximation scheme is effective.
3. Our efficient approximation algorithm, TNF, is substantially faster compared with running an ILP solver that exactly solves Max-Trans-Graph.
4. TNF results in only a slight decrease in quality of learned graphs compared with Exact-graph.
5. TNF, which respects the transitivity constraint, learns graphs of higher quality compared with the local baseline, No-trans.
These findings lead us to believe that the algorithms presented in Section 3 can scale to large entailment graphs. In the next section, we empirically evaluate on a much larger data set for which using ILP solvers is impractical. Table 7 demonstrates cases where TNCF correctly omitted an edge and consequently decomposed a weakly connected component into two. These are classical cases where the global constraint of transitivity helps the algorithm avoid errors. In Example 1, although sij = 0.83 (which is higher than λ = 0.2), the edge is not inserted because this would cause the addition of wrong edges from the LHS component that deals with seeing and visiting to the RHS component, which is concerned with help and support. Example 2 is a case of a pair of predicates that are co-hyponyms or even perhaps antonyms. Transitivity helps TNCF avoid this erroneous rule.
To conclude, our analysis reveals that applying structural constraints encourages global algorithms to omit edges. Although this is often desirable, it can also prevent correct rules from being discovered because of problems of ambiguity. Thus, as discussed in Section 5.3, we believe that an important direction for future research is to apply our global graph learning methods over context-sensitive representations (such as the one presented by Melamud et al. [2013]).
Question 2: Node re-attachment. We would like to understand the effect of applying TNF over the entailment graph, or more precisely, check whether TNF substantially alters the set of graph edges compared with an initialization with HTL-FRG. First, we run both
HTL-FRG and TNF with sparseness parameter λ = 0.2 and compute the proportion of edges that is in their intersection. HTL-FRG learns 48,986 edges and TNF learns 61,186 edges. The number of edges in the intersection is 35,636. This means that 27% of the edges learned by HTL-FRG were removed (13,350 edges) and in addition 25,550 edges were added into the graph. Clearly, this indicates that TNF changes the learned graph substantially.
We exemplify this with a few fragments of graphs learned by HTL-FRG and TNF. Figure 14 shows 11 predicates,11 where the black solid edges correspond to HTL-FRG edges and red dashed edges correspond to TNF edges. Nodes that contain more than a single predicate represent an agreement between TNF and HTL-FRG that all predicates in the node entail one another. Clearly, TNF substantially changes the initialization generated by HTL-FRG. HTL-FRG creates an erroneous component in which give a lot of and generate a lot of are synonymous and mean a lot of entails them. TNF disassembles this component and determines that give a lot of is equivalent to offer plenty of, while generate a lot of entails cause of. Moreover, TNF disconnects the predicate mean a lot of completely. TNF also identifies that the predicates offer a wealth of, provide a wealth of, provide a lot of, and provide plenty of entail offer plenty of. Of course, TNF sometimes causes errors—for example, HTL-FRG decided that cause and cause of are equivalent, but TNF deleted the correct entailment rule cause of ⇒ cause.
Figure 15 provides another interesting example for an improvement in the graph due to the application of TNF. In the initialization, HTL-FRG determined that the predicate encrypt entails the predicates convert and convince rather than the predicates code and encode. This is because the local score provided by the classifier for (encrypt, convert) is 0.997—slightly higher than the local score for (encrypt, encode), which is 0.995. Therefore, HTL-FRG inserts the wrong edge encrypt⇒ convert and then it is unable to insert the correct rule encrypt⇒ encode because of the FRG assumption. After HTL-FRG terminates, TNF goes over the nodes one by one and is able to determine that overall the predicate encrypt fits better as a synonym of the predicates code and encode and reattaches it in the correct location. As an aside, notice that both HTL-FRG and TNF are able to identify in this case the correct directionality of the rule compress⇒ encode. The
11 We do not explicitly write the X and Y variables for brevity, and thus we focus only on predicates where the X variable is a subject.
local score given by the classifier for (compress, encode) is 0.65, whereas the local score for (encode, compress) is –0.11.
As a last example, Figure 16 demonstrates again the problem of predicate ambiguity. The predicate depress can occur in two contexts that, though related, instigate different entailment relations. The first context is when the object of the predicate (the Y argument) is a person and then the meaning of depress is similar to discourage. A second context is when the object is some phenomenon, for example, The serious economical situation depresses consumption levels. In initialization, HTL-FRG generates a set of edges that is compatible with the latter context, determining that the predicate depress entails the predicates lower and bring down. TNF re-attaches depress as a synonym of discourage and deter (because this improves the objective function), resulting in a set of edges that corresponds to the first meaning of depress mentioned above. The methods we use in this article do not allow learning the rules for both contexts because this would result in violations of the transitivity and FRG assumptions. This again emphasizes the importance of learning graphs where predicates are marked by their various senses, which will result in a model that can directly benefit from the methods suggested in this article. 5 experimental evaluation: untyped entailment graphs :In this section we evaluate our methods on a graph with 20,000 nodes. Again, we describe how the nodes V and the weighting function w are constructed. The nodes of the entailment graph we learn in this section are not typed. Although ambiguity is a problem in this setting, we will show that nevertheless transitivity constraints can improve results compared with a state-of-the-art local entailment classifier.
The nodes of the graph were generated from a set of two billion tuples of the form (arg1,predicate,arg2) that were extracted by the Reverb open information extraction system (Fader, Soderland, and Etzioni 2011) over the Clueweb096 data set and were generously provided by the authors of Reverb.
We performed some simple preprocessing over the extracted tuples. Each tuple is accompanied by a confidence value and we discarded tuples with confidence smaller than 0.5. Next, we normalized predicates using a procedure that omits unnecessary content such as modals and adverbs. The entire normalization procedure is publicly available as part of the Reverb project.7 Last, we normalized arguments by replacing pronouns and determiners by the tokens PRONOUN and DET. We also ran the BIU
6 http://lemurproject.org/clueweb09.php/. 7 https://github.com/knowitall/reverb-core/blob/master/core/src/main/java/edu/washington/ cs/knowitall/normalization/RelationString.java.
number normalizer8 and replaced numbers larger than one by the token NUM. After these three steps, we are left with a total of 960 million tuples of which 291 million are distinct.
The number of distinct predicates in the set of extracted tuples is 103,315. Because this is still a large number, we restrict the predicate set to the 10,000 predicates that appear with the highest number of pairs of arguments. As previously explained, each predicate is split into two (for example, defeat is split into X defeat Y and Y defeat X), and so the final number of entailment graph nodes is 20,000. As with typed entailment graphs, the weighting function w is obtained by training a classifier that provides a score sij for all pairs of predicates9 and defining wij = sij − λ. This setting, computing the scores sij, involves the following steps:
1. Training set generation. We use the data set released by Zeichner, Berant, and Dagan (2012), which contains 6,567 entailment rule applications annotated for their validity by crowdsourcing. For example, the data set marks that The exercises alleviate pain⇒ The exercises help ease pain is a valid rule application, whereas Obama want to boost the defense budget⇒ Obama increase the defense budget is an invalid rule application. We extract a single rule from each rule application, for example, from the rule application The exercises alleviate pain⇒ The exercises help ease pain we extract the rule X alleviate Y⇒ X help ease Y. We use half of the data set for training, resulting in 1,224 positive examples and 2,060 negative examples. Another two training examples are X unable to pay Y⇒ X owe Y and X own Y ; Y be sold to X.
2. Feature representation. Each pair of predicates (p1, p2) is represented by a feature vector where the first six are distributional similarity features identical to the first six features described in Section 4.2. In addition, for pairs of predicates for which at least one distributional similarity feature is non-zero, we add lexicographic features computed from WordNet (Fellbaum 1998), VerbOcean (Chklovski and Pantel 2004), and CatVar (Habash and Dorr 2003), as well as string-similarity features. Table 2 provides the exact details of these features. A feature is computed for a pair of predicates (p1, p2) where in this context a predicate is a pair (pred,rev): pred is the lexical realization of the predicate, and rev is a Boolean indicating whether arg1 is X and arg2 is Y or vice versa. Overall, each pair of predicates is represented by 27 features.
3. Training. After obtaining a feature representation for every pair of predicates, we train a Gaussian kernel SVM classifier that optimizes F1 (SVMperf implementation [Joachims 2005]), and tune the parameters C and γ by a grid search combined with 5-fold cross validation. We use the trained local classifier to compute a score sij for all pairs of predicates.
8 http://u.cs.biu.ac.il/~nlp/downloads/normalizer.html. 9 We do not need to run the classifier on 10, 0002 pairs of predicates, because for the majority of predicate
pairs the features are all zeros, and for this set of pairs we can run the classifier only once.
(c) passive-active: The passive-active constraint holds for (p1, p2) and (p2, p1) if p2 is the passive form of p1.
Again, local constraints are integrated into our model by changing the score sij =∞ for positive local constraints and sij = −∞ for negative local constraints. To evaluate performance we use the second half of the data set released by Zeichner, Berant, and Dagan (2012) as a test set. This test set contains 3,283 annotated examples, where 1,734 are covered by the 10,000 nodes of our graph (649 positive examples and 1,085 negative examples). As usual, for each algorithm we compute recall and precision with respect to the gold standard at various points by varying the sparseness parameter λ.
Most methods used over typed graphs in Section 4 (Exact-graph, Exact-forest, LPrelax, and GNF) are intractable because they require an ILP solver and our graph does not decompose into components that are small enough. Therefore, we compare the Notrans local baseline, where transitivity is not used, with the efficient global methods TNF and TNCF. Recall that in Section 4 we initialized TNF by applying an ILP solver in a sparse configuration on each graph component. In this experiment the graph is too large to use an ILP solver and so we initialize TNF and TNCF with a simple heuristic we describe next.
We begin with an empty graph and first sort all pairs of predicates (i, j) for which wij > 0, according to their weight w. Then, we go over predicate pairs one-by-one and perform two operations. First, we verify that inserting (i, j) into the graph does not violate the FRG assumption. This is done by going over all edges (i, k) ∈ E and checking that for every k either ( j, k) ∈ E or (k, j) ∈ E. If this is the case, then the edge (i, j) is a valid candidate edge as the resulting reduced graph will remain a directed forest, otherwise adding it will result in i having two parents in the reduced graph, and we do not insert this edge. Second, we compute the transitive closure Tij, which contains all node pairs that must be edges in the graph in case we insert (i, j), due to transitivity. We compute the change in the value of the objective function if we were to insert Tij into the graph. If this change improves the objective function value, we insert Tij into the graph (and so the value of the objective function increases monotonically). We keep going over the candidate edges from highest score to lowest until the objective function can no longer be improved by inserting a candidate edge. We term this initialization HTL-FRG, because we scan edges from high-score to low-score and maintain an FRG. We add this initialization procedure as another baseline.
Figure 11 presents the precision-recall curve of all global algorithms compared with the local classifier for 0.05 ≤ λ ≤ 2.5. One evident property is that No-trans reaches higher recall values than the global algorithms, which means that adding a global transitivity constraint prevents adding correct edges that have positive weight, since this would cause the addition of many other edges that have negative weight. A possible reason for that is predicate ambiguity, where positive weight edges connect connectivity components through an ambiguous predicate. This suggests that a natural direction for future research is to develop algorithms that do not impose transitivity constraints for rules i⇒ j and j⇒ k, if each rule refers to a different meaning of j. One possible direction is to learn latent “sense” or “topic” variables for each rule (as suggested by Melamud et al. Melamud et al. [2013]). Then, we can use the methods presented in this article
to impose transitivity constraints for each sense separately, and easily allow violations of transitivity across different senses. We provide concrete examples for such cases of ambiguity in our qualitative analysis in Section 5.4.
Although global algorithms are limited in their recall, for recall values of 0.1– 0.2 they substantially improve precision over No-trans. To better observe this result, Table 5 presents the results for the recall range 0.1–0.25. Comparing TNCF and Notrans shows that the TNCF algorithm improves over No-trans by 5–10 precision points:
In the recall range of 0.1–.2, the precision of No-trans is 0.68–0.75, whereas the precision of TNCF is 0.74–0.8. The number of rules learned by TNCF in this recall range is about 100,000.
We use HTL-FRG as an initialization method for TNCF, which on its own is not a very good approximation algorithm. Comparing HTL-FRG with No-trans, we observe that it is unable to reach high recall values and it only marginally improves precision compared with No-trans. Applying TNF over HTL-FRG also provides disappointing results—it increases recall compared with HTL-FRG, but precision is not better than No-trans. Indeed, it seems that although performing node re-attachments monotonically improves the objective function value, it is not powerful enough to learn a graph that is better with respect to our gold standard. On the other hand, adding component re-attachments (TNCF) allows the algorithm to better explore the space of graphs and learn a graph with higher recall and precision than HTL-FRG.
Figure 12 presents the runtime for HTL-FRG, TNF, and TNCF (y-axis in logarithmic scale), all of which yield an FRG and can be run on a large untyped graph. Again, we were unable to obtain results with Exact-graph, Exact-forest, LP-relax, and GNF, because using an ILP solver on a graph with 20,000 nodes was impossible. As expected, HTL-FRG is much faster than both TNF and TNCF, and TNCF is somewhat slower than TNF. However, as mentioned earlier, TNCF is able to improve both precision and recall compared with TNF and HTL-FRG. Note that when λ = 1.2, runtime increases suddenly for both TNF and TNCF. This is because at this point a large connected component is formed, as we explain next.
Figure 13 presents the results of graph decomposition as described in Section 3.1. Evidently, when λ = 0 the largest component constitutes almost all of the graph, which contains 20,000 nodes. Note that when λ = 1.2, the size of the largest component increases suddenly to more than half of the graph nodes. Additionally, TNCF obtained recall of 0.1–0.2 when λ ≤ 0.2, and in this case the number of nodes in the largest
component is 90% of the total number of graph nodes. Thus, contrary to the typed graphs, untyped graphs that contain ambiguous predicates do not decompose well into small components.
To summarize, we demonstrated that our efficient global methods scale to a graph with 20,000 nodes and observed that even when predicates are ambiguous we are able to improve precision for moderate values of recall. Nevertheless, there are indications that our modeling assumptions are violated when applied over a large graph with ambiguous predicates. Transitivity and FRG constraints limit the recall we are able to reach, and the graph does not decompose very well into components. Thus, in future work we would like to apply our algorithms over graphs where predicate ambiguity is modeled and so transitivity constraints can be properly applied. In the previous section we quantitatively evaluated global methods for learning entailment graphs over a large set of untyped predicates. Our results revealed interesting issues that we would like to further investigate:
1. Why do methods that assume entailment graphs are transitive and forest-reducible have limited recall?
2. How much does node re-attachment affect graph structure?
In this section, we try to answer these questions by performing a qualitative analysis that allows us to gain better insight on the behavior of our algorithms.
Question 1: Transitivity and FRG assumptions. We would like to understand why using global algorithms such as TNCF results in limited recall. Our hypothesis is that because predicates in the graph are untyped and potentially ambiguous, violations of the
assumptions are possible, which results in recall reduction. To examine this hypothesis, we perform two qualitative analyses:
1. We analyze graphs containing only gold-standard edges, and manually identify cases where the modeling assumptions are violated.
2. We compare the local algorithm, No-trans, to the global algorithm, TNCF. If our hypothesis is correct, then we expect TNCF to remove correct edges that are inserted by No-trans due to ambiguous predicates.
We start by constructing a graph with all edges from our gold-standard data set and search for transitivity and FRG violations caused by ambiguous predicates. Note that the manually annotated gold-standard edges are only a subset of the full set of correct edges, which is actually much larger. Thus, we cannot automatically identify violations, because the graph is missing many true edges. Instead, we manually go over candidate violations and check if indeed this case is a result of a violation of our modeling assumptions, and if so then our global algorithm cannot predict the correct set of edges.
One example is the pair of rules seek ⇒ apply for and apply for ⇒ file for. The first rule was annotated in the rule application Others seek medical care ⇒ Others apply for medical care (Recall that crowdsourcing was used to annotate rule applications, and rules were extracted from these applications.) The second rule was extracted from the rule application Students apply for more than one award⇒ Student file for more than one award. This causes a transitivity violation because Students seek more than one award ; Student file for more than one award. Evidently, the meaning of the predicate apply for is contextdependent. Another example is the pair of rules contain ⇒ supply and supply ⇒ serve annotated in the applications The page contains links⇒ The page supplies links and Users supply information or material⇒Users serve information or material. Clearly, contain ; serve and the transitivity violation is caused by the fact that the predicate supply is contextdependent—in the first context the subject is inanimate, whereas in the second context the subject is human.
A good example for an FRG violation is the predicate come from. The following three entailment rules are annotated by the turkers: come from ⇒ be raised in, come from⇒ be derived from, and come from⇒ come out of. These correspond to the following rule applications: The Messiah comes from Judah ⇒ The Messiah was raised in Judah, The colors come from the sun light ⇒ The colors are derived from the sun light, and The truth comes from the book ⇒ The truth comes out of the book. Clearly, be raised in ; be derived from and be derived from ; be raised in, so this is an FRG violation. Indeed, the three applications correspond to different meanings of the predicate come from, which depend on context. A second example is the pair of rules be dedicated to⇒ be committed to and be dedicated to⇒ contain information on. Again, this is an FRG violation because be committed to ; contain information on and contain information on ; be committed to. This violation occurs because of the ambiguity of the predicate be dedicated to, which can be resolved by knowing whether the subject is human or is an object that carries information (for example, The web site is dedicated to the life of lizards).
This first analysis revealed particular cases where the transitivity and FRG assumptions do not hold. In the next analysis, we would like to directly compare cases where global and local methods disagree with one another. We hypothesize that because predicates are not disambiguated, then global algorithms will delete correct edges due
to an ambiguous predicate. We set the sparseness parameter λ = 0.2 and compare the set of edges learned by No-trans to the set of edges learned by TNCF.
Examining the disagreements, we observe that, as expected, in 90% of the cases TNCF deletes an edge that was inserted by No-trans. We divide these deleted edges into two categories in the following manner: We compute the weakly connected components of the graph learned by TNCF,10 and then for each edge inserted by No-trans and deleted by TNCF we check whether it connects two different weakly connected components or not. There are two reasons for focusing on this property of deleted edges. First, we hypothesize that most deleted edges will connect two different connected components, because such edges result in many violations of transitivity. Second, weakly connected components usually correspond to separate semantic meanings, and thus a correct edge that connects two weakly connected components probably involves an ambiguous predicate.
Indeed, out of all edges where No-trans inserts an edge and TNCF does not, 76.4% connect weakly connected components in the global graph. This proves that deleting such edges is the main cause for disagreement between TNCF and No-trans. Now, we can take a closer look at these edges, and check whether indeed when a correct edge is deleted, it is because of an ambiguous predicate.
We randomly sample five cases where TNCF erroneously deleted an edge connecting weakly connected components—these are cases where ambiguity is likely to occur. For comparison, we also sample five cases where TNCF was right and No-trans erred. Table 6 shows the samples where No-trans is correct. The first column describes the rule and the second column specifies the score sij provided by the local classifier No-trans. The last two columns detail two examples for predicates that are in the same weakly connected component with the rule LHS and RHS in the graph learned by TNCF. These two columns provide a flavor for the overall meaning of predicates that belong to this component. Table 7 is equivalent and shows the samples where TNCF is correct.
Example 1 in Table 6 already demonstrates the problem of ambiguity. The LHS for this rule application in the crowdsourcing experiment was your parents turn off comments and indeed in this case the RHS your parents cut off comments is inferred. However, we can see that the LHS component revolves around the turning off or shutting off of appliances, for example, whereas the RHS component deals with a more explicit and physical meaning of cutting. This is because the predicate X cut off Y is ambiguous and consequently TNCF omits this correct edge. Example 5 also demonstrates well the problem of ambiguity—the LHS of this rule application is This site was put by fans, which in this context entails the RHS This site was run by fans. However, the more common sense of be put by does not entail the predicate be run by; this sense is captured by the predicates in the LHS component Y help put X and X be placed by Y. Example 4 is another good example where the ambiguous predicate is X be open to Y—the RHS component is concerned with the physical state of being opened, whereas the LHS component has a more abstract sense of availability. The LHS of the rule application in this case was The services are available to museums. Examples 2 and 3 are cases where the problem is not in predicate ambiguity but in the connected component itself. Thus, out of five randomly sampled examples where TNCF deleted an edge that connects two weakly connected components, three can be attributed to predicate ambiguity.
10 A weekly connected component in a directed graph is a set of nodes S such that for every pair of nodes u, v ∈ S there is an undirected path from u to v and from v to u, that is, there is a path when edge directions are ignored. 6 conclusions :The problem of language variability is at the heart of many semantic applications such as Information Extraction, Question Answering, Semantic Parsing, and more. Consequently, learning broad coverage knowledge bases of entailment rules and paraphrases (Ganitkevitch, Van Durme, and Callison-Burch 2013) has proved crucial for systems that perform inference over textual representations (Stern and Dagan 2012; Angeli and Manning 2014).
In this article, we have presented algorithms for learning entailment rules between predicates that can scale to large sets of predicates. Our work builds on prior work by Berant, Dagan, and Goldberger (2012), who defined the concept of entailment graphs, and formulated entailment rule learning as a graph optimization problem, where the graph has to obey the structural constraint of transitivity.
Our main contribution is a heuristic polynomial approximation algorithm that can learn entailment graphs containing tens of thousands of nodes. The algorithm is based on two main observations: first, that entailment graphs are sparse, and second, that entailment graphs are forest-reducible. This leads to an efficient algorithm in which at the first step the graph is decomposed into smaller components, and in the second step we find a set of edges for each component. The second step is an iterative approximation algorithm in which at each iteration a node or component of the graph is deleted and then re-attached in a way that improves the overall objective function.
We performed a thorough empirical evaluation, in which we ran our algorithm on two separate data sets. On the first data set we witnessed a substantial improvement in runtime at a relatively small cost in performance. On the second data set, we show that our method scales to an untyped entailment graph containing 20,000 nodes, much larger than was previously possible using state-of-the-art global methods. We see that using a global algorithm improves precision for modest values of recall, but that local methods can achieve higher recall than global methods. We perform an analysis and conclude that the main reason for this limitation is predicate ambiguity, which results in graphs that do not adhere to the structural constraints of transitivity and forest reducibility.
In future work, we hope to address the problem of predicate ambiguity. A natural direction is to combine methods that model the context-sensitivity of predicates with global structural assumptions. In this way, one can apply structural constraints for a set of predicates, given that they appear in similar contexts. A second important direction for future work is to demonstrate that using entailment rules learned by our method improves performance in downstream semantic applications such as QA, IE, and Recognizing Textual Entailment."
"1 introduction :The distributional approach to the semantic modeling of natural language, inspired by the notion—presented by Firth (1957) and Harris (1968)—that the meaning of a word is tightly related to its context of use, has grown in popularity as a method of semantic representation. It draws from the frequent use of vector-based document models in information retrieval, modeling the meaning of words as vectors based on the distribution of co-occurring terms within the context of a word.
Using various vector similarity metrics as a measure of semantic similarity, these distributional semantic models (DSMs) are used for a variety of NLP tasks, from
∗ DeepMind Technologies Ltd, 5 New Street Square, London EC4A 3 TW. E-mail: etg@google.com.
∗∗ School of Electronic Engineering and Computer Science, Queen Mary University of London, Mile End Road, London E1 4NS, United Kingdom. E-mail: mehrnoosh.sadrzadeh@qmul.ac.uk.
† The work described in this article was performed while the authors were at the University of Oxford. 1 Support from EPSRC grant EP/J002607/1 is acknowledged.
Submission received: 26 September 2012; revised submission received: 31 October 2013; accepted for publication: 5 April 2014.
doi:10.1162/COLI a 00209
© 2015 Association for Computational Linguistics
automated thesaurus building (Grefenstette 1994; Curran 2004) to automated essay marking (Landauer and Dumais 1997). The broader connection to information retrieval and its applications is also discussed by Manning, Raghavan, and Schütze (2011). The success of DSMs in essentially word-based tasks such as thesaurus extraction and construction (Grefenstette 1994; Curran 2004) invites an investigation into how DSMs can be applied to NLP and information retrieval (IR) tasks revolving around larger units of text, using semantic representations for phrases, sentences, or documents, constructed from lemma vectors. However, the problem of compositionality in DSMs—of how to go from word to sentence and beyond—has proved to be non-trivial.
A new framework, which we refer to as DisCoCat, initially presented in Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) reconciles distributional approaches to natural language semantics with the structured, logical nature of formal semantic models. This framework is abstract; its theoretical predictions have not been evaluated on real data, and its applications to empirical natural language processing tasks have not been studied.
This article is the journal version of Grefenstette and Sadrzadeh (2011a, 2011b), which fill this gap in the DisCoCat literature; in it, we develop a concrete model and an unsupervised learning algorithm to instantiate the abstract vectors, linear maps, and vector spaces of the theoretical framework; we develop a series of empirical natural language processing experiments and data sets and implement our algorithm on large scale real data; we analyze the outputs of the algorithm in terms of linear algebraic equations; and we evaluate the model on these experiments and compare the results with other competing unsupervised models. Furthermore, we provide a linear algebraic analysis of the algorithm of Grefenstette and Sadrzadeh (2011a) and present an in-depth study of the better performance of the method of Grefenstette and Sadrzadeh (2011b).
We begin in Section 2 by presenting the background to the task of developing compositional distributional models. We briefly introduce two approaches to semantic modeling: formal semantic models and distributional semantic models. We discuss their differences, and relative advantages and disadvantages. We then present and critique various approaches to bridging the gap between these models, and their limitations. In Section 3, we summarize the categorical compositional distributional framework of Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) and provide the theoretical background necessary to understand it; we also sketch the road map of the literature leading to the development of this setting and outline the contributions of this paper to the field. In Section 4, we present the details of an implementation of this framework, and introduce learning algorithms used to build semantic representations in this implementation. In Section 5, we present a series of experiments designed to evaluate this implementation against other unsupervised distributional compositional models. Finally, in Section 6 we discuss these results, and posit future directions for this research area.","2 background :Compositional formal semantic models represent words as parts of logical expressions, and compose them according to grammatical structure. They stem from classical ideas in logic and philosophy of language, mainly Frege’s principle that the meaning of a sentence is a function of the meaning of its parts (Frege 1892). These models relate to well-known and robust logical formalisms, hence offering a scalable theory of meaning that can be used to reason about language using logical tools of proof and inference. Distributional models are a more recent approach to semantic modeling, representing
the meaning of words as vectors learned empirically from corpora. They have found their way into real-world applications such as thesaurus extraction (Grefenstette 1994; Curran 2004) or automated essay marking (Landauer and Dumais 1997), and have connections to semantically motivated information retrieval (Manning, Raghavan, and Schütze 2011). This two-sortedness of defining properties of meaning: “logical form” versus “contextual use,” has left the quest for “what is the foundational structure of meaning?”—a question initially the concern of solely linguists and philosophers of language—even more of a challenge.
In this section, we present a short overview of the background to the work developed in this article by briefly describing formal and distributional approaches to natural language semantics, and providing a non-exhaustive list of some approaches to compositional distributional semantics. For a more complete review of the topic, we encourage the reader to consult Turney (2012) or Clark (2013). Formal semantic models provide methods for translating sentences of natural language into logical formulae, which can then be fed to computer-aided automation tools to reason about them (Alshawi 1992).
To compute the meaning of a sentence consisting of n words, meanings of these words must interact with one another. In formal semantics, this further interaction is represented as a function derived from the grammatical structure of the sentence. Such models consist of a pairing of syntactic analysis rules (in the form of a grammar) with semantic interpretation rules, as exemplified by the simple model presented on the left of Figure 1.
The semantic representations of words are lambda expressions over parts of logical formulae, which can be combined with one another to form well-formed logical expressions. The function | − | : L → M maps elements of the lexicon L to their interpretation (e.g., predicates, relations, domain objects) in the logical model M used. Nouns are typically just logical atoms, whereas adjectives and verbs and other relational words are interpreted as predicates and relations. The parse of a sentence such as “cats like milk,” represented here as a binarized parse tree, is used to produce its semantic interpretation by substituting semantic representations for their grammatical constituents and applying β-reduction where needed. Such a derivation is shown on the right of Figure 1.
What makes this class of models attractive is that it reduces language meaning to logical expressions, a subject well studied by philosophers of language, logicians, and linguists. Its properties are well known, and it becomes simple to evaluate the meaning of a sentence if given a logical model and domain, as well as verify whether or not one sentence entails another according to the rules of logical consequence and deduction.
However, such logical analysis says nothing about the closeness in meaning or topic of expressions beyond their truth-conditions and which models satisfy these truth conditions. Hence, formal semantic approaches to modeling language meaning do not perform well on language tasks where the notion of similarity is not strictly based on truth conditions, such as document retrieval, topic classification, and so forth. Furthermore, an underlying domain of objects and a valuation function must be provided, as with any logic, leaving open the question of how we might learn the meaning of language using such a model, rather than just use it. A popular way of representing the meaning of words in lexical semantics is as distributions in a high-dimensional vector space. This approach is based on the distributional hypothesis of Harris (1968), who postulated that the meaning of a word was dictated by the context of its use. The more famous dictum stating this hypothesis is the statement of Firth (1957) that “You shall know a word by the company it keeps.” This view of semantics has furthermore been associated (Grefenstette 2009; Turney and Pantel 2010) with earlier work in philosophy of language by Wittgenstein (presented in Wittgenstein 1953), who stated that language meaning was equivalent to its real world use.
Practically speaking, the meaning of a word can be learned from a corpus by looking at what other words occur with it within a certain context, and the resulting distribution can be represented as a vector in a semantic vector space. This vectorial representation is convenient because vectors are a familiar structure with a rich set of ways of computing vector distance, allowing us to experiment with different word similarity metrics. The geometric nature of this representation entails that we can not only compare individual words’ meanings with various levels of granularity (e.g., we might, for example, be able to show that cats are closer to kittens than to dogs, but that all three are mutually closer than cats and steam engines), but also apply methods frequently called upon in IR tasks, such as those described by Manning, Raghavan, and Schütze (2011), to group concepts by topic, sentiment, and so on.
The distribution underlying word meaning is a vector in a vector space, the basis vectors of which are dictated by the context. In simple models, the basis vectors will be annotated with words from the lexicon. Traditionally, the vector spaces used in such models are Hilbert spaces (i.e., vector spaces with orthogonal bases, such that the inner product of any one basis vector with another [other than itself] is zero). The semantic vector for any word can be represented as the weighted sum of the basis vectors:
−−−−−−−→ some word = ∑ i ci −→ni
where {−→ni }i is the basis of the vector space the meaning of the word lives in, and ci ∈ R is the weight associated with basis vector −→ni .
The construction of the vector for a word is done by counting, for each lexicon word ni associated with basis vector
−→ni , how many times it occurs in the context of each occurrence of the word for which we are constructing the vector. This count is then typically adjusted according to a weighting scheme (e.g., TF-IDF). The “context” of a word can be something as simple as the other words occurring in the same sentence as
the word or within k words of it, or something more complex, such as using dependency relations (Padó and Lapata 2007) or other syntactic features.
Commonly, the similarity of two semantic vectors is computed by taking their cosine measure, which is the sum of the product of the basis weights of the vectors:
cosine(−→a ,−→b ) = ∑ i c a i c b i√∑
i (c a i )
2 ∑
i (c b i ) 2
where cai and c b i are the basis weights for −→a and −→b , respectively. However, other options may be a better fit for certain implementations, typically dependent on the weighting scheme.
Readers interested in learning more about these aspects of distributional lexical semantics are invited to consult Curran (2004), which contains an extensive overview of implementation options for distributional models of word meaning. In the previous overview of distributional semantic models of lexical semantics, we have seen that DSMs are a rich and tractable way of learning word meaning from a corpus, and obtaining a measure of semantic similarity of words or groups of words. However, it should be fairly obvious that the same method cannot be applied to sentences, whereby the meaning of a sentence would be given by the distribution of other sentences with which it occurs.
First and foremost, a sentence typically occurs only once in a corpus, and hence substantial and informative distributions cannot be created in this manner. More importantly, human ability to understand new sentences is a compositional mechanism: We understand sentences we have never seen before because we can generate sentence meaning from the words used, and how they are put into relation. To go from word vectors to sentence vectors, we must provide a composition operation allowing us to construct a sentence vector from a collection of word vectors. In this section, we will discuss several approaches to solving this problem, their advantages, and their limitations.
2.3.1 Additive Models. The simplest composition operation that comes to mind is straightforward vector addition, such that:
−→ ab = −→a +−→b
Conceptually speaking, if we view word vectors as semantic information distributed across a set of properties associated with basis vectors, using vector addition as a semantic composition operation states that the information of a set of lemmas in a sentence is simply the sum of the information of the individual lemmas. Although crude, this approach is computationally cheap, and appears sufficient for certain NLP tasks: Landauer and Dumais (1997) show it to be sufficient for automated essay marking tasks, and Grefenstette (2009) shows it to perform better than a collection of other simple similarity metrics for summarization, sentence paraphrase, and document paraphrase detection tasks.
However, there are two principal objections to additive models of composition: first, vector addition is commutative, therefore, −−−−−−−−−−−−→ John drank wine = −−→ John + −−−→ drank + −−−→ wine =
−−−−−−−−−−−−→ Wine drank John, and thus vector addition ignores syntactic structure completely; and second, vector addition sums the information contained in the vectors, effectively jumbling the meaning of words together as sentence length grows.
The first objection is problematic, as the syntactic insensitivity of additive models leads them to equate the representation of sentences with patently different meanings. Mitchell and Lapata (2008) propose to add some degree of syntactic sensitivity—namely, accounting for word order—by weighting word vectors according to their order of appearance in a sentence as follows:
−→ ab = α−→a + β−→b
where α,β∈R. Consequently, −−−−−−−−−−−−→John drank wine=α · −−→John + β · −−−→drank + γ·−−−→wine would not have the same representation as −−−−−−−−−−−−→ Wine drank John = α · −−−→wine + β · −−−→drank + γ · −−→John.
The question of how to obtain weights and whether they are only used to reflect word order or can be extended to cover more subtle syntactic information is open, but it is not immediately clear how such weights may be obtained empirically and whether this mode of composition scales well with sentence length and increase in syntactic complexity. Guevara (2010) suggests using machine-learning methods such as partial least squares regression to determine the weights empirically, but states that this approach enjoys little success beyond minor composition such as adjective-noun or noun-verb composition, and that there is a dearth of metrics by which to evaluate such machine learning–based systems, stunting their growth and development.
The second objection states that vector addition leads to increase in ambiguity as we construct sentences, rather than decrease in ambiguity as we would expect from giving words a context; and for this reason Mitchell and Lapata (2008) suggest replacing additive models with multiplicative models as discussed in Section 2.3.2, or combining them with multiplicative models to form mixture models as discussed in Section 2.3.3.
2.3.2 Multiplicative Models. The multiplicative model of Mitchell and Lapata (2008) is an attempt to solve the ambiguity problem discussed in Section 2.3.1 and provide implicit disambiguation during composition. The composition operation proposed is the component-wise multiplication ( ) of two vectors: Vectors are expressed as the weighted sum of their basis vectors, and the weight of the basis vectors of the composed vector is the product of the weights of the original vectors; for −→a =∑i ci−→ni , and−→ b = ∑ i c ′ i −→ni , we have
−→ ab = −→a −→b = ∑ i cic′i −→ni
Such multiplicative models are shown by Mitchell and Lapata (2008) to perform better at verb disambiguation tasks than additive models for noun-verb composition, against a baseline set by the original verb vectors. The experiment they use to show this will also serve to evaluate our own models, and form the basis for further experiments, as discussed in Section 5.
This approach to compositionality still suffers from two conceptual problems: First, component-wise multiplication is still commutative and hence word order is not accounted for; second, rather than “diluting” information during large compositions
and creating ambiguity, it may remove too much through the “filtering” effect of component-wise multiplication.
The first problem is more difficult to deal with for multiplicative models than for additive models, because both scalar multiplication and component-wise multiplication are commutative and hence α−→a β−→b = α−→b β−→a and thus word order cannot be taken into account using scalar weights.
To illustrate how the second problem entails that multiplicative models do not scale well with sentence length, we look at the structure of component-wise multiplication again: −→a −→b =∑i cic′i −→ni . For any i, if ci = 0 or c′i = 0, then cic′i = 0, and therefore for any composition, the number of non-zero basis weights of the produced vector is less than or equal to the number of non-zero basis weights of the original vectors: At each composition step information is filtered out (or preserved, but never increased). Hence, as the number of vectors to be composed grows, the number of non-zero basis weights of the product vector stays the same or—more realistically—decreases. Therefore, for any composition of the form −−−−−−−−→a1 . . . ai . . . an = −→a1 . . . −→ai . . . −→an , if for any i, −→ai is orthogonal to −−−−−−→a1 . . . ai−1 then −−−−−−−−→a1 . . . ai . . . an = −→0 . It follows that purely multiplicative models alone are not apt as a single mode of composition beyond nounverb (or adjective-noun) composition operations.
One solution to this second problem not discussed by Mitchell and Lapata (2008) would be to introduce some smoothing factor s ∈ R+ for point-wise multiplication such that −→a −→b =∑i (ci + s)(c′i + s)−→ni , ensuring that information is never completely filtered out. Seeing how the problem of syntactic insensitivity still stands in the way of full-blown compositionality for multiplicative models, we leave it to those interested in salvaging purely multiplicative models to determine whether some suitable value of s can be determined.
2.3.3 Mixture Models. The problems faced by multiplicative models presented in Section 2.3.2 are acknowledged in passing by Mitchell and Lapata (2008), who propose mixing additive and multiplicative models in the hope of leveraging the advantage of each while doing away with their pitfalls. This is simply expressed as the weighted sum of additive and multiplicative models:
−→ ab = α−→a + β−→b + γ(−→a −→b )
where α, β, and γ are predetermined scalar weights. The problems for these models are threefold. First, the question of how scalar weights are to be obtained still needs to be determined. Mitchell and Lapata (2008) concede that one advantage of purely multiplicative models over weighted additive or mixture models is that the lack of scalar weights removes the need to optimize the scalar weights for particular tasks (at the cost of not accounting for syntactic structure), and avoids the methodological concerns accompanying this requirement.
Second, the question of how well this process scales from noun-verb composition to more syntactically rich expressions must be addressed. Using scalar weights to account for word order seems ad hoc and superficial, as there is more to syntactic structure than the mere ordering of words. Therefore, an account of how to build sentence vectors for sentences such as The dog bit the man and The man was bitten by the dog in order to give both sentences the same (or a similar) representation would need to give a richer role to scalar weights than mere token order. Perhaps specific
weights could be given to particular syntactic classes (such as nouns) to introduce a more complex syntactic element into vector composition, but it is clear that this alone is not a solution, as the weight for nouns dog and man would be the same, allowing for the same commutative degeneracy observed in non-weighted additive models, in which −−−−−−−−−−−−−−→ the dog bit the man = −−−−−−−−−−−−−−→ the man bit the dog. Introducing a mixture of weighting systems accounting for both word order and syntactic roles may be a solution, but it is not only ad hoc but also arguably only partially reflects the syntactic structure of the sentence.
The third problem is that Mitchell and Lapata (2008) show that in practice, although mixture models perform better at verb disambiguation tasks than additive models and weighted additive models, they perform equivalently to purely multiplicative models with the added burden of requiring parametric optimization of the scalar weights.
Therefore, whereas mixture models aim to take the best of additive and multiplicative models while avoiding their problems, they are only partly successful in achieving the latter goal, and demonstrably do little better in achieving the former.
2.3.4 Tensor-Based Models. From Sections 2.3.1–2.3.3 we observe that the need for incorporating syntactic information into DSMs to achieve true compositionality is pressing, if only to develop a non-commutative composition operation that can take into account word order without the need for adhoc weighting schemes, and hopefully richer syntactic information as well.
An early proposal by Smolensky and colleagues (Smolensky 1990; Smolensky and Legendre 2006) to use linear algebraic tensors as a composition operation solves the problem of finding non-commutative vector composition operators. The composition of two vectors is their tensor product, sometimes called the kronecker product when ap-
plied to vectors rather than vector spaces. For −→a ∈ V =∑i ci−→ni , and −→b ∈ W =∑j c′j−→n′j , we have:
−→ ab = −→a ⊗−→b = ∑ ij cic′j −→ni ⊗ −→ n′j
The composition operation takes the original vectors and maps them to a vector in a larger vector space V ⊗ W, which is the tensor space of the original vectors’ spaces. Here the second instance of ⊗ is not a recursive application of the kronecker product, but rather the pairing of basis elements of V and W to form a basis element of V ⊗ W. The shared notation and occasional conflation of kronecker and tensor products may seem confusing, but is fairly standard in multilinear algebra.
The advantage of this approach is twofold: First, vectors for different words need not live in the same spaces but can be composed nonetheless. This allows us to represent vectors for different word classes (topics, syntactic roles, etc.) in different spaces with different bases, which was not possible under additive or multiplicative models. Second, because the product vector lives in a larger space, we obtain the intuitive notion that the information of the whole is richer and more complex than the information of the parts.
Dimensionality Problems. However, as observed and commented upon by Smolensky himself, this increase in dimensionality brings two rather large problems for tensor based models. The first is computational: The size of the product vector space is the product of the size of the original vector spaces. If we assume that all words live in the
same space N of dimensionality dim(N), then the dimensionality of an n-word sentence vector is dim(N)n. If we have as many basis vectors for our word semantic space as there are lexemes in our vocabulary (e.g., approximately 170k in English2), then the size of our sentence vectors quickly reaches magnitudes for which vector comparison (or even storage) are computationally intractable.3 Even if, as most DSM implementations do, we restrict the basis vectors of word semantic spaces to the k (e.g., k = 2,000) most frequent words in a corpus, the sentence vector size still grows exponentially with sentence length, and the implementation problems remain.
The second problem is mathematical: Sentences of different length live in different spaces, and if we assign different vector spaces to different word types (e.g., syntactic classes), then sentences of different syntactic structure live in different vector spaces, and hence cannot be compared directly using inner product or cosine measures, leaving us with no obvious mode of semantic comparison for sentence vectors. If any model wishes to use tensor products in composition operations, it must find some way of reducing the dimensionality of product vectors to some common vector space so that they may be directly compared.
One notable method by which these dimensionality problems can be solved in general are the holographic reduced representations proposed by Plate (1991). The product vector of two vectors is projected into a space of smaller dimensionality by circular convolution to produce a trace vector. The circular correlation of the trace vector and one of the original vectors produces a noisy version of the other original vector. The noisy vector can be used to recover the clean original vector by comparing it with a predefined set of candidates (e.g., the set of word vectors if our original vectors are word meanings). Traces can be summed to form new traces effectively containing several vector pairs from which original vectors can be recovered. Using this encoding/decoding mechanism, the tensor product of sets of vectors can be encoded in a space of smaller dimensionality, and then recovered for computation without ever having to fully represent or store the full tensor product, as discussed by Widdows (2008).
There are problems with this approach that make it unsuitable for our purposes, some of which are discussed in Mitchell and Lapata (2010). First, there is a limit to the information that can be stored in traces, which is independent of the size of the vectors stored, but is a logarithmic function of their number. As we wish to be able to store information for sentences of variable word length without having to directly represent the tensored sentence vector, setting an upper bound to the number of vectors that can be composed in this manner limits the length of the sentences we can represent compositionally using this method.
Second, and perhaps more importantly, there are restrictions on the nature of the vectors that can be encoded in such a way: The vectors must be independently distributed such that the mean Euclidean length of each vector is 1. Such conditions are unlikely to be met in word semantic vectors obtained from a corpus; and as the failure to do so affects the system’s ability to recover clean vectors, holographic reduced representations are not prima facie usable for compositional DSMs, although it is important to note that Widdows (2008) considers possible application areas where they may be of use, although once again these mostly involve noun-verb and adjectivenoun compositionality rather than full blown sentence vector construction. We retain
2 Source: http://www.oxforddictionaries.com/page/howmanywords. 3 At four bytes per integer, and one integer per basis vector weight, the vector for John loves Mary
would require roughly (170, 000 · 4)3 ≈ 280 petabytes of storage, which is over ten times the data Google processes on a daily basis according to Dean and Ghemawat (2008).
from Plate (1991) the importance of finding methods by which to project the tensored sentence vectors into a common space for direct comparison, as will be discussed further in Section 3.
Syntactic Expressivity. An additional problem of a more conceptual nature is that using the tensor product as a composition operation simply preserves word order. As we discussed in Section 2.3.3, this is not enough on its own to model sentence meaning. We need to have some means by which to incorporate syntactic analysis into composition operations.
Early work on including syntactic sensitivity into DSMs by Grefenstette (1992) suggests using crude syntactic relations to determine the frame in which the distributions for word vectors are collected from the corpus, thereby embedding syntactic information into the word vectors. This idea was already present in the work of Smolensky, who used sums of pairs of vector representations and their roles, obtained by taking their tensor products, to obtain a vector representation for a compound. The application of these ideas to DSMs was studied by Clark and Pulman (2007), who suggest instantiating the roles to dependency relations and using the distributional representations of words as the vectors. For example, in the sentence Simon loves red wine, Simon is the subject of loves, wine is its object, and red is an adjective describing wine. Hence, from the dependency tree with loves as root node, its subject and object as children, and their adjectival descriptors (if any) as their children, we read the following structure: −−−→ loves ⊗−−→subj ⊗−−−−→Simon ⊗−→obj ⊗−−−→wine ⊗−→adj ⊗−→red. Using the equality relation for inner products of tensor products:
〈−→a ⊗−→b | −→c ⊗−→d 〉 = 〈−→a | −→c 〉 × 〈−→b | −→d 〉
We can therefore express inner-products of sentence vectors efficiently without ever having to actually represent the tensored sentence vector:
〈−−−−−−−−−−−−→Simon loves red wine | −−−−−−−−−−−−−−→Mary likes delicious rosé〉 = 〈−−→loves ⊗−→subj ⊗−−→Simon ⊗−→obj ⊗−−→wine ⊗−→adj ⊗−→red | −→likes ⊗−→subj ⊗−−→Mary ⊗−→obj ⊗−→rosé ⊗−→adj ⊗−−−−→delicious〉 = 〈−−→loves | −→likes〉× 〈−→subj | −→subj〉 × 〈−−→Simon | −−→Mary〉 × 〈−→obj | −→obj〉 × 〈−−→wine | −→rosé〉× 〈−→adj | −→adj 〉 × 〈−→red | −−−−→delicious〉 = 〈−−→loves | −→likes〉× 〈−−→Simon | −−→Mary〉× 〈−−→wine | −→rosé〉 × 〈−→red | −−−−→delicious〉
This example shows that this formalism allows for sentence comparison of sentences with identical dependency trees to be broken down to term-to-term comparison without the need for the tensor products to ever be computed or stored, reducing computation to inner product calculations.
However, although matching terms with identical syntactic roles in the sentence works well in the given example, this model suffers from the same problems as the original tensor-based compositionality of Smolensky (1990) in that, by the authors’ own admission, sentences of different syntactic structure live in spaces of different dimensionality and thus cannot be directly compared. Hence we cannot use this to measure the similarity between even small variations in sentence structure, such as the pair “Simon likes red wine” and “Simon likes wine.”
2.3.5 SVS Models. The idea of including syntactic relations to other lemmas in word representations discussed in Section 2.3.4 is applied differently in the structured vector
space model presented by Erk and Padó (2008). They propose to represent word meanings not as simple vectors, but as triplets:
w = (v, R, R−1)
where v is the word vector, constructed as in any other DSM, R and R−1 are selectional preferences, and take the form of R → D maps where R is the set of dependency relations, and D is the set of word vectors. Selectional preferences are used to encode the lemmas that w is typically the parent of in the dependency trees of the corpus in the case of R, and typically the child of in the case of R−1.
Composition takes the form of vector updates according to the following protocol. Let a = (va, Ra, R−1a ) and b = (vb, Rb, R −1 b ) be two words being composed, and let r be the dependency relation linking a to b. The vector update procedure is as follows:
a′ = (va R−1b (r), Ra − {r}, R−1a ) b′ = (vb Ra(r), Rb, R−1b − {r})
where a′, b′ are the updated word meanings, and is whichever vector composition (addition, component-wise multiplication) we wish to use. The word vectors in the triplets are effectively filtered by combination with the lemma which the word they are being composed with expects to bear relation r to, and this relation between the composed words a and b is considered to be used and hence removed from the domain of the selectional preference functions used in composition.
This mechanism is therefore a more sophisticated version of the compositional disambiguation mechanism discussed by Mitchell and Lapata (2008) in that the combination of words filters the meaning of the original vectors that may be ambiguous (e.g., if we have one vector for bank); but contrary to Mitchell and Lapata (2008), the information of the original vectors is modified but essentially preserved, allowing for further combination with other terms, rather than directly producing a joint vector for the composed words. The added fact that R and R−1 are partial functions associated with specific lemmas forces grammaticality during composition, since if a holds a dependency relation r to b which it never expects to hold (for example a verb having as its subject another verb, rather than the reverse), then Ra and R −1 b are undefined for r and the update fails. However, there are some problems with this approach if our goal is true compositionality.
First, this model does not allow some of the “repeated compositionality” we need because of the update of R and R−1. For example, we expect that an adjective composed with a noun produces something like a noun in order to be further composed with a verb or even another adjective. However, here, because the relation adj would be removed from R−1b for some noun b composed with an adjective a, this new representation b
′ would not have the properties of a noun in that it would no longer expect composition with an adjective, rendering representations of simple expressions like “the new red car” impossible. Of course, we could remove the update of the selectional preference functions from the compositional mechanism, but then we would lose this attractive feature of grammaticality enforcement through the partial functionality of R and R−1.
Second, this model does little more than represent the implicit disambiguation that is expected during composition, rather than actually provide a full blown compositional model. The inability of this system to provide a novel mechanism by which to obtain
a joint vector for two composed lemmas—thereby building towards sentence vectors— entails that this system provides no means by which to obtain semantic representations of larger syntactic structures that can be compared by inner product or cosine measure as is done with any other DSM. Of course, this model could be combined with the compositional models presented in Sections 2.3.1–2.3.3 to produce sentence vectors, but whereas some syntactic sensitivity would have been obtained, the word ordering and other problems of the aforementioned models would still remain, and little progress would have been made towards true compositionality.
We retain from this attempt to introduce compositionality in DSMs that including information obtained from syntactic dependency relations is important for proper disambiguation, and that having some mechanism by which the grammaticality of the expression being composed is a precondition for its composition is a desirable feature for any compositional mechanism. The final class of approaches to vector composition we wish to discuss are three matrix-based models.
Generic Additive Model. The first is the Generic Additive Model of Zanzotto et al. (2010). This is a generalization of the weighted additive model presented in Section 2.3.1. In this model, rather than multiplying lexical vectors by fixed parameters α and β before adding them to form the representation of their combination, they are instead the arguments of matrix multiplication by square matrices A and B:
−→ ab = A−→a + B−→b
Here, A and B represent the added information provided by putting two words into relation.
The numerical content of A and B is learned by linear regression over triplets (−→a ,−→b ,−→c ) where −→a and −→b are lexical semantic vectors, and −→c is the expected output of the combination of −→a and −→b . This learning system thereby requires the provision of labeled data for linear regression to be performed. Zanzotto et al. (2010) suggest several sources for this labeled data, such as dictionary definitions and word etymologies.
This approach is richer than the weighted additive models because the matrices act as linear maps on the vectors they take as “arguments,” and thus can encode more subtle syntactic or semantic relations. However, this model treats all word combinations as the same operation—for example, treating the combination of an adjective with its argument and a verb with its subject as the same sort of composition. Because of the diverse ways there are of training such supervised models, we leave it to those who wish to further develop this specific line of research to perform such evaluations.
Adjective Matrices. The second approach is the matrix-composition model of Baroni and Zamparelli (2010), which they develop only for the case of adjective-noun composition, although their approach can seamlessly be used for any other predicate-argument composition. Contrary to most of the earlier approaches proposed, which aim to combine two lexical vectors to form a lexical vector for their combination, Baroni and Zamparelli suggest giving different semantic representations to different types, or more specifically to adjectives and nouns.
In this model, nouns are lexical vectors, as with other models. However, embracing a view of adjectives that is more in line with formal semantics than with distributional semantics, they model adjectives as linear maps taking lexical vectors as input and producing lexical vectors as output. Such linear maps can be encoded as square matrices, and applied to their arguments by matrix multiplication. Concretely, let Madjective be the matrix encoding the adjective’s linear map, and −−→noun be the lexical semantic vector for a noun; their combination is simply
−−−−−−−−−→ adjective noun = Madjective ×−−→noun
Similarly to the Generic Additive Model, the matrix for each adjective is learned by linear regression over a set of pairs (−−→noun,−→c ) where the vectors −−→noun are the lexical semantic vectors for the arguments of the adjective in a corpus, and −→c is the semantic vector corresponding to the expected output of the composition of the adjective with that noun.
This may, at first blush, also appear to be a supervised training method for learning adjective matrices from “labeled data,” seeing how the expected output vectors are needed. However, Baroni and Zamparelli (2010) work around this constraint by automatically producing the labeled data from the corpus by treating the adjectivenoun compound as a single token, and learning its vector using the same distributional learning methods they used to learn the vectors for nouns. This same approach can be extended to other unary relations without change and, using the general framework of the current article, an extension of it to binary predicates has been presented in Grefenstette et al. (2013), using multistep regression. For a direct comparison of the results of this approach with some of the results of the current article, we refer the reader to Grefenstette et al. (2013).
Recursive Matrix-Vector Model. The third approach is the recently developed Recursive Matrix-Vector Model (MV-RNN) of Socher et al. (2012), which claims the two matrixbased models described here as special cases. In MV-RNN, words are represented as a pairing of a lexical semantic vector −→a with an operation matrix A. Within this model, given the parse of a sentence in the form of a binarized tree, the semantic representation (−→c , C) of each non-terminal node in the tree is produced by performing the following two operations on its children (−→a , A) and (−→b , B).
First, the vector component −→c is produced by applying the operation matrix of one child to the vector of the other, and vice versa, and then projecting both of the products back into the same vector space as the child vectors using a projection matrix W, which must also be learned:
−→c = W × [
B ×−→a A ×−→b
]
Second, the matrix C is calculated by projecting the pairing of matrices A and B back into the same space, using a projection matrix WM, which must also be learned:
C = WM × [
A B
]
The pairing (−→c , C) obtained through these operations forms the semantic representation of the phrase falling under the scope of the segment of the parse tree below that node.
This approach to compositionality yields good results in the experiments described in Socher et al. (2012). It furthermore has appealing characteristics, such as treating relational words differently through their operation matrices, and allowing for recursive composition, as the output of each composition operation is of the same type of object as its inputs. This approach is highly general and has excellent coverage of different syntactic types, while leaving much room for parametric optimization.
The principal mathematical difference with the compositional framework presented subsequently is that composition in MV-RNN is always a binary operation; for example, to compose a transitive verb with its subject and object one would first need to compose it with its object, and then compose the output of that operation with the subject. The framework we discuss in this article allows for the construction of larger representations for relations of larger arities, permitting the simultaneous composition of a verb with its subject and object. Whether or not this theoretical difference leads to significant differences in composition quality requires joint evaluation. Additionally, the description of MV-RNN models in Socher et al. (2012) specifies the need for a source of learning error during training, which is easy to measure in the case of label prediction experiments such as sentiment prediction, but non-trivial in the case of paraphrase detection where no objective label exists. A direct comparison to MV-RNN methods within the context of experiments similar to those presented in this article has been produced by Blacoe and Lapata (2012), showing that simple operations perform on par with the earlier complex deep learning architectures produced by Socher and colleagues; we leave direct comparisons to future work. Early work has shown that the addition of a hidden layer with non-linearities to these simple models will improve the results. Domains and Functions. In recent work, Turney (2012) suggests modeling word representations not as a single semantic vector, but as a pair of vectors: one containing the information of the word relative to its domain (the other words that are ontologically similar), and another containing information relating to its function. The former vector is learned by looking at what nouns a word co-occurs with, and the latter is learned by looking at what verb-based patterns the word occurs in. Similarity between sets of words is not determined by a single similarity function, but rather through a combination of comparisons of the domain components of words’ representations with the function components of the words’ representations. Such combinations are designed on a task-specific basis. Although Turney’s work does not directly deal with vector composition of the sort we explore in this article, Turney shows that similarity measures can be designed for tasks similar to those presented here. The particular limitation of his approach, which Turney discusses, is that similarity measures must be specified for each task, whereas most of the compositional models described herein produce representations that can be compared in a task-independent manner (e.g., through cosine similarity). Nonetheless, this approach is innovative, and will merit further attention in future work in this area.
Language as Algebra. A theoretical model of meaning as context has been proposed in Clarke (2009, 2012). In that model, the meaning of any string of words is a vector built
from the occurrence of the string in a corpus. This is the most natural extension of distributional models from words to strings of words: in that model, one builds vectors for strings of words in exactly the same way as one does for words. The main problem, however, is that of data sparsity for the occurrences of strings of words. Words do appear repeatedly in a document, but strings of words, especially for longer strings, rarely do so; for instance, it hardly happens that an exact same sentence appears more than once in a document. To overcome this problem, the model is based on the hypothetical concept of an infinite corpus, an assumption that prevents it from being applied to real-world corpora and experimented within natural language processing tasks. On the positive side, the model provides a theoretical study of the abstract properties of a general bilinear associative composition operator; in particular, it is shown that this operator encompasses other composition operators, such as addition, multiplication, and even tensor product.","3 discocat :In Section 2, we discussed lexical DSMs and the problems faced by attempts to provide a vector composition operation that would allow us to form distributional sentence representations as a function of word meaning. In this section, we will present an existing formalism aimed at solving this compositionality problem, as well as the mathematical background required to understand it and further extensions, building on the features and failures of previously discussed attempts at syntactically sensitive compositionality.
Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) propose adapting a category theoretic model, inspired by the categorical compositional vector space model of quantum protocols (Abramsky and Coecke 2004), to the task of compositionality of semantic vectors. Syntactic analysis in the form of pregroup grammars—a categorial grammar—is given categorical semantics in order to be represented as a compact closed category P (a concept explained subsequently), the objects of which are syntactic types and the morphisms of which are the reductions forming the basis of syntactic analysis. This syntactic category is then mapped onto the semantic compact closed category FVect of finite dimensional vector spaces and linear maps. The mapping is done in the product category FVect × P via the following procedure. Each syntactic type is interpreted as a vector space in which semantic vectors for words with that particular syntactic type live; the reductions between the syntactic types are interpreted as linear maps between the interpreted vector spaces of the syntactic types.
The key feature of category theory exploited here is its ability to relate in a canonical way different mathematical formalisms that have similar structures, even if the original formalisms belong in different branches of mathematics. In this context, it has enabled us to relate syntactic types and reductions to vector spaces and linear maps and obtain a mechanism by which syntactic analysis guides semantic composition operations.
This pairing of syntactic analysis and semantic composition ensures both that grammaticality restrictions are in place as in the model of Erk and Padó (2008) and syntactically driven semantic composition in the form of inner-products provide the implicit disambiguation features of the compositional models of Erk and Padó (2008) and Mitchell and Lapata (2008). The composition mechanism also involves the projection of tensored vectors into a common semantic space without the need for full representation of the tensored vectors in a manner similar to Plate (1991), without restriction to the nature of the vector spaces it can be applied to. This avoids the problems faced by other tensor-based composition mechanisms such as Smolensky (1990) and Clark and Pulman (2007).
The word vectors can be specified model-theoretically and the sentence space can be defined over Boolean values to obtain grammatically driven truth-theoretic semantics in the style of Montague (1974), as proposed by Clark, Coecke, and Sadrzadeh (2008). Some logical operators can be emulated in this setting, such as using swap matrices for negation as shown by Coecke, Sadrzadeh, and Clark (2010). Alternatively, corpus-based variations on this formalism have been proposed by Grefenstette et al. (2011) to obtain a non-truth theoretic semantic model of sentence meaning for which logical operations have yet to be defined.
Before explaining how this formalism works, in Section 3.3, we will introduce the notions of pregroup grammars in Section 3.1, and the required basics of category theory in Section 3.2. Presented by Lambek (1999, 2008) as a successor to his syntactic calculus (Lambek 1958), pregroup grammars are a class of categorial type grammars with pregroup algebras as semantics. Pregroups are particularly interesting within the context of this work because of their well-studied algebraic structure, which can trivially be mapped onto the structure of the category of vector spaces, as will be discussed subsequently. Logically speaking, a pregroup is a non-commutative form of Linear Logic (Girard 1987) in which the tensor and its dual par coincide; this logic is sometimes referred to as Bi-Compact Linear Logic (Lambek 1999). The formalism works alongside the general guidelines of other categorial grammars, for instance, those of the combinatory categorial grammar (CCG) designed by Steedman (2001) and Steedman and Baldridge (2011). They consist of atomic grammatical types that combine to form compound types. A series of CCGlike application rules allow for type-reductions, forming the basis of syntactic analysis. As our first step, we show how this syntactic analysis formalism works by presenting an introduction to pregroup algebras.
Pregroups. A pregroup is an algebraic structure of the form (P,≤, ·, 1, (−)l, (−)r). Its elements are defined as follows:
P is a set {a, b, c, . . .}. The relation ≤ is a partial ordering on P. The binary operation · is an associative, non-commutative monoid multiplication with the type − · − : P × P → P, such that if a, b ∈ P then a · b ∈ P. In other words, P is closed under this operation.
1 ∈ P is the unit, satisfying a · 1 = a = 1 · a for all a ∈ P. (−)l and (−)r are maps with types (−)l : P → P and (−)r : P → P such that for any a ∈ P, we have that al, ar ∈ P. The images of these maps are referred to as the left and the right adjoints. These are unique and satisfy the following conditions:
– Reversal: if a ≤ b then bl ≤ al (and similarly for ar, br). – Ordering: a · ar ≤ 1 ≤ ar · a and al · a ≤ 1 ≤ a · al. – Cancellation: alr = a = arl. – Self-adjointness of identity: 1r = 1 = 1l. – Self-adjointness of multiplication: (a · b)r = br · ar.
As a notational simplification we write ab for a · b, and if abcd ≤ cd we write abcd → cd and call this a reduction, omitting the identity wherever it might appear. Monoid multiplication is associative, so parentheses may be added or removed for notational clarity without changing the meaning of the expression as long as they are not directly under the scope of an adjoint operator.
An example reduction in pregroup might be:
aarbclc → bclc → b
We note here that the reduction order is not always unique, as we could have reduced as follows: aarbclc → aarb → b. As a further notational simplification, if there exists a chain of reductions a → . . . → b we may simply write a → b (in virtue of the transitivity of partial ordering relations). Hence in our given example, we can express both reduction paths as aarbclc → b.
Pregroups and Syntactic Analysis. Pregroups are used for grammatical analysis by freely generating the set P of a pregroup from the basic syntactic types n, s, . . . , where here n stands for the type of both a noun and a noun phrase and s for that of a sentence. The conflation of nouns and noun phrases suggested here is done to keep the work discussed in this article as simple as possible, but we could of course model them as different types in a more sophisticated version of this pregroup grammar. As in any categorial grammar, words of the lexicon are assigned one or more possible types (corresponding to different syntactic roles) in a predefined type dictionary, and the grammaticality of a sentence is verified by demonstrating the existence of a reduction from the combination of the types of words within the sentence to the sentence type s.
For example, having assigned to noun phrases the type n and to sentences the type s, the transitive verbs will have the compound type nrsnl. We can read from the type of a transitive verb that it is the type of a word which “expects” a noun phrase on its left and a noun phrase on its right, in order to produce a sentence. A sample reduction of John loves cake with John and cake being noun phrases of type n and loves being a verb of type nrsnl is as follows:
n(nrsnl)n → s
We see that the transitive verb has combined with the subject and object to reduce to a sentence. Because the combination of the types of the words in the string John loves cake reduces to s, we say that this string of words is a grammatical sentence. As for more examples, we recall that intransitive verbs can be given the type nrs such that John sleeps would be analyzed in terms of the reduction n(nrs) → s. Adjectives can be given the type nnl such that red round rubber ball would be analyzed by (nnl)(nnl)(nnl)n → n. And so on and so forth for other syntactic classes.
Lambek (2008) presents the details of a slightly more complex pregroup grammar with a richer set of types than presented here. This grammar is hand-constructed and iteratively extended by expanding the type assignments as more sophisticated grammatical constructions are discussed. No general mechanism is proposed to cover all such types of assignments for larger fragments (e.g., as seen in empirical data). Pregroup grammars have been proven to be learnable by Béchet, Foret, and Tellier (2007), who also discuss the difficulty of this task and the nontractability of the procedure. Because of these constraints and lack of a workable pregroup parser, the pregroup grammars
we will use in our categorical formalism are derived from CCG types, as we explain in the following.
Pregroup Grammars and Other Categorial Grammars. Pregroup grammars, in contrast with other categorial grammars such as CCG, do not yet have a large set of tools for parsing available. If quick implementation of the formalism described later in this paper is required, it would be useful to be able to leverage the mature state of parsing tools available for other categorial grammars, such as the Clark and Curran (2007) statistical CCG parser, as well as Hockenmaier’s CCG lexicon and treebank (Hockenmaier 2003; Hockenmaier and Steedman 2007). In other words, is there any way we can translate at least some subset of CCG types into pregroup types?
There are some theoretical obstacles to consider first: Pregroup grammars and CCG are not equivalent. Buszkowski (2001) shows pregroup grammars to be equivalent to context-free grammars, whereas Joshi, Vijay-Shanker, and Weir (1989) show CCG to be weakly equivalent to more expressive mildly context-sensitive grammars. However, if our goal is to exploit the CCG used in Clark and Curran’s parser, or Hockenmaier’s lexicon and treebank, we may be in luck: Fowler and Penn (2010) prove that some CCGs, such as those used in the aforementioned tools, are strongly context-free and thus expressively equivalent to pregroup grammars. In order to be able to apply the parsing tools for CCGs to our setting, we use a translation mechanism from CCG types to pregroup types based on the Lambek-calculus-to-pregroup-grammar translation originally presented in Lambek (1999). In this mechanism, each atomic CCG type X is assigned a unique pregroup type x; for any X/Y in CCG we have xyl in the pregroup grammar; and for any X\Y in CCG we have yrx in pregroup grammar. Therefore, by assigning NP to n and S to s we could, for example, translate the CCG transitive verb type (S\NP)/NP into the pregroup type nrsnl, which corresponds to the pregroup type we used for transitive verbs in Section 3.1. Wherever type replacement (e.g., N → NP) is allowed in CCG we set an ordering relation in the pregroup grammar (e.g., n̄ ≤ n, where n̄ is the pregroup type associated with N). Because forward and backward slash “operators” in CCG are not associative whereas monoid multiplication in pregroups is, it is evident that some information is lost during the translation process. But because the translation we need is one-way, we may ignore this problem and use CCG parsing tools to obtain pregroup parses. Another concern lies with CCG’s crossed composition and substitution rules. The translations of these rules do not in general hold in a pregroup; this is not a surprise as pregroups are a simplification of the Lambek Calculus and these rules did not hold in the Lambek Calculus either, as shown in Moortgat (1997), for example. However, for the phenomena modeled in this paper, the CCG rules without the backward cross rules will suffice. In general for the case of English, one can avoid the use of these rules by overloading the lexicon and using additional categories. To deal with languages that have cross dependancies, such as Dutch, various solutions have been suggested (e.g., see Genkin, Francez, and Kaminski 2010; Preller 2010). Category theory is a branch of pure mathematics that allows for a general and uniform formulation of various different mathematical formalisms in terms of their main structural properties using a few abstract concepts such as objects, arrows, and combinations and compositions of these. This uniform conceptual language allows for derivation of new properties of existing formalisms and for relating these to properties of other formalisms, if they bear similar categorical representation. In this function, it has been
at the center of recent work in unifying two orthogonal models of meaning, a qualitative categorial grammar model and a quantitative distributional model (Clark, Coecke, and Sadrzadeh 2008; Coecke, Sadrzadeh, and Clark 2010). Moreover, the unifying categorical structures at work here were inspired by the ones used in the foundations of physics and the modeling of quantum information flow, as presented in Abramsky and Coecke (2004), where they relate the logical structure of quantum protocols to their state-based vector spaces data. The connection between the mathematics used for this branch of physics and those potentially useful for linguistic modeling has also been noted by several sources, such as Widdows (2005), Lambek (2010), and Van Rijsbergen (2004).
In this section, we will briefly examine the basics of category theory, monoidal categories, and compact closed categories. The focus will be on defining enough basic concepts to proceed rather than provide a full-blown tutorial on category theory and the modeling of information flow, as several excellent sources already cover both aspects (e.g., Mac Lane 1998; Walters 1991; Coecke and Paquette 2011). A categoriesin-a-nutshell crash course is also provided in Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010).
The Basics of Category Theory. A basic category C is defined in terms of the following elements:
A collection of objects ob(C). A collection of morphisms hom(C). A morphism composition operation ◦.
Each morphism f has a domain dom( f ) ∈ ob(C) and a codomain codom( f ) ∈ ob(C). For dom( f ) = A and codom( f ) = B we abbreviate these definitions as f : A → B. Despite the notational similarity to function definitions (and sets and functions being an example of a category), it is important to state that nothing else is presupposed about morphisms, and we should not treat them a priori as functions.
The following axioms hold in every category C :
For any f : A → B and g : B → C there exists h : A → C and h = g ◦ f . For any f : A → B, g : B → C and h : C → D, ◦ satisfies (h ◦ g) ◦ f = h ◦ (g ◦ f ). For every A ∈ ob(C) there is an identity morphism idA : A → A such that for any f : A → B, f ◦ idA = f = idB ◦ f .
We can model various mathematical formalisms using these basic concepts, and verify that these axioms hold for them. For example, category Set with sets as objects and functions as morphisms, or category Rel with sets as objects and relations as morphisms, category Pos with posets as objects and order-preserving maps as morphisms, and category Group with groups as objects and group homomorphisms as morphisms, to name a few.
The product C × D of two categories C and D is a category with pairs (A, B) as objects, where A ∈ ob(C) and B ∈ ob(D). There exists a morphism (f, g) : (A, B) → (C, D) in C × D if and only if there exists f : A → C ∈ hom(C) and g : B → D ∈ hom(D). Product categories allow us to relate objects and morphisms of one mathematical formalism to those in another, in this example those of C to D.
Compact Closed Categories. A monoidal category C is a basic category to which we add a monoidal tensor ⊗ such that:
For all A, B ∈ ob(C) there is an object A ⊗ B ∈ ob(C). For all A, B, C ∈ ob(C), we have (A ⊗ B) ⊗ C = A ⊗ (B ⊗ C). There exists some I ∈ ob(C) such that for any A ∈ ob(C), we have I ⊗ A = A = A ⊗ I. For f : A → C and g : B → D in hom(C) there is f ⊗ g : A ⊗ B → C ⊗ D in hom(C). For f1 : A → C, f2 : B → D, g1 : C → E and g2 : D → F the following equality holds:
(g1 ⊗ g2) ◦ ( f1 ⊗ f2) = (g1 ◦ f1) ⊗ (g2 ◦ f2) The product category of two monoidal categories has a monodial tensor, defined pointwisely by (a, A) ⊗ (b, B) := (a ⊗ b, A ⊗ B).
A compact closed category C is a monoidal category with the following additional axioms:
Each object A ∈ ob(C) has left and right “adjoint” objects Al and Ar in ob(C).
There exist four structural morphisms for each object A ∈ ob(C): – ηlA : I → A ⊗ Al. – ηrA : I → Ar ⊗ A. – lA : A
l ⊗ A → I. – rA : A ⊗ Ar → I.
The previous structural morphisms satisfy the following equalities:
– (1A ⊗ lA) ◦ (ηlA ⊗ 1A) = 1A. – ( rA ⊗ 1A) ◦ (1A ⊗ ηrA) = 1A. – (1Ar ⊗ rA) ◦ (ηr ⊗ 1Ar ) = 1Ar . – ( lA ⊗ 1Al ) ◦ (1Al ⊗ ηlA) = 1Al .
Compact closed categories come equipped with complete graphical calculi, surveyed in Selinger (2010). These calculi visualize and simplify the axiomatic reasoning within the category to a great extent. In particular, Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) show how they depict the pregroup grammatical reductions and visualize the flow of information in composing meanings of single words and forming meanings for sentences. Although useful at an abstract level, these calculi do not play the same simplifying role when it comes to the concrete and empirical computations; therefore we will not discuss them in this article.
A very relevant example of a compact closed category is a pregroup algebra P. Elements of a pregroup are objects of the category, the partial ordering relation provides the morphisms, 1 is I, and monoidal multiplication is the monoidal tensor.
Another very relevant example is the category FVect of finite dimensional Hilbert spaces and linear maps over R—that is, vector spaces over R with orthogonal bases of finite dimension, and an inner product operation 〈− | −〉 : A × A → R for every vector space A. The objects of FVect are vector spaces, and the morphisms are linear
maps between vector spaces. The unit object is R and the monoidal tensor is the linear algebraic tensor product. FVect is degenerate in its adjoints, in that for any vector space A, we have Al = Ar = A∗, where A∗ is the dual space of A. Moreover, by fixing a basis we obtain that A∗ ∼= A. As such, we can effectively do away with adjoints in this category, and “collapse” l, r, ηl, and ηr maps into “adjoint-free” and η maps. In this category, the maps are inner product operations, A : A ⊗ A → R, and the η maps η : R → A ⊗ A generate maximally entangled states, also referred to as Bell-states. In Section 3.2 we showed how a pregroup algebra and vector spaces can be modeled as compact closed categories and how product categories allow us to relate the objects and morphisms of one category to those of another. In this section, we will present how Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) suggest building on this by using categories to relate semantic composition to syntactic analysis in order to achieve syntax-sensitive composition in DSMs.
3.3.1 Syntax Guides Semantics. The product category FVect × P has as object pairs (A, a), where A is a vector space and a is a pregroup grammatical type, and as morphism pairs ( f,≤) where f is a linear map and ≤ a pregroup ordering relation. By the definition of product categories, for any two vector space-type pairs (A, a) and (B, b), there exists a morphism (A, a) → (B, b) only if there is both a linear map from A into B and a partial ordering a → b. If we view these pairings as the association of syntactic types with vector spaces containing semantic vectors for words of that type, this restriction effectively states that a linear map from A to B is only “permitted” in the product category if a reduces to b.
Both P and FVect being compact closed, it is simple to show that FVect × P is as well, by considering the pairs of unit objects and structural morphisms from the separate categories: I is now (R, 1), and the structural morphisms are ( A, la), ( A, r a), (ηA,η l a), and (ηA,ηra). We are particularly interested in the maps, which are defined as follows (from the definition of product categories):
( A, lA) : (A ⊗ A, ala) → (R, 1) ( A, rA) : (A ⊗ A, aar) → (R, 1)
This states that whenever there is a reduction step in the grammatical analysis of a sentence, there is a composition operation in the form of an inner product on the semantic front. Hence, if nouns of type n live in some noun space N and transitive verbs of type nlsnr live in some space N ⊗ S ⊗ N, then there must be some structural morphism of the form:
( N ⊗ 1S ⊗ N, rn1s ln) : (N ⊗ (N ⊗ S ⊗ N) ⊗ N, n(nrsnl)n) → (S, s)
We can read from this morphism the functions required to compose a sentence with a noun, a transitive verb, and an object to obtain a vector living in some sentence space S, namely, ( N ⊗ 1S ⊗ N ).
The form of a syntactic type is therefore what dictates the structure of the semantic space associated with it. The structural morphisms of the product category guarantee that for every syntactic reduction there is a semantic composition morphism provided by the product category: syntactic analysis guides semantic composition.
3.3.2 Example. To give an example, we give syntactic type n to nouns, and nrs to intransitive verbs. The grammatical reduction for kittens sleep, namely, nnrs → s, corresponds to the morphism rn ⊗ 1s in P. The syntactic types dictate that the noun −−−−→ kittens lives in some vector space N, and the intransitive verb −−−→ sleep in N ⊗ S. The reduction morphism gives us the composition morphism ( N ⊗ 1S), which we can apply to −−−−→ kittens ⊗−−−→sleep. Because we can express any vector as the weighted sum of its basis vectors, let us expand −−−−→ kittens = ∑ i c kittens i −→ni and −−−→ sleep = ∑ ij c sleep ij
−→ni ⊗−→sj ; then we can express the composition as follows:
−−−−−−−−→ kittens sleep = ( N ⊗ 1S)( −−−−→ kittens ⊗−−−→sleep)
= ( N ⊗ 1S) ⎛ ⎝∑
i
ckittensi −→ni ⊗ ∑ jk csleepjk −→nj ⊗−→sk
⎞ ⎠
= ( N ⊗ 1S) ⎛ ⎝∑
ijk
ckittensi c sleep jk −→ni ⊗−→nj ⊗−→sk
⎞ ⎠
= ∑
ijk
ckittensi c sleep jk 〈−→ni | −→nj 〉−→sk
= ∑
ik
ckittensi c sleep ik −→sk
In these equations, we have expressed the vectors in their explicit form, we have consolidated the sums by virtue of distributivity of linear algebraic tensor product over addition, we have applied the tensored linear maps to the vector components (as the weights are scalars), and finally, we have simplified the indices since 〈−→ni | −→nj 〉 = 1 if−→ni = −→nj and 0 otherwise. As a result of these, we have obtained a vector that lives in sentence space S.
Transitive sentences can be dealt with in a similar fashion:
−−−−−−−−−−−−−→ kittens chase mice
= ( N ⊗ 1S ⊗ N)( −−−−→ kittens ⊗−−−→chase ⊗−−→mice)
= ( N ⊗ 1S ⊗ N) ⎛ ⎝∑
i
ckittensi −→ni ⊗
⎛ ⎝∑
jkl
cchasejkl −→nj ⊗−→sk ⊗−→nl ⎞ ⎠⊗∑
m
cmicem −→nm
⎞ ⎠
= ( N ⊗ 1S ⊗ N) ⎛ ⎝∑
ijklm
ckittensi c chase jkl c mice m −→ni ⊗−→nj ⊗−→sk ⊗−→nl ⊗−→nm
⎞ ⎠
= ∑ ijklm ckittensi c chase jkl c mice m 〈−→ni | −→nj 〉−→sk 〈−→nl | −→nm〉
= ∑ ikm ckittensi c chase ikm c mice m −→sk
In both cases, it is important to note that the tensor product passed as argument to the composition morphism, namely,
−−−−→ kittens ⊗−−−→sleep in the intransitive case and−−−−→
kittens ⊗−−−→chase ⊗−−→mice in the transitive case, never needs to be computed. We can treat the tensor products here as commas separating function arguments, thereby avoiding the dimensionality problems presented by earlier tensor-based approaches to compositionality. As elaborated on in Section 2.3.4, the first general setting for pairing meaning vectors with syntactic types was proposed in Clark and Pulman (2007). The setting of a DisCoCat generalized this by making the meaning derivation process rely on a syntactic type system, hence overcoming its central problem whereby the vector representations of strings of words with different grammatical structure lived in different spaces. A preliminary version of a DisCoCat was developed in Clark, Coecke, and Sadrzadeh (2008), a full version was elaborated on in Coecke, Sadrzadeh, and Clark (2010), where, based on the developments of Preller and Sadrzadeh (2010), it was also exemplified how the vector space model may be instantiated in a truth theoretic setting where meanings of words were sets of their denotations and meanings of sentences were their truth values. A nontechnical description of this theoretical setting was presented in Clark (2013), where a plausibility truth-theoretic model for sentence spaces was worked out and exemplified. The work of Grefenstette et al. (2011) focused on a tangential branch and developed a toy example where neither words nor sentence spaces were Boolean. The applicability of the theoretical setting to a real empirical natural language processing task and data from a large scale corpus was demonstrated in Grefenstette and Sadrzadeh (2011a, 2011b). There, we presented a general algorithm to build vector representations for words with simple and complex types and the sentences containing them; then applied the algorithm to a disambiguation task performed on the British National Corpus (BNC). We also investigated the vector representation of transitive verbs and showed how a number of single operations may optimize the performance. We discuss these developments in detail in the following sections.","4 concrete semantic spaces :In Section 3.3 we presented a categorical formalism that relates syntactic analysis steps to semantic composition operations. The structure of our syntactic representation dictates the structure of our semantic spaces, but in exchange, we are provided with composition functions by the syntactic analysis, rather than having to stipulate them ad hoc. Whereas the approaches to compositional DSMs presented in Section 2 either failed to take syntax into account during composition, or did so at the cost of not being able to compare sentences of different structure in a common space, this categorical approach projects all sentences into a common sentence space where they can be directly compared. However, this alone does not give us a compositional DSM.
As we have seen in the previous examples, the structure of semantic spaces varies with syntactic types. We therefore cannot construct vectors for different syntactic types in the same way, as they live in spaces of different structure and dimensionality. Furthermore, nothing has yet been said about the structure of the sentence space S into which expressions reducing to type s are projected. If we wish to have a compositional DSM
that leverages all the benefits of lexical DSMs and ports them to sentence-level distributional representations, we must specify a new sort of vector construction procedure.
In the original formulation of this formalism by Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010), examples of how such a compositional DSM could be used for logical evaluation are presented, where S is defined as a Boolean space with True and False as basis vectors. However, the word vectors used are handwritten and specified model-theoretically, as the authors leave it for future research to determine how such vectors might be obtained from a corpus. In this section, we will discuss a new way of constructing vectors for compositional DSMs, and of defining sentence space S, in order to reconcile this powerful categorical formalism with the applicability and flexibility of standard distributional models. Assume the following sentences are all true:
1. The dogs chased the cats.
2. The dogs annoyed the cats.
3. The puppies followed the kittens.
4. The train left the station.
5. The president followed his agenda.
If asked which sentences have similar meaning, we would most likely point to the pair (1) and (3), and perhaps to a lesser degree (1) and (2), and (2) and (3). Sentences (4) and (5) obviously speak of a different state of the world as the other sentences.
If we compare these by truth value, we obviously have no means of making such distinctions. If we compare these by lexical similarity, (1) and (2) seem to be a closer match than (1) and (3). If we are classifying these sentences by some higher order relation such as “topic,” (5) might end up closer to (3) than (1). What, then, might cause us to pair (1) and (3)?
Intuitively, this similarity seems to be because the subjects and objects brought into relation by similar verbs are themselves similar. Abstracting away from tokens to some notion of property, we might say that both (1) and (3) express the fact that something furry and feline and furtive is being pursued by something aggressive (or playful) and canine. Playing along with the idea that lexical distributional semantics present concepts (word meanings) as “messy bundles of properties,” it seems only natural to have the way these properties are acted upon, qualified, and related as the basis for sentence-level distributional representations. In this respect, we here suggest that the sentence space S, instead of qualifying the truth value of a sentence, should express how the properties of the semantic objects within are qualified or brought into relation by verbs, adjectives, and other predicates and relations.
More specifically, we examine two suggestions for defining the sentence space, namely, SI ∼= N for sentences with intransitive verbs and ST ∼= N ⊗ N for sentences with transitive verbs. These definitions mean that the sentence space’s dimensions are commensurate with either those of N, or those of N ⊗ N. These are by no means the only options, but as we will discuss here, they offer practical benefits.
In the case of SI, the basis elements are labeled with unique basis elements of N, hence, −→s1 = −→n1 , −→s2 = −→n2, and so on. In the case of ST, the basis elements are labeled with
unique ordered pairs of elements from N, for example, −→s1 = −−−−→ (n1, n1), −→s2 = −−−−→ (n2, n1), −→s3 =−−−−→ (n1, n2), and so on, following the same matrix flattening structure used in the proof of the equal cardinality of N and Q. Because of the isomorphism between ST and N ⊗ N, we will use the notations −−−−→ (ni, nj) and
−→ni ⊗−→nj interchangeably, as both constitute appropriate ways of representing the basis elements of such a space. To propagate this distinction on the syntactic level, we define types sI and sT for intransitive and transitive sentences, respectively.
In creating this distinction, we lost one of the most appealing features of the framework of Coecke, Sadrzadeh, and Clark (2010), namely, the result that all sentence vectors live in the same sentence space. A mathematical solution to this two-space problem was suggested in Grefenstette et al. (2011), and a variant of the models presented in this article permitting the non-problematic projection into a single sentence space (S ∼= N) has been presented by Grefenstette et al. (2013). Keeping this separation allows us to deal with both the transitive and intransitive cases in a simpler manner, and because the experiments in this article only compare intransitive sentences with intransitive sentences, and transitive sentences with transitive sentences, we will not address the issue of unification here. While Lambek’s pregroup types presented in Lambek (2008) include a rich array of basic types and hand-designed compound types in order to capture specific grammatic properties, for the sake of simplicity we will use a simpler set of grammatical types for experimental purposes, similar to some common types found in the CCG-bank (Steedman 2001).
We assign a basic pregroup type n for all nouns, with an associated vector space N for their semantic representations. Furthermore, we will treat noun-phrases as nouns, assigning to them the same pregroup type and semantic space.
CCG treats intransitive verbs as functions NP\S that consume a noun phrase and return a sentence, and transitive verbs as functions (NP\S)/NP that consume a noun phrase and return an intransitive verb function, which in turn consumes a noun phrase and returns a sentence. Using our distinction between intransitive sentences, we give intransitive verbs the type nrsI associated with the semantic space N ⊗ SI, and transitive verbs the type nrsTnl associated with the semantic space N ⊗ ST ⊗ N.
Adjectives, in CCG, are treated as functions NP/NP, consuming a noun phrase and returning a noun phrase; and hence we give them the type nnl and associated semantic space N ⊗ N.
With the provision of a learning procedure for vectors in these semantic spaces, we can use these types to construct sentence vector representations for simple intransitive verb–based and transitive verb–based sentences, with and without adjectives applied to subjects and objects. To begin, we construct the semantic space N for all nouns in our lexicon (typically limited by the words available in the corpus used). Any distributional semantic model can be used for this stage, such as those presented in Curran (2004), or the lexical semantic models used by Mitchell and Lapata (2008). It seems reasonable to assume that higher quality lexical semantic vectors—as measured by metrics such as the WordSim353 test
of Finkelstein et al. (2001)—will produce better relational vectors from the procedure designed subsequently. We will not test the hypothesis here, but note that it is an underlying assumption in most of the current literature on the subject (Erk and Padó 2008; Mitchell and Lapata 2008; Baroni and Zamparelli 2010).
Building upon the foundation of the thus constructed noun vectors, we construct semantic representations for relational words. In pregroup grammars (or other combinatorial grammars such as CCG), we can view such words as functions, taking as arguments those types present as adjoints in the compound pregroup type, and returning a syntactic object whose type is that of the corresponding reduction. For example, an adjective nnl takes a noun or noun phrase n and returns a noun phrase n from the reduction (nnl)n → n. It can also compose with another adjective to return an adjective (nnl)(nnl) → nnl. We wish for our semantic representations to be viewed in the same way, such that the composition of an adjective with a noun (1N ⊗ N)((N ⊗ N) ⊗ N) can be viewed as the application of a function f : N → N to its argument of type N.
To learn the representation of such functions, we assume that their meaning can be characterized by the properties that their arguments hold in the corpus, rather than just by their context as is the case in lexical distributional semantic models. To give an example, rather than learning what the adjective angry means by observing that it co-occurs with words such as fighting, aggressive, or mean, we learn its meaning by observing that it typically takes as argument words that have the property of being (i.e., co-occur with words such as) fighting, aggressive, and mean. Whereas in the lexical semantic case, such associations might only rarely occur in the corpus, in this indirect method we learn what properties the adjective relates to even if they do not co-occur with it directly.
In turn, through composition with its argument, we expect the function for such an adjective to strengthen the properties that characterize it in the object it takes as argument. If, indeed, angry is characterized by arguments that have high basis weights for basis elements corresponding to the concepts (or context words) fighting, aggressive, and mean, and relatively low counts for semantically different concepts such as passive, peaceful, and loves, then when we apply angry to dog the vector for the compound angry dog should contain some of the information found in the vector for dog. But this vector should also have higher values for the basis weights of fighting, aggressive, and mean, and correspondingly lower values for the basis weights of passive, peaceful, and loves.
To turn this idea into a concrete algorithm for constructing the semantic representation for relations of any arity, as first presented in Grefenstette et al. (2011), we examine how we would deal with this for binary relations such as transitive verbs. If a transitive verb of semantic type N ⊗ ST ⊗ N is viewed as a function f : N × N → ST that expresses the extent to which the properties of subject and object are brought into relation by the verb, we learn the meaning of the verb by looking at what properties are brought into relation by its arguments in a corpus. Recall that the vector for a verb v,−→v ∈ N ⊗ ST ⊗ N, can be expressed as the weighted sum of its basis elements:
−→v = ∑
ijk
cvijk −→ni ⊗−→sj ⊗−→nk
We take the set of vectors for the subject and object of v in the corpus to be argv = {(−−−−→SUBJl, −−−→ OBJl )}l. We wish to calculate the basis weightings {cvijk}ijk for v. Exploiting our earlier definition of the basis {sj}j of ST, which states that for any value of i and k there
is some value of j such that sj = (ni, nk), we define Δijk = 1 if indeed sj = (ni, nk) and 0 otherwise. Using all of this, we define the calculation of each basis weight cvijk as:
cvijk = ∑
l
Δijkc SUBJl i c OBJl k
This allows for a full formulation of −→v as follows: −→v =
∑ l ∑ ijk Δijkc SUBJl i c OBJl k −→ni ⊗−→sj ⊗−→nk
The nested sums here may seem computationally inefficient, seeing how this would involve computing size(argv) × dim(N)2 × dim(S) = size(argv) × dim(N)4 products. However, using the decomposition of basis elements of S into pairs of basis elements of N (effectively basis elements of N ⊗ N), we can remove the Δijk term and ignore all values of j where sj = (ni, nk), because the basis weight for this combination of indices would be 0. Hence we simplify:
−→v = ∑
l ∑ ik cSUBJli c OBJl k −→ni ⊗ −−−−→ (ni, nk) ⊗−→nk
This representation is still bloated: We perform fewer calculations, but still obtain a vector in which all the basis weights where sj = (ni, nk) are 0, hence where only dim(N)2 of the dim(N)4 values are non-zero. In short, the vector weights for −→v are, under this learning algorithm, entirely characterized by the values of a dim(N) by dim(N) matrix, the entries of which are products cSUBJli c OBJl k where i and k have become row and column indices. Using this and our definition of ST as a space isomorphic to N ⊗ N, we can formulate a compact expression of −→v as follows. Let the kronecker product of two vectors−→u ,−→w ∈ N, written −→u ⊗−→w ∈ N ⊗ N, be as follows: −→u ⊗−→w =
∑ ij cui c w j −→ni ⊗−→nj
Equipped with this definition, we can formulate the compact form of −→v :
−→v = ∑
l ∑ ik cSUBJli c OBJl k −→ni ⊗−→nk
= ∑
l
−−−→ SUBJl ⊗ −−→ OBJl
In short, we are only required to iterate through the corpus once, taking for each instance of a transitive verb v the kronecker product of its subject and object, and summing these across all instances of v. It is simple to see that no information was discarded relative to the previous definition of −→v : The dimensionality reduction by a factor of dim(N)2 simply discards all basis elements for which the basis weight was 0 by default.
This raises a small problem though: This compact representation can no longer be used in the compositional mechanism presented in Section 3.3, as the dimensions of−→v no longer match those that it is required to have according to its syntactic type. However, a solution can be devised if we return to the sample calculation shown in Section 3.3.2 of the composition of a transitive verb with its arguments. The composition reduces as follows:
( N ⊗ 1S ⊗ N)( −−−→ SUBJ ⊗−→v ⊗−−→OBJ) = ∑ ikm cSUBJi c v ikmc OBJ m −→sk
where the verb v is represented in its non-compact form. If we introduce the compactness given to us by our isomorphism ST ∼= N ⊗ N we can express this as
−−−−−−−−→ SUBJ v OBJ = ∑ im cSUBJi c v imc OBJ m −→ni ⊗−→nm
where v is represented in its compact form. Furthermore, by introducing the component wise multiplication operation :
−→u −→v = ∑
i
cui c v i −→ni
we can show the general form of transitive verb composition with the reduced verb representation to be as follows:
−−−−−−−−→ SUBJ v OBJ = ∑ im cSUBJi c v imc OBJ m −→ni ⊗−→nm
= (∑ im cvim −→ni ⊗−→nm ) (∑ im cSUBJi c OBJ m −→ni ⊗−→nm )
= −→v (−−−→ SUBJ ⊗−−→OBJ )
To summarize what we have done with transitive verbs:
1. We have treated them as functions, taking two nouns and returning a sentence in a space ST ∼= N ⊗ N.
2. We have built them by counting what properties of subject and object noun vectors are related by the verb in transitive sentences in a corpus.
3. We have assumed that output properties were a function of input properties by the use of the Δijk function—for example, the weights associated with ni from the subject argument and with nk from the object argument only affect the “output” weight for the sentence basis element −→sj = −→ni ⊗−→nk .
4. We have shown that this leads to a compact representation of the verb’s semantic form, and an optimized learning procedure.
5. We have shown that the composition operations of our formalism can be adapted to this compact representation.
The compact representation and amended composition operation hinge on the choice of ST ∼= N ⊗ N as output type for N ⊗ N as an input type (the pair of arguments the verb takes), justifying our choice of a transitive sentence space. In the intransitive case, the same phenomenon can be observed, since such a verb takes as argument a vector in N and produces a vector in SI ∼= N. Furthermore, our choice to make all other types dependent on one base type—namely, n (with associated semantic space N)— yields the same property for every relational word we wish to learn: The output type is the same as the concatenated (on the syntactic level) or tensored (on the semantic level) input types. It is this symmetry between input and output types that guarantees that any m-ary relation, expressed in the original formulation as an element of tensor space N ⊗ . . .⊗ N︸ ︷︷ ︸
2m
, has a compact representation in N ⊗ . . .⊗ N︸ ︷︷ ︸ m , where the ith basis weight
of the reduced representation stands for the degree to which the ith element of the input vector affects the ith element of the output vector.
This allows the specification of the generalized learning algorithm for reduced representations, first presented in Grefenstette and Sadrzadeh (2011a), which is as follows. Each relational word P with grammatical type π and m adjoint types α1,α2, · · ·,αm is encoded as an (r × . . .× r) multi-dimensional array with m degrees of freedom (i.e., a rank m tensor). Because our vector space N has a fixed basis, each such array is represented in vector form as follows:
−→ P = ∑ ij · · ·ζ︸ ︷︷ ︸
m
cij···ζ ( −→n i ⊗−→n j ⊗ · · · ⊗ −→n ζ)︸ ︷︷ ︸
m
This vector lives in the tensor space N ⊗ N ⊗ · · · ⊗ N︸ ︷︷ ︸ m . Each cij···ζ is computed according to the procedure described in Figure 2.
Linear algebraically, this procedure corresponds to computing the following
−→ P = ∑ k (−→w1 ⊗−→w2 ⊗ · · · ⊗ −→wm)k
The general formulation of composing a relational word P with its arguments arg1, . . . , argm is now expressed as
−→ P (−−→arg1 ⊗ . . .⊗−−→argm)
For example, the computation of furry cats nag angry dogs would correspond to the following operations:
−−−−−−−−−−−−−−−−−−→ furry cats nag angry dogs = −−→nag ((−−−→ furry −→cat ) ⊗ (−−−→angry −−→dog))
This generalized learning algorithm effectively extends the coverage of our approach to any sentence for which we have a pregroup parse (e.g., as might be obtained by our translation mechanism from CCG). For example, determiners would have a type nnl, allowing us to model them as matrices in N ⊗ N; adverbs would be of type (nrs)r(nrs), and hence be elements of S ⊗ N ⊗ N ⊗ S. We could learn them using the procedure defined earlier, although for words like determiners and conjunctions, it is unclear whether we would want to learn such logical words or design them by hand, as was done in Coecke, Sadrzadeh, and Clark (2010) for negation. Nonetheless, the procedure given here allows us to generalize the work presented in this article to sentences of any length or structure based on the pregroup types of the words they contain. To conclude this section with an example taken from Grefenstette and Sadrzadeh (2011a), we demonstrate how the meaning of the word show might be learned from a corpus and then composed.
Suppose there are two instances of the verb show in the corpus: s1 = table show result s2 = map show location
The vector of show is
−−−→ show = −−→ table ⊗−−−→result + −−→map ⊗−−−−−→location
Consider a vector space N with four basis vectors far, room, scientific, and elect. The TF/IDF-weighted values for vectors of selected nouns (built from the BNC) are as shown in Table 1. Part of the matrix compact representation of show is presented in Table 2.
As a sample computation, the weight c11 for vector ( −→n1,−→n1 ), that is, ( −→ far, −→ far), is computed by multiplying weights of table and result on −→ far (6.6 × 7), multiplying weights of map and location on −→ far (5.6 × 5.9), and then adding these (46.2 + 33.04) and obtaining the total weight 79.24. Similarly, the weight c21 for vector ( −→n2,−→n1 ), that is, (−−−→room, −→ far), is computed by multiplying the weight of −−−→room for table by the weight of −→far for result (27 × 7), then multiplying the weight of −−−→room for map by the weight of −→far for location (7.4 × 5.9), and then adding these (189 + 43.66) to obtain the total weight of 232.66.
We now wish to compute the vector for the sentence [the] master shows [his] dog, omitting the determiner and possessive for simplicity, as we have left open the question as to whether or not we would want to learn them using the generalized procedure from Section 4.3 or specify them by design due to their logical structure. The calculation according to the vectors in Table 1 and the matrix in Table 2 will be:
−−−−−−−−−−−−→ master show dog
= −−−→ show (−−−−→ master ⊗−−→dog )
= ⎡ ⎢⎢⎣ 79.24 47.41 119.96 27.72 232.66 80.75 396.14 113.2 32.94 31.86 32.94 0
0 0 0 0
⎤ ⎥⎥⎦ ⎡ ⎢⎢⎣
23.65 32.68 0 0 50.05 69.16 0 0 22.55 31.16 0 0 34.1 47.12 0 0
⎤ ⎥⎥⎦
= ⎡ ⎢⎢⎣ 1,874.03 1,549.36 0 0 11,644.63 5,584.67 0 0
742.80 992.76 0 0 0 0 0 0
⎤ ⎥⎥⎦
Row-wise, flattening the final matrix representation gives us the result we seek, namely, the sentence vector in ST for [the] master showed [his] dog:
−−−−−−−−−−−−→ master show dog = [1,874, 1,549, 0, 0, 11,645, 5,585, 0, 0, 743, 993, 0, 0, 0, 0, 0, 0]","5 experiments :Evaluating compositional models of semantics is no easy task: First, we are trying to evaluate how well the compositional process works; second, we also are trying to determine how useful the final representation—the output of the composition—is, relative to our needs.
The scope of this second problem covers most phrase and sentence-level semantic models, from bag-of-words approaches in information retrieval to logic-based formal semantic models, via language models used for machine translation. It is heavily task dependent, in that a representation that is suitable for machine translation may not be appropriate for textual inference tasks, and one that is appropriate for IR may not be ideal for paraphrase detection. Therefore this aspect of semantic model evaluation ideally should take the form of application-oriented testing. For instance, to test semantic representations designed for machine translation purposes, we should use a machine translation evaluation task.
The DisCoCat framework (Clark, Coecke, and Sadrzadeh 2008; Coecke, Sadrzadeh, and Clark 2010), described in Section 3, allows for the composition of any words of any syntactic type. The general learning algorithm presented in Section 4.3 technically can be applied to learn and model relations of any semantic type. However, many open questions remain, such as how to deal with logical words, determiners, and quantification, and how to reconcile the different semantic types used for sentences with transitive and intransitive sentences. We will leave these questions for future work, briefly discussed in Section 6. In the meantime, we concretely are left with a way of satisfactorily modeling only simple sentences without having to answer these bigger questions.
With this in mind, in this section we present a series of experiments centered around evaluating how well various models of semantic vector composition perform (along with the one described in Section 4) in a phrase similarity comparison task. This task aims to test the quality of a compositional process by determining how well it forms a clear joint meaning for potentially ambiguous words. The intuition here is that tokens, on their own, can have several meanings; and that it is through the compositional process—through giving them context—that we understand their specific meaning. For example, bank itself could (among other meanings) mean a river bank or a financial bank; yet in the context of a sentence such as The bank refunded the deposit, it is likely we are talking about the financial institution.
In this section, we present three data sets designed to evaluate how well word-sense disambiguation occurs as a byproduct of composition. We begin by describing the first data set, based around noun-intransitive verb phrases, in Section 5.1. In Section 5.2, we present a data set based around short transitive-verb phrases (a transitive verb with subject and object). In Section 5.3, we discuss a new data set, based around short transitive-verb phrases where the subject and object are qualified by adjectives. We leave discussion of these results for Section 6. This first experiment, originally presented in Mitchell and Lapata (2008), evaluates the degree to which an ambiguous intransitive verb (e.g., draws) is disambiguated by combination with its subject.
Data set Description. The data set4 comprises 120 pairs of intransitive sentences, each of the form NOUN VERB. These sentence pairs are generated according to the following
4 Available at http://homepages.inf.ed.ac.uk/s0453356/results.
procedure, which will be the basis for the construction of the other data sets discussed subsequently:
1. A number of ambiguous intransitive verbs (15, in this case) are selected from frequently occurring verbs in the corpus.
2. For each verb V, two mutually exclusive synonyms V1 and V2 of the verb are produced, and each is paired with the original verb separately (for a total of 30 verb pairs). These are generated by taking maximally distant pairs of synonyms of the verb on WordNet, but any method could be used here.
3. For each pair of verb pairs (V, V1) and (V, V2), two frequently occurring nouns N1 and N2 are picked from the corpus, one for each synonym of V. For example, if V is glow and the synonyms V1 and V2 are beam and burn, we might choose face as N1, because a face glowing and a face beaming mean roughly the same thing; and fire as N2, because a fire glowing and a fire burning mean roughly the same thing.
4. By combining the nouns with the verb pairs, we form two high similarity triplets (V, V1, N1) and (V, V2, N2), and two low similarity triplets (V, V1, N2) and (V, V2, N1).
The last two steps can be repeated to form more than four triplets per pair of verb pairs. In Mitchell and Lapata (2008), eight triplets are generated for each pair of verb pairs, obtaining a total of 120 triplets from the 15 original verbs. Each triplet, along with its HIGH or LOW classification (based on the choice of noun for the verb pair) is an entry in the data set, and can be read as a pair of sentences: (V, Vi, N) translates into the intransitive sentences N V and N Vi.
Finally, the data set is presented, without the HIGH/LOW ratings, to human annotators. These annotators are asked to rate the similarity of meaning of the pairs of sentences in each entry on a scale of 1 (low similarity) to 7 (high similarity). The final form of the data set is a set of lines each containing:
Sentence 1 Sentence 2
butler bow butler submit head bow head stoop company bow company submit government bow government stoop
Evaluation Methodology. This data set is used to compare various compositional distributional semantic models, according to the following procedure:
1. For each entry in the data set, the representation of the two sentences N V and N Vi formed from the entry triple (V, Vi, N), which we will name S1 and S2, is constructed by the model.
2. The similarity of these sentences according to the model’s semantic distance measure constitutes the model score for the entry.
3. The rank correlation of entry model scores against entry annotator scores is calculated using Spearman’s rank correlation coefficient ρ.
The Spearman ρ scores are values between −1 (perfect negative correlation) and 1 (perfect correlation). The higher the ρ score, the higher the compositional model can be said to produce sentence representations that match human understanding of sentence meaning, when it comes to comparing the meaning of sentences. As such, we will rank the models evaluated using the task by decreasing order of ρ score.
One of the principal appealing features of Spearman’s ρ is that the coefficient is rank-based: It does not require models’ semantic similarity metrics to be normalized for a comparison to be made. One consequence is that a model providing excellent rank correlation with human scores, but producing model scores on a small scale (e.g., values between 0.5 and 0.6), will obtain a higher ρ score than a model producing model scores on a larger scale (e.g., between 0 and 1) but with less perfect rank correlation. If we wished to then use the former model in a task requiring some greater degree of numerical separation (let us say 0 for non-similar sentences and 1 for completely similar sentences), we could simply renormalize the model scores to fit the scale. By eschewing score normalization as an evaluation factor, we minimize the risk of erroneously ranking one model over another.
Finally, in addition to computing the rank alignment coefficient between model scores and annotator scores, Mitchell and Lapata (2008) calculate the mean model scores for entries labeled HIGH, and for entries labeled LOW. This information is reported in their paper as additional means for model comparison. However, for the same reason we considered Spearman’s ρ to be a fair means of model comparison— namely, in that it required no model score normalization procedure and thus was less likely to introduce error by adding such a degree of freedom—we consider the HIGH/LOW means to be inadequate grounds for comparison, precisely because it requires normalized model scores for comparison to be meaningful. As such, we will not include these mean scores in the presentation of this, or any further experiments in this article.
Models Compared. In this experiment, we compare the best and worst performing models of Mitchell and Lapata (2008) to our own. We begin by building a vector space W for all words in the corpus, using standard distributional semantic model construction procedures (n-word window) with the parameters of Mitchell and Lapata (2008). These parameters are as follows: The basis elements of this vector space are the 2,000 most frequently occurring words in the BNC, excluding common stop words. For our evaluation, the corpus was lemmatized using Carroll’s Morpha (Minnen, Carroll, and Pearce 2001), which was applied as a byproduct of our parsing of the BNC with C&C Tools (Curran, Clark, and Bos 2007).
The context of each occurrence of a word in the corpus was defined to be five words on either side. After the vector for each word w was produced, the basis weights cwi associated with context word bi were normalized by the following ratio of probabilities weighting scheme:
cwi = P(bi|w) P(w)
Let −−→ verb and −−→noun be the lexical semantic vectors for the verb and the noun, from W. It should be evident that there is no significant difference in using W for nouns in lieu of building a separate vector space N strictly for noun vectors.
As a first baseline, Verb Baseline, we ignored the information provided by the noun in constructing the sentence representation, effectively comparing the semantic content of the verbs:
Verb Baseline : −−−−−−→ noun verb = −−→ verb
The models from Mitchell and Lapata (2008) we evaluate here are those which were strictly unsupervised (i.e., no free parameters for composition). These are the additive model Add, wherein
Add : −−−−−−→ noun verb = −−→noun +−−→verb
and the multiplicative model Multiply, wherein
Multiply : −−−−−−→ noun verb = −−→noun −−→verb
Other models with parameters that must be optimized against a held-out section of the data set are presented in Mitchell and Lapata (2008). We omitted them here principally because they do not perform as well as Multiply, but also because the need to optimize the free parameters for this and other data sets makes fair comparison with completely unsupervised models more difficult, and less fair.
We also evaluate the Categorical model from Section 4 here, wherein
Categorical : −−−−−−→ noun verb = −−−→ verbcat −−→noun
where −−−→ verbcat is the compact representation of the relation the verb stands for, computed according to the procedure described in Section 4.3. It should be noted that for the case of intransitive verbs, this composition operation is mathematically equivalent to that of the Multiply model (as component-wise multiplication is a commutative operation), the difference being the learning procedure for the verb vector.
For all such vector-based models, the similarity of the two sentences s1 and s2 is taken to be the cosine similarity of their vectors, defined in Section 2.2:
similarity(s1, s2) = cosine( −→s1 ,−→s2 )
In addition to these vector-based methods, we define an additional baseline and an upper-bound. The additional baseline, Bigram Baseline, is a bigram-based language
model trained on the BNC with SRILM (Stolcke 2002), using the standard language model settings for computing log-probabilities of bigrams. To determine the semantic similarity of sentences s1 = noun verb1 and s2 = noun verb2, we assumed sentences have mutually conditionally independent properties, and computed the joint probability:
similarity(s1, s2) = logP(s1 ∧ s2) = log(P(s1)P(s2)) = logP(s1) + logP(s2)
The reasoning behind this baseline is as follows. The sentence formed by combining the first verb with its arguments is, by the design of this data set, a semantically coherent sentence (e.g., the head bowed and the government bowed both make sense). We therefore expect language models to treat this sort of bigram as having a higher probability than bigrams that are not semantically coherent sentences (and therefore unlikely to be observed in the corpus). A similar (relative) high probability is to be expected when the sentence formed by taking the second verb in each entry and combining it with the verb yields a sentence similar to the first in meaning (e.g., as would be the case with the head stooped), whereas the probability of a semantically incoherent sentence (e.g., the government stooped) is expected to be low relative to that of the first sentence. By taking the sum of log probabilities of sentences, we compute the log of the product of the probabilities, which we expect to be low when the probability of the second sentence is low, and high when it is high, with the probability of the first sentence acting as a normalizing factor. To summarize, while this bigram-based measure only tracks a tangential aspect of semantic similarity, it can be one which plays an artificially important role in experiments with a predetermined structure such as the one described in this section. For this reason, we use this bigram baseline for this experiment and all that follow.
The upper bound of the data set, UpperBound, was taken to be the inter-annotator agreement: the average of how each annotator’s score aligns with other annotator scores, using Spearman’s ρ.
Results. The results5 of the first experiment are shown in Table 4. As expected from the fact that Multiply and Categorical differ only in how the verb vector is learned, the results of these two models are virtually identical, outperforming both baselines and the Additive model by a significant margin. However, the distance from these models to the upper bound is even greater, demonstrating that there is still a lot of progress to be made. The first experiment was followed by a second similar experiment, in Mitchell and Lapata (2010), covering different sorts of composition operations for binary combinations of syntactic types (adjective-noun, noun-noun, verb-object). Such further experiments are interesting, but rather than continue down this binary road, we now turn to the development of our second experiment, involving sentences with larger syntactic structures, to examine how well various compositional models cope with more complex syntactic and semantic relations.
5 The results were state of the art when the experiments were run in 2011. The binary composition data set used here are popular; now there are a host of new state-of-the-art systems available in the literature. We invite the reader to check references to Mitchell and Lapata (2008) to find the current state of the art.
This second experiment, which we initially presented in Grefenstette and Sadrzadeh (2011a), is an extension of the first in the case of sentences centered around transitive verbs, composed with a subject and an object. The results of the first experiment did not demonstrate any difference between the multiplicative model, which takes into account no syntactic information or word ordering, and our syntactically motivated categorical compositional model. By running the same experiment over a new data set, where the relations expressed by the verb have a higher arity than in the first, we hope to demonstrate that added structure leads to better results for our syntax-sensitive model.
Data Set Description. The construction procedure for this data set6 is almost exactly as for the first data set, with the following differences:
Verbs are transitive instead of intransitive. We arbitrarily took 10 verbs from the most frequent verbs in the BNC, and for each verb, took two maximally distant synonyms (again, using WordNet) to obtain 10 pairs of pairs.
For each pair of verb pairs, we selected a set of subject and object nouns to use as context, as opposed to just a subject noun.
Each subject-object pair was manually chosen so that one of the verb pairs would have high similarity in the context of that subject and object, and the other would have low similarity. We used these choices to annotate the entries with HIGH and LOW tags.
Each combination of a verb pair with a subject-object pair constitutes an entry of our data set, of which there are 200.
As a form of quality control, we inserted “gold standard” sentences in the form of identical sentence pairs and rejected annotators who did not score these gold standard
6 The data set, reannotated by Turkers in 2013, is available at http://www.cs.ox.ac.uk/activities/compdistmeaning/GS2013data.txt.
sentences with a high score of 6 or 7.7 Lemmatized sentences from sample entries of this data set and our HIGH-LOW tags for them are shown in Table 5.
The data set was passed to a group of 50 annotators on Amazon Mechanical Turk, as for the previous data set. The annotators were shown, for each entry, a pair of sentences created by adding the in front of the subject and object nouns and putting the verbs in the past tense,8 and were instructed to score each pair of sentences on the same scale of 1 (not similar in meaning) to 7 (similar in meaning), based on how similar in meaning they believed the sentence pair was.
Evaluation Methodology. The methodology for this experiment is exactly that of the previous experiment. Models compositionally construct sentence representations, and compare them using a distance metric (all vector-based models once again used cosine similarity). The rank correlation of models scores with annotator scores is calculated using Spearman’s ρ, which is in turn used to rank models.
Models Compared. The models compared in this experiment are those of the first experiment, with the addition of an extra trigram-based baseline (trained with SRILM, using the addition of log-probability of the sentence as a similarity metric), and a variation on our categorical model, presented subsequently. With W as the distributional semantic space for all words in the corpus, trained using the same parameters as in the first experiment, and −−→ subj, −−→ verb, −−−→ object ∈ W as the vectors for subject, verb, and object of a sentence, respectively, and with verbcat as the compact representation of a transitive verb learned using the algorithm presented in this paper, we have the following compositional methods:
Add : −−−−−−−−−−−−→ subject verb object = −−−−→ subject + −−→ verb + −−−→ object Multiply : −−−−−−−−−−−−→ subject verb object = −−−−→ subject −−→verb −−−→object Categorical : −−−−−−−−−−−−→ subject verb object = −−−→ verbcat (−−−−→ subject ⊗−−−→object
) Kronecker : −−−−−−−−−−−−→ subject verb object = (−−→ verb ⊗−−→verb ) (−−−−→ subject ⊗−−−→object )
7 During the 2013 reannotation of this data set, we rejected no Turker contributions, as the answers to the gold standard sentence pairs were aligned with our expectations. We attribute this to our adding the requirement that Turkers be based in the US or the UK and have English as the first language. 8 For example, the entry draw table eye depict would yield the sentences The table drew the eye and The table depicted the eye.
All of these models have been explained earlier, with the exception of Kronecker. We first presented this new addition in Grefenstette and Sadrzadeh (2011b), where we observed that the compact representation of a verb in the DisCoCat framework, under the assumptions presented in Section 4, can be viewed as dim(N) × dim(N) matrices in N ⊗ N. We considered alternatives to the algorithm presented earlier for the construction of such matrices, and were surprised by the results of the Kronecker method, wherein we replaced the matrix learned by our algorithm with the Kronecker product of the lexical semantic vectors for the verb. Further analysis performed since the publication of that paper can help to understand why this method might work. Using the following property, for any vectors −→a ,−→b ,−→c ,−→d in a vector space
(−→a ⊗−→b ) (−→c ⊗−→d ) = (−→a −→c ) ⊗ (−→b −→d )
we can see that the Kronecker model’s composition operation can be expressed as
Kronecker : −−−−−−−−−−−−→ subject verb object = (−−→ verb −−−−→subject ) ⊗ (−−→ verb −−−→object )
Bearing in mind that the cosine measure we are using as a similarity metric is equivalent to the inner product of two vectors normalized by the product of their length
cosine(−→a ,−→b ) = 〈 −→a |−→b 〉
‖−→a ‖ × ‖−→b ‖
and the following property of the inner product of kronecker products
〈−→a ⊗−→b |−→c ⊗−→d 〉 = 〈−→a |−→c 〉 × 〈−→b |−→d 〉
we finally observe that comparing two sentences under Kronecker corresponds to the following computation:
cosine( −−−−−−−−−−−−→ subject verb1 object, −−−−−−−−−−−−→ subject verb2 object)
= α 〈(−−→ verb1 ⊗ −−→ verb1 ) (−−−−→ subject ⊗−−−→object ) | (−−→ verb2 ⊗ −−→ verb2 ) (−−−−→ subject ⊗−−−→object )〉
= α 〈(−−→ verb1 −−−−→ subject ) ⊗ (−−→ verb1 −−−→ object ) | (−−→ verb2 −−−−→ subject ) ⊗ (−−→ verb2 −−−→ object )〉 = α 〈(−−→ verb1 −−−−→ subject ) | (−−→ verb2 −−−−→ subject )〉〈(−−→ verb1 −−−→ object ) | (−−→ verb2 −−−→ object
)〉 where α is the normalization factor
α = 1 ‖−−−−−−−−−−−−→subject verb1 object‖ × ‖ −−−−−−−−−−−−→ subject verb2 object‖
We note here that the Kronecker is effectively a parallel application of the Multiply model, combining the subject and verb, and object and verb separately. Within the context of this task, the comparison of two sentences boils down to how well each verb combines with the subject multiplied by how well it combines with the object, with the
joint product forming the overall model score. In short, it more or less constitutes the introduction of some mild syntactic sensitivity into the multiplicative model of Mitchell and Lapata (2008), although it remains to be seen how well this scales with syntactic complexity (e.g., extended to ditransitive verbs, or other relations of higher arity).
The UpperBound here is, again, the inter-annotator agreement, calculated by computing the pairwise alignment of each annotator with every other, and averaging the resulting rank correlation coefficients.
Results. The results of the second experiment are shown in Table 6. The baseline scores fall in the 0.14–0.16 range, with best results being obtained for the Bigram Baseline and Trigram Baseline, the difference between both models not being statistically significant. The additive model Add performs on par with the first experiment. The multiplicative model Multiply and our Categorical model perform on par with the version used for the intransitive experiment, but obtain a score comparable to the baselines. The best performing model here is our newly introduced Kronecker model, which leads the pack by a steady margin, with a score of 0.26. The inter-annotator agreement UpperBound is much higher in this experiment than in the previous experiment, indicating even more room for improvement.
We therefore learn here that the Categorical model continues to operate on par with the multiplicative model, and that the Kronecker model provides the highest results, while being as simple in its construction as the multiplicative model, requiring only the learning of lexical vectors for the verb. The third and final experiment we present is a modified version of the second data set presented earlier, where the nouns in each entry are under the scope of adjectives applied to them.The intuition behind the data sets presented in Section 5.1 and Section 5.2 was that ambiguous verbs are disambiguated through composition with nouns. These nouns themselves may also be ambiguous, and a good compositional model will be capable of separating the noise produced by other meanings through its compositional mechanism to produce unambiguous phrase representations. The intuition behind this data set is similar, in that adjectives provide both additional information for disambiguation of the nouns they apply to, but also additional semantic noise. Therefore, a
good model will also be able to separate the useful information of the adjective from its semantic noise when composing it with its argument, in addition to doing this when composing the noun phrases with the verb.
Data Set Description. The construction procedure for this data set was to take the data set from Section 5.2, and, for each entry, add a pair of adjectives from those most frequently occurring in the corpus. The first adjective from the pair is applied to the first noun (subject) of the entry when forming the sentences, and the second adjective is applied to the second noun (object). For each entry, we chose adjectives which best preserved the meaning of the phrase constructed by combining the first verb with its subject and object.
This new data set9 was then annotated again by a group of 50 annotators using Amazon’s Mechanical Turk service. The annotators were shown, for each entry, a pair of sentences created by adding the in front of the subject and object noun phrases and putting the verbs in the past tense,10 and asked to give each sentence pair a meaning similarity score between 1 and 7, as for the previous data sets. We applied the same quality control mechanism as in the second experiment. Some 94 users returned annotations, of which we kept 50 according to our gold standard tests. We are unaware of whether or not it was applied in the production of the first data set, but believe that this can only lead to the production of higher quality annotations.
Sample sentences from this data set are shown in Table 7.
Evaluation Methodology. The evaluation methodology in this experiment is identical to that of the previous experiments.
Models Compared. In this experiment, in lieu of simply comparing compositional models “across the board” (e.g., using the multiplicative model for both adjective-noun composition and verb-argument composition), we experimented with different combinations of models. This evaluation procedure was chosen because we believe that adjective-noun composition need not necessarily be the same kind of compositional process as subjectverb-object composition, and also because different models may latch onto different semantic features during the compositional process, and it would be interesting to see what model mixtures work well together. Naturally, this is not a viable approach to selecting composition operations in general, as we will not have the luxury of trying every combination of composition operations for every combination of syntactic types, but it is worthwhile performing these tests to at least verify the hypothesis that
9 Available at http://www.cs.ox.ac.uk/activities/compdistmeaning/GS2012data.txt. 10 For example, the entry statistical table show express good result would yield the sentences The statistical table
showed the good result and The statistical table expressed the good result.
operation-specific composition operations (or parameters) is a good thing. Notably, this idea has been explored very recently, within the context of deep learning networks, by Chen et al. (2013).
Each mixed model has two components: a verb-argument composition model and a adjective-noun composition model. For verb-argument composition, we used the three best models from the previous experiment, namely, Multiply, Categorical, and Kronecker. For adjective composition we used three different methods of adjectivenoun composition. With −−−−−→ adjective and −−→noun being the vectors for an adjective and a noun in our distributional lexical semantic space W (built using the same procedure as the previous experiments) and −−→ adjcat being the compact representation in the Categorical model, built according to the algorithm from Section 4.3, we have the following models:
AdjMult : −−−−−−−−−→ adjective noun = −−−−−→ adjective −−→noun Categorical : −−−−−−−−−→ adjective noun = −−−−−−→ adjectivecat −−→noun
The third model, AdjNoun, is a holistic (non-compositional) model, wherein the adjective-noun compound was treated as a single token, as its semantic vector−−−−−−−−−−−−→ (adjective noun)lex ∈ W was learned from the corpus using the same learning procedure applied to construct other vectors in W. Hence, the model defines adjective-noun “composition” as:
AdjNoun : −−−−−−−−−→ adjective noun = −−−−−−−−−−−−→ (adjective noun)lex
In addition to these models, we also evaluated three baselines: Verb Baseline, Bigram Baseline, and Trigram Baseline. As in previous experiments, the verb baseline uses the verb vector as sentence vector, ignoring the information provided by other words. The bigram and trigram baselines are calculated from the same language model as used in the second experiment. In both cases, the log-probability of each sentence is calculated using SRLIM, and the sum of log-probabilities of two sentences is used as a similarity measure.
Finally, for comparison, we also considered the full additive model:
Additive : −−−−−→ sentence = −−−−−→ adjective1 + −−−→noun1 + −−→ verb + −−−−−→ adjective2 + −−−→noun2
Results. The results for the third experiment are shown in Table 8. The best performing adjective-noun combination operations for each verb-argument combination operation are shown in bold. Going through the combined models, we notice that in most cases the results stay the same whether the adjective-noun combination method is AdjMult or CategoricalAdj. This is because, as was shown in the first experiment, composition of a unary-relation such as an adjective or intransitive verb with its sole argument, under the categorical model with reduced representations, is mathematically equivalent to the multiplicative model. The sole difference is the way the adjective or intransitive verb vector is constructed. We note, however, that with Categorical as a verb-argument composition method, the CategoricalAdj outperforms AdjMult by a non-negligible margin (0.19 vs. 0.14), indicating that the difference in learning procedure can lead to different results depending on what other models it is combined with.
Overall, the best results are obtained for AdjMult+Kronecker (ρ = 0.26) and CategoricalAdj+Kronecker (ρ = 0.27). Combinations of the adjective composition methods with other composition methods at best matches the best-performing baseline, Verb Baseline. In all cases, the holistic model AdjNoun provides the worst results.","6 discussion :In this article, we presented various approaches to compositionality in distributional semantics. We discussed what mathematical operations could underlie vector combination, and several different ways of including syntactic information into the combinatory process. We reviewed an existing compositional framework that leverages the ability to communicate information across mathematical structures provided by category theory in order to define a general way of linking syntactic structure to syntax-sensitive composition operations. We presented concrete ways to apply this framework to linguistic tasks and evaluate it, and developed algorithms to construct semantic representations for words and relations within this framework. In this section, we first briefly comment upon the combined results of all three experiments, and then conclude by discussing what aspects of compositionality require further attention, and how experiment design should adapt towards this goal.
Results Commentary. We evaluated this framework against other unsupervised compositional distributional models, using non-compositional models and n-gram language models as baselines, within the context of three experiments. These experiments show that the concrete categorical model developed here, and the Kronecker-based variant
presented alongside it, outperform all other models in each experiment save the first, where they perform on par with what was the leading model at the time this experiment was performed (namely, 2011). As the experiments involved progressively more syntactically complicated sentences, the increased reliability of our categorical approaches relative to competing models as sentence complexity rises seems to indicate that both the categorical and Kronecker model successfully leverage the added information provided by additional terms and syntactic structures.
The third experiment also served to show that using different combinations of composition operations, depending on the syntactic type of the terms being combined, can yield better results, and that some models combine better than others. Notably, the adjective-noun combination models AdjMult and CategoricalAdj, despite their mathematical similarity, produce noticeably different results when combined with the categorical verb-argument composition operation, while they perform equally with most other verb-argument composition operations. We can conclude that different models combine different semantic aspects more prominently than others, and that through combination we can find better performance by assuming that different kinds of composition play on different semantic properties. For example, predicates such as intersective adjectives add information to their argument (a red ball is a ball that is also red). This raises the question of how to design models of composition that systematically select which operations will match the semantic aspects of the words being combined based on their syntactic type. This is an open question, which we believe warrants further investigation.
Overall, we observe two advantages of these models over those presented in the early part of this article, and those evaluated in the later part. First, we have shown that a linguistically motivated and mathematically sound framework can be implemented and learned. Second, we showed that simple methods such as the Kronecker model perform well, especially when combined with the multiplicative model (effectively a unary instance of the Kronecker model) for sentences with adjectives. Considering the “simple is best” position recently argued for experimentally by Blacoe and Lapata (2012), this is an interesting candidate for dealing with binary relation composition.
Future Work. These experiments showcased the ability of various compositional models to produce sentence representations that were less ambiguous than the words that formed them. They also more generally demonstrated that concrete models could be built from the general categorical framework and perform adequately in simple paraphrase detection tasks. However, various aspects of compositionality were not evaluated here. First, syntax sensitivity is not as important in these experiments as it might be in “real” language. For instance, whereas models that treat adjectives, nouns, and verbs differently tended to perform better than those that did not, the actual capacity of a model to use the syntactic structure was not tested. For instance, models that ignore word order and syntax altogether were not particularly penalized, as might be done if some sentences were simply the reverse of another sentence they are paired with (where syntax insensitive models would erroneously give such sentence pairs a high similarity score). One way of testing word order while expanding the data set was suggested by Turney (2012), who for every entry in a phrase similarity test such as Mitchell and Lapata’s, discussed herein, creates a new entry where one of the sentences has the order of its words reversed, and is assigned an artificial annotator score of 1 (the minimum). This is an interesting approach, but we would prefer to see such reversals designed into the data set and seen by annotators, for instance,
in the recent data set presented in Pham et al. (2013). This data set has entries such as guitar played man and man played guitar, where the subjects and objects of verbs are indeed reversed but the two sentences do not necessarily express opposite meanings; we will be looking into expanding such data sets to ones where both sentences make sense and have opposite meanings, such as in the pair man bites dog and dog bites man.
Furthermore, sentences all have the exact same sentence structure, and therefore words are aligned. Models that might indirectly benefit from this alignment, such as Kronecker, may have been given an advantage due to this, as opposed to models such as Categorical, which are designed to compare sentences of different length or syntactic structure, when resolving the difference between intransitive and transitive sentence spaces as has been done in Grefenstette et al. (2011). Future experiments should aim to do away with this automated alignment, and include sentence comparisons that would penalize models which do not leverage syntactic information.
Furthermore, each of these experiments dealt with the comparison of a certain type of sentence (transitive, intransitive). This was convenient for our concrete categorical model, as we defined the sentence space differently based on the valency of the head verb. However, sentences with different sorts of verbs should be able to be directly compared. Not only do several models, both non-syntax sensitive (additive, multiplicative) and syntax-sensitive (Baroni and Zamparelli 2010; Socher et al. 2012), not face this problem, as the product of composition is either naturally in the same space as the vectors being composed or is projected back into it, but the categorical framework the concrete categorical models were derived from does not commit us to different sentence spaces either. The topic of how to solve this problem for the concrete models developed here is beyond the scope of this article, but various options exist, such as the projection of ST into SI, the embedding of SI into ST, the use of a combined sentence space S = SI ⊕ ST, and so on. It is clear that future experiments should not give this degree of convenience to the models being evaluated (and their designers), by comparing sentences of different types, lengths, and structure so as to more fairly evaluate the capacity of models to produce meaningful semantic representations in a single space, which can be directly compared regardless of syntactic structure and verb-type.
Finally, we mentioned at the beginning of Section 5 that evaluations need to be application-oriented. The class of experiments presented in this article specifically see how well disambiguation occurs as a byproduct of composition. We presented concrete models that would offer ways of combining relations underlying verbs and adjectives, and tested how well these relations disambiguated their arguments and were in turn disambiguated by their arguments. We did not address other aspects of language, such as quantification, logical operations, relative clauses, intensional clauses, embedded sentences, and many other linguistic aspects of spoken and written language that have complex syntactic structure and complicated semantic roles. Addressing these sorts of problems requires determining how negation, conjunction, and other logical relations should be modeled within compositional distributional formalisms, a problem we share with other similar approaches. The development of newer models within the categorical framework we based our work on, and the further development of other approaches mentioned here, must be driven by new experimental goals different from those offered by the task and experiments discussed in this article. Thinking not only about how to address these aspects of language within compositional models, but how to evaluate them, should be a priority for all those interested in further developing this field.",,,"Modeling compositional meaning for sentences using empirical distributional methods has been a challenge for computational linguists. The categorical model of Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) provides a solution by unifying a categorial grammar and a distributional model of meaning. It takes into account syntactic relations during semantic vector composition operations. But the setting is abstract: It has not been evaluated on empirical data and applied to any language tasks. We generate concrete models for this setting by developing algorithms to construct tensors and linear maps and instantiate the abstract parameters using empirical data. We then evaluate our concrete models against several experiments, both existing and new, based on measuring how well models align with human judgments in a paraphrase detection task. Our results show the implementation of this general abstract framework to perform on par with or outperform other leading models in these experiments.1","[{""affiliations"": [], ""name"": ""Edward Grefenstette""}, {""affiliations"": [], ""name"": ""Google DeepMind""}, {""affiliations"": [], ""name"": ""Mehrnoosh Sadrzadeh""}]",SP:2c6b675504d6903e070d034b855e39f296f5ccd4,"[{""authors"": [""S. Abramsky"", ""B. Coecke.""], ""title"": ""A categorical semantics of quantum protocols"", ""venue"": ""Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science, pages 415\u2013425, Turku, Finland."", ""year"": 2004}, {""authors"": [""H. Alshawi""], ""title"": ""The Core Language Engine"", ""year"": 1992}, {""authors"": [""M. Baroni"", ""R. Zamparelli.""], ""title"": ""Nouns are vectors, adjectives are matrices: Representing adjective-noun constructions in semantic space"", ""venue"": ""Proceedings of the 2010 Conference on Empirical Methods in Natural"", ""year"": 2010}, {""authors"": [""D. B\u00e9chet"", ""A. Foret"", ""I. Tellier.""], ""title"": ""Learnability of Pregroup Grammars"", ""venue"": ""Studia Logica, 87(2\u20133), pages 225\u2013252."", ""year"": 2007}, {""authors"": [""W. Blacoe"", ""M. Lapata.""], ""title"": ""A comparison of vector-based representations for semantic composition"", ""venue"": ""Proceedings of the 2012 Conference on Empirical Methods in Natural Language"", ""year"": 2012}, {""authors"": [""W. Buszkowski""], ""title"": ""Lambek grammars based on pregroups"", ""venue"": ""P. de Groote, G. Morrill, and C. Retor\u00e9, editors, Logical Aspects of Computational Linguistics. Lecture Notes in Computer Science,"", ""year"": 2001}, {""authors"": [""Chen"", ""Danqi"", ""Richard Socher"", ""Christopher D. Manning"", ""Andrew Y. Ng.""], ""title"": ""Learning new facts from knowledge bases with neural tensor networks and semantic word vectors"", ""venue"": ""arXiv preprint"", ""year"": 2013}, {""authors"": [""S. Clark""], ""title"": ""Type-driven syntax and semantics for composing meaning vectors"", ""venue"": ""Quantum Physics and Linguistics: A Compositional, Diagrammatic Discourse. Oxford University Press."", ""year"": 2013}, {""authors"": [""S. Clark"", ""B. Coecke"", ""M. Sadrzadeh.""], ""title"": ""A compositional distributional model of meaning"", ""venue"": ""Proceedings of the Second Quantum Interaction Symposium (QI-2008), Oxford."", ""year"": 2008}, {""authors"": [""S. Clark"", ""J.R. Curran.""], ""title"": ""Wide-coverage efficient statistical parsing with CCG and log-linear models"", ""venue"": ""Computational Linguistics, 33:493\u2013552."", ""year"": 2007}, {""authors"": [""S. Clark"", ""S. Pulman.""], ""title"": ""Combining symbolic and distributional models of meaning"", ""venue"": ""AAAI Spring Symposium on Quantum Interaction, Stanford, USA."", ""year"": 2007}, {""authors"": [""Clarke"", ""Daoud.""], ""title"": ""Context-theoretic semantics for natural language: An overview"", ""venue"": ""Proceedings of the Workshop on Geometrical Models of Natural Language Semantics, pages 112\u2013119, Edinburgh."", ""year"": 2009}, {""authors"": [""Clarke"", ""Daoud.""], ""title"": ""A context-theoretic framework for compositionality in distributional semantics"", ""venue"": ""Computational Linguistics, 38(1):41\u201371."", ""year"": 2012}, {""authors"": [""B. Coecke"", ""\u00c9.O. Paquette.""], ""title"": ""Categories for the practising physicist"", ""venue"": ""B. Coecke, editor, New Structures for Physics. Lecture Notes in Physics, volume 813, pages 173\u2013286, Springer,"", ""year"": 2011}, {""authors"": [""B. Coecke"", ""M. Sadrzadeh"", ""S. Clark.""], ""title"": ""Mathematical Foundations for a Compositional Distributional Model of Meaning"", ""venue"": ""Linguistic Analysis, 36:345\u2013384."", ""year"": 2010}, {""authors"": [""Curran"", ""James"", ""Stephen Clark"", ""Johan Bos.""], ""title"": ""Linguistically motivated large-scale NLP with C&C and boxer"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational"", ""year"": 2007}, {""authors"": [""J.R. Curran""], ""title"": ""From distributional to semantic similarity"", ""venue"": ""Ph.D. thesis, School of Informatics, University of Edinburgh."", ""year"": 2004}, {""authors"": [""Dean"", ""Jeffrey"", ""Sanjay Ghemawat.""], ""title"": ""MapReduce: Simplified data processing on large clusters"", ""venue"": ""Communications of the ACM, 51(1):107\u2013113."", ""year"": 2008}, {""authors"": [""Erk"", ""Katrin"", ""Sebastian Pad\u00f3.""], ""title"": ""A structured vector space model for word meaning in context"", ""venue"": ""Proceedings of the Conference on Empirical Methods in Natural Language Processing - EMNLP \u201908,"", ""year"": 2008}, {""authors"": [""J.R. Firth""], ""title"": ""A synopsis of linguistic theory 1930-1955"", ""venue"": ""Studies in linguistic analysis."", ""year"": 1957}, {""authors"": [""T.A.D. Fowler"", ""G. Penn.""], ""title"": ""Accurate context-free parsing with combinatory categorial grammar"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 335\u2013344,"", ""year"": 2010}, {""authors"": [""G. Frege""], ""title"": ""Uber sinn und bedeutung"", ""venue"": ""Zeitschrift f\u00fcr Philosophie und philosophische Kritik, 100(1):25\u201350."", ""year"": 1892}, {""authors"": [""Genkin"", ""Daniel"", ""Nissim Francez"", ""Michael Kaminski.""], ""title"": ""Mildly context-sensitive languages via buffer augmented pregroup grammars"", ""venue"": ""Z. Manna and D. A. Peled, editors, Time for Verification, pages 144\u2013166,"", ""year"": 2010}, {""authors"": [""Girard"", ""Jean-Yves.""], ""title"": ""Linear logic"", ""venue"": ""Theoretical Computer Science, 50:1\u2013102."", ""year"": 1987}, {""authors"": [""E. Grefenstette""], ""title"": ""Analysing document similarity measures"", ""venue"": ""Master\u2019s thesis, University of Oxford."", ""year"": 2009}, {""authors"": [""E. Grefenstette"", ""G. Dinu"", ""Y. Zhang"", ""M. Sadrzadeh"", ""M. Baroni.""], ""title"": ""Multistep regression learning for compositional distributional semantics"", ""venue"": ""Proceedings of the Tenth International Conference on"", ""year"": 2013}, {""authors"": [""E. Grefenstette"", ""M. Sadrzadeh.""], ""title"": ""Experimental support for a categorical compositional distributional model of meaning"", ""venue"": ""Proceedings of the 2011 Conference on Empirical Methods"", ""year"": 2011}, {""authors"": [""E. Grefenstette"", ""M. Sadrzadeh.""], ""title"": ""Experimenting with transitive verbs in a DisCoCat"", ""venue"": ""Proceedings of the 2011 EMNLP Workshop on Geometric Models of Natural Language Semantics, pages 62\u201366,"", ""year"": 2011}, {""authors"": [""E. Grefenstette"", ""M. Sadrzadeh"", ""S. Clark"", ""B. Coecke"", ""S. Pulman.""], ""title"": ""Concrete sentence spaces for compositional distributional models of meaning"", ""venue"": ""Proceedings of the Ninth International"", ""year"": 2011}, {""authors"": [""G. Grefenstette""], ""title"": ""Use of syntactic context to produce term association lists for text retrieval"", ""venue"": ""Proceedings of the 15th Annual International ACM SIGIR Conference on Research and Development in Information"", ""year"": 1992}, {""authors"": [""G. Grefenstette""], ""title"": ""Explorations in automatic thesaurus discovery"", ""year"": 1994}, {""authors"": [""E. Guevara""], ""title"": ""A regression model of adjective-noun compositionality in distributional semantics"", ""venue"": ""Proceedings of the 2010 Workshop on Geometrical Models of Natural Language Semantics,"", ""year"": 2010}, {""authors"": [""Z.S. Harris""], ""title"": ""Mathematical structures of language"", ""venue"": ""Wiley."", ""year"": 1968}, {""authors"": [""Hockenmaier"", ""Julia.""], ""title"": ""Data and Models for Statistical Parsing with Combinatory Categorial Grammar"", ""venue"": ""Ph.D. thesis, School of Informatics, University of Edinburgh."", ""year"": 2003}, {""authors"": [""Hockenmaier"", ""Julia"", ""Mark Steedman.""], ""title"": ""CCGBank: A corpus of CCG derivations and dependency structures extracted from the Penn treebank"", ""venue"": ""Computational Linguistics, 33(3):355\u2013396."", ""year"": 2007}, {""authors"": [""A.K. Joshi"", ""K. Vijay-Shanker"", ""D.J. Weir.""], ""title"": ""The convergence of mildly context-sensitive grammar formalisms"", ""venue"": ""Working paper, University of Pennsylvania, School of Engineering"", ""year"": 1989}, {""authors"": [""J. Lambek""], ""title"": ""The mathematics of sentence structure"", ""venue"": ""The American Mathematical Monthly, 65(3):154\u2013170."", ""year"": 1958}, {""authors"": [""J. Lambek""], ""title"": ""Type grammar revisited"", ""venue"": ""Logical Aspects of Computational Linguistics, pages 1\u201327."", ""year"": 1999}, {""authors"": [""J. Lambek""], ""title"": ""From word to sentence"", ""venue"": ""A computational algebraic approach to grammar. Milan, Polimetrica."", ""year"": 2008}, {""authors"": [""J. Lambek""], ""title"": ""Compact Monoidal Categories from Linguistics to Physics, pages 451\u2013469"", ""year"": 2010}, {""authors"": [""T.K. Landauer"", ""S.T. Dumais""], ""title"": ""A solution to Plato\u2019s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge"", ""year"": 1997}, {""authors"": [""S. Mac Lane""], ""title"": ""Categories for the Working Mathematician"", ""venue"": ""Springer Verlag."", ""year"": 1998}, {""authors"": [""C.D. Manning"", ""P. Raghavan"", ""H. Sch\u00fctze.""], ""title"": ""Introduction to information retrieval"", ""venue"": ""Cambridge University Press, New York, NY."", ""year"": 2011}, {""authors"": [""G. Minnen"", ""J. Carroll"", ""D. Pearce.""], ""title"": ""Applied morphological processing of English"", ""venue"": ""Natural Language Engineering, 7(03):207\u2013223."", ""year"": 2001}, {""authors"": [""J. Mitchell"", ""M. Lapata.""], ""title"": ""Vectorbased models of semantic composition"", ""venue"": ""Proceedings of ACL, volume 8, pages 236\u2013244, Columbus, OH. 117"", ""year"": 2008}, {""authors"": [""J. Mitchell"", ""M. Lapata.""], ""title"": ""Composition in Distributional Models of Semantics"", ""venue"": ""Cognitive Science 34(8):1388\u20131429."", ""year"": 2010}, {""authors"": [""R. Montague""], ""title"": ""English as a Formal Language"", ""venue"": ""R. H. Thomason, editor, Formal Semantics: The Essential Readings."", ""year"": 1974}, {""authors"": [""M. Moortgat""], ""title"": ""Categorial type logics"", ""venue"": ""H. van Ditmarsch and L. S. Moss, editors, Handbook of Logic and Language. Elsevier, pages 93\u2013177."", ""year"": 1997}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Mirella Lapata.""], ""title"": ""Dependency-based construction of semantic space models"", ""venue"": ""Computational Linguistics, 33(2):161\u2013199."", ""year"": 2007}, {""authors"": [""N. Pham"", ""R. Bernardi"", ""Y.-Z. Zhang"", ""M. Baroni.""], ""title"": ""Sentence paraphrase detection: When determiners and word order make the difference"", ""venue"": ""Proceedings of the Towards a Formal Distributional"", ""year"": 2013}, {""authors"": [""T.A. Plate""], ""title"": ""Holographic reduced representations: Convolution algebra for compositional distributed representations"", ""venue"": ""Proceedings of the 12th International Joint Conference on Artificial Intelligence,"", ""year"": 1991}, {""authors"": [""A. Preller""], ""title"": ""Polynomial pregroup grammars parse context sensitive languages"", ""venue"": ""Linguistic Analysis, 36:483\u2013516."", ""year"": 2010}, {""authors"": [""A. Preller"", ""M. Sadrzadeh.""], ""title"": ""Bell states and negative sentences in the distributed model of meaning"", ""venue"": ""P. Selinger, B. Coecke, P. Panangaden, editors, Electronic Notes in Theoretical"", ""year"": 2010}, {""authors"": [""P. Selinger""], ""title"": ""A survey of graphical languages for monoidal categories"", ""venue"": ""New Structures for Physics, pages 275\u2013337."", ""year"": 2010}, {""authors"": [""P. Smolensky""], ""title"": ""Tensor product variable binding and the representation of symbolic structures in connectionist systems"", ""venue"": ""Artificial Intelligence, 46(1-2):159\u2013216."", ""year"": 1990}, {""authors"": [""P. Smolensky"", ""G. Legendre""], ""title"": ""The Harmonic Mind: From Neural Computation"", ""year"": 2006}, {""authors"": [""R. Socher"", ""B. Huval"", ""C.D. Manning"", ""A.Y. Ng.""], ""title"": ""Semantic compositionality through recursive matrix-vector spaces"", ""venue"": ""Proceedings of the 2012 Conference on Empirical Methods in Natural Language"", ""year"": 2012}, {""authors"": [""M. Steedman""], ""title"": ""The Syntactic Process"", ""venue"": ""The MIT Press."", ""year"": 2001}, {""authors"": [""M. Steedman"", ""J. Baldridge.""], ""title"": ""Combinatory Categorial Grammar"", ""venue"": ""Wiley-Blackwell."", ""year"": 2011}, {""authors"": [""A. Stolcke""], ""title"": ""SRILM\u2014An extensible language modeling toolkit"", ""venue"": ""Seventh International Conference on Spoken Language Processing."", ""year"": 2002}, {""authors"": [""P.D. Turney"", ""P. Pantel.""], ""title"": ""From frequency to meaning: Vector space models of semantics"", ""venue"": ""Journal of Artificial Intelligence Research, 37(1):141\u2013188."", ""year"": 2010}, {""authors"": [""Turney"", ""Peter D.""], ""title"": ""Domain and function: A dual-space model of semantic relations and compositions"", ""venue"": ""Journal of Artificial Intelligence Research, 44:533\u2013585."", ""year"": 2012}, {""authors"": [""C.J. Van Rijsbergen""], ""title"": ""The Geometry of Information Retrieval"", ""venue"": ""Cambridge University Press."", ""year"": 2004}, {""authors"": [""R.F. Walters""], ""title"": ""Categories and Computer Science"", ""venue"": ""Cambridge University Press."", ""year"": 1991}, {""authors"": [""D. Widdows""], ""title"": ""Geometry and Meaning"", ""venue"": ""University of Chicago Press."", ""year"": 2005}, {""authors"": [""D. Widdows""], ""title"": ""Semantic vector products: Some initial investigations"", ""venue"": ""Proceedings of the Second Quantum Interaction Symposium (QI-2008). College Publications. CITESEER, Oxford."", ""year"": 2008}, {""authors"": [""L. Wittgenstein""], ""title"": ""Philosophical Investigations"", ""venue"": ""Blackwell."", ""year"": 1953}, {""authors"": [""F.M. Zanzotto"", ""I. Korkontzelos"", ""F. Fallucchi"", ""S. Manandhar.""], ""title"": ""Estimating linear models for compositional distributional semantics"", ""venue"": ""Proceedings of the 23rd International Conference"", ""year"": 2010}]","acknowledgments  :We thank the EPSRC for funding the research leading to this article via EPSRC grant EP/J002607/1. Furthermore, we would like to thank John Harding and the equational theorem prover “Prover 9”11 for checking the identities occurred as a result of translating CCG’s rules to the language of a pregroup grammar. We would also like to thank Stephen Clark, Stephen Pulman, Bob Coecke, Karl Moritz Hermann, Richard Socher, and Dimitri Kartsaklis for valuable discussions and comments.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :The distributional approach to the semantic modeling of natural language, inspired by the notion—presented by Firth (1957) and Harris (1968)—that the meaning of a word is tightly related to its context of use, has grown in popularity as a method of semantic representation. It draws from the frequent use of vector-based document models in information retrieval, modeling the meaning of words as vectors based on the distribution of co-occurring terms within the context of a word.
Using various vector similarity metrics as a measure of semantic similarity, these distributional semantic models (DSMs) are used for a variety of NLP tasks, from
∗ DeepMind Technologies Ltd, 5 New Street Square, London EC4A 3 TW. E-mail: etg@google.com.
∗∗ School of Electronic Engineering and Computer Science, Queen Mary University of London, Mile End Road, London E1 4NS, United Kingdom. E-mail: mehrnoosh.sadrzadeh@qmul.ac.uk.
† The work described in this article was performed while the authors were at the University of Oxford. 1 Support from EPSRC grant EP/J002607/1 is acknowledged.
Submission received: 26 September 2012; revised submission received: 31 October 2013; accepted for publication: 5 April 2014.
doi:10.1162/COLI a 00209
© 2015 Association for Computational Linguistics
automated thesaurus building (Grefenstette 1994; Curran 2004) to automated essay marking (Landauer and Dumais 1997). The broader connection to information retrieval and its applications is also discussed by Manning, Raghavan, and Schütze (2011). The success of DSMs in essentially word-based tasks such as thesaurus extraction and construction (Grefenstette 1994; Curran 2004) invites an investigation into how DSMs can be applied to NLP and information retrieval (IR) tasks revolving around larger units of text, using semantic representations for phrases, sentences, or documents, constructed from lemma vectors. However, the problem of compositionality in DSMs—of how to go from word to sentence and beyond—has proved to be non-trivial.
A new framework, which we refer to as DisCoCat, initially presented in Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) reconciles distributional approaches to natural language semantics with the structured, logical nature of formal semantic models. This framework is abstract; its theoretical predictions have not been evaluated on real data, and its applications to empirical natural language processing tasks have not been studied.
This article is the journal version of Grefenstette and Sadrzadeh (2011a, 2011b), which fill this gap in the DisCoCat literature; in it, we develop a concrete model and an unsupervised learning algorithm to instantiate the abstract vectors, linear maps, and vector spaces of the theoretical framework; we develop a series of empirical natural language processing experiments and data sets and implement our algorithm on large scale real data; we analyze the outputs of the algorithm in terms of linear algebraic equations; and we evaluate the model on these experiments and compare the results with other competing unsupervised models. Furthermore, we provide a linear algebraic analysis of the algorithm of Grefenstette and Sadrzadeh (2011a) and present an in-depth study of the better performance of the method of Grefenstette and Sadrzadeh (2011b).
We begin in Section 2 by presenting the background to the task of developing compositional distributional models. We briefly introduce two approaches to semantic modeling: formal semantic models and distributional semantic models. We discuss their differences, and relative advantages and disadvantages. We then present and critique various approaches to bridging the gap between these models, and their limitations. In Section 3, we summarize the categorical compositional distributional framework of Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) and provide the theoretical background necessary to understand it; we also sketch the road map of the literature leading to the development of this setting and outline the contributions of this paper to the field. In Section 4, we present the details of an implementation of this framework, and introduce learning algorithms used to build semantic representations in this implementation. In Section 5, we present a series of experiments designed to evaluate this implementation against other unsupervised distributional compositional models. Finally, in Section 6 we discuss these results, and posit future directions for this research area. 2 background :Compositional formal semantic models represent words as parts of logical expressions, and compose them according to grammatical structure. They stem from classical ideas in logic and philosophy of language, mainly Frege’s principle that the meaning of a sentence is a function of the meaning of its parts (Frege 1892). These models relate to well-known and robust logical formalisms, hence offering a scalable theory of meaning that can be used to reason about language using logical tools of proof and inference. Distributional models are a more recent approach to semantic modeling, representing
the meaning of words as vectors learned empirically from corpora. They have found their way into real-world applications such as thesaurus extraction (Grefenstette 1994; Curran 2004) or automated essay marking (Landauer and Dumais 1997), and have connections to semantically motivated information retrieval (Manning, Raghavan, and Schütze 2011). This two-sortedness of defining properties of meaning: “logical form” versus “contextual use,” has left the quest for “what is the foundational structure of meaning?”—a question initially the concern of solely linguists and philosophers of language—even more of a challenge.
In this section, we present a short overview of the background to the work developed in this article by briefly describing formal and distributional approaches to natural language semantics, and providing a non-exhaustive list of some approaches to compositional distributional semantics. For a more complete review of the topic, we encourage the reader to consult Turney (2012) or Clark (2013). Formal semantic models provide methods for translating sentences of natural language into logical formulae, which can then be fed to computer-aided automation tools to reason about them (Alshawi 1992).
To compute the meaning of a sentence consisting of n words, meanings of these words must interact with one another. In formal semantics, this further interaction is represented as a function derived from the grammatical structure of the sentence. Such models consist of a pairing of syntactic analysis rules (in the form of a grammar) with semantic interpretation rules, as exemplified by the simple model presented on the left of Figure 1.
The semantic representations of words are lambda expressions over parts of logical formulae, which can be combined with one another to form well-formed logical expressions. The function | − | : L → M maps elements of the lexicon L to their interpretation (e.g., predicates, relations, domain objects) in the logical model M used. Nouns are typically just logical atoms, whereas adjectives and verbs and other relational words are interpreted as predicates and relations. The parse of a sentence such as “cats like milk,” represented here as a binarized parse tree, is used to produce its semantic interpretation by substituting semantic representations for their grammatical constituents and applying β-reduction where needed. Such a derivation is shown on the right of Figure 1.
What makes this class of models attractive is that it reduces language meaning to logical expressions, a subject well studied by philosophers of language, logicians, and linguists. Its properties are well known, and it becomes simple to evaluate the meaning of a sentence if given a logical model and domain, as well as verify whether or not one sentence entails another according to the rules of logical consequence and deduction.
However, such logical analysis says nothing about the closeness in meaning or topic of expressions beyond their truth-conditions and which models satisfy these truth conditions. Hence, formal semantic approaches to modeling language meaning do not perform well on language tasks where the notion of similarity is not strictly based on truth conditions, such as document retrieval, topic classification, and so forth. Furthermore, an underlying domain of objects and a valuation function must be provided, as with any logic, leaving open the question of how we might learn the meaning of language using such a model, rather than just use it. A popular way of representing the meaning of words in lexical semantics is as distributions in a high-dimensional vector space. This approach is based on the distributional hypothesis of Harris (1968), who postulated that the meaning of a word was dictated by the context of its use. The more famous dictum stating this hypothesis is the statement of Firth (1957) that “You shall know a word by the company it keeps.” This view of semantics has furthermore been associated (Grefenstette 2009; Turney and Pantel 2010) with earlier work in philosophy of language by Wittgenstein (presented in Wittgenstein 1953), who stated that language meaning was equivalent to its real world use.
Practically speaking, the meaning of a word can be learned from a corpus by looking at what other words occur with it within a certain context, and the resulting distribution can be represented as a vector in a semantic vector space. This vectorial representation is convenient because vectors are a familiar structure with a rich set of ways of computing vector distance, allowing us to experiment with different word similarity metrics. The geometric nature of this representation entails that we can not only compare individual words’ meanings with various levels of granularity (e.g., we might, for example, be able to show that cats are closer to kittens than to dogs, but that all three are mutually closer than cats and steam engines), but also apply methods frequently called upon in IR tasks, such as those described by Manning, Raghavan, and Schütze (2011), to group concepts by topic, sentiment, and so on.
The distribution underlying word meaning is a vector in a vector space, the basis vectors of which are dictated by the context. In simple models, the basis vectors will be annotated with words from the lexicon. Traditionally, the vector spaces used in such models are Hilbert spaces (i.e., vector spaces with orthogonal bases, such that the inner product of any one basis vector with another [other than itself] is zero). The semantic vector for any word can be represented as the weighted sum of the basis vectors:
−−−−−−−→ some word = ∑ i ci −→ni
where {−→ni }i is the basis of the vector space the meaning of the word lives in, and ci ∈ R is the weight associated with basis vector −→ni .
The construction of the vector for a word is done by counting, for each lexicon word ni associated with basis vector
−→ni , how many times it occurs in the context of each occurrence of the word for which we are constructing the vector. This count is then typically adjusted according to a weighting scheme (e.g., TF-IDF). The “context” of a word can be something as simple as the other words occurring in the same sentence as
the word or within k words of it, or something more complex, such as using dependency relations (Padó and Lapata 2007) or other syntactic features.
Commonly, the similarity of two semantic vectors is computed by taking their cosine measure, which is the sum of the product of the basis weights of the vectors:
cosine(−→a ,−→b ) = ∑ i c a i c b i√∑
i (c a i )
2 ∑
i (c b i ) 2
where cai and c b i are the basis weights for −→a and −→b , respectively. However, other options may be a better fit for certain implementations, typically dependent on the weighting scheme.
Readers interested in learning more about these aspects of distributional lexical semantics are invited to consult Curran (2004), which contains an extensive overview of implementation options for distributional models of word meaning. In the previous overview of distributional semantic models of lexical semantics, we have seen that DSMs are a rich and tractable way of learning word meaning from a corpus, and obtaining a measure of semantic similarity of words or groups of words. However, it should be fairly obvious that the same method cannot be applied to sentences, whereby the meaning of a sentence would be given by the distribution of other sentences with which it occurs.
First and foremost, a sentence typically occurs only once in a corpus, and hence substantial and informative distributions cannot be created in this manner. More importantly, human ability to understand new sentences is a compositional mechanism: We understand sentences we have never seen before because we can generate sentence meaning from the words used, and how they are put into relation. To go from word vectors to sentence vectors, we must provide a composition operation allowing us to construct a sentence vector from a collection of word vectors. In this section, we will discuss several approaches to solving this problem, their advantages, and their limitations.
2.3.1 Additive Models. The simplest composition operation that comes to mind is straightforward vector addition, such that:
−→ ab = −→a +−→b
Conceptually speaking, if we view word vectors as semantic information distributed across a set of properties associated with basis vectors, using vector addition as a semantic composition operation states that the information of a set of lemmas in a sentence is simply the sum of the information of the individual lemmas. Although crude, this approach is computationally cheap, and appears sufficient for certain NLP tasks: Landauer and Dumais (1997) show it to be sufficient for automated essay marking tasks, and Grefenstette (2009) shows it to perform better than a collection of other simple similarity metrics for summarization, sentence paraphrase, and document paraphrase detection tasks.
However, there are two principal objections to additive models of composition: first, vector addition is commutative, therefore, −−−−−−−−−−−−→ John drank wine = −−→ John + −−−→ drank + −−−→ wine =
−−−−−−−−−−−−→ Wine drank John, and thus vector addition ignores syntactic structure completely; and second, vector addition sums the information contained in the vectors, effectively jumbling the meaning of words together as sentence length grows.
The first objection is problematic, as the syntactic insensitivity of additive models leads them to equate the representation of sentences with patently different meanings. Mitchell and Lapata (2008) propose to add some degree of syntactic sensitivity—namely, accounting for word order—by weighting word vectors according to their order of appearance in a sentence as follows:
−→ ab = α−→a + β−→b
where α,β∈R. Consequently, −−−−−−−−−−−−→John drank wine=α · −−→John + β · −−−→drank + γ·−−−→wine would not have the same representation as −−−−−−−−−−−−→ Wine drank John = α · −−−→wine + β · −−−→drank + γ · −−→John.
The question of how to obtain weights and whether they are only used to reflect word order or can be extended to cover more subtle syntactic information is open, but it is not immediately clear how such weights may be obtained empirically and whether this mode of composition scales well with sentence length and increase in syntactic complexity. Guevara (2010) suggests using machine-learning methods such as partial least squares regression to determine the weights empirically, but states that this approach enjoys little success beyond minor composition such as adjective-noun or noun-verb composition, and that there is a dearth of metrics by which to evaluate such machine learning–based systems, stunting their growth and development.
The second objection states that vector addition leads to increase in ambiguity as we construct sentences, rather than decrease in ambiguity as we would expect from giving words a context; and for this reason Mitchell and Lapata (2008) suggest replacing additive models with multiplicative models as discussed in Section 2.3.2, or combining them with multiplicative models to form mixture models as discussed in Section 2.3.3.
2.3.2 Multiplicative Models. The multiplicative model of Mitchell and Lapata (2008) is an attempt to solve the ambiguity problem discussed in Section 2.3.1 and provide implicit disambiguation during composition. The composition operation proposed is the component-wise multiplication ( ) of two vectors: Vectors are expressed as the weighted sum of their basis vectors, and the weight of the basis vectors of the composed vector is the product of the weights of the original vectors; for −→a =∑i ci−→ni , and−→ b = ∑ i c ′ i −→ni , we have
−→ ab = −→a −→b = ∑ i cic′i −→ni
Such multiplicative models are shown by Mitchell and Lapata (2008) to perform better at verb disambiguation tasks than additive models for noun-verb composition, against a baseline set by the original verb vectors. The experiment they use to show this will also serve to evaluate our own models, and form the basis for further experiments, as discussed in Section 5.
This approach to compositionality still suffers from two conceptual problems: First, component-wise multiplication is still commutative and hence word order is not accounted for; second, rather than “diluting” information during large compositions
and creating ambiguity, it may remove too much through the “filtering” effect of component-wise multiplication.
The first problem is more difficult to deal with for multiplicative models than for additive models, because both scalar multiplication and component-wise multiplication are commutative and hence α−→a β−→b = α−→b β−→a and thus word order cannot be taken into account using scalar weights.
To illustrate how the second problem entails that multiplicative models do not scale well with sentence length, we look at the structure of component-wise multiplication again: −→a −→b =∑i cic′i −→ni . For any i, if ci = 0 or c′i = 0, then cic′i = 0, and therefore for any composition, the number of non-zero basis weights of the produced vector is less than or equal to the number of non-zero basis weights of the original vectors: At each composition step information is filtered out (or preserved, but never increased). Hence, as the number of vectors to be composed grows, the number of non-zero basis weights of the product vector stays the same or—more realistically—decreases. Therefore, for any composition of the form −−−−−−−−→a1 . . . ai . . . an = −→a1 . . . −→ai . . . −→an , if for any i, −→ai is orthogonal to −−−−−−→a1 . . . ai−1 then −−−−−−−−→a1 . . . ai . . . an = −→0 . It follows that purely multiplicative models alone are not apt as a single mode of composition beyond nounverb (or adjective-noun) composition operations.
One solution to this second problem not discussed by Mitchell and Lapata (2008) would be to introduce some smoothing factor s ∈ R+ for point-wise multiplication such that −→a −→b =∑i (ci + s)(c′i + s)−→ni , ensuring that information is never completely filtered out. Seeing how the problem of syntactic insensitivity still stands in the way of full-blown compositionality for multiplicative models, we leave it to those interested in salvaging purely multiplicative models to determine whether some suitable value of s can be determined.
2.3.3 Mixture Models. The problems faced by multiplicative models presented in Section 2.3.2 are acknowledged in passing by Mitchell and Lapata (2008), who propose mixing additive and multiplicative models in the hope of leveraging the advantage of each while doing away with their pitfalls. This is simply expressed as the weighted sum of additive and multiplicative models:
−→ ab = α−→a + β−→b + γ(−→a −→b )
where α, β, and γ are predetermined scalar weights. The problems for these models are threefold. First, the question of how scalar weights are to be obtained still needs to be determined. Mitchell and Lapata (2008) concede that one advantage of purely multiplicative models over weighted additive or mixture models is that the lack of scalar weights removes the need to optimize the scalar weights for particular tasks (at the cost of not accounting for syntactic structure), and avoids the methodological concerns accompanying this requirement.
Second, the question of how well this process scales from noun-verb composition to more syntactically rich expressions must be addressed. Using scalar weights to account for word order seems ad hoc and superficial, as there is more to syntactic structure than the mere ordering of words. Therefore, an account of how to build sentence vectors for sentences such as The dog bit the man and The man was bitten by the dog in order to give both sentences the same (or a similar) representation would need to give a richer role to scalar weights than mere token order. Perhaps specific
weights could be given to particular syntactic classes (such as nouns) to introduce a more complex syntactic element into vector composition, but it is clear that this alone is not a solution, as the weight for nouns dog and man would be the same, allowing for the same commutative degeneracy observed in non-weighted additive models, in which −−−−−−−−−−−−−−→ the dog bit the man = −−−−−−−−−−−−−−→ the man bit the dog. Introducing a mixture of weighting systems accounting for both word order and syntactic roles may be a solution, but it is not only ad hoc but also arguably only partially reflects the syntactic structure of the sentence.
The third problem is that Mitchell and Lapata (2008) show that in practice, although mixture models perform better at verb disambiguation tasks than additive models and weighted additive models, they perform equivalently to purely multiplicative models with the added burden of requiring parametric optimization of the scalar weights.
Therefore, whereas mixture models aim to take the best of additive and multiplicative models while avoiding their problems, they are only partly successful in achieving the latter goal, and demonstrably do little better in achieving the former.
2.3.4 Tensor-Based Models. From Sections 2.3.1–2.3.3 we observe that the need for incorporating syntactic information into DSMs to achieve true compositionality is pressing, if only to develop a non-commutative composition operation that can take into account word order without the need for adhoc weighting schemes, and hopefully richer syntactic information as well.
An early proposal by Smolensky and colleagues (Smolensky 1990; Smolensky and Legendre 2006) to use linear algebraic tensors as a composition operation solves the problem of finding non-commutative vector composition operators. The composition of two vectors is their tensor product, sometimes called the kronecker product when ap-
plied to vectors rather than vector spaces. For −→a ∈ V =∑i ci−→ni , and −→b ∈ W =∑j c′j−→n′j , we have:
−→ ab = −→a ⊗−→b = ∑ ij cic′j −→ni ⊗ −→ n′j
The composition operation takes the original vectors and maps them to a vector in a larger vector space V ⊗ W, which is the tensor space of the original vectors’ spaces. Here the second instance of ⊗ is not a recursive application of the kronecker product, but rather the pairing of basis elements of V and W to form a basis element of V ⊗ W. The shared notation and occasional conflation of kronecker and tensor products may seem confusing, but is fairly standard in multilinear algebra.
The advantage of this approach is twofold: First, vectors for different words need not live in the same spaces but can be composed nonetheless. This allows us to represent vectors for different word classes (topics, syntactic roles, etc.) in different spaces with different bases, which was not possible under additive or multiplicative models. Second, because the product vector lives in a larger space, we obtain the intuitive notion that the information of the whole is richer and more complex than the information of the parts.
Dimensionality Problems. However, as observed and commented upon by Smolensky himself, this increase in dimensionality brings two rather large problems for tensor based models. The first is computational: The size of the product vector space is the product of the size of the original vector spaces. If we assume that all words live in the
same space N of dimensionality dim(N), then the dimensionality of an n-word sentence vector is dim(N)n. If we have as many basis vectors for our word semantic space as there are lexemes in our vocabulary (e.g., approximately 170k in English2), then the size of our sentence vectors quickly reaches magnitudes for which vector comparison (or even storage) are computationally intractable.3 Even if, as most DSM implementations do, we restrict the basis vectors of word semantic spaces to the k (e.g., k = 2,000) most frequent words in a corpus, the sentence vector size still grows exponentially with sentence length, and the implementation problems remain.
The second problem is mathematical: Sentences of different length live in different spaces, and if we assign different vector spaces to different word types (e.g., syntactic classes), then sentences of different syntactic structure live in different vector spaces, and hence cannot be compared directly using inner product or cosine measures, leaving us with no obvious mode of semantic comparison for sentence vectors. If any model wishes to use tensor products in composition operations, it must find some way of reducing the dimensionality of product vectors to some common vector space so that they may be directly compared.
One notable method by which these dimensionality problems can be solved in general are the holographic reduced representations proposed by Plate (1991). The product vector of two vectors is projected into a space of smaller dimensionality by circular convolution to produce a trace vector. The circular correlation of the trace vector and one of the original vectors produces a noisy version of the other original vector. The noisy vector can be used to recover the clean original vector by comparing it with a predefined set of candidates (e.g., the set of word vectors if our original vectors are word meanings). Traces can be summed to form new traces effectively containing several vector pairs from which original vectors can be recovered. Using this encoding/decoding mechanism, the tensor product of sets of vectors can be encoded in a space of smaller dimensionality, and then recovered for computation without ever having to fully represent or store the full tensor product, as discussed by Widdows (2008).
There are problems with this approach that make it unsuitable for our purposes, some of which are discussed in Mitchell and Lapata (2010). First, there is a limit to the information that can be stored in traces, which is independent of the size of the vectors stored, but is a logarithmic function of their number. As we wish to be able to store information for sentences of variable word length without having to directly represent the tensored sentence vector, setting an upper bound to the number of vectors that can be composed in this manner limits the length of the sentences we can represent compositionally using this method.
Second, and perhaps more importantly, there are restrictions on the nature of the vectors that can be encoded in such a way: The vectors must be independently distributed such that the mean Euclidean length of each vector is 1. Such conditions are unlikely to be met in word semantic vectors obtained from a corpus; and as the failure to do so affects the system’s ability to recover clean vectors, holographic reduced representations are not prima facie usable for compositional DSMs, although it is important to note that Widdows (2008) considers possible application areas where they may be of use, although once again these mostly involve noun-verb and adjectivenoun compositionality rather than full blown sentence vector construction. We retain
2 Source: http://www.oxforddictionaries.com/page/howmanywords. 3 At four bytes per integer, and one integer per basis vector weight, the vector for John loves Mary
would require roughly (170, 000 · 4)3 ≈ 280 petabytes of storage, which is over ten times the data Google processes on a daily basis according to Dean and Ghemawat (2008).
from Plate (1991) the importance of finding methods by which to project the tensored sentence vectors into a common space for direct comparison, as will be discussed further in Section 3.
Syntactic Expressivity. An additional problem of a more conceptual nature is that using the tensor product as a composition operation simply preserves word order. As we discussed in Section 2.3.3, this is not enough on its own to model sentence meaning. We need to have some means by which to incorporate syntactic analysis into composition operations.
Early work on including syntactic sensitivity into DSMs by Grefenstette (1992) suggests using crude syntactic relations to determine the frame in which the distributions for word vectors are collected from the corpus, thereby embedding syntactic information into the word vectors. This idea was already present in the work of Smolensky, who used sums of pairs of vector representations and their roles, obtained by taking their tensor products, to obtain a vector representation for a compound. The application of these ideas to DSMs was studied by Clark and Pulman (2007), who suggest instantiating the roles to dependency relations and using the distributional representations of words as the vectors. For example, in the sentence Simon loves red wine, Simon is the subject of loves, wine is its object, and red is an adjective describing wine. Hence, from the dependency tree with loves as root node, its subject and object as children, and their adjectival descriptors (if any) as their children, we read the following structure: −−−→ loves ⊗−−→subj ⊗−−−−→Simon ⊗−→obj ⊗−−−→wine ⊗−→adj ⊗−→red. Using the equality relation for inner products of tensor products:
〈−→a ⊗−→b | −→c ⊗−→d 〉 = 〈−→a | −→c 〉 × 〈−→b | −→d 〉
We can therefore express inner-products of sentence vectors efficiently without ever having to actually represent the tensored sentence vector:
〈−−−−−−−−−−−−→Simon loves red wine | −−−−−−−−−−−−−−→Mary likes delicious rosé〉 = 〈−−→loves ⊗−→subj ⊗−−→Simon ⊗−→obj ⊗−−→wine ⊗−→adj ⊗−→red | −→likes ⊗−→subj ⊗−−→Mary ⊗−→obj ⊗−→rosé ⊗−→adj ⊗−−−−→delicious〉 = 〈−−→loves | −→likes〉× 〈−→subj | −→subj〉 × 〈−−→Simon | −−→Mary〉 × 〈−→obj | −→obj〉 × 〈−−→wine | −→rosé〉× 〈−→adj | −→adj 〉 × 〈−→red | −−−−→delicious〉 = 〈−−→loves | −→likes〉× 〈−−→Simon | −−→Mary〉× 〈−−→wine | −→rosé〉 × 〈−→red | −−−−→delicious〉
This example shows that this formalism allows for sentence comparison of sentences with identical dependency trees to be broken down to term-to-term comparison without the need for the tensor products to ever be computed or stored, reducing computation to inner product calculations.
However, although matching terms with identical syntactic roles in the sentence works well in the given example, this model suffers from the same problems as the original tensor-based compositionality of Smolensky (1990) in that, by the authors’ own admission, sentences of different syntactic structure live in spaces of different dimensionality and thus cannot be directly compared. Hence we cannot use this to measure the similarity between even small variations in sentence structure, such as the pair “Simon likes red wine” and “Simon likes wine.”
2.3.5 SVS Models. The idea of including syntactic relations to other lemmas in word representations discussed in Section 2.3.4 is applied differently in the structured vector
space model presented by Erk and Padó (2008). They propose to represent word meanings not as simple vectors, but as triplets:
w = (v, R, R−1)
where v is the word vector, constructed as in any other DSM, R and R−1 are selectional preferences, and take the form of R → D maps where R is the set of dependency relations, and D is the set of word vectors. Selectional preferences are used to encode the lemmas that w is typically the parent of in the dependency trees of the corpus in the case of R, and typically the child of in the case of R−1.
Composition takes the form of vector updates according to the following protocol. Let a = (va, Ra, R−1a ) and b = (vb, Rb, R −1 b ) be two words being composed, and let r be the dependency relation linking a to b. The vector update procedure is as follows:
a′ = (va R−1b (r), Ra − {r}, R−1a ) b′ = (vb Ra(r), Rb, R−1b − {r})
where a′, b′ are the updated word meanings, and is whichever vector composition (addition, component-wise multiplication) we wish to use. The word vectors in the triplets are effectively filtered by combination with the lemma which the word they are being composed with expects to bear relation r to, and this relation between the composed words a and b is considered to be used and hence removed from the domain of the selectional preference functions used in composition.
This mechanism is therefore a more sophisticated version of the compositional disambiguation mechanism discussed by Mitchell and Lapata (2008) in that the combination of words filters the meaning of the original vectors that may be ambiguous (e.g., if we have one vector for bank); but contrary to Mitchell and Lapata (2008), the information of the original vectors is modified but essentially preserved, allowing for further combination with other terms, rather than directly producing a joint vector for the composed words. The added fact that R and R−1 are partial functions associated with specific lemmas forces grammaticality during composition, since if a holds a dependency relation r to b which it never expects to hold (for example a verb having as its subject another verb, rather than the reverse), then Ra and R −1 b are undefined for r and the update fails. However, there are some problems with this approach if our goal is true compositionality.
First, this model does not allow some of the “repeated compositionality” we need because of the update of R and R−1. For example, we expect that an adjective composed with a noun produces something like a noun in order to be further composed with a verb or even another adjective. However, here, because the relation adj would be removed from R−1b for some noun b composed with an adjective a, this new representation b
′ would not have the properties of a noun in that it would no longer expect composition with an adjective, rendering representations of simple expressions like “the new red car” impossible. Of course, we could remove the update of the selectional preference functions from the compositional mechanism, but then we would lose this attractive feature of grammaticality enforcement through the partial functionality of R and R−1.
Second, this model does little more than represent the implicit disambiguation that is expected during composition, rather than actually provide a full blown compositional model. The inability of this system to provide a novel mechanism by which to obtain
a joint vector for two composed lemmas—thereby building towards sentence vectors— entails that this system provides no means by which to obtain semantic representations of larger syntactic structures that can be compared by inner product or cosine measure as is done with any other DSM. Of course, this model could be combined with the compositional models presented in Sections 2.3.1–2.3.3 to produce sentence vectors, but whereas some syntactic sensitivity would have been obtained, the word ordering and other problems of the aforementioned models would still remain, and little progress would have been made towards true compositionality.
We retain from this attempt to introduce compositionality in DSMs that including information obtained from syntactic dependency relations is important for proper disambiguation, and that having some mechanism by which the grammaticality of the expression being composed is a precondition for its composition is a desirable feature for any compositional mechanism. The final class of approaches to vector composition we wish to discuss are three matrix-based models.
Generic Additive Model. The first is the Generic Additive Model of Zanzotto et al. (2010). This is a generalization of the weighted additive model presented in Section 2.3.1. In this model, rather than multiplying lexical vectors by fixed parameters α and β before adding them to form the representation of their combination, they are instead the arguments of matrix multiplication by square matrices A and B:
−→ ab = A−→a + B−→b
Here, A and B represent the added information provided by putting two words into relation.
The numerical content of A and B is learned by linear regression over triplets (−→a ,−→b ,−→c ) where −→a and −→b are lexical semantic vectors, and −→c is the expected output of the combination of −→a and −→b . This learning system thereby requires the provision of labeled data for linear regression to be performed. Zanzotto et al. (2010) suggest several sources for this labeled data, such as dictionary definitions and word etymologies.
This approach is richer than the weighted additive models because the matrices act as linear maps on the vectors they take as “arguments,” and thus can encode more subtle syntactic or semantic relations. However, this model treats all word combinations as the same operation—for example, treating the combination of an adjective with its argument and a verb with its subject as the same sort of composition. Because of the diverse ways there are of training such supervised models, we leave it to those who wish to further develop this specific line of research to perform such evaluations.
Adjective Matrices. The second approach is the matrix-composition model of Baroni and Zamparelli (2010), which they develop only for the case of adjective-noun composition, although their approach can seamlessly be used for any other predicate-argument composition. Contrary to most of the earlier approaches proposed, which aim to combine two lexical vectors to form a lexical vector for their combination, Baroni and Zamparelli suggest giving different semantic representations to different types, or more specifically to adjectives and nouns.
In this model, nouns are lexical vectors, as with other models. However, embracing a view of adjectives that is more in line with formal semantics than with distributional semantics, they model adjectives as linear maps taking lexical vectors as input and producing lexical vectors as output. Such linear maps can be encoded as square matrices, and applied to their arguments by matrix multiplication. Concretely, let Madjective be the matrix encoding the adjective’s linear map, and −−→noun be the lexical semantic vector for a noun; their combination is simply
−−−−−−−−−→ adjective noun = Madjective ×−−→noun
Similarly to the Generic Additive Model, the matrix for each adjective is learned by linear regression over a set of pairs (−−→noun,−→c ) where the vectors −−→noun are the lexical semantic vectors for the arguments of the adjective in a corpus, and −→c is the semantic vector corresponding to the expected output of the composition of the adjective with that noun.
This may, at first blush, also appear to be a supervised training method for learning adjective matrices from “labeled data,” seeing how the expected output vectors are needed. However, Baroni and Zamparelli (2010) work around this constraint by automatically producing the labeled data from the corpus by treating the adjectivenoun compound as a single token, and learning its vector using the same distributional learning methods they used to learn the vectors for nouns. This same approach can be extended to other unary relations without change and, using the general framework of the current article, an extension of it to binary predicates has been presented in Grefenstette et al. (2013), using multistep regression. For a direct comparison of the results of this approach with some of the results of the current article, we refer the reader to Grefenstette et al. (2013).
Recursive Matrix-Vector Model. The third approach is the recently developed Recursive Matrix-Vector Model (MV-RNN) of Socher et al. (2012), which claims the two matrixbased models described here as special cases. In MV-RNN, words are represented as a pairing of a lexical semantic vector −→a with an operation matrix A. Within this model, given the parse of a sentence in the form of a binarized tree, the semantic representation (−→c , C) of each non-terminal node in the tree is produced by performing the following two operations on its children (−→a , A) and (−→b , B).
First, the vector component −→c is produced by applying the operation matrix of one child to the vector of the other, and vice versa, and then projecting both of the products back into the same vector space as the child vectors using a projection matrix W, which must also be learned:
−→c = W × [
B ×−→a A ×−→b
]
Second, the matrix C is calculated by projecting the pairing of matrices A and B back into the same space, using a projection matrix WM, which must also be learned:
C = WM × [
A B
]
The pairing (−→c , C) obtained through these operations forms the semantic representation of the phrase falling under the scope of the segment of the parse tree below that node.
This approach to compositionality yields good results in the experiments described in Socher et al. (2012). It furthermore has appealing characteristics, such as treating relational words differently through their operation matrices, and allowing for recursive composition, as the output of each composition operation is of the same type of object as its inputs. This approach is highly general and has excellent coverage of different syntactic types, while leaving much room for parametric optimization.
The principal mathematical difference with the compositional framework presented subsequently is that composition in MV-RNN is always a binary operation; for example, to compose a transitive verb with its subject and object one would first need to compose it with its object, and then compose the output of that operation with the subject. The framework we discuss in this article allows for the construction of larger representations for relations of larger arities, permitting the simultaneous composition of a verb with its subject and object. Whether or not this theoretical difference leads to significant differences in composition quality requires joint evaluation. Additionally, the description of MV-RNN models in Socher et al. (2012) specifies the need for a source of learning error during training, which is easy to measure in the case of label prediction experiments such as sentiment prediction, but non-trivial in the case of paraphrase detection where no objective label exists. A direct comparison to MV-RNN methods within the context of experiments similar to those presented in this article has been produced by Blacoe and Lapata (2012), showing that simple operations perform on par with the earlier complex deep learning architectures produced by Socher and colleagues; we leave direct comparisons to future work. Early work has shown that the addition of a hidden layer with non-linearities to these simple models will improve the results. Domains and Functions. In recent work, Turney (2012) suggests modeling word representations not as a single semantic vector, but as a pair of vectors: one containing the information of the word relative to its domain (the other words that are ontologically similar), and another containing information relating to its function. The former vector is learned by looking at what nouns a word co-occurs with, and the latter is learned by looking at what verb-based patterns the word occurs in. Similarity between sets of words is not determined by a single similarity function, but rather through a combination of comparisons of the domain components of words’ representations with the function components of the words’ representations. Such combinations are designed on a task-specific basis. Although Turney’s work does not directly deal with vector composition of the sort we explore in this article, Turney shows that similarity measures can be designed for tasks similar to those presented here. The particular limitation of his approach, which Turney discusses, is that similarity measures must be specified for each task, whereas most of the compositional models described herein produce representations that can be compared in a task-independent manner (e.g., through cosine similarity). Nonetheless, this approach is innovative, and will merit further attention in future work in this area.
Language as Algebra. A theoretical model of meaning as context has been proposed in Clarke (2009, 2012). In that model, the meaning of any string of words is a vector built
from the occurrence of the string in a corpus. This is the most natural extension of distributional models from words to strings of words: in that model, one builds vectors for strings of words in exactly the same way as one does for words. The main problem, however, is that of data sparsity for the occurrences of strings of words. Words do appear repeatedly in a document, but strings of words, especially for longer strings, rarely do so; for instance, it hardly happens that an exact same sentence appears more than once in a document. To overcome this problem, the model is based on the hypothetical concept of an infinite corpus, an assumption that prevents it from being applied to real-world corpora and experimented within natural language processing tasks. On the positive side, the model provides a theoretical study of the abstract properties of a general bilinear associative composition operator; in particular, it is shown that this operator encompasses other composition operators, such as addition, multiplication, and even tensor product. 3 discocat :In Section 2, we discussed lexical DSMs and the problems faced by attempts to provide a vector composition operation that would allow us to form distributional sentence representations as a function of word meaning. In this section, we will present an existing formalism aimed at solving this compositionality problem, as well as the mathematical background required to understand it and further extensions, building on the features and failures of previously discussed attempts at syntactically sensitive compositionality.
Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) propose adapting a category theoretic model, inspired by the categorical compositional vector space model of quantum protocols (Abramsky and Coecke 2004), to the task of compositionality of semantic vectors. Syntactic analysis in the form of pregroup grammars—a categorial grammar—is given categorical semantics in order to be represented as a compact closed category P (a concept explained subsequently), the objects of which are syntactic types and the morphisms of which are the reductions forming the basis of syntactic analysis. This syntactic category is then mapped onto the semantic compact closed category FVect of finite dimensional vector spaces and linear maps. The mapping is done in the product category FVect × P via the following procedure. Each syntactic type is interpreted as a vector space in which semantic vectors for words with that particular syntactic type live; the reductions between the syntactic types are interpreted as linear maps between the interpreted vector spaces of the syntactic types.
The key feature of category theory exploited here is its ability to relate in a canonical way different mathematical formalisms that have similar structures, even if the original formalisms belong in different branches of mathematics. In this context, it has enabled us to relate syntactic types and reductions to vector spaces and linear maps and obtain a mechanism by which syntactic analysis guides semantic composition operations.
This pairing of syntactic analysis and semantic composition ensures both that grammaticality restrictions are in place as in the model of Erk and Padó (2008) and syntactically driven semantic composition in the form of inner-products provide the implicit disambiguation features of the compositional models of Erk and Padó (2008) and Mitchell and Lapata (2008). The composition mechanism also involves the projection of tensored vectors into a common semantic space without the need for full representation of the tensored vectors in a manner similar to Plate (1991), without restriction to the nature of the vector spaces it can be applied to. This avoids the problems faced by other tensor-based composition mechanisms such as Smolensky (1990) and Clark and Pulman (2007).
The word vectors can be specified model-theoretically and the sentence space can be defined over Boolean values to obtain grammatically driven truth-theoretic semantics in the style of Montague (1974), as proposed by Clark, Coecke, and Sadrzadeh (2008). Some logical operators can be emulated in this setting, such as using swap matrices for negation as shown by Coecke, Sadrzadeh, and Clark (2010). Alternatively, corpus-based variations on this formalism have been proposed by Grefenstette et al. (2011) to obtain a non-truth theoretic semantic model of sentence meaning for which logical operations have yet to be defined.
Before explaining how this formalism works, in Section 3.3, we will introduce the notions of pregroup grammars in Section 3.1, and the required basics of category theory in Section 3.2. Presented by Lambek (1999, 2008) as a successor to his syntactic calculus (Lambek 1958), pregroup grammars are a class of categorial type grammars with pregroup algebras as semantics. Pregroups are particularly interesting within the context of this work because of their well-studied algebraic structure, which can trivially be mapped onto the structure of the category of vector spaces, as will be discussed subsequently. Logically speaking, a pregroup is a non-commutative form of Linear Logic (Girard 1987) in which the tensor and its dual par coincide; this logic is sometimes referred to as Bi-Compact Linear Logic (Lambek 1999). The formalism works alongside the general guidelines of other categorial grammars, for instance, those of the combinatory categorial grammar (CCG) designed by Steedman (2001) and Steedman and Baldridge (2011). They consist of atomic grammatical types that combine to form compound types. A series of CCGlike application rules allow for type-reductions, forming the basis of syntactic analysis. As our first step, we show how this syntactic analysis formalism works by presenting an introduction to pregroup algebras.
Pregroups. A pregroup is an algebraic structure of the form (P,≤, ·, 1, (−)l, (−)r). Its elements are defined as follows:
P is a set {a, b, c, . . .}. The relation ≤ is a partial ordering on P. The binary operation · is an associative, non-commutative monoid multiplication with the type − · − : P × P → P, such that if a, b ∈ P then a · b ∈ P. In other words, P is closed under this operation.
1 ∈ P is the unit, satisfying a · 1 = a = 1 · a for all a ∈ P. (−)l and (−)r are maps with types (−)l : P → P and (−)r : P → P such that for any a ∈ P, we have that al, ar ∈ P. The images of these maps are referred to as the left and the right adjoints. These are unique and satisfy the following conditions:
– Reversal: if a ≤ b then bl ≤ al (and similarly for ar, br). – Ordering: a · ar ≤ 1 ≤ ar · a and al · a ≤ 1 ≤ a · al. – Cancellation: alr = a = arl. – Self-adjointness of identity: 1r = 1 = 1l. – Self-adjointness of multiplication: (a · b)r = br · ar.
As a notational simplification we write ab for a · b, and if abcd ≤ cd we write abcd → cd and call this a reduction, omitting the identity wherever it might appear. Monoid multiplication is associative, so parentheses may be added or removed for notational clarity without changing the meaning of the expression as long as they are not directly under the scope of an adjoint operator.
An example reduction in pregroup might be:
aarbclc → bclc → b
We note here that the reduction order is not always unique, as we could have reduced as follows: aarbclc → aarb → b. As a further notational simplification, if there exists a chain of reductions a → . . . → b we may simply write a → b (in virtue of the transitivity of partial ordering relations). Hence in our given example, we can express both reduction paths as aarbclc → b.
Pregroups and Syntactic Analysis. Pregroups are used for grammatical analysis by freely generating the set P of a pregroup from the basic syntactic types n, s, . . . , where here n stands for the type of both a noun and a noun phrase and s for that of a sentence. The conflation of nouns and noun phrases suggested here is done to keep the work discussed in this article as simple as possible, but we could of course model them as different types in a more sophisticated version of this pregroup grammar. As in any categorial grammar, words of the lexicon are assigned one or more possible types (corresponding to different syntactic roles) in a predefined type dictionary, and the grammaticality of a sentence is verified by demonstrating the existence of a reduction from the combination of the types of words within the sentence to the sentence type s.
For example, having assigned to noun phrases the type n and to sentences the type s, the transitive verbs will have the compound type nrsnl. We can read from the type of a transitive verb that it is the type of a word which “expects” a noun phrase on its left and a noun phrase on its right, in order to produce a sentence. A sample reduction of John loves cake with John and cake being noun phrases of type n and loves being a verb of type nrsnl is as follows:
n(nrsnl)n → s
We see that the transitive verb has combined with the subject and object to reduce to a sentence. Because the combination of the types of the words in the string John loves cake reduces to s, we say that this string of words is a grammatical sentence. As for more examples, we recall that intransitive verbs can be given the type nrs such that John sleeps would be analyzed in terms of the reduction n(nrs) → s. Adjectives can be given the type nnl such that red round rubber ball would be analyzed by (nnl)(nnl)(nnl)n → n. And so on and so forth for other syntactic classes.
Lambek (2008) presents the details of a slightly more complex pregroup grammar with a richer set of types than presented here. This grammar is hand-constructed and iteratively extended by expanding the type assignments as more sophisticated grammatical constructions are discussed. No general mechanism is proposed to cover all such types of assignments for larger fragments (e.g., as seen in empirical data). Pregroup grammars have been proven to be learnable by Béchet, Foret, and Tellier (2007), who also discuss the difficulty of this task and the nontractability of the procedure. Because of these constraints and lack of a workable pregroup parser, the pregroup grammars
we will use in our categorical formalism are derived from CCG types, as we explain in the following.
Pregroup Grammars and Other Categorial Grammars. Pregroup grammars, in contrast with other categorial grammars such as CCG, do not yet have a large set of tools for parsing available. If quick implementation of the formalism described later in this paper is required, it would be useful to be able to leverage the mature state of parsing tools available for other categorial grammars, such as the Clark and Curran (2007) statistical CCG parser, as well as Hockenmaier’s CCG lexicon and treebank (Hockenmaier 2003; Hockenmaier and Steedman 2007). In other words, is there any way we can translate at least some subset of CCG types into pregroup types?
There are some theoretical obstacles to consider first: Pregroup grammars and CCG are not equivalent. Buszkowski (2001) shows pregroup grammars to be equivalent to context-free grammars, whereas Joshi, Vijay-Shanker, and Weir (1989) show CCG to be weakly equivalent to more expressive mildly context-sensitive grammars. However, if our goal is to exploit the CCG used in Clark and Curran’s parser, or Hockenmaier’s lexicon and treebank, we may be in luck: Fowler and Penn (2010) prove that some CCGs, such as those used in the aforementioned tools, are strongly context-free and thus expressively equivalent to pregroup grammars. In order to be able to apply the parsing tools for CCGs to our setting, we use a translation mechanism from CCG types to pregroup types based on the Lambek-calculus-to-pregroup-grammar translation originally presented in Lambek (1999). In this mechanism, each atomic CCG type X is assigned a unique pregroup type x; for any X/Y in CCG we have xyl in the pregroup grammar; and for any X\Y in CCG we have yrx in pregroup grammar. Therefore, by assigning NP to n and S to s we could, for example, translate the CCG transitive verb type (S\NP)/NP into the pregroup type nrsnl, which corresponds to the pregroup type we used for transitive verbs in Section 3.1. Wherever type replacement (e.g., N → NP) is allowed in CCG we set an ordering relation in the pregroup grammar (e.g., n̄ ≤ n, where n̄ is the pregroup type associated with N). Because forward and backward slash “operators” in CCG are not associative whereas monoid multiplication in pregroups is, it is evident that some information is lost during the translation process. But because the translation we need is one-way, we may ignore this problem and use CCG parsing tools to obtain pregroup parses. Another concern lies with CCG’s crossed composition and substitution rules. The translations of these rules do not in general hold in a pregroup; this is not a surprise as pregroups are a simplification of the Lambek Calculus and these rules did not hold in the Lambek Calculus either, as shown in Moortgat (1997), for example. However, for the phenomena modeled in this paper, the CCG rules without the backward cross rules will suffice. In general for the case of English, one can avoid the use of these rules by overloading the lexicon and using additional categories. To deal with languages that have cross dependancies, such as Dutch, various solutions have been suggested (e.g., see Genkin, Francez, and Kaminski 2010; Preller 2010). Category theory is a branch of pure mathematics that allows for a general and uniform formulation of various different mathematical formalisms in terms of their main structural properties using a few abstract concepts such as objects, arrows, and combinations and compositions of these. This uniform conceptual language allows for derivation of new properties of existing formalisms and for relating these to properties of other formalisms, if they bear similar categorical representation. In this function, it has been
at the center of recent work in unifying two orthogonal models of meaning, a qualitative categorial grammar model and a quantitative distributional model (Clark, Coecke, and Sadrzadeh 2008; Coecke, Sadrzadeh, and Clark 2010). Moreover, the unifying categorical structures at work here were inspired by the ones used in the foundations of physics and the modeling of quantum information flow, as presented in Abramsky and Coecke (2004), where they relate the logical structure of quantum protocols to their state-based vector spaces data. The connection between the mathematics used for this branch of physics and those potentially useful for linguistic modeling has also been noted by several sources, such as Widdows (2005), Lambek (2010), and Van Rijsbergen (2004).
In this section, we will briefly examine the basics of category theory, monoidal categories, and compact closed categories. The focus will be on defining enough basic concepts to proceed rather than provide a full-blown tutorial on category theory and the modeling of information flow, as several excellent sources already cover both aspects (e.g., Mac Lane 1998; Walters 1991; Coecke and Paquette 2011). A categoriesin-a-nutshell crash course is also provided in Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010).
The Basics of Category Theory. A basic category C is defined in terms of the following elements:
A collection of objects ob(C). A collection of morphisms hom(C). A morphism composition operation ◦.
Each morphism f has a domain dom( f ) ∈ ob(C) and a codomain codom( f ) ∈ ob(C). For dom( f ) = A and codom( f ) = B we abbreviate these definitions as f : A → B. Despite the notational similarity to function definitions (and sets and functions being an example of a category), it is important to state that nothing else is presupposed about morphisms, and we should not treat them a priori as functions.
The following axioms hold in every category C :
For any f : A → B and g : B → C there exists h : A → C and h = g ◦ f . For any f : A → B, g : B → C and h : C → D, ◦ satisfies (h ◦ g) ◦ f = h ◦ (g ◦ f ). For every A ∈ ob(C) there is an identity morphism idA : A → A such that for any f : A → B, f ◦ idA = f = idB ◦ f .
We can model various mathematical formalisms using these basic concepts, and verify that these axioms hold for them. For example, category Set with sets as objects and functions as morphisms, or category Rel with sets as objects and relations as morphisms, category Pos with posets as objects and order-preserving maps as morphisms, and category Group with groups as objects and group homomorphisms as morphisms, to name a few.
The product C × D of two categories C and D is a category with pairs (A, B) as objects, where A ∈ ob(C) and B ∈ ob(D). There exists a morphism (f, g) : (A, B) → (C, D) in C × D if and only if there exists f : A → C ∈ hom(C) and g : B → D ∈ hom(D). Product categories allow us to relate objects and morphisms of one mathematical formalism to those in another, in this example those of C to D.
Compact Closed Categories. A monoidal category C is a basic category to which we add a monoidal tensor ⊗ such that:
For all A, B ∈ ob(C) there is an object A ⊗ B ∈ ob(C). For all A, B, C ∈ ob(C), we have (A ⊗ B) ⊗ C = A ⊗ (B ⊗ C). There exists some I ∈ ob(C) such that for any A ∈ ob(C), we have I ⊗ A = A = A ⊗ I. For f : A → C and g : B → D in hom(C) there is f ⊗ g : A ⊗ B → C ⊗ D in hom(C). For f1 : A → C, f2 : B → D, g1 : C → E and g2 : D → F the following equality holds:
(g1 ⊗ g2) ◦ ( f1 ⊗ f2) = (g1 ◦ f1) ⊗ (g2 ◦ f2) The product category of two monoidal categories has a monodial tensor, defined pointwisely by (a, A) ⊗ (b, B) := (a ⊗ b, A ⊗ B).
A compact closed category C is a monoidal category with the following additional axioms:
Each object A ∈ ob(C) has left and right “adjoint” objects Al and Ar in ob(C).
There exist four structural morphisms for each object A ∈ ob(C): – ηlA : I → A ⊗ Al. – ηrA : I → Ar ⊗ A. – lA : A
l ⊗ A → I. – rA : A ⊗ Ar → I.
The previous structural morphisms satisfy the following equalities:
– (1A ⊗ lA) ◦ (ηlA ⊗ 1A) = 1A. – ( rA ⊗ 1A) ◦ (1A ⊗ ηrA) = 1A. – (1Ar ⊗ rA) ◦ (ηr ⊗ 1Ar ) = 1Ar . – ( lA ⊗ 1Al ) ◦ (1Al ⊗ ηlA) = 1Al .
Compact closed categories come equipped with complete graphical calculi, surveyed in Selinger (2010). These calculi visualize and simplify the axiomatic reasoning within the category to a great extent. In particular, Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) show how they depict the pregroup grammatical reductions and visualize the flow of information in composing meanings of single words and forming meanings for sentences. Although useful at an abstract level, these calculi do not play the same simplifying role when it comes to the concrete and empirical computations; therefore we will not discuss them in this article.
A very relevant example of a compact closed category is a pregroup algebra P. Elements of a pregroup are objects of the category, the partial ordering relation provides the morphisms, 1 is I, and monoidal multiplication is the monoidal tensor.
Another very relevant example is the category FVect of finite dimensional Hilbert spaces and linear maps over R—that is, vector spaces over R with orthogonal bases of finite dimension, and an inner product operation 〈− | −〉 : A × A → R for every vector space A. The objects of FVect are vector spaces, and the morphisms are linear
maps between vector spaces. The unit object is R and the monoidal tensor is the linear algebraic tensor product. FVect is degenerate in its adjoints, in that for any vector space A, we have Al = Ar = A∗, where A∗ is the dual space of A. Moreover, by fixing a basis we obtain that A∗ ∼= A. As such, we can effectively do away with adjoints in this category, and “collapse” l, r, ηl, and ηr maps into “adjoint-free” and η maps. In this category, the maps are inner product operations, A : A ⊗ A → R, and the η maps η : R → A ⊗ A generate maximally entangled states, also referred to as Bell-states. In Section 3.2 we showed how a pregroup algebra and vector spaces can be modeled as compact closed categories and how product categories allow us to relate the objects and morphisms of one category to those of another. In this section, we will present how Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) suggest building on this by using categories to relate semantic composition to syntactic analysis in order to achieve syntax-sensitive composition in DSMs.
3.3.1 Syntax Guides Semantics. The product category FVect × P has as object pairs (A, a), where A is a vector space and a is a pregroup grammatical type, and as morphism pairs ( f,≤) where f is a linear map and ≤ a pregroup ordering relation. By the definition of product categories, for any two vector space-type pairs (A, a) and (B, b), there exists a morphism (A, a) → (B, b) only if there is both a linear map from A into B and a partial ordering a → b. If we view these pairings as the association of syntactic types with vector spaces containing semantic vectors for words of that type, this restriction effectively states that a linear map from A to B is only “permitted” in the product category if a reduces to b.
Both P and FVect being compact closed, it is simple to show that FVect × P is as well, by considering the pairs of unit objects and structural morphisms from the separate categories: I is now (R, 1), and the structural morphisms are ( A, la), ( A, r a), (ηA,η l a), and (ηA,ηra). We are particularly interested in the maps, which are defined as follows (from the definition of product categories):
( A, lA) : (A ⊗ A, ala) → (R, 1) ( A, rA) : (A ⊗ A, aar) → (R, 1)
This states that whenever there is a reduction step in the grammatical analysis of a sentence, there is a composition operation in the form of an inner product on the semantic front. Hence, if nouns of type n live in some noun space N and transitive verbs of type nlsnr live in some space N ⊗ S ⊗ N, then there must be some structural morphism of the form:
( N ⊗ 1S ⊗ N, rn1s ln) : (N ⊗ (N ⊗ S ⊗ N) ⊗ N, n(nrsnl)n) → (S, s)
We can read from this morphism the functions required to compose a sentence with a noun, a transitive verb, and an object to obtain a vector living in some sentence space S, namely, ( N ⊗ 1S ⊗ N ).
The form of a syntactic type is therefore what dictates the structure of the semantic space associated with it. The structural morphisms of the product category guarantee that for every syntactic reduction there is a semantic composition morphism provided by the product category: syntactic analysis guides semantic composition.
3.3.2 Example. To give an example, we give syntactic type n to nouns, and nrs to intransitive verbs. The grammatical reduction for kittens sleep, namely, nnrs → s, corresponds to the morphism rn ⊗ 1s in P. The syntactic types dictate that the noun −−−−→ kittens lives in some vector space N, and the intransitive verb −−−→ sleep in N ⊗ S. The reduction morphism gives us the composition morphism ( N ⊗ 1S), which we can apply to −−−−→ kittens ⊗−−−→sleep. Because we can express any vector as the weighted sum of its basis vectors, let us expand −−−−→ kittens = ∑ i c kittens i −→ni and −−−→ sleep = ∑ ij c sleep ij
−→ni ⊗−→sj ; then we can express the composition as follows:
−−−−−−−−→ kittens sleep = ( N ⊗ 1S)( −−−−→ kittens ⊗−−−→sleep)
= ( N ⊗ 1S) ⎛ ⎝∑
i
ckittensi −→ni ⊗ ∑ jk csleepjk −→nj ⊗−→sk
⎞ ⎠
= ( N ⊗ 1S) ⎛ ⎝∑
ijk
ckittensi c sleep jk −→ni ⊗−→nj ⊗−→sk
⎞ ⎠
= ∑
ijk
ckittensi c sleep jk 〈−→ni | −→nj 〉−→sk
= ∑
ik
ckittensi c sleep ik −→sk
In these equations, we have expressed the vectors in their explicit form, we have consolidated the sums by virtue of distributivity of linear algebraic tensor product over addition, we have applied the tensored linear maps to the vector components (as the weights are scalars), and finally, we have simplified the indices since 〈−→ni | −→nj 〉 = 1 if−→ni = −→nj and 0 otherwise. As a result of these, we have obtained a vector that lives in sentence space S.
Transitive sentences can be dealt with in a similar fashion:
−−−−−−−−−−−−−→ kittens chase mice
= ( N ⊗ 1S ⊗ N)( −−−−→ kittens ⊗−−−→chase ⊗−−→mice)
= ( N ⊗ 1S ⊗ N) ⎛ ⎝∑
i
ckittensi −→ni ⊗
⎛ ⎝∑
jkl
cchasejkl −→nj ⊗−→sk ⊗−→nl ⎞ ⎠⊗∑
m
cmicem −→nm
⎞ ⎠
= ( N ⊗ 1S ⊗ N) ⎛ ⎝∑
ijklm
ckittensi c chase jkl c mice m −→ni ⊗−→nj ⊗−→sk ⊗−→nl ⊗−→nm
⎞ ⎠
= ∑ ijklm ckittensi c chase jkl c mice m 〈−→ni | −→nj 〉−→sk 〈−→nl | −→nm〉
= ∑ ikm ckittensi c chase ikm c mice m −→sk
In both cases, it is important to note that the tensor product passed as argument to the composition morphism, namely,
−−−−→ kittens ⊗−−−→sleep in the intransitive case and−−−−→
kittens ⊗−−−→chase ⊗−−→mice in the transitive case, never needs to be computed. We can treat the tensor products here as commas separating function arguments, thereby avoiding the dimensionality problems presented by earlier tensor-based approaches to compositionality. As elaborated on in Section 2.3.4, the first general setting for pairing meaning vectors with syntactic types was proposed in Clark and Pulman (2007). The setting of a DisCoCat generalized this by making the meaning derivation process rely on a syntactic type system, hence overcoming its central problem whereby the vector representations of strings of words with different grammatical structure lived in different spaces. A preliminary version of a DisCoCat was developed in Clark, Coecke, and Sadrzadeh (2008), a full version was elaborated on in Coecke, Sadrzadeh, and Clark (2010), where, based on the developments of Preller and Sadrzadeh (2010), it was also exemplified how the vector space model may be instantiated in a truth theoretic setting where meanings of words were sets of their denotations and meanings of sentences were their truth values. A nontechnical description of this theoretical setting was presented in Clark (2013), where a plausibility truth-theoretic model for sentence spaces was worked out and exemplified. The work of Grefenstette et al. (2011) focused on a tangential branch and developed a toy example where neither words nor sentence spaces were Boolean. The applicability of the theoretical setting to a real empirical natural language processing task and data from a large scale corpus was demonstrated in Grefenstette and Sadrzadeh (2011a, 2011b). There, we presented a general algorithm to build vector representations for words with simple and complex types and the sentences containing them; then applied the algorithm to a disambiguation task performed on the British National Corpus (BNC). We also investigated the vector representation of transitive verbs and showed how a number of single operations may optimize the performance. We discuss these developments in detail in the following sections. 4 concrete semantic spaces :In Section 3.3 we presented a categorical formalism that relates syntactic analysis steps to semantic composition operations. The structure of our syntactic representation dictates the structure of our semantic spaces, but in exchange, we are provided with composition functions by the syntactic analysis, rather than having to stipulate them ad hoc. Whereas the approaches to compositional DSMs presented in Section 2 either failed to take syntax into account during composition, or did so at the cost of not being able to compare sentences of different structure in a common space, this categorical approach projects all sentences into a common sentence space where they can be directly compared. However, this alone does not give us a compositional DSM.
As we have seen in the previous examples, the structure of semantic spaces varies with syntactic types. We therefore cannot construct vectors for different syntactic types in the same way, as they live in spaces of different structure and dimensionality. Furthermore, nothing has yet been said about the structure of the sentence space S into which expressions reducing to type s are projected. If we wish to have a compositional DSM
that leverages all the benefits of lexical DSMs and ports them to sentence-level distributional representations, we must specify a new sort of vector construction procedure.
In the original formulation of this formalism by Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010), examples of how such a compositional DSM could be used for logical evaluation are presented, where S is defined as a Boolean space with True and False as basis vectors. However, the word vectors used are handwritten and specified model-theoretically, as the authors leave it for future research to determine how such vectors might be obtained from a corpus. In this section, we will discuss a new way of constructing vectors for compositional DSMs, and of defining sentence space S, in order to reconcile this powerful categorical formalism with the applicability and flexibility of standard distributional models. Assume the following sentences are all true:
1. The dogs chased the cats.
2. The dogs annoyed the cats.
3. The puppies followed the kittens.
4. The train left the station.
5. The president followed his agenda.
If asked which sentences have similar meaning, we would most likely point to the pair (1) and (3), and perhaps to a lesser degree (1) and (2), and (2) and (3). Sentences (4) and (5) obviously speak of a different state of the world as the other sentences.
If we compare these by truth value, we obviously have no means of making such distinctions. If we compare these by lexical similarity, (1) and (2) seem to be a closer match than (1) and (3). If we are classifying these sentences by some higher order relation such as “topic,” (5) might end up closer to (3) than (1). What, then, might cause us to pair (1) and (3)?
Intuitively, this similarity seems to be because the subjects and objects brought into relation by similar verbs are themselves similar. Abstracting away from tokens to some notion of property, we might say that both (1) and (3) express the fact that something furry and feline and furtive is being pursued by something aggressive (or playful) and canine. Playing along with the idea that lexical distributional semantics present concepts (word meanings) as “messy bundles of properties,” it seems only natural to have the way these properties are acted upon, qualified, and related as the basis for sentence-level distributional representations. In this respect, we here suggest that the sentence space S, instead of qualifying the truth value of a sentence, should express how the properties of the semantic objects within are qualified or brought into relation by verbs, adjectives, and other predicates and relations.
More specifically, we examine two suggestions for defining the sentence space, namely, SI ∼= N for sentences with intransitive verbs and ST ∼= N ⊗ N for sentences with transitive verbs. These definitions mean that the sentence space’s dimensions are commensurate with either those of N, or those of N ⊗ N. These are by no means the only options, but as we will discuss here, they offer practical benefits.
In the case of SI, the basis elements are labeled with unique basis elements of N, hence, −→s1 = −→n1 , −→s2 = −→n2, and so on. In the case of ST, the basis elements are labeled with
unique ordered pairs of elements from N, for example, −→s1 = −−−−→ (n1, n1), −→s2 = −−−−→ (n2, n1), −→s3 =−−−−→ (n1, n2), and so on, following the same matrix flattening structure used in the proof of the equal cardinality of N and Q. Because of the isomorphism between ST and N ⊗ N, we will use the notations −−−−→ (ni, nj) and
−→ni ⊗−→nj interchangeably, as both constitute appropriate ways of representing the basis elements of such a space. To propagate this distinction on the syntactic level, we define types sI and sT for intransitive and transitive sentences, respectively.
In creating this distinction, we lost one of the most appealing features of the framework of Coecke, Sadrzadeh, and Clark (2010), namely, the result that all sentence vectors live in the same sentence space. A mathematical solution to this two-space problem was suggested in Grefenstette et al. (2011), and a variant of the models presented in this article permitting the non-problematic projection into a single sentence space (S ∼= N) has been presented by Grefenstette et al. (2013). Keeping this separation allows us to deal with both the transitive and intransitive cases in a simpler manner, and because the experiments in this article only compare intransitive sentences with intransitive sentences, and transitive sentences with transitive sentences, we will not address the issue of unification here. While Lambek’s pregroup types presented in Lambek (2008) include a rich array of basic types and hand-designed compound types in order to capture specific grammatic properties, for the sake of simplicity we will use a simpler set of grammatical types for experimental purposes, similar to some common types found in the CCG-bank (Steedman 2001).
We assign a basic pregroup type n for all nouns, with an associated vector space N for their semantic representations. Furthermore, we will treat noun-phrases as nouns, assigning to them the same pregroup type and semantic space.
CCG treats intransitive verbs as functions NP\S that consume a noun phrase and return a sentence, and transitive verbs as functions (NP\S)/NP that consume a noun phrase and return an intransitive verb function, which in turn consumes a noun phrase and returns a sentence. Using our distinction between intransitive sentences, we give intransitive verbs the type nrsI associated with the semantic space N ⊗ SI, and transitive verbs the type nrsTnl associated with the semantic space N ⊗ ST ⊗ N.
Adjectives, in CCG, are treated as functions NP/NP, consuming a noun phrase and returning a noun phrase; and hence we give them the type nnl and associated semantic space N ⊗ N.
With the provision of a learning procedure for vectors in these semantic spaces, we can use these types to construct sentence vector representations for simple intransitive verb–based and transitive verb–based sentences, with and without adjectives applied to subjects and objects. To begin, we construct the semantic space N for all nouns in our lexicon (typically limited by the words available in the corpus used). Any distributional semantic model can be used for this stage, such as those presented in Curran (2004), or the lexical semantic models used by Mitchell and Lapata (2008). It seems reasonable to assume that higher quality lexical semantic vectors—as measured by metrics such as the WordSim353 test
of Finkelstein et al. (2001)—will produce better relational vectors from the procedure designed subsequently. We will not test the hypothesis here, but note that it is an underlying assumption in most of the current literature on the subject (Erk and Padó 2008; Mitchell and Lapata 2008; Baroni and Zamparelli 2010).
Building upon the foundation of the thus constructed noun vectors, we construct semantic representations for relational words. In pregroup grammars (or other combinatorial grammars such as CCG), we can view such words as functions, taking as arguments those types present as adjoints in the compound pregroup type, and returning a syntactic object whose type is that of the corresponding reduction. For example, an adjective nnl takes a noun or noun phrase n and returns a noun phrase n from the reduction (nnl)n → n. It can also compose with another adjective to return an adjective (nnl)(nnl) → nnl. We wish for our semantic representations to be viewed in the same way, such that the composition of an adjective with a noun (1N ⊗ N)((N ⊗ N) ⊗ N) can be viewed as the application of a function f : N → N to its argument of type N.
To learn the representation of such functions, we assume that their meaning can be characterized by the properties that their arguments hold in the corpus, rather than just by their context as is the case in lexical distributional semantic models. To give an example, rather than learning what the adjective angry means by observing that it co-occurs with words such as fighting, aggressive, or mean, we learn its meaning by observing that it typically takes as argument words that have the property of being (i.e., co-occur with words such as) fighting, aggressive, and mean. Whereas in the lexical semantic case, such associations might only rarely occur in the corpus, in this indirect method we learn what properties the adjective relates to even if they do not co-occur with it directly.
In turn, through composition with its argument, we expect the function for such an adjective to strengthen the properties that characterize it in the object it takes as argument. If, indeed, angry is characterized by arguments that have high basis weights for basis elements corresponding to the concepts (or context words) fighting, aggressive, and mean, and relatively low counts for semantically different concepts such as passive, peaceful, and loves, then when we apply angry to dog the vector for the compound angry dog should contain some of the information found in the vector for dog. But this vector should also have higher values for the basis weights of fighting, aggressive, and mean, and correspondingly lower values for the basis weights of passive, peaceful, and loves.
To turn this idea into a concrete algorithm for constructing the semantic representation for relations of any arity, as first presented in Grefenstette et al. (2011), we examine how we would deal with this for binary relations such as transitive verbs. If a transitive verb of semantic type N ⊗ ST ⊗ N is viewed as a function f : N × N → ST that expresses the extent to which the properties of subject and object are brought into relation by the verb, we learn the meaning of the verb by looking at what properties are brought into relation by its arguments in a corpus. Recall that the vector for a verb v,−→v ∈ N ⊗ ST ⊗ N, can be expressed as the weighted sum of its basis elements:
−→v = ∑
ijk
cvijk −→ni ⊗−→sj ⊗−→nk
We take the set of vectors for the subject and object of v in the corpus to be argv = {(−−−−→SUBJl, −−−→ OBJl )}l. We wish to calculate the basis weightings {cvijk}ijk for v. Exploiting our earlier definition of the basis {sj}j of ST, which states that for any value of i and k there
is some value of j such that sj = (ni, nk), we define Δijk = 1 if indeed sj = (ni, nk) and 0 otherwise. Using all of this, we define the calculation of each basis weight cvijk as:
cvijk = ∑
l
Δijkc SUBJl i c OBJl k
This allows for a full formulation of −→v as follows: −→v =
∑ l ∑ ijk Δijkc SUBJl i c OBJl k −→ni ⊗−→sj ⊗−→nk
The nested sums here may seem computationally inefficient, seeing how this would involve computing size(argv) × dim(N)2 × dim(S) = size(argv) × dim(N)4 products. However, using the decomposition of basis elements of S into pairs of basis elements of N (effectively basis elements of N ⊗ N), we can remove the Δijk term and ignore all values of j where sj = (ni, nk), because the basis weight for this combination of indices would be 0. Hence we simplify:
−→v = ∑
l ∑ ik cSUBJli c OBJl k −→ni ⊗ −−−−→ (ni, nk) ⊗−→nk
This representation is still bloated: We perform fewer calculations, but still obtain a vector in which all the basis weights where sj = (ni, nk) are 0, hence where only dim(N)2 of the dim(N)4 values are non-zero. In short, the vector weights for −→v are, under this learning algorithm, entirely characterized by the values of a dim(N) by dim(N) matrix, the entries of which are products cSUBJli c OBJl k where i and k have become row and column indices. Using this and our definition of ST as a space isomorphic to N ⊗ N, we can formulate a compact expression of −→v as follows. Let the kronecker product of two vectors−→u ,−→w ∈ N, written −→u ⊗−→w ∈ N ⊗ N, be as follows: −→u ⊗−→w =
∑ ij cui c w j −→ni ⊗−→nj
Equipped with this definition, we can formulate the compact form of −→v :
−→v = ∑
l ∑ ik cSUBJli c OBJl k −→ni ⊗−→nk
= ∑
l
−−−→ SUBJl ⊗ −−→ OBJl
In short, we are only required to iterate through the corpus once, taking for each instance of a transitive verb v the kronecker product of its subject and object, and summing these across all instances of v. It is simple to see that no information was discarded relative to the previous definition of −→v : The dimensionality reduction by a factor of dim(N)2 simply discards all basis elements for which the basis weight was 0 by default.
This raises a small problem though: This compact representation can no longer be used in the compositional mechanism presented in Section 3.3, as the dimensions of−→v no longer match those that it is required to have according to its syntactic type. However, a solution can be devised if we return to the sample calculation shown in Section 3.3.2 of the composition of a transitive verb with its arguments. The composition reduces as follows:
( N ⊗ 1S ⊗ N)( −−−→ SUBJ ⊗−→v ⊗−−→OBJ) = ∑ ikm cSUBJi c v ikmc OBJ m −→sk
where the verb v is represented in its non-compact form. If we introduce the compactness given to us by our isomorphism ST ∼= N ⊗ N we can express this as
−−−−−−−−→ SUBJ v OBJ = ∑ im cSUBJi c v imc OBJ m −→ni ⊗−→nm
where v is represented in its compact form. Furthermore, by introducing the component wise multiplication operation :
−→u −→v = ∑
i
cui c v i −→ni
we can show the general form of transitive verb composition with the reduced verb representation to be as follows:
−−−−−−−−→ SUBJ v OBJ = ∑ im cSUBJi c v imc OBJ m −→ni ⊗−→nm
= (∑ im cvim −→ni ⊗−→nm ) (∑ im cSUBJi c OBJ m −→ni ⊗−→nm )
= −→v (−−−→ SUBJ ⊗−−→OBJ )
To summarize what we have done with transitive verbs:
1. We have treated them as functions, taking two nouns and returning a sentence in a space ST ∼= N ⊗ N.
2. We have built them by counting what properties of subject and object noun vectors are related by the verb in transitive sentences in a corpus.
3. We have assumed that output properties were a function of input properties by the use of the Δijk function—for example, the weights associated with ni from the subject argument and with nk from the object argument only affect the “output” weight for the sentence basis element −→sj = −→ni ⊗−→nk .
4. We have shown that this leads to a compact representation of the verb’s semantic form, and an optimized learning procedure.
5. We have shown that the composition operations of our formalism can be adapted to this compact representation.
The compact representation and amended composition operation hinge on the choice of ST ∼= N ⊗ N as output type for N ⊗ N as an input type (the pair of arguments the verb takes), justifying our choice of a transitive sentence space. In the intransitive case, the same phenomenon can be observed, since such a verb takes as argument a vector in N and produces a vector in SI ∼= N. Furthermore, our choice to make all other types dependent on one base type—namely, n (with associated semantic space N)— yields the same property for every relational word we wish to learn: The output type is the same as the concatenated (on the syntactic level) or tensored (on the semantic level) input types. It is this symmetry between input and output types that guarantees that any m-ary relation, expressed in the original formulation as an element of tensor space N ⊗ . . .⊗ N︸ ︷︷ ︸
2m
, has a compact representation in N ⊗ . . .⊗ N︸ ︷︷ ︸ m , where the ith basis weight
of the reduced representation stands for the degree to which the ith element of the input vector affects the ith element of the output vector.
This allows the specification of the generalized learning algorithm for reduced representations, first presented in Grefenstette and Sadrzadeh (2011a), which is as follows. Each relational word P with grammatical type π and m adjoint types α1,α2, · · ·,αm is encoded as an (r × . . .× r) multi-dimensional array with m degrees of freedom (i.e., a rank m tensor). Because our vector space N has a fixed basis, each such array is represented in vector form as follows:
−→ P = ∑ ij · · ·ζ︸ ︷︷ ︸
m
cij···ζ ( −→n i ⊗−→n j ⊗ · · · ⊗ −→n ζ)︸ ︷︷ ︸
m
This vector lives in the tensor space N ⊗ N ⊗ · · · ⊗ N︸ ︷︷ ︸ m . Each cij···ζ is computed according to the procedure described in Figure 2.
Linear algebraically, this procedure corresponds to computing the following
−→ P = ∑ k (−→w1 ⊗−→w2 ⊗ · · · ⊗ −→wm)k
The general formulation of composing a relational word P with its arguments arg1, . . . , argm is now expressed as
−→ P (−−→arg1 ⊗ . . .⊗−−→argm)
For example, the computation of furry cats nag angry dogs would correspond to the following operations:
−−−−−−−−−−−−−−−−−−→ furry cats nag angry dogs = −−→nag ((−−−→ furry −→cat ) ⊗ (−−−→angry −−→dog))
This generalized learning algorithm effectively extends the coverage of our approach to any sentence for which we have a pregroup parse (e.g., as might be obtained by our translation mechanism from CCG). For example, determiners would have a type nnl, allowing us to model them as matrices in N ⊗ N; adverbs would be of type (nrs)r(nrs), and hence be elements of S ⊗ N ⊗ N ⊗ S. We could learn them using the procedure defined earlier, although for words like determiners and conjunctions, it is unclear whether we would want to learn such logical words or design them by hand, as was done in Coecke, Sadrzadeh, and Clark (2010) for negation. Nonetheless, the procedure given here allows us to generalize the work presented in this article to sentences of any length or structure based on the pregroup types of the words they contain. To conclude this section with an example taken from Grefenstette and Sadrzadeh (2011a), we demonstrate how the meaning of the word show might be learned from a corpus and then composed.
Suppose there are two instances of the verb show in the corpus: s1 = table show result s2 = map show location
The vector of show is
−−−→ show = −−→ table ⊗−−−→result + −−→map ⊗−−−−−→location
Consider a vector space N with four basis vectors far, room, scientific, and elect. The TF/IDF-weighted values for vectors of selected nouns (built from the BNC) are as shown in Table 1. Part of the matrix compact representation of show is presented in Table 2.
As a sample computation, the weight c11 for vector ( −→n1,−→n1 ), that is, ( −→ far, −→ far), is computed by multiplying weights of table and result on −→ far (6.6 × 7), multiplying weights of map and location on −→ far (5.6 × 5.9), and then adding these (46.2 + 33.04) and obtaining the total weight 79.24. Similarly, the weight c21 for vector ( −→n2,−→n1 ), that is, (−−−→room, −→ far), is computed by multiplying the weight of −−−→room for table by the weight of −→far for result (27 × 7), then multiplying the weight of −−−→room for map by the weight of −→far for location (7.4 × 5.9), and then adding these (189 + 43.66) to obtain the total weight of 232.66.
We now wish to compute the vector for the sentence [the] master shows [his] dog, omitting the determiner and possessive for simplicity, as we have left open the question as to whether or not we would want to learn them using the generalized procedure from Section 4.3 or specify them by design due to their logical structure. The calculation according to the vectors in Table 1 and the matrix in Table 2 will be:
−−−−−−−−−−−−→ master show dog
= −−−→ show (−−−−→ master ⊗−−→dog )
= ⎡ ⎢⎢⎣ 79.24 47.41 119.96 27.72 232.66 80.75 396.14 113.2 32.94 31.86 32.94 0
0 0 0 0
⎤ ⎥⎥⎦ ⎡ ⎢⎢⎣
23.65 32.68 0 0 50.05 69.16 0 0 22.55 31.16 0 0 34.1 47.12 0 0
⎤ ⎥⎥⎦
= ⎡ ⎢⎢⎣ 1,874.03 1,549.36 0 0 11,644.63 5,584.67 0 0
742.80 992.76 0 0 0 0 0 0
⎤ ⎥⎥⎦
Row-wise, flattening the final matrix representation gives us the result we seek, namely, the sentence vector in ST for [the] master showed [his] dog:
−−−−−−−−−−−−→ master show dog = [1,874, 1,549, 0, 0, 11,645, 5,585, 0, 0, 743, 993, 0, 0, 0, 0, 0, 0] 5 experiments :Evaluating compositional models of semantics is no easy task: First, we are trying to evaluate how well the compositional process works; second, we also are trying to determine how useful the final representation—the output of the composition—is, relative to our needs.
The scope of this second problem covers most phrase and sentence-level semantic models, from bag-of-words approaches in information retrieval to logic-based formal semantic models, via language models used for machine translation. It is heavily task dependent, in that a representation that is suitable for machine translation may not be appropriate for textual inference tasks, and one that is appropriate for IR may not be ideal for paraphrase detection. Therefore this aspect of semantic model evaluation ideally should take the form of application-oriented testing. For instance, to test semantic representations designed for machine translation purposes, we should use a machine translation evaluation task.
The DisCoCat framework (Clark, Coecke, and Sadrzadeh 2008; Coecke, Sadrzadeh, and Clark 2010), described in Section 3, allows for the composition of any words of any syntactic type. The general learning algorithm presented in Section 4.3 technically can be applied to learn and model relations of any semantic type. However, many open questions remain, such as how to deal with logical words, determiners, and quantification, and how to reconcile the different semantic types used for sentences with transitive and intransitive sentences. We will leave these questions for future work, briefly discussed in Section 6. In the meantime, we concretely are left with a way of satisfactorily modeling only simple sentences without having to answer these bigger questions.
With this in mind, in this section we present a series of experiments centered around evaluating how well various models of semantic vector composition perform (along with the one described in Section 4) in a phrase similarity comparison task. This task aims to test the quality of a compositional process by determining how well it forms a clear joint meaning for potentially ambiguous words. The intuition here is that tokens, on their own, can have several meanings; and that it is through the compositional process—through giving them context—that we understand their specific meaning. For example, bank itself could (among other meanings) mean a river bank or a financial bank; yet in the context of a sentence such as The bank refunded the deposit, it is likely we are talking about the financial institution.
In this section, we present three data sets designed to evaluate how well word-sense disambiguation occurs as a byproduct of composition. We begin by describing the first data set, based around noun-intransitive verb phrases, in Section 5.1. In Section 5.2, we present a data set based around short transitive-verb phrases (a transitive verb with subject and object). In Section 5.3, we discuss a new data set, based around short transitive-verb phrases where the subject and object are qualified by adjectives. We leave discussion of these results for Section 6. This first experiment, originally presented in Mitchell and Lapata (2008), evaluates the degree to which an ambiguous intransitive verb (e.g., draws) is disambiguated by combination with its subject.
Data set Description. The data set4 comprises 120 pairs of intransitive sentences, each of the form NOUN VERB. These sentence pairs are generated according to the following
4 Available at http://homepages.inf.ed.ac.uk/s0453356/results.
procedure, which will be the basis for the construction of the other data sets discussed subsequently:
1. A number of ambiguous intransitive verbs (15, in this case) are selected from frequently occurring verbs in the corpus.
2. For each verb V, two mutually exclusive synonyms V1 and V2 of the verb are produced, and each is paired with the original verb separately (for a total of 30 verb pairs). These are generated by taking maximally distant pairs of synonyms of the verb on WordNet, but any method could be used here.
3. For each pair of verb pairs (V, V1) and (V, V2), two frequently occurring nouns N1 and N2 are picked from the corpus, one for each synonym of V. For example, if V is glow and the synonyms V1 and V2 are beam and burn, we might choose face as N1, because a face glowing and a face beaming mean roughly the same thing; and fire as N2, because a fire glowing and a fire burning mean roughly the same thing.
4. By combining the nouns with the verb pairs, we form two high similarity triplets (V, V1, N1) and (V, V2, N2), and two low similarity triplets (V, V1, N2) and (V, V2, N1).
The last two steps can be repeated to form more than four triplets per pair of verb pairs. In Mitchell and Lapata (2008), eight triplets are generated for each pair of verb pairs, obtaining a total of 120 triplets from the 15 original verbs. Each triplet, along with its HIGH or LOW classification (based on the choice of noun for the verb pair) is an entry in the data set, and can be read as a pair of sentences: (V, Vi, N) translates into the intransitive sentences N V and N Vi.
Finally, the data set is presented, without the HIGH/LOW ratings, to human annotators. These annotators are asked to rate the similarity of meaning of the pairs of sentences in each entry on a scale of 1 (low similarity) to 7 (high similarity). The final form of the data set is a set of lines each containing:
Sentence 1 Sentence 2
butler bow butler submit head bow head stoop company bow company submit government bow government stoop
Evaluation Methodology. This data set is used to compare various compositional distributional semantic models, according to the following procedure:
1. For each entry in the data set, the representation of the two sentences N V and N Vi formed from the entry triple (V, Vi, N), which we will name S1 and S2, is constructed by the model.
2. The similarity of these sentences according to the model’s semantic distance measure constitutes the model score for the entry.
3. The rank correlation of entry model scores against entry annotator scores is calculated using Spearman’s rank correlation coefficient ρ.
The Spearman ρ scores are values between −1 (perfect negative correlation) and 1 (perfect correlation). The higher the ρ score, the higher the compositional model can be said to produce sentence representations that match human understanding of sentence meaning, when it comes to comparing the meaning of sentences. As such, we will rank the models evaluated using the task by decreasing order of ρ score.
One of the principal appealing features of Spearman’s ρ is that the coefficient is rank-based: It does not require models’ semantic similarity metrics to be normalized for a comparison to be made. One consequence is that a model providing excellent rank correlation with human scores, but producing model scores on a small scale (e.g., values between 0.5 and 0.6), will obtain a higher ρ score than a model producing model scores on a larger scale (e.g., between 0 and 1) but with less perfect rank correlation. If we wished to then use the former model in a task requiring some greater degree of numerical separation (let us say 0 for non-similar sentences and 1 for completely similar sentences), we could simply renormalize the model scores to fit the scale. By eschewing score normalization as an evaluation factor, we minimize the risk of erroneously ranking one model over another.
Finally, in addition to computing the rank alignment coefficient between model scores and annotator scores, Mitchell and Lapata (2008) calculate the mean model scores for entries labeled HIGH, and for entries labeled LOW. This information is reported in their paper as additional means for model comparison. However, for the same reason we considered Spearman’s ρ to be a fair means of model comparison— namely, in that it required no model score normalization procedure and thus was less likely to introduce error by adding such a degree of freedom—we consider the HIGH/LOW means to be inadequate grounds for comparison, precisely because it requires normalized model scores for comparison to be meaningful. As such, we will not include these mean scores in the presentation of this, or any further experiments in this article.
Models Compared. In this experiment, we compare the best and worst performing models of Mitchell and Lapata (2008) to our own. We begin by building a vector space W for all words in the corpus, using standard distributional semantic model construction procedures (n-word window) with the parameters of Mitchell and Lapata (2008). These parameters are as follows: The basis elements of this vector space are the 2,000 most frequently occurring words in the BNC, excluding common stop words. For our evaluation, the corpus was lemmatized using Carroll’s Morpha (Minnen, Carroll, and Pearce 2001), which was applied as a byproduct of our parsing of the BNC with C&C Tools (Curran, Clark, and Bos 2007).
The context of each occurrence of a word in the corpus was defined to be five words on either side. After the vector for each word w was produced, the basis weights cwi associated with context word bi were normalized by the following ratio of probabilities weighting scheme:
cwi = P(bi|w) P(w)
Let −−→ verb and −−→noun be the lexical semantic vectors for the verb and the noun, from W. It should be evident that there is no significant difference in using W for nouns in lieu of building a separate vector space N strictly for noun vectors.
As a first baseline, Verb Baseline, we ignored the information provided by the noun in constructing the sentence representation, effectively comparing the semantic content of the verbs:
Verb Baseline : −−−−−−→ noun verb = −−→ verb
The models from Mitchell and Lapata (2008) we evaluate here are those which were strictly unsupervised (i.e., no free parameters for composition). These are the additive model Add, wherein
Add : −−−−−−→ noun verb = −−→noun +−−→verb
and the multiplicative model Multiply, wherein
Multiply : −−−−−−→ noun verb = −−→noun −−→verb
Other models with parameters that must be optimized against a held-out section of the data set are presented in Mitchell and Lapata (2008). We omitted them here principally because they do not perform as well as Multiply, but also because the need to optimize the free parameters for this and other data sets makes fair comparison with completely unsupervised models more difficult, and less fair.
We also evaluate the Categorical model from Section 4 here, wherein
Categorical : −−−−−−→ noun verb = −−−→ verbcat −−→noun
where −−−→ verbcat is the compact representation of the relation the verb stands for, computed according to the procedure described in Section 4.3. It should be noted that for the case of intransitive verbs, this composition operation is mathematically equivalent to that of the Multiply model (as component-wise multiplication is a commutative operation), the difference being the learning procedure for the verb vector.
For all such vector-based models, the similarity of the two sentences s1 and s2 is taken to be the cosine similarity of their vectors, defined in Section 2.2:
similarity(s1, s2) = cosine( −→s1 ,−→s2 )
In addition to these vector-based methods, we define an additional baseline and an upper-bound. The additional baseline, Bigram Baseline, is a bigram-based language
model trained on the BNC with SRILM (Stolcke 2002), using the standard language model settings for computing log-probabilities of bigrams. To determine the semantic similarity of sentences s1 = noun verb1 and s2 = noun verb2, we assumed sentences have mutually conditionally independent properties, and computed the joint probability:
similarity(s1, s2) = logP(s1 ∧ s2) = log(P(s1)P(s2)) = logP(s1) + logP(s2)
The reasoning behind this baseline is as follows. The sentence formed by combining the first verb with its arguments is, by the design of this data set, a semantically coherent sentence (e.g., the head bowed and the government bowed both make sense). We therefore expect language models to treat this sort of bigram as having a higher probability than bigrams that are not semantically coherent sentences (and therefore unlikely to be observed in the corpus). A similar (relative) high probability is to be expected when the sentence formed by taking the second verb in each entry and combining it with the verb yields a sentence similar to the first in meaning (e.g., as would be the case with the head stooped), whereas the probability of a semantically incoherent sentence (e.g., the government stooped) is expected to be low relative to that of the first sentence. By taking the sum of log probabilities of sentences, we compute the log of the product of the probabilities, which we expect to be low when the probability of the second sentence is low, and high when it is high, with the probability of the first sentence acting as a normalizing factor. To summarize, while this bigram-based measure only tracks a tangential aspect of semantic similarity, it can be one which plays an artificially important role in experiments with a predetermined structure such as the one described in this section. For this reason, we use this bigram baseline for this experiment and all that follow.
The upper bound of the data set, UpperBound, was taken to be the inter-annotator agreement: the average of how each annotator’s score aligns with other annotator scores, using Spearman’s ρ.
Results. The results5 of the first experiment are shown in Table 4. As expected from the fact that Multiply and Categorical differ only in how the verb vector is learned, the results of these two models are virtually identical, outperforming both baselines and the Additive model by a significant margin. However, the distance from these models to the upper bound is even greater, demonstrating that there is still a lot of progress to be made. The first experiment was followed by a second similar experiment, in Mitchell and Lapata (2010), covering different sorts of composition operations for binary combinations of syntactic types (adjective-noun, noun-noun, verb-object). Such further experiments are interesting, but rather than continue down this binary road, we now turn to the development of our second experiment, involving sentences with larger syntactic structures, to examine how well various compositional models cope with more complex syntactic and semantic relations.
5 The results were state of the art when the experiments were run in 2011. The binary composition data set used here are popular; now there are a host of new state-of-the-art systems available in the literature. We invite the reader to check references to Mitchell and Lapata (2008) to find the current state of the art.
This second experiment, which we initially presented in Grefenstette and Sadrzadeh (2011a), is an extension of the first in the case of sentences centered around transitive verbs, composed with a subject and an object. The results of the first experiment did not demonstrate any difference between the multiplicative model, which takes into account no syntactic information or word ordering, and our syntactically motivated categorical compositional model. By running the same experiment over a new data set, where the relations expressed by the verb have a higher arity than in the first, we hope to demonstrate that added structure leads to better results for our syntax-sensitive model.
Data Set Description. The construction procedure for this data set6 is almost exactly as for the first data set, with the following differences:
Verbs are transitive instead of intransitive. We arbitrarily took 10 verbs from the most frequent verbs in the BNC, and for each verb, took two maximally distant synonyms (again, using WordNet) to obtain 10 pairs of pairs.
For each pair of verb pairs, we selected a set of subject and object nouns to use as context, as opposed to just a subject noun.
Each subject-object pair was manually chosen so that one of the verb pairs would have high similarity in the context of that subject and object, and the other would have low similarity. We used these choices to annotate the entries with HIGH and LOW tags.
Each combination of a verb pair with a subject-object pair constitutes an entry of our data set, of which there are 200.
As a form of quality control, we inserted “gold standard” sentences in the form of identical sentence pairs and rejected annotators who did not score these gold standard
6 The data set, reannotated by Turkers in 2013, is available at http://www.cs.ox.ac.uk/activities/compdistmeaning/GS2013data.txt.
sentences with a high score of 6 or 7.7 Lemmatized sentences from sample entries of this data set and our HIGH-LOW tags for them are shown in Table 5.
The data set was passed to a group of 50 annotators on Amazon Mechanical Turk, as for the previous data set. The annotators were shown, for each entry, a pair of sentences created by adding the in front of the subject and object nouns and putting the verbs in the past tense,8 and were instructed to score each pair of sentences on the same scale of 1 (not similar in meaning) to 7 (similar in meaning), based on how similar in meaning they believed the sentence pair was.
Evaluation Methodology. The methodology for this experiment is exactly that of the previous experiment. Models compositionally construct sentence representations, and compare them using a distance metric (all vector-based models once again used cosine similarity). The rank correlation of models scores with annotator scores is calculated using Spearman’s ρ, which is in turn used to rank models.
Models Compared. The models compared in this experiment are those of the first experiment, with the addition of an extra trigram-based baseline (trained with SRILM, using the addition of log-probability of the sentence as a similarity metric), and a variation on our categorical model, presented subsequently. With W as the distributional semantic space for all words in the corpus, trained using the same parameters as in the first experiment, and −−→ subj, −−→ verb, −−−→ object ∈ W as the vectors for subject, verb, and object of a sentence, respectively, and with verbcat as the compact representation of a transitive verb learned using the algorithm presented in this paper, we have the following compositional methods:
Add : −−−−−−−−−−−−→ subject verb object = −−−−→ subject + −−→ verb + −−−→ object Multiply : −−−−−−−−−−−−→ subject verb object = −−−−→ subject −−→verb −−−→object Categorical : −−−−−−−−−−−−→ subject verb object = −−−→ verbcat (−−−−→ subject ⊗−−−→object
) Kronecker : −−−−−−−−−−−−→ subject verb object = (−−→ verb ⊗−−→verb ) (−−−−→ subject ⊗−−−→object )
7 During the 2013 reannotation of this data set, we rejected no Turker contributions, as the answers to the gold standard sentence pairs were aligned with our expectations. We attribute this to our adding the requirement that Turkers be based in the US or the UK and have English as the first language. 8 For example, the entry draw table eye depict would yield the sentences The table drew the eye and The table depicted the eye.
All of these models have been explained earlier, with the exception of Kronecker. We first presented this new addition in Grefenstette and Sadrzadeh (2011b), where we observed that the compact representation of a verb in the DisCoCat framework, under the assumptions presented in Section 4, can be viewed as dim(N) × dim(N) matrices in N ⊗ N. We considered alternatives to the algorithm presented earlier for the construction of such matrices, and were surprised by the results of the Kronecker method, wherein we replaced the matrix learned by our algorithm with the Kronecker product of the lexical semantic vectors for the verb. Further analysis performed since the publication of that paper can help to understand why this method might work. Using the following property, for any vectors −→a ,−→b ,−→c ,−→d in a vector space
(−→a ⊗−→b ) (−→c ⊗−→d ) = (−→a −→c ) ⊗ (−→b −→d )
we can see that the Kronecker model’s composition operation can be expressed as
Kronecker : −−−−−−−−−−−−→ subject verb object = (−−→ verb −−−−→subject ) ⊗ (−−→ verb −−−→object )
Bearing in mind that the cosine measure we are using as a similarity metric is equivalent to the inner product of two vectors normalized by the product of their length
cosine(−→a ,−→b ) = 〈 −→a |−→b 〉
‖−→a ‖ × ‖−→b ‖
and the following property of the inner product of kronecker products
〈−→a ⊗−→b |−→c ⊗−→d 〉 = 〈−→a |−→c 〉 × 〈−→b |−→d 〉
we finally observe that comparing two sentences under Kronecker corresponds to the following computation:
cosine( −−−−−−−−−−−−→ subject verb1 object, −−−−−−−−−−−−→ subject verb2 object)
= α 〈(−−→ verb1 ⊗ −−→ verb1 ) (−−−−→ subject ⊗−−−→object ) | (−−→ verb2 ⊗ −−→ verb2 ) (−−−−→ subject ⊗−−−→object )〉
= α 〈(−−→ verb1 −−−−→ subject ) ⊗ (−−→ verb1 −−−→ object ) | (−−→ verb2 −−−−→ subject ) ⊗ (−−→ verb2 −−−→ object )〉 = α 〈(−−→ verb1 −−−−→ subject ) | (−−→ verb2 −−−−→ subject )〉〈(−−→ verb1 −−−→ object ) | (−−→ verb2 −−−→ object
)〉 where α is the normalization factor
α = 1 ‖−−−−−−−−−−−−→subject verb1 object‖ × ‖ −−−−−−−−−−−−→ subject verb2 object‖
We note here that the Kronecker is effectively a parallel application of the Multiply model, combining the subject and verb, and object and verb separately. Within the context of this task, the comparison of two sentences boils down to how well each verb combines with the subject multiplied by how well it combines with the object, with the
joint product forming the overall model score. In short, it more or less constitutes the introduction of some mild syntactic sensitivity into the multiplicative model of Mitchell and Lapata (2008), although it remains to be seen how well this scales with syntactic complexity (e.g., extended to ditransitive verbs, or other relations of higher arity).
The UpperBound here is, again, the inter-annotator agreement, calculated by computing the pairwise alignment of each annotator with every other, and averaging the resulting rank correlation coefficients.
Results. The results of the second experiment are shown in Table 6. The baseline scores fall in the 0.14–0.16 range, with best results being obtained for the Bigram Baseline and Trigram Baseline, the difference between both models not being statistically significant. The additive model Add performs on par with the first experiment. The multiplicative model Multiply and our Categorical model perform on par with the version used for the intransitive experiment, but obtain a score comparable to the baselines. The best performing model here is our newly introduced Kronecker model, which leads the pack by a steady margin, with a score of 0.26. The inter-annotator agreement UpperBound is much higher in this experiment than in the previous experiment, indicating even more room for improvement.
We therefore learn here that the Categorical model continues to operate on par with the multiplicative model, and that the Kronecker model provides the highest results, while being as simple in its construction as the multiplicative model, requiring only the learning of lexical vectors for the verb. The third and final experiment we present is a modified version of the second data set presented earlier, where the nouns in each entry are under the scope of adjectives applied to them.The intuition behind the data sets presented in Section 5.1 and Section 5.2 was that ambiguous verbs are disambiguated through composition with nouns. These nouns themselves may also be ambiguous, and a good compositional model will be capable of separating the noise produced by other meanings through its compositional mechanism to produce unambiguous phrase representations. The intuition behind this data set is similar, in that adjectives provide both additional information for disambiguation of the nouns they apply to, but also additional semantic noise. Therefore, a
good model will also be able to separate the useful information of the adjective from its semantic noise when composing it with its argument, in addition to doing this when composing the noun phrases with the verb.
Data Set Description. The construction procedure for this data set was to take the data set from Section 5.2, and, for each entry, add a pair of adjectives from those most frequently occurring in the corpus. The first adjective from the pair is applied to the first noun (subject) of the entry when forming the sentences, and the second adjective is applied to the second noun (object). For each entry, we chose adjectives which best preserved the meaning of the phrase constructed by combining the first verb with its subject and object.
This new data set9 was then annotated again by a group of 50 annotators using Amazon’s Mechanical Turk service. The annotators were shown, for each entry, a pair of sentences created by adding the in front of the subject and object noun phrases and putting the verbs in the past tense,10 and asked to give each sentence pair a meaning similarity score between 1 and 7, as for the previous data sets. We applied the same quality control mechanism as in the second experiment. Some 94 users returned annotations, of which we kept 50 according to our gold standard tests. We are unaware of whether or not it was applied in the production of the first data set, but believe that this can only lead to the production of higher quality annotations.
Sample sentences from this data set are shown in Table 7.
Evaluation Methodology. The evaluation methodology in this experiment is identical to that of the previous experiments.
Models Compared. In this experiment, in lieu of simply comparing compositional models “across the board” (e.g., using the multiplicative model for both adjective-noun composition and verb-argument composition), we experimented with different combinations of models. This evaluation procedure was chosen because we believe that adjective-noun composition need not necessarily be the same kind of compositional process as subjectverb-object composition, and also because different models may latch onto different semantic features during the compositional process, and it would be interesting to see what model mixtures work well together. Naturally, this is not a viable approach to selecting composition operations in general, as we will not have the luxury of trying every combination of composition operations for every combination of syntactic types, but it is worthwhile performing these tests to at least verify the hypothesis that
9 Available at http://www.cs.ox.ac.uk/activities/compdistmeaning/GS2012data.txt. 10 For example, the entry statistical table show express good result would yield the sentences The statistical table
showed the good result and The statistical table expressed the good result.
operation-specific composition operations (or parameters) is a good thing. Notably, this idea has been explored very recently, within the context of deep learning networks, by Chen et al. (2013).
Each mixed model has two components: a verb-argument composition model and a adjective-noun composition model. For verb-argument composition, we used the three best models from the previous experiment, namely, Multiply, Categorical, and Kronecker. For adjective composition we used three different methods of adjectivenoun composition. With −−−−−→ adjective and −−→noun being the vectors for an adjective and a noun in our distributional lexical semantic space W (built using the same procedure as the previous experiments) and −−→ adjcat being the compact representation in the Categorical model, built according to the algorithm from Section 4.3, we have the following models:
AdjMult : −−−−−−−−−→ adjective noun = −−−−−→ adjective −−→noun Categorical : −−−−−−−−−→ adjective noun = −−−−−−→ adjectivecat −−→noun
The third model, AdjNoun, is a holistic (non-compositional) model, wherein the adjective-noun compound was treated as a single token, as its semantic vector−−−−−−−−−−−−→ (adjective noun)lex ∈ W was learned from the corpus using the same learning procedure applied to construct other vectors in W. Hence, the model defines adjective-noun “composition” as:
AdjNoun : −−−−−−−−−→ adjective noun = −−−−−−−−−−−−→ (adjective noun)lex
In addition to these models, we also evaluated three baselines: Verb Baseline, Bigram Baseline, and Trigram Baseline. As in previous experiments, the verb baseline uses the verb vector as sentence vector, ignoring the information provided by other words. The bigram and trigram baselines are calculated from the same language model as used in the second experiment. In both cases, the log-probability of each sentence is calculated using SRLIM, and the sum of log-probabilities of two sentences is used as a similarity measure.
Finally, for comparison, we also considered the full additive model:
Additive : −−−−−→ sentence = −−−−−→ adjective1 + −−−→noun1 + −−→ verb + −−−−−→ adjective2 + −−−→noun2
Results. The results for the third experiment are shown in Table 8. The best performing adjective-noun combination operations for each verb-argument combination operation are shown in bold. Going through the combined models, we notice that in most cases the results stay the same whether the adjective-noun combination method is AdjMult or CategoricalAdj. This is because, as was shown in the first experiment, composition of a unary-relation such as an adjective or intransitive verb with its sole argument, under the categorical model with reduced representations, is mathematically equivalent to the multiplicative model. The sole difference is the way the adjective or intransitive verb vector is constructed. We note, however, that with Categorical as a verb-argument composition method, the CategoricalAdj outperforms AdjMult by a non-negligible margin (0.19 vs. 0.14), indicating that the difference in learning procedure can lead to different results depending on what other models it is combined with.
Overall, the best results are obtained for AdjMult+Kronecker (ρ = 0.26) and CategoricalAdj+Kronecker (ρ = 0.27). Combinations of the adjective composition methods with other composition methods at best matches the best-performing baseline, Verb Baseline. In all cases, the holistic model AdjNoun provides the worst results. 6 discussion :In this article, we presented various approaches to compositionality in distributional semantics. We discussed what mathematical operations could underlie vector combination, and several different ways of including syntactic information into the combinatory process. We reviewed an existing compositional framework that leverages the ability to communicate information across mathematical structures provided by category theory in order to define a general way of linking syntactic structure to syntax-sensitive composition operations. We presented concrete ways to apply this framework to linguistic tasks and evaluate it, and developed algorithms to construct semantic representations for words and relations within this framework. In this section, we first briefly comment upon the combined results of all three experiments, and then conclude by discussing what aspects of compositionality require further attention, and how experiment design should adapt towards this goal.
Results Commentary. We evaluated this framework against other unsupervised compositional distributional models, using non-compositional models and n-gram language models as baselines, within the context of three experiments. These experiments show that the concrete categorical model developed here, and the Kronecker-based variant
presented alongside it, outperform all other models in each experiment save the first, where they perform on par with what was the leading model at the time this experiment was performed (namely, 2011). As the experiments involved progressively more syntactically complicated sentences, the increased reliability of our categorical approaches relative to competing models as sentence complexity rises seems to indicate that both the categorical and Kronecker model successfully leverage the added information provided by additional terms and syntactic structures.
The third experiment also served to show that using different combinations of composition operations, depending on the syntactic type of the terms being combined, can yield better results, and that some models combine better than others. Notably, the adjective-noun combination models AdjMult and CategoricalAdj, despite their mathematical similarity, produce noticeably different results when combined with the categorical verb-argument composition operation, while they perform equally with most other verb-argument composition operations. We can conclude that different models combine different semantic aspects more prominently than others, and that through combination we can find better performance by assuming that different kinds of composition play on different semantic properties. For example, predicates such as intersective adjectives add information to their argument (a red ball is a ball that is also red). This raises the question of how to design models of composition that systematically select which operations will match the semantic aspects of the words being combined based on their syntactic type. This is an open question, which we believe warrants further investigation.
Overall, we observe two advantages of these models over those presented in the early part of this article, and those evaluated in the later part. First, we have shown that a linguistically motivated and mathematically sound framework can be implemented and learned. Second, we showed that simple methods such as the Kronecker model perform well, especially when combined with the multiplicative model (effectively a unary instance of the Kronecker model) for sentences with adjectives. Considering the “simple is best” position recently argued for experimentally by Blacoe and Lapata (2012), this is an interesting candidate for dealing with binary relation composition.
Future Work. These experiments showcased the ability of various compositional models to produce sentence representations that were less ambiguous than the words that formed them. They also more generally demonstrated that concrete models could be built from the general categorical framework and perform adequately in simple paraphrase detection tasks. However, various aspects of compositionality were not evaluated here. First, syntax sensitivity is not as important in these experiments as it might be in “real” language. For instance, whereas models that treat adjectives, nouns, and verbs differently tended to perform better than those that did not, the actual capacity of a model to use the syntactic structure was not tested. For instance, models that ignore word order and syntax altogether were not particularly penalized, as might be done if some sentences were simply the reverse of another sentence they are paired with (where syntax insensitive models would erroneously give such sentence pairs a high similarity score). One way of testing word order while expanding the data set was suggested by Turney (2012), who for every entry in a phrase similarity test such as Mitchell and Lapata’s, discussed herein, creates a new entry where one of the sentences has the order of its words reversed, and is assigned an artificial annotator score of 1 (the minimum). This is an interesting approach, but we would prefer to see such reversals designed into the data set and seen by annotators, for instance,
in the recent data set presented in Pham et al. (2013). This data set has entries such as guitar played man and man played guitar, where the subjects and objects of verbs are indeed reversed but the two sentences do not necessarily express opposite meanings; we will be looking into expanding such data sets to ones where both sentences make sense and have opposite meanings, such as in the pair man bites dog and dog bites man.
Furthermore, sentences all have the exact same sentence structure, and therefore words are aligned. Models that might indirectly benefit from this alignment, such as Kronecker, may have been given an advantage due to this, as opposed to models such as Categorical, which are designed to compare sentences of different length or syntactic structure, when resolving the difference between intransitive and transitive sentence spaces as has been done in Grefenstette et al. (2011). Future experiments should aim to do away with this automated alignment, and include sentence comparisons that would penalize models which do not leverage syntactic information.
Furthermore, each of these experiments dealt with the comparison of a certain type of sentence (transitive, intransitive). This was convenient for our concrete categorical model, as we defined the sentence space differently based on the valency of the head verb. However, sentences with different sorts of verbs should be able to be directly compared. Not only do several models, both non-syntax sensitive (additive, multiplicative) and syntax-sensitive (Baroni and Zamparelli 2010; Socher et al. 2012), not face this problem, as the product of composition is either naturally in the same space as the vectors being composed or is projected back into it, but the categorical framework the concrete categorical models were derived from does not commit us to different sentence spaces either. The topic of how to solve this problem for the concrete models developed here is beyond the scope of this article, but various options exist, such as the projection of ST into SI, the embedding of SI into ST, the use of a combined sentence space S = SI ⊕ ST, and so on. It is clear that future experiments should not give this degree of convenience to the models being evaluated (and their designers), by comparing sentences of different types, lengths, and structure so as to more fairly evaluate the capacity of models to produce meaningful semantic representations in a single space, which can be directly compared regardless of syntactic structure and verb-type.
Finally, we mentioned at the beginning of Section 5 that evaluations need to be application-oriented. The class of experiments presented in this article specifically see how well disambiguation occurs as a byproduct of composition. We presented concrete models that would offer ways of combining relations underlying verbs and adjectives, and tested how well these relations disambiguated their arguments and were in turn disambiguated by their arguments. We did not address other aspects of language, such as quantification, logical operations, relative clauses, intensional clauses, embedded sentences, and many other linguistic aspects of spoken and written language that have complex syntactic structure and complicated semantic roles. Addressing these sorts of problems requires determining how negation, conjunction, and other logical relations should be modeled within compositional distributional formalisms, a problem we share with other similar approaches. The development of newer models within the categorical framework we based our work on, and the further development of other approaches mentioned here, must be driven by new experimental goals different from those offered by the task and experiments discussed in this article. Thinking not only about how to address these aspects of language within compositional models, but how to evaluate them, should be a priority for all those interested in further developing this field. Modeling compositional meaning for sentences using empirical distributional methods has been a challenge for computational linguists. The categorical model of Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) provides a solution by unifying a categorial grammar and a distributional model of meaning. It takes into account syntactic relations during semantic vector composition operations. But the setting is abstract: It has not been evaluated on empirical data and applied to any language tasks. We generate concrete models for this setting by developing algorithms to construct tensors and linear maps and instantiate the abstract parameters using empirical data. We then evaluate our concrete models against several experiments, both existing and new, based on measuring how well models align with human judgments in a paraphrase detection task. Our results show the implementation of this general abstract framework to perform on par with or outperform other leading models in these experiments.1 [{""affiliations"": [], ""name"": ""Edward Grefenstette""}, {""affiliations"": [], ""name"": ""Google DeepMind""}, {""affiliations"": [], ""name"": ""Mehrnoosh Sadrzadeh""}] SP:2c6b675504d6903e070d034b855e39f296f5ccd4 [{""authors"": [""S. Abramsky"", ""B. Coecke.""], ""title"": ""A categorical semantics of quantum protocols"", ""venue"": ""Proceedings of the 19th Annual IEEE Symposium on Logic in Computer Science, pages 415\u2013425, Turku, Finland."", ""year"": 2004}, {""authors"": [""H. Alshawi""], ""title"": ""The Core Language Engine"", ""year"": 1992}, {""authors"": [""M. Baroni"", ""R. Zamparelli.""], ""title"": ""Nouns are vectors, adjectives are matrices: Representing adjective-noun constructions in semantic space"", ""venue"": ""Proceedings of the 2010 Conference on Empirical Methods in Natural"", ""year"": 2010}, {""authors"": [""D. B\u00e9chet"", ""A. Foret"", ""I. Tellier.""], ""title"": ""Learnability of Pregroup Grammars"", ""venue"": ""Studia Logica, 87(2\u20133), pages 225\u2013252."", ""year"": 2007}, {""authors"": [""W. Blacoe"", ""M. Lapata.""], ""title"": ""A comparison of vector-based representations for semantic composition"", ""venue"": ""Proceedings of the 2012 Conference on Empirical Methods in Natural Language"", ""year"": 2012}, {""authors"": [""W. Buszkowski""], ""title"": ""Lambek grammars based on pregroups"", ""venue"": ""P. de Groote, G. Morrill, and C. Retor\u00e9, editors, Logical Aspects of Computational Linguistics. Lecture Notes in Computer Science,"", ""year"": 2001}, {""authors"": [""Chen"", ""Danqi"", ""Richard Socher"", ""Christopher D. Manning"", ""Andrew Y. Ng.""], ""title"": ""Learning new facts from knowledge bases with neural tensor networks and semantic word vectors"", ""venue"": ""arXiv preprint"", ""year"": 2013}, {""authors"": [""S. Clark""], ""title"": ""Type-driven syntax and semantics for composing meaning vectors"", ""venue"": ""Quantum Physics and Linguistics: A Compositional, Diagrammatic Discourse. Oxford University Press."", ""year"": 2013}, {""authors"": [""S. Clark"", ""B. Coecke"", ""M. Sadrzadeh.""], ""title"": ""A compositional distributional model of meaning"", ""venue"": ""Proceedings of the Second Quantum Interaction Symposium (QI-2008), Oxford."", ""year"": 2008}, {""authors"": [""S. Clark"", ""J.R. Curran.""], ""title"": ""Wide-coverage efficient statistical parsing with CCG and log-linear models"", ""venue"": ""Computational Linguistics, 33:493\u2013552."", ""year"": 2007}, {""authors"": [""S. Clark"", ""S. Pulman.""], ""title"": ""Combining symbolic and distributional models of meaning"", ""venue"": ""AAAI Spring Symposium on Quantum Interaction, Stanford, USA."", ""year"": 2007}, {""authors"": [""Clarke"", ""Daoud.""], ""title"": ""Context-theoretic semantics for natural language: An overview"", ""venue"": ""Proceedings of the Workshop on Geometrical Models of Natural Language Semantics, pages 112\u2013119, Edinburgh."", ""year"": 2009}, {""authors"": [""Clarke"", ""Daoud.""], ""title"": ""A context-theoretic framework for compositionality in distributional semantics"", ""venue"": ""Computational Linguistics, 38(1):41\u201371."", ""year"": 2012}, {""authors"": [""B. Coecke"", ""\u00c9.O. Paquette.""], ""title"": ""Categories for the practising physicist"", ""venue"": ""B. Coecke, editor, New Structures for Physics. Lecture Notes in Physics, volume 813, pages 173\u2013286, Springer,"", ""year"": 2011}, {""authors"": [""B. Coecke"", ""M. Sadrzadeh"", ""S. Clark.""], ""title"": ""Mathematical Foundations for a Compositional Distributional Model of Meaning"", ""venue"": ""Linguistic Analysis, 36:345\u2013384."", ""year"": 2010}, {""authors"": [""Curran"", ""James"", ""Stephen Clark"", ""Johan Bos.""], ""title"": ""Linguistically motivated large-scale NLP with C&C and boxer"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational"", ""year"": 2007}, {""authors"": [""J.R. Curran""], ""title"": ""From distributional to semantic similarity"", ""venue"": ""Ph.D. thesis, School of Informatics, University of Edinburgh."", ""year"": 2004}, {""authors"": [""Dean"", ""Jeffrey"", ""Sanjay Ghemawat.""], ""title"": ""MapReduce: Simplified data processing on large clusters"", ""venue"": ""Communications of the ACM, 51(1):107\u2013113."", ""year"": 2008}, {""authors"": [""Erk"", ""Katrin"", ""Sebastian Pad\u00f3.""], ""title"": ""A structured vector space model for word meaning in context"", ""venue"": ""Proceedings of the Conference on Empirical Methods in Natural Language Processing - EMNLP \u201908,"", ""year"": 2008}, {""authors"": [""J.R. Firth""], ""title"": ""A synopsis of linguistic theory 1930-1955"", ""venue"": ""Studies in linguistic analysis."", ""year"": 1957}, {""authors"": [""T.A.D. Fowler"", ""G. Penn.""], ""title"": ""Accurate context-free parsing with combinatory categorial grammar"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 335\u2013344,"", ""year"": 2010}, {""authors"": [""G. Frege""], ""title"": ""Uber sinn und bedeutung"", ""venue"": ""Zeitschrift f\u00fcr Philosophie und philosophische Kritik, 100(1):25\u201350."", ""year"": 1892}, {""authors"": [""Genkin"", ""Daniel"", ""Nissim Francez"", ""Michael Kaminski.""], ""title"": ""Mildly context-sensitive languages via buffer augmented pregroup grammars"", ""venue"": ""Z. Manna and D. A. Peled, editors, Time for Verification, pages 144\u2013166,"", ""year"": 2010}, {""authors"": [""Girard"", ""Jean-Yves.""], ""title"": ""Linear logic"", ""venue"": ""Theoretical Computer Science, 50:1\u2013102."", ""year"": 1987}, {""authors"": [""E. Grefenstette""], ""title"": ""Analysing document similarity measures"", ""venue"": ""Master\u2019s thesis, University of Oxford."", ""year"": 2009}, {""authors"": [""E. Grefenstette"", ""G. Dinu"", ""Y. Zhang"", ""M. Sadrzadeh"", ""M. Baroni.""], ""title"": ""Multistep regression learning for compositional distributional semantics"", ""venue"": ""Proceedings of the Tenth International Conference on"", ""year"": 2013}, {""authors"": [""E. Grefenstette"", ""M. Sadrzadeh.""], ""title"": ""Experimental support for a categorical compositional distributional model of meaning"", ""venue"": ""Proceedings of the 2011 Conference on Empirical Methods"", ""year"": 2011}, {""authors"": [""E. Grefenstette"", ""M. Sadrzadeh.""], ""title"": ""Experimenting with transitive verbs in a DisCoCat"", ""venue"": ""Proceedings of the 2011 EMNLP Workshop on Geometric Models of Natural Language Semantics, pages 62\u201366,"", ""year"": 2011}, {""authors"": [""E. Grefenstette"", ""M. Sadrzadeh"", ""S. Clark"", ""B. Coecke"", ""S. Pulman.""], ""title"": ""Concrete sentence spaces for compositional distributional models of meaning"", ""venue"": ""Proceedings of the Ninth International"", ""year"": 2011}, {""authors"": [""G. Grefenstette""], ""title"": ""Use of syntactic context to produce term association lists for text retrieval"", ""venue"": ""Proceedings of the 15th Annual International ACM SIGIR Conference on Research and Development in Information"", ""year"": 1992}, {""authors"": [""G. Grefenstette""], ""title"": ""Explorations in automatic thesaurus discovery"", ""year"": 1994}, {""authors"": [""E. Guevara""], ""title"": ""A regression model of adjective-noun compositionality in distributional semantics"", ""venue"": ""Proceedings of the 2010 Workshop on Geometrical Models of Natural Language Semantics,"", ""year"": 2010}, {""authors"": [""Z.S. Harris""], ""title"": ""Mathematical structures of language"", ""venue"": ""Wiley."", ""year"": 1968}, {""authors"": [""Hockenmaier"", ""Julia.""], ""title"": ""Data and Models for Statistical Parsing with Combinatory Categorial Grammar"", ""venue"": ""Ph.D. thesis, School of Informatics, University of Edinburgh."", ""year"": 2003}, {""authors"": [""Hockenmaier"", ""Julia"", ""Mark Steedman.""], ""title"": ""CCGBank: A corpus of CCG derivations and dependency structures extracted from the Penn treebank"", ""venue"": ""Computational Linguistics, 33(3):355\u2013396."", ""year"": 2007}, {""authors"": [""A.K. Joshi"", ""K. Vijay-Shanker"", ""D.J. Weir.""], ""title"": ""The convergence of mildly context-sensitive grammar formalisms"", ""venue"": ""Working paper, University of Pennsylvania, School of Engineering"", ""year"": 1989}, {""authors"": [""J. Lambek""], ""title"": ""The mathematics of sentence structure"", ""venue"": ""The American Mathematical Monthly, 65(3):154\u2013170."", ""year"": 1958}, {""authors"": [""J. Lambek""], ""title"": ""Type grammar revisited"", ""venue"": ""Logical Aspects of Computational Linguistics, pages 1\u201327."", ""year"": 1999}, {""authors"": [""J. Lambek""], ""title"": ""From word to sentence"", ""venue"": ""A computational algebraic approach to grammar. Milan, Polimetrica."", ""year"": 2008}, {""authors"": [""J. Lambek""], ""title"": ""Compact Monoidal Categories from Linguistics to Physics, pages 451\u2013469"", ""year"": 2010}, {""authors"": [""T.K. Landauer"", ""S.T. Dumais""], ""title"": ""A solution to Plato\u2019s problem: The latent semantic analysis theory of acquisition, induction, and representation of knowledge"", ""year"": 1997}, {""authors"": [""S. Mac Lane""], ""title"": ""Categories for the Working Mathematician"", ""venue"": ""Springer Verlag."", ""year"": 1998}, {""authors"": [""C.D. Manning"", ""P. Raghavan"", ""H. Sch\u00fctze.""], ""title"": ""Introduction to information retrieval"", ""venue"": ""Cambridge University Press, New York, NY."", ""year"": 2011}, {""authors"": [""G. Minnen"", ""J. Carroll"", ""D. Pearce.""], ""title"": ""Applied morphological processing of English"", ""venue"": ""Natural Language Engineering, 7(03):207\u2013223."", ""year"": 2001}, {""authors"": [""J. Mitchell"", ""M. Lapata.""], ""title"": ""Vectorbased models of semantic composition"", ""venue"": ""Proceedings of ACL, volume 8, pages 236\u2013244, Columbus, OH. 117"", ""year"": 2008}, {""authors"": [""J. Mitchell"", ""M. Lapata.""], ""title"": ""Composition in Distributional Models of Semantics"", ""venue"": ""Cognitive Science 34(8):1388\u20131429."", ""year"": 2010}, {""authors"": [""R. Montague""], ""title"": ""English as a Formal Language"", ""venue"": ""R. H. Thomason, editor, Formal Semantics: The Essential Readings."", ""year"": 1974}, {""authors"": [""M. Moortgat""], ""title"": ""Categorial type logics"", ""venue"": ""H. van Ditmarsch and L. S. Moss, editors, Handbook of Logic and Language. Elsevier, pages 93\u2013177."", ""year"": 1997}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Mirella Lapata.""], ""title"": ""Dependency-based construction of semantic space models"", ""venue"": ""Computational Linguistics, 33(2):161\u2013199."", ""year"": 2007}, {""authors"": [""N. Pham"", ""R. Bernardi"", ""Y.-Z. Zhang"", ""M. Baroni.""], ""title"": ""Sentence paraphrase detection: When determiners and word order make the difference"", ""venue"": ""Proceedings of the Towards a Formal Distributional"", ""year"": 2013}, {""authors"": [""T.A. Plate""], ""title"": ""Holographic reduced representations: Convolution algebra for compositional distributed representations"", ""venue"": ""Proceedings of the 12th International Joint Conference on Artificial Intelligence,"", ""year"": 1991}, {""authors"": [""A. Preller""], ""title"": ""Polynomial pregroup grammars parse context sensitive languages"", ""venue"": ""Linguistic Analysis, 36:483\u2013516."", ""year"": 2010}, {""authors"": [""A. Preller"", ""M. Sadrzadeh.""], ""title"": ""Bell states and negative sentences in the distributed model of meaning"", ""venue"": ""P. Selinger, B. Coecke, P. Panangaden, editors, Electronic Notes in Theoretical"", ""year"": 2010}, {""authors"": [""P. Selinger""], ""title"": ""A survey of graphical languages for monoidal categories"", ""venue"": ""New Structures for Physics, pages 275\u2013337."", ""year"": 2010}, {""authors"": [""P. Smolensky""], ""title"": ""Tensor product variable binding and the representation of symbolic structures in connectionist systems"", ""venue"": ""Artificial Intelligence, 46(1-2):159\u2013216."", ""year"": 1990}, {""authors"": [""P. Smolensky"", ""G. Legendre""], ""title"": ""The Harmonic Mind: From Neural Computation"", ""year"": 2006}, {""authors"": [""R. Socher"", ""B. Huval"", ""C.D. Manning"", ""A.Y. Ng.""], ""title"": ""Semantic compositionality through recursive matrix-vector spaces"", ""venue"": ""Proceedings of the 2012 Conference on Empirical Methods in Natural Language"", ""year"": 2012}, {""authors"": [""M. Steedman""], ""title"": ""The Syntactic Process"", ""venue"": ""The MIT Press."", ""year"": 2001}, {""authors"": [""M. Steedman"", ""J. Baldridge.""], ""title"": ""Combinatory Categorial Grammar"", ""venue"": ""Wiley-Blackwell."", ""year"": 2011}, {""authors"": [""A. Stolcke""], ""title"": ""SRILM\u2014An extensible language modeling toolkit"", ""venue"": ""Seventh International Conference on Spoken Language Processing."", ""year"": 2002}, {""authors"": [""P.D. Turney"", ""P. Pantel.""], ""title"": ""From frequency to meaning: Vector space models of semantics"", ""venue"": ""Journal of Artificial Intelligence Research, 37(1):141\u2013188."", ""year"": 2010}, {""authors"": [""Turney"", ""Peter D.""], ""title"": ""Domain and function: A dual-space model of semantic relations and compositions"", ""venue"": ""Journal of Artificial Intelligence Research, 44:533\u2013585."", ""year"": 2012}, {""authors"": [""C.J. Van Rijsbergen""], ""title"": ""The Geometry of Information Retrieval"", ""venue"": ""Cambridge University Press."", ""year"": 2004}, {""authors"": [""R.F. Walters""], ""title"": ""Categories and Computer Science"", ""venue"": ""Cambridge University Press."", ""year"": 1991}, {""authors"": [""D. Widdows""], ""title"": ""Geometry and Meaning"", ""venue"": ""University of Chicago Press."", ""year"": 2005}, {""authors"": [""D. Widdows""], ""title"": ""Semantic vector products: Some initial investigations"", ""venue"": ""Proceedings of the Second Quantum Interaction Symposium (QI-2008). College Publications. CITESEER, Oxford."", ""year"": 2008}, {""authors"": [""L. Wittgenstein""], ""title"": ""Philosophical Investigations"", ""venue"": ""Blackwell."", ""year"": 1953}, {""authors"": [""F.M. Zanzotto"", ""I. Korkontzelos"", ""F. Fallucchi"", ""S. Manandhar.""], ""title"": ""Estimating linear models for compositional distributional semantics"", ""venue"": ""Proceedings of the 23rd International Conference"", ""year"": 2010}] acknowledgments  :We thank the EPSRC for funding the research leading to this article via EPSRC grant EP/J002607/1. Furthermore, we would like to thank John Harding and the equational theorem prover “Prover 9”11 for checking the identities occurred as a result of translating CCG’s rules to the language of a pregroup grammar. We would also like to thank Stephen Clark, Stephen Pulman, Bob Coecke, Karl Moritz Hermann, Richard Socher, and Dimitri Kartsaklis for valuable discussions and comments.","1 introduction :The distributional approach to the semantic modeling of natural language, inspired by the notion—presented by Firth (1957) and Harris (1968)—that the meaning of a word is tightly related to its context of use, has grown in popularity as a method of semantic representation. It draws from the frequent use of vector-based document models in information retrieval, modeling the meaning of words as vectors based on the distribution of co-occurring terms within the context of a word.
Using various vector similarity metrics as a measure of semantic similarity, these distributional semantic models (DSMs) are used for a variety of NLP tasks, from
∗ DeepMind Technologies Ltd, 5 New Street Square, London EC4A 3 TW. E-mail: etg@google.com.
∗∗ School of Electronic Engineering and Computer Science, Queen Mary University of London, Mile End Road, London E1 4NS, United Kingdom. E-mail: mehrnoosh.sadrzadeh@qmul.ac.uk.
† The work described in this article was performed while the authors were at the University of Oxford. 1 Support from EPSRC grant EP/J002607/1 is acknowledged.
Submission received: 26 September 2012; revised submission received: 31 October 2013; accepted for publication: 5 April 2014.
doi:10.1162/COLI a 00209
© 2015 Association for Computational Linguistics
automated thesaurus building (Grefenstette 1994; Curran 2004) to automated essay marking (Landauer and Dumais 1997). The broader connection to information retrieval and its applications is also discussed by Manning, Raghavan, and Schütze (2011). The success of DSMs in essentially word-based tasks such as thesaurus extraction and construction (Grefenstette 1994; Curran 2004) invites an investigation into how DSMs can be applied to NLP and information retrieval (IR) tasks revolving around larger units of text, using semantic representations for phrases, sentences, or documents, constructed from lemma vectors. However, the problem of compositionality in DSMs—of how to go from word to sentence and beyond—has proved to be non-trivial.
A new framework, which we refer to as DisCoCat, initially presented in Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) reconciles distributional approaches to natural language semantics with the structured, logical nature of formal semantic models. This framework is abstract; its theoretical predictions have not been evaluated on real data, and its applications to empirical natural language processing tasks have not been studied.
This article is the journal version of Grefenstette and Sadrzadeh (2011a, 2011b), which fill this gap in the DisCoCat literature; in it, we develop a concrete model and an unsupervised learning algorithm to instantiate the abstract vectors, linear maps, and vector spaces of the theoretical framework; we develop a series of empirical natural language processing experiments and data sets and implement our algorithm on large scale real data; we analyze the outputs of the algorithm in terms of linear algebraic equations; and we evaluate the model on these experiments and compare the results with other competing unsupervised models. Furthermore, we provide a linear algebraic analysis of the algorithm of Grefenstette and Sadrzadeh (2011a) and present an in-depth study of the better performance of the method of Grefenstette and Sadrzadeh (2011b).
We begin in Section 2 by presenting the background to the task of developing compositional distributional models. We briefly introduce two approaches to semantic modeling: formal semantic models and distributional semantic models. We discuss their differences, and relative advantages and disadvantages. We then present and critique various approaches to bridging the gap between these models, and their limitations. In Section 3, we summarize the categorical compositional distributional framework of Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010) and provide the theoretical background necessary to understand it; we also sketch the road map of the literature leading to the development of this setting and outline the contributions of this paper to the field. In Section 4, we present the details of an implementation of this framework, and introduce learning algorithms used to build semantic representations in this implementation. In Section 5, we present a series of experiments designed to evaluate this implementation against other unsupervised distributional compositional models. Finally, in Section 6 we discuss these results, and posit future directions for this research area. 2 background :Compositional formal semantic models represent words as parts of logical expressions, and compose them according to grammatical structure. They stem from classical ideas in logic and philosophy of language, mainly Frege’s principle that the meaning of a sentence is a function of the meaning of its parts (Frege 1892). These models relate to well-known and robust logical formalisms, hence offering a scalable theory of meaning that can be used to reason about language using logical tools of proof and inference. Distributional models are a more recent approach to semantic modeling, representing
the meaning of words as vectors learned empirically from corpora. They have found their way into real-world applications such as thesaurus extraction (Grefenstette 1994; Curran 2004) or automated essay marking (Landauer and Dumais 1997), and have connections to semantically motivated information retrieval (Manning, Raghavan, and Schütze 2011). This two-sortedness of defining properties of meaning: “logical form” versus “contextual use,” has left the quest for “what is the foundational structure of meaning?”—a question initially the concern of solely linguists and philosophers of language—even more of a challenge.
In this section, we present a short overview of the background to the work developed in this article by briefly describing formal and distributional approaches to natural language semantics, and providing a non-exhaustive list of some approaches to compositional distributional semantics. For a more complete review of the topic, we encourage the reader to consult Turney (2012) or Clark (2013). Formal semantic models provide methods for translating sentences of natural language into logical formulae, which can then be fed to computer-aided automation tools to reason about them (Alshawi 1992).
To compute the meaning of a sentence consisting of n words, meanings of these words must interact with one another. In formal semantics, this further interaction is represented as a function derived from the grammatical structure of the sentence. Such models consist of a pairing of syntactic analysis rules (in the form of a grammar) with semantic interpretation rules, as exemplified by the simple model presented on the left of Figure 1.
The semantic representations of words are lambda expressions over parts of logical formulae, which can be combined with one another to form well-formed logical expressions. The function | − | : L → M maps elements of the lexicon L to their interpretation (e.g., predicates, relations, domain objects) in the logical model M used. Nouns are typically just logical atoms, whereas adjectives and verbs and other relational words are interpreted as predicates and relations. The parse of a sentence such as “cats like milk,” represented here as a binarized parse tree, is used to produce its semantic interpretation by substituting semantic representations for their grammatical constituents and applying β-reduction where needed. Such a derivation is shown on the right of Figure 1.
What makes this class of models attractive is that it reduces language meaning to logical expressions, a subject well studied by philosophers of language, logicians, and linguists. Its properties are well known, and it becomes simple to evaluate the meaning of a sentence if given a logical model and domain, as well as verify whether or not one sentence entails another according to the rules of logical consequence and deduction.
However, such logical analysis says nothing about the closeness in meaning or topic of expressions beyond their truth-conditions and which models satisfy these truth conditions. Hence, formal semantic approaches to modeling language meaning do not perform well on language tasks where the notion of similarity is not strictly based on truth conditions, such as document retrieval, topic classification, and so forth. Furthermore, an underlying domain of objects and a valuation function must be provided, as with any logic, leaving open the question of how we might learn the meaning of language using such a model, rather than just use it. A popular way of representing the meaning of words in lexical semantics is as distributions in a high-dimensional vector space. This approach is based on the distributional hypothesis of Harris (1968), who postulated that the meaning of a word was dictated by the context of its use. The more famous dictum stating this hypothesis is the statement of Firth (1957) that “You shall know a word by the company it keeps.” This view of semantics has furthermore been associated (Grefenstette 2009; Turney and Pantel 2010) with earlier work in philosophy of language by Wittgenstein (presented in Wittgenstein 1953), who stated that language meaning was equivalent to its real world use.
Practically speaking, the meaning of a word can be learned from a corpus by looking at what other words occur with it within a certain context, and the resulting distribution can be represented as a vector in a semantic vector space. This vectorial representation is convenient because vectors are a familiar structure with a rich set of ways of computing vector distance, allowing us to experiment with different word similarity metrics. The geometric nature of this representation entails that we can not only compare individual words’ meanings with various levels of granularity (e.g., we might, for example, be able to show that cats are closer to kittens than to dogs, but that all three are mutually closer than cats and steam engines), but also apply methods frequently called upon in IR tasks, such as those described by Manning, Raghavan, and Schütze (2011), to group concepts by topic, sentiment, and so on.
The distribution underlying word meaning is a vector in a vector space, the basis vectors of which are dictated by the context. In simple models, the basis vectors will be annotated with words from the lexicon. Traditionally, the vector spaces used in such models are Hilbert spaces (i.e., vector spaces with orthogonal bases, such that the inner product of any one basis vector with another [other than itself] is zero). The semantic vector for any word can be represented as the weighted sum of the basis vectors:
−−−−−−−→ some word = ∑ i ci −→ni
where {−→ni }i is the basis of the vector space the meaning of the word lives in, and ci ∈ R is the weight associated with basis vector −→ni .
The construction of the vector for a word is done by counting, for each lexicon word ni associated with basis vector
−→ni , how many times it occurs in the context of each occurrence of the word for which we are constructing the vector. This count is then typically adjusted according to a weighting scheme (e.g., TF-IDF). The “context” of a word can be something as simple as the other words occurring in the same sentence as
the word or within k words of it, or something more complex, such as using dependency relations (Padó and Lapata 2007) or other syntactic features.
Commonly, the similarity of two semantic vectors is computed by taking their cosine measure, which is the sum of the product of the basis weights of the vectors:
cosine(−→a ,−→b ) = ∑ i c a i c b i√∑
i (c a i )
2 ∑
i (c b i ) 2
where cai and c b i are the basis weights for −→a and −→b , respectively. However, other options may be a better fit for certain implementations, typically dependent on the weighting scheme.
Readers interested in learning more about these aspects of distributional lexical semantics are invited to consult Curran (2004), which contains an extensive overview of implementation options for distributional models of word meaning. In the previous overview of distributional semantic models of lexical semantics, we have seen that DSMs are a rich and tractable way of learning word meaning from a corpus, and obtaining a measure of semantic similarity of words or groups of words. However, it should be fairly obvious that the same method cannot be applied to sentences, whereby the meaning of a sentence would be given by the distribution of other sentences with which it occurs.
First and foremost, a sentence typically occurs only once in a corpus, and hence substantial and informative distributions cannot be created in this manner. More importantly, human ability to understand new sentences is a compositional mechanism: We understand sentences we have never seen before because we can generate sentence meaning from the words used, and how they are put into relation. To go from word vectors to sentence vectors, we must provide a composition operation allowing us to construct a sentence vector from a collection of word vectors. In this section, we will discuss several approaches to solving this problem, their advantages, and their limitations.
2.3.1 Additive Models. The simplest composition operation that comes to mind is straightforward vector addition, such that:
−→ ab = −→a +−→b
Conceptually speaking, if we view word vectors as semantic information distributed across a set of properties associated with basis vectors, using vector addition as a semantic composition operation states that the information of a set of lemmas in a sentence is simply the sum of the information of the individual lemmas. Although crude, this approach is computationally cheap, and appears sufficient for certain NLP tasks: Landauer and Dumais (1997) show it to be sufficient for automated essay marking tasks, and Grefenstette (2009) shows it to perform better than a collection of other simple similarity metrics for summarization, sentence paraphrase, and document paraphrase detection tasks.
However, there are two principal objections to additive models of composition: first, vector addition is commutative, therefore, −−−−−−−−−−−−→ John drank wine = −−→ John + −−−→ drank + −−−→ wine =
−−−−−−−−−−−−→ Wine drank John, and thus vector addition ignores syntactic structure completely; and second, vector addition sums the information contained in the vectors, effectively jumbling the meaning of words together as sentence length grows.
The first objection is problematic, as the syntactic insensitivity of additive models leads them to equate the representation of sentences with patently different meanings. Mitchell and Lapata (2008) propose to add some degree of syntactic sensitivity—namely, accounting for word order—by weighting word vectors according to their order of appearance in a sentence as follows:
−→ ab = α−→a + β−→b
where α,β∈R. Consequently, −−−−−−−−−−−−→John drank wine=α · −−→John + β · −−−→drank + γ·−−−→wine would not have the same representation as −−−−−−−−−−−−→ Wine drank John = α · −−−→wine + β · −−−→drank + γ · −−→John.
The question of how to obtain weights and whether they are only used to reflect word order or can be extended to cover more subtle syntactic information is open, but it is not immediately clear how such weights may be obtained empirically and whether this mode of composition scales well with sentence length and increase in syntactic complexity. Guevara (2010) suggests using machine-learning methods such as partial least squares regression to determine the weights empirically, but states that this approach enjoys little success beyond minor composition such as adjective-noun or noun-verb composition, and that there is a dearth of metrics by which to evaluate such machine learning–based systems, stunting their growth and development.
The second objection states that vector addition leads to increase in ambiguity as we construct sentences, rather than decrease in ambiguity as we would expect from giving words a context; and for this reason Mitchell and Lapata (2008) suggest replacing additive models with multiplicative models as discussed in Section 2.3.2, or combining them with multiplicative models to form mixture models as discussed in Section 2.3.3.
2.3.2 Multiplicative Models. The multiplicative model of Mitchell and Lapata (2008) is an attempt to solve the ambiguity problem discussed in Section 2.3.1 and provide implicit disambiguation during composition. The composition operation proposed is the component-wise multiplication ( ) of two vectors: Vectors are expressed as the weighted sum of their basis vectors, and the weight of the basis vectors of the composed vector is the product of the weights of the original vectors; for −→a =∑i ci−→ni , and−→ b = ∑ i c ′ i −→ni , we have
−→ ab = −→a −→b = ∑ i cic′i −→ni
Such multiplicative models are shown by Mitchell and Lapata (2008) to perform better at verb disambiguation tasks than additive models for noun-verb composition, against a baseline set by the original verb vectors. The experiment they use to show this will also serve to evaluate our own models, and form the basis for further experiments, as discussed in Section 5.
This approach to compositionality still suffers from two conceptual problems: First, component-wise multiplication is still commutative and hence word order is not accounted for; second, rather than “diluting” information during large compositions
and creating ambiguity, it may remove too much through the “filtering” effect of component-wise multiplication.
The first problem is more difficult to deal with for multiplicative models than for additive models, because both scalar multiplication and component-wise multiplication are commutative and hence α−→a β−→b = α−→b β−→a and thus word order cannot be taken into account using scalar weights.
To illustrate how the second problem entails that multiplicative models do not scale well with sentence length, we look at the structure of component-wise multiplication again: −→a −→b =∑i cic′i −→ni . For any i, if ci = 0 or c′i = 0, then cic′i = 0, and therefore for any composition, the number of non-zero basis weights of the produced vector is less than or equal to the number of non-zero basis weights of the original vectors: At each composition step information is filtered out (or preserved, but never increased). Hence, as the number of vectors to be composed grows, the number of non-zero basis weights of the product vector stays the same or—more realistically—decreases. Therefore, for any composition of the form −−−−−−−−→a1 . . . ai . . . an = −→a1 . . . −→ai . . . −→an , if for any i, −→ai is orthogonal to −−−−−−→a1 . . . ai−1 then −−−−−−−−→a1 . . . ai . . . an = −→0 . It follows that purely multiplicative models alone are not apt as a single mode of composition beyond nounverb (or adjective-noun) composition operations.
One solution to this second problem not discussed by Mitchell and Lapata (2008) would be to introduce some smoothing factor s ∈ R+ for point-wise multiplication such that −→a −→b =∑i (ci + s)(c′i + s)−→ni , ensuring that information is never completely filtered out. Seeing how the problem of syntactic insensitivity still stands in the way of full-blown compositionality for multiplicative models, we leave it to those interested in salvaging purely multiplicative models to determine whether some suitable value of s can be determined.
2.3.3 Mixture Models. The problems faced by multiplicative models presented in Section 2.3.2 are acknowledged in passing by Mitchell and Lapata (2008), who propose mixing additive and multiplicative models in the hope of leveraging the advantage of each while doing away with their pitfalls. This is simply expressed as the weighted sum of additive and multiplicative models:
−→ ab = α−→a + β−→b + γ(−→a −→b )
where α, β, and γ are predetermined scalar weights. The problems for these models are threefold. First, the question of how scalar weights are to be obtained still needs to be determined. Mitchell and Lapata (2008) concede that one advantage of purely multiplicative models over weighted additive or mixture models is that the lack of scalar weights removes the need to optimize the scalar weights for particular tasks (at the cost of not accounting for syntactic structure), and avoids the methodological concerns accompanying this requirement.
Second, the question of how well this process scales from noun-verb composition to more syntactically rich expressions must be addressed. Using scalar weights to account for word order seems ad hoc and superficial, as there is more to syntactic structure than the mere ordering of words. Therefore, an account of how to build sentence vectors for sentences such as The dog bit the man and The man was bitten by the dog in order to give both sentences the same (or a similar) representation would need to give a richer role to scalar weights than mere token order. Perhaps specific
weights could be given to particular syntactic classes (such as nouns) to introduce a more complex syntactic element into vector composition, but it is clear that this alone is not a solution, as the weight for nouns dog and man would be the same, allowing for the same commutative degeneracy observed in non-weighted additive models, in which −−−−−−−−−−−−−−→ the dog bit the man = −−−−−−−−−−−−−−→ the man bit the dog. Introducing a mixture of weighting systems accounting for both word order and syntactic roles may be a solution, but it is not only ad hoc but also arguably only partially reflects the syntactic structure of the sentence.
The third problem is that Mitchell and Lapata (2008) show that in practice, although mixture models perform better at verb disambiguation tasks than additive models and weighted additive models, they perform equivalently to purely multiplicative models with the added burden of requiring parametric optimization of the scalar weights.
Therefore, whereas mixture models aim to take the best of additive and multiplicative models while avoiding their problems, they are only partly successful in achieving the latter goal, and demonstrably do little better in achieving the former.
2.3.4 Tensor-Based Models. From Sections 2.3.1–2.3.3 we observe that the need for incorporating syntactic information into DSMs to achieve true compositionality is pressing, if only to develop a non-commutative composition operation that can take into account word order without the need for adhoc weighting schemes, and hopefully richer syntactic information as well.
An early proposal by Smolensky and colleagues (Smolensky 1990; Smolensky and Legendre 2006) to use linear algebraic tensors as a composition operation solves the problem of finding non-commutative vector composition operators. The composition of two vectors is their tensor product, sometimes called the kronecker product when ap-
plied to vectors rather than vector spaces. For −→a ∈ V =∑i ci−→ni , and −→b ∈ W =∑j c′j−→n′j , we have:
−→ ab = −→a ⊗−→b = ∑ ij cic′j −→ni ⊗ −→ n′j
The composition operation takes the original vectors and maps them to a vector in a larger vector space V ⊗ W, which is the tensor space of the original vectors’ spaces. Here the second instance of ⊗ is not a recursive application of the kronecker product, but rather the pairing of basis elements of V and W to form a basis element of V ⊗ W. The shared notation and occasional conflation of kronecker and tensor products may seem confusing, but is fairly standard in multilinear algebra.
The advantage of this approach is twofold: First, vectors for different words need not live in the same spaces but can be composed nonetheless. This allows us to represent vectors for different word classes (topics, syntactic roles, etc.) in different spaces with different bases, which was not possible under additive or multiplicative models. Second, because the product vector lives in a larger space, we obtain the intuitive notion that the information of the whole is richer and more complex than the information of the parts.
Dimensionality Problems. However, as observed and commented upon by Smolensky himself, this increase in dimensionality brings two rather large problems for tensor based models. The first is computational: The size of the product vector space is the product of the size of the original vector spaces. If we assume that all words live in the
same space N of dimensionality dim(N), then the dimensionality of an n-word sentence vector is dim(N)n. If we have as many basis vectors for our word semantic space as there are lexemes in our vocabulary (e.g., approximately 170k in English2), then the size of our sentence vectors quickly reaches magnitudes for which vector comparison (or even storage) are computationally intractable.3 Even if, as most DSM implementations do, we restrict the basis vectors of word semantic spaces to the k (e.g., k = 2,000) most frequent words in a corpus, the sentence vector size still grows exponentially with sentence length, and the implementation problems remain.
The second problem is mathematical: Sentences of different length live in different spaces, and if we assign different vector spaces to different word types (e.g., syntactic classes), then sentences of different syntactic structure live in different vector spaces, and hence cannot be compared directly using inner product or cosine measures, leaving us with no obvious mode of semantic comparison for sentence vectors. If any model wishes to use tensor products in composition operations, it must find some way of reducing the dimensionality of product vectors to some common vector space so that they may be directly compared.
One notable method by which these dimensionality problems can be solved in general are the holographic reduced representations proposed by Plate (1991). The product vector of two vectors is projected into a space of smaller dimensionality by circular convolution to produce a trace vector. The circular correlation of the trace vector and one of the original vectors produces a noisy version of the other original vector. The noisy vector can be used to recover the clean original vector by comparing it with a predefined set of candidates (e.g., the set of word vectors if our original vectors are word meanings). Traces can be summed to form new traces effectively containing several vector pairs from which original vectors can be recovered. Using this encoding/decoding mechanism, the tensor product of sets of vectors can be encoded in a space of smaller dimensionality, and then recovered for computation without ever having to fully represent or store the full tensor product, as discussed by Widdows (2008).
There are problems with this approach that make it unsuitable for our purposes, some of which are discussed in Mitchell and Lapata (2010). First, there is a limit to the information that can be stored in traces, which is independent of the size of the vectors stored, but is a logarithmic function of their number. As we wish to be able to store information for sentences of variable word length without having to directly represent the tensored sentence vector, setting an upper bound to the number of vectors that can be composed in this manner limits the length of the sentences we can represent compositionally using this method.
Second, and perhaps more importantly, there are restrictions on the nature of the vectors that can be encoded in such a way: The vectors must be independently distributed such that the mean Euclidean length of each vector is 1. Such conditions are unlikely to be met in word semantic vectors obtained from a corpus; and as the failure to do so affects the system’s ability to recover clean vectors, holographic reduced representations are not prima facie usable for compositional DSMs, although it is important to note that Widdows (2008) considers possible application areas where they may be of use, although once again these mostly involve noun-verb and adjectivenoun compositionality rather than full blown sentence vector construction. We retain
2 Source: http://www.oxforddictionaries.com/page/howmanywords. 3 At four bytes per integer, and one integer per basis vector weight, the vector for John loves Mary
would require roughly (170, 000 · 4)3 ≈ 280 petabytes of storage, which is over ten times the data Google processes on a daily basis according to Dean and Ghemawat (2008).
from Plate (1991) the importance of finding methods by which to project the tensored sentence vectors into a common space for direct comparison, as will be discussed further in Section 3.
Syntactic Expressivity. An additional problem of a more conceptual nature is that using the tensor product as a composition operation simply preserves word order. As we discussed in Section 2.3.3, this is not enough on its own to model sentence meaning. We need to have some means by which to incorporate syntactic analysis into composition operations.
Early work on including syntactic sensitivity into DSMs by Grefenstette (1992) suggests using crude syntactic relations to determine the frame in which the distributions for word vectors are collected from the corpus, thereby embedding syntactic information into the word vectors. This idea was already present in the work of Smolensky, who used sums of pairs of vector representations and their roles, obtained by taking their tensor products, to obtain a vector representation for a compound. The application of these ideas to DSMs was studied by Clark and Pulman (2007), who suggest instantiating the roles to dependency relations and using the distributional representations of words as the vectors. For example, in the sentence Simon loves red wine, Simon is the subject of loves, wine is its object, and red is an adjective describing wine. Hence, from the dependency tree with loves as root node, its subject and object as children, and their adjectival descriptors (if any) as their children, we read the following structure: −−−→ loves ⊗−−→subj ⊗−−−−→Simon ⊗−→obj ⊗−−−→wine ⊗−→adj ⊗−→red. Using the equality relation for inner products of tensor products:
〈−→a ⊗−→b | −→c ⊗−→d 〉 = 〈−→a | −→c 〉 × 〈−→b | −→d 〉
We can therefore express inner-products of sentence vectors efficiently without ever having to actually represent the tensored sentence vector:
〈−−−−−−−−−−−−→Simon loves red wine | −−−−−−−−−−−−−−→Mary likes delicious rosé〉 = 〈−−→loves ⊗−→subj ⊗−−→Simon ⊗−→obj ⊗−−→wine ⊗−→adj ⊗−→red | −→likes ⊗−→subj ⊗−−→Mary ⊗−→obj ⊗−→rosé ⊗−→adj ⊗−−−−→delicious〉 = 〈−−→loves | −→likes〉× 〈−→subj | −→subj〉 × 〈−−→Simon | −−→Mary〉 × 〈−→obj | −→obj〉 × 〈−−→wine | −→rosé〉× 〈−→adj | −→adj 〉 × 〈−→red | −−−−→delicious〉 = 〈−−→loves | −→likes〉× 〈−−→Simon | −−→Mary〉× 〈−−→wine | −→rosé〉 × 〈−→red | −−−−→delicious〉
This example shows that this formalism allows for sentence comparison of sentences with identical dependency trees to be broken down to term-to-term comparison without the need for the tensor products to ever be computed or stored, reducing computation to inner product calculations.
However, although matching terms with identical syntactic roles in the sentence works well in the given example, this model suffers from the same problems as the original tensor-based compositionality of Smolensky (1990) in that, by the authors’ own admission, sentences of different syntactic structure live in spaces of different dimensionality and thus cannot be directly compared. Hence we cannot use this to measure the similarity between even small variations in sentence structure, such as the pair “Simon likes red wine” and “Simon likes wine.”
2.3.5 SVS Models. The idea of including syntactic relations to other lemmas in word representations discussed in Section 2.3.4 is applied differently in the structured vector
space model presented by Erk and Padó (2008). They propose to represent word meanings not as simple vectors, but as triplets:
w = (v, R, R−1)
where v is the word vector, constructed as in any other DSM, R and R−1 are selectional preferences, and take the form of R → D maps where R is the set of dependency relations, and D is the set of word vectors. Selectional preferences are used to encode the lemmas that w is typically the parent of in the dependency trees of the corpus in the case of R, and typically the child of in the case of R−1.
Composition takes the form of vector updates according to the following protocol. Let a = (va, Ra, R−1a ) and b = (vb, Rb, R −1 b ) be two words being composed, and let r be the dependency relation linking a to b. The vector update procedure is as follows:
a′ = (va R−1b (r), Ra − {r}, R−1a ) b′ = (vb Ra(r), Rb, R−1b − {r})
where a′, b′ are the updated word meanings, and is whichever vector composition (addition, component-wise multiplication) we wish to use. The word vectors in the triplets are effectively filtered by combination with the lemma which the word they are being composed with expects to bear relation r to, and this relation between the composed words a and b is considered to be used and hence removed from the domain of the selectional preference functions used in composition.
This mechanism is therefore a more sophisticated version of the compositional disambiguation mechanism discussed by Mitchell and Lapata (2008) in that the combination of words filters the meaning of the original vectors that may be ambiguous (e.g., if we have one vector for bank); but contrary to Mitchell and Lapata (2008), the information of the original vectors is modified but essentially preserved, allowing for further combination with other terms, rather than directly producing a joint vector for the composed words. The added fact that R and R−1 are partial functions associated with specific lemmas forces grammaticality during composition, since if a holds a dependency relation r to b which it never expects to hold (for example a verb having as its subject another verb, rather than the reverse), then Ra and R −1 b are undefined for r and the update fails. However, there are some problems with this approach if our goal is true compositionality.
First, this model does not allow some of the “repeated compositionality” we need because of the update of R and R−1. For example, we expect that an adjective composed with a noun produces something like a noun in order to be further composed with a verb or even another adjective. However, here, because the relation adj would be removed from R−1b for some noun b composed with an adjective a, this new representation b
′ would not have the properties of a noun in that it would no longer expect composition with an adjective, rendering representations of simple expressions like “the new red car” impossible. Of course, we could remove the update of the selectional preference functions from the compositional mechanism, but then we would lose this attractive feature of grammaticality enforcement through the partial functionality of R and R−1.
Second, this model does little more than represent the implicit disambiguation that is expected during composition, rather than actually provide a full blown compositional model. The inability of this system to provide a novel mechanism by which to obtain
a joint vector for two composed lemmas—thereby building towards sentence vectors— entails that this system provides no means by which to obtain semantic representations of larger syntactic structures that can be compared by inner product or cosine measure as is done with any other DSM. Of course, this model could be combined with the compositional models presented in Sections 2.3.1–2.3.3 to produce sentence vectors, but whereas some syntactic sensitivity would have been obtained, the word ordering and other problems of the aforementioned models would still remain, and little progress would have been made towards true compositionality.
We retain from this attempt to introduce compositionality in DSMs that including information obtained from syntactic dependency relations is important for proper disambiguation, and that having some mechanism by which the grammaticality of the expression being composed is a precondition for its composition is a desirable feature for any compositional mechanism. The final class of approaches to vector composition we wish to discuss are three matrix-based models.
Generic Additive Model. The first is the Generic Additive Model of Zanzotto et al. (2010). This is a generalization of the weighted additive model presented in Section 2.3.1. In this model, rather than multiplying lexical vectors by fixed parameters α and β before adding them to form the representation of their combination, they are instead the arguments of matrix multiplication by square matrices A and B:
−→ ab = A−→a + B−→b
Here, A and B represent the added information provided by putting two words into relation.
The numerical content of A and B is learned by linear regression over triplets (−→a ,−→b ,−→c ) where −→a and −→b are lexical semantic vectors, and −→c is the expected output of the combination of −→a and −→b . This learning system thereby requires the provision of labeled data for linear regression to be performed. Zanzotto et al. (2010) suggest several sources for this labeled data, such as dictionary definitions and word etymologies.
This approach is richer than the weighted additive models because the matrices act as linear maps on the vectors they take as “arguments,” and thus can encode more subtle syntactic or semantic relations. However, this model treats all word combinations as the same operation—for example, treating the combination of an adjective with its argument and a verb with its subject as the same sort of composition. Because of the diverse ways there are of training such supervised models, we leave it to those who wish to further develop this specific line of research to perform such evaluations.
Adjective Matrices. The second approach is the matrix-composition model of Baroni and Zamparelli (2010), which they develop only for the case of adjective-noun composition, although their approach can seamlessly be used for any other predicate-argument composition. Contrary to most of the earlier approaches proposed, which aim to combine two lexical vectors to form a lexical vector for their combination, Baroni and Zamparelli suggest giving different semantic representations to different types, or more specifically to adjectives and nouns.
In this model, nouns are lexical vectors, as with other models. However, embracing a view of adjectives that is more in line with formal semantics than with distributional semantics, they model adjectives as linear maps taking lexical vectors as input and producing lexical vectors as output. Such linear maps can be encoded as square matrices, and applied to their arguments by matrix multiplication. Concretely, let Madjective be the matrix encoding the adjective’s linear map, and −−→noun be the lexical semantic vector for a noun; their combination is simply
−−−−−−−−−→ adjective noun = Madjective ×−−→noun
Similarly to the Generic Additive Model, the matrix for each adjective is learned by linear regression over a set of pairs (−−→noun,−→c ) where the vectors −−→noun are the lexical semantic vectors for the arguments of the adjective in a corpus, and −→c is the semantic vector corresponding to the expected output of the composition of the adjective with that noun.
This may, at first blush, also appear to be a supervised training method for learning adjective matrices from “labeled data,” seeing how the expected output vectors are needed. However, Baroni and Zamparelli (2010) work around this constraint by automatically producing the labeled data from the corpus by treating the adjectivenoun compound as a single token, and learning its vector using the same distributional learning methods they used to learn the vectors for nouns. This same approach can be extended to other unary relations without change and, using the general framework of the current article, an extension of it to binary predicates has been presented in Grefenstette et al. (2013), using multistep regression. For a direct comparison of the results of this approach with some of the results of the current article, we refer the reader to Grefenstette et al. (2013).
Recursive Matrix-Vector Model. The third approach is the recently developed Recursive Matrix-Vector Model (MV-RNN) of Socher et al. (2012), which claims the two matrixbased models described here as special cases. In MV-RNN, words are represented as a pairing of a lexical semantic vector −→a with an operation matrix A. Within this model, given the parse of a sentence in the form of a binarized tree, the semantic representation (−→c , C) of each non-terminal node in the tree is produced by performing the following two operations on its children (−→a , A) and (−→b , B).
First, the vector component −→c is produced by applying the operation matrix of one child to the vector of the other, and vice versa, and then projecting both of the products back into the same vector space as the child vectors using a projection matrix W, which must also be learned:
−→c = W × [
B ×−→a A ×−→b
]
Second, the matrix C is calculated by projecting the pairing of matrices A and B back into the same space, using a projection matrix WM, which must also be learned:
C = WM × [
A B
]
The pairing (−→c , C) obtained through these operations forms the semantic representation of the phrase falling under the scope of the segment of the parse tree below that node.
This approach to compositionality yields good results in the experiments described in Socher et al. (2012). It furthermore has appealing characteristics, such as treating relational words differently through their operation matrices, and allowing for recursive composition, as the output of each composition operation is of the same type of object as its inputs. This approach is highly general and has excellent coverage of different syntactic types, while leaving much room for parametric optimization.
The principal mathematical difference with the compositional framework presented subsequently is that composition in MV-RNN is always a binary operation; for example, to compose a transitive verb with its subject and object one would first need to compose it with its object, and then compose the output of that operation with the subject. The framework we discuss in this article allows for the construction of larger representations for relations of larger arities, permitting the simultaneous composition of a verb with its subject and object. Whether or not this theoretical difference leads to significant differences in composition quality requires joint evaluation. Additionally, the description of MV-RNN models in Socher et al. (2012) specifies the need for a source of learning error during training, which is easy to measure in the case of label prediction experiments such as sentiment prediction, but non-trivial in the case of paraphrase detection where no objective label exists. A direct comparison to MV-RNN methods within the context of experiments similar to those presented in this article has been produced by Blacoe and Lapata (2012), showing that simple operations perform on par with the earlier complex deep learning architectures produced by Socher and colleagues; we leave direct comparisons to future work. Early work has shown that the addition of a hidden layer with non-linearities to these simple models will improve the results. Domains and Functions. In recent work, Turney (2012) suggests modeling word representations not as a single semantic vector, but as a pair of vectors: one containing the information of the word relative to its domain (the other words that are ontologically similar), and another containing information relating to its function. The former vector is learned by looking at what nouns a word co-occurs with, and the latter is learned by looking at what verb-based patterns the word occurs in. Similarity between sets of words is not determined by a single similarity function, but rather through a combination of comparisons of the domain components of words’ representations with the function components of the words’ representations. Such combinations are designed on a task-specific basis. Although Turney’s work does not directly deal with vector composition of the sort we explore in this article, Turney shows that similarity measures can be designed for tasks similar to those presented here. The particular limitation of his approach, which Turney discusses, is that similarity measures must be specified for each task, whereas most of the compositional models described herein produce representations that can be compared in a task-independent manner (e.g., through cosine similarity). Nonetheless, this approach is innovative, and will merit further attention in future work in this area.
Language as Algebra. A theoretical model of meaning as context has been proposed in Clarke (2009, 2012). In that model, the meaning of any string of words is a vector built
from the occurrence of the string in a corpus. This is the most natural extension of distributional models from words to strings of words: in that model, one builds vectors for strings of words in exactly the same way as one does for words. The main problem, however, is that of data sparsity for the occurrences of strings of words. Words do appear repeatedly in a document, but strings of words, especially for longer strings, rarely do so; for instance, it hardly happens that an exact same sentence appears more than once in a document. To overcome this problem, the model is based on the hypothetical concept of an infinite corpus, an assumption that prevents it from being applied to real-world corpora and experimented within natural language processing tasks. On the positive side, the model provides a theoretical study of the abstract properties of a general bilinear associative composition operator; in particular, it is shown that this operator encompasses other composition operators, such as addition, multiplication, and even tensor product. 4 concrete semantic spaces :In Section 3.3 we presented a categorical formalism that relates syntactic analysis steps to semantic composition operations. The structure of our syntactic representation dictates the structure of our semantic spaces, but in exchange, we are provided with composition functions by the syntactic analysis, rather than having to stipulate them ad hoc. Whereas the approaches to compositional DSMs presented in Section 2 either failed to take syntax into account during composition, or did so at the cost of not being able to compare sentences of different structure in a common space, this categorical approach projects all sentences into a common sentence space where they can be directly compared. However, this alone does not give us a compositional DSM.
As we have seen in the previous examples, the structure of semantic spaces varies with syntactic types. We therefore cannot construct vectors for different syntactic types in the same way, as they live in spaces of different structure and dimensionality. Furthermore, nothing has yet been said about the structure of the sentence space S into which expressions reducing to type s are projected. If we wish to have a compositional DSM
that leverages all the benefits of lexical DSMs and ports them to sentence-level distributional representations, we must specify a new sort of vector construction procedure.
In the original formulation of this formalism by Clark, Coecke, and Sadrzadeh (2008) and Coecke, Sadrzadeh, and Clark (2010), examples of how such a compositional DSM could be used for logical evaluation are presented, where S is defined as a Boolean space with True and False as basis vectors. However, the word vectors used are handwritten and specified model-theoretically, as the authors leave it for future research to determine how such vectors might be obtained from a corpus. In this section, we will discuss a new way of constructing vectors for compositional DSMs, and of defining sentence space S, in order to reconcile this powerful categorical formalism with the applicability and flexibility of standard distributional models. Assume the following sentences are all true:
1. The dogs chased the cats.
2. The dogs annoyed the cats.
3. The puppies followed the kittens.
4. The train left the station.
5. The president followed his agenda.
If asked which sentences have similar meaning, we would most likely point to the pair (1) and (3), and perhaps to a lesser degree (1) and (2), and (2) and (3). Sentences (4) and (5) obviously speak of a different state of the world as the other sentences.
If we compare these by truth value, we obviously have no means of making such distinctions. If we compare these by lexical similarity, (1) and (2) seem to be a closer match than (1) and (3). If we are classifying these sentences by some higher order relation such as “topic,” (5) might end up closer to (3) than (1). What, then, might cause us to pair (1) and (3)?
Intuitively, this similarity seems to be because the subjects and objects brought into relation by similar verbs are themselves similar. Abstracting away from tokens to some notion of property, we might say that both (1) and (3) express the fact that something furry and feline and furtive is being pursued by something aggressive (or playful) and canine. Playing along with the idea that lexical distributional semantics present concepts (word meanings) as “messy bundles of properties,” it seems only natural to have the way these properties are acted upon, qualified, and related as the basis for sentence-level distributional representations. In this respect, we here suggest that the sentence space S, instead of qualifying the truth value of a sentence, should express how the properties of the semantic objects within are qualified or brought into relation by verbs, adjectives, and other predicates and relations.
More specifically, we examine two suggestions for defining the sentence space, namely, SI ∼= N for sentences with intransitive verbs and ST ∼= N ⊗ N for sentences with transitive verbs. These definitions mean that the sentence space’s dimensions are commensurate with either those of N, or those of N ⊗ N. These are by no means the only options, but as we will discuss here, they offer practical benefits.
In the case of SI, the basis elements are labeled with unique basis elements of N, hence, −→s1 = −→n1 , −→s2 = −→n2, and so on. In the case of ST, the basis elements are labeled with
unique ordered pairs of elements from N, for example, −→s1 = −−−−→ (n1, n1), −→s2 = −−−−→ (n2, n1), −→s3 =−−−−→ (n1, n2), and so on, following the same matrix flattening structure used in the proof of the equal cardinality of N and Q. Because of the isomorphism between ST and N ⊗ N, we will use the notations −−−−→ (ni, nj) and
−→ni ⊗−→nj interchangeably, as both constitute appropriate ways of representing the basis elements of such a space. To propagate this distinction on the syntactic level, we define types sI and sT for intransitive and transitive sentences, respectively.
In creating this distinction, we lost one of the most appealing features of the framework of Coecke, Sadrzadeh, and Clark (2010), namely, the result that all sentence vectors live in the same sentence space. A mathematical solution to this two-space problem was suggested in Grefenstette et al. (2011), and a variant of the models presented in this article permitting the non-problematic projection into a single sentence space (S ∼= N) has been presented by Grefenstette et al. (2013). Keeping this separation allows us to deal with both the transitive and intransitive cases in a simpler manner, and because the experiments in this article only compare intransitive sentences with intransitive sentences, and transitive sentences with transitive sentences, we will not address the issue of unification here. While Lambek’s pregroup types presented in Lambek (2008) include a rich array of basic types and hand-designed compound types in order to capture specific grammatic properties, for the sake of simplicity we will use a simpler set of grammatical types for experimental purposes, similar to some common types found in the CCG-bank (Steedman 2001).
We assign a basic pregroup type n for all nouns, with an associated vector space N for their semantic representations. Furthermore, we will treat noun-phrases as nouns, assigning to them the same pregroup type and semantic space.
CCG treats intransitive verbs as functions NP\S that consume a noun phrase and return a sentence, and transitive verbs as functions (NP\S)/NP that consume a noun phrase and return an intransitive verb function, which in turn consumes a noun phrase and returns a sentence. Using our distinction between intransitive sentences, we give intransitive verbs the type nrsI associated with the semantic space N ⊗ SI, and transitive verbs the type nrsTnl associated with the semantic space N ⊗ ST ⊗ N.
Adjectives, in CCG, are treated as functions NP/NP, consuming a noun phrase and returning a noun phrase; and hence we give them the type nnl and associated semantic space N ⊗ N.
With the provision of a learning procedure for vectors in these semantic spaces, we can use these types to construct sentence vector representations for simple intransitive verb–based and transitive verb–based sentences, with and without adjectives applied to subjects and objects. To begin, we construct the semantic space N for all nouns in our lexicon (typically limited by the words available in the corpus used). Any distributional semantic model can be used for this stage, such as those presented in Curran (2004), or the lexical semantic models used by Mitchell and Lapata (2008). It seems reasonable to assume that higher quality lexical semantic vectors—as measured by metrics such as the WordSim353 test
of Finkelstein et al. (2001)—will produce better relational vectors from the procedure designed subsequently. We will not test the hypothesis here, but note that it is an underlying assumption in most of the current literature on the subject (Erk and Padó 2008; Mitchell and Lapata 2008; Baroni and Zamparelli 2010).
Building upon the foundation of the thus constructed noun vectors, we construct semantic representations for relational words. In pregroup grammars (or other combinatorial grammars such as CCG), we can view such words as functions, taking as arguments those types present as adjoints in the compound pregroup type, and returning a syntactic object whose type is that of the corresponding reduction. For example, an adjective nnl takes a noun or noun phrase n and returns a noun phrase n from the reduction (nnl)n → n. It can also compose with another adjective to return an adjective (nnl)(nnl) → nnl. We wish for our semantic representations to be viewed in the same way, such that the composition of an adjective with a noun (1N ⊗ N)((N ⊗ N) ⊗ N) can be viewed as the application of a function f : N → N to its argument of type N.
To learn the representation of such functions, we assume that their meaning can be characterized by the properties that their arguments hold in the corpus, rather than just by their context as is the case in lexical distributional semantic models. To give an example, rather than learning what the adjective angry means by observing that it co-occurs with words such as fighting, aggressive, or mean, we learn its meaning by observing that it typically takes as argument words that have the property of being (i.e., co-occur with words such as) fighting, aggressive, and mean. Whereas in the lexical semantic case, such associations might only rarely occur in the corpus, in this indirect method we learn what properties the adjective relates to even if they do not co-occur with it directly.
In turn, through composition with its argument, we expect the function for such an adjective to strengthen the properties that characterize it in the object it takes as argument. If, indeed, angry is characterized by arguments that have high basis weights for basis elements corresponding to the concepts (or context words) fighting, aggressive, and mean, and relatively low counts for semantically different concepts such as passive, peaceful, and loves, then when we apply angry to dog the vector for the compound angry dog should contain some of the information found in the vector for dog. But this vector should also have higher values for the basis weights of fighting, aggressive, and mean, and correspondingly lower values for the basis weights of passive, peaceful, and loves.
To turn this idea into a concrete algorithm for constructing the semantic representation for relations of any arity, as first presented in Grefenstette et al. (2011), we examine how we would deal with this for binary relations such as transitive verbs. If a transitive verb of semantic type N ⊗ ST ⊗ N is viewed as a function f : N × N → ST that expresses the extent to which the properties of subject and object are brought into relation by the verb, we learn the meaning of the verb by looking at what properties are brought into relation by its arguments in a corpus. Recall that the vector for a verb v,−→v ∈ N ⊗ ST ⊗ N, can be expressed as the weighted sum of its basis elements:
−→v = ∑
ijk
cvijk −→ni ⊗−→sj ⊗−→nk
We take the set of vectors for the subject and object of v in the corpus to be argv = {(−−−−→SUBJl, −−−→ OBJl )}l. We wish to calculate the basis weightings {cvijk}ijk for v. Exploiting our earlier definition of the basis {sj}j of ST, which states that for any value of i and k there
is some value of j such that sj = (ni, nk), we define Δijk = 1 if indeed sj = (ni, nk) and 0 otherwise. Using all of this, we define the calculation of each basis weight cvijk as:
cvijk = ∑
l
Δijkc SUBJl i c OBJl k
This allows for a full formulation of −→v as follows: −→v =
∑ l ∑ ijk Δijkc SUBJl i c OBJl k −→ni ⊗−→sj ⊗−→nk
The nested sums here may seem computationally inefficient, seeing how this would involve computing size(argv) × dim(N)2 × dim(S) = size(argv) × dim(N)4 products. However, using the decomposition of basis elements of S into pairs of basis elements of N (effectively basis elements of N ⊗ N), we can remove the Δijk term and ignore all values of j where sj = (ni, nk), because the basis weight for this combination of indices would be 0. Hence we simplify:
−→v = ∑
l ∑ ik cSUBJli c OBJl k −→ni ⊗ −−−−→ (ni, nk) ⊗−→nk
This representation is still bloated: We perform fewer calculations, but still obtain a vector in which all the basis weights where sj = (ni, nk) are 0, hence where only dim(N)2 of the dim(N)4 values are non-zero. In short, the vector weights for −→v are, under this learning algorithm, entirely characterized by the values of a dim(N) by dim(N) matrix, the entries of which are products cSUBJli c OBJl k where i and k have become row and column indices. Using this and our definition of ST as a space isomorphic to N ⊗ N, we can formulate a compact expression of −→v as follows. Let the kronecker product of two vectors−→u ,−→w ∈ N, written −→u ⊗−→w ∈ N ⊗ N, be as follows: −→u ⊗−→w =
∑ ij cui c w j −→ni ⊗−→nj
Equipped with this definition, we can formulate the compact form of −→v :
−→v = ∑
l ∑ ik cSUBJli c OBJl k −→ni ⊗−→nk
= ∑
l
−−−→ SUBJl ⊗ −−→ OBJl
In short, we are only required to iterate through the corpus once, taking for each instance of a transitive verb v the kronecker product of its subject and object, and summing these across all instances of v. It is simple to see that no information was discarded relative to the previous definition of −→v : The dimensionality reduction by a factor of dim(N)2 simply discards all basis elements for which the basis weight was 0 by default.
This raises a small problem though: This compact representation can no longer be used in the compositional mechanism presented in Section 3.3, as the dimensions of−→v no longer match those that it is required to have according to its syntactic type. However, a solution can be devised if we return to the sample calculation shown in Section 3.3.2 of the composition of a transitive verb with its arguments. The composition reduces as follows:
( N ⊗ 1S ⊗ N)( −−−→ SUBJ ⊗−→v ⊗−−→OBJ) = ∑ ikm cSUBJi c v ikmc OBJ m −→sk
where the verb v is represented in its non-compact form. If we introduce the compactness given to us by our isomorphism ST ∼= N ⊗ N we can express this as
−−−−−−−−→ SUBJ v OBJ = ∑ im cSUBJi c v imc OBJ m −→ni ⊗−→nm
where v is represented in its compact form. Furthermore, by introducing the component wise multiplication operation :
−→u −→v = ∑
i
cui c v i −→ni
we can show the general form of transitive verb composition with the reduced verb representation to be as follows:
−−−−−−−−→ SUBJ v OBJ = ∑ im cSUBJi c v imc OBJ m −→ni ⊗−→nm
= (∑ im cvim −→ni ⊗−→nm ) (∑ im cSUBJi c OBJ m −→ni ⊗−→nm )
= −→v (−−−→ SUBJ ⊗−−→OBJ )
To summarize what we have done with transitive verbs:
1. We have treated them as functions, taking two nouns and returning a sentence in a space ST ∼= N ⊗ N.
2. We have built them by counting what properties of subject and object noun vectors are related by the verb in transitive sentences in a corpus.
3. We have assumed that output properties were a function of input properties by the use of the Δijk function—for example, the weights associated with ni from the subject argument and with nk from the object argument only affect the “output” weight for the sentence basis element −→sj = −→ni ⊗−→nk .
4. We have shown that this leads to a compact representation of the verb’s semantic form, and an optimized learning procedure.
5. We have shown that the composition operations of our formalism can be adapted to this compact representation.
The compact representation and amended composition operation hinge on the choice of ST ∼= N ⊗ N as output type for N ⊗ N as an input type (the pair of arguments the verb takes), justifying our choice of a transitive sentence space. In the intransitive case, the same phenomenon can be observed, since such a verb takes as argument a vector in N and produces a vector in SI ∼= N. Furthermore, our choice to make all other types dependent on one base type—namely, n (with associated semantic space N)— yields the same property for every relational word we wish to learn: The output type is the same as the concatenated (on the syntactic level) or tensored (on the semantic level) input types. It is this symmetry between input and output types that guarantees that any m-ary relation, expressed in the original formulation as an element of tensor space N ⊗ . . .⊗ N︸ ︷︷ ︸
2m
, has a compact representation in N ⊗ . . .⊗ N︸ ︷︷ ︸ m , where the ith basis weight
of the reduced representation stands for the degree to which the ith element of the input vector affects the ith element of the output vector.
This allows the specification of the generalized learning algorithm for reduced representations, first presented in Grefenstette and Sadrzadeh (2011a), which is as follows. Each relational word P with grammatical type π and m adjoint types α1,α2, · · ·,αm is encoded as an (r × . . .× r) multi-dimensional array with m degrees of freedom (i.e., a rank m tensor). Because our vector space N has a fixed basis, each such array is represented in vector form as follows:
−→ P = ∑ ij · · ·ζ︸ ︷︷ ︸
m
cij···ζ ( −→n i ⊗−→n j ⊗ · · · ⊗ −→n ζ)︸ ︷︷ ︸
m
This vector lives in the tensor space N ⊗ N ⊗ · · · ⊗ N︸ ︷︷ ︸ m . Each cij···ζ is computed according to the procedure described in Figure 2.
Linear algebraically, this procedure corresponds to computing the following
−→ P = ∑ k (−→w1 ⊗−→w2 ⊗ · · · ⊗ −→wm)k
The general formulation of composing a relational word P with its arguments arg1, . . . , argm is now expressed as
−→ P (−−→arg1 ⊗ . . .⊗−−→argm)
For example, the computation of furry cats nag angry dogs would correspond to the following operations:
−−−−−−−−−−−−−−−−−−→ furry cats nag angry dogs = −−→nag ((−−−→ furry −→cat ) ⊗ (−−−→angry −−→dog))
This generalized learning algorithm effectively extends the coverage of our approach to any sentence for which we have a pregroup parse (e.g., as might be obtained by our translation mechanism from CCG). For example, determiners would have a type nnl, allowing us to model them as matrices in N ⊗ N; adverbs would be of type (nrs)r(nrs), and hence be elements of S ⊗ N ⊗ N ⊗ S. We could learn them using the procedure defined earlier, although for words like determiners and conjunctions, it is unclear whether we would want to learn such logical words or design them by hand, as was done in Coecke, Sadrzadeh, and Clark (2010) for negation. Nonetheless, the procedure given here allows us to generalize the work presented in this article to sentences of any length or structure based on the pregroup types of the words they contain. To conclude this section with an example taken from Grefenstette and Sadrzadeh (2011a), we demonstrate how the meaning of the word show might be learned from a corpus and then composed.
Suppose there are two instances of the verb show in the corpus: s1 = table show result s2 = map show location
The vector of show is
−−−→ show = −−→ table ⊗−−−→result + −−→map ⊗−−−−−→location
Consider a vector space N with four basis vectors far, room, scientific, and elect. The TF/IDF-weighted values for vectors of selected nouns (built from the BNC) are as shown in Table 1. Part of the matrix compact representation of show is presented in Table 2.
As a sample computation, the weight c11 for vector ( −→n1,−→n1 ), that is, ( −→ far, −→ far), is computed by multiplying weights of table and result on −→ far (6.6 × 7), multiplying weights of map and location on −→ far (5.6 × 5.9), and then adding these (46.2 + 33.04) and obtaining the total weight 79.24. Similarly, the weight c21 for vector ( −→n2,−→n1 ), that is, (−−−→room, −→ far), is computed by multiplying the weight of −−−→room for table by the weight of −→far for result (27 × 7), then multiplying the weight of −−−→room for map by the weight of −→far for location (7.4 × 5.9), and then adding these (189 + 43.66) to obtain the total weight of 232.66.
We now wish to compute the vector for the sentence [the] master shows [his] dog, omitting the determiner and possessive for simplicity, as we have left open the question as to whether or not we would want to learn them using the generalized procedure from Section 4.3 or specify them by design due to their logical structure. The calculation according to the vectors in Table 1 and the matrix in Table 2 will be:
−−−−−−−−−−−−→ master show dog
= −−−→ show (−−−−→ master ⊗−−→dog )
= ⎡ ⎢⎢⎣ 79.24 47.41 119.96 27.72 232.66 80.75 396.14 113.2 32.94 31.86 32.94 0
0 0 0 0
⎤ ⎥⎥⎦ ⎡ ⎢⎢⎣
23.65 32.68 0 0 50.05 69.16 0 0 22.55 31.16 0 0 34.1 47.12 0 0
⎤ ⎥⎥⎦
= ⎡ ⎢⎢⎣ 1,874.03 1,549.36 0 0 11,644.63 5,584.67 0 0
742.80 992.76 0 0 0 0 0 0
⎤ ⎥⎥⎦
Row-wise, flattening the final matrix representation gives us the result we seek, namely, the sentence vector in ST for [the] master showed [his] dog:
−−−−−−−−−−−−→ master show dog = [1,874, 1,549, 0, 0, 11,645, 5,585, 0, 0, 743, 993, 0, 0, 0, 0, 0, 0] 5 experiments :Evaluating compositional models of semantics is no easy task: First, we are trying to evaluate how well the compositional process works; second, we also are trying to determine how useful the final representation—the output of the composition—is, relative to our needs.
The scope of this second problem covers most phrase and sentence-level semantic models, from bag-of-words approaches in information retrieval to logic-based formal semantic models, via language models used for machine translation. It is heavily task dependent, in that a representation that is suitable for machine translation may not be appropriate for textual inference tasks, and one that is appropriate for IR may not be ideal for paraphrase detection. Therefore this aspect of semantic model evaluation ideally should take the form of application-oriented testing. For instance, to test semantic representations designed for machine translation purposes, we should use a machine translation evaluation task.
The DisCoCat framework (Clark, Coecke, and Sadrzadeh 2008; Coecke, Sadrzadeh, and Clark 2010), described in Section 3, allows for the composition of any words of any syntactic type. The general learning algorithm presented in Section 4.3 technically can be applied to learn and model relations of any semantic type. However, many open questions remain, such as how to deal with logical words, determiners, and quantification, and how to reconcile the different semantic types used for sentences with transitive and intransitive sentences. We will leave these questions for future work, briefly discussed in Section 6. In the meantime, we concretely are left with a way of satisfactorily modeling only simple sentences without having to answer these bigger questions.
With this in mind, in this section we present a series of experiments centered around evaluating how well various models of semantic vector composition perform (along with the one described in Section 4) in a phrase similarity comparison task. This task aims to test the quality of a compositional process by determining how well it forms a clear joint meaning for potentially ambiguous words. The intuition here is that tokens, on their own, can have several meanings; and that it is through the compositional process—through giving them context—that we understand their specific meaning. For example, bank itself could (among other meanings) mean a river bank or a financial bank; yet in the context of a sentence such as The bank refunded the deposit, it is likely we are talking about the financial institution.
In this section, we present three data sets designed to evaluate how well word-sense disambiguation occurs as a byproduct of composition. We begin by describing the first data set, based around noun-intransitive verb phrases, in Section 5.1. In Section 5.2, we present a data set based around short transitive-verb phrases (a transitive verb with subject and object). In Section 5.3, we discuss a new data set, based around short transitive-verb phrases where the subject and object are qualified by adjectives. We leave discussion of these results for Section 6. This first experiment, originally presented in Mitchell and Lapata (2008), evaluates the degree to which an ambiguous intransitive verb (e.g., draws) is disambiguated by combination with its subject.
Data set Description. The data set4 comprises 120 pairs of intransitive sentences, each of the form NOUN VERB. These sentence pairs are generated according to the following
4 Available at http://homepages.inf.ed.ac.uk/s0453356/results.
procedure, which will be the basis for the construction of the other data sets discussed subsequently:
1. A number of ambiguous intransitive verbs (15, in this case) are selected from frequently occurring verbs in the corpus.
2. For each verb V, two mutually exclusive synonyms V1 and V2 of the verb are produced, and each is paired with the original verb separately (for a total of 30 verb pairs). These are generated by taking maximally distant pairs of synonyms of the verb on WordNet, but any method could be used here.
3. For each pair of verb pairs (V, V1) and (V, V2), two frequently occurring nouns N1 and N2 are picked from the corpus, one for each synonym of V. For example, if V is glow and the synonyms V1 and V2 are beam and burn, we might choose face as N1, because a face glowing and a face beaming mean roughly the same thing; and fire as N2, because a fire glowing and a fire burning mean roughly the same thing.
4. By combining the nouns with the verb pairs, we form two high similarity triplets (V, V1, N1) and (V, V2, N2), and two low similarity triplets (V, V1, N2) and (V, V2, N1).
The last two steps can be repeated to form more than four triplets per pair of verb pairs. In Mitchell and Lapata (2008), eight triplets are generated for each pair of verb pairs, obtaining a total of 120 triplets from the 15 original verbs. Each triplet, along with its HIGH or LOW classification (based on the choice of noun for the verb pair) is an entry in the data set, and can be read as a pair of sentences: (V, Vi, N) translates into the intransitive sentences N V and N Vi.
Finally, the data set is presented, without the HIGH/LOW ratings, to human annotators. These annotators are asked to rate the similarity of meaning of the pairs of sentences in each entry on a scale of 1 (low similarity) to 7 (high similarity). The final form of the data set is a set of lines each containing:
Sentence 1 Sentence 2
butler bow butler submit head bow head stoop company bow company submit government bow government stoop
Evaluation Methodology. This data set is used to compare various compositional distributional semantic models, according to the following procedure:
1. For each entry in the data set, the representation of the two sentences N V and N Vi formed from the entry triple (V, Vi, N), which we will name S1 and S2, is constructed by the model.
2. The similarity of these sentences according to the model’s semantic distance measure constitutes the model score for the entry.
3. The rank correlation of entry model scores against entry annotator scores is calculated using Spearman’s rank correlation coefficient ρ.
The Spearman ρ scores are values between −1 (perfect negative correlation) and 1 (perfect correlation). The higher the ρ score, the higher the compositional model can be said to produce sentence representations that match human understanding of sentence meaning, when it comes to comparing the meaning of sentences. As such, we will rank the models evaluated using the task by decreasing order of ρ score.
One of the principal appealing features of Spearman’s ρ is that the coefficient is rank-based: It does not require models’ semantic similarity metrics to be normalized for a comparison to be made. One consequence is that a model providing excellent rank correlation with human scores, but producing model scores on a small scale (e.g., values between 0.5 and 0.6), will obtain a higher ρ score than a model producing model scores on a larger scale (e.g., between 0 and 1) but with less perfect rank correlation. If we wished to then use the former model in a task requiring some greater degree of numerical separation (let us say 0 for non-similar sentences and 1 for completely similar sentences), we could simply renormalize the model scores to fit the scale. By eschewing score normalization as an evaluation factor, we minimize the risk of erroneously ranking one model over another.
Finally, in addition to computing the rank alignment coefficient between model scores and annotator scores, Mitchell and Lapata (2008) calculate the mean model scores for entries labeled HIGH, and for entries labeled LOW. This information is reported in their paper as additional means for model comparison. However, for the same reason we considered Spearman’s ρ to be a fair means of model comparison— namely, in that it required no model score normalization procedure and thus was less likely to introduce error by adding such a degree of freedom—we consider the HIGH/LOW means to be inadequate grounds for comparison, precisely because it requires normalized model scores for comparison to be meaningful. As such, we will not include these mean scores in the presentation of this, or any further experiments in this article.
Models Compared. In this experiment, we compare the best and worst performing models of Mitchell and Lapata (2008) to our own. We begin by building a vector space W for all words in the corpus, using standard distributional semantic model construction procedures (n-word window) with the parameters of Mitchell and Lapata (2008). These parameters are as follows: The basis elements of this vector space are the 2,000 most frequently occurring words in the BNC, excluding common stop words. For our evaluation, the corpus was lemmatized using Carroll’s Morpha (Minnen, Carroll, and Pearce 2001), which was applied as a byproduct of our parsing of the BNC with C&C Tools (Curran, Clark, and Bos 2007).
The context of each occurrence of a word in the corpus was defined to be five words on either side. After the vector for each word w was produced, the basis weights cwi associated with context word bi were normalized by the following ratio of probabilities weighting scheme:
cwi = P(bi|w) P(w)
Let −−→ verb and −−→noun be the lexical semantic vectors for the verb and the noun, from W. It should be evident that there is no significant difference in using W for nouns in lieu of building a separate vector space N strictly for noun vectors.
As a first baseline, Verb Baseline, we ignored the information provided by the noun in constructing the sentence representation, effectively comparing the semantic content of the verbs:
Verb Baseline : −−−−−−→ noun verb = −−→ verb
The models from Mitchell and Lapata (2008) we evaluate here are those which were strictly unsupervised (i.e., no free parameters for composition). These are the additive model Add, wherein
Add : −−−−−−→ noun verb = −−→noun +−−→verb
and the multiplicative model Multiply, wherein
Multiply : −−−−−−→ noun verb = −−→noun −−→verb
Other models with parameters that must be optimized against a held-out section of the data set are presented in Mitchell and Lapata (2008). We omitted them here principally because they do not perform as well as Multiply, but also because the need to optimize the free parameters for this and other data sets makes fair comparison with completely unsupervised models more difficult, and less fair.
We also evaluate the Categorical model from Section 4 here, wherein
Categorical : −−−−−−→ noun verb = −−−→ verbcat −−→noun
where −−−→ verbcat is the compact representation of the relation the verb stands for, computed according to the procedure described in Section 4.3. It should be noted that for the case of intransitive verbs, this composition operation is mathematically equivalent to that of the Multiply model (as component-wise multiplication is a commutative operation), the difference being the learning procedure for the verb vector.
For all such vector-based models, the similarity of the two sentences s1 and s2 is taken to be the cosine similarity of their vectors, defined in Section 2.2:
similarity(s1, s2) = cosine( −→s1 ,−→s2 )
In addition to these vector-based methods, we define an additional baseline and an upper-bound. The additional baseline, Bigram Baseline, is a bigram-based language
model trained on the BNC with SRILM (Stolcke 2002), using the standard language model settings for computing log-probabilities of bigrams. To determine the semantic similarity of sentences s1 = noun verb1 and s2 = noun verb2, we assumed sentences have mutually conditionally independent properties, and computed the joint probability:
similarity(s1, s2) = logP(s1 ∧ s2) = log(P(s1)P(s2)) = logP(s1) + logP(s2)
The reasoning behind this baseline is as follows. The sentence formed by combining the first verb with its arguments is, by the design of this data set, a semantically coherent sentence (e.g., the head bowed and the government bowed both make sense). We therefore expect language models to treat this sort of bigram as having a higher probability than bigrams that are not semantically coherent sentences (and therefore unlikely to be observed in the corpus). A similar (relative) high probability is to be expected when the sentence formed by taking the second verb in each entry and combining it with the verb yields a sentence similar to the first in meaning (e.g., as would be the case with the head stooped), whereas the probability of a semantically incoherent sentence (e.g., the government stooped) is expected to be low relative to that of the first sentence. By taking the sum of log probabilities of sentences, we compute the log of the product of the probabilities, which we expect to be low when the probability of the second sentence is low, and high when it is high, with the probability of the first sentence acting as a normalizing factor. To summarize, while this bigram-based measure only tracks a tangential aspect of semantic similarity, it can be one which plays an artificially important role in experiments with a predetermined structure such as the one described in this section. For this reason, we use this bigram baseline for this experiment and all that follow.
The upper bound of the data set, UpperBound, was taken to be the inter-annotator agreement: the average of how each annotator’s score aligns with other annotator scores, using Spearman’s ρ.
Results. The results5 of the first experiment are shown in Table 4. As expected from the fact that Multiply and Categorical differ only in how the verb vector is learned, the results of these two models are virtually identical, outperforming both baselines and the Additive model by a significant margin. However, the distance from these models to the upper bound is even greater, demonstrating that there is still a lot of progress to be made. The first experiment was followed by a second similar experiment, in Mitchell and Lapata (2010), covering different sorts of composition operations for binary combinations of syntactic types (adjective-noun, noun-noun, verb-object). Such further experiments are interesting, but rather than continue down this binary road, we now turn to the development of our second experiment, involving sentences with larger syntactic structures, to examine how well various compositional models cope with more complex syntactic and semantic relations.
5 The results were state of the art when the experiments were run in 2011. The binary composition data set used here are popular; now there are a host of new state-of-the-art systems available in the literature. We invite the reader to check references to Mitchell and Lapata (2008) to find the current state of the art.
This second experiment, which we initially presented in Grefenstette and Sadrzadeh (2011a), is an extension of the first in the case of sentences centered around transitive verbs, composed with a subject and an object. The results of the first experiment did not demonstrate any difference between the multiplicative model, which takes into account no syntactic information or word ordering, and our syntactically motivated categorical compositional model. By running the same experiment over a new data set, where the relations expressed by the verb have a higher arity than in the first, we hope to demonstrate that added structure leads to better results for our syntax-sensitive model.
Data Set Description. The construction procedure for this data set6 is almost exactly as for the first data set, with the following differences:
Verbs are transitive instead of intransitive. We arbitrarily took 10 verbs from the most frequent verbs in the BNC, and for each verb, took two maximally distant synonyms (again, using WordNet) to obtain 10 pairs of pairs.
For each pair of verb pairs, we selected a set of subject and object nouns to use as context, as opposed to just a subject noun.
Each subject-object pair was manually chosen so that one of the verb pairs would have high similarity in the context of that subject and object, and the other would have low similarity. We used these choices to annotate the entries with HIGH and LOW tags.
Each combination of a verb pair with a subject-object pair constitutes an entry of our data set, of which there are 200.
As a form of quality control, we inserted “gold standard” sentences in the form of identical sentence pairs and rejected annotators who did not score these gold standard
6 The data set, reannotated by Turkers in 2013, is available at http://www.cs.ox.ac.uk/activities/compdistmeaning/GS2013data.txt.
sentences with a high score of 6 or 7.7 Lemmatized sentences from sample entries of this data set and our HIGH-LOW tags for them are shown in Table 5.
The data set was passed to a group of 50 annotators on Amazon Mechanical Turk, as for the previous data set. The annotators were shown, for each entry, a pair of sentences created by adding the in front of the subject and object nouns and putting the verbs in the past tense,8 and were instructed to score each pair of sentences on the same scale of 1 (not similar in meaning) to 7 (similar in meaning), based on how similar in meaning they believed the sentence pair was.
Evaluation Methodology. The methodology for this experiment is exactly that of the previous experiment. Models compositionally construct sentence representations, and compare them using a distance metric (all vector-based models once again used cosine similarity). The rank correlation of models scores with annotator scores is calculated using Spearman’s ρ, which is in turn used to rank models.
Models Compared. The models compared in this experiment are those of the first experiment, with the addition of an extra trigram-based baseline (trained with SRILM, using the addition of log-probability of the sentence as a similarity metric), and a variation on our categorical model, presented subsequently. With W as the distributional semantic space for all words in the corpus, trained using the same parameters as in the first experiment, and −−→ subj, −−→ verb, −−−→ object ∈ W as the vectors for subject, verb, and object of a sentence, respectively, and with verbcat as the compact representation of a transitive verb learned using the algorithm presented in this paper, we have the following compositional methods:
Add : −−−−−−−−−−−−→ subject verb object = −−−−→ subject + −−→ verb + −−−→ object Multiply : −−−−−−−−−−−−→ subject verb object = −−−−→ subject −−→verb −−−→object Categorical : −−−−−−−−−−−−→ subject verb object = −−−→ verbcat (−−−−→ subject ⊗−−−→object
) Kronecker : −−−−−−−−−−−−→ subject verb object = (−−→ verb ⊗−−→verb ) (−−−−→ subject ⊗−−−→object )
7 During the 2013 reannotation of this data set, we rejected no Turker contributions, as the answers to the gold standard sentence pairs were aligned with our expectations. We attribute this to our adding the requirement that Turkers be based in the US or the UK and have English as the first language. 8 For example, the entry draw table eye depict would yield the sentences The table drew the eye and The table depicted the eye.
All of these models have been explained earlier, with the exception of Kronecker. We first presented this new addition in Grefenstette and Sadrzadeh (2011b), where we observed that the compact representation of a verb in the DisCoCat framework, under the assumptions presented in Section 4, can be viewed as dim(N) × dim(N) matrices in N ⊗ N. We considered alternatives to the algorithm presented earlier for the construction of such matrices, and were surprised by the results of the Kronecker method, wherein we replaced the matrix learned by our algorithm with the Kronecker product of the lexical semantic vectors for the verb. Further analysis performed since the publication of that paper can help to understand why this method might work. Using the following property, for any vectors −→a ,−→b ,−→c ,−→d in a vector space
(−→a ⊗−→b ) (−→c ⊗−→d ) = (−→a −→c ) ⊗ (−→b −→d )
we can see that the Kronecker model’s composition operation can be expressed as
Kronecker : −−−−−−−−−−−−→ subject verb object = (−−→ verb −−−−→subject ) ⊗ (−−→ verb −−−→object )
Bearing in mind that the cosine measure we are using as a similarity metric is equivalent to the inner product of two vectors normalized by the product of their length
cosine(−→a ,−→b ) = 〈 −→a |−→b 〉
‖−→a ‖ × ‖−→b ‖
and the following property of the inner product of kronecker products
〈−→a ⊗−→b |−→c ⊗−→d 〉 = 〈−→a |−→c 〉 × 〈−→b |−→d 〉
we finally observe that comparing two sentences under Kronecker corresponds to the following computation:
cosine( −−−−−−−−−−−−→ subject verb1 object, −−−−−−−−−−−−→ subject verb2 object)
= α 〈(−−→ verb1 ⊗ −−→ verb1 ) (−−−−→ subject ⊗−−−→object ) | (−−→ verb2 ⊗ −−→ verb2 ) (−−−−→ subject ⊗−−−→object )〉
= α 〈(−−→ verb1 −−−−→ subject ) ⊗ (−−→ verb1 −−−→ object ) | (−−→ verb2 −−−−→ subject ) ⊗ (−−→ verb2 −−−→ object )〉 = α 〈(−−→ verb1 −−−−→ subject ) | (−−→ verb2 −−−−→ subject )〉〈(−−→ verb1 −−−→ object ) | (−−→ verb2 −−−→ object
)〉 where α is the normalization factor
α = 1 ‖−−−−−−−−−−−−→subject verb1 object‖ × ‖ −−−−−−−−−−−−→ subject verb2 object‖
We note here that the Kronecker is effectively a parallel application of the Multiply model, combining the subject and verb, and object and verb separately. Within the context of this task, the comparison of two sentences boils down to how well each verb combines with the subject multiplied by how well it combines with the object, with the
joint product forming the overall model score. In short, it more or less constitutes the introduction of some mild syntactic sensitivity into the multiplicative model of Mitchell and Lapata (2008), although it remains to be seen how well this scales with syntactic complexity (e.g., extended to ditransitive verbs, or other relations of higher arity).
The UpperBound here is, again, the inter-annotator agreement, calculated by computing the pairwise alignment of each annotator with every other, and averaging the resulting rank correlation coefficients.
Results. The results of the second experiment are shown in Table 6. The baseline scores fall in the 0.14–0.16 range, with best results being obtained for the Bigram Baseline and Trigram Baseline, the difference between both models not being statistically significant. The additive model Add performs on par with the first experiment. The multiplicative model Multiply and our Categorical model perform on par with the version used for the intransitive experiment, but obtain a score comparable to the baselines. The best performing model here is our newly introduced Kronecker model, which leads the pack by a steady margin, with a score of 0.26. The inter-annotator agreement UpperBound is much higher in this experiment than in the previous experiment, indicating even more room for improvement.
We therefore learn here that the Categorical model continues to operate on par with the multiplicative model, and that the Kronecker model provides the highest results, while being as simple in its construction as the multiplicative model, requiring only the learning of lexical vectors for the verb. The third and final experiment we present is a modified version of the second data set presented earlier, where the nouns in each entry are under the scope of adjectives applied to them.The intuition behind the data sets presented in Section 5.1 and Section 5.2 was that ambiguous verbs are disambiguated through composition with nouns. These nouns themselves may also be ambiguous, and a good compositional model will be capable of separating the noise produced by other meanings through its compositional mechanism to produce unambiguous phrase representations. The intuition behind this data set is similar, in that adjectives provide both additional information for disambiguation of the nouns they apply to, but also additional semantic noise. Therefore, a
good model will also be able to separate the useful information of the adjective from its semantic noise when composing it with its argument, in addition to doing this when composing the noun phrases with the verb.
Data Set Description. The construction procedure for this data set was to take the data set from Section 5.2, and, for each entry, add a pair of adjectives from those most frequently occurring in the corpus. The first adjective from the pair is applied to the first noun (subject) of the entry when forming the sentences, and the second adjective is applied to the second noun (object). For each entry, we chose adjectives which best preserved the meaning of the phrase constructed by combining the first verb with its subject and object.
This new data set9 was then annotated again by a group of 50 annotators using Amazon’s Mechanical Turk service. The annotators were shown, for each entry, a pair of sentences created by adding the in front of the subject and object noun phrases and putting the verbs in the past tense,10 and asked to give each sentence pair a meaning similarity score between 1 and 7, as for the previous data sets. We applied the same quality control mechanism as in the second experiment. Some 94 users returned annotations, of which we kept 50 according to our gold standard tests. We are unaware of whether or not it was applied in the production of the first data set, but believe that this can only lead to the production of higher quality annotations.
Sample sentences from this data set are shown in Table 7.
Evaluation Methodology. The evaluation methodology in this experiment is identical to that of the previous experiments.
Models Compared. In this experiment, in lieu of simply comparing compositional models “across the board” (e.g., using the multiplicative model for both adjective-noun composition and verb-argument composition), we experimented with different combinations of models. This evaluation procedure was chosen because we believe that adjective-noun composition need not necessarily be the same kind of compositional process as subjectverb-object composition, and also because different models may latch onto different semantic features during the compositional process, and it would be interesting to see what model mixtures work well together. Naturally, this is not a viable approach to selecting composition operations in general, as we will not have the luxury of trying every combination of composition operations for every combination of syntactic types, but it is worthwhile performing these tests to at least verify the hypothesis that
9 Available at http://www.cs.ox.ac.uk/activities/compdistmeaning/GS2012data.txt. 10 For example, the entry statistical table show express good result would yield the sentences The statistical table
showed the good result and The statistical table expressed the good result.
operation-specific composition operations (or parameters) is a good thing. Notably, this idea has been explored very recently, within the context of deep learning networks, by Chen et al. (2013).
Each mixed model has two components: a verb-argument composition model and a adjective-noun composition model. For verb-argument composition, we used the three best models from the previous experiment, namely, Multiply, Categorical, and Kronecker. For adjective composition we used three different methods of adjectivenoun composition. With −−−−−→ adjective and −−→noun being the vectors for an adjective and a noun in our distributional lexical semantic space W (built using the same procedure as the previous experiments) and −−→ adjcat being the compact representation in the Categorical model, built according to the algorithm from Section 4.3, we have the following models:
AdjMult : −−−−−−−−−→ adjective noun = −−−−−→ adjective −−→noun Categorical : −−−−−−−−−→ adjective noun = −−−−−−→ adjectivecat −−→noun
The third model, AdjNoun, is a holistic (non-compositional) model, wherein the adjective-noun compound was treated as a single token, as its semantic vector−−−−−−−−−−−−→ (adjective noun)lex ∈ W was learned from the corpus using the same learning procedure applied to construct other vectors in W. Hence, the model defines adjective-noun “composition” as:
AdjNoun : −−−−−−−−−→ adjective noun = −−−−−−−−−−−−→ (adjective noun)lex
In addition to these models, we also evaluated three baselines: Verb Baseline, Bigram Baseline, and Trigram Baseline. As in previous experiments, the verb baseline uses the verb vector as sentence vector, ignoring the information provided by other words. The bigram and trigram baselines are calculated from the same language model as used in the second experiment. In both cases, the log-probability of each sentence is calculated using SRLIM, and the sum of log-probabilities of two sentences is used as a similarity measure.
Finally, for comparison, we also considered the full additive model:
Additive : −−−−−→ sentence = −−−−−→ adjective1 + −−−→noun1 + −−→ verb + −−−−−→ adjective2 + −−−→noun2
Results. The results for the third experiment are shown in Table 8. The best performing adjective-noun combination operations for each verb-argument combination operation are shown in bold. Going through the combined models, we notice that in most cases the results stay the same whether the adjective-noun combination method is AdjMult or CategoricalAdj. This is because, as was shown in the first experiment, composition of a unary-relation such as an adjective or intransitive verb with its sole argument, under the categorical model with reduced representations, is mathematically equivalent to the multiplicative model. The sole difference is the way the adjective or intransitive verb vector is constructed. We note, however, that with Categorical as a verb-argument composition method, the CategoricalAdj outperforms AdjMult by a non-negligible margin (0.19 vs. 0.14), indicating that the difference in learning procedure can lead to different results depending on what other models it is combined with.
Overall, the best results are obtained for AdjMult+Kronecker (ρ = 0.26) and CategoricalAdj+Kronecker (ρ = 0.27). Combinations of the adjective composition methods with other composition methods at best matches the best-performing baseline, Verb Baseline. In all cases, the holistic model AdjNoun provides the worst results. 6 discussion :In this article, we presented various approaches to compositionality in distributional semantics. We discussed what mathematical operations could underlie vector combination, and several different ways of including syntactic information into the combinatory process. We reviewed an existing compositional framework that leverages the ability to communicate information across mathematical structures provided by category theory in order to define a general way of linking syntactic structure to syntax-sensitive composition operations. We presented concrete ways to apply this framework to linguistic tasks and evaluate it, and developed algorithms to construct semantic representations for words and relations within this framework. In this section, we first briefly comment upon the combined results of all three experiments, and then conclude by discussing what aspects of compositionality require further attention, and how experiment design should adapt towards this goal.
Results Commentary. We evaluated this framework against other unsupervised compositional distributional models, using non-compositional models and n-gram language models as baselines, within the context of three experiments. These experiments show that the concrete categorical model developed here, and the Kronecker-based variant
presented alongside it, outperform all other models in each experiment save the first, where they perform on par with what was the leading model at the time this experiment was performed (namely, 2011). As the experiments involved progressively more syntactically complicated sentences, the increased reliability of our categorical approaches relative to competing models as sentence complexity rises seems to indicate that both the categorical and Kronecker model successfully leverage the added information provided by additional terms and syntactic structures.
The third experiment also served to show that using different combinations of composition operations, depending on the syntactic type of the terms being combined, can yield better results, and that some models combine better than others. Notably, the adjective-noun combination models AdjMult and CategoricalAdj, despite their mathematical similarity, produce noticeably different results when combined with the categorical verb-argument composition operation, while they perform equally with most other verb-argument composition operations. We can conclude that different models combine different semantic aspects more prominently than others, and that through combination we can find better performance by assuming that different kinds of composition play on different semantic properties. For example, predicates such as intersective adjectives add information to their argument (a red ball is a ball that is also red). This raises the question of how to design models of composition that systematically select which operations will match the semantic aspects of the words being combined based on their syntactic type. This is an open question, which we believe warrants further investigation.
Overall, we observe two advantages of these models over those presented in the early part of this article, and those evaluated in the later part. First, we have shown that a linguistically motivated and mathematically sound framework can be implemented and learned. Second, we showed that simple methods such as the Kronecker model perform well, especially when combined with the multiplicative model (effectively a unary instance of the Kronecker model) for sentences with adjectives. Considering the “simple is best” position recently argued for experimentally by Blacoe and Lapata (2012), this is an interesting candidate for dealing with binary relation composition.
Future Work. These experiments showcased the ability of various compositional models to produce sentence representations that were less ambiguous than the words that formed them. They also more generally demonstrated that concrete models could be built from the general categorical framework and perform adequately in simple paraphrase detection tasks. However, various aspects of compositionality were not evaluated here. First, syntax sensitivity is not as important in these experiments as it might be in “real” language. For instance, whereas models that treat adjectives, nouns, and verbs differently tended to perform better than those that did not, the actual capacity of a model to use the syntactic structure was not tested. For instance, models that ignore word order and syntax altogether were not particularly penalized, as might be done if some sentences were simply the reverse of another sentence they are paired with (where syntax insensitive models would erroneously give such sentence pairs a high similarity score). One way of testing word order while expanding the data set was suggested by Turney (2012), who for every entry in a phrase similarity test such as Mitchell and Lapata’s, discussed herein, creates a new entry where one of the sentences has the order of its words reversed, and is assigned an artificial annotator score of 1 (the minimum). This is an interesting approach, but we would prefer to see such reversals designed into the data set and seen by annotators, for instance,
in the recent data set presented in Pham et al. (2013). This data set has entries such as guitar played man and man played guitar, where the subjects and objects of verbs are indeed reversed but the two sentences do not necessarily express opposite meanings; we will be looking into expanding such data sets to ones where both sentences make sense and have opposite meanings, such as in the pair man bites dog and dog bites man.
Furthermore, sentences all have the exact same sentence structure, and therefore words are aligned. Models that might indirectly benefit from this alignment, such as Kronecker, may have been given an advantage due to this, as opposed to models such as Categorical, which are designed to compare sentences of different length or syntactic structure, when resolving the difference between intransitive and transitive sentence spaces as has been done in Grefenstette et al. (2011). Future experiments should aim to do away with this automated alignment, and include sentence comparisons that would penalize models which do not leverage syntactic information.
Furthermore, each of these experiments dealt with the comparison of a certain type of sentence (transitive, intransitive). This was convenient for our concrete categorical model, as we defined the sentence space differently based on the valency of the head verb. However, sentences with different sorts of verbs should be able to be directly compared. Not only do several models, both non-syntax sensitive (additive, multiplicative) and syntax-sensitive (Baroni and Zamparelli 2010; Socher et al. 2012), not face this problem, as the product of composition is either naturally in the same space as the vectors being composed or is projected back into it, but the categorical framework the concrete categorical models were derived from does not commit us to different sentence spaces either. The topic of how to solve this problem for the concrete models developed here is beyond the scope of this article, but various options exist, such as the projection of ST into SI, the embedding of SI into ST, the use of a combined sentence space S = SI ⊕ ST, and so on. It is clear that future experiments should not give this degree of convenience to the models being evaluated (and their designers), by comparing sentences of different types, lengths, and structure so as to more fairly evaluate the capacity of models to produce meaningful semantic representations in a single space, which can be directly compared regardless of syntactic structure and verb-type.
Finally, we mentioned at the beginning of Section 5 that evaluations need to be application-oriented. The class of experiments presented in this article specifically see how well disambiguation occurs as a byproduct of composition. We presented concrete models that would offer ways of combining relations underlying verbs and adjectives, and tested how well these relations disambiguated their arguments and were in turn disambiguated by their arguments. We did not address other aspects of language, such as quantification, logical operations, relative clauses, intensional clauses, embedded sentences, and many other linguistic aspects of spoken and written language that have complex syntactic structure and complicated semantic roles. Addressing these sorts of problems requires determining how negation, conjunction, and other logical relations should be modeled within compositional distributional formalisms, a problem we share with other similar approaches. The development of newer models within the categorical framework we based our work on, and the further development of other approaches mentioned here, must be driven by new experimental goals different from those offered by the task and experiments discussed in this article. Thinking not only about how to address these aspects of language within compositional models, but how to evaluate them, should be a priority for all those interested in further developing this field."
,,,,,,,,"Although Prolog was intended to be used in building natural language processing (NLP) applications, the language has often been neglected in many modern NLP introductory courses and texts. Because of the popularity of machine learning and statistical approaches to language processing problems over the past decades, researchers and practitioners tend to use other modern programming languages such as C++ and Java for developing NLP applications. However, with a growing interest in semantic processing and knowledge base construction in recent years, researchers have found Prolog to be an indispensable tool for tasks such as searching databases and performing symbolic reasoning. We have already seen recent successful deployment of Prolog-based NLP systems in industry. An example would be the use of Prolog in IBM’s DeepQA project to express pattern-matching rules.1 Pierre M. Nugues’ comprehensive textbook Language Processing with Perl and Prolog provides a timely, in-depth introduction to the field of NLP using Prolog and, to a lesser extent, Perl. This book is suitable for anyone wanting to enter NLP through learning to prototype NLP modules in Prolog.","[{""affiliations"": [], ""name"": ""Pierre M. Nugues""}]",SP:33e4b5aa7f18907299f2e831a450682f8afc4ebb,[],,,,,,,,,,,,,,,,book review :,,,"reviewed by wei lu singapore university of technology and design :Although Prolog was intended to be used in building natural language processing (NLP) applications, the language has often been neglected in many modern NLP introductory courses and texts. Because of the popularity of machine learning and statistical approaches to language processing problems over the past decades, researchers and practitioners tend to use other modern programming languages such as C++ and Java for developing NLP applications. However, with a growing interest in semantic processing and knowledge base construction in recent years, researchers have found Prolog to be an indispensable tool for tasks such as searching databases and performing symbolic reasoning. We have already seen recent successful deployment of Prolog-based NLP systems in industry. An example would be the use of Prolog in IBM’s DeepQA project to express pattern-matching rules.1 Pierre M. Nugues’ comprehensive textbook Language Processing with Perl and Prolog provides a timely, in-depth introduction to the field of NLP using Prolog and, to a lesser extent, Perl. This book is suitable for anyone wanting to enter NLP through learning to prototype NLP modules in Prolog.
© 2015 Association for Computational Linguistics
I approve of how Nugues has structured his textbook. It starts from first principles, moves on to concepts such as words, syntax, and semantics, and concludes with a discussion on the more advanced and open topics of discourse and dialogues.
Chapter 1 gives an overview of the field of natural language processing. Chapters 2 and 3 then provide a technical discussion of corpus processing and data representations.
Chapter 4 focuses on information theory and machine learning. The important concepts of entropy and perplexity are introduced first. They are followed by a discussion on the learning of a decision tree model. One minor quibble is that general machine learning concepts such as supervised and unsupervised learning are presented under the subsection “Machine Learning” within “Entropy and Decision Trees.” A brief introduction to “Linear Models” is given before moving on to regression and classification models, including perceptrons, support vector machines, and logistic regression. I liked the way that such concepts were elucidated—I thought the use of automatic language detection as an example to illustrate how the different models worked was a nice
1 http://www.cs.nmsu.edu/ALP/2011/03/natural-language-processing-with-prolog-in-the-ibmwatson-system/.
doi:10.1162/COLI r 00238
touch. Although advanced readers might find this section unsatisfying since it does not cover more sophisticated models and algorithms, I believe this section is ideal for beginners, as it covers the most fundamental and essential topics for computational linguistics.
Chapters 5 to 8 discuss word-level concepts. Chapter 5 touches on some fundamental issues related to word-level processing, including how to count the words in sentences, tokenization issues, and how to model word sequences. Next, the concepts of parts of speech (POS), lexicon, and morphology are introduced in Chapter 6. An indepth treatment of morphology is also given in this chapter, which includes a review of finite state automata and finite state transducers. The chapter mainly focuses on European languages, but I believe additional discussions on other language families would be helpful. Chapter 7 presents rule-based techniques for performing POS tagging. In this chapter, Nugues describes Brill’s tagger, one of the earliest successful approaches to POS tagging. Recent developments in multilingual POS tag sets are also introduced. In Chapter 8, statistical approaches to POS tagging are discussed. Methods based on noisy-channel models, hidden Markov models (HMMs), perceptrons, and decision trees are presented. HMMs, in particular, are illustrated with detailed inference and decoding algorithms and well-chosen examples. Although HMMs are influential graphical models, I feel that its discriminative counterpart, linear-chain conditional random fields (CRFs), which are now widely used in the NLP community, could also be covered in the chapter. At the end of the chapter, two applications of the noisy-channel model are given. One is on spellchecking, and the other on machine translation. The treatment of machine translation in this example constitutes most of the discussions of this topic in the text.
Chapters 9 to 13 focus on syntactic parsing. Chapter 9 begins with the discussion of the definite clause grammar (DCG) used in Prolog to define grammars. DCG is then linked to phrase-structure rules used for syntactic parsing of natural languages. Extensive Prolog code examples on parsing based on DCG are presented in this chapter. A small section in Chapter 9 is devoted to the interesting topic of semantic parsing, where the λ-calculus as a semantic formalism is introduced and its implementation discussed. The topics of the next few chapters include partial parsing, constituency parsing, and dependency parsing.
Chapter 10 discusses partial parsing (also known as shallow parsing or chunking). In this chapter, Nugues explains the many concepts related to information extraction, with a focus on named entities and multiword expressions. Real examples taken from the CONLL shared tasks are used when introducing such concepts.
Chapter 11 gives an overview of syntactic formalisms. The chapter starts with an introduction to Chomsky’s ideas on constituency parsing, followed by unification-based grammars. Next, dependency grammars are introduced in detail. At the end of the chapter, the connections between constituency grammars and dependency grammars are highlighted.
Chapter 12 focuses on constituency parsing: Classic PCFG parsing algorithms such as the CYK algorithm and the Earley algorithm are discussed. These algorithms are treated in a non-probabilistic manner and are illustrated extensively in Prolog. However, not much is written about probabilistic constituency parsing apart from the subsection “Adding Probabilities to the CYK Parser,” which contains no code examples. Nugues also writes about lexicalized PCFG parsing algorithms, with special attention given to Charniak’s parser. In his discussion of dependency parsing in Chapter 13, Nugues covers Nivre’s shift-reduce style parser, Covington’s non-projective parser, as well as Eisner’s cubic-time projective parser.
After discussing syntactic processing, Nugues moves on to semantics in Chapters 14 and 15. The former focuses on formal semantics and the latter on lexical semantics. Chapter 14 starts with predicate logic, which can be naturally implemented with Prolog. The λ-calculus expressions are then revisited, followed by a wide range of technical issues related to representing words and phrases with logical forms. The issue of using semantic representations to interact with databases for question-answering is also covered in this chapter. The area of semantic parsing—mapping natural languages to their semantic representations or denotations—has recently received a lot of attention. Although some recent developments in the field are not covered in these sections, the discussion should already serve as a good introduction to readers interested in this growing area.
Chapter 15 presents lexical semantics. Concepts such as ontology and case grammar are discussed. Some useful resources such as WordNet, FrameNet, and Propositional Bank are also introduced here.
In the last two chapters, Nugues takes one step further by moving language processing beyond the sentence level. Chapter 16 focuses on discourse analysis. Topics under this theme include coreference, rhetoric, as well as event and temporal information processing. The final chapter focuses on the more advanced topic of dialogues. This can be regarded as a shift from a single, static monologue to dynamic, interacting dialogues. Many open issues remain to be addressed on such a topic, but some discussions on building simple dialogue systems are given in this chapter.
The Appendix explains some of the concepts of Prolog. Since the information contained in the Appendix is necessary to understand the examples within the text, it might be better to have it at the front of the book.
Overall, Nugues presents us with a very useful and comprehensive textbook with adequate depth and breadth. The book presents a nice synthesis of actionable code and theory and its well-considered structure allows different concepts in NLP to be strung together. This text would serve as a solid stepping-stone for NLP beginners, and is therefore suitable for graduate or senior undergraduate students who are interested in entering into the field. It would also serve as an excellent reference for researchers as well as a good guide for practitioners who are interested in building NLP applications.
Wei Lu is an assistant professor at Singapore University of Technology and Design (SUTD). His primary research interest is in the application of machine learning techniques to natural language processing problems. He is particularly interested in the modeling of semantics of natural language texts. Lu’s email address is luwei@sutd.edu.sg.",,,,,,,,,,,,,,,,,,,,,,"language processing with perl and prolog: theories, implemetation, and application :Pierre M. Nugues (Lund University, Sweden)
Springer (Cognitive technologies series, edited by A. Bundy et. al), 2014, Second Edition, xxv+662 pp; hardcover, ISBN 978-3-642-41463-3, $89.99; ebook, ISBN 978-3-642-41464-0, $69.99; doi 10.1007/978-3-642-41464-0",,,,,,,,,"Although Prolog was intended to be used in building natural language processing (NLP) applications, the language has often been neglected in many modern NLP introductory courses and texts. Because of the popularity of machine learning and statistical approaches to language processing problems over the past decades, researchers and practitioners tend to use other modern programming languages such as C++ and Java for developing NLP applications. However, with a growing interest in semantic processing and knowledge base construction in recent years, researchers have found Prolog to be an indispensable tool for tasks such as searching databases and performing symbolic reasoning. We have already seen recent successful deployment of Prolog-based NLP systems in industry. An example would be the use of Prolog in IBM’s DeepQA project to express pattern-matching rules.1 Pierre M. Nugues’ comprehensive textbook Language Processing with Perl and Prolog provides a timely, in-depth introduction to the field of NLP using Prolog and, to a lesser extent, Perl. This book is suitable for anyone wanting to enter NLP through learning to prototype NLP modules in Prolog. [{""affiliations"": [], ""name"": ""Pierre M. Nugues""}] SP:33e4b5aa7f18907299f2e831a450682f8afc4ebb [] book review : reviewed by wei lu singapore university of technology and design :Although Prolog was intended to be used in building natural language processing (NLP) applications, the language has often been neglected in many modern NLP introductory courses and texts. Because of the popularity of machine learning and statistical approaches to language processing problems over the past decades, researchers and practitioners tend to use other modern programming languages such as C++ and Java for developing NLP applications. However, with a growing interest in semantic processing and knowledge base construction in recent years, researchers have found Prolog to be an indispensable tool for tasks such as searching databases and performing symbolic reasoning. We have already seen recent successful deployment of Prolog-based NLP systems in industry. An example would be the use of Prolog in IBM’s DeepQA project to express pattern-matching rules.1 Pierre M. Nugues’ comprehensive textbook Language Processing with Perl and Prolog provides a timely, in-depth introduction to the field of NLP using Prolog and, to a lesser extent, Perl. This book is suitable for anyone wanting to enter NLP through learning to prototype NLP modules in Prolog.
© 2015 Association for Computational Linguistics
I approve of how Nugues has structured his textbook. It starts from first principles, moves on to concepts such as words, syntax, and semantics, and concludes with a discussion on the more advanced and open topics of discourse and dialogues.
Chapter 1 gives an overview of the field of natural language processing. Chapters 2 and 3 then provide a technical discussion of corpus processing and data representations.
Chapter 4 focuses on information theory and machine learning. The important concepts of entropy and perplexity are introduced first. They are followed by a discussion on the learning of a decision tree model. One minor quibble is that general machine learning concepts such as supervised and unsupervised learning are presented under the subsection “Machine Learning” within “Entropy and Decision Trees.” A brief introduction to “Linear Models” is given before moving on to regression and classification models, including perceptrons, support vector machines, and logistic regression. I liked the way that such concepts were elucidated—I thought the use of automatic language detection as an example to illustrate how the different models worked was a nice
1 http://www.cs.nmsu.edu/ALP/2011/03/natural-language-processing-with-prolog-in-the-ibmwatson-system/.
doi:10.1162/COLI r 00238
touch. Although advanced readers might find this section unsatisfying since it does not cover more sophisticated models and algorithms, I believe this section is ideal for beginners, as it covers the most fundamental and essential topics for computational linguistics.
Chapters 5 to 8 discuss word-level concepts. Chapter 5 touches on some fundamental issues related to word-level processing, including how to count the words in sentences, tokenization issues, and how to model word sequences. Next, the concepts of parts of speech (POS), lexicon, and morphology are introduced in Chapter 6. An indepth treatment of morphology is also given in this chapter, which includes a review of finite state automata and finite state transducers. The chapter mainly focuses on European languages, but I believe additional discussions on other language families would be helpful. Chapter 7 presents rule-based techniques for performing POS tagging. In this chapter, Nugues describes Brill’s tagger, one of the earliest successful approaches to POS tagging. Recent developments in multilingual POS tag sets are also introduced. In Chapter 8, statistical approaches to POS tagging are discussed. Methods based on noisy-channel models, hidden Markov models (HMMs), perceptrons, and decision trees are presented. HMMs, in particular, are illustrated with detailed inference and decoding algorithms and well-chosen examples. Although HMMs are influential graphical models, I feel that its discriminative counterpart, linear-chain conditional random fields (CRFs), which are now widely used in the NLP community, could also be covered in the chapter. At the end of the chapter, two applications of the noisy-channel model are given. One is on spellchecking, and the other on machine translation. The treatment of machine translation in this example constitutes most of the discussions of this topic in the text.
Chapters 9 to 13 focus on syntactic parsing. Chapter 9 begins with the discussion of the definite clause grammar (DCG) used in Prolog to define grammars. DCG is then linked to phrase-structure rules used for syntactic parsing of natural languages. Extensive Prolog code examples on parsing based on DCG are presented in this chapter. A small section in Chapter 9 is devoted to the interesting topic of semantic parsing, where the λ-calculus as a semantic formalism is introduced and its implementation discussed. The topics of the next few chapters include partial parsing, constituency parsing, and dependency parsing.
Chapter 10 discusses partial parsing (also known as shallow parsing or chunking). In this chapter, Nugues explains the many concepts related to information extraction, with a focus on named entities and multiword expressions. Real examples taken from the CONLL shared tasks are used when introducing such concepts.
Chapter 11 gives an overview of syntactic formalisms. The chapter starts with an introduction to Chomsky’s ideas on constituency parsing, followed by unification-based grammars. Next, dependency grammars are introduced in detail. At the end of the chapter, the connections between constituency grammars and dependency grammars are highlighted.
Chapter 12 focuses on constituency parsing: Classic PCFG parsing algorithms such as the CYK algorithm and the Earley algorithm are discussed. These algorithms are treated in a non-probabilistic manner and are illustrated extensively in Prolog. However, not much is written about probabilistic constituency parsing apart from the subsection “Adding Probabilities to the CYK Parser,” which contains no code examples. Nugues also writes about lexicalized PCFG parsing algorithms, with special attention given to Charniak’s parser. In his discussion of dependency parsing in Chapter 13, Nugues covers Nivre’s shift-reduce style parser, Covington’s non-projective parser, as well as Eisner’s cubic-time projective parser.
After discussing syntactic processing, Nugues moves on to semantics in Chapters 14 and 15. The former focuses on formal semantics and the latter on lexical semantics. Chapter 14 starts with predicate logic, which can be naturally implemented with Prolog. The λ-calculus expressions are then revisited, followed by a wide range of technical issues related to representing words and phrases with logical forms. The issue of using semantic representations to interact with databases for question-answering is also covered in this chapter. The area of semantic parsing—mapping natural languages to their semantic representations or denotations—has recently received a lot of attention. Although some recent developments in the field are not covered in these sections, the discussion should already serve as a good introduction to readers interested in this growing area.
Chapter 15 presents lexical semantics. Concepts such as ontology and case grammar are discussed. Some useful resources such as WordNet, FrameNet, and Propositional Bank are also introduced here.
In the last two chapters, Nugues takes one step further by moving language processing beyond the sentence level. Chapter 16 focuses on discourse analysis. Topics under this theme include coreference, rhetoric, as well as event and temporal information processing. The final chapter focuses on the more advanced topic of dialogues. This can be regarded as a shift from a single, static monologue to dynamic, interacting dialogues. Many open issues remain to be addressed on such a topic, but some discussions on building simple dialogue systems are given in this chapter.
The Appendix explains some of the concepts of Prolog. Since the information contained in the Appendix is necessary to understand the examples within the text, it might be better to have it at the front of the book.
Overall, Nugues presents us with a very useful and comprehensive textbook with adequate depth and breadth. The book presents a nice synthesis of actionable code and theory and its well-considered structure allows different concepts in NLP to be strung together. This text would serve as a solid stepping-stone for NLP beginners, and is therefore suitable for graduate or senior undergraduate students who are interested in entering into the field. It would also serve as an excellent reference for researchers as well as a good guide for practitioners who are interested in building NLP applications.
Wei Lu is an assistant professor at Singapore University of Technology and Design (SUTD). His primary research interest is in the application of machine learning techniques to natural language processing problems. He is particularly interested in the modeling of semantics of natural language texts. Lu’s email address is luwei@sutd.edu.sg. language processing with perl and prolog: theories, implemetation, and application :Pierre M. Nugues (Lund University, Sweden)
Springer (Cognitive technologies series, edited by A. Bundy et. al), 2014, Second Edition, xxv+662 pp; hardcover, ISBN 978-3-642-41463-3, $89.99; ebook, ISBN 978-3-642-41464-0, $69.99; doi 10.1007/978-3-642-41464-0",
"1 introduction :Metaphor enriches our communication with a more diverse imagery and provides an important mechanism for reasoning about concepts. At the same time, it is also a very common linguistic device that has long become a part of our everyday language. Metaphors arise through systematic associations between distinct, and seemingly unrelated, concepts. For example, when we say “The wheels of Stalin’s regime were well-oiled and already turning,” we view a political system in terms of a mechanism, it can function, break, have wheels, and so forth. The existence of this association allows us to transfer knowledge and inferences from the domain of mechanisms to that of political systems. As a result, we reason about political systems in terms of mechanisms and discuss them using the mechanism terminology, giving rise to a variety of metaphorical expressions.
This view of metaphor is widely known as Conceptual Metaphor Theory (CMT). It was proposed by Lakoff and Johnson (1980), who claimed that metaphor is not merely a property of language, but rather a cognitive mechanism that structures our conceptual system in a certain way. Lakoff and Johnson explained metaphor through a presence of a mapping between two domains of experience: the target (e.g., politics) and the source (e.g., mechanism). Metaphor is thus not limited to meaning extensions of individual
∗ International Computer Science Institute, 1947 Center St. Ste. 600, Berkeley, CA 94704, USA. E-mail: katia@icsi.berkeley.edu.
Submission received: 9 October 2013; revised version received: 4 March 2015; accepted for publication: 10 June 2015.
doi:10.1162/COLI a 00233
© 2015 Association for Computational Linguistics
words, but rather involves a complex cross-domain knowledge projection process. Let us consider a few more examples.
(1) “President Obama is rebuilding the campaign machinery that vaulted him into office” (New York Times 2011)
(2) 20 steps towards a modern, working democracy
(3) Time to mend our foreign policy.
(4) “She knows the nuts and bolts, and it’s not the nuts and bolts inside legislation, it’s the nuts and bolts of raising money, preparing the party for elections, a political consultant kind of politics.” (Bzdek 2008)
These examples demonstrate how multiple properties and inferences from the domain of mechanisms are systematically projected onto our knowledge about politics. Lakoff and Johnson coined the term conceptual metaphor to describe such mappings from the source domain to the target. The view of an inter-conceptual mapping as a basis of metaphor was echoed by other prominent theories in the field. These include, most notably, the comparison view, formulated in the Structure–Mapping Theory of Gentner (1983), and the interaction view (Black 1962; Hesse 1966). However, the principles of CMT have inspired and influenced much of the computational work on metaphor, thus becoming more central to this paper. Conceptual metaphor manifests itself in language in the form of linguistic metaphor, or metaphorical expressions. These in turn include lexical metaphor, that is, single-word meaning extensions (as in Examples (2) and (3)); multi-word metaphorical expressions (e.g., “the government turned a blind eye to corruption”); or extended metaphor, that spans longer discourse fragments.
Manifestations of metaphor are frequent in language, appearing on average in every third sentence of general-domain text, according to corpus studies (Cameron 2003; Martin 2006; Shutova and Teufel 2010; Steen et al. 2010). This makes metaphor an important subject of linguistic research and makes its accurate processing essential for a range of practical NLP applications. These include, for example, (1) machine translation (MT): Because a large number of metaphorical expressions are culture-specific, they represent a considerable challenge for MT (e.g., the English metaphor “to shoot down someone’s arguments” cannot be literally translated into German as “Argumente abschießen” and metaphor interpretation is required); (2) opinion mining: Metaphorical expressions tend to contain a strong emotional component—for example, compare the metaphorical expression “Government loosened stranglehold on business” and its literal counterpart Government deregulated business (Narayanan 1999); (3) information retrieval (IR): Non-literal language without appropriate disambiguation may lead to false positives in information retrieval (e.g., documents describing “old school gentlemen” should not be returned for the query school [Korkontzelos et al. 2013]); and many others.
Because metaphor interpretation requires complex analogical comparisons and the projection of inference structures across domains, the task of automatic metaphor processing is challenging. For many years, computational work on metaphor evolved around the use of hand-coded knowledge and rules to model metaphorical associations, making the systems hard to scale. Recent years have seen a growing interest in statistical modeling of metaphor (Mason 2004; Gedigian et al. 2006; Shutova 2010; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Heintz et al. 2013; Hovy et al. 2013; Li, Zhu, and Wang 2013; Mohler et al. 2013; Shutova and Sun 2013;
Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with many new techniques opening routes for improving system accuracy and robustness. A wide range of methods have been proposed and investigated by the community, including supervised (Gedigian et al. 2006; Dunn 2013a; Hovy et al. 2013; Mohler et al. 2013; Tsvetkov, Mukomel, and Gershman 2013) and unsupervised (Heintz et al. 2013; Shutova and Sun 2013) learning, distributional approaches (Shutova 2010, 2013; Shutova, Van de Cruys, and Korhonen 2012), lexical resource-based methods (Krishnakumaran and Zhu 2007; Wilks et al. 2013), psycholinguistic features (Turney et al. 2011; Gandy et al. 2013; Neuman et al. 2013; Strzalkowski et al. 2013), and Web search (Veale and Hao 2008; Bollegala and Shutova 2013; Li, Zhu, and Wang 2013). Although individual approaches tackling individual aspects of metaphor have met with success, the insights gained from these experiments are still difficult to integrate into a single computational metaphor modeling landscape, because of the lack of a unified task definition, a shared data set, and well-defined evaluation standards. This hampers our progress as a community in this area. In this paper we take a step towards closing this gap: We review the recent work on computational modeling of metaphor, the tasks addressed, the system features proposed, and the evaluations conducted, and analyze the relevance of different linguistic aspects of metaphor for system performance and applicability, with the aim of identifying the desired properties of metaphor processing systems and a set of requirements for their evaluation.","2 considerations in the design of a metaphor processing system :When designing a metaphor processing system, one faces a number of choices. Some stem from the linguistic and cognitive properties of metaphor, others concern the applicability and usefulness of the system in the wider NLP context. In this section, we analyze individual aspects of metaphor and their relevance to computational modeling, as well as their interplay in the design of a real-world system. Linguistic considerations that inform the design of metaphor processing systems concern primarily the choice of the level (or levels) of analysis. The levels of metaphor analysis include (1) linguistic metaphor (or metaphorical expressions), (2) conceptual metaphor, (3) extended metaphor, and (4) metaphorical inference. Let us consider an example of manifestations of the conceptual metaphor EUROPEAN INTEGRATION as a TRAIN JOURNEY, popular in the early 1990s, at various levels.r Conceptual: EUROPEAN INTEGRATION as a TRAIN JOURNEYr Linguistic: The coupling of the carriages may not be reliably secure, but the
pan-European express is in motion.r Extended metaphor: “There is a fear that the European train will thunder forward, laden with its customary cargo of gravy, towards a destination neither wished for nor understood by electorates. But the train can be stopped.” (Margaret Thatcher, Sunday Times, 20 Sept 1992)r Metaphorical inference: The fact that expensive tracks have to be laid for the train to move forward means that someone has to fund the process of European integration.
2.1.1 Linguistic metaphor. Linguistic metaphor, or metaphorical expressions, concern the surface realization of metaphorical mechanisms, and have been unsurprisingly central to metaphor processing research to date (Birke and Sarkar 2006; Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Gandy et al. 2013; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013). Metaphorical expressions represent the way in which text-processing systems encounter metaphor, and there is little doubt that any real-world metaphor processing system, whatever the approach or the application, ultimately needs to be able to identify and interpret them.
When focusing on linguistic metaphor, one needs to further take into account the level of conventionality of the expressions; how different syntactic constructions are used to convey metaphorical meanings and at what level metaphor annotation needs to be done (word, relation, or sentence level). Thus the following considerations become important for the task and system design:r Level of conventionality of the metaphors accepted: Metaphor is a
productive phenomenon, that is, its novel examples continue to emerge in language. However, a large number of metaphorical expressions become conventionalized over time (e.g., “I cannot grasp his way of thinking”). Although metaphorical in nature, their meanings are deeply entrenched in everyday use, and their comprehension is likened to that of literally used terms (Nunberg 1987). According to Gibbs (1984), metaphorical expressions are spread along a continuum from highly conventional, lexicalized metaphors to entirely novel and creative ones. Gibbs thus suggests that there is no clear demarcation line between literal and metaphorical language, and the distinction between them is rather governed by the level of conventionality of metaphorical expressions. From the usage perspective, metaphoricity may be viewed as a gradient phenomenon rather than a binary one (Dunn 2011), and conventionality becomes an important factor in the design and evaluation of metaphor processing systems. It is not yet clear where on the metaphorical–literal continuum the system should draw the line between what it considers metaphorical and what it considers literal. The answer to this question most likely depends on the NLP application in mind. However, generally speaking, real-world NLP applications are unlikely to be concerned with historical aspects of metaphor, but rather with the identification of figurative language that needs to be interpreted differently from the literal language. We therefore suggest that NLP applications do not necessarily need to address highly conventional and lexicalized metaphors that can be interpreted using standard word sense disambiguation techniques, but rather would benefit from the identification of less conventional and more creative language. Much of the metaphor processing work has focused on conventional metaphor, though in principle capable of identifying novel metaphor as well (Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Gandy et al. 2013; Heintz et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with few approaches modeling only novel metaphor (Desalle, Gaume, and Duvignau 2009) or discriminating between conventional and novel metaphors (Krishnakumaran and Zhu 2007). Other approaches have
looked at literal versus non-literal distinction defined more broadly (Birke and Sarkar 2006; Li and Sporleder 2009, 2010).r Syntactic constructions covered: Metaphors vary with respect to how they are expressed in language grammatically. The grammatical structure of metaphorical expressions is tightly coupled with the inference processes that produce them and are guided by the semantic frames they evoke (Sullivan 2007, 2013). According to Sullivan, conceptual metaphors are realized in language via mapping semantic roles in the source frame onto roles in the target frame. This suggests that both the source and the target domain impose a set of constraints on the roles that can be mapped, and thus on the syntactic options for expressing the metaphorical meaning. There has not yet been a computational approach investigating the interplay of frame semantics and surface structure of metaphorical language. However, the community has addressed modeling linguistic metaphor in a range of syntactic constructions. Verbal or adjectival metaphors are particularly widely embraced by NLP researchers (most approaches), with a few works focusing on copula constructions (Krishnakumaran and Zhu 2007; Gandy et al. 2013; Li, Zhu, and Wang 2013; Neuman et al. 2013) or other nominal metaphors (Heintz et al. 2013; Hovy et al. 2013; Li, Zhu, and Wang 2013; Mohler et al. 2013; Strzalkowski et al. 2013). A number of approaches to metaphor identification have also addressed multiword metaphors (Li and Sporleder 2010; Heintz et al. 2013; Hovy et al. 2013; Mohler et al. 2013). Corpus-linguistic research has shown that verbs and adjectives account for a large proportion of metaphorical expressions observed in the data (Cameron 2003; Shutova and Teufel 2010). However, a recent study (Jamrozik et al. 2013) has also shown that relational words tend to have a higher metaphorical potential. This corresponds to the data on verbal metaphor frequency, and it also suggests that relational nouns are important (e.g., “words are friends of translators”).r Lexical, relation, or sentence level: Finally, one needs to decide if metaphor should be annotated at the word level (i.e., tagging the source domain words alone), relation level (i.e., tagging both source and target words in a particular grammatical relation), or sentence level (i.e., tagging sentences that contain metaphorical language, without explicit annotation of source and target domain words). The word or relation levels provide the most information and have been the focus of the majority of approaches (Gedigian et al. 2006; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Gandy et al. 2013; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Shutova and Sun 2013; Wilks et al. 2013). However, some works annotated metaphor at the sentence level (Krishnakumaran and Zhu 2007; Dunn 2013a; Li, Zhu, and Wang 2013; Mohler et al. 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013).
2.1.2 Conceptual metaphor. Conceptual metaphor represents a cognitive and conceptual mechanism by which humans produce and comprehend metaphorical expressions. Manifestations of conceptual metaphor are ubiquitous in language, communication, and even decision-making (Thibodeau and Boroditsky 2011). Here are a few examples of
common metaphorical mappings: TIME is MONEY (e.g., “That flat tire cost me an hour”), IDEAS are PHYSICAL OBJECTS (e.g., “I cannot grasp his way of thinking”), EMOTIONS ARE VEHICLES (e.g., “she was transported with pleasure”), FEELINGS ARE LIQUIDS (e.g., “all of this stirred an unfathomable excitement in her”), LIFE IS A JOURNEY (e.g., “He arrived at the end of his life with very little emotional baggage”); ARGUMENT IS A WAR (e.g., “He shot down all of my arguments,” “He attacked every weak point in my argument”).
On one hand, few would disagree that the metaphor processing systems capable of understanding and applying conceptual metaphor should be in a better position to accurately handle linguistic metaphors as well. However, a series of questions arise when designing a model of conceptual metaphor. For example, how does one represent conceptual metaphors in the system? What labels does one assign to source and target domains? Is it even possible to name all the conceptual metaphors that humans use and is it necessary to do so? A computational model needs a clear definition of what constitutes the source and target domains (whether they are manually listed or automatically learned) and the consistency and coverage of source and target domain categories would play a crucial role in how well the model can account for real-world data. Previous research on annotation of conceptual metaphor (Shutova and Teufel 2010) has shown that the annotators tend to disagree on the assignment of source and target domain categories. The most variation stems from the level of generality of the selected categories, indicating that while cross-domain mappings are intuitive to humans (i.e., they can be annotated in arbitrary text in principle), labeling source and target domains consistently appears to be a challenging task. What this suggests is that, although the mechanism of conceptual metaphor may be helpful to the system, the question of how to best represent source and target domains within the system remains open. A predefined set of categories, such as those widely discussed in the linguistic literature on CMT, may not be sufficient or even suitable for a computational model. And despite the validity of the main principles of CMT as a linguistic theory, it is not straightforward to port it to computational modeling of metaphor. A more flexible, and potentially datadriven, representation of source and target domain categories is needed for the latter purpose. A data-driven representation would also be better suited to account for the freedom of interpretation that some metaphors allow, since more flexible structures can be dynamically learned from the data. So far, the community has attempted assigning manually created labels to metaphorical mappings (Mason 2004; Baumer, Tomlinson, and Richland 2009), harvesting fine-grained mappings between individual nouns (Li, Zhu, and Wang 2013), using lexical resources to define or expand source and target domain categories (Gandy et al. 2013; Mohler et al. 2013), representing source and target concepts as word clusters (Shutova and Sun 2013) or automatically learned topics (Heintz et al. 2013), and learning metaphorical mappings implicitly within the model without explicit labeling (Shutova, Sun, and Korhonen 2010).
2.1.3 Extended metaphor. Extended metaphor refers to the use of metaphor at the discourse level. It manifests itself in discourse via a sequence of metaphorically used language, yielded by the same conceptual metaphor, whereby a continuous scenario from the source domain is metaphorically projected onto the target. For instance, viewing European integration as a train journey led to numerous metaphorical expressions in political discourse. Each of them mapped certain properties of train journeys to political processes—with countries as loosely connected carriages, peoples of different countries as passengers of their respective carriages, expensive tracks that have to be laid for the train to move forward, and the final destination not very well understood. This metaphor has been dominating the debate, with the European leaders arguing over its details
(Beigman Klebanov and Beigman 2010). One can see this metaphor frequently reappear and evolve over time, helping politicians defend their agendas and, not least, shedding some clarity on an otherwise uncertain future.
Research in linguistics and political science (Musolff 2000; Lakoff 2008; Lakoff and Wehling 2012) suggests that the use of a particular metaphor often guides the speakers’ argumentation strategy throughout a piece of discourse, as well as participants’ behavior in a dialogue. Beigman Klebanov and Beigman (2010) investigated extended metaphor within a game-theoretic framework, demonstrating that maintaining the metaphorical frame in a debate is rationalizable in terms of the gains the participants may get from doing so and their potential losses from swerving away from the metaphor. Their approach reverse-engineers the motivations behind the use of extended metaphor within a formal framework. Beigman Klebanov and Beigman’s work was an important advance, enhancing our understanding of the inner workings of extended metaphor, motivation behind its use, and its effects on social dynamics. However, a computational method for identification and interpretation of extended metaphor in real-world discourse is yet to be proposed. A discourse-level metaphor processing system would need to identify a chain of metaphorical expressions in a text, which indicates a systematic association of the text topic with a particular domain. These chains would then demonstrate how continuous scenarios can be transferred across domains. Recovering this information from the data would allow us to better understand the structure behind metaphorical associations, as well as the inferential process by which knowledge is projected across domains. This system would also find application in social science, where metaphorical framing is widely studied as an indicator of the underlying cultural and moral models (Lakoff and Wehling 2012).
2.1.4 Metaphorical inference. When projecting knowledge from one domain to another, a set of complex inferences take place. Metaphorical inferences are grounded in the source domain and result in the production of surface structures we observe in language as metaphorical expressions. Metaphorical mappings are thus realized via projecting inference structures from the source domain onto the target. For example, when European integration is metaphorically viewed as a train journey, our knowledge of typical events and their consequences from the domain of train journey are projected onto our reasoning about the process of European integration. For instance, if we know that expensive tracks need to be laid before a train can move forward, we can infer that someone also needs to fund the process of European integration. Interestingly, in the presence of a conceptual metaphor such inference can take place even without any linguistic metaphor referring to the tracks being present, but rather on the basis of our common sense knowledge about the functioning of trains.
Besides allowing us to derive new information about the target domain, projecting the inferential structures from the source domain also invokes an emotional response coming from the source domain—for example, an unknown destination or a great expense make one feel uneasy when referred to both literally and metaphorically. Such a transfer of inferential processes and emotional content is believed by some to be one of the central purposes of metaphor (Kovecses 2005; Feldman 2006; Thibodeau and Boroditsky 2011). Psychologists Thibodeau and Boroditsky (2011) investigated how metaphor and metaphorical inference affect decision-making. Their hypothesis was that metaphors in language “instantiate frame-consistent knowledge structures [from the source domain] and invite structurally consistent inferences [about the target domain].” They used two groups of subjects, who were presented with two different texts about crime. In the first text crime was metaphorically portrayed as a VIRUS and
in the second as a BEAST. The two groups were then asked a set of questions on how to tackle crime in the city. It turned out that, whereas the first group tended to opt for preventive measures in tackling crime (e.g., stronger social policies), the second group converged on punishment- or restraint-oriented measures. According to Thibodeau and Boroditsky, their results demonstrate that metaphors have profound influence on how we conceptualize and act with respect to societal issues. Interestingly, the participants would explain their decisions via arguments unrelated to the metaphor, showing that the effect of metaphorical inference in guiding human reasoning is rather covert.
Being able to reproduce these inferences is likely to make automatic metaphor understanding better informed and hence more accurate. It may also provide a mechanism of representing source–target domain mappings, that themselves are generalizations over a set of inferences transferred from one domain to another. And finally, these inferences provide a platform for metaphor interpretation, namely, deriving the meaning of a metaphorical expression and the additional connotations it introduces (that are likely to originate from the source domain). There is a consensus among cognitive linguists that it is metaphorical inference that provides for the very texture behind the use of metaphor (Hobbs 1981; Carbonell 1982; Rohrer 1997; Turner and Fauconnier 2003; Feldman 2006). Although uncovering this texture is certainly one of the main objectives of computational metaphor understanding, it is at the same time a very challenging undertaking. Reproducing metaphorical inferences would require the ability to learn vast amounts of world knowledge from the data, as well as performing complex crossdomain comparisons. And despite being a very promising route, it has not yet been attempted in NLP. Another set of considerations in the design of metaphor processing systems stems from the needs of real-world NLP. The high frequency of metaphorical language in textual data makes accurate metaphor processing desired for a number of NLP applications. Thus, the format of metaphor understanding that metaphor processing systems provide should ideally be informed by the requirements of external NLP and a number of considerations arise in that respect:r Metaphor processing typically involves two tasks: metaphor
identification (distinguishing between literal and metaphorical language in text) and metaphor interpretation (identifying the intended meaning of the metaphor). Both of these provide useful information for language understanding and need to be addressed, either independently or as a single process.r A metaphor processing system should provide a representation of metaphor interpretation that can be easily integrated with other NLP systems. This criterion places constraints on how the metaphor processing task should be defined. The most universally applicable metaphor interpretation would be in the text-to-text form. This means that a metaphor processing system would take raw text as input and provide textual output, in which metaphors are interpreted.r In order to be useful for real-world NLP, the system needs to be capable of processing real-world data, and thus operate on unrestricted, continuous text. Rather than only dealing with individual carefully selected clear-cut
examples, the system should be fully implemented and tested on free, naturally occurring text.r To enable wide applicability, the system needs to be open-domain—that is, operate in all domains, genres, and topics. Therefore, ideally it should not rely on any domain-specific information or focus on individual types of instances (e.g., a limited set of hand-chosen source-target domain mappings).r To be easily adaptable to new domains, the system should not rely on task-specific hand-coded knowledge. This means it needs to be either data-driven and be able to automatically acquire the knowledge it needs from text corpora, or rely only on large-scale, general-domain lexical resources (that are already in existence and do not need to be created in a costly manner). However, it would be an advantage if no such resource is required and the system can dynamically induce meanings in context.r To be robust, the system needs to be able to deal with metaphors represented by all word classes and syntactic constructions. Many existing models are designed with specific kinds of metaphorical expressions in mind, for instance nominal metaphors in copula constructions or verbal metaphors in verb–object relations. To be applicable to support real-world NLP applications, these models need to be extended beyond those specific word classes and syntactic constructions, and be able to process any kind of metaphorical language.
When designing a metaphor processing task, methodology, and evaluation strategy, one thus needs to keep these criteria in mind. Although modeling all of the phenomena described in this section within a single system is by no means a requirement, it is critically important to be aware of all the guises that metaphor may take, both conceptually and empirically.","3 metaphor annotation and resources : Metaphor annotation studies have typically focused on one (or both) of the following tasks: (1) identification of metaphorical senses in text (i.e., distinguishing between literal and non-literal meanings), and (2) assignment of the corresponding source– target domain mappings. The majority of corpus-linguistic studies were concerned with metaphorical expressions and mappings within a limited domain—for example, WAR, BUSINESS, FOOD, or PLANT metaphors (Santa Ana 1999; Izwaini 2003; Koller 2004; Skorczynska Sznajder and Pique-Angordans 2004; Chung, Ahrens, and Huang 2005; Hardie et al. 2007; Gong, Ahrens, and Huang 2008; Lu and Ahrens 2008; Low et al. 2010), in a particular genre or type of discourse (Charteris-Black 2000; Cameron 2003; Izwaini 2003; Koller 2004; Skorczynska Sznajder and Pique-Angordans 2004; Martin 2006; Hardie et al. 2007; Lu and Ahrens 2008; Beigman Klebanov and Flor 2013), or in individual examples in isolation from wider context (Wikberg 2006; LönnekerRodman 2008). In addition, these approaches often focused on a small predefined set of source and target domains. Another vein of corpus-based research concerned crosslinguistic differences in the use of metaphor, also in a specific domain—for example,
financial discourse (Charteris-Black and Ennis 2001), metaphors describing FEELINGS (Stefanowitsch 2004; Diaz-Vera and Caballero 2013), or metaphorical expressions referring to body parts (Deignan and Potter 2004). Three recent studies are notable in that they moved away from investigating particular domains to a more general study of how metaphor behaves in unrestricted continuous text. Wallington et al. (2003), Shutova and Teufel (2010), and Steen et al. (2010) conducted consecutive metaphor annotation in open-domain texts.
Wallington et al. (2003) used two teams of annotators and compared externally prescribed definitions of metaphor with intuitive internal ones. Team A was asked to annotate “interesting stretches,” whereby a phrase was considered interesting if (1) its significance in the document was non-physical, (2) it could have a physical significance in another context with a similar syntactic frame, and (3) this physical significance was related to the abstract one. Team B had to annotate phrases according to their own intuitive definition of metaphor. Apart from metaphorical expressions, the respective source–target domain mappings were also to be annotated. For this latter task, the annotators were given a set of mappings from the Master Metaphor List and were asked to assign the most suitable ones. However, the authors do not report the level of interannotator agreement, nor the coverage of the mappings in the Master Metaphor List on their data. The fact that the method is limited to a set of mappings exemplified in the Master Metaphor List suggests that it may not scale well to real-world data, because the predefined inventory of mappings is unlikely to be sufficient to cover the majority of metaphorical expressions in arbitrary text.
Steen and his colleagues (Pragglejaz Group 2007; Steen et al. 2010) proposed a metaphor identification procedure (MIP). In the framework of this procedure, the sense of every word in the text is considered as a potential metaphor. Every word is then tagged as literal or metaphorical, based on whether is has a “more basic, contemporary meaning” in other contexts than the current one. The summary of their annotation procedure is presented is Figure 1. In a sense, such annotation can be viewed as a form of word sense disambiguation with an emphasis on metaphoricity. Steen and colleagues ran a reliability study involving near-native speaker annotators (strongly relying on dictionary definitions) and report an interannotator agreement of 0.85 in terms of Fleiss’ kappa. MIP laid the basis for the creation of the VU Amsterdam Metaphor Corpus1 (Steen et al. 2010). This corpus is a subset of BNC Baby2 annotated for linguistic metaphor. Its size is 200,000 words and it comprises four genres: news text, academic text, fiction, and conversations. The corpus has already found application in computational metaphor processing research (Dunn 2013b; Niculae and Yaneva 2013), as well as inspiring metaphor annotation efforts in other languages (Badryzlova et al. 2013).
The study of Shutova and Teufel (2010) was concerned with annotation of both metaphorical expressions and metaphorical mappings in continuous text. Their annotation procedure is based on MIP, modifying and extending it to the identification of conceptual metaphors along with the linguistic ones. Following MIP, the annotators were asked to identify the more basic sense of the word, and then label the context in which the word occurs in the basic sense as the source domain, and the current context as the target. They were provided with a list of suggested common source
1 http://www.ota.ox.ac.uk/headers/2541.xml. 2 BNC Baby is a 4-million-word subset of the British National Corpus (BNC) (Burnard 2007), comprising
four different genres: academic, fiction, newspaper, and conversation. For more information, see http://www.natcorp.ox.ac.uk/corpus/babyinfo.html.
and target domains, but were also allowed to introduce domains of their own to match their intuitions. Shutova and Teufel’s corpus is a subset of the BNC sampling various genres: fiction, newspaper/journal articles, essays on politics, international relations and sociology, and radio broadcast (transcribed speech). The size of the corpus is 13,642 words, containing 241 metaphorical expressions in total. Table 1 shows the breakdown of metaphors by type, as well as their variation across genres. Shutova and colleagues used the corpus data as a testbed in a number of computational experiments (Shutova 2010; Shutova, Van de Cruys, and Korhonen 2012; Shutova 2013; Shutova, Teufel, and Korhonen 2013). Lakoff and colleagues organized their ideas in a resource called the Master Metaphor List (MML) (Lakoff, Espenson, and Schwartz 1991). The list is a collection of source– target domain mappings (mainly those related to mind, feelings, and emotions) with corresponding examples of language use. The mappings in the list are organized in an ontology—for example, the metaphor PURPOSES ARE DESTINATIONS is a special case of a more general metaphor STATES ARE LOCATIONS. The resource has been criticized for the lack of clear structuring principles of the mapping ontology (Lönneker-Rodman 2008). However, to date MML is the most comprehensive resource for conceptual metaphor in the linguistic literature, and the examples from the list have been used by computational approaches (Mason 2004; Krishnakumaran and Zhu 2007; Li, Zhu, and Wang 2013), both for development and evaluation purposes. The MML also inspired the creation of other resources, including resources in multiple languages that could facilitate cross-linguistic research on metaphor. One such example is the Hamburg Metaphor Database (Lönneker 2004; Reining and Lönneker-Rodman 2007), which contains examples of metaphorical expressions in German and French. The expressions are mapped to senses from EuroWordNet3 and annotated with source–target domain mappings taken from the MML.","4 metaphor identification systems :Early approaches to metaphor relied on information in handcrafted knowledge bases, followed by metaphor identification in and with the help of lexical resources. Recent years have witnessed a growing interest in statistical and machine learning approaches to metaphor identification. As the field of computational semantics—in particular, robust parsing and lexical acquisition techniques—have progressed to the point where it is possible to accurately acquire lexical, domain, and relational information from corpora, this opened many new avenues for large-scale statistical metaphor identification. The vast majority of systems identify metaphor at the linguistic level (Birke and Sarkar 2006; Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with very few focusing on the conceptual level (Mason 2004; Baumer, Tomlinson, and Richland 2009) or identifying both (Gandy et al. 2013; Li, Zhu, and Wang 2013; Shutova and Sun 2013). This section will first present computational approaches to linguistic metaphor identification, then move on to conceptual metaphor. 4.1.1 Approaches Using Hand-Coded Knowledge and Lexical Resources. One of the first approaches to identify and interpret metaphorical expressions in text was proposed by Fass (1991) in his met* system. This system relies on the hypothesis that metaphors often represent a violation of selectional preferences in a given context (Wilks 1975,
3 EuroWordNet is a multilingual database with wordnets for several European languages (Dutch, Italian, Spanish, German, French, Czech, and Estonian). The wordnets are structured in the same way as the Princeton WordNet for English. http://www.illc.uva.nl/EuroWordNet/.
1978). Selectional preferences are the semantic constraints that a predicate places onto its arguments. Consider the following metaphorical expression.
(5) My car drinks gasoline. (Wilks 1978)
The verb drink normally requires a grammatical subject of type ANIMATE and a grammatical object of type LIQUID. Therefore, drink taking a car as a subject in Example (5) is an anomaly, which, according to Wilks, indicates a metaphorical use of drink. met* detects non-literalness via selectional preference violation, utilizing handcrafted descriptions of selectional preferences. In case of a violation, the respective phrase is first tested for being metonymic, using hand-coded patterns (e.g., CONTAINER-FOR-CONTENT). If this fails, the system searches the knowledge base for a relevant analogy in order to discriminate metaphorical relations from the anomalous ones. For example, the sentence “My car drinks gasoline” would be represented in this framework as (car,drink,gasoline), which does not satisfy the preference (animal,drink,liquid), as car is not a hyponym of animal. met* then searches its knowledge base for a triple containing a hypernym of both the actual argument and the desired argument and finds (thing,use,energy source), which represents the metaphorical interpretation. Fass (1991) presented the approach itself, but reported no evaluation results.
More recently, Wilks et al. (2013) revisited this idea, acquiring selectional preferences from lexical resources, namely VerbNet and WordNet. They focused on conventionalized metaphors included in lexical resources and proposed a technique for their automatic identification. They see this work as complementary to the approaches that perform data-driven learning of selectional preferences. The latter, according to the authors, is likely to miss conventional metaphors because of their widespread presence in the data. Wilks and colleagues expect that selectional preferences acquired from term definitions in lexical resources would circumvent this issue and enable them to efficiently detect highly conventionalized metaphors. The main hypothesis behind their approach is that if the first (main) WordNet sense of a word does not satisfy the preferences of its context in a given sentence, but has a lower (less frequent) sense in WordNet that satisfies the preference, then that use of the word and that WordNet sense are likely to be metaphorical. For instance, in the example “Mary married a brick”, the first sense of brick is ‘a physical object,’ thus violating the preference of marry that selects for people, but the second sense of brick as ‘a reliable person’ satisfies this preference. To implement this approach, Wilks and colleagues acquire typical preferences of concepts (i.e., word senses) from WordNet glosses. They use a semantic parser (Allen, Swift, and de Beaumont 2008) to identify the nominal arguments of the verbs in glosses and their semantic roles and then abstract to their higher-level hypernyms in WordNet, which define the preferences. They compared the performance of their system to a baseline using hand-coded verb preferences in VerbNet. The evaluation was carried out on a set of 122 sentences from the domain of Governance, manually annotated for metaphoricity and selected so that the data set contains 50% metaphorical instances and 50% literal ones. They report an F-score of 0.49 for the VerbNet-based system and 0.67 for the WordNet-based one, the latter showing higher recall and the former higher precision. The approach of Wilks et al. rests on the assumption that WordNet sense ranking corresponds somewhat to the literal-to-metaphorical scale, as well as the assumption that there is only one literal sense for the given word. Although this may be true for the majority of senses, it is relatively easy to find counter-examples. For instance, the first WordNet sense of the verb erase is metaphorical, defined as “remove from memory or existence, e.g., The Turks erased the Armenians in 1915,” with the literal sense ranked
second. The reliance on WordNet sense numbering is thus a limitation of the presented approach. Another issue (that the authors point out themselves) is that this approach is likely to detect metonymic uses along with metaphor, and a method to discriminate between the two is still needed.
The method of Krishnakumaran and Zhu (2007) uses hyponymy relation in WordNet and word bigram counts to annotate metaphor at the sentence level. Given an IS-A metaphor (e.g., The world is a stage4) they first verify if the two nouns involved are in hyponymy relation in WordNet, and if this is not the case then the sentence is tagged as containing a metaphor. Along with this they consider expressions containing a verb or an adjective used metaphorically (e.g., “He planted good ideas in their minds” or “He has a fertile imagination”). In such cases, they calculate bigram probabilities of verb– noun and adjective–noun pairs (including the hyponyms/hypernyms of the noun in question). If the combination is not observed in the data with sufficient frequency, the system tags the sentence containing it as metaphorical. This idea follows the intuition of Wilks. However, by using bigram counts over verb–noun pairs Krishnakumaran and Zhu (2007) lose a great deal of information compared with a system extracting selectional preferences for specific grammatical relations from parsed text. The authors evaluated their system on a set of example sentences compiled from the MML, whereby highly conventionalized metaphors (or dead metaphors) are taken to be negative examples, reporting an accuracy of 0.58. Thus Krishnakumaran and Zhu do not deal with literal examples as such: Essentially, the distinction they are making is between the senses included in WordNet, even if they are conventional metaphors, and those not included in WordNet.
4.1.2 Statistical Learning for Metaphor Identification. The first statistical approach to metaphor is the TroFi system (Trope Finder) of Birke and Sarkar (2006). Their method is based on sentence clustering, originating from a similarity-based word sense disambiguation method developed by Karov and Edelman (1998). The method uses a set of seed sentences, where the senses are annotated, computes similarity between the sentence containing the word to be disambiguated and all of the seed sentences, and selects the sense corresponding to the annotation in the most similar seed sentences. Birke and Sarkar adapt this algorithm to perform two-way classification: literal versus non-literal, and they do not clearly define the kinds of tropes they aim to discover. They evaluated their system on a set of 25 verbs (such as absorb, die, touch, knock, strike, pour, etc.), for each of which they extracted a set of sentences containing its literal and figurative uses, 1,298 in total, from the Wall Street Journal corpus. An example for the verb pour in their data set is shown in Figure 2. Two annotators annotated the sentences for literalness, achieving an agreement of κ = 0.77. The authors report a system performance of 53.8% in terms of F-score on this data set.
The metaphor identification system of Shutova, Sun, and Korhonen (2010) also uses clustering techniques, but performs word clustering to discover verb–subject and verb– object metaphors in unrestricted text. It starts from a small seed set of metaphorical expressions, learns the analogies involved in their production, and extends the set of analogies by means of verb and noun clustering. The method is based on the hypothesis of “clustering by association”—namely, that in the course of distributional noun clustering, abstract concepts tend to cluster together if they are associated with the same source domain, whereas concrete concepts cluster by meaning similarity. For
4 William Shakespeare.
instance, democracy and marriage get clustered together, because both are associated with mechanisms, and as such appear with the mechanism terminology in the corpus. This allows the system to discover new, previously unseen conceptual and linguistic metaphors—for example, having seen the seed metaphor “mend marriage” it infers that “the functioning of democracy” is also used metaphorically. This is how the system expands from the seed set to new concepts. Shutova, Sun and Korhonen used a spectral clustering algorithm with lexico-syntactic features to cluster verbs and nouns. They applied their system to continuous text (the whole BNC) and evaluated its performance on a random sample of the extracted metaphors against human judgments. They report a precision of 0.79 with an inter-judge agreement of k = 0.63 among five annotators. Their data-driven system favorably compares to a WordNet-based baseline, where synsets are used in place of automatically derived clusters. Shutova and colleagues have shown that the clustering-based solution has a significantly wider coverage, capturing new metaphors rather than the synonymous ones, as well as yielding a 35% increase in precision. However, Shutova, Sun and Korhonen did not evaluate the recall of their system, which is likely to be dependent on the size of the seed set and a relatively large and representative seed set is needed to achieve full coverage.
Turney et al. (2011) classify verbs and adjectives as literal or metaphorical based on their level of concreteness or abstractness in relation to the noun they appear with. They learn concreteness rankings for words automatically (starting from a set of examples) and then search for expressions where a concrete adjective or verb is used with an abstract noun (e.g., “dark humor” is tagged as a metaphor and “dark hair” is not). They used the data set of Birke and Sarkar (2006) for evaluation of verb metaphors and attain an F-score of 0.68, which favorably compares to that of Birke and Sarkar. For adjectives, they have created their own data set of selected individual adjective–noun pairs for five adjectives: dark, deep, hard, sweet, and warm; 100 phrases in total. These were then manually annotated for metaphoricity. As compared to these annotations, the accuracy of adjective classification is 0.79. However, the adjective data set was constructed with the concreteness feature in mind, and therefore the results reported for verb metaphors are likely to be more objective.
Neuman et al. (2013) proposed an extension to the method of Turney et al. (2011) by incorporating the concept of selectional preferences into the concreteness-based model of metaphor. Their goal was to improve the performance of Turney’s algorithm by covering metaphors formed of concrete concepts only (e.g., “broken heart”) by detecting selectional preference violations. The authors address three types of metaphor introduced by Krishnakumaran and Zhu (2007) and they claim to have expanded on Turney’s work by carrying out a more comprehensive evaluation of the abstractness–concreteness algorithm. However, the evaluation was done on only five
target concepts: governance, government, God, mother, and father. Sentences describing these concepts have been extracted from the Reuters (Lewis et al. 2004) and New York Times (Sandhaus 2008) corpora and annotated for metaphoricity. The authors measured the average precision of their system on this data set at 0.72 and the average recall at 0.80. The improvement over Turney’s evaluation set-up was the annotation of complete sentences rather than isolated phrases. However, it should be noted that the system was evaluated on selected examples rather than continuous text.
Heintz et al. (2013) applied Latent Dirichlet Allocation (LDA) topic modeling (Blei, Ng, and Jordan 2003) to the problem of metaphor identification in experiments with English and Spanish. Their goal was to create a minimally supervised metaphor processing system that can be applied to low-resource languages. The hypothesis behind their system is that a sentence that contains both source and target domain vocabulary contains a metaphor. The authors focused on the target domain of governance and have manually compiled a set of source concepts with which governance can be associated. They use LDA topics as proxies for source and target concepts, and if vocabulary from both source and target topics is present in a sentence, this sentence is tagged as containing a metaphor. The topics are learned from Wikipedia and then aligned to source and target concepts using sets of human-created seed words. When the metaphorical sentences are retrieved, the source topics that are common in the document are excluded, thus ensuring that the source vocabulary is transferred from a new domain. Although this allows the authors to filter out some literal uses, this may also lead to discarding cases of extended metaphor. The authors collected the data for their experiments from news Web sites and governance-related blogs in English and Spanish. They ran their system on this data, and output a ranked set of metaphorical examples. They carried out two types of evaluation: (1) top five examples for each conceptual metaphor judged by two annotators, reporting an F-score of 0.59 for English (κ = 0.48); and (2) 250 top-ranked examples in system output annotated for metaphoricity using Amazon Mechanical Turk, yielding a mean metaphoricity of 0.41 (standard deviation = 0.33) in English and 0.33 (standard deviation = 0.23) in Spanish. One of the assumptions behind Heintz et al.’s method is that the same source–target domain mappings manifest themselves across languages. Although this is likely to be true for primary metaphors (Grady 1997), as the authors point out themselves, this assumption may not extend to a broader spectrum of metaphors, and thus may lead to limited coverage in some languages, as well as false positives. Another issue that comes to mind concerns the learning of the topics themselves: Because a large number of metaphors are used conventionally within a particular topic (e.g. “cut taxes”), in principle such an approach would learn them as part of the target domain topics and may thus fail to recognize them as source domain terms. However, the authors do not comment on how often this was observed in their data.
The method of Strzalkowski et al. (2013) also relies on modeling the topical structure of text, although using different techniques from Heintz et al. (2013). If Heintz et al. used LDA-acquired topics as approximations of concepts, Strzalkowski and colleagues identify topical chains (Broadwell et al. 2013) by looking for sequences of concepts in text. They also experiment within a limited domain, the target domain of governance. Their method first identifies sentences containing target domain vocabulary and extracts the surrounding five-sentence passage. They then identify topical chains in that passage, by linking the occurrences of nouns and verbs, including repetition, lexical variants, pronominal references, and WordNet synonyms and hyponyms. By virtue of this linking, the authors claim to “uncover the topical structure [of the text] that holds the narrative together.” Their main hypothesis is that metaphorically used terms
typically occur outside the core topical structure of the text, because they represent vocabulary imported from a different domain. For all the words that are found outside the topical chains, Strzalkowski et al. compute imageability scores and retain the highestscoring ones as candidate metaphors, if they are in a syntactic relation with any of the target domain terms. The authors then extract common contexts in which the candidates are used in text corpora and cluster these contexts in order to identify potential source domains, the so-called “proto-sources.” Strzalkowski et al. (2013) evaluated the performance of their method on four languages: English, Spanish, Russian, and Farsi. The evaluation was carried out against human judgments of system output that were obtained using Amazon Mechanical Turk. The authors report a metaphor identification accuracy of 71% in English, 80% in Spanish, 69% in Russian, and 78% in Farsi. According to the paper, “hundreds” of instances were annotated in each language, although the exact number of instances is not reported. While the system performance is high, it should be noted that the experiments were carried out within a limited domain, and it is possible that the approach is not equally applicable to all domains. Because of its high reliance on imageability scores, it is likely to be able to delineate metaphorical language reasonably well for the abstract target domains, but less so for the concrete target domains. In the latter case, the target domain words may also exhibit high imageability, and the system would then rely solely on topic chain extraction to differentiate between literal and metaphorical language. The performance of the generalized system is thus dependent on the accuracy of topic chain extraction, which has not been evaluated independently. In addition, the current method ignores low-imageability metaphors, which abound even within the studied domain (e.g., “invent a new form of governance”). Despite the lack of generality, Strzalkowski et al.’s work, however, makes important contributions in that it addresses (though indirectly) the behavior of metaphor in discourse, and their framework can be viewed as a step towards modeling extended metaphor.
Many other statistical methods treated metaphor identification as a classification problem. Such methods are described in the following section.
4.1.3 Metaphor Identification as a Classification Problem. Gedigian et al. (2006) presented a method that discriminates between literal and metaphorical language, using a maximum entropy classifier. They obtained their training and test data by extracting the lexical items whose frames are related to MOTION and CURE from FrameNet (Fillmore, Johnson, and Petruck 2003). They then searched the PropBank Wall Street Journal corpus (Kingsbury and Palmer 2002) for sentences containing such lexical items and annotated them with respect to metaphoricity. They used PropBank annotation (arguments and their semantic types) as features to train the classifier and report an accuracy of 95.12%. This result is, however, only a little higher than the performance of the naive baseline assigning majority class to all instances (92.90%). These numbers can be explained by the fact that 92.00% of the verbs of MOTION and CURE in the Wall Street Journal corpus are used metaphorically, thus making the data set unbalanced with respect to the target categories and the task easier.
The system of Li and Sporleder (2009, 2010) detects idioms by measuring semantic similarity within and between the literal and non-literal parts of an utterance. The nonliteral language considered by their model includes metaphors, as well as other types of figurative language. Their main assumption is that figurative uses break cohesion in the sentence, which is defined by a similarity measure. This idea also goes back to Wilk’s selectional preference violation approach to metaphor; however, combinations of word usages with larger sentential context are considered to determine the mismatch
(or violation). Li and Sporleder used Normalized Google Distance (Cilibrasi and Vitanyi 2007) as a similarity measure and a combination of classifiers (support vector machines [SVM] and Gaussian mixture models [GMM]) using similarity (or cohesion) information as features to learn idiomaticity scores. They evaluated their system on a data set of 17 idioms and their literal and non-literal contexts. For each expression, its occurrences were extracted from the Gigaword corpus along with five paragraphs of context. These examples were then annotated for literalness with an inter-annotator agreement of κ = 0.7. There were 3,964 examples in total, with approximately 80% of them being non-literal. They evaluated the method using 10-fold cross-validation and report an F-score of 0.75. However, they did not evaluate their system on metaphorical language independently.
Dunn (2013a, 2013b) presented an ontology-based domain interaction system MIMIL (Measuring and Identifying Metaphor in Language), which identifies metaphorical expressions at the utterance level. Dunn’s system first maps the lexical items in the given utterance to concepts from SUMO ontology (Niles and Pease 2001, 2003), assuming that each lexical item is used in its default sense (i.e., no sense disambiguation is performed). The system then extracts the properties of concepts from the ontology, such as their domain type (ABSTRACT, PHYSICAL, SOCIAL, MENTAL) and event status (PROCESS, STATE, OBJECT). Those properties are then combined into feature-vector representations of the utterances. Dunn then applied a logistic regression classifier implemented in Weka (Witten and Frank 2005), using these features to perform metaphor identification. The work of Dunn (2013a, 2013b) is notable as he conducted evaluation of four types of approaches and compared their performance on the same task (identification of metaphorical expressions in continuous text) and on the same data (Corpus of Contemporary American English [CoCA] [Davies 2009] and VU Amsterdam Metaphor Corpus [Steen et al. 2010]). The evaluated approaches included the semantic similarity measurement method of Li and Sporleder (2009, 2010); the concreteness-based method of Turney et al. (2011); the clustering-based method of Shutova, Sun, and Korhonen (2010) modeling source–target domain mappings; and his own domain interaction method. Dunn re-implemented the four approaches as closely as possible to the original systems, although with some adjustments. A number of Dunn’s adjustments were operational (e.g., using logistic regression instead of SVM for the implementation of the similaritybased method of Li and Sporleder (2009); using a k-means clustering approach instead of spectral clustering for the method of Shutova, Sun, and Korhonen (2010); and using a different semantic relatedness measure for the method of Li and Sporleder (2009)). However, some adjustments were conceptual, for instance, using bag-of-words based semantic relatedness instead of dependency-based distributional similarity in the reimplementation of the clustering system. Admitting that these adjustments may have impacted the results and, as such, may not be an accurate reflection of the performance of the original algorithms in full, Dunn’s comparison of individual system features that were re-implemented nonetheless sheds light on the importance of particular properties of concepts for metaphor identification. In his first study (Dunn 2013a), he evaluated the systems on the CoCA data, where the sentences were annotated as metaphorical, literal, or humorous (however, neither the size of the data set nor the annotation procedure are described in Dunn’s article). On this data set, the clustering-based system and the domain-interaction method significantly outperformed the other two systems, as shown in Figure 3. Dunn explains such discrepancy by the fact that the former systems are both theory-based and aim to model the underlying mechanisms of metaphor, while the similarity-based and abstractness-based systems model its surface realizations. In his second study, conducted on the VU Amsterdam Metaphor corpus, Dunn (2013b)
Shutova Design and Evaluation of Metaphor Processing Systems
15
testing; the evaluations were performed using cross-validation (100 folds). The F-measures reported here are for metaphor classification only (i.e., precision for non-metaphor is not directly considered because this inflates the performance of the systems). This is done because some of the systems greatly over-identify literal utterances; however, because literal utterances dominate the evaluation data set, the over-identification of literal utterances would disproportionately raise
the average F-measure for all classes in these systems. The feature vectors and other material
used in the evaluation can be found at http://www.jdunn.name.
humor, and literal utterances, the similarity and abstractness systems performed very poorly,
essentially identifying no metaphors. The joint system performed worse than the domain
interaction system, showing that the abstractness and similarity features reduce performance. As
shown in later tests, the measurements of abstractness and semantic similarity, both at the word-
level, simply do not distinguish between metaphor and non-metaphor in a realistic data set. The
reports different results, however. He evaluated the systems on two versions of the data set, one where named entities have been recognized during pre-processing and one without named-entity recognition. The results are shown in Tables 2 and 3, respectively. Here, the domain-interaction and abstractness-based methods are leading, with the clustering-based method coming third. The difference in the results may be explained by the properties of the VU Amsterdam corpus. The corpus was compiled with an interest in historic aspects of metaphor, and, therefore, highly conventional and lexicalized metaphors account for a large proportion of the data. What this suggests is that the domain-interaction and abstractness-based approaches are perhaps better-suited for processing lexicalized metaphors, whereas the clustering and similarity-based systems may fail to identify those due to their high frequency and the near-literal behavior in the data. In contrast, the domain-interaction system, which is knowledge-based, and the abstractness system, which relies on a non-changing property of concepts (i.e., concreteness), thus appear to be well-suited for handling lexicalized metaphors.
Tsvetkov, Mukomel, and Gershman (2013) presented a supervised learning approach that makes use of coarse semantic features. They experimented with metaphor
identification in English and Russian, first training a classifier on English data only, and then projecting the trained model to Russian using a dictionary. They abstracted from the words in the English data to their higher level features, such as concreteness, animateness, named entity labels, and coarse-grained WordNet categories (corresponding to WN lexicographer files,5 e.g., noun.artifact, noun.body, verb.motion, verb.cognition). They focused on subject–verb–object constructions and annotated metaphor at the sentence level. The authors used a logistic regression classifier and the combination of coarse semantic features for this purpose. They evaluated their model on the TroFi data set (Birke and Sarkar 2006) for English and a self-constructed data set of 140 sentences for Russian, attaining the F-scores of 0.78 and 0.76, respectively. Tsvetkov et al. (2014) extended this experiment to identify adjective–noun metaphors using similar features, as well as porting the model to two further languages (Spanish and Farsi), achieving F-scores in the range of 0.72 to 0.85. The results are encouraging and show that porting coarse-grained semantic knowledge across languages is feasible. However, it should be noted that the generalization to coarse semantic features inevitably only captures shallow behavior of metaphorical expressions in the data and bypasses conceptual information. In reality, as confirmed by corpus-linguistic studies (Charteris-Black and Ennis 2001; Kovecses 2005; Diaz-Vera and Caballero 2013), there is considerable variation in metaphorical language across cultures, which makes training only on one language and simply translating the model less suitable for modeling conceptual structure behind metaphor, which is one of the limitations of this approach. However, the experiments of Tsvetkov and colleagues suggest that coarse semantic features could be a useful component of a more complex system.
The approach of Mohler et al. (2013) relied on the concept of semantic signature of a text. The authors defined semantic signatures as a set of highly related and interlinked WordNet senses. They induced domain-sensitive semantic signatures of texts and then trained a set of classifiers to detect metaphoricity within a text by comparing its semantic signature to a set of known metaphors. The main intuition behind this approach is that the texts whose semantic signature closely matches the signature of a known metaphor is likely to represent an instance of the same conceptual metaphor. Mohler and colleagues conducted their experiments within a limited domain (the target domain of governance) and manually constructed an index of known metaphors for this domain. They then automatically created the target domain signature and a signature for each source domain among the known metaphors in the index. This was done by means of semantic expansion of domain terms using WordNet, Wikipedia links, and corpus cooccurrence statistics. Given an input text their method first identified all target domain terms using the target domain signature, then disambiguated the remaining terms using sense clustering and classified them according to their proximity to the source domains listed in the index. For the latter purpose, the authors experimented with a set of classifiers, including a maximum entropy classifier, an unpruned decision tree classifier, support vector machines, a random forest classifier, as well as the combination thereof. They evaluated their system on a balanced data set containing 241 metaphorical and 241 literal examples, and obtained the highest result of F-score of 0.70 using the decision tree classifier.
Hovy et al. (2013) used the idea of selectional preference violation as the indicator of metaphor, taking it to the next level. They trained an SVM classifier (Cortes and Vapnik 1995) with tree kernels (Moschitti, Pighin, and Basili 2006) to capture compositional
5 http://wordnet.princeton.edu/man/lexnames.5WN.html.
Figure 1: Examples of a metaphor seed, the matching Brown sentences, and their annotations
thus probably not very helpful. However, we would like to capture semantic aspects of the word and represent it in an expressive way. We use the existing vector representation SENNA (Collobert et al., 2011) which is derived from contextual similarity. In it, semantically similar words are represented by similar vectors, without us having to define similarity or looking at the word itself. In initial tests, these vectors performed better than binary vectors straightforwardly derived from features of the word in context.
2.3 Constructing Trees
a) b) c)like
I people
the sweet in
Boston
NNS
DT JJ IN
n.group
O adj.all O
NNP n.location
VB
PRP
v.emotion
O
Figure 3: Graphic demonstration of our approach. a) dependency tree over words, with node of interest labeled. b) as POS representation. c) as supersense representation
The intuition behind our approach is that metaphorical use differs from literal use in certain syntactic relations. For example, the only difference between the two sentences “I like the sweet people in Boston” and “I like the sweet pies in Boston” is the head of “sweet”. Our assumption is that—given enough examples—certain patterns emerge (e.g., that “sweet” in combination with food nouns is literal, but is metaphorical if governed by a noun denoting people).
We assume that these patterns occur on different levels, and mainly between syntactically related words. We thus need a data representation to capture these patterns. We borrow its structure from
dependency trees, and the different levels from various annotations. We parse the input sentence with the FANSE parser (Tratz and Hovy, 2011)6. It provides the dependency structure, POS tags, and other information.
To construct the different tree representations, we replace each node in the tree with its word, lemma, POS tag, dependency label, or supersense (the WordNet lexicographer name of the word’s first sense (Fellbaum, 1998)), and mark the word in question with a special node. See Figure 3 for a graphical representation. These trees are used in addition to the vectors.
This approach is similar to the ones described in (Moschitti et al., 2006; Qian et al., 2008; Hovy et al., 2012).
2.4 Classification Models A tree kernel is simply a similarity matrix over tree instances. It computes the similarity between two trees T1, T2 based on the number of shared subtrees.
We want to make use of the information encoded in the different tree representations during classification, i.e., a forest of tree kernels. We thus combine the contributions of the individual tree representation kernels via addition. We use kernels over the lemma, POS tag, and supersense tree representations, the combination which performed best on the dev set in terms of accuracy.
We use the SVMlight TK implementation by Moschitti (2006).7 We left most parameters set to default values, but tuned the weight of the contribution of the trees and the cost factor on the dev set. We set the multiplicative constant for the trees to 2.0, and the cost factor for errors on positive examples to 1.7.
6http://www.isi.edu/publications/ licensed-sw/fanseparser/index.html
7http://disi.unitn.it/moschitti/ Tree-Kernel.htm
54
properties of metaphorical language. Their hypothesis is that unusual semantic compositions in the data may be indicative of the use of metaphor. They trained the model on labeled examples of literal and metaphorical uses of 329 words (3,872 sentences in total), with an expectation to learn the differences in their compositional behavior in the given lexico-syntactic contexts. The choice of dependency-tree kernels helped to capture such compositional properties, according to the authors. The authors constructed their data set by extracting sentences from the Brown corpus (Francis and Kucera 1979) that containe the words of interest, and annotating them f r metaphoricity using Amazon Mechanical Turk. Example entries for the adjective br ght is shown i Figure 4. Eighty percent of the data were used for training purposes, 10% for parameter tuning, and 10% for the evaluation. The learning was carried out using word vectors, as w ll as lexical, part-of-speech tags, and WordNet supersense representations of sente ce trees as features, as shown in Figure 5. The authors reported encouraging results (F-score = 0.75), which is an indication of the importance of syntactic information and compositionality in metaphor identification. The first method for automatic identification of conceptual metaphor was the CorMet system of Mason (2004). CorMet induced metaphorical mappings by identifying systematic variations in domain-specific selectional preferences, which were learned in a data-driven way. For example, the verb pour has a strong selectional preference for
B low he could see the bright torches lighting the riverbank . no
Her bright eyes were twinkling . yes Washed , they came out surprisingly clear and bright . no
Figure 1: Exampl s of a metaphor seed, the matching Brown sentences, and their annotations
thus probably not very helpf l. However, we woul like to capture semantic aspects of the word and epresent it in an expressive way. We use t xi t-
ing vector representation SENNA (Collobert et al., 2011) which s derived from contextual simila ity. In it, semantically similar words are represe ted by similar vectors, without us having to define similarity or looking at the word itself. In initial tests, these vectors performed better than binary vectors straightforwardly derived from features f the word in context.
2.3 Constructing Trees
a) b) c)like
I people
the sweet in
Boston
NNS
DT JJ IN
n.group
O adj.all O
NNP n.location
VB
PRP
v.emotion
O
Figure 3: Graphic demonstration of our approach. a) de-
pendency tree over words, with node of interest labeled.
b) as POS representation. c) as supersense representation
The intuition behind our approach is that metaphorical use differs from literal use in certain syntactic relations. For example, the only difference between the two sentences “I like the sweet people in Boston” and “I like the sweet pies in Boston” is the head of “sweet”. Our assumption is that—given enough examples—certain patterns emerge (e.g., that “sweet” in combination with food nouns is literal, but is metaphorical if governed by a noun denoting people).
We assume that these patterns occur on different levels, and mainly between syntactically related words. We thus need a data representation to capture these patterns. We borrow its structure from
dependency trees, and the different levels from various annotations. We parse the input sentence with the FANSE parser (Tratz and Hovy, 2011)6. It provides the dependency structure, POS tags, and other information.
To construct the different tree representations, we replace each node in the tree with its word, lemma, POS tag, dependency label, or supersense (the WordNet lexicographer name of the word’s first sense (Fellbaum, 1998)), and mark the word in questio with a special node. See Figure 3 for a graphical representation. These trees are used in addition to the vectors.
This approach is similar to the ones described in (Moschitti et al., 2006; Qian et al., 2008; Hovy et al., 2012).
2.4 Classification Models A tree kernel is simply a similarity matrix over tree instances. It computes the similarity between two trees T1, T2 based on the number of shared subtrees.
We want to make use of the information en-
coded in the different tree representations during
classification, i.e., a forest of tree kernels. We thus combine the contributions of the individual tree representation kernels via addition. We use kernels over the lemma, POS tag, and supersense tree representations, the combination which performed best on the dev set in terms of accuracy.
We use the SVMlight TK implementation by Moschitti (2006).7 We left most parameters set to default values, but tuned the weight of the contribution of the trees and the cost factor on the dev set. We set the multiplicative constant for the trees to 2.0, and the cost factor for errors on positive examples to 1.7.
6http://www.isi.edu/publications/ licensed-sw/fanseparser/index.html
7http://disi.unitn.it/moschitti/ Tree-Kernel.htm
Figure 5 Dependency trees with lexical, part-of-speech, and WordNet supersense features from Hovy et al. (2013).
599
objects of type liquid in the LAB domain, and for money in the FINANCE domain. From this Mason’s system inferred the domain mapping FINANCE—LAB and the concept mapping money–liquid. Mason used WordNet for acquisition of selectional preference classes and, therefore, the source and target domain categories were represented as clusters of WordNet synsets. The domain-specific corpora were obtained by searching the Web for specific terms of interest. Mason conducted two types of evaluation: (1) against the MML, where he manually mapped his output (WordNet synsets) to concrete concepts described in the MML (13 mappings in total) and then measured the accuracy at 77% (a mapping discovered by CorMet was considered correct if submappings specified in the MML were mostly present with high salience and incorrect submappings were present with relatively low salience); and (2) by compiling a list of mappings at random (assumed to be incorrect) and showing that the system assigned low scores to those. Baumer, Tomlinson, and Richland (2009) reimplemented the method of Mason (2004) in the framework of computational metaphor identification (CMI) procedure, and applied it to two types of corpora: student essays and political blogs. The authors presented some interesting examples of conceptual metaphors the system extracted, which they claim may foster critical thinking in social science. However, they did not carry out any quantitative evaluation.
Li, Zhu, and Wang (2013) proposed a method that performs metaphor identification using an “is-a” knowledge base. The authors automatically created two probabilistic knowledge bases by querying the Web using lexico-syntactic patterns. The first knowledge base contained hypernym–hyponym relations and was acquired using Hearst patterns (Hearst 1992). The second knowledge base contained metaphors in the form 〈target is a source〉 learned using a “*BE/VB like*” pattern. The second database was then filtered by removing the hypernym–hyponym relations present in the first database, as well as symmetric relations, to form a metaphor knowledge base. The authors applied the resulting metaphor knowledge base to perform metaphor recognition and explanation. They experimented with nominal metaphors (e.g., “Juliet is the sun”) and verbal metaphors (e.g., “My car drinks gasoline”). In the case of nominal metaphors, the database was queried directly and the corresponding metaphor was either retrieved or not. In the case of verbal metaphors, where the noun denoting the source concept was not explicitly present in the sentence, it was derived based on the selectional preferences of the verbs. The authors computed selectional preferences of the given verb for the nouns present in the knowledge base, and “explained” the given metaphor by the noun exhibiting the highest selectional association with the metaphorical verb. For example, it outputs an explanation “car is a horse” for the metaphor in “my car drinks gasoline,” since the conceptual metaphor CAR IS A HORSE is present in the knowledge base and horse satisfies the subject preference of drink. The authors evaluated their approach on a manually constructed data set of 200 randomly sampled sentences containing “is-a” constructions and 1,000 sentences containing metaphorical and literal uses of verbs. The annotation was carried out at the sentence level (i.e., complete sentences were annotated as metaphorical or not). The authors report an F-score of 69% on the recognition of “is-a” metaphors and that of 58% on the recognition of the verbal ones. Metaphor explanation performance (i.e., the source–target domain mappings generated for each recognized metaphor) was evaluated separately on 214 sentences extracted from linguistic literature (Lakoff and Johnson 1980) and the top-rank precision of 43% is reported. Intuitively, a purely simile-based approach to metaphor is likely to both undergenerate (a large number of metaphors would never be manifested in similelike constructions) and overgenerate (“A is like B” pattern may describe other relations than metaphor). The key contribution of Li, Zhu, and Wang appears to be the filtering
method they introduce, as well as the selectional preference extension of the knowledge base to identify verbal metaphors.
The method of Shutova and Sun (2013) learns metaphorical associations between concepts from the data in an unsupervised way. They created a network (or a graph) of concepts, using hierarchical graph factorization clustering of nouns, and quantified the strength of association between concepts in this graph. Concrete concepts exhibited well-defined association patterns mainly based on subsumption within one domain, whereas abstract concepts tended to have both within-domain and cross-domain associations: the literal ones and the metaphorical ones. For example, the abstract concept of democracy was literally associated with a more general concept of political system, as well as metaphorically associated with the concept of mechanism. Because we often discuss political systems using the mechanism terminology, a corpus-based distributional learning approach learns that they share features with political systems (from their literal uses), as well as with mechanisms (from their metaphorical uses). The system of Shutova and Sun (2013) automatically discovered such association patterns within the graph and used them to identify metaphorical mappings. The mappings were represented in their system as cross-level, one-directional connections between clusters in the hierarchical graph (e.g., the feeling cluster was strongly associated with fire). Example output for the source concepts of fire and disease is shown in Figure 6. To identify metaphorical expressions representing a given mapping, Shutova and Sun used the features that resulted in strong metaphorical associations between the clusters in question (e.g., “passion flared” for FEELING IS FIRE), as shown in Figure 7. The authors evaluated the quality of metaphorical mappings and metaphorical expressions identified by the system against human judgments, as follows: (1) the human judges were presented with a random sample of system-produced metaphorical mappings between the clusters of nouns, as well as the corresponding metaphorical expressions, and asked to mark the ones they considered valid as correct; (2) the human annotators were presented with a set of source domain concepts and asked to write down all target concepts they associated with a given source, thus creating a gold standard. Shutova and Sun report the precision of 0.69 for metaphorical associations and 0.65 for metaphorical expressions, as evaluated against human judgments, and the recall of 0.61 for metaphorical associations, as evaluated against a human-created gold standard. These results are encouraging in that they show that it is possible to induce information about metaphorical mechanisms from distributional properties of concepts alone, without
the use of hand-coded knowledge. Nevertheless, the fact that clustering techniques are typically applied to a limited set of concepts (i.e., cluster a limited set of nouns) somewhat constrains this approach. For instance, whereas common concepts (that are well represented in the data) can be clustered with a high accuracy, this is not always the case for rare concepts for which feature vectors are sparse. Thus an additional technique is needed to map new, unseen concepts to the concepts present in the graph.
Gandy et al. (2013) presented a system that first discovers metaphorical expressions using concreteness algorithm of Turney et al. (2011) and then assigns the corresponding metaphorical mappings using lexical resources and context clustering. They focused on the three types of metaphor defined by Krishnakumaran and Zhu (2007). Once the metaphorical expressions had been identified, Gandy et al. extracted the nouns that the metaphorical words, or facets, tend to co-occur within a large corpus (e.g., the nominal arguments of open in “open government”). The goal of this process was to form candidate nominal analogies between the target noun in the metaphor and the extracted nouns. For example, the expression “open government” suggests an analogy “government∼ door,” according to the authors. Figure 8 shows how nominal analogies are formed based on
metaphorical usage cannot exist for one sense only. Then, the algorithm verifies that the noun N belongs to at least one high-level semantic category (using the WordStat dictionary of semantic categories based on Wordnet1). If not, the algorithm cannot make a decision and stops. Otherwise, it identifies the n nouns most frequently collocated with A, and chooses the k most concrete nouns (using the abstractness scale of (Turney et al. 2011)). The high-level semantic categories represented in this list by at least i nouns each are selected; if N does not belong to the any of them, the phrase is labeled as metaphorical, otherwise it is labeled as non-metaphorical.
Based on exploratory analysis of a development dataset, separate from our test set, we set n = 1000; k = 100; and i = 16.
Nominal Analogies Once a set of metaphorical expressions SM = {hf, nti} (each a pair of a facet and a target noun) is identified, we seek nominal analogies which relate two specific nouns, a source and a target. For example, if the system finds many different linguistic metaphors which can be interpreted as viewing governments as doors, we may have the nominal analogy “government ⇠ door.”
The basic idea behind the nominal analogy finding algorithm is that, since the metaphorical facets of a target noun are to be understood in reference to the source noun, we must find source/target pairs such that many of the metaphorical facets of the target are associated in literal senses with the source (see Figure 3). The stronger and more numerous these associations, the more likely the nominal analogy is to be real.
It should be noted that the system will not only find frequent associations. Since the association strength of each facet-noun pairing is measured by PMI, not its raw frequency, infrequent pairings can (and do) pop up as significant.
Candidate generation. We first find a set SC of nouns as candidate source terms for nominal analogies. This set is the set of all nouns (in a given corpus) which have strong nonmetaphoric associations with facets (adjectives or verbs) that are used in the linguistic metaphors in SM . We start with the set of all facets used in the linguistic metaphors in SM which have a positive point-wise mutual information (PMI; cf. (Pantel and Lin 2002)) with some target term in the set. We then define the set SC to consist of all other (non-target) nouns (in a given large corpus) associated with each of those facets (in the appropriate syntactic relationships) with a positive PMI. A higher threshold than 0 can also be set, though we haven’t seen that increase precision, and it may reduce recall.
For example, consider finding candidate source terms for the target term “government.” We first find all facets in linguistic metaphors identified by the system that are associated with “government” with a positive PMI. These include such terms as “better”, “big”, “small”, “divided”, “open”,
1See http://provalisresearch.com/.
Figure 3: A schematic partial view of the nominal analogy
“government ⇠ door.” Facets in the middle are associated
with the target noun on the left metaphorically, and with the source noun on the right literally.
“closed”, and “limited.” Each of these facets also are associated with other nouns in non-metaphorical senses. For instance, the terms “open” and “closed” are associated with the words “door”, “table”, “sky”, “arms”, and “house”.
Identification. To identify which pairs of candidate source terms in SC and target terms are likely nominal analogies, we seek a measurement of the similarity of the nonmetaphorical facets of each candidate with the metaphorical facets of each target.
To begin, we define for each facet fi and noun (source or target) ni the pair’s association score aij as its PMI, and the pair’s metaphoricity score, mij where mij = 1 if the pair is judged by the system to be metaphorical and mij = 0 if it is judged to be non-metaphorical. In our current system, all metaphoricity scores are either 1 or 0, but (as described below) we plan to generalize the scheme to allow different levels of confidence in linguistic metaphor classification.
We then define the a metaphoric facet distribution (MFT) for facets given target terms by normalizing the product of association and metaphoricity scores:
PM (fi|nj) = aijmijP ij aijmij
as well as a literal facet distribution (LFT), similarly, as:
PL(fi|nj) = aij(1 mij)P ij aij(1 mij)
As noted above, we seek source term / target term pairs such that the metaphorical facets of the target term are likely to be literal facets of the source term and vice versa. A natural measure of this tendency for a candidate source noun ns and candidate target noun nt is the Jensen-Shannon divergence between the LFT of ns and the MFT of nt,
DJS (PL(·|ns) k PM (·|nt)) The larger the J-S divergence, the less likely the pair is to be a good nominal analogy.
Figure 8 Nominal analogy inducti n from Gandy et al. (2013).
602
a collection of metaphorical expressions. The individual (related) nominal analogies were then clustered together to identify conceptual metaphors, as shown in Figure 9. The authors evaluated their system by annotating metaphorical expressions for five target concepts (government, governance, god, father, and mother) in selected sentences from the Reuters corpus (Lewis et al. 2004). They report very encouraging results: Precision (P) = 0.76, Recall (R) = 0.82 for verb metaphors; P = 0.54, R = 0.43 for adjectival metaphors; and P = 0.84, R = 0.97 for copula constructions. The authors also evaluated the quality of conceptual metaphors produced by the system against human judgments and attained a precision of 0.65. However, the scope of the experiment is only limited to the given five concepts and it is not clear how well the method would generalize beyond these. Although the approach of Gandy et al. (2013) seems very promising, a comprehensive evaluation on open-domain corpus data is still necessary to prove its viability.","5 metaphor interpretation systems :In one of the first approaches to metaphor interpretation, Martin (1990) presented a Metaphor Interpretation, Denotation, and Acquisition System (MIDAS), which explained linguistic metaphors through finding the corresponding conceptual metaphor. The method is based on the idea of hierarchical organization of conventional metaphors, namely, that more specific conventional metaphors descend from the general ones. Given an example of a metaphorical expression, MIDAS searched its database for a corresponding metaphor that would explain the anomaly. If it did not find any, it abstracted from the example to more general concepts and repeated the search. If it found a suitable general metaphor, it created a mapping for its descendant, a more specific metaphor, based on the given example. This was also how novel metaphors were acquired. MIDAS was integrated with the Unix Consultant (UC), the system that answers users’ questions about Unix. The UC first tried to find a literal answer to the question. Failing to do so, it called MIDAS, which detected metaphorical expressions via selectional preference violation and searched its database for a metaphor explaining the anomaly in the question.
Another branch of early work on metaphor interpretation relied on performing inferences about entities and events in the source and target domains. The most prominent approaches include the KARMA system (Narayanan 1997, 1999; Feldman and
Narayanan 2004) and the ATT-Meta project (Barnden and Lee 2002; Agerri et al. 2007). Within both systems the authors developed a metaphor-based reasoning framework in accordance with the theory of conceptual metaphor. The reasoning process relied on manually constructed knowledge about the world and operated mainly in the source domain. The results were then projected onto the target domain using the conceptual mapping representation. The ATT-Meta project concerned metaphorical and metonymic description of mental states and reasoning about mental states using first order logic. Their system, however, did not take natural language sentences as input, but logical expressions that are representations of small discourse fragments. KARMA in turn dealt with a broad range of abstract actions and events and took parsed text as input.
Since then the field moved towards acquiring the knowledge necessary for metaphor interpretation automatically (and at a larger scale) from lexical resources, corpora, and the Web. Veale and Hao (2008) derived a “fluid knowledge representation for metaphor interpretation and generation,” called Talking Points. Talking Points are a set of characteristics of concepts belonging to source and target domains and related facts about the world which the authors acquired automatically from WordNet and from the Web. Talking Points were then organized in Slipnet, a framework that allowed for a number of insertions, deletions, and substitutions in definitions of such characteristics in order to establish a connection between the target and the source concepts. This work built on the idea of slippage in knowledge representation for understanding analogies in abstract domains (Hofstadter and Mitchell 1994; Hofstadter 1995). Example (6) demonstrates how slippage operates to explain the metaphor Make-up is a Western burqa.
(6) Make-up => ≡ typically worn by women ≈ expected to be worn by women ≈must be worn by women ≈must be worn by Muslim women
Burqa <=
By doing insertions and substitutions the system arrived from the definition typically worn by women to that of must be worn by Muslim women, and thus established a link between the concepts of make-up and burqa. Veale and Hao (2008), however, did not evaluate to what extent their knowledge base of Talking Points and the associated reasoning framework are useful to interpret metaphorical expressions occurring in text.
Shutova (2010) defined metaphor interpretation as a paraphrasing task and presented a method for deriving literal paraphrases for metaphorical expressions from the BNC. For example, for the metaphors in “All of this stirred an unfathomable excitement in her” or “a carelessly leaked report,” their system produced interpretations All of this provoked an unfathomable excitement in her and a carelessly disclosed report, respectively. They first applied a probabilistic model to rank all possible paraphrases for the metaphorical expressions, given the context; and then used automatically induced selectional preferences to discriminate between figurative and literal paraphrases. The selectional preference distribution was defined in terms of selectional association measures introduced by Resnik (1993) over the noun classes automatically produced by Sun and Korhonen (2009). Shutova (2010) tested her system only on metaphors expressed by a verb and reports an accuracy of 0.81, as evaluated on top-ranked paraphrases produced by the system. However, she used WordNet for supervision, which limits the number and range of paraphrases that can be identified by her method. Shutova, Van de Cruys,
and Korhonen (2012) and Bollegala and Shutova (2013) expanded on this work, addressing the metaphor paraphrasing task in an unsupervised setting and extending the coverage. The method of Shutova, Van de Cruys, and Korhonen (2012) first computed candidate paraphrases according to the context in which the metaphor appeared, using a vector space model. It then used a selectional preference model to measure the degree of literalness of the paraphrases. The authors evaluated their method on the metaphor paraphrasing data set of Shutova (2010) and reported a top-rank precision of 0.52. Bollegala and Shutova (2013) used a similar experimental set-up, however, their method extracted a set of candidate paraphrases from the Web using lexico-syntactic patterns as queries and ranked them based on search engine hits, attaining a precision of 0.42.
Shutova, Teufel, and Korhonen (2013) combined metaphor identification (Shutova, Sun, and Korhonen 2010) and interpretation (Shutova 2010) to perform text-to-text metaphor processing. The resulting system could take arbitrary text as input, parse it using a syntactic parser, identify metaphorical expressions in it, retrieve their literal paraphrases, and output a new version of the text in which metaphors were interpreted. The motivation behind such a set-up was that it allowed for a relatively straightforward integration with external NLP applications. To evaluate the system, the authors extracted a random sample of 200 metaphorical expressions the system identified in the BNC and applied the paraphrasing method to them. They evaluated the accuracy of metaphor identification and interpretation when performed simultaneously, as well as the system’s applicability. The applicability was defined as the proportion of cases where the paraphrase was literal and the meaning of the phrase was retained, indicating whether this type of system paraphrasing would result in an error when hypothetically integrated with an external NLP application. In 54% of cases, the system both identified and interpreted the metaphor correctly, which is a promising result. In a further 13% of cases, the system produced a correct, literal paraphrase for a literal expression erroneously identified as a metaphor, leading to the overall applicability of integrated metaphor processing at 67%. Although the system is easy to integrate with external NLP applications that could benefit from metaphor resolution, it should be noted that some information conveyed by the metaphor is inevitably lost during literal paraphrasing. Metaphor paraphrasing as an approach thus rests on a crucial assumption that the benefit of correct metaphor understanding would outweigh the loss of additional connotations and rhetorical elements. This assumption is yet to be verified through an integration of this technology into real-world NLP, however.
Shutova (2013) presented a computational method that identified metaphorical expressions in unrestricted text by means of their interpretation. She again treated metaphor interpretation as paraphrasing and introduced the concept of symmetric reverse paraphrasing as a criterion for metaphor identification. The hypothesis behind the method is that literal paraphrases of literally used words should yield the original phrase when paraphrased in reverse. For example, when the expression clean the house is paraphrased as tidy the house, the reverse paraphrasing of tidy would generate clean as one of possible paraphrases. Shutova’s expectation was that such symmetry in paraphrasing is indicative of literal use. The metaphorically used words are unlikely to exhibit this symmetry property when paraphrased in reverse. For example, the literal paraphrasing of the verb stir in “stir excitement” would yield “provoke excitement,” but the reverse paraphrasing of provoke would not retrieve stir, indicating the non-literal use of stir. Shutova experimentally verified this hypothesis in a setting involving singleword metaphors expressed by a verb in verb–subject and verb–direct object relations. She applied the selectional preference-based metaphor paraphrasing method (Shutova 2010) to retrieve literal paraphrases of all input verbs and extended the method to
perform metaphor identification by reverse paraphrasing. She evaluated the performance of the system on verb–subject and verb–object relations using the manually annotated metaphor corpus of Shutova and Teufel (2010), reporting a precision of 0.68 and a recall of 0.66. The system outperformed a baseline using selectional preference violation as an indicator of metaphor, that only attained a precision of 0.17 and a recall of 0.55. Some examples of metaphorical expressions identified by the system and their literal paraphrases are shown in Figure 10.","6 investigated techniques and lessons learned :The community has investigated a wide range of techniques and features for metaphor identification and interpretation in a variety of experimental settings. The majority of identification systems focus on the linguistic level, identifying either linguistic metaphor or non-literal language more generally, with a few identifying conceptual metaphor. Table 4 presents a summary of the tasks addressed. Some systems annotated metaphorical expressions at the word level, whereas others opted for the relation level or carried out sentence-level annotation. Individual approaches frequently limited the scope of their experiments to metaphors expressed by a particular part of speech and syntactic construction, as shown in Table 5. The majority of the systems focused on metaphorically used verbs or adjectives, with a few also considering nouns (in modifier or copula constructions) and multiword metaphors. The systems that identified conceptual metaphor also exhibit some variation in the representations they used. Source and target domains were represented as WordNet synsets (Mason 2004); individual nouns (Li, Zhu, and Wang 2013); or clusters of nouns (Shutova and Sun 2013; Gandy et al. 2013). Recent work on metaphor interpretation unfolded along two main axes: metaphor explanation (i.e., identifying the properties of concepts that the metaphor
highlights and the comparisons it involves [Veale and Hao 2008]) and metaphor paraphrasing (i.e., identifying a literal [or more conventional] paraphrase of the metaphorical expression [Shutova 2010; Bollegala and Shutova 2013]).
Identification and interpretation approaches investigated a range of properties of metaphor and implemented them in a variety of system components. The most
Ta bl
e 6
Id en
ti fic
at io
n sy
st em
s: Sy
st em
fe at
ur es
an d
te ch
ni qu
es .
Sy st
em /
fe at
ur es
Se l. pr ef
. V
io la
ti on
C on
cr et
en es
s To
pi ca l st ru ct ur e
C lu
st er
in g
M L
C la
ss ifi
er Le
x re
so ur
ce W
eb se
ar ch
su pe
rv is
ed w
ea kl ysu pe rv is ed
un su
pe rv
is ed
M as
on (2
00 4)
4 –
– –
4 –
4 4
4 – – Bi rk e an d Sa rk ar (2 00 6) – – – – 4 – – – – 4 – G ed ig ia n et al .( 20 06 ) – – – – – 4 4 – 4 – – K ri sh na ku m ar an an d Z hu 20 07 ) – 4 – – – – 4 – 4 – – Sh ut ov a et al .( 20 10 ) 4 – – – 4 – – – – 4 – Li an d Sp or le de r (2 00 9, 20 10 ) – 4 – – – 4 – – 4 – 4 Tu rn ey et al .( 20 11 ) – – 4 – – 4 – – 4 – – N eu m an et al .( 20 13 a, b) 4 4 4 – – – – – 4 – – D un n (2 01 3) – – – – – 4 – – 4 – – Ts ve tk ov et al .( 20 13 ,2 01 4) – – 4 – – 4 4 – 4 – – M oh le r et al .( 20 13 ) – – – 4 4 4 4 – 4 – – H ei nt z et al .( 20 13 ) – – – 4 – – – – – 4 – H ov y et al .( 20 13 ) – 4 – – – 4 4 – 4 – – W ilk s et al .( 20 13 ) 4 4 – – – – 4 – 4 – – St rz al ko w sk ie ta l. (2 01 3) – – 4 4 4 – 4 – 4 – – Sh ut ov a an d Su n (2 01 3) 4 – – – 4 – – – – – 4 G an dy et al .( 20 13 ) – – 4 – 4 – 4 – 4 – – Li et al .( 20 13 ) 4 – – – – – – 4 – 4 – Sh ut ov a (2 01 3) 4 – – – – – 4 – 4 – –
prominent ones include selectional preferences (Martin 1990; Fass 1991; Mason 2004; Krishnakumaran and Zhu 2007; Li and Sporleder 2009, 2010; Shutova 2010; Shutova, Sun, and Korhonen 2010; Hovy et al. 2013; Li, Zhu, and Wang 2013; Wilks et al. 2013); semantic properties of concepts, such as imageability and concreteness (Turney et al. 2011; Gandy et al. 2013; Neuman et al. 2013; Strzalkowski et al. 2013); and topical structure of text (Heintz et al. 2013; Strzalkowski et al. 2013). The common methods used include supervised classification (Gedigian et al. 2006; Dunn 2013a; Hovy et al. 2013; Mohler et al. 2013; Tsvetkov, Mukomel, and Gershman 2013); clustering (Gandy et al. 2013; Shutova and Sun 2013; Shutova, Sun, and Korhonen 2010; Strzalkowski et al. 2013); vector space models (Shutova, Van de Cruys, and Korhonen 2012); the use of lexical resources and ontologies (Mason 2004; Krishnakumaran and Zhu 2007; Dunn 2013b; Gandy et al. 2013; Hovy et al. 2013; Mohler et al. 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013; Wilks et al. 2013); and Web search (Veale and Hao 2008; Bollegala and Shutova 2013; Li, Zhu, and Wang 2013). A summary of techniques investigated by the community is presented in Table 6. In what follows we will discuss the main trends in metaphor processing research and the usefulness of individual types of techniques. Selectional preferences have long established themselves as one of the central components in metaphor-processing research. Wilks’ (1978) selectional preference violation view of metaphor has been highly influential, with numerous approaches to metaphor identification implementing it directly or indirectly (Fass 1991; Martin 1990; Wilks et al. 2013). Other approaches modified this view and treated metaphor as a violation of semantic norm construed more broadly—for example, searching for expressions with low bi-gram probabilities (Krishnakumaran and Zhu 2007), identifying units that break sentence cohesion (Li and Sporleder 2009, 2010), or detecting unusual patterns in words’ compositional behavior (Hovy et al. 2013).
Generally speaking, selectional preference violations (or other semantic violations mentioned above) are a property of surface realization of metaphor rather than its underlying conceptual mechanisms. One needs to bear this in mind when using this as a heuristic. On one hand, such violations are indicative of any kind of non-literalness (i.e., not only metaphor, but also, for instance, metonymy) or anomaly in language and the approach is likely to overgenerate. On the other hand, in the case of most conventional metaphors that are highly frequent, no statistically significant violation can be detected in the data, and the approach would bypass many such metaphors. Shutova (2013) conducted a data-driven study, where verb preferences were automatically acquired from the data and all the nominal arguments below a certain selectional association threshold were considered to represent a violation and were tagged as metaphorical. Such a technique attained a precision of 0.17 and a recall of 0.55, suggesting that the selectional preference violation hypothesis does not port well beyond handcrafted descriptions to large-scale, data-driven techniques.
In contrast, other, “non-violation” applications of selectional preferences have been fruitful in metaphor modeling. Mason (2004) automatically acquired domain-specific selectional preferences of verbs, and then, by mapping their common nominal arguments in different domains, arrived at the corresponding metaphorical mappings. Shutova (2010) presented a modification of Wilks’ view, treating a strong selectional preference fit as a likely indicator of literalness or conventionality. In her metaphor paraphrasing system, Shutova ranks candidate paraphrases based on how well the
context fits their preferences, thus determining their literalness. In their metaphor identification system, Shutova, Sun, and Korhonen (2010) filtered out verbs that have weak selectional preferences, that is, that are equally associated with many argument classes (e.g., choose or remember), as having a lower metaphorical potential. Li, Zhu, and Wang (2013) used selectional preferences to assign the corresponding conceptual metaphor to metaphorical expressions. Although the idea of violation (i.e., treating metaphor as merely a context outlier) is controversial and should be applied with care, selectional preferences themselves are an important source of semantic information about the properties of concepts, which can be successfully exploited in metaphor processing in a variety of ways. Two approaches (Heintz et al. 2013; Strzalkowski et al. 2013) focused on modeling topical structure of text to identify metaphor. The main hypothesis behind these methods is that metaphorical language (coming from a different domain) would represent atypical vocabulary within the topical structure of the text. This intuition is somewhat similar to the idea of semantic norm violation as an indicator of metaphor, although it is different in two crucial ways: (1) topical structure-based approaches explicitly model the interaction of vocabulary from two different domains (i.e., the source and the target); and (2) these approaches take into account domain interactions over extended discourse fragments, rather than individual expressions, thus utilizing information from wider context. Exploiting the wider topical structure of text is a promising avenue for metaphor processing. However, one needs to keep in mind that distributional similaritybased methods risk assigning frequent metaphors to target domains (as is the case for other semantic violation-based methods). For instance, cut may appear more frequently within the domain of economics and finance, rather than its original source domain. The choice of data for training such a model thus becomes crucial, and an appropriately balanced data set is needed. Investigating the topical structure of text is also an important step towards modeling extended metaphor, which interweaves the narrative in complex, but systematic ways. Turney et al. (2011) introduced the idea of measuring concreteness of concepts to predict metaphorical use. The intuition behind their approach is that metaphor is commonly used to describe abstract concepts in terms of more concrete or physical experiences. Thus, Turney and colleagues expect that there would be some discrepancy in the level of concreteness of source and target terms in the metaphor. Neuman et al. (2013) and Gandy et al. (2013) followed in Turney’s steps, reporting promising results. Tsvetkov, Mukomel, and Gershman (2013) took a different route, and used concreteness as one of the features to train a classifier. Strzalkowski et al. (2013) experimented with the imageability feature (that indicates how easy it is to visualize a concept) and demonstrated its relevance to metaphor identification.
Based on the results of these experiments, concreteness is likely to be a practically useful feature for metaphor processing. However, it should be noted that Turney’s hypothesis (that target words tend to be abstract and source words tend to be concrete) explains only a fraction of metaphors and does not always hold. For example, one can use concrete–concrete metaphors (e.g., “broken heart”), abstract–abstract metaphors (“diagnose corruption”), and even abstract–concrete metaphors (“invent a soup”).
However, it may be the case that within the concrete–abstract class of metaphor, the method operates with reasonable performance. Thus concreteness may become a useful feature of a more complex system that takes multiple factors into account, but is unlikely to be a reliable indicator of metaphor on its own. A number of approaches trained classifiers on manually annotated data to recognize metaphor. The key question that supervised classification poses is what features are indicative of metaphor and how can one abstract from individual expressions to its high-level mechanisms? The community has experimented with a number of features, including lexical and syntactic information; higher-level features such as semantic roles, WordNet supersenses, named-entity types, and domain types extracted from ontologies; and semantic properties of concepts, such as animateness and concreteness. Gedigian et al. (2006) classified verb uses as literal or metaphorical, using the verbs’ nominal arguments and their semantic roles (as annotated in PropBank) as features. They reported unusually high performance scores, although the narrow focus on specific lexical items makes it possible for the system to learn a model for individual words rather than performing generalization. In contrast, Dunn (2013a, 2013b) experimented with a wide range of metaphorical expressions from the VU Amsterdam Metaphor corpus, using high-level properties of concepts extracted an ontology, such as domain type and event status. Tsvetkov, Mukomel, and Gershman (2013) and Tsvetkov et al. (2014) used coarse semantic features, such as concreteness, animateness, named-entity types, and WordNet supersenses. What is particularly interesting about this work is that the authors have shown that the model learned with such coarse semantic features is portable across languages, thus suggesting that the chosen features successfully capture some of the properties of metaphor (even if the shallow ones). The work of Hovy et al. (2013) is notable as they focused on compositional rather than categorical features. They trained an SVM with dependency-tree kernels to capture compositional information using lexical, part-of-speech tags, and WordNet supersense representations of sentence trees, achieving successful results. The system of Mohler et al. (2013) aimed at modeling conceptual information in the form of semantic signatures of domains and metaphors. Such rich semantic information is likely to be a successful feature in metaphor recognition: however, Mohler and colleagues experimented within a limited domain and it is not clear how scalable such features would be.
To reliably capture the patterns of the use of metaphor in the data at a large scale, one needs to address conceptual properties of metaphor, along with the surface ones. Thus the models making generalizations at the level of metaphorical mappings and coarse-grained classes of concepts, in essence representing different domains, are likely to yield the optimal framework for the task. However, this hypothesis is yet to be experimentally verified. Clustering techniques were used in numerous approaches, predominantly to identify concepts similar or related to each other. Mason (2004) performed WordNet sense clustering to obtain selectional preference classes, whereas Mohler et al. (2013) used it to determine similarity between concepts and to link them in semantic signatures. Strzalkowski et al. (2013) and Gandy et al. (2013) clustered metaphorically used terms to
form potential source domains. Birke and Sarkar (2006) clustered sentences containing metaphorical and literal uses of verbs.
Another line of research focused on the use of clustering methods to investigate how metaphor partitions the linguistic feature space. Shutova, Sun, and Korhonen (2010) pointed out that the metaphorical uses of words constitute a large portion of the co-occurrence features extracted for abstract concepts from the data. For example, the feature vector for politics would contain GAME or MECHANISM terms among the frequent features. As a result, distributional clustering of abstract nouns with such features identifies groups of diverse concepts metaphorically associated with the same source domain (or sets of source domains). Shutova, Sun, and Korhonen exploit this property of co-occurrence vectors to identify new metaphorical mappings starting from a set of examples. The work of Shutova and Sun (2013) is based on the same observation. Through the use of hierarchical clustering techniques they derive a network of concepts in which metaphorical associations are exhibited at different levels of generality. Peters and Peters (2000) and Wilks et al. (2013) detected metaphor directly in lexical resources. Peters and Peters mine WordNet for examples of systematic polysemy, which allows them to capture metonymic and metaphorical relations. Their system searches for nodes that are relatively high in the WordNet hierarchy (i.e., are relatively general) and that share a set of common word forms among their descendants. Peters and Peters found that such nodes often happen to be in a metonymic (e.g., publisher – publication) or a metaphorical (e.g. theory – supporting structure) relation. Wilks et al. (2013) used WordNet glosses to learn selectional preferences of verbs, which were then used to annotate senses as literal or metaphorical, based on the selectional preference–violation hypothesis. Krishnakumaran and Zhu (2007) use hyponymy relation in WordNet to detect semantic violations. Shutova (2010, 2013) also relied on the hierarchical structure of WordNet, but to identify concepts that share common features (defined as sharing a common hypernym within three levels of the hierarchy).
WordNet synsets were also used to form selectional preference classes in SP-based methods (Mason 2004) or to detect semantically related concepts (Mohler et al. 2013; Strzalkowski et al. 2013; Gandy et al. 2013). Other researchers used WordNet to identify high-level properties of concepts, most notably WordNet supersenses, that served as features for classification (Tsvetkov, Mukomel, and Gershman 2013; Hovy et al. 2013). Because metaphor is a knowledge-intensive phenomenon, multiple approaches attempted to acquire the knowledge necessary for its identification and interpretation from the Web (Veale and Hao 2008; Bollegala and Shutova 2013; Li, Zhu, and Wang 2013). Web search engines provide a flexible tool for retrieving information that matches specific lexico-syntactic patterns (used as queries) and quantifying co-occurrence. Veale and Hao (2008) query the Web to harvest properties of concepts and cultural stereotypes, such as has magical skill for Wizard or has brave spirit for Lion, which are then used to perform metaphor interpretation through property comparison and substitution. Bollegala and Shutova (2013) use the Web to extract co-occurrence information for verbs and nouns, which allows them to generate a set of candidate paraphrases for metaphorical verbs. Li, Zhu, and Wang (2013) query the Web with Hearst patterns to acquire a large knowledge base of hyponymy relations; and simile patterns
(* is like *) to acquire a set of potential conceptual metaphors. Along with the flexibility and convenience of information retrieval tools, these three approaches also boast of wide coverage that the use of the Web allows them to achieve. The knowledge contained on the Web is not merely vast, but it is also constantly updated, which allows the system to stay on par with the current events and trends. Because metaphor is a productive and dynamic phenomenon (new metaphors arise as new events take place), such scalability and ongoing expansion of the Web make it an attractive corpus for metaphor research.","7 system evaluation :System evaluation methodologies continue to be debated in many areas of computational semantics. Identifying a comprehensive and fair evaluation strategy for the task in mind is crucial for the development of fully functional NLP systems. In that light, a number of shared tasks have been proposed over the years at Workshops on Semantic Evaluation (SemEval) that enabled performance comparison across systems and methods. Such tasks as sentiment analysis, word similarity, word sense induction and disambiguation, coreference resolution, lexical substitution, and many others are commonly addressed at SemEval and have a number of benchmark data sets created for them (Agirre and Soroa 2007; McCarthy and Navigli 2007; Lefever and Hoste 2010; Manandhar et al. 2010; Mihalcea, Sinha, and McCarthy 2010; Recasens et al. 2010; Nakov et al. 2013), against which the systems are evaluated and compared. Computational work on metaphor, on the contrary, is considerably more fragmented than similar research efforts in other areas of NLP. With the lack of an established data set, the community has utilized a variety of evaluation strategies, including the use of annotated corpora, human judgments of system output, evaluation against the MML, and annotation of individual selected examples (usually phrases or sentences) via Amazon Mechanical Turk. With a few exceptions, the majority of approaches created their own test sets, making the results not directly comparable.
The most desirable type of evaluation is that conducted against an annotated full-text corpus, namely, naturally occurring, continuous text manually annotated for metaphor. Ideally, such a corpus should be open-domain and representative of a range of genres, making the results indicative of the likely performance of the system on arbitrary text. Another benefit of this type of evaluation is that it allows one to assess both the precision and the recall of the system. However, only two of the presented approaches (Dunn 2013b; Shutova 2013) conducted this type of evaluation, as shown in Table 7. More typically, approaches were instead evaluated on a random sample of system output against human judgments (Mason 2004; Shutova, Sun, and Korhonen 2010; Heintz et al. 2013; Shutova and Sun 2013; Strzalkowski et al. 2013). Although this type of evaluation allows one to measure the precision of the system on a random sample, it does not provide any information about the possible recall. Shutova, Sun, and Korhonen (2010) and Shutova and Sun (2013) applied their methods to a generaldomain corpus (the BNC), from which a random sample of metaphorical expressions annotated by the system was then extracted for evaluation. In contrast, Heintz et al. (2013) and Strzalkowski et al. (2013) collected their data with a focus on a limited domain. The experiments of Mason (2004) and Shutova and Sun (2013) were concerned with conceptual metaphor, and a random sample of the metaphorical mappings identified by the systems was extracted and evaluated against human judgments in terms of precision. However, the latter two approaches also measured recall, by manually compiling a gold-standard of metaphorical mappings for the concepts of interest (Shutova and Sun 2013) or against the MML (Mason 2004).
The majority of approaches (see Table 7) did not apply their systems to continuous text, but rather to a set of pre-selected examples (phrases, sentences, or occasionally paragraphs) in isolation from wider context. Such examples were annotated as metaphorical or literal by independent expert annotators or via Amazon Mechanical Turk. The benefit of this set-up (as opposed to the evaluation on a random sample of system output) is that it allows one to measure both precision and recall. However, the method used for selection of individual examples may introduce a bias into the evaluation and provide an unfair advantage to the system, unless the test sample was selected randomly from arbitrary text. In other words, this type of evaluation is likely to be less objective than the evaluation on continuous corpus text or a random sample.
A number of approaches (Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Gandy et al. 2013; Heintz et al. 2013; Mohler et al. 2013; Neuman et al. 2013; Strzalkowski et al. 2013; Wilks et al. 2013) conducted their experiments within a limited domain. Despite allowing for an in-depth investigation of domain-specific patterns of metaphor use, such evaluations are problematic as they provide no indication of the scalability of the method beyond the studied domain to real-world data. This criticism also applies to the evaluations against MML, as the list is limited in domain coverage and the type of metaphors it provides.
Two data sets stand out as having been repeatedly adopted for metaphor research, enabling direct system comparison. These include the TroFi data set of Birke and Sarkar (2006) and the metaphor paraphrasing data set of Shutova (2010). The TroFi dataset consists of 25 verbs and example sentences, containing their metaphorical and literal use. It was adopted by Turney et al. (2011) and Tsvetkov, Mukomel, and Gershman (2013) in their metaphor identification experiments. The paraphrasing data set and gold standard of Shutova (2010) consists of 52 metaphorically used verbs and their humanderived literal (or more conventional) paraphrases in the given context. The data set
has been used in multiple metaphor interpretation experiments (Shutova 2010; Shutova, Van de Cruys, and Korhonen 2012; Bollegala and Shutova 2013).
The evaluations of metaphor identification tend to be conducted in terms of precision and recall, and occasionally, accuracy. Table 8 presents a summary of results of metaphor identification experiments, classified by domain coverage. The most successful systems attain an F-score in the range of 70–78%, with the highest precision reported for the methods of Tsvetkov, Mukomel, and Gershman (2013), Shutova, Sun, and Korhonen (2010), and Gandy et al. (2013); and the highest recall for the methods of Mohler et al. (2013), Wilks et al. (2013), Gandy et al. (2013) (limited domain) and Tsvetkov, Mukomel, and Gershman (2013) and Hovy et al. (2013) (open domain). However, because the methods were evaluated on data sets of different size and balance of categories, as well as created with different criteria in mind, this comparison is only approximate. A comprehensive evaluation on the same large data set is needed to determine the best performing techniques. For example, the accuracy of 95% reported by Gedigian et al. (2006) was measured on a data set dominated by metaphorical expressions (all metaphor baseline achieves 92%). This result cannot be directly compared to that of a system evaluated on a test set with a balance of metaphorical and literal instances.
One of the key difficulties metaphor processing research is facing today is that of a lack of annotated data. The data sets used in the experiments are typically too
small to obtain a generalizable result. Another issue is that many data sets created to evaluate individual methods were designed for a specific task, focusing on a particular type of metaphor and using an annotation scheme of their own. This makes the results not directly comparable and the overall performance landscape difficult to interpret. This situation calls for the consolidation of the insights gained from the current experiments into a single task definition and the creation of a large data set with this task in mind. Ideally, the systems should be evaluated on an expertannotated corpus, containing continuous, general-domain text. In order to be indicative of the likely performance of the system on real-world data, the corpus needs to have a comprehensive coverage of registers and genres. Finally, a large corpus annotated for metaphor would enable reliable evaluation both in terms of precision and recall.
The work of Dunn (2013b) was exemplary in that he evaluated and compared four different methods on the same data set. Dunn used the VU Amsterdam Metaphor Corpus (Steen et al. 2010), which is currently the largest metaphor corpus in existence, containing approximately 200,000 words. However, because Steen and colleagues were interested in historical aspects of metaphor along with its use in modern language, the VU metaphor corpus contains a large proportion of lexicalized metaphors. Their meanings are ingrained in everyday use and can be interpreted via established techniques (e.g., word sense disambiguation), and their metaphorical nature may or may not be of interest to wider NLP. Whether metaphor processing systems should address highly conventional and dead metaphors depends on the task and application in mind, and corpus annotation should reflect this task definition in a consistent way. As the field moves forward, it would also be desirable to conduct extrinsic evaluations of the metaphor processing systems, in order to determine their usefulness for external NLP applications. One such experiment has already been carried out by Agerri (2008), who has demonstrated that metaphor interpretation plays an important role in textual entailment resolution.","8 conclusion :Metaphor makes our thoughts move vivid and enriches our communication with novel imagery, but most importantly it plays a fundamental structural role in our cognition, helping us organize and project knowledge. As a result, its manifestations are pervasive in language and reasoning, making its computational processing an imperative task within NLP and intelligent systems engineering at large. Despite involving complex comparisons and information transfers, metaphor is a well-structured and systematic phenomenon, highly suitable for computational modeling. Focusing primarily on linguistic metaphor, the community has investigated a range of its aspects, implemented in a variety of system features. Among the most successful features are concreteness, distributional behavior of source and target domain vocabulary, selectional preferences, textual coherence, and topical properties of source and target words. The field has evolved from the widespread use of hand-coded knowledge to mainly datadriven research. Balanced corpora, the Web, Wikipedia and, sometimes, domain-specific corpora have become the primary source of knowledge for metaphor processing. The community has investigated supervised learning, clustering, topic modeling, and pattern-based search to acquire lexical, relational, and domain knowledge from these corpora. Yielding promising results, this was a significant advance in computational modeling of metaphor, allowing for the application of the systems to real-world data.
These experiments also provided new insights on the behavior of metaphor across domains, genres, and types of discourse.
The research on mining metaphorical associations from the data sheds light on how metaphors structure our conceptual system, as well as how specific conceptual metaphors are realized in language, revealing new information about their cognitive processing. Large-scale, automatic identification of conceptual metaphor thus provides a bridge between computational and cognitive research in this area, and has a wider scientific relevance beyond NLP. An interdisciplinary approach, leveraging knowledge from linguistics, cognitive science, psychology, neuroscience, and computer science, would be well-positioned to advance our understanding of metaphorical mechanisms and take the performance of metaphor processing systems to the next level. Two areas that are particularly likely to benefit from an interdisciplinary approach are metaphorical inference and extended metaphor, which have so far escaped attention in NLP. Recent advances in processing linguistic and conceptual metaphor, however, bring us a step closer to understanding and modeling these phenomena.
Despite the promising experimental results reported by the community, little attention has yet been given to real-world applications of metaphor processing. Possible applications include other semantic tasks within NLP and data mining, as well as social science and educational applications. Within NLP, most applications that need to access semantic knowledge would benefit from robust and accurate metaphor resolution. These include, for instance, machine translation, sentiment analysis, or text classification. Because the metaphors we use are known to be indicative of our underlying viewpoints, metaphor processing is likely to be fruitful in determining political affiliation from text or pinning down cross-cultural and crosspopulation differences, and thus become a useful tool in data mining. In social science, metaphor is extensively studied as a way to frame cultural and moral models, and to predict social choice (Landau, Sullivan, and Greenberg 2009; Thibodeau and Boroditsky 2011; Lakoff and Wehling 2012). Metaphor is also widely viewed as a creative tool. Its knowledge projection mechanisms help us to grasp new concepts and generate innovative ideas. This opens many avenues for the creation of computational tools that foster creativity (Veale 2011, 2014) and support assessment in education (Burstein et al. 2013).
The design of the metaphor processing task should thus be informed by the possible applications. The application in mind may place particular requirements on the types of metaphor the system needs to address and the output representations it is expected to produce. Whereas an NLP application, such as machine translation, would be primarily concerned with linguistic metaphors and, possibly, the more creative instances thereof, a data mining application, aiming to detect a set of trends, may find the identification of conceptual metaphors prominent in the data more informative. The formulation of the task and experimental design, in turn, predetermine how system performance is best evaluated. So far, the lack of a common task definition and a shared data set have hampered our progress as a community in this area. This calls for a unification of the task definition and a large-scale annotation effort that would provide a data set for metaphor system evaluation, built with the insights gained from the present studies. The main purpose of this paper is to provide a platform for debate that would assist us in formulating the overall goals of metaphor processing and devising an optimal experimental strategy, enabling us as a community to make significant progress in this important and fascinating area.","ness–concreteness algorithm. However, the evaluation was done on only five","[{""affiliations"": [], ""name"": ""Ekaterina Shutova""}]",SP:dfd9f1276e2968c35fcb8cfca252fc00ef19d9e7,"[{""authors"": [""Agerri"", ""Rodrigo.""], ""title"": ""Metaphor in textual entailment"", ""venue"": ""Proceedings of COLING 2008, pages 3\u20136, Manchester, UK."", ""year"": 2008}, {""authors"": [""Agerri"", ""Rodrigo"", ""John Barnden"", ""Mark Lee"", ""Alan Wallington.""], ""title"": ""Metaphor, inference and domain-independent mappings"", ""venue"": ""Proceedings of RANLP-2007, pages 17\u201323, Borovets."", ""year"": 2007}, {""authors"": [""Agirre"", ""Eneko"", ""Aitor Soroa.""], ""title"": ""Semeval-2007 task 02: Evaluating word sense induction and discrimination systems"", ""venue"": ""Proceedings of the Fourth International Workshop on Semantic"", ""year"": 2007}, {""authors"": [""Allen"", ""James F."", ""Mary Swift"", ""Will de Beaumont.""], ""title"": ""Deep semantic analysis of text"", ""venue"": ""Proceedings of the 2008 Conference on Semantics in Text Processing, STEP \u201908, pages 343\u2013354, Venice."", ""year"": 2008}, {""authors"": [""Badryzlova"", ""Yulia"", ""Natalia Shekhtman"", ""Yekaterina Isaeva"", ""Ruslan Kerimov.""], ""title"": ""Annotating a Russian corpus of conceptual metaphor: A bottom\u2013up approach"", ""venue"": ""Proceedings of the First"", ""year"": 2013}, {""authors"": [""Barnden"", ""John"", ""Mark Lee.""], ""title"": ""An artificial intelligence approach to metaphor understanding"", ""venue"": ""Theoria et Historia Scientiarum, 6(1):399\u2013412."", ""year"": 2002}, {""authors"": [""Baumer"", ""Eric"", ""Bill Tomlinson"", ""Lindsey Richland.""], ""title"": ""Computational metaphor identification: A method for identifying conceptual metaphors in written text"", ""venue"": ""Proceedings of Analogy \u201909, pages 20\u201329,"", ""year"": 2009}, {""authors"": [""Beigman Klebanov"", ""Beata"", ""Eyal Beigman.""], ""title"": ""A game-theoretic model of metaphorical bargaining"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2010}, {""authors"": [""Beigman Klebanov"", ""Beata"", ""Michael Flor.""], ""title"": ""Argumentation-relevant metaphors in test-taker essays"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 11\u201320, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Birke"", ""Julia"", ""Anoop Sarkar.""], ""title"": ""A clustering approach for the nearly unsupervised recognition of nonliteral language"", ""venue"": ""Proceedings of EACL-06, pages 329\u2013336, Trento."", ""year"": 2006}, {""authors"": [""Black"", ""Max.""], ""title"": ""Models and Metaphors"", ""venue"": ""Cornell University Press. Ithaca, NY."", ""year"": 1962}, {""authors"": [""Bollegala"", ""Danushka"", ""Ekaterina Shutova.""], ""title"": ""Metaphor interpretation using paraphrases extracted from the web"", ""venue"": ""PLoS ONE, 8(9):e74304."", ""year"": 2013}, {""authors"": [""Burnard"", ""Lou.""], ""title"": ""Reference Guide for the British National Corpus (XML Edition)"", ""venue"": ""Available from www.natcorp.ox.ac.uk/ corpus/babyinfo.html."", ""year"": 2007}, {""authors"": [""Burstein"", ""Jill"", ""John Sabatini"", ""Jane Shore"", ""Brad Moulder"", ""Jennifer Lentini""], ""title"": ""A user study: Technology to increase teachers\u2019 linguistic awareness to improve instructional language support"", ""year"": 2013}, {""authors"": [""Bzdek"", ""Vincent.""], ""title"": ""Woman of the House: The Rise of Nancy Pelosi"", ""venue"": ""Palgrave Macmillan, Basingstoke, New York."", ""year"": 2008}, {""authors"": [""Cameron"", ""Lynne.""], ""title"": ""Metaphor in Educational Discourse"", ""venue"": ""Continuum, London."", ""year"": 2003}, {""authors"": [""Carbonell"", ""Jaime.""], ""title"": ""Metaphor: An inescapable phenomenon in natural language comprehension"", ""venue"": ""Wendy Lehnert and Martin Ringle, editors, Strategies for Natural Language Processing."", ""year"": 1982}, {""authors"": [""Charteris-Black"", ""Jonathan.""], ""title"": ""Metaphor and vocabulary teaching in ESP economics"", ""venue"": ""English for Specific Purposes, 19(2):149\u2013165."", ""year"": 2000}, {""authors"": [""Charteris-Black"", ""Jonathan"", ""Timothy Ennis.""], ""title"": ""A comparative study of metaphor in Spanish and English financial reporting"", ""venue"": ""English for Specific Purposes, 20(3):249\u2013266."", ""year"": 2001}, {""authors"": [""Chung"", ""Siaw-Fong"", ""Kathleen Ahrens"", ""Chu-Ren Huang.""], ""title"": ""Source domains as concept domains in metaphorical expressions"", ""venue"": ""International Journal of Computational Linguistics and Chinese"", ""year"": 2005}, {""authors"": [""Cilibrasi"", ""Rudi L."", ""Paul M.B. Vitanyi.""], ""title"": ""The Google similarity distance"", ""venue"": ""IEEE Transactional on Knowledge and Data Engineering, 19(3):370\u2013383."", ""year"": 2007}, {""authors"": [""Cortes"", ""Corinna"", ""Vladimir Vapnik.""], ""title"": ""Support-vector networks"", ""venue"": ""Machine Learning, 20(3):273\u2013297. 618"", ""year"": 1995}, {""authors"": [""Davies"", ""Mark.""], ""title"": ""The 385+ million word Corpus of Contemporary American English (1990\u20132008): Design, architecture, and linguistic insights"", ""venue"": ""International Journal of Corpus Linguistics,"", ""year"": 2009}, {""authors"": [""A. Deignan"", ""L. Potter.""], ""title"": ""A corpus study of metaphors and metonyms in English and Italian"", ""venue"": ""Journal of Pragmatics, 36:1231\u20131252."", ""year"": 2004}, {""authors"": [""Desalle"", ""Yann"", ""Bruno Gaume"", ""Karine Duvignau.""], ""title"": ""Slam: Solutions lexicales automatique pour m\u00e9taphores"", ""venue"": ""Traitement Automatique des Langues, 50(1):145\u2013175."", ""year"": 2009}, {""authors"": [""Diaz-Vera"", ""Javier"", ""Rosario Caballero.""], ""title"": ""Exploring the feeling-emotions continuum across cultures: Jealousy in English and Spanish"", ""venue"": ""Intercultural Pragmatics, 10(2):265\u2013294."", ""year"": 2013}, {""authors"": [""Dunn"", ""Jonathan.""], ""title"": ""Gradient semantic intuitions of metaphoric expressions"", ""venue"": ""Metaphor and Symbol, 26(1):53\u201367."", ""year"": 2011}, {""authors"": [""Dunn"", ""Jonathan.""], ""title"": ""Evaluating the premises and results of four metaphor identification systems"", ""venue"": ""Proceedings of CICLing\u201913, pages 471\u2013486, Samos."", ""year"": 2013}, {""authors"": [""Dunn"", ""Jonathan.""], ""title"": ""What metaphor identification systems can tell us about metaphor-in-language"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 1\u201310, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Fass"", ""Dan.""], ""title"": ""met*: A method for discriminating metonymy and metaphor by computer"", ""venue"": ""Computational Linguistics, 17(1):49\u201390."", ""year"": 1991}, {""authors"": [""Feldman"", ""Jerome.""], ""title"": ""From Molecule to Metaphor: A Neural Theory of Language"", ""venue"": ""The MIT Press, Cambridge, MA."", ""year"": 2006}, {""authors"": [""Feldman"", ""Jerome"", ""Srini Narayanan.""], ""title"": ""Embodied meaning in a neural theory of language"", ""venue"": ""Brain and Language, 89(2):385\u2013392."", ""year"": 2004}, {""authors"": [""Fillmore"", ""Charles"", ""Christopher Johnson"", ""Miriam Petruck.""], ""title"": ""Background to FrameNet"", ""venue"": ""International Journal of Lexicography, 16(3):235\u2013250."", ""year"": 2003}, {""authors"": [""Francis"", ""W. Nelson"", ""Henry Kucera.""], ""title"": ""A standard corpus of present-day edited American English"", ""venue"": ""Technical report, Brown University, Providence, RI. Available from www.helsinki.fi/varieng/CoRD/"", ""year"": 1979}, {""authors"": [""Gandy"", ""Lisa"", ""Nadji Allan"", ""Mark Atallah"", ""Ophir Frieder"", ""Newton Howard"", ""Sergey Kanareykin"", ""Moshe Koppel"", ""Mark Last"", ""Yair Neuman"", ""Shlomo Argamon""], ""title"": ""Automatic identification of conceptual"", ""year"": 2013}, {""authors"": [""Gedigian"", ""Matt"", ""John Bryant"", ""Srini Narayanan"", ""Branimir Ciric.""], ""title"": ""Catching metaphors"", ""venue"": ""Proceedings of the 3rd Workshop on Scalable Natural Language Understanding, pages 41\u201348, New York."", ""year"": 2006}, {""authors"": [""Gentner"", ""Deirdre.""], ""title"": ""Structure mapping: A theoretical framework for analogy"", ""venue"": ""Cognitive Science, 7:155\u2013170."", ""year"": 1983}, {""authors"": [""R. Gibbs""], ""title"": ""Literal meaning and psychological theory"", ""venue"": ""Cognitive Science, 8:275\u2013304."", ""year"": 1984}, {""authors"": [""Gong"", ""Shu-Ping"", ""Kathleen Ahrens"", ""Chu-Ren Huang""], ""title"": ""Chinese word sketch and mapping principles: A corpus-based study of conceptual metaphors using the building source"", ""year"": 2008}, {""authors"": [""Grady"", ""Joe.""], ""title"": ""Foundations of Meaning: Primary Metaphors and Primary Scenes"", ""venue"": ""Ph.D. thesis, University of California at Berkeley."", ""year"": 1997}, {""authors"": [""Hardie"", ""Andrew"", ""Veronika Koller"", ""Paul Rayson"", ""Elena Semino.""], ""title"": ""Exploiting a semantic annotation tool for metaphor analysis"", ""venue"": ""Proceedings of the Corpus Linguistics Conference, pages 1\u201312,"", ""year"": 2007}, {""authors"": [""Hearst"", ""Marti A.""], ""title"": ""Automatic acquisition of hyponyms from large text corpora"", ""venue"": ""Proceedings of the 14th Conference on Computational Linguistics - Volume 2, COLING \u201992, pages 539\u2013545, Nantes."", ""year"": 1992}, {""authors"": [""Heintz"", ""Ilana"", ""Ryan Gabbard"", ""Mahesh Srivastava"", ""Dave Barner"", ""Donald Black"", ""Majorie Friedman"", ""Ralph Weischedel.""], ""title"": ""Automatic extraction of linguistic metaphors with lda topic modeling"", ""venue"": ""In"", ""year"": 2013}, {""authors"": [""Hesse"", ""Mary.""], ""title"": ""Models and Analogies in Science"", ""venue"": ""Notre Dame University Press, South Bend, NA."", ""year"": 1966}, {""authors"": [""Hobbs"", ""Jerry R.""], ""title"": ""Metaphor interpretation as selective inferencing"", ""venue"": ""Proceedings of the 7th International Joint Conference on Artificial Intelligence Volume 1, IJCAI\u201981, pages 85\u201391,"", ""year"": 1981}, {""authors"": [""Hofstadter"", ""Douglas.""], ""title"": ""Fluid Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought"", ""venue"": ""HarperCollins Publishers."", ""year"": 1995}, {""authors"": [""Hofstadter"", ""Douglas"", ""Melanie Mitchell.""], ""title"": ""The Copycat Project: A model of mental fluidity and analogy-making"", ""venue"": ""K. J. Holyoak and J. A. Barnden, editors, 619"", ""year"": 1994}, {""authors"": [""Izwaini"", ""Sattar.""], ""title"": ""Corpus-based study of metaphor in information technology"", ""venue"": ""Proceedings of the Workshop on Corpus-based Approaches to Figurative Language, Corpus Linguistics 2003, pages 1\u20138, Lancaster."", ""year"": 2003}, {""authors"": [""Jamrozik"", ""Anja"", ""Eyal Sagi"", ""Micah Goldwater"", ""Dedre Gentner.""], ""title"": ""Relational words have high metaphoric potential"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 21\u201326,"", ""year"": 2013}, {""authors"": [""Karov"", ""Yael"", ""Shimon Edelman.""], ""title"": ""Similarity-based word sense disambiguation"", ""venue"": ""Computational Linguistics, 24(1):41\u201359."", ""year"": 1998}, {""authors"": [""Kingsbury"", ""Paul"", ""Martha Palmer.""], ""title"": ""From TreeBank to PropBank"", ""venue"": ""Proceedings of LREC-2002, pages 1989\u20131993, Gran Canaria."", ""year"": 2002}, {""authors"": [""Koller"", ""Veronika.""], ""title"": ""Metaphor and Gender in Business Media Discourse: A Critical Cognitive Study"", ""venue"": ""Palgrave Macmillan, Basingstoke, New York."", ""year"": 2004}, {""authors"": [""Korkontzelos"", ""Ioannis"", ""Torsten Zesch"", ""Fabio Massimo Zanzotto"", ""Chris Biemann.""], ""title"": ""Semeval-2013 task 5: Evaluating phrasal semantics"", ""venue"": ""Second Joint Conference on Lexical and Computational"", ""year"": 2013}, {""authors"": [""Z. Kovecses""], ""title"": ""Metaphor in Culture: Universality and Variation"", ""venue"": ""Cambridge University Press, Cambridge."", ""year"": 2005}, {""authors"": [""Krishnakumaran"", ""Saisuresh"", ""Xiaojin Zhu.""], ""title"": ""Hunting elusive metaphors using lexical resources"", ""venue"": ""Proceedings of the Workshop on Computational Approaches to Figurative Language, pages 13\u201320,"", ""year"": 2007}, {""authors"": [""Lakoff"", ""George.""], ""title"": ""The Political Mind"", ""venue"": ""Viking, New York."", ""year"": 2008}, {""authors"": [""Lakoff"", ""George"", ""Jane Espenson"", ""Alan Schwartz.""], ""title"": ""The master metaphor list"", ""venue"": ""Technical report, University of California at Berkeley. Available from araw.mede.uic.edu/valansz/"", ""year"": 1991}, {""authors"": [""Lakoff"", ""George"", ""Mark Johnson.""], ""title"": ""Metaphors We Live By"", ""venue"": ""University of Chicago Press, Chicago."", ""year"": 1980}, {""authors"": [""Lakoff"", ""George"", ""Elisabeth Wehling.""], ""title"": ""The Little Blue Book: The Essential Guide to Thinking and Talking Democratic"", ""venue"": ""Free Press, New York."", ""year"": 2012}, {""authors"": [""Landau"", ""Mark J."", ""Daniel Sullivan"", ""Jeff Greenberg.""], ""title"": ""Evidence that self-relevant motives and metaphoric framing interact to influence political and social attitudes"", ""venue"": ""Psychological"", ""year"": 2009}, {""authors"": [""Lefever"", ""Els"", ""V\u00e9ronique Hoste.""], ""title"": ""Semeval-2010 task 3: Cross-lingual word sense disambiguation"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic Evaluation, pages 15\u201320, Uppsala."", ""year"": 2010}, {""authors"": [""Lewis"", ""David D."", ""Yiming Yang"", ""Tony G. Rose"", ""Fan Li.""], ""title"": ""Rcv1: A new benchmark collection for text categorization research"", ""venue"": ""Journal of Machine Learning Research, 5:361\u2013397."", ""year"": 2004}, {""authors"": [""Li"", ""Hongsong"", ""Kenny Q. Zhu"", ""Haixun Wang.""], ""title"": ""Data-driven metaphor recognition and explanation"", ""venue"": ""Transactions of the Association for Computational Linguistics, 1:379\u2013390."", ""year"": 2013}, {""authors"": [""Li"", ""Linlin"", ""Caroline Sporleder.""], ""title"": ""Classifier combination for contextual idiom detection without labelled data"", ""venue"": ""Proceedings of the 2009 Conference on Empirical Methods in Natural Language"", ""year"": 2009}, {""authors"": [""Li"", ""Linlin"", ""Caroline Sporleder.""], ""title"": ""Using Gaussian mixture models to detect figurative language in context"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North"", ""year"": 2010}, {""authors"": [""L\u00f6nneker"", ""Birte.""], ""title"": ""Lexical databases as resources for linguistic creativity: Focus on metaphor"", ""venue"": ""Proceedings of the LREC 2004 Workshop on Language Resources for Linguistic Creativity,"", ""year"": 2004}, {""authors"": [""L\u00f6nneker-Rodman"", ""Birte.""], ""title"": ""The Hamburg metaphor database project: Issues in resource creation"", ""venue"": ""Language Resources and Evaluation, 42(3):293\u2013318."", ""year"": 2008}, {""authors"": [""Low"", ""Graham"", ""Zazie Todd"", ""Alice Deignan"", ""Lynne Cameron.""], ""title"": ""Researching and Applying Metaphor in the Real World"", ""venue"": ""John Benjamins, Amsterdam/Philadelphia."", ""year"": 2010}, {""authors"": [""Lu"", ""Louis"", ""Kathleen Ahrens""], ""title"": ""Ideological influences on building metaphors in Taiwanese presidential"", ""year"": 2008}, {""authors"": [""Manandhar"", ""Suresh"", ""Ioannis Klapaftis"", ""Dmitriy Dligach"", ""Sameer Pradhan.""], ""title"": ""Semeval-2010 task 14: Word sense induction & disambiguation"", ""venue"": ""Proceedings of the 5th International Workshop on"", ""year"": 2010}, {""authors"": [""Martin"", ""James.""], ""title"": ""A Computational Model of Metaphor Interpretation"", ""venue"": ""Academic Press Professional, Inc., San Diego, CA."", ""year"": 1990}, {""authors"": [""Martin"", ""James.""], ""title"": ""A corpus-based analysis of context effects on metaphor comprehension"", ""venue"": ""A. Stefanowitsch and S. T. Gries, editors, Corpus-Based Approaches to Metaphor and Metonymy."", ""year"": 2006}, {""authors"": [""Mason"", ""Zachary.""], ""title"": ""Cormet: A computational, corpus-based conventional metaphor extraction system"", ""venue"": ""Computational Linguistics, 30(1):23\u201344."", ""year"": 2004}, {""authors"": [""McCarthy"", ""Diana"", ""Roberto Navigli.""], ""title"": ""Semeval-2007 task 10: English lexical substitution task"", ""venue"": ""Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007), pages 48\u201353,"", ""year"": 2007}, {""authors"": [""Mihalcea"", ""Rada"", ""Ravi Sinha"", ""Diana McCarthy.""], ""title"": ""Semeval-2010 task 2: Cross-lingual lexical substitution"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic Evaluation, pages 9\u201314,"", ""year"": 2010}, {""authors"": [""Mohler"", ""Michael"", ""David Bracewell"", ""Marc Tomlinson"", ""David Hinote.""], ""title"": ""Semantic signatures for example-based linguistic metaphor detection"", ""venue"": ""Proceedings of the First Workshop on Metaphor"", ""year"": 2013}, {""authors"": [""Moschitti"", ""Ro"", ""Daniele Pighin"", ""Roberto Basili.""], ""title"": ""Tree kernel engineering for proposition re-ranking"", ""venue"": ""Proceedings of Mining and Learning with Graphs (MLG), pages 165\u2013172, Berlin."", ""year"": 2006}, {""authors"": [""Musolff"", ""Andreas.""], ""title"": ""Mirror Images of Europe: Metaphors in the Public Debate about Europe in Britain and Germany"", ""venue"": ""Iudicium, Muenchen."", ""year"": 2000}, {""authors"": [""Nakov"", ""Preslav"", ""Sara Rosenthal"", ""Zornitsa Kozareva"", ""Veselin Stoyanov"", ""Alan Ritter"", ""Theresa Wilson.""], ""title"": ""Semeval-2013 task 2: Sentiment analysis in twitter"", ""venue"": ""Second Joint Conference on Lexical and"", ""year"": 2013}, {""authors"": [""Narayanan"", ""Srini""], ""title"": ""Knowledge-Based Action Representations for Metaphor and"", ""year"": 1997}, {""authors"": [""Narayanan"", ""Srini.""], ""title"": ""Moving right along: A computational model of metaphoric reasoning about events"", ""venue"": ""Proceedings of AAAI 99), pages 121\u2013128, Orlando, FL."", ""year"": 1999}, {""authors"": [""Neuman"", ""Yair"", ""Dan Assaf"", ""Yohai Cohen"", ""Mark Last"", ""Shlomo Argamon"", ""Newton Howard"", ""Ophir Frieder.""], ""title"": ""Metaphor identification in large texts corpora"", ""venue"": ""PLoS ONE, 8(4):e62343."", ""year"": 2013}, {""authors"": [""Niculae"", ""Vlad"", ""Victoria Yaneva.""], ""title"": ""Computational considerations of comparisons and similes"", ""venue"": ""Proceedings of ACL (Student Research Workshop), pages 89\u201395, Sophia."", ""year"": 2013}, {""authors"": [""Niles"", ""Ian"", ""Adam Pease.""], ""title"": ""Towards a standard upper ontology"", ""venue"": ""Proceedings of the International Conference on Formal Ontology in Information Systems - Volume 2001, FOIS \u201901, pages 2\u20139, New York, NY."", ""year"": 2001}, {""authors"": [""Niles"", ""Ian"", ""Adam Pease.""], ""title"": ""Linking lexicons and ontologies: Mapping WordNet to the suggested upper merged ontology"", ""venue"": ""Proceedings of the International Conference on Information and Knowledge"", ""year"": 2003}, {""authors"": [""Nunberg"", ""Geoffrey.""], ""title"": ""Poetic and prosaic metaphors"", ""venue"": ""Proceedings of the 1987 Workshop on Theoretical Issues in Natural Language Processing, pages 198\u2013201, Stroudsburg, PA."", ""year"": 1987}, {""authors"": [""Peters"", ""Wim"", ""Ivonne Peters.""], ""title"": ""Lexicalised systematic polysemy in WordNet"", ""venue"": ""Proceedings of LREC 2000, pages 1\u20137, Athens."", ""year"": 2000}, {""authors"": [""Pragglejaz Group.""], ""title"": ""MIP: A method for identifying metaphorically used words in discourse"", ""venue"": ""Metaphor and Symbol, 22:1\u201339."", ""year"": 2007}, {""authors"": [""Recasens"", ""Marta"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Emili Sapena"", ""M. Ant\u00f2nia Mart\u0131"", ""Mariona Taul\u00e9"", ""V\u00e9ronique Hoste"", ""Massimo Poesio"", ""Yannick Versley""], ""title"": ""Semeval-2010 task 1: Coreference resolution"", ""year"": 2010}, {""authors"": [""Reining"", ""Astrid"", ""Birte L\u00f6nneker-Rodman.""], ""title"": ""Corpus-driven metaphor harvesting"", ""venue"": ""Proceedings of the HLT/NAACL-07 Workshop on Computational Approaches to Figurative Language, pages 5\u201312,"", ""year"": 2007}, {""authors"": [""Resnik"", ""Philip.""], ""title"": ""Selection and Information: A Class-Based Approach to Lexical Relationships"", ""venue"": ""Ph.D. thesis, University of Pennsylvania. 621"", ""year"": 1993}, {""authors"": [""Rohrer"", ""Tim.""], ""title"": ""Conceptual blending on the information highway: How metaphorical inferences work"", ""venue"": ""W. Liebert, G. Redeker, and L. Waugh, editors, Discourse and Perspective in"", ""year"": 1997}, {""authors"": [""Sandhaus"", ""Evan.""], ""title"": ""The New York Times Annotated Corpus"", ""venue"": ""Available from https://catalog/lde.upenn.edu/ LDC2008T19."", ""year"": 2008}, {""authors"": [""Santa Ana"", ""Otto.""], ""title"": ""Like an animal I was treated?: Anti-immigrant metaphor in US public discourse"", ""venue"": ""Discourse Society, 10(2):191\u2013224."", ""year"": 1999}, {""authors"": [""Shutova"", ""Ekaterina.""], ""title"": ""Automatic metaphor interpretation as a paraphrasing task"", ""venue"": ""Proceedings of NAACL 2010, pages 1029\u20131037, Los Angeles, CA."", ""year"": 2010}, {""authors"": [""Shutova"", ""Ekaterina.""], ""title"": ""Metaphor identification as interpretation"", ""venue"": ""Proceedings of *SEM 2013, pages 276\u2013285, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Shutova"", ""Ekaterina"", ""Lin Sun.""], ""title"": ""Unsupervised metaphor identification using hierarchical graph factorization clustering"", ""venue"": ""Proceedings of NAACL 2013, pages 978\u2013988, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Shutova"", ""Ekaterina"", ""Lin Sun"", ""Anna Korhonen.""], ""title"": ""Metaphor identification using verb and noun clustering"", ""venue"": ""Proceedings of Conference on Computational Linguistics 2010, pages 1002\u20131010,"", ""year"": 2010}, {""authors"": [""Shutova"", ""Ekaterina"", ""Simone Teufel.""], ""title"": ""Metaphor corpus annotated for source\u2013 target domain mappings"", ""venue"": ""Proceedings of LREC 2010, pages 3255\u20133261, Malta."", ""year"": 2010}, {""authors"": [""Shutova"", ""Ekaterina"", ""Simone Teufel"", ""Anna Korhonen.""], ""title"": ""Statistical metaphor processing"", ""venue"": ""Computational Linguistics, 39(2):301\u2013353."", ""year"": 2013}, {""authors"": [""Shutova"", ""Ekaterina"", ""Tim Van de Cruys"", ""Anna Korhonen.""], ""title"": ""Unsupervised metaphor paraphrasing using a vector space model"", ""venue"": ""Proceedings of COLING 2012, pages 1121\u20131130, Mumbai."", ""year"": 2012}, {""authors"": [""Skorczynska Sznajder"", ""Hanna"", ""Jordi Pique-Angordans.""], ""title"": ""A corpus-based description of metaphorical marking patterns in scientific and popular business discourse"", ""venue"": ""Proceedings of European"", ""year"": 2004}, {""authors"": [""Trijntje Pasma""], ""title"": ""A Method for Linguistic Metaphor Identification: From MIP to MIPVU"", ""year"": 2010}, {""authors"": [""A. Stefanowitsch""], ""title"": ""Happiness in English and German: A metaphorical-pattern analysis"", ""venue"": ""M. Altenberg and S. Kemmer, editors, Language, Culture, and Mind. CSLI Publications Stanford,"", ""year"": 2004}, {""authors"": [""Strzalkowski"", ""Tomek"", ""George Aaron Broadwell"", ""Sarah Taylor"", ""Laurie Feldman"", ""Samira Shaikh"", ""Ting Liu"", ""Boris Yamrom"", ""Kit Cho"", ""Umit Boz"", ""Ignacio Cases"", ""Kyle Elliot""], ""title"": ""Robust extraction"", ""year"": 2013}, {""authors"": [""Sullivan"", ""Karen.""], ""title"": ""Grammar in metaphor: A construction grammar account of metaphoric language"", ""venue"": ""Ph.D. thesis, University of California, Berkeley."", ""year"": 2007}, {""authors"": [""Sullivan"", ""Karen.""], ""title"": ""Frames and constructions in metaphoric language"", ""venue"": ""John Benjamins, Amsterdam."", ""year"": 2013}, {""authors"": [""Sun"", ""Lin"", ""Anna Korhonen.""], ""title"": ""Improving verb clustering with automatically acquired selectional preferences"", ""venue"": ""Proceedings of EMNLP 2009, pages 638\u2013647, Singapore."", ""year"": 2009}, {""authors"": [""Thibodeau"", ""Paul H."", ""Lera Boroditsky.""], ""title"": ""Metaphors we think with: The role of metaphor in reasoning"", ""venue"": ""PLoS ONE, 6(2):e16782, 02."", ""year"": 2011}, {""authors"": [""Tsvetkov"", ""Yulia"", ""Leonid Boytsov"", ""Anatole Gershman"", ""Eric Nyberg"", ""Chris Dyer.""], ""title"": ""Metaphor detection with crosslingual model transfer"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association"", ""year"": 2014}, {""authors"": [""Tsvetkov"", ""Yulia"", ""Elena Mukomel"", ""Anatole Gershman.""], ""title"": ""Cross-lingual metaphor detection using common semantic features"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 45\u201351, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Turner"", ""Mark"", ""Gilles Fauconnier.""], ""title"": ""Metaphor, metonymy, and binding"", ""venue"": ""A. Barcelona, editor, Metaphor and Metonymy at the Crossroads: A Cognitive Perspective. Walter de Gruyter GmbH,"", ""year"": 2003}, {""authors"": [""Turney"", ""Peter D."", ""Yair Neuman"", ""Dan Assaf"", ""Yohai Cohen.""], ""title"": ""Literal and metaphorical sense identification through concrete and abstract context"", ""venue"": ""Proceedings of the Conference on Empirical"", ""year"": 2011}, {""authors"": [""Veale"", ""Tony.""], ""title"": ""Creative language retrieval: A robust hybrid of information retrieval and linguistic creativity"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human"", ""year"": 2011}, {""authors"": [""Veale"", ""Tony.""], ""title"": ""A service-oriented architecture for metaphor processing"", ""venue"": ""Proceedings of the Second Workshop on Metaphor in NLP, pages 52\u201360, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Veale"", ""Tony"", ""Yanfen Hao.""], ""title"": ""A fluid knowledge representation for understanding and generating creative metaphors"", ""venue"": ""Proceedings of COLING 2008, pages 945\u2013952, Manchester,"", ""year"": 2008}, {""authors"": [""A.M. Wallington"", ""J.A. Barnden"", ""P. Buchlovsky"", ""L. Fellows"", ""S.R. Glasbey.""], ""title"": ""Metaphor annotation: A systematic study"", ""venue"": ""Technical Report"", ""year"": 2003}, {""authors"": [""Wikberg"", ""Kay.""], ""title"": ""The role of corpus studies in metaphor research"", ""venue"": ""2006 Stockholm Metaphor Festival, pages 33\u201348, Stockholm."", ""year"": 2006}, {""authors"": [""Wilks"", ""Yorick.""], ""title"": ""A preferential pattern-seeking semantics for natural language inference"", ""venue"": ""Artificial Intelligence, 6:53\u201374."", ""year"": 1975}, {""authors"": [""Wilks"", ""Yorick.""], ""title"": ""Making preferences more active"", ""venue"": ""Artificial Intelligence, 11(3):197\u2013223."", ""year"": 1978}, {""authors"": [""Wilks"", ""Yorick"", ""Adam Dalton"", ""James Allen"", ""Lucian Galescu.""], ""title"": ""Automatic metaphor detection using large-scale lexical resources and conventional metaphor extraction"", ""venue"": ""Proceedings"", ""year"": 2013}, {""authors"": [""Witten"", ""Ian H."", ""Eibe Frank.""], ""title"": ""Data Mining: Practical Machine Learning Tools and Techniques"", ""venue"": ""Morgan Kaufmann Series in Data Management Systems. Morgan Kaufmann, 2nd edition."", ""year"": 2005}]",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :Metaphor enriches our communication with a more diverse imagery and provides an important mechanism for reasoning about concepts. At the same time, it is also a very common linguistic device that has long become a part of our everyday language. Metaphors arise through systematic associations between distinct, and seemingly unrelated, concepts. For example, when we say “The wheels of Stalin’s regime were well-oiled and already turning,” we view a political system in terms of a mechanism, it can function, break, have wheels, and so forth. The existence of this association allows us to transfer knowledge and inferences from the domain of mechanisms to that of political systems. As a result, we reason about political systems in terms of mechanisms and discuss them using the mechanism terminology, giving rise to a variety of metaphorical expressions.
This view of metaphor is widely known as Conceptual Metaphor Theory (CMT). It was proposed by Lakoff and Johnson (1980), who claimed that metaphor is not merely a property of language, but rather a cognitive mechanism that structures our conceptual system in a certain way. Lakoff and Johnson explained metaphor through a presence of a mapping between two domains of experience: the target (e.g., politics) and the source (e.g., mechanism). Metaphor is thus not limited to meaning extensions of individual
∗ International Computer Science Institute, 1947 Center St. Ste. 600, Berkeley, CA 94704, USA. E-mail: katia@icsi.berkeley.edu.
Submission received: 9 October 2013; revised version received: 4 March 2015; accepted for publication: 10 June 2015.
doi:10.1162/COLI a 00233
© 2015 Association for Computational Linguistics
words, but rather involves a complex cross-domain knowledge projection process. Let us consider a few more examples.
(1) “President Obama is rebuilding the campaign machinery that vaulted him into office” (New York Times 2011)
(2) 20 steps towards a modern, working democracy
(3) Time to mend our foreign policy.
(4) “She knows the nuts and bolts, and it’s not the nuts and bolts inside legislation, it’s the nuts and bolts of raising money, preparing the party for elections, a political consultant kind of politics.” (Bzdek 2008)
These examples demonstrate how multiple properties and inferences from the domain of mechanisms are systematically projected onto our knowledge about politics. Lakoff and Johnson coined the term conceptual metaphor to describe such mappings from the source domain to the target. The view of an inter-conceptual mapping as a basis of metaphor was echoed by other prominent theories in the field. These include, most notably, the comparison view, formulated in the Structure–Mapping Theory of Gentner (1983), and the interaction view (Black 1962; Hesse 1966). However, the principles of CMT have inspired and influenced much of the computational work on metaphor, thus becoming more central to this paper. Conceptual metaphor manifests itself in language in the form of linguistic metaphor, or metaphorical expressions. These in turn include lexical metaphor, that is, single-word meaning extensions (as in Examples (2) and (3)); multi-word metaphorical expressions (e.g., “the government turned a blind eye to corruption”); or extended metaphor, that spans longer discourse fragments.
Manifestations of metaphor are frequent in language, appearing on average in every third sentence of general-domain text, according to corpus studies (Cameron 2003; Martin 2006; Shutova and Teufel 2010; Steen et al. 2010). This makes metaphor an important subject of linguistic research and makes its accurate processing essential for a range of practical NLP applications. These include, for example, (1) machine translation (MT): Because a large number of metaphorical expressions are culture-specific, they represent a considerable challenge for MT (e.g., the English metaphor “to shoot down someone’s arguments” cannot be literally translated into German as “Argumente abschießen” and metaphor interpretation is required); (2) opinion mining: Metaphorical expressions tend to contain a strong emotional component—for example, compare the metaphorical expression “Government loosened stranglehold on business” and its literal counterpart Government deregulated business (Narayanan 1999); (3) information retrieval (IR): Non-literal language without appropriate disambiguation may lead to false positives in information retrieval (e.g., documents describing “old school gentlemen” should not be returned for the query school [Korkontzelos et al. 2013]); and many others.
Because metaphor interpretation requires complex analogical comparisons and the projection of inference structures across domains, the task of automatic metaphor processing is challenging. For many years, computational work on metaphor evolved around the use of hand-coded knowledge and rules to model metaphorical associations, making the systems hard to scale. Recent years have seen a growing interest in statistical modeling of metaphor (Mason 2004; Gedigian et al. 2006; Shutova 2010; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Heintz et al. 2013; Hovy et al. 2013; Li, Zhu, and Wang 2013; Mohler et al. 2013; Shutova and Sun 2013;
Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with many new techniques opening routes for improving system accuracy and robustness. A wide range of methods have been proposed and investigated by the community, including supervised (Gedigian et al. 2006; Dunn 2013a; Hovy et al. 2013; Mohler et al. 2013; Tsvetkov, Mukomel, and Gershman 2013) and unsupervised (Heintz et al. 2013; Shutova and Sun 2013) learning, distributional approaches (Shutova 2010, 2013; Shutova, Van de Cruys, and Korhonen 2012), lexical resource-based methods (Krishnakumaran and Zhu 2007; Wilks et al. 2013), psycholinguistic features (Turney et al. 2011; Gandy et al. 2013; Neuman et al. 2013; Strzalkowski et al. 2013), and Web search (Veale and Hao 2008; Bollegala and Shutova 2013; Li, Zhu, and Wang 2013). Although individual approaches tackling individual aspects of metaphor have met with success, the insights gained from these experiments are still difficult to integrate into a single computational metaphor modeling landscape, because of the lack of a unified task definition, a shared data set, and well-defined evaluation standards. This hampers our progress as a community in this area. In this paper we take a step towards closing this gap: We review the recent work on computational modeling of metaphor, the tasks addressed, the system features proposed, and the evaluations conducted, and analyze the relevance of different linguistic aspects of metaphor for system performance and applicability, with the aim of identifying the desired properties of metaphor processing systems and a set of requirements for their evaluation. 2 considerations in the design of a metaphor processing system :When designing a metaphor processing system, one faces a number of choices. Some stem from the linguistic and cognitive properties of metaphor, others concern the applicability and usefulness of the system in the wider NLP context. In this section, we analyze individual aspects of metaphor and their relevance to computational modeling, as well as their interplay in the design of a real-world system. Linguistic considerations that inform the design of metaphor processing systems concern primarily the choice of the level (or levels) of analysis. The levels of metaphor analysis include (1) linguistic metaphor (or metaphorical expressions), (2) conceptual metaphor, (3) extended metaphor, and (4) metaphorical inference. Let us consider an example of manifestations of the conceptual metaphor EUROPEAN INTEGRATION as a TRAIN JOURNEY, popular in the early 1990s, at various levels.r Conceptual: EUROPEAN INTEGRATION as a TRAIN JOURNEYr Linguistic: The coupling of the carriages may not be reliably secure, but the
pan-European express is in motion.r Extended metaphor: “There is a fear that the European train will thunder forward, laden with its customary cargo of gravy, towards a destination neither wished for nor understood by electorates. But the train can be stopped.” (Margaret Thatcher, Sunday Times, 20 Sept 1992)r Metaphorical inference: The fact that expensive tracks have to be laid for the train to move forward means that someone has to fund the process of European integration.
2.1.1 Linguistic metaphor. Linguistic metaphor, or metaphorical expressions, concern the surface realization of metaphorical mechanisms, and have been unsurprisingly central to metaphor processing research to date (Birke and Sarkar 2006; Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Gandy et al. 2013; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013). Metaphorical expressions represent the way in which text-processing systems encounter metaphor, and there is little doubt that any real-world metaphor processing system, whatever the approach or the application, ultimately needs to be able to identify and interpret them.
When focusing on linguistic metaphor, one needs to further take into account the level of conventionality of the expressions; how different syntactic constructions are used to convey metaphorical meanings and at what level metaphor annotation needs to be done (word, relation, or sentence level). Thus the following considerations become important for the task and system design:r Level of conventionality of the metaphors accepted: Metaphor is a
productive phenomenon, that is, its novel examples continue to emerge in language. However, a large number of metaphorical expressions become conventionalized over time (e.g., “I cannot grasp his way of thinking”). Although metaphorical in nature, their meanings are deeply entrenched in everyday use, and their comprehension is likened to that of literally used terms (Nunberg 1987). According to Gibbs (1984), metaphorical expressions are spread along a continuum from highly conventional, lexicalized metaphors to entirely novel and creative ones. Gibbs thus suggests that there is no clear demarcation line between literal and metaphorical language, and the distinction between them is rather governed by the level of conventionality of metaphorical expressions. From the usage perspective, metaphoricity may be viewed as a gradient phenomenon rather than a binary one (Dunn 2011), and conventionality becomes an important factor in the design and evaluation of metaphor processing systems. It is not yet clear where on the metaphorical–literal continuum the system should draw the line between what it considers metaphorical and what it considers literal. The answer to this question most likely depends on the NLP application in mind. However, generally speaking, real-world NLP applications are unlikely to be concerned with historical aspects of metaphor, but rather with the identification of figurative language that needs to be interpreted differently from the literal language. We therefore suggest that NLP applications do not necessarily need to address highly conventional and lexicalized metaphors that can be interpreted using standard word sense disambiguation techniques, but rather would benefit from the identification of less conventional and more creative language. Much of the metaphor processing work has focused on conventional metaphor, though in principle capable of identifying novel metaphor as well (Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Gandy et al. 2013; Heintz et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with few approaches modeling only novel metaphor (Desalle, Gaume, and Duvignau 2009) or discriminating between conventional and novel metaphors (Krishnakumaran and Zhu 2007). Other approaches have
looked at literal versus non-literal distinction defined more broadly (Birke and Sarkar 2006; Li and Sporleder 2009, 2010).r Syntactic constructions covered: Metaphors vary with respect to how they are expressed in language grammatically. The grammatical structure of metaphorical expressions is tightly coupled with the inference processes that produce them and are guided by the semantic frames they evoke (Sullivan 2007, 2013). According to Sullivan, conceptual metaphors are realized in language via mapping semantic roles in the source frame onto roles in the target frame. This suggests that both the source and the target domain impose a set of constraints on the roles that can be mapped, and thus on the syntactic options for expressing the metaphorical meaning. There has not yet been a computational approach investigating the interplay of frame semantics and surface structure of metaphorical language. However, the community has addressed modeling linguistic metaphor in a range of syntactic constructions. Verbal or adjectival metaphors are particularly widely embraced by NLP researchers (most approaches), with a few works focusing on copula constructions (Krishnakumaran and Zhu 2007; Gandy et al. 2013; Li, Zhu, and Wang 2013; Neuman et al. 2013) or other nominal metaphors (Heintz et al. 2013; Hovy et al. 2013; Li, Zhu, and Wang 2013; Mohler et al. 2013; Strzalkowski et al. 2013). A number of approaches to metaphor identification have also addressed multiword metaphors (Li and Sporleder 2010; Heintz et al. 2013; Hovy et al. 2013; Mohler et al. 2013). Corpus-linguistic research has shown that verbs and adjectives account for a large proportion of metaphorical expressions observed in the data (Cameron 2003; Shutova and Teufel 2010). However, a recent study (Jamrozik et al. 2013) has also shown that relational words tend to have a higher metaphorical potential. This corresponds to the data on verbal metaphor frequency, and it also suggests that relational nouns are important (e.g., “words are friends of translators”).r Lexical, relation, or sentence level: Finally, one needs to decide if metaphor should be annotated at the word level (i.e., tagging the source domain words alone), relation level (i.e., tagging both source and target words in a particular grammatical relation), or sentence level (i.e., tagging sentences that contain metaphorical language, without explicit annotation of source and target domain words). The word or relation levels provide the most information and have been the focus of the majority of approaches (Gedigian et al. 2006; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Gandy et al. 2013; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Shutova and Sun 2013; Wilks et al. 2013). However, some works annotated metaphor at the sentence level (Krishnakumaran and Zhu 2007; Dunn 2013a; Li, Zhu, and Wang 2013; Mohler et al. 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013).
2.1.2 Conceptual metaphor. Conceptual metaphor represents a cognitive and conceptual mechanism by which humans produce and comprehend metaphorical expressions. Manifestations of conceptual metaphor are ubiquitous in language, communication, and even decision-making (Thibodeau and Boroditsky 2011). Here are a few examples of
common metaphorical mappings: TIME is MONEY (e.g., “That flat tire cost me an hour”), IDEAS are PHYSICAL OBJECTS (e.g., “I cannot grasp his way of thinking”), EMOTIONS ARE VEHICLES (e.g., “she was transported with pleasure”), FEELINGS ARE LIQUIDS (e.g., “all of this stirred an unfathomable excitement in her”), LIFE IS A JOURNEY (e.g., “He arrived at the end of his life with very little emotional baggage”); ARGUMENT IS A WAR (e.g., “He shot down all of my arguments,” “He attacked every weak point in my argument”).
On one hand, few would disagree that the metaphor processing systems capable of understanding and applying conceptual metaphor should be in a better position to accurately handle linguistic metaphors as well. However, a series of questions arise when designing a model of conceptual metaphor. For example, how does one represent conceptual metaphors in the system? What labels does one assign to source and target domains? Is it even possible to name all the conceptual metaphors that humans use and is it necessary to do so? A computational model needs a clear definition of what constitutes the source and target domains (whether they are manually listed or automatically learned) and the consistency and coverage of source and target domain categories would play a crucial role in how well the model can account for real-world data. Previous research on annotation of conceptual metaphor (Shutova and Teufel 2010) has shown that the annotators tend to disagree on the assignment of source and target domain categories. The most variation stems from the level of generality of the selected categories, indicating that while cross-domain mappings are intuitive to humans (i.e., they can be annotated in arbitrary text in principle), labeling source and target domains consistently appears to be a challenging task. What this suggests is that, although the mechanism of conceptual metaphor may be helpful to the system, the question of how to best represent source and target domains within the system remains open. A predefined set of categories, such as those widely discussed in the linguistic literature on CMT, may not be sufficient or even suitable for a computational model. And despite the validity of the main principles of CMT as a linguistic theory, it is not straightforward to port it to computational modeling of metaphor. A more flexible, and potentially datadriven, representation of source and target domain categories is needed for the latter purpose. A data-driven representation would also be better suited to account for the freedom of interpretation that some metaphors allow, since more flexible structures can be dynamically learned from the data. So far, the community has attempted assigning manually created labels to metaphorical mappings (Mason 2004; Baumer, Tomlinson, and Richland 2009), harvesting fine-grained mappings between individual nouns (Li, Zhu, and Wang 2013), using lexical resources to define or expand source and target domain categories (Gandy et al. 2013; Mohler et al. 2013), representing source and target concepts as word clusters (Shutova and Sun 2013) or automatically learned topics (Heintz et al. 2013), and learning metaphorical mappings implicitly within the model without explicit labeling (Shutova, Sun, and Korhonen 2010).
2.1.3 Extended metaphor. Extended metaphor refers to the use of metaphor at the discourse level. It manifests itself in discourse via a sequence of metaphorically used language, yielded by the same conceptual metaphor, whereby a continuous scenario from the source domain is metaphorically projected onto the target. For instance, viewing European integration as a train journey led to numerous metaphorical expressions in political discourse. Each of them mapped certain properties of train journeys to political processes—with countries as loosely connected carriages, peoples of different countries as passengers of their respective carriages, expensive tracks that have to be laid for the train to move forward, and the final destination not very well understood. This metaphor has been dominating the debate, with the European leaders arguing over its details
(Beigman Klebanov and Beigman 2010). One can see this metaphor frequently reappear and evolve over time, helping politicians defend their agendas and, not least, shedding some clarity on an otherwise uncertain future.
Research in linguistics and political science (Musolff 2000; Lakoff 2008; Lakoff and Wehling 2012) suggests that the use of a particular metaphor often guides the speakers’ argumentation strategy throughout a piece of discourse, as well as participants’ behavior in a dialogue. Beigman Klebanov and Beigman (2010) investigated extended metaphor within a game-theoretic framework, demonstrating that maintaining the metaphorical frame in a debate is rationalizable in terms of the gains the participants may get from doing so and their potential losses from swerving away from the metaphor. Their approach reverse-engineers the motivations behind the use of extended metaphor within a formal framework. Beigman Klebanov and Beigman’s work was an important advance, enhancing our understanding of the inner workings of extended metaphor, motivation behind its use, and its effects on social dynamics. However, a computational method for identification and interpretation of extended metaphor in real-world discourse is yet to be proposed. A discourse-level metaphor processing system would need to identify a chain of metaphorical expressions in a text, which indicates a systematic association of the text topic with a particular domain. These chains would then demonstrate how continuous scenarios can be transferred across domains. Recovering this information from the data would allow us to better understand the structure behind metaphorical associations, as well as the inferential process by which knowledge is projected across domains. This system would also find application in social science, where metaphorical framing is widely studied as an indicator of the underlying cultural and moral models (Lakoff and Wehling 2012).
2.1.4 Metaphorical inference. When projecting knowledge from one domain to another, a set of complex inferences take place. Metaphorical inferences are grounded in the source domain and result in the production of surface structures we observe in language as metaphorical expressions. Metaphorical mappings are thus realized via projecting inference structures from the source domain onto the target. For example, when European integration is metaphorically viewed as a train journey, our knowledge of typical events and their consequences from the domain of train journey are projected onto our reasoning about the process of European integration. For instance, if we know that expensive tracks need to be laid before a train can move forward, we can infer that someone also needs to fund the process of European integration. Interestingly, in the presence of a conceptual metaphor such inference can take place even without any linguistic metaphor referring to the tracks being present, but rather on the basis of our common sense knowledge about the functioning of trains.
Besides allowing us to derive new information about the target domain, projecting the inferential structures from the source domain also invokes an emotional response coming from the source domain—for example, an unknown destination or a great expense make one feel uneasy when referred to both literally and metaphorically. Such a transfer of inferential processes and emotional content is believed by some to be one of the central purposes of metaphor (Kovecses 2005; Feldman 2006; Thibodeau and Boroditsky 2011). Psychologists Thibodeau and Boroditsky (2011) investigated how metaphor and metaphorical inference affect decision-making. Their hypothesis was that metaphors in language “instantiate frame-consistent knowledge structures [from the source domain] and invite structurally consistent inferences [about the target domain].” They used two groups of subjects, who were presented with two different texts about crime. In the first text crime was metaphorically portrayed as a VIRUS and
in the second as a BEAST. The two groups were then asked a set of questions on how to tackle crime in the city. It turned out that, whereas the first group tended to opt for preventive measures in tackling crime (e.g., stronger social policies), the second group converged on punishment- or restraint-oriented measures. According to Thibodeau and Boroditsky, their results demonstrate that metaphors have profound influence on how we conceptualize and act with respect to societal issues. Interestingly, the participants would explain their decisions via arguments unrelated to the metaphor, showing that the effect of metaphorical inference in guiding human reasoning is rather covert.
Being able to reproduce these inferences is likely to make automatic metaphor understanding better informed and hence more accurate. It may also provide a mechanism of representing source–target domain mappings, that themselves are generalizations over a set of inferences transferred from one domain to another. And finally, these inferences provide a platform for metaphor interpretation, namely, deriving the meaning of a metaphorical expression and the additional connotations it introduces (that are likely to originate from the source domain). There is a consensus among cognitive linguists that it is metaphorical inference that provides for the very texture behind the use of metaphor (Hobbs 1981; Carbonell 1982; Rohrer 1997; Turner and Fauconnier 2003; Feldman 2006). Although uncovering this texture is certainly one of the main objectives of computational metaphor understanding, it is at the same time a very challenging undertaking. Reproducing metaphorical inferences would require the ability to learn vast amounts of world knowledge from the data, as well as performing complex crossdomain comparisons. And despite being a very promising route, it has not yet been attempted in NLP. Another set of considerations in the design of metaphor processing systems stems from the needs of real-world NLP. The high frequency of metaphorical language in textual data makes accurate metaphor processing desired for a number of NLP applications. Thus, the format of metaphor understanding that metaphor processing systems provide should ideally be informed by the requirements of external NLP and a number of considerations arise in that respect:r Metaphor processing typically involves two tasks: metaphor
identification (distinguishing between literal and metaphorical language in text) and metaphor interpretation (identifying the intended meaning of the metaphor). Both of these provide useful information for language understanding and need to be addressed, either independently or as a single process.r A metaphor processing system should provide a representation of metaphor interpretation that can be easily integrated with other NLP systems. This criterion places constraints on how the metaphor processing task should be defined. The most universally applicable metaphor interpretation would be in the text-to-text form. This means that a metaphor processing system would take raw text as input and provide textual output, in which metaphors are interpreted.r In order to be useful for real-world NLP, the system needs to be capable of processing real-world data, and thus operate on unrestricted, continuous text. Rather than only dealing with individual carefully selected clear-cut
examples, the system should be fully implemented and tested on free, naturally occurring text.r To enable wide applicability, the system needs to be open-domain—that is, operate in all domains, genres, and topics. Therefore, ideally it should not rely on any domain-specific information or focus on individual types of instances (e.g., a limited set of hand-chosen source-target domain mappings).r To be easily adaptable to new domains, the system should not rely on task-specific hand-coded knowledge. This means it needs to be either data-driven and be able to automatically acquire the knowledge it needs from text corpora, or rely only on large-scale, general-domain lexical resources (that are already in existence and do not need to be created in a costly manner). However, it would be an advantage if no such resource is required and the system can dynamically induce meanings in context.r To be robust, the system needs to be able to deal with metaphors represented by all word classes and syntactic constructions. Many existing models are designed with specific kinds of metaphorical expressions in mind, for instance nominal metaphors in copula constructions or verbal metaphors in verb–object relations. To be applicable to support real-world NLP applications, these models need to be extended beyond those specific word classes and syntactic constructions, and be able to process any kind of metaphorical language.
When designing a metaphor processing task, methodology, and evaluation strategy, one thus needs to keep these criteria in mind. Although modeling all of the phenomena described in this section within a single system is by no means a requirement, it is critically important to be aware of all the guises that metaphor may take, both conceptually and empirically. 3 metaphor annotation and resources : Metaphor annotation studies have typically focused on one (or both) of the following tasks: (1) identification of metaphorical senses in text (i.e., distinguishing between literal and non-literal meanings), and (2) assignment of the corresponding source– target domain mappings. The majority of corpus-linguistic studies were concerned with metaphorical expressions and mappings within a limited domain—for example, WAR, BUSINESS, FOOD, or PLANT metaphors (Santa Ana 1999; Izwaini 2003; Koller 2004; Skorczynska Sznajder and Pique-Angordans 2004; Chung, Ahrens, and Huang 2005; Hardie et al. 2007; Gong, Ahrens, and Huang 2008; Lu and Ahrens 2008; Low et al. 2010), in a particular genre or type of discourse (Charteris-Black 2000; Cameron 2003; Izwaini 2003; Koller 2004; Skorczynska Sznajder and Pique-Angordans 2004; Martin 2006; Hardie et al. 2007; Lu and Ahrens 2008; Beigman Klebanov and Flor 2013), or in individual examples in isolation from wider context (Wikberg 2006; LönnekerRodman 2008). In addition, these approaches often focused on a small predefined set of source and target domains. Another vein of corpus-based research concerned crosslinguistic differences in the use of metaphor, also in a specific domain—for example,
financial discourse (Charteris-Black and Ennis 2001), metaphors describing FEELINGS (Stefanowitsch 2004; Diaz-Vera and Caballero 2013), or metaphorical expressions referring to body parts (Deignan and Potter 2004). Three recent studies are notable in that they moved away from investigating particular domains to a more general study of how metaphor behaves in unrestricted continuous text. Wallington et al. (2003), Shutova and Teufel (2010), and Steen et al. (2010) conducted consecutive metaphor annotation in open-domain texts.
Wallington et al. (2003) used two teams of annotators and compared externally prescribed definitions of metaphor with intuitive internal ones. Team A was asked to annotate “interesting stretches,” whereby a phrase was considered interesting if (1) its significance in the document was non-physical, (2) it could have a physical significance in another context with a similar syntactic frame, and (3) this physical significance was related to the abstract one. Team B had to annotate phrases according to their own intuitive definition of metaphor. Apart from metaphorical expressions, the respective source–target domain mappings were also to be annotated. For this latter task, the annotators were given a set of mappings from the Master Metaphor List and were asked to assign the most suitable ones. However, the authors do not report the level of interannotator agreement, nor the coverage of the mappings in the Master Metaphor List on their data. The fact that the method is limited to a set of mappings exemplified in the Master Metaphor List suggests that it may not scale well to real-world data, because the predefined inventory of mappings is unlikely to be sufficient to cover the majority of metaphorical expressions in arbitrary text.
Steen and his colleagues (Pragglejaz Group 2007; Steen et al. 2010) proposed a metaphor identification procedure (MIP). In the framework of this procedure, the sense of every word in the text is considered as a potential metaphor. Every word is then tagged as literal or metaphorical, based on whether is has a “more basic, contemporary meaning” in other contexts than the current one. The summary of their annotation procedure is presented is Figure 1. In a sense, such annotation can be viewed as a form of word sense disambiguation with an emphasis on metaphoricity. Steen and colleagues ran a reliability study involving near-native speaker annotators (strongly relying on dictionary definitions) and report an interannotator agreement of 0.85 in terms of Fleiss’ kappa. MIP laid the basis for the creation of the VU Amsterdam Metaphor Corpus1 (Steen et al. 2010). This corpus is a subset of BNC Baby2 annotated for linguistic metaphor. Its size is 200,000 words and it comprises four genres: news text, academic text, fiction, and conversations. The corpus has already found application in computational metaphor processing research (Dunn 2013b; Niculae and Yaneva 2013), as well as inspiring metaphor annotation efforts in other languages (Badryzlova et al. 2013).
The study of Shutova and Teufel (2010) was concerned with annotation of both metaphorical expressions and metaphorical mappings in continuous text. Their annotation procedure is based on MIP, modifying and extending it to the identification of conceptual metaphors along with the linguistic ones. Following MIP, the annotators were asked to identify the more basic sense of the word, and then label the context in which the word occurs in the basic sense as the source domain, and the current context as the target. They were provided with a list of suggested common source
1 http://www.ota.ox.ac.uk/headers/2541.xml. 2 BNC Baby is a 4-million-word subset of the British National Corpus (BNC) (Burnard 2007), comprising
four different genres: academic, fiction, newspaper, and conversation. For more information, see http://www.natcorp.ox.ac.uk/corpus/babyinfo.html.
and target domains, but were also allowed to introduce domains of their own to match their intuitions. Shutova and Teufel’s corpus is a subset of the BNC sampling various genres: fiction, newspaper/journal articles, essays on politics, international relations and sociology, and radio broadcast (transcribed speech). The size of the corpus is 13,642 words, containing 241 metaphorical expressions in total. Table 1 shows the breakdown of metaphors by type, as well as their variation across genres. Shutova and colleagues used the corpus data as a testbed in a number of computational experiments (Shutova 2010; Shutova, Van de Cruys, and Korhonen 2012; Shutova 2013; Shutova, Teufel, and Korhonen 2013). Lakoff and colleagues organized their ideas in a resource called the Master Metaphor List (MML) (Lakoff, Espenson, and Schwartz 1991). The list is a collection of source– target domain mappings (mainly those related to mind, feelings, and emotions) with corresponding examples of language use. The mappings in the list are organized in an ontology—for example, the metaphor PURPOSES ARE DESTINATIONS is a special case of a more general metaphor STATES ARE LOCATIONS. The resource has been criticized for the lack of clear structuring principles of the mapping ontology (Lönneker-Rodman 2008). However, to date MML is the most comprehensive resource for conceptual metaphor in the linguistic literature, and the examples from the list have been used by computational approaches (Mason 2004; Krishnakumaran and Zhu 2007; Li, Zhu, and Wang 2013), both for development and evaluation purposes. The MML also inspired the creation of other resources, including resources in multiple languages that could facilitate cross-linguistic research on metaphor. One such example is the Hamburg Metaphor Database (Lönneker 2004; Reining and Lönneker-Rodman 2007), which contains examples of metaphorical expressions in German and French. The expressions are mapped to senses from EuroWordNet3 and annotated with source–target domain mappings taken from the MML. 4 metaphor identification systems :Early approaches to metaphor relied on information in handcrafted knowledge bases, followed by metaphor identification in and with the help of lexical resources. Recent years have witnessed a growing interest in statistical and machine learning approaches to metaphor identification. As the field of computational semantics—in particular, robust parsing and lexical acquisition techniques—have progressed to the point where it is possible to accurately acquire lexical, domain, and relational information from corpora, this opened many new avenues for large-scale statistical metaphor identification. The vast majority of systems identify metaphor at the linguistic level (Birke and Sarkar 2006; Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with very few focusing on the conceptual level (Mason 2004; Baumer, Tomlinson, and Richland 2009) or identifying both (Gandy et al. 2013; Li, Zhu, and Wang 2013; Shutova and Sun 2013). This section will first present computational approaches to linguistic metaphor identification, then move on to conceptual metaphor. 4.1.1 Approaches Using Hand-Coded Knowledge and Lexical Resources. One of the first approaches to identify and interpret metaphorical expressions in text was proposed by Fass (1991) in his met* system. This system relies on the hypothesis that metaphors often represent a violation of selectional preferences in a given context (Wilks 1975,
3 EuroWordNet is a multilingual database with wordnets for several European languages (Dutch, Italian, Spanish, German, French, Czech, and Estonian). The wordnets are structured in the same way as the Princeton WordNet for English. http://www.illc.uva.nl/EuroWordNet/.
1978). Selectional preferences are the semantic constraints that a predicate places onto its arguments. Consider the following metaphorical expression.
(5) My car drinks gasoline. (Wilks 1978)
The verb drink normally requires a grammatical subject of type ANIMATE and a grammatical object of type LIQUID. Therefore, drink taking a car as a subject in Example (5) is an anomaly, which, according to Wilks, indicates a metaphorical use of drink. met* detects non-literalness via selectional preference violation, utilizing handcrafted descriptions of selectional preferences. In case of a violation, the respective phrase is first tested for being metonymic, using hand-coded patterns (e.g., CONTAINER-FOR-CONTENT). If this fails, the system searches the knowledge base for a relevant analogy in order to discriminate metaphorical relations from the anomalous ones. For example, the sentence “My car drinks gasoline” would be represented in this framework as (car,drink,gasoline), which does not satisfy the preference (animal,drink,liquid), as car is not a hyponym of animal. met* then searches its knowledge base for a triple containing a hypernym of both the actual argument and the desired argument and finds (thing,use,energy source), which represents the metaphorical interpretation. Fass (1991) presented the approach itself, but reported no evaluation results.
More recently, Wilks et al. (2013) revisited this idea, acquiring selectional preferences from lexical resources, namely VerbNet and WordNet. They focused on conventionalized metaphors included in lexical resources and proposed a technique for their automatic identification. They see this work as complementary to the approaches that perform data-driven learning of selectional preferences. The latter, according to the authors, is likely to miss conventional metaphors because of their widespread presence in the data. Wilks and colleagues expect that selectional preferences acquired from term definitions in lexical resources would circumvent this issue and enable them to efficiently detect highly conventionalized metaphors. The main hypothesis behind their approach is that if the first (main) WordNet sense of a word does not satisfy the preferences of its context in a given sentence, but has a lower (less frequent) sense in WordNet that satisfies the preference, then that use of the word and that WordNet sense are likely to be metaphorical. For instance, in the example “Mary married a brick”, the first sense of brick is ‘a physical object,’ thus violating the preference of marry that selects for people, but the second sense of brick as ‘a reliable person’ satisfies this preference. To implement this approach, Wilks and colleagues acquire typical preferences of concepts (i.e., word senses) from WordNet glosses. They use a semantic parser (Allen, Swift, and de Beaumont 2008) to identify the nominal arguments of the verbs in glosses and their semantic roles and then abstract to their higher-level hypernyms in WordNet, which define the preferences. They compared the performance of their system to a baseline using hand-coded verb preferences in VerbNet. The evaluation was carried out on a set of 122 sentences from the domain of Governance, manually annotated for metaphoricity and selected so that the data set contains 50% metaphorical instances and 50% literal ones. They report an F-score of 0.49 for the VerbNet-based system and 0.67 for the WordNet-based one, the latter showing higher recall and the former higher precision. The approach of Wilks et al. rests on the assumption that WordNet sense ranking corresponds somewhat to the literal-to-metaphorical scale, as well as the assumption that there is only one literal sense for the given word. Although this may be true for the majority of senses, it is relatively easy to find counter-examples. For instance, the first WordNet sense of the verb erase is metaphorical, defined as “remove from memory or existence, e.g., The Turks erased the Armenians in 1915,” with the literal sense ranked
second. The reliance on WordNet sense numbering is thus a limitation of the presented approach. Another issue (that the authors point out themselves) is that this approach is likely to detect metonymic uses along with metaphor, and a method to discriminate between the two is still needed.
The method of Krishnakumaran and Zhu (2007) uses hyponymy relation in WordNet and word bigram counts to annotate metaphor at the sentence level. Given an IS-A metaphor (e.g., The world is a stage4) they first verify if the two nouns involved are in hyponymy relation in WordNet, and if this is not the case then the sentence is tagged as containing a metaphor. Along with this they consider expressions containing a verb or an adjective used metaphorically (e.g., “He planted good ideas in their minds” or “He has a fertile imagination”). In such cases, they calculate bigram probabilities of verb– noun and adjective–noun pairs (including the hyponyms/hypernyms of the noun in question). If the combination is not observed in the data with sufficient frequency, the system tags the sentence containing it as metaphorical. This idea follows the intuition of Wilks. However, by using bigram counts over verb–noun pairs Krishnakumaran and Zhu (2007) lose a great deal of information compared with a system extracting selectional preferences for specific grammatical relations from parsed text. The authors evaluated their system on a set of example sentences compiled from the MML, whereby highly conventionalized metaphors (or dead metaphors) are taken to be negative examples, reporting an accuracy of 0.58. Thus Krishnakumaran and Zhu do not deal with literal examples as such: Essentially, the distinction they are making is between the senses included in WordNet, even if they are conventional metaphors, and those not included in WordNet.
4.1.2 Statistical Learning for Metaphor Identification. The first statistical approach to metaphor is the TroFi system (Trope Finder) of Birke and Sarkar (2006). Their method is based on sentence clustering, originating from a similarity-based word sense disambiguation method developed by Karov and Edelman (1998). The method uses a set of seed sentences, where the senses are annotated, computes similarity between the sentence containing the word to be disambiguated and all of the seed sentences, and selects the sense corresponding to the annotation in the most similar seed sentences. Birke and Sarkar adapt this algorithm to perform two-way classification: literal versus non-literal, and they do not clearly define the kinds of tropes they aim to discover. They evaluated their system on a set of 25 verbs (such as absorb, die, touch, knock, strike, pour, etc.), for each of which they extracted a set of sentences containing its literal and figurative uses, 1,298 in total, from the Wall Street Journal corpus. An example for the verb pour in their data set is shown in Figure 2. Two annotators annotated the sentences for literalness, achieving an agreement of κ = 0.77. The authors report a system performance of 53.8% in terms of F-score on this data set.
The metaphor identification system of Shutova, Sun, and Korhonen (2010) also uses clustering techniques, but performs word clustering to discover verb–subject and verb– object metaphors in unrestricted text. It starts from a small seed set of metaphorical expressions, learns the analogies involved in their production, and extends the set of analogies by means of verb and noun clustering. The method is based on the hypothesis of “clustering by association”—namely, that in the course of distributional noun clustering, abstract concepts tend to cluster together if they are associated with the same source domain, whereas concrete concepts cluster by meaning similarity. For
4 William Shakespeare.
instance, democracy and marriage get clustered together, because both are associated with mechanisms, and as such appear with the mechanism terminology in the corpus. This allows the system to discover new, previously unseen conceptual and linguistic metaphors—for example, having seen the seed metaphor “mend marriage” it infers that “the functioning of democracy” is also used metaphorically. This is how the system expands from the seed set to new concepts. Shutova, Sun and Korhonen used a spectral clustering algorithm with lexico-syntactic features to cluster verbs and nouns. They applied their system to continuous text (the whole BNC) and evaluated its performance on a random sample of the extracted metaphors against human judgments. They report a precision of 0.79 with an inter-judge agreement of k = 0.63 among five annotators. Their data-driven system favorably compares to a WordNet-based baseline, where synsets are used in place of automatically derived clusters. Shutova and colleagues have shown that the clustering-based solution has a significantly wider coverage, capturing new metaphors rather than the synonymous ones, as well as yielding a 35% increase in precision. However, Shutova, Sun and Korhonen did not evaluate the recall of their system, which is likely to be dependent on the size of the seed set and a relatively large and representative seed set is needed to achieve full coverage.
Turney et al. (2011) classify verbs and adjectives as literal or metaphorical based on their level of concreteness or abstractness in relation to the noun they appear with. They learn concreteness rankings for words automatically (starting from a set of examples) and then search for expressions where a concrete adjective or verb is used with an abstract noun (e.g., “dark humor” is tagged as a metaphor and “dark hair” is not). They used the data set of Birke and Sarkar (2006) for evaluation of verb metaphors and attain an F-score of 0.68, which favorably compares to that of Birke and Sarkar. For adjectives, they have created their own data set of selected individual adjective–noun pairs for five adjectives: dark, deep, hard, sweet, and warm; 100 phrases in total. These were then manually annotated for metaphoricity. As compared to these annotations, the accuracy of adjective classification is 0.79. However, the adjective data set was constructed with the concreteness feature in mind, and therefore the results reported for verb metaphors are likely to be more objective.
Neuman et al. (2013) proposed an extension to the method of Turney et al. (2011) by incorporating the concept of selectional preferences into the concreteness-based model of metaphor. Their goal was to improve the performance of Turney’s algorithm by covering metaphors formed of concrete concepts only (e.g., “broken heart”) by detecting selectional preference violations. The authors address three types of metaphor introduced by Krishnakumaran and Zhu (2007) and they claim to have expanded on Turney’s work by carrying out a more comprehensive evaluation of the abstractness–concreteness algorithm. However, the evaluation was done on only five
target concepts: governance, government, God, mother, and father. Sentences describing these concepts have been extracted from the Reuters (Lewis et al. 2004) and New York Times (Sandhaus 2008) corpora and annotated for metaphoricity. The authors measured the average precision of their system on this data set at 0.72 and the average recall at 0.80. The improvement over Turney’s evaluation set-up was the annotation of complete sentences rather than isolated phrases. However, it should be noted that the system was evaluated on selected examples rather than continuous text.
Heintz et al. (2013) applied Latent Dirichlet Allocation (LDA) topic modeling (Blei, Ng, and Jordan 2003) to the problem of metaphor identification in experiments with English and Spanish. Their goal was to create a minimally supervised metaphor processing system that can be applied to low-resource languages. The hypothesis behind their system is that a sentence that contains both source and target domain vocabulary contains a metaphor. The authors focused on the target domain of governance and have manually compiled a set of source concepts with which governance can be associated. They use LDA topics as proxies for source and target concepts, and if vocabulary from both source and target topics is present in a sentence, this sentence is tagged as containing a metaphor. The topics are learned from Wikipedia and then aligned to source and target concepts using sets of human-created seed words. When the metaphorical sentences are retrieved, the source topics that are common in the document are excluded, thus ensuring that the source vocabulary is transferred from a new domain. Although this allows the authors to filter out some literal uses, this may also lead to discarding cases of extended metaphor. The authors collected the data for their experiments from news Web sites and governance-related blogs in English and Spanish. They ran their system on this data, and output a ranked set of metaphorical examples. They carried out two types of evaluation: (1) top five examples for each conceptual metaphor judged by two annotators, reporting an F-score of 0.59 for English (κ = 0.48); and (2) 250 top-ranked examples in system output annotated for metaphoricity using Amazon Mechanical Turk, yielding a mean metaphoricity of 0.41 (standard deviation = 0.33) in English and 0.33 (standard deviation = 0.23) in Spanish. One of the assumptions behind Heintz et al.’s method is that the same source–target domain mappings manifest themselves across languages. Although this is likely to be true for primary metaphors (Grady 1997), as the authors point out themselves, this assumption may not extend to a broader spectrum of metaphors, and thus may lead to limited coverage in some languages, as well as false positives. Another issue that comes to mind concerns the learning of the topics themselves: Because a large number of metaphors are used conventionally within a particular topic (e.g. “cut taxes”), in principle such an approach would learn them as part of the target domain topics and may thus fail to recognize them as source domain terms. However, the authors do not comment on how often this was observed in their data.
The method of Strzalkowski et al. (2013) also relies on modeling the topical structure of text, although using different techniques from Heintz et al. (2013). If Heintz et al. used LDA-acquired topics as approximations of concepts, Strzalkowski and colleagues identify topical chains (Broadwell et al. 2013) by looking for sequences of concepts in text. They also experiment within a limited domain, the target domain of governance. Their method first identifies sentences containing target domain vocabulary and extracts the surrounding five-sentence passage. They then identify topical chains in that passage, by linking the occurrences of nouns and verbs, including repetition, lexical variants, pronominal references, and WordNet synonyms and hyponyms. By virtue of this linking, the authors claim to “uncover the topical structure [of the text] that holds the narrative together.” Their main hypothesis is that metaphorically used terms
typically occur outside the core topical structure of the text, because they represent vocabulary imported from a different domain. For all the words that are found outside the topical chains, Strzalkowski et al. compute imageability scores and retain the highestscoring ones as candidate metaphors, if they are in a syntactic relation with any of the target domain terms. The authors then extract common contexts in which the candidates are used in text corpora and cluster these contexts in order to identify potential source domains, the so-called “proto-sources.” Strzalkowski et al. (2013) evaluated the performance of their method on four languages: English, Spanish, Russian, and Farsi. The evaluation was carried out against human judgments of system output that were obtained using Amazon Mechanical Turk. The authors report a metaphor identification accuracy of 71% in English, 80% in Spanish, 69% in Russian, and 78% in Farsi. According to the paper, “hundreds” of instances were annotated in each language, although the exact number of instances is not reported. While the system performance is high, it should be noted that the experiments were carried out within a limited domain, and it is possible that the approach is not equally applicable to all domains. Because of its high reliance on imageability scores, it is likely to be able to delineate metaphorical language reasonably well for the abstract target domains, but less so for the concrete target domains. In the latter case, the target domain words may also exhibit high imageability, and the system would then rely solely on topic chain extraction to differentiate between literal and metaphorical language. The performance of the generalized system is thus dependent on the accuracy of topic chain extraction, which has not been evaluated independently. In addition, the current method ignores low-imageability metaphors, which abound even within the studied domain (e.g., “invent a new form of governance”). Despite the lack of generality, Strzalkowski et al.’s work, however, makes important contributions in that it addresses (though indirectly) the behavior of metaphor in discourse, and their framework can be viewed as a step towards modeling extended metaphor.
Many other statistical methods treated metaphor identification as a classification problem. Such methods are described in the following section.
4.1.3 Metaphor Identification as a Classification Problem. Gedigian et al. (2006) presented a method that discriminates between literal and metaphorical language, using a maximum entropy classifier. They obtained their training and test data by extracting the lexical items whose frames are related to MOTION and CURE from FrameNet (Fillmore, Johnson, and Petruck 2003). They then searched the PropBank Wall Street Journal corpus (Kingsbury and Palmer 2002) for sentences containing such lexical items and annotated them with respect to metaphoricity. They used PropBank annotation (arguments and their semantic types) as features to train the classifier and report an accuracy of 95.12%. This result is, however, only a little higher than the performance of the naive baseline assigning majority class to all instances (92.90%). These numbers can be explained by the fact that 92.00% of the verbs of MOTION and CURE in the Wall Street Journal corpus are used metaphorically, thus making the data set unbalanced with respect to the target categories and the task easier.
The system of Li and Sporleder (2009, 2010) detects idioms by measuring semantic similarity within and between the literal and non-literal parts of an utterance. The nonliteral language considered by their model includes metaphors, as well as other types of figurative language. Their main assumption is that figurative uses break cohesion in the sentence, which is defined by a similarity measure. This idea also goes back to Wilk’s selectional preference violation approach to metaphor; however, combinations of word usages with larger sentential context are considered to determine the mismatch
(or violation). Li and Sporleder used Normalized Google Distance (Cilibrasi and Vitanyi 2007) as a similarity measure and a combination of classifiers (support vector machines [SVM] and Gaussian mixture models [GMM]) using similarity (or cohesion) information as features to learn idiomaticity scores. They evaluated their system on a data set of 17 idioms and their literal and non-literal contexts. For each expression, its occurrences were extracted from the Gigaword corpus along with five paragraphs of context. These examples were then annotated for literalness with an inter-annotator agreement of κ = 0.7. There were 3,964 examples in total, with approximately 80% of them being non-literal. They evaluated the method using 10-fold cross-validation and report an F-score of 0.75. However, they did not evaluate their system on metaphorical language independently.
Dunn (2013a, 2013b) presented an ontology-based domain interaction system MIMIL (Measuring and Identifying Metaphor in Language), which identifies metaphorical expressions at the utterance level. Dunn’s system first maps the lexical items in the given utterance to concepts from SUMO ontology (Niles and Pease 2001, 2003), assuming that each lexical item is used in its default sense (i.e., no sense disambiguation is performed). The system then extracts the properties of concepts from the ontology, such as their domain type (ABSTRACT, PHYSICAL, SOCIAL, MENTAL) and event status (PROCESS, STATE, OBJECT). Those properties are then combined into feature-vector representations of the utterances. Dunn then applied a logistic regression classifier implemented in Weka (Witten and Frank 2005), using these features to perform metaphor identification. The work of Dunn (2013a, 2013b) is notable as he conducted evaluation of four types of approaches and compared their performance on the same task (identification of metaphorical expressions in continuous text) and on the same data (Corpus of Contemporary American English [CoCA] [Davies 2009] and VU Amsterdam Metaphor Corpus [Steen et al. 2010]). The evaluated approaches included the semantic similarity measurement method of Li and Sporleder (2009, 2010); the concreteness-based method of Turney et al. (2011); the clustering-based method of Shutova, Sun, and Korhonen (2010) modeling source–target domain mappings; and his own domain interaction method. Dunn re-implemented the four approaches as closely as possible to the original systems, although with some adjustments. A number of Dunn’s adjustments were operational (e.g., using logistic regression instead of SVM for the implementation of the similaritybased method of Li and Sporleder (2009); using a k-means clustering approach instead of spectral clustering for the method of Shutova, Sun, and Korhonen (2010); and using a different semantic relatedness measure for the method of Li and Sporleder (2009)). However, some adjustments were conceptual, for instance, using bag-of-words based semantic relatedness instead of dependency-based distributional similarity in the reimplementation of the clustering system. Admitting that these adjustments may have impacted the results and, as such, may not be an accurate reflection of the performance of the original algorithms in full, Dunn’s comparison of individual system features that were re-implemented nonetheless sheds light on the importance of particular properties of concepts for metaphor identification. In his first study (Dunn 2013a), he evaluated the systems on the CoCA data, where the sentences were annotated as metaphorical, literal, or humorous (however, neither the size of the data set nor the annotation procedure are described in Dunn’s article). On this data set, the clustering-based system and the domain-interaction method significantly outperformed the other two systems, as shown in Figure 3. Dunn explains such discrepancy by the fact that the former systems are both theory-based and aim to model the underlying mechanisms of metaphor, while the similarity-based and abstractness-based systems model its surface realizations. In his second study, conducted on the VU Amsterdam Metaphor corpus, Dunn (2013b)
Shutova Design and Evaluation of Metaphor Processing Systems
15
testing; the evaluations were performed using cross-validation (100 folds). The F-measures reported here are for metaphor classification only (i.e., precision for non-metaphor is not directly considered because this inflates the performance of the systems). This is done because some of the systems greatly over-identify literal utterances; however, because literal utterances dominate the evaluation data set, the over-identification of literal utterances would disproportionately raise
the average F-measure for all classes in these systems. The feature vectors and other material
used in the evaluation can be found at http://www.jdunn.name.
humor, and literal utterances, the similarity and abstractness systems performed very poorly,
essentially identifying no metaphors. The joint system performed worse than the domain
interaction system, showing that the abstractness and similarity features reduce performance. As
shown in later tests, the measurements of abstractness and semantic similarity, both at the word-
level, simply do not distinguish between metaphor and non-metaphor in a realistic data set. The
reports different results, however. He evaluated the systems on two versions of the data set, one where named entities have been recognized during pre-processing and one without named-entity recognition. The results are shown in Tables 2 and 3, respectively. Here, the domain-interaction and abstractness-based methods are leading, with the clustering-based method coming third. The difference in the results may be explained by the properties of the VU Amsterdam corpus. The corpus was compiled with an interest in historic aspects of metaphor, and, therefore, highly conventional and lexicalized metaphors account for a large proportion of the data. What this suggests is that the domain-interaction and abstractness-based approaches are perhaps better-suited for processing lexicalized metaphors, whereas the clustering and similarity-based systems may fail to identify those due to their high frequency and the near-literal behavior in the data. In contrast, the domain-interaction system, which is knowledge-based, and the abstractness system, which relies on a non-changing property of concepts (i.e., concreteness), thus appear to be well-suited for handling lexicalized metaphors.
Tsvetkov, Mukomel, and Gershman (2013) presented a supervised learning approach that makes use of coarse semantic features. They experimented with metaphor
identification in English and Russian, first training a classifier on English data only, and then projecting the trained model to Russian using a dictionary. They abstracted from the words in the English data to their higher level features, such as concreteness, animateness, named entity labels, and coarse-grained WordNet categories (corresponding to WN lexicographer files,5 e.g., noun.artifact, noun.body, verb.motion, verb.cognition). They focused on subject–verb–object constructions and annotated metaphor at the sentence level. The authors used a logistic regression classifier and the combination of coarse semantic features for this purpose. They evaluated their model on the TroFi data set (Birke and Sarkar 2006) for English and a self-constructed data set of 140 sentences for Russian, attaining the F-scores of 0.78 and 0.76, respectively. Tsvetkov et al. (2014) extended this experiment to identify adjective–noun metaphors using similar features, as well as porting the model to two further languages (Spanish and Farsi), achieving F-scores in the range of 0.72 to 0.85. The results are encouraging and show that porting coarse-grained semantic knowledge across languages is feasible. However, it should be noted that the generalization to coarse semantic features inevitably only captures shallow behavior of metaphorical expressions in the data and bypasses conceptual information. In reality, as confirmed by corpus-linguistic studies (Charteris-Black and Ennis 2001; Kovecses 2005; Diaz-Vera and Caballero 2013), there is considerable variation in metaphorical language across cultures, which makes training only on one language and simply translating the model less suitable for modeling conceptual structure behind metaphor, which is one of the limitations of this approach. However, the experiments of Tsvetkov and colleagues suggest that coarse semantic features could be a useful component of a more complex system.
The approach of Mohler et al. (2013) relied on the concept of semantic signature of a text. The authors defined semantic signatures as a set of highly related and interlinked WordNet senses. They induced domain-sensitive semantic signatures of texts and then trained a set of classifiers to detect metaphoricity within a text by comparing its semantic signature to a set of known metaphors. The main intuition behind this approach is that the texts whose semantic signature closely matches the signature of a known metaphor is likely to represent an instance of the same conceptual metaphor. Mohler and colleagues conducted their experiments within a limited domain (the target domain of governance) and manually constructed an index of known metaphors for this domain. They then automatically created the target domain signature and a signature for each source domain among the known metaphors in the index. This was done by means of semantic expansion of domain terms using WordNet, Wikipedia links, and corpus cooccurrence statistics. Given an input text their method first identified all target domain terms using the target domain signature, then disambiguated the remaining terms using sense clustering and classified them according to their proximity to the source domains listed in the index. For the latter purpose, the authors experimented with a set of classifiers, including a maximum entropy classifier, an unpruned decision tree classifier, support vector machines, a random forest classifier, as well as the combination thereof. They evaluated their system on a balanced data set containing 241 metaphorical and 241 literal examples, and obtained the highest result of F-score of 0.70 using the decision tree classifier.
Hovy et al. (2013) used the idea of selectional preference violation as the indicator of metaphor, taking it to the next level. They trained an SVM classifier (Cortes and Vapnik 1995) with tree kernels (Moschitti, Pighin, and Basili 2006) to capture compositional
5 http://wordnet.princeton.edu/man/lexnames.5WN.html.
Figure 1: Examples of a metaphor seed, the matching Brown sentences, and their annotations
thus probably not very helpful. However, we would like to capture semantic aspects of the word and represent it in an expressive way. We use the existing vector representation SENNA (Collobert et al., 2011) which is derived from contextual similarity. In it, semantically similar words are represented by similar vectors, without us having to define similarity or looking at the word itself. In initial tests, these vectors performed better than binary vectors straightforwardly derived from features of the word in context.
2.3 Constructing Trees
a) b) c)like
I people
the sweet in
Boston
NNS
DT JJ IN
n.group
O adj.all O
NNP n.location
VB
PRP
v.emotion
O
Figure 3: Graphic demonstration of our approach. a) dependency tree over words, with node of interest labeled. b) as POS representation. c) as supersense representation
The intuition behind our approach is that metaphorical use differs from literal use in certain syntactic relations. For example, the only difference between the two sentences “I like the sweet people in Boston” and “I like the sweet pies in Boston” is the head of “sweet”. Our assumption is that—given enough examples—certain patterns emerge (e.g., that “sweet” in combination with food nouns is literal, but is metaphorical if governed by a noun denoting people).
We assume that these patterns occur on different levels, and mainly between syntactically related words. We thus need a data representation to capture these patterns. We borrow its structure from
dependency trees, and the different levels from various annotations. We parse the input sentence with the FANSE parser (Tratz and Hovy, 2011)6. It provides the dependency structure, POS tags, and other information.
To construct the different tree representations, we replace each node in the tree with its word, lemma, POS tag, dependency label, or supersense (the WordNet lexicographer name of the word’s first sense (Fellbaum, 1998)), and mark the word in question with a special node. See Figure 3 for a graphical representation. These trees are used in addition to the vectors.
This approach is similar to the ones described in (Moschitti et al., 2006; Qian et al., 2008; Hovy et al., 2012).
2.4 Classification Models A tree kernel is simply a similarity matrix over tree instances. It computes the similarity between two trees T1, T2 based on the number of shared subtrees.
We want to make use of the information encoded in the different tree representations during classification, i.e., a forest of tree kernels. We thus combine the contributions of the individual tree representation kernels via addition. We use kernels over the lemma, POS tag, and supersense tree representations, the combination which performed best on the dev set in terms of accuracy.
We use the SVMlight TK implementation by Moschitti (2006).7 We left most parameters set to default values, but tuned the weight of the contribution of the trees and the cost factor on the dev set. We set the multiplicative constant for the trees to 2.0, and the cost factor for errors on positive examples to 1.7.
6http://www.isi.edu/publications/ licensed-sw/fanseparser/index.html
7http://disi.unitn.it/moschitti/ Tree-Kernel.htm
54
properties of metaphorical language. Their hypothesis is that unusual semantic compositions in the data may be indicative of the use of metaphor. They trained the model on labeled examples of literal and metaphorical uses of 329 words (3,872 sentences in total), with an expectation to learn the differences in their compositional behavior in the given lexico-syntactic contexts. The choice of dependency-tree kernels helped to capture such compositional properties, according to the authors. The authors constructed their data set by extracting sentences from the Brown corpus (Francis and Kucera 1979) that containe the words of interest, and annotating them f r metaphoricity using Amazon Mechanical Turk. Example entries for the adjective br ght is shown i Figure 4. Eighty percent of the data were used for training purposes, 10% for parameter tuning, and 10% for the evaluation. The learning was carried out using word vectors, as w ll as lexical, part-of-speech tags, and WordNet supersense representations of sente ce trees as features, as shown in Figure 5. The authors reported encouraging results (F-score = 0.75), which is an indication of the importance of syntactic information and compositionality in metaphor identification. The first method for automatic identification of conceptual metaphor was the CorMet system of Mason (2004). CorMet induced metaphorical mappings by identifying systematic variations in domain-specific selectional preferences, which were learned in a data-driven way. For example, the verb pour has a strong selectional preference for
B low he could see the bright torches lighting the riverbank . no
Her bright eyes were twinkling . yes Washed , they came out surprisingly clear and bright . no
Figure 1: Exampl s of a metaphor seed, the matching Brown sentences, and their annotations
thus probably not very helpf l. However, we woul like to capture semantic aspects of the word and epresent it in an expressive way. We use t xi t-
ing vector representation SENNA (Collobert et al., 2011) which s derived from contextual simila ity. In it, semantically similar words are represe ted by similar vectors, without us having to define similarity or looking at the word itself. In initial tests, these vectors performed better than binary vectors straightforwardly derived from features f the word in context.
2.3 Constructing Trees
a) b) c)like
I people
the sweet in
Boston
NNS
DT JJ IN
n.group
O adj.all O
NNP n.location
VB
PRP
v.emotion
O
Figure 3: Graphic demonstration of our approach. a) de-
pendency tree over words, with node of interest labeled.
b) as POS representation. c) as supersense representation
The intuition behind our approach is that metaphorical use differs from literal use in certain syntactic relations. For example, the only difference between the two sentences “I like the sweet people in Boston” and “I like the sweet pies in Boston” is the head of “sweet”. Our assumption is that—given enough examples—certain patterns emerge (e.g., that “sweet” in combination with food nouns is literal, but is metaphorical if governed by a noun denoting people).
We assume that these patterns occur on different levels, and mainly between syntactically related words. We thus need a data representation to capture these patterns. We borrow its structure from
dependency trees, and the different levels from various annotations. We parse the input sentence with the FANSE parser (Tratz and Hovy, 2011)6. It provides the dependency structure, POS tags, and other information.
To construct the different tree representations, we replace each node in the tree with its word, lemma, POS tag, dependency label, or supersense (the WordNet lexicographer name of the word’s first sense (Fellbaum, 1998)), and mark the word in questio with a special node. See Figure 3 for a graphical representation. These trees are used in addition to the vectors.
This approach is similar to the ones described in (Moschitti et al., 2006; Qian et al., 2008; Hovy et al., 2012).
2.4 Classification Models A tree kernel is simply a similarity matrix over tree instances. It computes the similarity between two trees T1, T2 based on the number of shared subtrees.
We want to make use of the information en-
coded in the different tree representations during
classification, i.e., a forest of tree kernels. We thus combine the contributions of the individual tree representation kernels via addition. We use kernels over the lemma, POS tag, and supersense tree representations, the combination which performed best on the dev set in terms of accuracy.
We use the SVMlight TK implementation by Moschitti (2006).7 We left most parameters set to default values, but tuned the weight of the contribution of the trees and the cost factor on the dev set. We set the multiplicative constant for the trees to 2.0, and the cost factor for errors on positive examples to 1.7.
6http://www.isi.edu/publications/ licensed-sw/fanseparser/index.html
7http://disi.unitn.it/moschitti/ Tree-Kernel.htm
Figure 5 Dependency trees with lexical, part-of-speech, and WordNet supersense features from Hovy et al. (2013).
599
objects of type liquid in the LAB domain, and for money in the FINANCE domain. From this Mason’s system inferred the domain mapping FINANCE—LAB and the concept mapping money–liquid. Mason used WordNet for acquisition of selectional preference classes and, therefore, the source and target domain categories were represented as clusters of WordNet synsets. The domain-specific corpora were obtained by searching the Web for specific terms of interest. Mason conducted two types of evaluation: (1) against the MML, where he manually mapped his output (WordNet synsets) to concrete concepts described in the MML (13 mappings in total) and then measured the accuracy at 77% (a mapping discovered by CorMet was considered correct if submappings specified in the MML were mostly present with high salience and incorrect submappings were present with relatively low salience); and (2) by compiling a list of mappings at random (assumed to be incorrect) and showing that the system assigned low scores to those. Baumer, Tomlinson, and Richland (2009) reimplemented the method of Mason (2004) in the framework of computational metaphor identification (CMI) procedure, and applied it to two types of corpora: student essays and political blogs. The authors presented some interesting examples of conceptual metaphors the system extracted, which they claim may foster critical thinking in social science. However, they did not carry out any quantitative evaluation.
Li, Zhu, and Wang (2013) proposed a method that performs metaphor identification using an “is-a” knowledge base. The authors automatically created two probabilistic knowledge bases by querying the Web using lexico-syntactic patterns. The first knowledge base contained hypernym–hyponym relations and was acquired using Hearst patterns (Hearst 1992). The second knowledge base contained metaphors in the form 〈target is a source〉 learned using a “*BE/VB like*” pattern. The second database was then filtered by removing the hypernym–hyponym relations present in the first database, as well as symmetric relations, to form a metaphor knowledge base. The authors applied the resulting metaphor knowledge base to perform metaphor recognition and explanation. They experimented with nominal metaphors (e.g., “Juliet is the sun”) and verbal metaphors (e.g., “My car drinks gasoline”). In the case of nominal metaphors, the database was queried directly and the corresponding metaphor was either retrieved or not. In the case of verbal metaphors, where the noun denoting the source concept was not explicitly present in the sentence, it was derived based on the selectional preferences of the verbs. The authors computed selectional preferences of the given verb for the nouns present in the knowledge base, and “explained” the given metaphor by the noun exhibiting the highest selectional association with the metaphorical verb. For example, it outputs an explanation “car is a horse” for the metaphor in “my car drinks gasoline,” since the conceptual metaphor CAR IS A HORSE is present in the knowledge base and horse satisfies the subject preference of drink. The authors evaluated their approach on a manually constructed data set of 200 randomly sampled sentences containing “is-a” constructions and 1,000 sentences containing metaphorical and literal uses of verbs. The annotation was carried out at the sentence level (i.e., complete sentences were annotated as metaphorical or not). The authors report an F-score of 69% on the recognition of “is-a” metaphors and that of 58% on the recognition of the verbal ones. Metaphor explanation performance (i.e., the source–target domain mappings generated for each recognized metaphor) was evaluated separately on 214 sentences extracted from linguistic literature (Lakoff and Johnson 1980) and the top-rank precision of 43% is reported. Intuitively, a purely simile-based approach to metaphor is likely to both undergenerate (a large number of metaphors would never be manifested in similelike constructions) and overgenerate (“A is like B” pattern may describe other relations than metaphor). The key contribution of Li, Zhu, and Wang appears to be the filtering
method they introduce, as well as the selectional preference extension of the knowledge base to identify verbal metaphors.
The method of Shutova and Sun (2013) learns metaphorical associations between concepts from the data in an unsupervised way. They created a network (or a graph) of concepts, using hierarchical graph factorization clustering of nouns, and quantified the strength of association between concepts in this graph. Concrete concepts exhibited well-defined association patterns mainly based on subsumption within one domain, whereas abstract concepts tended to have both within-domain and cross-domain associations: the literal ones and the metaphorical ones. For example, the abstract concept of democracy was literally associated with a more general concept of political system, as well as metaphorically associated with the concept of mechanism. Because we often discuss political systems using the mechanism terminology, a corpus-based distributional learning approach learns that they share features with political systems (from their literal uses), as well as with mechanisms (from their metaphorical uses). The system of Shutova and Sun (2013) automatically discovered such association patterns within the graph and used them to identify metaphorical mappings. The mappings were represented in their system as cross-level, one-directional connections between clusters in the hierarchical graph (e.g., the feeling cluster was strongly associated with fire). Example output for the source concepts of fire and disease is shown in Figure 6. To identify metaphorical expressions representing a given mapping, Shutova and Sun used the features that resulted in strong metaphorical associations between the clusters in question (e.g., “passion flared” for FEELING IS FIRE), as shown in Figure 7. The authors evaluated the quality of metaphorical mappings and metaphorical expressions identified by the system against human judgments, as follows: (1) the human judges were presented with a random sample of system-produced metaphorical mappings between the clusters of nouns, as well as the corresponding metaphorical expressions, and asked to mark the ones they considered valid as correct; (2) the human annotators were presented with a set of source domain concepts and asked to write down all target concepts they associated with a given source, thus creating a gold standard. Shutova and Sun report the precision of 0.69 for metaphorical associations and 0.65 for metaphorical expressions, as evaluated against human judgments, and the recall of 0.61 for metaphorical associations, as evaluated against a human-created gold standard. These results are encouraging in that they show that it is possible to induce information about metaphorical mechanisms from distributional properties of concepts alone, without
the use of hand-coded knowledge. Nevertheless, the fact that clustering techniques are typically applied to a limited set of concepts (i.e., cluster a limited set of nouns) somewhat constrains this approach. For instance, whereas common concepts (that are well represented in the data) can be clustered with a high accuracy, this is not always the case for rare concepts for which feature vectors are sparse. Thus an additional technique is needed to map new, unseen concepts to the concepts present in the graph.
Gandy et al. (2013) presented a system that first discovers metaphorical expressions using concreteness algorithm of Turney et al. (2011) and then assigns the corresponding metaphorical mappings using lexical resources and context clustering. They focused on the three types of metaphor defined by Krishnakumaran and Zhu (2007). Once the metaphorical expressions had been identified, Gandy et al. extracted the nouns that the metaphorical words, or facets, tend to co-occur within a large corpus (e.g., the nominal arguments of open in “open government”). The goal of this process was to form candidate nominal analogies between the target noun in the metaphor and the extracted nouns. For example, the expression “open government” suggests an analogy “government∼ door,” according to the authors. Figure 8 shows how nominal analogies are formed based on
metaphorical usage cannot exist for one sense only. Then, the algorithm verifies that the noun N belongs to at least one high-level semantic category (using the WordStat dictionary of semantic categories based on Wordnet1). If not, the algorithm cannot make a decision and stops. Otherwise, it identifies the n nouns most frequently collocated with A, and chooses the k most concrete nouns (using the abstractness scale of (Turney et al. 2011)). The high-level semantic categories represented in this list by at least i nouns each are selected; if N does not belong to the any of them, the phrase is labeled as metaphorical, otherwise it is labeled as non-metaphorical.
Based on exploratory analysis of a development dataset, separate from our test set, we set n = 1000; k = 100; and i = 16.
Nominal Analogies Once a set of metaphorical expressions SM = {hf, nti} (each a pair of a facet and a target noun) is identified, we seek nominal analogies which relate two specific nouns, a source and a target. For example, if the system finds many different linguistic metaphors which can be interpreted as viewing governments as doors, we may have the nominal analogy “government ⇠ door.”
The basic idea behind the nominal analogy finding algorithm is that, since the metaphorical facets of a target noun are to be understood in reference to the source noun, we must find source/target pairs such that many of the metaphorical facets of the target are associated in literal senses with the source (see Figure 3). The stronger and more numerous these associations, the more likely the nominal analogy is to be real.
It should be noted that the system will not only find frequent associations. Since the association strength of each facet-noun pairing is measured by PMI, not its raw frequency, infrequent pairings can (and do) pop up as significant.
Candidate generation. We first find a set SC of nouns as candidate source terms for nominal analogies. This set is the set of all nouns (in a given corpus) which have strong nonmetaphoric associations with facets (adjectives or verbs) that are used in the linguistic metaphors in SM . We start with the set of all facets used in the linguistic metaphors in SM which have a positive point-wise mutual information (PMI; cf. (Pantel and Lin 2002)) with some target term in the set. We then define the set SC to consist of all other (non-target) nouns (in a given large corpus) associated with each of those facets (in the appropriate syntactic relationships) with a positive PMI. A higher threshold than 0 can also be set, though we haven’t seen that increase precision, and it may reduce recall.
For example, consider finding candidate source terms for the target term “government.” We first find all facets in linguistic metaphors identified by the system that are associated with “government” with a positive PMI. These include such terms as “better”, “big”, “small”, “divided”, “open”,
1See http://provalisresearch.com/.
Figure 3: A schematic partial view of the nominal analogy
“government ⇠ door.” Facets in the middle are associated
with the target noun on the left metaphorically, and with the source noun on the right literally.
“closed”, and “limited.” Each of these facets also are associated with other nouns in non-metaphorical senses. For instance, the terms “open” and “closed” are associated with the words “door”, “table”, “sky”, “arms”, and “house”.
Identification. To identify which pairs of candidate source terms in SC and target terms are likely nominal analogies, we seek a measurement of the similarity of the nonmetaphorical facets of each candidate with the metaphorical facets of each target.
To begin, we define for each facet fi and noun (source or target) ni the pair’s association score aij as its PMI, and the pair’s metaphoricity score, mij where mij = 1 if the pair is judged by the system to be metaphorical and mij = 0 if it is judged to be non-metaphorical. In our current system, all metaphoricity scores are either 1 or 0, but (as described below) we plan to generalize the scheme to allow different levels of confidence in linguistic metaphor classification.
We then define the a metaphoric facet distribution (MFT) for facets given target terms by normalizing the product of association and metaphoricity scores:
PM (fi|nj) = aijmijP ij aijmij
as well as a literal facet distribution (LFT), similarly, as:
PL(fi|nj) = aij(1 mij)P ij aij(1 mij)
As noted above, we seek source term / target term pairs such that the metaphorical facets of the target term are likely to be literal facets of the source term and vice versa. A natural measure of this tendency for a candidate source noun ns and candidate target noun nt is the Jensen-Shannon divergence between the LFT of ns and the MFT of nt,
DJS (PL(·|ns) k PM (·|nt)) The larger the J-S divergence, the less likely the pair is to be a good nominal analogy.
Figure 8 Nominal analogy inducti n from Gandy et al. (2013).
602
a collection of metaphorical expressions. The individual (related) nominal analogies were then clustered together to identify conceptual metaphors, as shown in Figure 9. The authors evaluated their system by annotating metaphorical expressions for five target concepts (government, governance, god, father, and mother) in selected sentences from the Reuters corpus (Lewis et al. 2004). They report very encouraging results: Precision (P) = 0.76, Recall (R) = 0.82 for verb metaphors; P = 0.54, R = 0.43 for adjectival metaphors; and P = 0.84, R = 0.97 for copula constructions. The authors also evaluated the quality of conceptual metaphors produced by the system against human judgments and attained a precision of 0.65. However, the scope of the experiment is only limited to the given five concepts and it is not clear how well the method would generalize beyond these. Although the approach of Gandy et al. (2013) seems very promising, a comprehensive evaluation on open-domain corpus data is still necessary to prove its viability. 5 metaphor interpretation systems :In one of the first approaches to metaphor interpretation, Martin (1990) presented a Metaphor Interpretation, Denotation, and Acquisition System (MIDAS), which explained linguistic metaphors through finding the corresponding conceptual metaphor. The method is based on the idea of hierarchical organization of conventional metaphors, namely, that more specific conventional metaphors descend from the general ones. Given an example of a metaphorical expression, MIDAS searched its database for a corresponding metaphor that would explain the anomaly. If it did not find any, it abstracted from the example to more general concepts and repeated the search. If it found a suitable general metaphor, it created a mapping for its descendant, a more specific metaphor, based on the given example. This was also how novel metaphors were acquired. MIDAS was integrated with the Unix Consultant (UC), the system that answers users’ questions about Unix. The UC first tried to find a literal answer to the question. Failing to do so, it called MIDAS, which detected metaphorical expressions via selectional preference violation and searched its database for a metaphor explaining the anomaly in the question.
Another branch of early work on metaphor interpretation relied on performing inferences about entities and events in the source and target domains. The most prominent approaches include the KARMA system (Narayanan 1997, 1999; Feldman and
Narayanan 2004) and the ATT-Meta project (Barnden and Lee 2002; Agerri et al. 2007). Within both systems the authors developed a metaphor-based reasoning framework in accordance with the theory of conceptual metaphor. The reasoning process relied on manually constructed knowledge about the world and operated mainly in the source domain. The results were then projected onto the target domain using the conceptual mapping representation. The ATT-Meta project concerned metaphorical and metonymic description of mental states and reasoning about mental states using first order logic. Their system, however, did not take natural language sentences as input, but logical expressions that are representations of small discourse fragments. KARMA in turn dealt with a broad range of abstract actions and events and took parsed text as input.
Since then the field moved towards acquiring the knowledge necessary for metaphor interpretation automatically (and at a larger scale) from lexical resources, corpora, and the Web. Veale and Hao (2008) derived a “fluid knowledge representation for metaphor interpretation and generation,” called Talking Points. Talking Points are a set of characteristics of concepts belonging to source and target domains and related facts about the world which the authors acquired automatically from WordNet and from the Web. Talking Points were then organized in Slipnet, a framework that allowed for a number of insertions, deletions, and substitutions in definitions of such characteristics in order to establish a connection between the target and the source concepts. This work built on the idea of slippage in knowledge representation for understanding analogies in abstract domains (Hofstadter and Mitchell 1994; Hofstadter 1995). Example (6) demonstrates how slippage operates to explain the metaphor Make-up is a Western burqa.
(6) Make-up => ≡ typically worn by women ≈ expected to be worn by women ≈must be worn by women ≈must be worn by Muslim women
Burqa <=
By doing insertions and substitutions the system arrived from the definition typically worn by women to that of must be worn by Muslim women, and thus established a link between the concepts of make-up and burqa. Veale and Hao (2008), however, did not evaluate to what extent their knowledge base of Talking Points and the associated reasoning framework are useful to interpret metaphorical expressions occurring in text.
Shutova (2010) defined metaphor interpretation as a paraphrasing task and presented a method for deriving literal paraphrases for metaphorical expressions from the BNC. For example, for the metaphors in “All of this stirred an unfathomable excitement in her” or “a carelessly leaked report,” their system produced interpretations All of this provoked an unfathomable excitement in her and a carelessly disclosed report, respectively. They first applied a probabilistic model to rank all possible paraphrases for the metaphorical expressions, given the context; and then used automatically induced selectional preferences to discriminate between figurative and literal paraphrases. The selectional preference distribution was defined in terms of selectional association measures introduced by Resnik (1993) over the noun classes automatically produced by Sun and Korhonen (2009). Shutova (2010) tested her system only on metaphors expressed by a verb and reports an accuracy of 0.81, as evaluated on top-ranked paraphrases produced by the system. However, she used WordNet for supervision, which limits the number and range of paraphrases that can be identified by her method. Shutova, Van de Cruys,
and Korhonen (2012) and Bollegala and Shutova (2013) expanded on this work, addressing the metaphor paraphrasing task in an unsupervised setting and extending the coverage. The method of Shutova, Van de Cruys, and Korhonen (2012) first computed candidate paraphrases according to the context in which the metaphor appeared, using a vector space model. It then used a selectional preference model to measure the degree of literalness of the paraphrases. The authors evaluated their method on the metaphor paraphrasing data set of Shutova (2010) and reported a top-rank precision of 0.52. Bollegala and Shutova (2013) used a similar experimental set-up, however, their method extracted a set of candidate paraphrases from the Web using lexico-syntactic patterns as queries and ranked them based on search engine hits, attaining a precision of 0.42.
Shutova, Teufel, and Korhonen (2013) combined metaphor identification (Shutova, Sun, and Korhonen 2010) and interpretation (Shutova 2010) to perform text-to-text metaphor processing. The resulting system could take arbitrary text as input, parse it using a syntactic parser, identify metaphorical expressions in it, retrieve their literal paraphrases, and output a new version of the text in which metaphors were interpreted. The motivation behind such a set-up was that it allowed for a relatively straightforward integration with external NLP applications. To evaluate the system, the authors extracted a random sample of 200 metaphorical expressions the system identified in the BNC and applied the paraphrasing method to them. They evaluated the accuracy of metaphor identification and interpretation when performed simultaneously, as well as the system’s applicability. The applicability was defined as the proportion of cases where the paraphrase was literal and the meaning of the phrase was retained, indicating whether this type of system paraphrasing would result in an error when hypothetically integrated with an external NLP application. In 54% of cases, the system both identified and interpreted the metaphor correctly, which is a promising result. In a further 13% of cases, the system produced a correct, literal paraphrase for a literal expression erroneously identified as a metaphor, leading to the overall applicability of integrated metaphor processing at 67%. Although the system is easy to integrate with external NLP applications that could benefit from metaphor resolution, it should be noted that some information conveyed by the metaphor is inevitably lost during literal paraphrasing. Metaphor paraphrasing as an approach thus rests on a crucial assumption that the benefit of correct metaphor understanding would outweigh the loss of additional connotations and rhetorical elements. This assumption is yet to be verified through an integration of this technology into real-world NLP, however.
Shutova (2013) presented a computational method that identified metaphorical expressions in unrestricted text by means of their interpretation. She again treated metaphor interpretation as paraphrasing and introduced the concept of symmetric reverse paraphrasing as a criterion for metaphor identification. The hypothesis behind the method is that literal paraphrases of literally used words should yield the original phrase when paraphrased in reverse. For example, when the expression clean the house is paraphrased as tidy the house, the reverse paraphrasing of tidy would generate clean as one of possible paraphrases. Shutova’s expectation was that such symmetry in paraphrasing is indicative of literal use. The metaphorically used words are unlikely to exhibit this symmetry property when paraphrased in reverse. For example, the literal paraphrasing of the verb stir in “stir excitement” would yield “provoke excitement,” but the reverse paraphrasing of provoke would not retrieve stir, indicating the non-literal use of stir. Shutova experimentally verified this hypothesis in a setting involving singleword metaphors expressed by a verb in verb–subject and verb–direct object relations. She applied the selectional preference-based metaphor paraphrasing method (Shutova 2010) to retrieve literal paraphrases of all input verbs and extended the method to
perform metaphor identification by reverse paraphrasing. She evaluated the performance of the system on verb–subject and verb–object relations using the manually annotated metaphor corpus of Shutova and Teufel (2010), reporting a precision of 0.68 and a recall of 0.66. The system outperformed a baseline using selectional preference violation as an indicator of metaphor, that only attained a precision of 0.17 and a recall of 0.55. Some examples of metaphorical expressions identified by the system and their literal paraphrases are shown in Figure 10. 6 investigated techniques and lessons learned :The community has investigated a wide range of techniques and features for metaphor identification and interpretation in a variety of experimental settings. The majority of identification systems focus on the linguistic level, identifying either linguistic metaphor or non-literal language more generally, with a few identifying conceptual metaphor. Table 4 presents a summary of the tasks addressed. Some systems annotated metaphorical expressions at the word level, whereas others opted for the relation level or carried out sentence-level annotation. Individual approaches frequently limited the scope of their experiments to metaphors expressed by a particular part of speech and syntactic construction, as shown in Table 5. The majority of the systems focused on metaphorically used verbs or adjectives, with a few also considering nouns (in modifier or copula constructions) and multiword metaphors. The systems that identified conceptual metaphor also exhibit some variation in the representations they used. Source and target domains were represented as WordNet synsets (Mason 2004); individual nouns (Li, Zhu, and Wang 2013); or clusters of nouns (Shutova and Sun 2013; Gandy et al. 2013). Recent work on metaphor interpretation unfolded along two main axes: metaphor explanation (i.e., identifying the properties of concepts that the metaphor
highlights and the comparisons it involves [Veale and Hao 2008]) and metaphor paraphrasing (i.e., identifying a literal [or more conventional] paraphrase of the metaphorical expression [Shutova 2010; Bollegala and Shutova 2013]).
Identification and interpretation approaches investigated a range of properties of metaphor and implemented them in a variety of system components. The most
Ta bl
e 6
Id en
ti fic
at io
n sy
st em
s: Sy
st em
fe at
ur es
an d
te ch
ni qu
es .
Sy st
em /
fe at
ur es
Se l. pr ef
. V
io la
ti on
C on
cr et
en es
s To
pi ca l st ru ct ur e
C lu
st er
in g
M L
C la
ss ifi
er Le
x re
so ur
ce W
eb se
ar ch
su pe
rv is
ed w
ea kl ysu pe rv is ed
un su
pe rv
is ed
M as
on (2
00 4)
4 –
– –
4 –
4 4
4 – – Bi rk e an d Sa rk ar (2 00 6) – – – – 4 – – – – 4 – G ed ig ia n et al .( 20 06 ) – – – – – 4 4 – 4 – – K ri sh na ku m ar an an d Z hu 20 07 ) – 4 – – – – 4 – 4 – – Sh ut ov a et al .( 20 10 ) 4 – – – 4 – – – – 4 – Li an d Sp or le de r (2 00 9, 20 10 ) – 4 – – – 4 – – 4 – 4 Tu rn ey et al .( 20 11 ) – – 4 – – 4 – – 4 – – N eu m an et al .( 20 13 a, b) 4 4 4 – – – – – 4 – – D un n (2 01 3) – – – – – 4 – – 4 – – Ts ve tk ov et al .( 20 13 ,2 01 4) – – 4 – – 4 4 – 4 – – M oh le r et al .( 20 13 ) – – – 4 4 4 4 – 4 – – H ei nt z et al .( 20 13 ) – – – 4 – – – – – 4 – H ov y et al .( 20 13 ) – 4 – – – 4 4 – 4 – – W ilk s et al .( 20 13 ) 4 4 – – – – 4 – 4 – – St rz al ko w sk ie ta l. (2 01 3) – – 4 4 4 – 4 – 4 – – Sh ut ov a an d Su n (2 01 3) 4 – – – 4 – – – – – 4 G an dy et al .( 20 13 ) – – 4 – 4 – 4 – 4 – – Li et al .( 20 13 ) 4 – – – – – – 4 – 4 – Sh ut ov a (2 01 3) 4 – – – – – 4 – 4 – –
prominent ones include selectional preferences (Martin 1990; Fass 1991; Mason 2004; Krishnakumaran and Zhu 2007; Li and Sporleder 2009, 2010; Shutova 2010; Shutova, Sun, and Korhonen 2010; Hovy et al. 2013; Li, Zhu, and Wang 2013; Wilks et al. 2013); semantic properties of concepts, such as imageability and concreteness (Turney et al. 2011; Gandy et al. 2013; Neuman et al. 2013; Strzalkowski et al. 2013); and topical structure of text (Heintz et al. 2013; Strzalkowski et al. 2013). The common methods used include supervised classification (Gedigian et al. 2006; Dunn 2013a; Hovy et al. 2013; Mohler et al. 2013; Tsvetkov, Mukomel, and Gershman 2013); clustering (Gandy et al. 2013; Shutova and Sun 2013; Shutova, Sun, and Korhonen 2010; Strzalkowski et al. 2013); vector space models (Shutova, Van de Cruys, and Korhonen 2012); the use of lexical resources and ontologies (Mason 2004; Krishnakumaran and Zhu 2007; Dunn 2013b; Gandy et al. 2013; Hovy et al. 2013; Mohler et al. 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013; Wilks et al. 2013); and Web search (Veale and Hao 2008; Bollegala and Shutova 2013; Li, Zhu, and Wang 2013). A summary of techniques investigated by the community is presented in Table 6. In what follows we will discuss the main trends in metaphor processing research and the usefulness of individual types of techniques. Selectional preferences have long established themselves as one of the central components in metaphor-processing research. Wilks’ (1978) selectional preference violation view of metaphor has been highly influential, with numerous approaches to metaphor identification implementing it directly or indirectly (Fass 1991; Martin 1990; Wilks et al. 2013). Other approaches modified this view and treated metaphor as a violation of semantic norm construed more broadly—for example, searching for expressions with low bi-gram probabilities (Krishnakumaran and Zhu 2007), identifying units that break sentence cohesion (Li and Sporleder 2009, 2010), or detecting unusual patterns in words’ compositional behavior (Hovy et al. 2013).
Generally speaking, selectional preference violations (or other semantic violations mentioned above) are a property of surface realization of metaphor rather than its underlying conceptual mechanisms. One needs to bear this in mind when using this as a heuristic. On one hand, such violations are indicative of any kind of non-literalness (i.e., not only metaphor, but also, for instance, metonymy) or anomaly in language and the approach is likely to overgenerate. On the other hand, in the case of most conventional metaphors that are highly frequent, no statistically significant violation can be detected in the data, and the approach would bypass many such metaphors. Shutova (2013) conducted a data-driven study, where verb preferences were automatically acquired from the data and all the nominal arguments below a certain selectional association threshold were considered to represent a violation and were tagged as metaphorical. Such a technique attained a precision of 0.17 and a recall of 0.55, suggesting that the selectional preference violation hypothesis does not port well beyond handcrafted descriptions to large-scale, data-driven techniques.
In contrast, other, “non-violation” applications of selectional preferences have been fruitful in metaphor modeling. Mason (2004) automatically acquired domain-specific selectional preferences of verbs, and then, by mapping their common nominal arguments in different domains, arrived at the corresponding metaphorical mappings. Shutova (2010) presented a modification of Wilks’ view, treating a strong selectional preference fit as a likely indicator of literalness or conventionality. In her metaphor paraphrasing system, Shutova ranks candidate paraphrases based on how well the
context fits their preferences, thus determining their literalness. In their metaphor identification system, Shutova, Sun, and Korhonen (2010) filtered out verbs that have weak selectional preferences, that is, that are equally associated with many argument classes (e.g., choose or remember), as having a lower metaphorical potential. Li, Zhu, and Wang (2013) used selectional preferences to assign the corresponding conceptual metaphor to metaphorical expressions. Although the idea of violation (i.e., treating metaphor as merely a context outlier) is controversial and should be applied with care, selectional preferences themselves are an important source of semantic information about the properties of concepts, which can be successfully exploited in metaphor processing in a variety of ways. Two approaches (Heintz et al. 2013; Strzalkowski et al. 2013) focused on modeling topical structure of text to identify metaphor. The main hypothesis behind these methods is that metaphorical language (coming from a different domain) would represent atypical vocabulary within the topical structure of the text. This intuition is somewhat similar to the idea of semantic norm violation as an indicator of metaphor, although it is different in two crucial ways: (1) topical structure-based approaches explicitly model the interaction of vocabulary from two different domains (i.e., the source and the target); and (2) these approaches take into account domain interactions over extended discourse fragments, rather than individual expressions, thus utilizing information from wider context. Exploiting the wider topical structure of text is a promising avenue for metaphor processing. However, one needs to keep in mind that distributional similaritybased methods risk assigning frequent metaphors to target domains (as is the case for other semantic violation-based methods). For instance, cut may appear more frequently within the domain of economics and finance, rather than its original source domain. The choice of data for training such a model thus becomes crucial, and an appropriately balanced data set is needed. Investigating the topical structure of text is also an important step towards modeling extended metaphor, which interweaves the narrative in complex, but systematic ways. Turney et al. (2011) introduced the idea of measuring concreteness of concepts to predict metaphorical use. The intuition behind their approach is that metaphor is commonly used to describe abstract concepts in terms of more concrete or physical experiences. Thus, Turney and colleagues expect that there would be some discrepancy in the level of concreteness of source and target terms in the metaphor. Neuman et al. (2013) and Gandy et al. (2013) followed in Turney’s steps, reporting promising results. Tsvetkov, Mukomel, and Gershman (2013) took a different route, and used concreteness as one of the features to train a classifier. Strzalkowski et al. (2013) experimented with the imageability feature (that indicates how easy it is to visualize a concept) and demonstrated its relevance to metaphor identification.
Based on the results of these experiments, concreteness is likely to be a practically useful feature for metaphor processing. However, it should be noted that Turney’s hypothesis (that target words tend to be abstract and source words tend to be concrete) explains only a fraction of metaphors and does not always hold. For example, one can use concrete–concrete metaphors (e.g., “broken heart”), abstract–abstract metaphors (“diagnose corruption”), and even abstract–concrete metaphors (“invent a soup”).
However, it may be the case that within the concrete–abstract class of metaphor, the method operates with reasonable performance. Thus concreteness may become a useful feature of a more complex system that takes multiple factors into account, but is unlikely to be a reliable indicator of metaphor on its own. A number of approaches trained classifiers on manually annotated data to recognize metaphor. The key question that supervised classification poses is what features are indicative of metaphor and how can one abstract from individual expressions to its high-level mechanisms? The community has experimented with a number of features, including lexical and syntactic information; higher-level features such as semantic roles, WordNet supersenses, named-entity types, and domain types extracted from ontologies; and semantic properties of concepts, such as animateness and concreteness. Gedigian et al. (2006) classified verb uses as literal or metaphorical, using the verbs’ nominal arguments and their semantic roles (as annotated in PropBank) as features. They reported unusually high performance scores, although the narrow focus on specific lexical items makes it possible for the system to learn a model for individual words rather than performing generalization. In contrast, Dunn (2013a, 2013b) experimented with a wide range of metaphorical expressions from the VU Amsterdam Metaphor corpus, using high-level properties of concepts extracted an ontology, such as domain type and event status. Tsvetkov, Mukomel, and Gershman (2013) and Tsvetkov et al. (2014) used coarse semantic features, such as concreteness, animateness, named-entity types, and WordNet supersenses. What is particularly interesting about this work is that the authors have shown that the model learned with such coarse semantic features is portable across languages, thus suggesting that the chosen features successfully capture some of the properties of metaphor (even if the shallow ones). The work of Hovy et al. (2013) is notable as they focused on compositional rather than categorical features. They trained an SVM with dependency-tree kernels to capture compositional information using lexical, part-of-speech tags, and WordNet supersense representations of sentence trees, achieving successful results. The system of Mohler et al. (2013) aimed at modeling conceptual information in the form of semantic signatures of domains and metaphors. Such rich semantic information is likely to be a successful feature in metaphor recognition: however, Mohler and colleagues experimented within a limited domain and it is not clear how scalable such features would be.
To reliably capture the patterns of the use of metaphor in the data at a large scale, one needs to address conceptual properties of metaphor, along with the surface ones. Thus the models making generalizations at the level of metaphorical mappings and coarse-grained classes of concepts, in essence representing different domains, are likely to yield the optimal framework for the task. However, this hypothesis is yet to be experimentally verified. Clustering techniques were used in numerous approaches, predominantly to identify concepts similar or related to each other. Mason (2004) performed WordNet sense clustering to obtain selectional preference classes, whereas Mohler et al. (2013) used it to determine similarity between concepts and to link them in semantic signatures. Strzalkowski et al. (2013) and Gandy et al. (2013) clustered metaphorically used terms to
form potential source domains. Birke and Sarkar (2006) clustered sentences containing metaphorical and literal uses of verbs.
Another line of research focused on the use of clustering methods to investigate how metaphor partitions the linguistic feature space. Shutova, Sun, and Korhonen (2010) pointed out that the metaphorical uses of words constitute a large portion of the co-occurrence features extracted for abstract concepts from the data. For example, the feature vector for politics would contain GAME or MECHANISM terms among the frequent features. As a result, distributional clustering of abstract nouns with such features identifies groups of diverse concepts metaphorically associated with the same source domain (or sets of source domains). Shutova, Sun, and Korhonen exploit this property of co-occurrence vectors to identify new metaphorical mappings starting from a set of examples. The work of Shutova and Sun (2013) is based on the same observation. Through the use of hierarchical clustering techniques they derive a network of concepts in which metaphorical associations are exhibited at different levels of generality. Peters and Peters (2000) and Wilks et al. (2013) detected metaphor directly in lexical resources. Peters and Peters mine WordNet for examples of systematic polysemy, which allows them to capture metonymic and metaphorical relations. Their system searches for nodes that are relatively high in the WordNet hierarchy (i.e., are relatively general) and that share a set of common word forms among their descendants. Peters and Peters found that such nodes often happen to be in a metonymic (e.g., publisher – publication) or a metaphorical (e.g. theory – supporting structure) relation. Wilks et al. (2013) used WordNet glosses to learn selectional preferences of verbs, which were then used to annotate senses as literal or metaphorical, based on the selectional preference–violation hypothesis. Krishnakumaran and Zhu (2007) use hyponymy relation in WordNet to detect semantic violations. Shutova (2010, 2013) also relied on the hierarchical structure of WordNet, but to identify concepts that share common features (defined as sharing a common hypernym within three levels of the hierarchy).
WordNet synsets were also used to form selectional preference classes in SP-based methods (Mason 2004) or to detect semantically related concepts (Mohler et al. 2013; Strzalkowski et al. 2013; Gandy et al. 2013). Other researchers used WordNet to identify high-level properties of concepts, most notably WordNet supersenses, that served as features for classification (Tsvetkov, Mukomel, and Gershman 2013; Hovy et al. 2013). Because metaphor is a knowledge-intensive phenomenon, multiple approaches attempted to acquire the knowledge necessary for its identification and interpretation from the Web (Veale and Hao 2008; Bollegala and Shutova 2013; Li, Zhu, and Wang 2013). Web search engines provide a flexible tool for retrieving information that matches specific lexico-syntactic patterns (used as queries) and quantifying co-occurrence. Veale and Hao (2008) query the Web to harvest properties of concepts and cultural stereotypes, such as has magical skill for Wizard or has brave spirit for Lion, which are then used to perform metaphor interpretation through property comparison and substitution. Bollegala and Shutova (2013) use the Web to extract co-occurrence information for verbs and nouns, which allows them to generate a set of candidate paraphrases for metaphorical verbs. Li, Zhu, and Wang (2013) query the Web with Hearst patterns to acquire a large knowledge base of hyponymy relations; and simile patterns
(* is like *) to acquire a set of potential conceptual metaphors. Along with the flexibility and convenience of information retrieval tools, these three approaches also boast of wide coverage that the use of the Web allows them to achieve. The knowledge contained on the Web is not merely vast, but it is also constantly updated, which allows the system to stay on par with the current events and trends. Because metaphor is a productive and dynamic phenomenon (new metaphors arise as new events take place), such scalability and ongoing expansion of the Web make it an attractive corpus for metaphor research. 7 system evaluation :System evaluation methodologies continue to be debated in many areas of computational semantics. Identifying a comprehensive and fair evaluation strategy for the task in mind is crucial for the development of fully functional NLP systems. In that light, a number of shared tasks have been proposed over the years at Workshops on Semantic Evaluation (SemEval) that enabled performance comparison across systems and methods. Such tasks as sentiment analysis, word similarity, word sense induction and disambiguation, coreference resolution, lexical substitution, and many others are commonly addressed at SemEval and have a number of benchmark data sets created for them (Agirre and Soroa 2007; McCarthy and Navigli 2007; Lefever and Hoste 2010; Manandhar et al. 2010; Mihalcea, Sinha, and McCarthy 2010; Recasens et al. 2010; Nakov et al. 2013), against which the systems are evaluated and compared. Computational work on metaphor, on the contrary, is considerably more fragmented than similar research efforts in other areas of NLP. With the lack of an established data set, the community has utilized a variety of evaluation strategies, including the use of annotated corpora, human judgments of system output, evaluation against the MML, and annotation of individual selected examples (usually phrases or sentences) via Amazon Mechanical Turk. With a few exceptions, the majority of approaches created their own test sets, making the results not directly comparable.
The most desirable type of evaluation is that conducted against an annotated full-text corpus, namely, naturally occurring, continuous text manually annotated for metaphor. Ideally, such a corpus should be open-domain and representative of a range of genres, making the results indicative of the likely performance of the system on arbitrary text. Another benefit of this type of evaluation is that it allows one to assess both the precision and the recall of the system. However, only two of the presented approaches (Dunn 2013b; Shutova 2013) conducted this type of evaluation, as shown in Table 7. More typically, approaches were instead evaluated on a random sample of system output against human judgments (Mason 2004; Shutova, Sun, and Korhonen 2010; Heintz et al. 2013; Shutova and Sun 2013; Strzalkowski et al. 2013). Although this type of evaluation allows one to measure the precision of the system on a random sample, it does not provide any information about the possible recall. Shutova, Sun, and Korhonen (2010) and Shutova and Sun (2013) applied their methods to a generaldomain corpus (the BNC), from which a random sample of metaphorical expressions annotated by the system was then extracted for evaluation. In contrast, Heintz et al. (2013) and Strzalkowski et al. (2013) collected their data with a focus on a limited domain. The experiments of Mason (2004) and Shutova and Sun (2013) were concerned with conceptual metaphor, and a random sample of the metaphorical mappings identified by the systems was extracted and evaluated against human judgments in terms of precision. However, the latter two approaches also measured recall, by manually compiling a gold-standard of metaphorical mappings for the concepts of interest (Shutova and Sun 2013) or against the MML (Mason 2004).
The majority of approaches (see Table 7) did not apply their systems to continuous text, but rather to a set of pre-selected examples (phrases, sentences, or occasionally paragraphs) in isolation from wider context. Such examples were annotated as metaphorical or literal by independent expert annotators or via Amazon Mechanical Turk. The benefit of this set-up (as opposed to the evaluation on a random sample of system output) is that it allows one to measure both precision and recall. However, the method used for selection of individual examples may introduce a bias into the evaluation and provide an unfair advantage to the system, unless the test sample was selected randomly from arbitrary text. In other words, this type of evaluation is likely to be less objective than the evaluation on continuous corpus text or a random sample.
A number of approaches (Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Gandy et al. 2013; Heintz et al. 2013; Mohler et al. 2013; Neuman et al. 2013; Strzalkowski et al. 2013; Wilks et al. 2013) conducted their experiments within a limited domain. Despite allowing for an in-depth investigation of domain-specific patterns of metaphor use, such evaluations are problematic as they provide no indication of the scalability of the method beyond the studied domain to real-world data. This criticism also applies to the evaluations against MML, as the list is limited in domain coverage and the type of metaphors it provides.
Two data sets stand out as having been repeatedly adopted for metaphor research, enabling direct system comparison. These include the TroFi data set of Birke and Sarkar (2006) and the metaphor paraphrasing data set of Shutova (2010). The TroFi dataset consists of 25 verbs and example sentences, containing their metaphorical and literal use. It was adopted by Turney et al. (2011) and Tsvetkov, Mukomel, and Gershman (2013) in their metaphor identification experiments. The paraphrasing data set and gold standard of Shutova (2010) consists of 52 metaphorically used verbs and their humanderived literal (or more conventional) paraphrases in the given context. The data set
has been used in multiple metaphor interpretation experiments (Shutova 2010; Shutova, Van de Cruys, and Korhonen 2012; Bollegala and Shutova 2013).
The evaluations of metaphor identification tend to be conducted in terms of precision and recall, and occasionally, accuracy. Table 8 presents a summary of results of metaphor identification experiments, classified by domain coverage. The most successful systems attain an F-score in the range of 70–78%, with the highest precision reported for the methods of Tsvetkov, Mukomel, and Gershman (2013), Shutova, Sun, and Korhonen (2010), and Gandy et al. (2013); and the highest recall for the methods of Mohler et al. (2013), Wilks et al. (2013), Gandy et al. (2013) (limited domain) and Tsvetkov, Mukomel, and Gershman (2013) and Hovy et al. (2013) (open domain). However, because the methods were evaluated on data sets of different size and balance of categories, as well as created with different criteria in mind, this comparison is only approximate. A comprehensive evaluation on the same large data set is needed to determine the best performing techniques. For example, the accuracy of 95% reported by Gedigian et al. (2006) was measured on a data set dominated by metaphorical expressions (all metaphor baseline achieves 92%). This result cannot be directly compared to that of a system evaluated on a test set with a balance of metaphorical and literal instances.
One of the key difficulties metaphor processing research is facing today is that of a lack of annotated data. The data sets used in the experiments are typically too
small to obtain a generalizable result. Another issue is that many data sets created to evaluate individual methods were designed for a specific task, focusing on a particular type of metaphor and using an annotation scheme of their own. This makes the results not directly comparable and the overall performance landscape difficult to interpret. This situation calls for the consolidation of the insights gained from the current experiments into a single task definition and the creation of a large data set with this task in mind. Ideally, the systems should be evaluated on an expertannotated corpus, containing continuous, general-domain text. In order to be indicative of the likely performance of the system on real-world data, the corpus needs to have a comprehensive coverage of registers and genres. Finally, a large corpus annotated for metaphor would enable reliable evaluation both in terms of precision and recall.
The work of Dunn (2013b) was exemplary in that he evaluated and compared four different methods on the same data set. Dunn used the VU Amsterdam Metaphor Corpus (Steen et al. 2010), which is currently the largest metaphor corpus in existence, containing approximately 200,000 words. However, because Steen and colleagues were interested in historical aspects of metaphor along with its use in modern language, the VU metaphor corpus contains a large proportion of lexicalized metaphors. Their meanings are ingrained in everyday use and can be interpreted via established techniques (e.g., word sense disambiguation), and their metaphorical nature may or may not be of interest to wider NLP. Whether metaphor processing systems should address highly conventional and dead metaphors depends on the task and application in mind, and corpus annotation should reflect this task definition in a consistent way. As the field moves forward, it would also be desirable to conduct extrinsic evaluations of the metaphor processing systems, in order to determine their usefulness for external NLP applications. One such experiment has already been carried out by Agerri (2008), who has demonstrated that metaphor interpretation plays an important role in textual entailment resolution. 8 conclusion :Metaphor makes our thoughts move vivid and enriches our communication with novel imagery, but most importantly it plays a fundamental structural role in our cognition, helping us organize and project knowledge. As a result, its manifestations are pervasive in language and reasoning, making its computational processing an imperative task within NLP and intelligent systems engineering at large. Despite involving complex comparisons and information transfers, metaphor is a well-structured and systematic phenomenon, highly suitable for computational modeling. Focusing primarily on linguistic metaphor, the community has investigated a range of its aspects, implemented in a variety of system features. Among the most successful features are concreteness, distributional behavior of source and target domain vocabulary, selectional preferences, textual coherence, and topical properties of source and target words. The field has evolved from the widespread use of hand-coded knowledge to mainly datadriven research. Balanced corpora, the Web, Wikipedia and, sometimes, domain-specific corpora have become the primary source of knowledge for metaphor processing. The community has investigated supervised learning, clustering, topic modeling, and pattern-based search to acquire lexical, relational, and domain knowledge from these corpora. Yielding promising results, this was a significant advance in computational modeling of metaphor, allowing for the application of the systems to real-world data.
These experiments also provided new insights on the behavior of metaphor across domains, genres, and types of discourse.
The research on mining metaphorical associations from the data sheds light on how metaphors structure our conceptual system, as well as how specific conceptual metaphors are realized in language, revealing new information about their cognitive processing. Large-scale, automatic identification of conceptual metaphor thus provides a bridge between computational and cognitive research in this area, and has a wider scientific relevance beyond NLP. An interdisciplinary approach, leveraging knowledge from linguistics, cognitive science, psychology, neuroscience, and computer science, would be well-positioned to advance our understanding of metaphorical mechanisms and take the performance of metaphor processing systems to the next level. Two areas that are particularly likely to benefit from an interdisciplinary approach are metaphorical inference and extended metaphor, which have so far escaped attention in NLP. Recent advances in processing linguistic and conceptual metaphor, however, bring us a step closer to understanding and modeling these phenomena.
Despite the promising experimental results reported by the community, little attention has yet been given to real-world applications of metaphor processing. Possible applications include other semantic tasks within NLP and data mining, as well as social science and educational applications. Within NLP, most applications that need to access semantic knowledge would benefit from robust and accurate metaphor resolution. These include, for instance, machine translation, sentiment analysis, or text classification. Because the metaphors we use are known to be indicative of our underlying viewpoints, metaphor processing is likely to be fruitful in determining political affiliation from text or pinning down cross-cultural and crosspopulation differences, and thus become a useful tool in data mining. In social science, metaphor is extensively studied as a way to frame cultural and moral models, and to predict social choice (Landau, Sullivan, and Greenberg 2009; Thibodeau and Boroditsky 2011; Lakoff and Wehling 2012). Metaphor is also widely viewed as a creative tool. Its knowledge projection mechanisms help us to grasp new concepts and generate innovative ideas. This opens many avenues for the creation of computational tools that foster creativity (Veale 2011, 2014) and support assessment in education (Burstein et al. 2013).
The design of the metaphor processing task should thus be informed by the possible applications. The application in mind may place particular requirements on the types of metaphor the system needs to address and the output representations it is expected to produce. Whereas an NLP application, such as machine translation, would be primarily concerned with linguistic metaphors and, possibly, the more creative instances thereof, a data mining application, aiming to detect a set of trends, may find the identification of conceptual metaphors prominent in the data more informative. The formulation of the task and experimental design, in turn, predetermine how system performance is best evaluated. So far, the lack of a common task definition and a shared data set have hampered our progress as a community in this area. This calls for a unification of the task definition and a large-scale annotation effort that would provide a data set for metaphor system evaluation, built with the insights gained from the present studies. The main purpose of this paper is to provide a platform for debate that would assist us in formulating the overall goals of metaphor processing and devising an optimal experimental strategy, enabling us as a community to make significant progress in this important and fascinating area. ness–concreteness algorithm. However, the evaluation was done on only five [{""affiliations"": [], ""name"": ""Ekaterina Shutova""}] SP:dfd9f1276e2968c35fcb8cfca252fc00ef19d9e7 [{""authors"": [""Agerri"", ""Rodrigo.""], ""title"": ""Metaphor in textual entailment"", ""venue"": ""Proceedings of COLING 2008, pages 3\u20136, Manchester, UK."", ""year"": 2008}, {""authors"": [""Agerri"", ""Rodrigo"", ""John Barnden"", ""Mark Lee"", ""Alan Wallington.""], ""title"": ""Metaphor, inference and domain-independent mappings"", ""venue"": ""Proceedings of RANLP-2007, pages 17\u201323, Borovets."", ""year"": 2007}, {""authors"": [""Agirre"", ""Eneko"", ""Aitor Soroa.""], ""title"": ""Semeval-2007 task 02: Evaluating word sense induction and discrimination systems"", ""venue"": ""Proceedings of the Fourth International Workshop on Semantic"", ""year"": 2007}, {""authors"": [""Allen"", ""James F."", ""Mary Swift"", ""Will de Beaumont.""], ""title"": ""Deep semantic analysis of text"", ""venue"": ""Proceedings of the 2008 Conference on Semantics in Text Processing, STEP \u201908, pages 343\u2013354, Venice."", ""year"": 2008}, {""authors"": [""Badryzlova"", ""Yulia"", ""Natalia Shekhtman"", ""Yekaterina Isaeva"", ""Ruslan Kerimov.""], ""title"": ""Annotating a Russian corpus of conceptual metaphor: A bottom\u2013up approach"", ""venue"": ""Proceedings of the First"", ""year"": 2013}, {""authors"": [""Barnden"", ""John"", ""Mark Lee.""], ""title"": ""An artificial intelligence approach to metaphor understanding"", ""venue"": ""Theoria et Historia Scientiarum, 6(1):399\u2013412."", ""year"": 2002}, {""authors"": [""Baumer"", ""Eric"", ""Bill Tomlinson"", ""Lindsey Richland.""], ""title"": ""Computational metaphor identification: A method for identifying conceptual metaphors in written text"", ""venue"": ""Proceedings of Analogy \u201909, pages 20\u201329,"", ""year"": 2009}, {""authors"": [""Beigman Klebanov"", ""Beata"", ""Eyal Beigman.""], ""title"": ""A game-theoretic model of metaphorical bargaining"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2010}, {""authors"": [""Beigman Klebanov"", ""Beata"", ""Michael Flor.""], ""title"": ""Argumentation-relevant metaphors in test-taker essays"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 11\u201320, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Birke"", ""Julia"", ""Anoop Sarkar.""], ""title"": ""A clustering approach for the nearly unsupervised recognition of nonliteral language"", ""venue"": ""Proceedings of EACL-06, pages 329\u2013336, Trento."", ""year"": 2006}, {""authors"": [""Black"", ""Max.""], ""title"": ""Models and Metaphors"", ""venue"": ""Cornell University Press. Ithaca, NY."", ""year"": 1962}, {""authors"": [""Bollegala"", ""Danushka"", ""Ekaterina Shutova.""], ""title"": ""Metaphor interpretation using paraphrases extracted from the web"", ""venue"": ""PLoS ONE, 8(9):e74304."", ""year"": 2013}, {""authors"": [""Burnard"", ""Lou.""], ""title"": ""Reference Guide for the British National Corpus (XML Edition)"", ""venue"": ""Available from www.natcorp.ox.ac.uk/ corpus/babyinfo.html."", ""year"": 2007}, {""authors"": [""Burstein"", ""Jill"", ""John Sabatini"", ""Jane Shore"", ""Brad Moulder"", ""Jennifer Lentini""], ""title"": ""A user study: Technology to increase teachers\u2019 linguistic awareness to improve instructional language support"", ""year"": 2013}, {""authors"": [""Bzdek"", ""Vincent.""], ""title"": ""Woman of the House: The Rise of Nancy Pelosi"", ""venue"": ""Palgrave Macmillan, Basingstoke, New York."", ""year"": 2008}, {""authors"": [""Cameron"", ""Lynne.""], ""title"": ""Metaphor in Educational Discourse"", ""venue"": ""Continuum, London."", ""year"": 2003}, {""authors"": [""Carbonell"", ""Jaime.""], ""title"": ""Metaphor: An inescapable phenomenon in natural language comprehension"", ""venue"": ""Wendy Lehnert and Martin Ringle, editors, Strategies for Natural Language Processing."", ""year"": 1982}, {""authors"": [""Charteris-Black"", ""Jonathan.""], ""title"": ""Metaphor and vocabulary teaching in ESP economics"", ""venue"": ""English for Specific Purposes, 19(2):149\u2013165."", ""year"": 2000}, {""authors"": [""Charteris-Black"", ""Jonathan"", ""Timothy Ennis.""], ""title"": ""A comparative study of metaphor in Spanish and English financial reporting"", ""venue"": ""English for Specific Purposes, 20(3):249\u2013266."", ""year"": 2001}, {""authors"": [""Chung"", ""Siaw-Fong"", ""Kathleen Ahrens"", ""Chu-Ren Huang.""], ""title"": ""Source domains as concept domains in metaphorical expressions"", ""venue"": ""International Journal of Computational Linguistics and Chinese"", ""year"": 2005}, {""authors"": [""Cilibrasi"", ""Rudi L."", ""Paul M.B. Vitanyi.""], ""title"": ""The Google similarity distance"", ""venue"": ""IEEE Transactional on Knowledge and Data Engineering, 19(3):370\u2013383."", ""year"": 2007}, {""authors"": [""Cortes"", ""Corinna"", ""Vladimir Vapnik.""], ""title"": ""Support-vector networks"", ""venue"": ""Machine Learning, 20(3):273\u2013297. 618"", ""year"": 1995}, {""authors"": [""Davies"", ""Mark.""], ""title"": ""The 385+ million word Corpus of Contemporary American English (1990\u20132008): Design, architecture, and linguistic insights"", ""venue"": ""International Journal of Corpus Linguistics,"", ""year"": 2009}, {""authors"": [""A. Deignan"", ""L. Potter.""], ""title"": ""A corpus study of metaphors and metonyms in English and Italian"", ""venue"": ""Journal of Pragmatics, 36:1231\u20131252."", ""year"": 2004}, {""authors"": [""Desalle"", ""Yann"", ""Bruno Gaume"", ""Karine Duvignau.""], ""title"": ""Slam: Solutions lexicales automatique pour m\u00e9taphores"", ""venue"": ""Traitement Automatique des Langues, 50(1):145\u2013175."", ""year"": 2009}, {""authors"": [""Diaz-Vera"", ""Javier"", ""Rosario Caballero.""], ""title"": ""Exploring the feeling-emotions continuum across cultures: Jealousy in English and Spanish"", ""venue"": ""Intercultural Pragmatics, 10(2):265\u2013294."", ""year"": 2013}, {""authors"": [""Dunn"", ""Jonathan.""], ""title"": ""Gradient semantic intuitions of metaphoric expressions"", ""venue"": ""Metaphor and Symbol, 26(1):53\u201367."", ""year"": 2011}, {""authors"": [""Dunn"", ""Jonathan.""], ""title"": ""Evaluating the premises and results of four metaphor identification systems"", ""venue"": ""Proceedings of CICLing\u201913, pages 471\u2013486, Samos."", ""year"": 2013}, {""authors"": [""Dunn"", ""Jonathan.""], ""title"": ""What metaphor identification systems can tell us about metaphor-in-language"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 1\u201310, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Fass"", ""Dan.""], ""title"": ""met*: A method for discriminating metonymy and metaphor by computer"", ""venue"": ""Computational Linguistics, 17(1):49\u201390."", ""year"": 1991}, {""authors"": [""Feldman"", ""Jerome.""], ""title"": ""From Molecule to Metaphor: A Neural Theory of Language"", ""venue"": ""The MIT Press, Cambridge, MA."", ""year"": 2006}, {""authors"": [""Feldman"", ""Jerome"", ""Srini Narayanan.""], ""title"": ""Embodied meaning in a neural theory of language"", ""venue"": ""Brain and Language, 89(2):385\u2013392."", ""year"": 2004}, {""authors"": [""Fillmore"", ""Charles"", ""Christopher Johnson"", ""Miriam Petruck.""], ""title"": ""Background to FrameNet"", ""venue"": ""International Journal of Lexicography, 16(3):235\u2013250."", ""year"": 2003}, {""authors"": [""Francis"", ""W. Nelson"", ""Henry Kucera.""], ""title"": ""A standard corpus of present-day edited American English"", ""venue"": ""Technical report, Brown University, Providence, RI. Available from www.helsinki.fi/varieng/CoRD/"", ""year"": 1979}, {""authors"": [""Gandy"", ""Lisa"", ""Nadji Allan"", ""Mark Atallah"", ""Ophir Frieder"", ""Newton Howard"", ""Sergey Kanareykin"", ""Moshe Koppel"", ""Mark Last"", ""Yair Neuman"", ""Shlomo Argamon""], ""title"": ""Automatic identification of conceptual"", ""year"": 2013}, {""authors"": [""Gedigian"", ""Matt"", ""John Bryant"", ""Srini Narayanan"", ""Branimir Ciric.""], ""title"": ""Catching metaphors"", ""venue"": ""Proceedings of the 3rd Workshop on Scalable Natural Language Understanding, pages 41\u201348, New York."", ""year"": 2006}, {""authors"": [""Gentner"", ""Deirdre.""], ""title"": ""Structure mapping: A theoretical framework for analogy"", ""venue"": ""Cognitive Science, 7:155\u2013170."", ""year"": 1983}, {""authors"": [""R. Gibbs""], ""title"": ""Literal meaning and psychological theory"", ""venue"": ""Cognitive Science, 8:275\u2013304."", ""year"": 1984}, {""authors"": [""Gong"", ""Shu-Ping"", ""Kathleen Ahrens"", ""Chu-Ren Huang""], ""title"": ""Chinese word sketch and mapping principles: A corpus-based study of conceptual metaphors using the building source"", ""year"": 2008}, {""authors"": [""Grady"", ""Joe.""], ""title"": ""Foundations of Meaning: Primary Metaphors and Primary Scenes"", ""venue"": ""Ph.D. thesis, University of California at Berkeley."", ""year"": 1997}, {""authors"": [""Hardie"", ""Andrew"", ""Veronika Koller"", ""Paul Rayson"", ""Elena Semino.""], ""title"": ""Exploiting a semantic annotation tool for metaphor analysis"", ""venue"": ""Proceedings of the Corpus Linguistics Conference, pages 1\u201312,"", ""year"": 2007}, {""authors"": [""Hearst"", ""Marti A.""], ""title"": ""Automatic acquisition of hyponyms from large text corpora"", ""venue"": ""Proceedings of the 14th Conference on Computational Linguistics - Volume 2, COLING \u201992, pages 539\u2013545, Nantes."", ""year"": 1992}, {""authors"": [""Heintz"", ""Ilana"", ""Ryan Gabbard"", ""Mahesh Srivastava"", ""Dave Barner"", ""Donald Black"", ""Majorie Friedman"", ""Ralph Weischedel.""], ""title"": ""Automatic extraction of linguistic metaphors with lda topic modeling"", ""venue"": ""In"", ""year"": 2013}, {""authors"": [""Hesse"", ""Mary.""], ""title"": ""Models and Analogies in Science"", ""venue"": ""Notre Dame University Press, South Bend, NA."", ""year"": 1966}, {""authors"": [""Hobbs"", ""Jerry R.""], ""title"": ""Metaphor interpretation as selective inferencing"", ""venue"": ""Proceedings of the 7th International Joint Conference on Artificial Intelligence Volume 1, IJCAI\u201981, pages 85\u201391,"", ""year"": 1981}, {""authors"": [""Hofstadter"", ""Douglas.""], ""title"": ""Fluid Concepts and Creative Analogies: Computer Models of the Fundamental Mechanisms of Thought"", ""venue"": ""HarperCollins Publishers."", ""year"": 1995}, {""authors"": [""Hofstadter"", ""Douglas"", ""Melanie Mitchell.""], ""title"": ""The Copycat Project: A model of mental fluidity and analogy-making"", ""venue"": ""K. J. Holyoak and J. A. Barnden, editors, 619"", ""year"": 1994}, {""authors"": [""Izwaini"", ""Sattar.""], ""title"": ""Corpus-based study of metaphor in information technology"", ""venue"": ""Proceedings of the Workshop on Corpus-based Approaches to Figurative Language, Corpus Linguistics 2003, pages 1\u20138, Lancaster."", ""year"": 2003}, {""authors"": [""Jamrozik"", ""Anja"", ""Eyal Sagi"", ""Micah Goldwater"", ""Dedre Gentner.""], ""title"": ""Relational words have high metaphoric potential"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 21\u201326,"", ""year"": 2013}, {""authors"": [""Karov"", ""Yael"", ""Shimon Edelman.""], ""title"": ""Similarity-based word sense disambiguation"", ""venue"": ""Computational Linguistics, 24(1):41\u201359."", ""year"": 1998}, {""authors"": [""Kingsbury"", ""Paul"", ""Martha Palmer.""], ""title"": ""From TreeBank to PropBank"", ""venue"": ""Proceedings of LREC-2002, pages 1989\u20131993, Gran Canaria."", ""year"": 2002}, {""authors"": [""Koller"", ""Veronika.""], ""title"": ""Metaphor and Gender in Business Media Discourse: A Critical Cognitive Study"", ""venue"": ""Palgrave Macmillan, Basingstoke, New York."", ""year"": 2004}, {""authors"": [""Korkontzelos"", ""Ioannis"", ""Torsten Zesch"", ""Fabio Massimo Zanzotto"", ""Chris Biemann.""], ""title"": ""Semeval-2013 task 5: Evaluating phrasal semantics"", ""venue"": ""Second Joint Conference on Lexical and Computational"", ""year"": 2013}, {""authors"": [""Z. Kovecses""], ""title"": ""Metaphor in Culture: Universality and Variation"", ""venue"": ""Cambridge University Press, Cambridge."", ""year"": 2005}, {""authors"": [""Krishnakumaran"", ""Saisuresh"", ""Xiaojin Zhu.""], ""title"": ""Hunting elusive metaphors using lexical resources"", ""venue"": ""Proceedings of the Workshop on Computational Approaches to Figurative Language, pages 13\u201320,"", ""year"": 2007}, {""authors"": [""Lakoff"", ""George.""], ""title"": ""The Political Mind"", ""venue"": ""Viking, New York."", ""year"": 2008}, {""authors"": [""Lakoff"", ""George"", ""Jane Espenson"", ""Alan Schwartz.""], ""title"": ""The master metaphor list"", ""venue"": ""Technical report, University of California at Berkeley. Available from araw.mede.uic.edu/valansz/"", ""year"": 1991}, {""authors"": [""Lakoff"", ""George"", ""Mark Johnson.""], ""title"": ""Metaphors We Live By"", ""venue"": ""University of Chicago Press, Chicago."", ""year"": 1980}, {""authors"": [""Lakoff"", ""George"", ""Elisabeth Wehling.""], ""title"": ""The Little Blue Book: The Essential Guide to Thinking and Talking Democratic"", ""venue"": ""Free Press, New York."", ""year"": 2012}, {""authors"": [""Landau"", ""Mark J."", ""Daniel Sullivan"", ""Jeff Greenberg.""], ""title"": ""Evidence that self-relevant motives and metaphoric framing interact to influence political and social attitudes"", ""venue"": ""Psychological"", ""year"": 2009}, {""authors"": [""Lefever"", ""Els"", ""V\u00e9ronique Hoste.""], ""title"": ""Semeval-2010 task 3: Cross-lingual word sense disambiguation"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic Evaluation, pages 15\u201320, Uppsala."", ""year"": 2010}, {""authors"": [""Lewis"", ""David D."", ""Yiming Yang"", ""Tony G. Rose"", ""Fan Li.""], ""title"": ""Rcv1: A new benchmark collection for text categorization research"", ""venue"": ""Journal of Machine Learning Research, 5:361\u2013397."", ""year"": 2004}, {""authors"": [""Li"", ""Hongsong"", ""Kenny Q. Zhu"", ""Haixun Wang.""], ""title"": ""Data-driven metaphor recognition and explanation"", ""venue"": ""Transactions of the Association for Computational Linguistics, 1:379\u2013390."", ""year"": 2013}, {""authors"": [""Li"", ""Linlin"", ""Caroline Sporleder.""], ""title"": ""Classifier combination for contextual idiom detection without labelled data"", ""venue"": ""Proceedings of the 2009 Conference on Empirical Methods in Natural Language"", ""year"": 2009}, {""authors"": [""Li"", ""Linlin"", ""Caroline Sporleder.""], ""title"": ""Using Gaussian mixture models to detect figurative language in context"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North"", ""year"": 2010}, {""authors"": [""L\u00f6nneker"", ""Birte.""], ""title"": ""Lexical databases as resources for linguistic creativity: Focus on metaphor"", ""venue"": ""Proceedings of the LREC 2004 Workshop on Language Resources for Linguistic Creativity,"", ""year"": 2004}, {""authors"": [""L\u00f6nneker-Rodman"", ""Birte.""], ""title"": ""The Hamburg metaphor database project: Issues in resource creation"", ""venue"": ""Language Resources and Evaluation, 42(3):293\u2013318."", ""year"": 2008}, {""authors"": [""Low"", ""Graham"", ""Zazie Todd"", ""Alice Deignan"", ""Lynne Cameron.""], ""title"": ""Researching and Applying Metaphor in the Real World"", ""venue"": ""John Benjamins, Amsterdam/Philadelphia."", ""year"": 2010}, {""authors"": [""Lu"", ""Louis"", ""Kathleen Ahrens""], ""title"": ""Ideological influences on building metaphors in Taiwanese presidential"", ""year"": 2008}, {""authors"": [""Manandhar"", ""Suresh"", ""Ioannis Klapaftis"", ""Dmitriy Dligach"", ""Sameer Pradhan.""], ""title"": ""Semeval-2010 task 14: Word sense induction & disambiguation"", ""venue"": ""Proceedings of the 5th International Workshop on"", ""year"": 2010}, {""authors"": [""Martin"", ""James.""], ""title"": ""A Computational Model of Metaphor Interpretation"", ""venue"": ""Academic Press Professional, Inc., San Diego, CA."", ""year"": 1990}, {""authors"": [""Martin"", ""James.""], ""title"": ""A corpus-based analysis of context effects on metaphor comprehension"", ""venue"": ""A. Stefanowitsch and S. T. Gries, editors, Corpus-Based Approaches to Metaphor and Metonymy."", ""year"": 2006}, {""authors"": [""Mason"", ""Zachary.""], ""title"": ""Cormet: A computational, corpus-based conventional metaphor extraction system"", ""venue"": ""Computational Linguistics, 30(1):23\u201344."", ""year"": 2004}, {""authors"": [""McCarthy"", ""Diana"", ""Roberto Navigli.""], ""title"": ""Semeval-2007 task 10: English lexical substitution task"", ""venue"": ""Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007), pages 48\u201353,"", ""year"": 2007}, {""authors"": [""Mihalcea"", ""Rada"", ""Ravi Sinha"", ""Diana McCarthy.""], ""title"": ""Semeval-2010 task 2: Cross-lingual lexical substitution"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic Evaluation, pages 9\u201314,"", ""year"": 2010}, {""authors"": [""Mohler"", ""Michael"", ""David Bracewell"", ""Marc Tomlinson"", ""David Hinote.""], ""title"": ""Semantic signatures for example-based linguistic metaphor detection"", ""venue"": ""Proceedings of the First Workshop on Metaphor"", ""year"": 2013}, {""authors"": [""Moschitti"", ""Ro"", ""Daniele Pighin"", ""Roberto Basili.""], ""title"": ""Tree kernel engineering for proposition re-ranking"", ""venue"": ""Proceedings of Mining and Learning with Graphs (MLG), pages 165\u2013172, Berlin."", ""year"": 2006}, {""authors"": [""Musolff"", ""Andreas.""], ""title"": ""Mirror Images of Europe: Metaphors in the Public Debate about Europe in Britain and Germany"", ""venue"": ""Iudicium, Muenchen."", ""year"": 2000}, {""authors"": [""Nakov"", ""Preslav"", ""Sara Rosenthal"", ""Zornitsa Kozareva"", ""Veselin Stoyanov"", ""Alan Ritter"", ""Theresa Wilson.""], ""title"": ""Semeval-2013 task 2: Sentiment analysis in twitter"", ""venue"": ""Second Joint Conference on Lexical and"", ""year"": 2013}, {""authors"": [""Narayanan"", ""Srini""], ""title"": ""Knowledge-Based Action Representations for Metaphor and"", ""year"": 1997}, {""authors"": [""Narayanan"", ""Srini.""], ""title"": ""Moving right along: A computational model of metaphoric reasoning about events"", ""venue"": ""Proceedings of AAAI 99), pages 121\u2013128, Orlando, FL."", ""year"": 1999}, {""authors"": [""Neuman"", ""Yair"", ""Dan Assaf"", ""Yohai Cohen"", ""Mark Last"", ""Shlomo Argamon"", ""Newton Howard"", ""Ophir Frieder.""], ""title"": ""Metaphor identification in large texts corpora"", ""venue"": ""PLoS ONE, 8(4):e62343."", ""year"": 2013}, {""authors"": [""Niculae"", ""Vlad"", ""Victoria Yaneva.""], ""title"": ""Computational considerations of comparisons and similes"", ""venue"": ""Proceedings of ACL (Student Research Workshop), pages 89\u201395, Sophia."", ""year"": 2013}, {""authors"": [""Niles"", ""Ian"", ""Adam Pease.""], ""title"": ""Towards a standard upper ontology"", ""venue"": ""Proceedings of the International Conference on Formal Ontology in Information Systems - Volume 2001, FOIS \u201901, pages 2\u20139, New York, NY."", ""year"": 2001}, {""authors"": [""Niles"", ""Ian"", ""Adam Pease.""], ""title"": ""Linking lexicons and ontologies: Mapping WordNet to the suggested upper merged ontology"", ""venue"": ""Proceedings of the International Conference on Information and Knowledge"", ""year"": 2003}, {""authors"": [""Nunberg"", ""Geoffrey.""], ""title"": ""Poetic and prosaic metaphors"", ""venue"": ""Proceedings of the 1987 Workshop on Theoretical Issues in Natural Language Processing, pages 198\u2013201, Stroudsburg, PA."", ""year"": 1987}, {""authors"": [""Peters"", ""Wim"", ""Ivonne Peters.""], ""title"": ""Lexicalised systematic polysemy in WordNet"", ""venue"": ""Proceedings of LREC 2000, pages 1\u20137, Athens."", ""year"": 2000}, {""authors"": [""Pragglejaz Group.""], ""title"": ""MIP: A method for identifying metaphorically used words in discourse"", ""venue"": ""Metaphor and Symbol, 22:1\u201339."", ""year"": 2007}, {""authors"": [""Recasens"", ""Marta"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Emili Sapena"", ""M. Ant\u00f2nia Mart\u0131"", ""Mariona Taul\u00e9"", ""V\u00e9ronique Hoste"", ""Massimo Poesio"", ""Yannick Versley""], ""title"": ""Semeval-2010 task 1: Coreference resolution"", ""year"": 2010}, {""authors"": [""Reining"", ""Astrid"", ""Birte L\u00f6nneker-Rodman.""], ""title"": ""Corpus-driven metaphor harvesting"", ""venue"": ""Proceedings of the HLT/NAACL-07 Workshop on Computational Approaches to Figurative Language, pages 5\u201312,"", ""year"": 2007}, {""authors"": [""Resnik"", ""Philip.""], ""title"": ""Selection and Information: A Class-Based Approach to Lexical Relationships"", ""venue"": ""Ph.D. thesis, University of Pennsylvania. 621"", ""year"": 1993}, {""authors"": [""Rohrer"", ""Tim.""], ""title"": ""Conceptual blending on the information highway: How metaphorical inferences work"", ""venue"": ""W. Liebert, G. Redeker, and L. Waugh, editors, Discourse and Perspective in"", ""year"": 1997}, {""authors"": [""Sandhaus"", ""Evan.""], ""title"": ""The New York Times Annotated Corpus"", ""venue"": ""Available from https://catalog/lde.upenn.edu/ LDC2008T19."", ""year"": 2008}, {""authors"": [""Santa Ana"", ""Otto.""], ""title"": ""Like an animal I was treated?: Anti-immigrant metaphor in US public discourse"", ""venue"": ""Discourse Society, 10(2):191\u2013224."", ""year"": 1999}, {""authors"": [""Shutova"", ""Ekaterina.""], ""title"": ""Automatic metaphor interpretation as a paraphrasing task"", ""venue"": ""Proceedings of NAACL 2010, pages 1029\u20131037, Los Angeles, CA."", ""year"": 2010}, {""authors"": [""Shutova"", ""Ekaterina.""], ""title"": ""Metaphor identification as interpretation"", ""venue"": ""Proceedings of *SEM 2013, pages 276\u2013285, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Shutova"", ""Ekaterina"", ""Lin Sun.""], ""title"": ""Unsupervised metaphor identification using hierarchical graph factorization clustering"", ""venue"": ""Proceedings of NAACL 2013, pages 978\u2013988, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Shutova"", ""Ekaterina"", ""Lin Sun"", ""Anna Korhonen.""], ""title"": ""Metaphor identification using verb and noun clustering"", ""venue"": ""Proceedings of Conference on Computational Linguistics 2010, pages 1002\u20131010,"", ""year"": 2010}, {""authors"": [""Shutova"", ""Ekaterina"", ""Simone Teufel.""], ""title"": ""Metaphor corpus annotated for source\u2013 target domain mappings"", ""venue"": ""Proceedings of LREC 2010, pages 3255\u20133261, Malta."", ""year"": 2010}, {""authors"": [""Shutova"", ""Ekaterina"", ""Simone Teufel"", ""Anna Korhonen.""], ""title"": ""Statistical metaphor processing"", ""venue"": ""Computational Linguistics, 39(2):301\u2013353."", ""year"": 2013}, {""authors"": [""Shutova"", ""Ekaterina"", ""Tim Van de Cruys"", ""Anna Korhonen.""], ""title"": ""Unsupervised metaphor paraphrasing using a vector space model"", ""venue"": ""Proceedings of COLING 2012, pages 1121\u20131130, Mumbai."", ""year"": 2012}, {""authors"": [""Skorczynska Sznajder"", ""Hanna"", ""Jordi Pique-Angordans.""], ""title"": ""A corpus-based description of metaphorical marking patterns in scientific and popular business discourse"", ""venue"": ""Proceedings of European"", ""year"": 2004}, {""authors"": [""Trijntje Pasma""], ""title"": ""A Method for Linguistic Metaphor Identification: From MIP to MIPVU"", ""year"": 2010}, {""authors"": [""A. Stefanowitsch""], ""title"": ""Happiness in English and German: A metaphorical-pattern analysis"", ""venue"": ""M. Altenberg and S. Kemmer, editors, Language, Culture, and Mind. CSLI Publications Stanford,"", ""year"": 2004}, {""authors"": [""Strzalkowski"", ""Tomek"", ""George Aaron Broadwell"", ""Sarah Taylor"", ""Laurie Feldman"", ""Samira Shaikh"", ""Ting Liu"", ""Boris Yamrom"", ""Kit Cho"", ""Umit Boz"", ""Ignacio Cases"", ""Kyle Elliot""], ""title"": ""Robust extraction"", ""year"": 2013}, {""authors"": [""Sullivan"", ""Karen.""], ""title"": ""Grammar in metaphor: A construction grammar account of metaphoric language"", ""venue"": ""Ph.D. thesis, University of California, Berkeley."", ""year"": 2007}, {""authors"": [""Sullivan"", ""Karen.""], ""title"": ""Frames and constructions in metaphoric language"", ""venue"": ""John Benjamins, Amsterdam."", ""year"": 2013}, {""authors"": [""Sun"", ""Lin"", ""Anna Korhonen.""], ""title"": ""Improving verb clustering with automatically acquired selectional preferences"", ""venue"": ""Proceedings of EMNLP 2009, pages 638\u2013647, Singapore."", ""year"": 2009}, {""authors"": [""Thibodeau"", ""Paul H."", ""Lera Boroditsky.""], ""title"": ""Metaphors we think with: The role of metaphor in reasoning"", ""venue"": ""PLoS ONE, 6(2):e16782, 02."", ""year"": 2011}, {""authors"": [""Tsvetkov"", ""Yulia"", ""Leonid Boytsov"", ""Anatole Gershman"", ""Eric Nyberg"", ""Chris Dyer.""], ""title"": ""Metaphor detection with crosslingual model transfer"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association"", ""year"": 2014}, {""authors"": [""Tsvetkov"", ""Yulia"", ""Elena Mukomel"", ""Anatole Gershman.""], ""title"": ""Cross-lingual metaphor detection using common semantic features"", ""venue"": ""Proceedings of the First Workshop on Metaphor in NLP, pages 45\u201351, Atlanta, GA."", ""year"": 2013}, {""authors"": [""Turner"", ""Mark"", ""Gilles Fauconnier.""], ""title"": ""Metaphor, metonymy, and binding"", ""venue"": ""A. Barcelona, editor, Metaphor and Metonymy at the Crossroads: A Cognitive Perspective. Walter de Gruyter GmbH,"", ""year"": 2003}, {""authors"": [""Turney"", ""Peter D."", ""Yair Neuman"", ""Dan Assaf"", ""Yohai Cohen.""], ""title"": ""Literal and metaphorical sense identification through concrete and abstract context"", ""venue"": ""Proceedings of the Conference on Empirical"", ""year"": 2011}, {""authors"": [""Veale"", ""Tony.""], ""title"": ""Creative language retrieval: A robust hybrid of information retrieval and linguistic creativity"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human"", ""year"": 2011}, {""authors"": [""Veale"", ""Tony.""], ""title"": ""A service-oriented architecture for metaphor processing"", ""venue"": ""Proceedings of the Second Workshop on Metaphor in NLP, pages 52\u201360, Baltimore, MD."", ""year"": 2014}, {""authors"": [""Veale"", ""Tony"", ""Yanfen Hao.""], ""title"": ""A fluid knowledge representation for understanding and generating creative metaphors"", ""venue"": ""Proceedings of COLING 2008, pages 945\u2013952, Manchester,"", ""year"": 2008}, {""authors"": [""A.M. Wallington"", ""J.A. Barnden"", ""P. Buchlovsky"", ""L. Fellows"", ""S.R. Glasbey.""], ""title"": ""Metaphor annotation: A systematic study"", ""venue"": ""Technical Report"", ""year"": 2003}, {""authors"": [""Wikberg"", ""Kay.""], ""title"": ""The role of corpus studies in metaphor research"", ""venue"": ""2006 Stockholm Metaphor Festival, pages 33\u201348, Stockholm."", ""year"": 2006}, {""authors"": [""Wilks"", ""Yorick.""], ""title"": ""A preferential pattern-seeking semantics for natural language inference"", ""venue"": ""Artificial Intelligence, 6:53\u201374."", ""year"": 1975}, {""authors"": [""Wilks"", ""Yorick.""], ""title"": ""Making preferences more active"", ""venue"": ""Artificial Intelligence, 11(3):197\u2013223."", ""year"": 1978}, {""authors"": [""Wilks"", ""Yorick"", ""Adam Dalton"", ""James Allen"", ""Lucian Galescu.""], ""title"": ""Automatic metaphor detection using large-scale lexical resources and conventional metaphor extraction"", ""venue"": ""Proceedings"", ""year"": 2013}, {""authors"": [""Witten"", ""Ian H."", ""Eibe Frank.""], ""title"": ""Data Mining: Practical Machine Learning Tools and Techniques"", ""venue"": ""Morgan Kaufmann Series in Data Management Systems. Morgan Kaufmann, 2nd edition."", ""year"": 2005}]","2 considerations in the design of a metaphor processing system :When designing a metaphor processing system, one faces a number of choices. Some stem from the linguistic and cognitive properties of metaphor, others concern the applicability and usefulness of the system in the wider NLP context. In this section, we analyze individual aspects of metaphor and their relevance to computational modeling, as well as their interplay in the design of a real-world system. Linguistic considerations that inform the design of metaphor processing systems concern primarily the choice of the level (or levels) of analysis. The levels of metaphor analysis include (1) linguistic metaphor (or metaphorical expressions), (2) conceptual metaphor, (3) extended metaphor, and (4) metaphorical inference. Let us consider an example of manifestations of the conceptual metaphor EUROPEAN INTEGRATION as a TRAIN JOURNEY, popular in the early 1990s, at various levels.r Conceptual: EUROPEAN INTEGRATION as a TRAIN JOURNEYr Linguistic: The coupling of the carriages may not be reliably secure, but the
pan-European express is in motion.r Extended metaphor: “There is a fear that the European train will thunder forward, laden with its customary cargo of gravy, towards a destination neither wished for nor understood by electorates. But the train can be stopped.” (Margaret Thatcher, Sunday Times, 20 Sept 1992)r Metaphorical inference: The fact that expensive tracks have to be laid for the train to move forward means that someone has to fund the process of European integration.
2.1.1 Linguistic metaphor. Linguistic metaphor, or metaphorical expressions, concern the surface realization of metaphorical mechanisms, and have been unsurprisingly central to metaphor processing research to date (Birke and Sarkar 2006; Gedigian et al. 2006; Krishnakumaran and Zhu 2007; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Gandy et al. 2013; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013). Metaphorical expressions represent the way in which text-processing systems encounter metaphor, and there is little doubt that any real-world metaphor processing system, whatever the approach or the application, ultimately needs to be able to identify and interpret them.
When focusing on linguistic metaphor, one needs to further take into account the level of conventionality of the expressions; how different syntactic constructions are used to convey metaphorical meanings and at what level metaphor annotation needs to be done (word, relation, or sentence level). Thus the following considerations become important for the task and system design:r Level of conventionality of the metaphors accepted: Metaphor is a
productive phenomenon, that is, its novel examples continue to emerge in language. However, a large number of metaphorical expressions become conventionalized over time (e.g., “I cannot grasp his way of thinking”). Although metaphorical in nature, their meanings are deeply entrenched in everyday use, and their comprehension is likened to that of literally used terms (Nunberg 1987). According to Gibbs (1984), metaphorical expressions are spread along a continuum from highly conventional, lexicalized metaphors to entirely novel and creative ones. Gibbs thus suggests that there is no clear demarcation line between literal and metaphorical language, and the distinction between them is rather governed by the level of conventionality of metaphorical expressions. From the usage perspective, metaphoricity may be viewed as a gradient phenomenon rather than a binary one (Dunn 2011), and conventionality becomes an important factor in the design and evaluation of metaphor processing systems. It is not yet clear where on the metaphorical–literal continuum the system should draw the line between what it considers metaphorical and what it considers literal. The answer to this question most likely depends on the NLP application in mind. However, generally speaking, real-world NLP applications are unlikely to be concerned with historical aspects of metaphor, but rather with the identification of figurative language that needs to be interpreted differently from the literal language. We therefore suggest that NLP applications do not necessarily need to address highly conventional and lexicalized metaphors that can be interpreted using standard word sense disambiguation techniques, but rather would benefit from the identification of less conventional and more creative language. Much of the metaphor processing work has focused on conventional metaphor, though in principle capable of identifying novel metaphor as well (Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Dunn 2013a; Gandy et al. 2013; Heintz et al. 2013; Neuman et al. 2013; Shutova 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013), with few approaches modeling only novel metaphor (Desalle, Gaume, and Duvignau 2009) or discriminating between conventional and novel metaphors (Krishnakumaran and Zhu 2007). Other approaches have
looked at literal versus non-literal distinction defined more broadly (Birke and Sarkar 2006; Li and Sporleder 2009, 2010).r Syntactic constructions covered: Metaphors vary with respect to how they are expressed in language grammatically. The grammatical structure of metaphorical expressions is tightly coupled with the inference processes that produce them and are guided by the semantic frames they evoke (Sullivan 2007, 2013). According to Sullivan, conceptual metaphors are realized in language via mapping semantic roles in the source frame onto roles in the target frame. This suggests that both the source and the target domain impose a set of constraints on the roles that can be mapped, and thus on the syntactic options for expressing the metaphorical meaning. There has not yet been a computational approach investigating the interplay of frame semantics and surface structure of metaphorical language. However, the community has addressed modeling linguistic metaphor in a range of syntactic constructions. Verbal or adjectival metaphors are particularly widely embraced by NLP researchers (most approaches), with a few works focusing on copula constructions (Krishnakumaran and Zhu 2007; Gandy et al. 2013; Li, Zhu, and Wang 2013; Neuman et al. 2013) or other nominal metaphors (Heintz et al. 2013; Hovy et al. 2013; Li, Zhu, and Wang 2013; Mohler et al. 2013; Strzalkowski et al. 2013). A number of approaches to metaphor identification have also addressed multiword metaphors (Li and Sporleder 2010; Heintz et al. 2013; Hovy et al. 2013; Mohler et al. 2013). Corpus-linguistic research has shown that verbs and adjectives account for a large proportion of metaphorical expressions observed in the data (Cameron 2003; Shutova and Teufel 2010). However, a recent study (Jamrozik et al. 2013) has also shown that relational words tend to have a higher metaphorical potential. This corresponds to the data on verbal metaphor frequency, and it also suggests that relational nouns are important (e.g., “words are friends of translators”).r Lexical, relation, or sentence level: Finally, one needs to decide if metaphor should be annotated at the word level (i.e., tagging the source domain words alone), relation level (i.e., tagging both source and target words in a particular grammatical relation), or sentence level (i.e., tagging sentences that contain metaphorical language, without explicit annotation of source and target domain words). The word or relation levels provide the most information and have been the focus of the majority of approaches (Gedigian et al. 2006; Shutova, Sun, and Korhonen 2010; Turney et al. 2011; Gandy et al. 2013; Heintz et al. 2013; Hovy et al. 2013; Neuman et al. 2013; Shutova 2013; Shutova and Sun 2013; Wilks et al. 2013). However, some works annotated metaphor at the sentence level (Krishnakumaran and Zhu 2007; Dunn 2013a; Li, Zhu, and Wang 2013; Mohler et al. 2013; Strzalkowski et al. 2013; Tsvetkov, Mukomel, and Gershman 2013).
2.1.2 Conceptual metaphor. Conceptual metaphor represents a cognitive and conceptual mechanism by which humans produce and comprehend metaphorical expressions. Manifestations of conceptual metaphor are ubiquitous in language, communication, and even decision-making (Thibodeau and Boroditsky 2011). Here are a few examples of
common metaphorical mappings: TIME is MONEY (e.g., “That flat tire cost me an hour”), IDEAS are PHYSICAL OBJECTS (e.g., “I cannot grasp his way of thinking”), EMOTIONS ARE VEHICLES (e.g., “she was transported with pleasure”), FEELINGS ARE LIQUIDS (e.g., “all of this stirred an unfathomable excitement in her”), LIFE IS A JOURNEY (e.g., “He arrived at the end of his life with very little emotional baggage”); ARGUMENT IS A WAR (e.g., “He shot down all of my arguments,” “He attacked every weak point in my argument”).
On one hand, few would disagree that the metaphor processing systems capable of understanding and applying conceptual metaphor should be in a better position to accurately handle linguistic metaphors as well. However, a series of questions arise when designing a model of conceptual metaphor. For example, how does one represent conceptual metaphors in the system? What labels does one assign to source and target domains? Is it even possible to name all the conceptual metaphors that humans use and is it necessary to do so? A computational model needs a clear definition of what constitutes the source and target domains (whether they are manually listed or automatically learned) and the consistency and coverage of source and target domain categories would play a crucial role in how well the model can account for real-world data. Previous research on annotation of conceptual metaphor (Shutova and Teufel 2010) has shown that the annotators tend to disagree on the assignment of source and target domain categories. The most variation stems from the level of generality of the selected categories, indicating that while cross-domain mappings are intuitive to humans (i.e., they can be annotated in arbitrary text in principle), labeling source and target domains consistently appears to be a challenging task. What this suggests is that, although the mechanism of conceptual metaphor may be helpful to the system, the question of how to best represent source and target domains within the system remains open. A predefined set of categories, such as those widely discussed in the linguistic literature on CMT, may not be sufficient or even suitable for a computational model. And despite the validity of the main principles of CMT as a linguistic theory, it is not straightforward to port it to computational modeling of metaphor. A more flexible, and potentially datadriven, representation of source and target domain categories is needed for the latter purpose. A data-driven representation would also be better suited to account for the freedom of interpretation that some metaphors allow, since more flexible structures can be dynamically learned from the data. So far, the community has attempted assigning manually created labels to metaphorical mappings (Mason 2004; Baumer, Tomlinson, and Richland 2009), harvesting fine-grained mappings between individual nouns (Li, Zhu, and Wang 2013), using lexical resources to define or expand source and target domain categories (Gandy et al. 2013; Mohler et al. 2013), representing source and target concepts as word clusters (Shutova and Sun 2013) or automatically learned topics (Heintz et al. 2013), and learning metaphorical mappings implicitly within the model without explicit labeling (Shutova, Sun, and Korhonen 2010).
2.1.3 Extended metaphor. Extended metaphor refers to the use of metaphor at the discourse level. It manifests itself in discourse via a sequence of metaphorically used language, yielded by the same conceptual metaphor, whereby a continuous scenario from the source domain is metaphorically projected onto the target. For instance, viewing European integration as a train journey led to numerous metaphorical expressions in political discourse. Each of them mapped certain properties of train journeys to political processes—with countries as loosely connected carriages, peoples of different countries as passengers of their respective carriages, expensive tracks that have to be laid for the train to move forward, and the final destination not very well understood. This metaphor has been dominating the debate, with the European leaders arguing over its details
(Beigman Klebanov and Beigman 2010). One can see this metaphor frequently reappear and evolve over time, helping politicians defend their agendas and, not least, shedding some clarity on an otherwise uncertain future.
Research in linguistics and political science (Musolff 2000; Lakoff 2008; Lakoff and Wehling 2012) suggests that the use of a particular metaphor often guides the speakers’ argumentation strategy throughout a piece of discourse, as well as participants’ behavior in a dialogue. Beigman Klebanov and Beigman (2010) investigated extended metaphor within a game-theoretic framework, demonstrating that maintaining the metaphorical frame in a debate is rationalizable in terms of the gains the participants may get from doing so and their potential losses from swerving away from the metaphor. Their approach reverse-engineers the motivations behind the use of extended metaphor within a formal framework. Beigman Klebanov and Beigman’s work was an important advance, enhancing our understanding of the inner workings of extended metaphor, motivation behind its use, and its effects on social dynamics. However, a computational method for identification and interpretation of extended metaphor in real-world discourse is yet to be proposed. A discourse-level metaphor processing system would need to identify a chain of metaphorical expressions in a text, which indicates a systematic association of the text topic with a particular domain. These chains would then demonstrate how continuous scenarios can be transferred across domains. Recovering this information from the data would allow us to better understand the structure behind metaphorical associations, as well as the inferential process by which knowledge is projected across domains. This system would also find application in social science, where metaphorical framing is widely studied as an indicator of the underlying cultural and moral models (Lakoff and Wehling 2012).
2.1.4 Metaphorical inference. When projecting knowledge from one domain to another, a set of complex inferences take place. Metaphorical inferences are grounded in the source domain and result in the production of surface structures we observe in language as metaphorical expressions. Metaphorical mappings are thus realized via projecting inference structures from the source domain onto the target. For example, when European integration is metaphorically viewed as a train journey, our knowledge of typical events and their consequences from the domain of train journey are projected onto our reasoning about the process of European integration. For instance, if we know that expensive tracks need to be laid before a train can move forward, we can infer that someone also needs to fund the process of European integration. Interestingly, in the presence of a conceptual metaphor such inference can take place even without any linguistic metaphor referring to the tracks being present, but rather on the basis of our common sense knowledge about the functioning of trains.
Besides allowing us to derive new information about the target domain, projecting the inferential structures from the source domain also invokes an emotional response coming from the source domain—for example, an unknown destination or a great expense make one feel uneasy when referred to both literally and metaphorically. Such a transfer of inferential processes and emotional content is believed by some to be one of the central purposes of metaphor (Kovecses 2005; Feldman 2006; Thibodeau and Boroditsky 2011). Psychologists Thibodeau and Boroditsky (2011) investigated how metaphor and metaphorical inference affect decision-making. Their hypothesis was that metaphors in language “instantiate frame-consistent knowledge structures [from the source domain] and invite structurally consistent inferences [about the target domain].” They used two groups of subjects, who were presented with two different texts about crime. In the first text crime was metaphorically portrayed as a VIRUS and
in the second as a BEAST. The two groups were then asked a set of questions on how to tackle crime in the city. It turned out that, whereas the first group tended to opt for preventive measures in tackling crime (e.g., stronger social policies), the second group converged on punishment- or restraint-oriented measures. According to Thibodeau and Boroditsky, their results demonstrate that metaphors have profound influence on how we conceptualize and act with respect to societal issues. Interestingly, the participants would explain their decisions via arguments unrelated to the metaphor, showing that the effect of metaphorical inference in guiding human reasoning is rather covert.
Being able to reproduce these inferences is likely to make automatic metaphor understanding better informed and hence more accurate. It may also provide a mechanism of representing source–target domain mappings, that themselves are generalizations over a set of inferences transferred from one domain to another. And finally, these inferences provide a platform for metaphor interpretation, namely, deriving the meaning of a metaphorical expression and the additional connotations it introduces (that are likely to originate from the source domain). There is a consensus among cognitive linguists that it is metaphorical inference that provides for the very texture behind the use of metaphor (Hobbs 1981; Carbonell 1982; Rohrer 1997; Turner and Fauconnier 2003; Feldman 2006). Although uncovering this texture is certainly one of the main objectives of computational metaphor understanding, it is at the same time a very challenging undertaking. Reproducing metaphorical inferences would require the ability to learn vast amounts of world knowledge from the data, as well as performing complex crossdomain comparisons. And despite being a very promising route, it has not yet been attempted in NLP. Another set of considerations in the design of metaphor processing systems stems from the needs of real-world NLP. The high frequency of metaphorical language in textual data makes accurate metaphor processing desired for a number of NLP applications. Thus, the format of metaphor understanding that metaphor processing systems provide should ideally be informed by the requirements of external NLP and a number of considerations arise in that respect:r Metaphor processing typically involves two tasks: metaphor
identification (distinguishing between literal and metaphorical language in text) and metaphor interpretation (identifying the intended meaning of the metaphor). Both of these provide useful information for language understanding and need to be addressed, either independently or as a single process.r A metaphor processing system should provide a representation of metaphor interpretation that can be easily integrated with other NLP systems. This criterion places constraints on how the metaphor processing task should be defined. The most universally applicable metaphor interpretation would be in the text-to-text form. This means that a metaphor processing system would take raw text as input and provide textual output, in which metaphors are interpreted.r In order to be useful for real-world NLP, the system needs to be capable of processing real-world data, and thus operate on unrestricted, continuous text. Rather than only dealing with individual carefully selected clear-cut
examples, the system should be fully implemented and tested on free, naturally occurring text.r To enable wide applicability, the system needs to be open-domain—that is, operate in all domains, genres, and topics. Therefore, ideally it should not rely on any domain-specific information or focus on individual types of instances (e.g., a limited set of hand-chosen source-target domain mappings).r To be easily adaptable to new domains, the system should not rely on task-specific hand-coded knowledge. This means it needs to be either data-driven and be able to automatically acquire the knowledge it needs from text corpora, or rely only on large-scale, general-domain lexical resources (that are already in existence and do not need to be created in a costly manner). However, it would be an advantage if no such resource is required and the system can dynamically induce meanings in context.r To be robust, the system needs to be able to deal with metaphors represented by all word classes and syntactic constructions. Many existing models are designed with specific kinds of metaphorical expressions in mind, for instance nominal metaphors in copula constructions or verbal metaphors in verb–object relations. To be applicable to support real-world NLP applications, these models need to be extended beyond those specific word classes and syntactic constructions, and be able to process any kind of metaphorical language.
When designing a metaphor processing task, methodology, and evaluation strategy, one thus needs to keep these criteria in mind. Although modeling all of the phenomena described in this section within a single system is by no means a requirement, it is critically important to be aware of all the guises that metaphor may take, both conceptually and empirically."
"1 introduction :Word ordering is a fundamental problem in natural language generation (NLG, Reiter and Dale 1997). In this article we focus on text generation: Starting with a bag of words, or lemmas, as input, the task is to generate a fluent and grammatical sentence using
∗ Singapore University of Technology and Design. 20 Dover Drive, Singapore. E-mail: yue zhang@sutd.edu.sg.
∗∗ University of Cambridge Computer Laboratory, William Gates Building, 15 JJ Thomson Avenue, Cambridge, UK. E-mail: stephen.clark@cl.cam.ac.uk.
Submission received: 17 April 2013; revised version received: 23 April 2014; accepted for publication: 22 June 2014.
doi:10.1162/COLI a 00229
© 2015 Association for Computational Linguistics
those words. Additional annotation may also be provided with the input—for example, part-of-speech (POS) tags or syntactic dependencies. Applications that can benefit from better text generation algorithms include machine translation (Koehn 2010), abstractive text summarization (Barzilay and McKeown 2005), and grammar correction (Lee and Seneff 2006). Typically, statistical machine translation (SMT) systems (Chiang 2007; Koehn 2010) perform generation into the target language as part of an integrated system, which avoids the high computational complexity of independent word ordering. On the other hand, performing word ordering separately in a pipeline has many potential advantages. For SMT, it offers better modularity between adequacy (translation) and fluency (linearization), and can potentially improve target grammaticality for syntactically different languages (e.g., Chinese and English). More importantly, a standalone word ordering component can in principle be applied to a wide range of text generation tasks, including transfer-based machine translation (Chang and Toutanova 2007).
Most word ordering systems use an n-gram language model, which is effective at controling local fluency. Syntax-based language models, in particular dependency language models (Xu, Chelba, and Jelinek 2002), are sometimes used in an attempt to improve global fluency through the capturing of long-range dependencies. In this article, we take a syntax-based approach and consider two grammar formalisms: Combinatory Categorial Grammar (CCG) and dependency grammar. Our system also employs a discriminative model. Coupled with heuristic search, a strength of the model is that arbitrary features can be defined to capture complex syntactic patterns in output hypotheses. The discriminative model is trained using syntactically annotated data.
From the perspective of search, word ordering is a computationally difficult problem. Finding the best permutation for a set of words according to a bigram language model, for example, is NP-hard, which can be proved by linear reduction from the traveling salesman problem. In practice, exploring the whole search space of permutations is often prevented by adding constraints. In phrase-based machine translation (Koehn, Och, and Marcu 2003; Koehn et al. 2007), a distortion limit is used to constrain the position of output phrases. In syntax-based machine translation systems such as Wu (1997) and Chiang (2007), synchronous grammars limit the search space so that polynomialtime inference is feasible. In fluency improvement (Blackwood, de Gispert, and Byrne 2010), parts of translation hypotheses identified as having high local confidence are held fixed, so that word ordering elsewhere is strictly local.
In this article we begin by proposing a general system to solve the word ordering problem, which does not rely on constraints (which are typically task-specific). In particular, we treat syntax-based word ordering as a structured prediction problem, for which the input is a multi-set (bag) of words and the output is an ordered sentence, together with its syntactic analysis (either CCG derivation or dependency tree, depending on the grammar formalism being used). Given an input, our system searches for the highestscored output, according to a syntax-based discriminative model. One advantage of this formulation of the reordering problem, which can perhaps be thought of as a “pure” text realization task, is that systems for solving it are easily evaluated, because all that is required is a set of sentences for reordering and a standard evaluation metric such as BLEU (Papineni et al. 2002). However, one potential criticism of the “pure” problem is that it is unclear how it relates to real realization tasks, since in practice (e.g., in statistical machine translation systems) the input does provide constraints on the possible output orderings. Our general formulation still allows task-specific contraints to be added if appropriate. Hence as a test of the flexibility of our system,
and a demonstration of the applicability of the system to more realistic text generation scenarios, we consider two further tasks for the dependency-based realization system.
The first task considers a variety of input conditions for the dependency-based system, determined by two parameters. The first is whether POS information is provided for each word in the input multi-set. The second is whether syntactic dependencies between the words are provided. The extreme case is when all dependencies are provided, in which case the problem reduces to the tree linearization problem (Filippova and Strube 2009; He et al. 2009). However, the input can also lie between the two extremes of no- and full-dependency information.
The second task is the NLG 2011 shared task, which provides a further demonstration of the practical utility of our system. The shared task is closer to a real realization scenario, in that lemmas, rather than inflected words, are provided as input. Hence some modifications are required to our system in order that it can perform some word inflection, as well as deciding on the ordering. The shared task data also uses labeled, rather than unlabeled, syntactic dependencies, and so the system was modified to incorporate labels. The final result is that our system gives competitive BLEU scores, compared to the best-performing systems on the shared task.
The structured prediction problem we solve is a very hard problem. Due to the use of syntax, and the search for a sentence together with a single CCG derivation or dependency tree, the search space is exponentially larger than the n-gram word permutation problem. No efficient algorithm exists for finding the optimal solution. Kay (1996) recognized the computational difficulty of chart-based generation, which has many similarities to the problem we address in his seminal paper. We tackle the high complexity by using learning-guided best-first search, exploring a small path in the whole search space, which contains the most likely structures according to the discriminative model. One of the contributions of this article is to introduce, and provide a discriminative solution to, this difficult structured prediction problem, which is an interesting machine learning problem in its own right.
This article is based on, and significantly extends, three conference papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012; Zhang 2013). It includes a more detailed description and discussion of our guided-search approach to syntax-based word ordering, bringing together the CCG- and dependency-based systems under one unified framework. In addition, we discuss the limitations of our previous work, and show that a better model can be developed through scaling of the feature vectors. The resulting model allows fair comparison of constituents of different sizes, and enables the learning algorithms to expand negative examples during training, which leads to significantly improved results over our previous work. The competitive results on the NLG 2011 shared task data are new for this article, and demonstrate the applicability of our system to more realistic text realization scenarios.
The contributions of this article can be summarized as follows: r We address the problem of syntax-based word ordering, drawing attention to this challenging language modeling task and offering a general solution that does not rely on constraints to limit the search space.r We present a novel method for solving the word ordering problem that gives the best reported accuracies to date on the standard Wall Street Journal data.
r We show how our system can be used with two different grammar formalisms: Combinatory Categorial Grammar and dependency grammar.r We show how syntactic constraints can be easily incorporated into the system, presenting results for the dependency-based system with a range of input conditions.r We demonstrate the applicability of the system to more realistic text realization scenarios by obtaining competitive results on the NLG 2011 shared task data, including performing some word inflection as part of a joint system that also performs word reordering.r More generally, we propose a learning-guided, best-first search algorithm for application of discriminative models to extremely large search spaces containing structures of varying sizes. This method could be applied to other complex structured prediction tasks in NLP and machine learning.","2 overview of the search and training algorithms :In this section, the CCG-based system is used to describe the search and training algorithms. However, the same approach can be used for the dependency-based system, as described in Section 4: Instead of building hypotheses by applying CCG rules in a bottom-up manner, the dependency-based system creates dependency links between words.
Given a bag of words, the goal is to put them into an ordered sentence that has a plausible CCG derivation. The search space of the decoding problem consists of all possible CCG derivations for all possible word permutations, and the search goal is to find the highest-scored derivation in the search space. This is an NP-hard problem, as mentioned in the Introduction. We apply learning-guided search to address the high complexity. The intuition is that, because the whole search space cannot be exhausted in order to find the optimal solution, we choose to explore a small area in the search space. A statistical model is used to guide the search, so that only a small portion of the search space containing the most plausible hypotheses is explored.
One natural choice for the decoding algorithm is best-first search, which uses an agenda to order hypotheses, and expands the highest-scored hypothesis on the agenda at each step. The resulting hypotheses after each hypothesis expansion are put back on the agenda, and the process repeats until a goal hypothesis (a full sentence) is found. This search process is guided by the current scores of the hypotheses, and the search path will contain the most plausible hypotheses if they are scored higher than implausible ones. An alternative to best-first search is A∗ search, which makes use of a heuristic function to estimate future scores. A∗ can potentially be more efficient given an effective heuristic function; however, it is not straightforward to define an admissible and accurate estimate of future scores for our problem, and we leave this research question to future work.
In our formulation of the word ordering problem, a hypothesis is a phrase or sentence together with its CCG derivation. Hypotheses are constructed bottom–up: starting from single words, smaller phrases are combined into larger ones according to CCG rules. To allow the combination of hypotheses, we use an additional structure to store a set of hypotheses that have been expanded, which we call accepted
hypotheses. When a hypothesis from the agenda is expanded, it is combined with all accepted hypotheses in all possible ways to produce new hypotheses. The data structure for accepted hypotheses is similar to that used for best-first parsing (Caraballo and Charniak 1998), and we adopt the term chart for this structure. However, note there are important differences to the parsing problem. First, the parsing problem has a fixed word order and is considerably simpler than the word ordering problem we are tackling. Second, although we use the term chart, the structure for accepted hypotheses is not a dynamic programming chart in the same way as for the parsing problem. In our previous papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012), we applied a set of beams to this structure, which makes it similar to the data structure used for phrase-based MT decoding (Koehn 2010). However, we will show later that this structure is unnecessary when the model has more discriminative power, and a conceptually simpler single beam can be used. We will also investigate the possibility of applying dynamic-programming-style pruning to the chart.
We now give an overview of the training algorithm, which is crucial to both the speed and accuracy of the resulting decoder. CCGBank (Hockenmaier and Steedman 2007) is used to train the model. For each training sentence, the corresponding CCGBank derivation together with all its sub-derivations are treated as gold-standard hypotheses. All other hypotheses that can be constructed from the same bag of words are non-gold hypotheses. From the generation perspective this assumption is too strong, because sentences can have multiple orderings (with multiple derivations) that are both grammatical and fluent. Nevertheless, it is the most feasible choice given the training data available.
The efficiency of the decoding algorithm is dependent on the training algorithm because the agenda is ordered according to the hypothesis scores. Hence, a better model will lead to a goal hypothesis being found more quickly. In the ideal situation, all goldstandard hypotheses are scored higher than all non-gold hypotheses, and therefore only gold-standard hypotheses are expanded before the gold-standard goal hypothesis is found. In this case, the minimum number of hypotheses is expanded and the output is correct. The best-first search decoder is optimal not only with respect to accuracy but also speed. This ideal situation can hardly be met in practice, but it determines the goal of the training algorithm: to find a model that scores gold-standard hypotheses higher than non-gold ones.
Learning-guided search places more challenges on the training of a discriminative model than standard structured prediction problems, for example, CKY parsing for CCG (Clark and Curran 2007b). If we take gold-standard hypotheses as positive training examples, and non-gold hypotheses as negative examples, then the training goal is to find a large separating margin between the scores of all positive examples and all negative examples. For CKY parsing, the highest-scored negative example can be found via optimal Viterbi decoding, according to the current model, and this negative example can be used in place of all negative examples during the updating of parameters. In contrast, our best-first search algorithm cannot find an output in reasonable time unless a good model has already been trained, and therefore we cannot run the decoding algorithm in the standard way during training. In our previous papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012), we proposed an approximate online training algorithm, which forces positive examples to be kept in the hypothesis space without being discarded, and prevents the expansion of negative examples during the training process (so that the hypothesis space does not get too large). This algorithm ensures training efficiency, but greatly limits the space of negative examples that is explored during training (and hence fails to replicate the conditions experienced at test
time). In this article, we will show that, with an improved scoring model, it is possible to expand negative examples, which leads to improved performance.
A second and more subtle challenge for our training problem is that we need a stronger model for learning-guided search than for dynamic programming (DP)– based search, such as CKY decoding. For CKY decoding, the model is used to compare hypotheses within each chart cell, which cover the same input words. In contrast, for the best-first search decoder, the model is used to order hypotheses on the agenda, which can cover different numbers of words. It needs much stronger discriminating power, so that it can determine whether a single-word phrase is better than, say, a 40-word sentence. In this article we use scaling of the hypothesis scores by size, so that hypotheses of different sizes can be fairly compared. We also find that, with this new approach, negative examples can be expanded during training and a single beam applied to the chart, resulting in a conceptually simpler and more effective training algorithm and decoder.","3 ccg-based word ordering : We were motivated to use CCG as one of the grammar formalisms for our syntax-based realization system because of its successful application to a number of related tasks, such as wide-coverage parsing (Hockenmaier 2003; Clark and Curran 2007b; Auli and Lopez 2011), semantic parsing (Zettlemoyer and Collins 2005), wide-coverage semantic analysis (Bos et al. 2004), and generation itself (Espinosa, White, and Mehay 2008). The grammar formalism has been described in detail in those papers, and so here we provide only a short description.
CCG (Steedman 2000) is a lexicalized grammar formalism that associates words with lexical categories. Lexical categories are detailed grammatical labels, typically expressing subcategorization information. During CCG parsing, and during our search procedure, categories are combined using CCG’s combinatory rules. For example, a verb phrase in English (S\NP) can combine with an NP to its left, in this case using the combinatory rule of (backward) function application:
NP S\NP⇒ S
In addition to binary rule instances, such as this one, there are also unary rules that operate on a single category in order to change its type. For example, forward typeraising can change a subject NP into a complex category looking to the right for a verb phrase:
NP⇒ S/(S\NP)
Such a type-raised category can then combine with a transitive verb type using the rule of forward composition:
S/(S\NP) (S\NP)/NP⇒ S/NP
Following Fowler and Penn (2010), we extract the grammar by reading rule instances directly from the derivations in CCGbank (Hockenmaier and Steedman 2007), rather than defining the combinatory rule schema manually as in Clark and Curran
(2007b). Hence the grammar we use can be thought of as a context-free approximation to the mildly content sensitive grammar arising from the use of generalized composition rules (Weir 1988). Hockenmaier (2003) contains a detailed description of the grammar that is obtained in this way, including the various unary type-changing rules, as well as additional rules needed to deal with naturally occurring text, such as punctuation rules. For the rest of this article, the term edge is used to refer to a hypothesis in the decoding algorithm. An edge corresponds to a sentence or phrase with a CCG derivation. Edges are built bottom–up, starting from leaf edges, which are constructed by assigning possible lexical categories to input words. Each leaf edge corresponds to an input word with a particular lexical category. Two existing edges can be combined if there exists a CCG rule (extracted from CCGbank, as described earlier) that combines their category labels, and if they do not contain the same input word more times than its total count in the input. The resulting edge is assigned a category label according to the CCG rule, and covers the concatenated surface strings of the two sub-edges in their order of combination. New edges can also be built by applying unary rules to a single existing edge. We define a goal edge as an edge that covers all input words.
Two edges are equivalent if they have the same surface string and identical CCG derivations. Edge equivalence is used for comparison with gold-standard edges. Two edges are DP-equivalent when they have the same DP-signature. Based on the feature templates in Table 1, we define the DP-signature of an edge as the CCG category at the
root of its derivation, the head word associated with the root category, and the multiset of words it contains, together with the word and POS bigrams on either side of its surface string. Edges are built bottom–up from input words or existing edges. If we treat the assignment of lexical categories to input words and the application of unary and binary CCG rules to existing edges as edge-building actions, the structure of an edge can be defined recursively as the sub-structure resulting from its top action plus the structure of its sub-edges (if any), as shown in Figure 1. Here the top action of an edge refers to the most recent action that has been applied to build the edge.
In our previous papers we used a global linear model to score edges, where the score of an edge e is defined as:
f (e) = Φ(e) · θ⃗
Φ(e) represents the feature vector of e and θ⃗ is the parameter vector of the model. Similar to the structure of e, the feature vector Φ(e) can be defined recursively:
Φ(e) = (∑
es∈e Φ(es)
) +ϕ(e)
= ( ∑
ers∈re ϕ(ers)
) +ϕ(e)
In this equation, es ∈ e represents a sub-edge of e. Leaf edges do not have any subedges. Unary-branching edges have one sub-edge, and binary-branching edges have two sub-edges. ers ∈r e represents a (strictly) recursive sub-edge of e. The feature vector ϕ(e) represents the structure of the top action of e; it is extracted according to the feature templates in Table 1. Example instances of the feature templates are given according to the example string and CCG derivation in Figure 2. For leaf edges, ϕ(e) includes information about the lexical category label; for unary-branching edges, ϕ(e) includes information from the unary rule; for binary-branching edges, ϕ(e) includes information from the binary rule, and additionally the token, POS, and lexical category bigrams and trigrams that result from the surface string concatenation of its sub-edges.
By the given definition of Φ(e), f (e), the score of edge e, can be computed recursively as e is built during the decoding process:
f (e) = Φ(e) · θ⃗ = ((∑
es∈e Φ(es)
) +ϕ(e) ) · θ⃗
= (∑
es∈e Φ(es) · θ⃗
) +ϕ(e) · θ⃗
= (∑
es∈e f (es)
) +ϕ(e) · θ⃗
When the top action is applied, the score of f (e) is computed as the sum of f (es) (for all es ∈ e) plus ϕ(e) · θ⃗.
An important aspect of the scoring model is that it is used to compare edges with different sizes during decoding. The size of an edge can be measured in terms of the number of words it contains, or the number of syntax rules in its structure. We define the size of an edge as the number of recursive sub-edges in the edge plus one (e.g., the size of a leaf edge is 1), which is equivalent to the number of actions (i.e., lexical category assignment for leaf edges, and rule application for unary/binary edges) that have been applied to construct the edge. Edges with different sizes can have significantly different numbers of features, which can make the training of a discriminative linear model more difficult. Note that it is common in structured prediction problems for feature vectors to have slightly different sizes because of variant feature instantiation conditions. In CKY parsing, for example, constituents with different numbers of unary rules can be kept in the same chart cell and compared with each other, provided that they cover the same span in the input. In our case, however, the sizes of two feature vectors under comparison can be very different indeed, since a leaf edge with one word can be compared with an edge over the entire input sentence.
In our previous papers we observed empirical convergence of online learning using this linear model, and obtained competitive results. However, as explained in Section 2, only positive examples were expanded during training, and the expansion of negative examples led to non-convergence and made online training infeasible. In this article, in order to increase the discriminating power of the model and to make use of negative examples during training, we apply length normalization to the scoring function, so that the score of an edge is independent of its size. To achieve this, we scale the original
linear model score by the number of recursive sub-edges in the edge plus one. For a given edge e, the new score f̂ (e) is defined as:
f̂ (e) = f (e) |e|
= Φ(e) · θ⃗ |e|
=
∑ ers∈re ϕ(e r s) · θ⃗+ϕ(e) · θ⃗
|{ers ∈r e}+ 1|
In the equation, |e| represents the size of e, which is equal to the number of actions that have been applied when e is constructed. By dividing the score f (e) by the size of e, the score f̂ (e) represents an averaged value of ϕ(ers) · θ⃗ (ers ∈r e) and ϕ(e) · θ⃗, averaged by the number of recursive sub-edges plus one (i.e., the total actions), and is independent of the size of e. Given normalized feature vectors, the training of the parameter vector θ⃗ needs to be adjusted correspondingly, which will be discussed subsequently. The decoding algorithm takes a multi-set of input words, turns them into a set of leaf edges, and searches for a goal edge by repeated expansion of existing edges. For bestfirst decoding, an agenda and a chart are used. The agenda is a priority queue on which edges to be expanded are ordered according to their current scores. The chart is a fixedsize beam used to record a limited number of accepted edges. During initialization, leaf edges are generated by assigning all possible lexical categories to each input word, before they are put on the agenda. During each step in the decoding process, the highestscored edge on the agenda is popped off and expanded. If it is a goal edge, it is returned as the output, and the decoding finishes. Otherwise it is extended with unary rules, and combined with existing edges in the chart, using binary rules to produce new edges. The resulting edges are scored and put on the agenda, and the original edge is put into the chart. The process repeats until a goal edge is found, or a timeout limit is reached.
For the timeout case, a default output is produced by greedily combining existing edges in the chart in descending order of size. In particular, edges in the chart are sorted by size, and the largest is taken as the current default output. Then the sorted list is traversed, with an attempt to greedily concatenate the current edges in the list to the right of the current default output. If the combination is not allowed (i.e., the two edges contain some input words more times than its count in the input), the current edge is discarded. Otherwise, the current default output is updated.
In our previous papers we used a set of beams for the chart, each storing a certain number of highest-scored edges that cover a particular number of words. This structure is similar to the chart used for phrase-based SMT decoding. The main reason for the multiple beams is the non-comparability of edges in different beams, which can have feature vectors of significantly different sizes. In this article, however, our chart is a single beam structure containing the top-scored accepted edges. This simple data structure is enabled by the use of the scaled linear model, and leads to comparable accuracies to the multiple-beam chart. In addition to its simplicity, it also fits well with the use of agenda-based search, because edges of different sizes will ultimately be compared with each other on the agenda.
We apply DP-style pruning to the chart, keeping only the highest-scored edge among those that have the same DP-signature. During decoding, before a newly constructued edge e is put into the chart, the chart is examined to check whether it contains an existing edge e0 with the same DP-signature as e. If such an edge exists, it is popped off the chart and compared with the newly constructed edge e, with the higher scored edge ẽ being put into the chart and the lower scored edge e′ being discarded. If the newly constructed edge e is not discarded, then we expand e to generate new edges.
It is worth noting that, in this case, a new edge that results from the expansion of e can have DP-equivalent edges in the agenda or the chart, which had been generated by expansion of its DP-equivalent predecessor e′ = e0. Putting such new edges on the agenda will result in the system keeping multiple edges with the same signature. However, because applying DP-style pruning to the agenda requires updating the whole agenda, and is computationally expensive, we choose to tolerate such DP-equivalent duplications in the agenda.
Pseudocode for the decoder is shown as Algorithm 1. INITAGENDA returns an initialized agenda with all leaf edges. INITCHART returns a cleared chart. TIMEOUT
Algorithm 1 The decoding algorithm. a← INITAGENDA( ) c← INITCHART( ) while not TIMEOUT( ) do
new← [] e← POPBEST(a) if GOALTEST(e) then
return e end if (e′, ẽ)← DPCHARTPRUNE(c, e) if e′ is e then
continue end if for e′ ∈ UNARY(e, grammar) do
APPEND(new, e′) end for for ẽ ∈ c do
if CANCOMBINE(e, ẽ, grammar) then e′← BINARY(e, ẽ, grammar) APPEND(new, e′) end if if CANCOMBINE(ẽ, e, grammar) then
e′← BINARY(ẽ, e, grammar) APPEND(new, e′)
end if end for for e′ ∈ new do
ADD(a, e′) end for ADD(c, e)
end while
returns true if the timeout limit has been reached, and false otherwise. POPBEST pops the top edge from the agenda and returns the edge. GOALTEST takes an edge and returns true if and only if the edge is a goal edge. DPCHARTPRUNE takes an edge e and checks whether there exists in the chart an edge e0 that is DP-equivalent to e. In case e0 exists, it is popped off the chart and compared with e, with the lower scored edge e′ being discarded, and the higher scored edge ẽ being put into the chart. The function returns the pair e′ and ẽ. CANCOMBINE checks whether two edges can be combined in a given order. Two edges can be combined if they do not contain an overlapping word (i.e., they do not contain a word more times than its count in the input), and their categories can be combined according to the CCG grammar. ADD inserts an edge into the agenda or the chart. In the former case, it is placed into the priority queue according to its score, and, in the latter case, the lowest scored edge in the beam is pruned when the chart is full. We begin by introducing the training algorithm of our previous papers, shown in Algorithm 2, which has the same fundamental structure as the training algorithm of this article but is simpler. The algorithm is based on the decoder, where an agenda is used as a priority queue of edges to be expanded, and a set of accepted edges is kept in a fixed-size chart. The functions INITAGENDA, INITCHART, TIMEOUT, POPBEST, GOALTEST, DPCHARTPRUNE, UNARY, CANCOMBINE, and BINARY are identical to those used in the decoding algorithm. GOLDSTANDARD takes an edge and returns true if and only if it is a gold-standard edge. MINGOLD returns the lowest scored goldstandard edge in the agenda. UPDATEPARAMETERS represents the parameter update algorithm. RECOMPUTESCORES updates the scores of edges in the agenda and chart after the model is updated.
Similar to the decoding algorithm, the agenda is intialized using all possible leaf edges. During each step, the edge e on top of the agenda is popped off. If it is a goldstandard edge, it is expanded in exactly the same way as in the decoder, with the newly generated edges being put on the agenda, and e being inserted into the chart. If e is not a gold-standard edge, we take it as a negative example e−, and take the lowest scored gold-standard edge on the agenda e+ as a positive example, in order to make an update to the parameter vector θ⃗. Note that there must exist a gold-standard edge in the agenda, which can be proved by contradiction.1
The two edges e+ and e− used to perform a model update can be radically different. For example, they may not cover the same words, or even the same number of words. This is different from online training for CKY parsing, for which both positive and negative examples used to adjust parameter vectors reside in the same chart cell, and cover the same sequence of words. The training goal of a typical CKY parser (Clark and Curran 2007a, 2007b) is to find a large separation margin between feature vectors of different derivations of the same sentence, which have comparable sizes. Our goal is to score all gold-standard edges higher than all non-gold edges regardless of their size, which is a more challenging goal. After updating the parameters, the scores of the agenda edges above and including e−, together with all chart edges, are updated, and e− is discarded before the start of the next processing step.
1 An example proof can be based on induction, where the basis is that the agenda contains all gold leaf edges at first, and the induction step is based on edge combination.
Algorithm 2 The training algorithm of our previous papers. 1: a← INITAGENDA( ) 2: c← INITCHART( ) 3: while not TIMEOUT( ) do 4: new← [] 5: e← POPBEST(a) 6: if GOLDSTANDARD(e) and GOALTEST(e) then 7: return e 8: end if 9: if not GOLDSTANDARD(e) then 10: e−← e 11: e+←MINGOLD(a) 12: UPDATEPARAMETERS(e+, e−) 13: RECOMPUTESCORES(a, c) 14: continue 15: end if 16: (e′, ẽ)← DPCHARTPRUNE(c, e) 17: if e′ is e then 18: continue 19: end if 20: for e′ ∈ UNARY(e, grammar) do 21: APPEND(new, e′) 22: end for 23: for ẽ ∈ c do 24: if CANCOMBINE(e, ẽ, grammar) then 25: e′← BINARY(e, ẽ, grammar) 26: APPEND(new, e′) 27: end if 28: if CANCOMBINE(ẽ, e, grammar) then 29: e′← BINARY(ẽ, e, grammar) 30: APPEND(new, e′) 31: end if 32: end for 33: for e′ ∈ new do 34: ADD(a, e′) 35: end for 36: ADD(c, e) 37: end while
One way of viewing the training process is that it pushes gold-standard edges towards the top of the agenda, and, crucially, pushes them above non-gold edges (Zhang and Clark 2011). Given a positive example e+ and a negative example e−, a perceptronstyle update is used to penalize the score for ϕ(e−) and reward the score of ϕ(e+):
θ⃗← θ⃗0 +ϕ(e+)−ϕ(e−)
Here θ⃗0 and θ⃗ denote the parameter vectors before and after the update, respectively. This method proved effective empirically (Zhang and Clark 2011), but it did not
converge well when an n-gram language model was integrated into the system (Zhang, Blackwood, and Clark 2012).
Hence we applied an alternative method for score updates that proved more effective than the perceptron update and enabled the incorporation of a large-scale language model (Zhang, Blackwood, and Clark 2012). This method treats parameter update as finding a separation between gold-standard and non-gold edges. Given a positive example e+ and a negative example e−, we make a minimum update to the parameters so that the score of e+ is higher than that of e− by a margin of 1:
θ⃗← arg min θ⃗′ ∥ θ⃗′ − θ⃗0 ∥, such that Φ(e+)θ⃗′ − Φ(e−)θ⃗′ ≥ 1
The update is similar to the parameter update of online large-margin learning algorithms, such as 1-best MIRA (Crammer et al. 2006), and has a closed-form solution:
θ⃗← θ⃗0 + f (e−)− f (e+) + 1 ∥ Φ(e+)− Φ(e−) ∥2
( Φ(e+)− Φ(e−) ) This online learning method proved more effective than the perceptron algorithm empirically, but still has an important shortcoming in that it did not provide competitive results when allowing the expansion of negative examples during training, which can potentially improve the discriminative model (since expanding negative examples can result in a more representative sample of the search space). We address this issue by introducing a scaled linear model in this article, which, when combined with the expansion of negative examples, significantly improves performance. We apply the same online large-margin training principle; however, the parameter update has to be adjusted for the scaled linear model. In particular, the new goal is to find a separation between f̂ (e+) and f̂ (e−) instead of f (e+) and f (e−), for which the optimization corresponding to the parameter update becomes:
θ⃗← arg min θ⃗′
∥ θ⃗′ − θ⃗0 ∥, such that Φ(e+)θ⃗′ |e+| − Φ(e−)θ⃗ ′ |e−| ≥ 1
where θ⃗0 and θ⃗ represent the parameter vectors before and after the update, respectively. The equation has a closed-form solution:
θ⃗← θ⃗0 + f̂ (e−)− f̂ (e+) + 1 ∥ Φ(e+ )|e+| − Φ(e− ) |e−| ∥ 2
(Φ(e+) |e+| − Φ(e−)|e−| )
Pseudocode for the new training algorithm of this article is shown in Algorithm 3, where MAXNONGOLD returns the highest-scored non-gold edge in the chart. In addition to the aforementioned difference in parameter updates, new code is added to perform additional updates when gold-standard edges are removed from the chart. In our previous work, parameter updates happen only when the top edge from the agenda is not a gold-standard edge. In this article, the expansion of negative training examples will lead to negative examples being put into the chart during training, and hence the possibility of gold-standard edges being removed from the chart. There are two situations when this can happen. First, if a non-gold edge is inserted into the chart,
Algorithm 3 The training algorithm of this article. 1: a← INITAGENDA( ) 2: c← INITCHART( ) 3: while not TIMEOUT( ) do 4: new← [] 5: e← POPBEST(a) 6: if GOLDSTANDARD(e) and GOALTEST(e) then 7: return e 8: end if 9: if not GOLDSTANDARD(e) then 10: e−← e 11: e+←MINGOLD(a) 12: UPDATEPARAMETERS(e+, e−) 13: RECOMPUTESCORES(a, c) 14: end if 15: (e′, ẽ)← DPCHARTPRUNE(c, e) 16: if GOLDSTANDARD(e′) then 17: UPDATEPARAMETERS(e′, ẽ) 18: REMOVE(c, ẽ) 19: ADD(c, e′) 20: else 21: if e′ is e then 22: continue 23: end if 24: end if 25: for e′ ∈ UNARY(e, grammar) do 26: APPEND(new, e′) 27: end for 28: for ẽ ∈ c do 29: if CANCOMBINE(e, ẽ, grammar) then 30: e′← BINARY(e, ẽ, grammar) 31: APPEND(new, e′) 32: end if 33: if CANCOMBINE(ẽ, e, grammar) then 34: e′← BINARY(ẽ, e, grammar) 35: APPEND(new, e′) 36: end if 37: end for 38: for e′ ∈ new do 39: ADD(a, e′) 40: end for 41: e′← ADD(c, e) 42: if GOLDSTANDARD(e′) then 43: ẽ←MAXNONGOLD(c) 44: UPDATEPARAMETERS(e′, ẽ) 45: ADD(c, e′) 46: end if 47: end while
and there exists a gold-standard edge in the chart with the same DP-signature but a lower score, the gold-standard edge will be removed from the chart because of DP-style pruning (since only the highest-scored edge with the same DP-signature is kept in the chart).
Second, if the chart is full when a non-gold edge is put into the chart (recall that the chart is a fixed-size beam), then the lowest scored edge on the chart will be removed. This edge can be a gold-standard edge. In both the first and second case, a gold-standard edge is pruned as the result of the expansion of a negative example. On the other hand, in order for the gold-standard goal edge to be constructed, all gold-standard edges that have been expanded must remain in the chart. As a result, our training algorithm triggers a parameter update whenever a gold-standard edge is removed from the chart, the scores of all chart edges are updated, and the original pruned gold edge is returned to the chart. The original pruned gold-standard edge is treated as the positive example for the update. For the first situation, the newly inserted non-gold edge with the same DP-signature is taken as the negative example, and will be discarded after the parameter update (with a new score that is lower than the new score of the corresponding goldstandard). In the second situation, the highest-scored non-gold edge in the chart is taken as the negative example, and removed from the chart after the update.
In summary, there are two main differences between Algorithms 2 and 3. First, line 14 in Algorithm 2, which skips the expansion of negative examples, is removed in Algorithm 3. Second, lines 16–20 and 42–46 are added in Algorithm 3, which correspond to the updating of parameters when a gold-standard edge is removed from the chart. In addition, the definitions of UPDATEPARAMETERS are different for the perceptron training algorithm (Zhang and Clark 2011), the large-margin training algorithm (Zhang, Blackwood, and Clark 2012), and the large-margin algorithm of this article, as explained earlier.","4 dependency-based word ordering and tree linearization :As well as CCG, the same approach can be applied to the word ordering problem using other grammar formalisms. In this section, we present a dependency-based word ordering system, where the input is again a multi-set of words with gold-standard POS, and the output is an ordered sentence together with its dependency parse. Except for necessary changes to the edge data structure and edge expansion, the same algorithm can be applied to this task.
In addition to abstract word ordering, our framework can be used to solve a more informed, dependency-based word ordering task: tree linearization (Filippova and Strube 2009; He et al. 2009), a task that is very similar to abstract word ordering from a computational perspective. Both tasks involve the permutation of a set of input words, and are NP-hard. The only difference is that, for tree linearization, full unordered dependency trees are given as input. As a result, the output word permutations are more constrained (under the projectivity assumption), and more information is available for search disambiguation.
Tree linearization can be treated as a special case of word ordering, where a grammar constraint is applied such that the output sentence has to be consistent with the input tree. There is a spectrum of grammar constraints between abstract word ordering (no constraints) and tree linearization (full tree constraints). For example, one constraint might consist of a set of dependency relations between input words, but which do not form a complete unordered spanning tree. We call this word ordering task the partial-tree linearization problem, a task that is perhaps closer to NLP applications
than both the abstract word ordering task and the full tree linearization problem, in the sense that NLG pipelines might provide some syntactic relations between words for the linearization step, but not the full spanning tree.
The main content of this section is based on a conference paper (Zhang 2013), which we extend by using the technique of expanding negative training examples (one of the overall contributions of this article). Given a multi-set of input words W and a set of head-dependent relations H between the words in W, the task is to find an ordered sentence consisting of all the words in W and a dependency tree that contains all the relations in H. If each word in W is given a POS tag and H covers all words in W, then the task is (full-)tree linearization; if not then the task is partial-tree linearization. For partial-tree linearization, a subset of W is given fixed POS tags. In all cases, a word either has exactly one (gold) POS tag, or no POS tags. Similar to the CCG case, edge refers to the data structure for a hypothesis in the decoding algorithm. Here a leaf edge refers to an input word with a POS tag, and a non-leaf edge refers to a phrase or sentence with its dependency tree. Edges are constructed bottom–up, by recursively joining two existing edges and adding an unlabeled dependency link between their head words.
As for the CCG system, edges are scored by a global linear model:
f (e) = Φ(e) · θ⃗|e|
where Φ(e) represents the feature vector of e and θ⃗ is the parameter vector of the model. Table 2 shows the feature templates we use, which are inspired by the rich feature templates used for dependency parsing (Koo and Collins 2010; Zhang and Nivre 2011). In the table, h, m, s, hl, hr, ml, and mr are the indices of words in the newly constructed edge, where h and m refer to the head and dependent of the newly constructed arc, s refers to the nearest sibling of m (on the same side of h), and hl, hr, ml, and mr refer to the left and rightmost dependents of h and m, respectively. WORD, POS, LVAL, and RVAL are maps from indices to word forms, POS, left valencies, and right valencies of words, respectively. Example feature instances extracted from the sentence in Figure 3 are shown in the example column. Because of the non-local nature of some of the feature templates we define, we do not apply DP-style pruning for dependency-based treelinearization. The decoding algorithm is similar to that of the CCG system, where an agenda is a priority queue for edges to expand, and chart is a fixed-size beam for a list of accepted edges. During initialization, input words are assigned possible POS tags, resulting in a set of leaf edges that are put onto the agenda. For words with POS constraints, only
the allowed POS tag is assigned. For unconstrained words, we assign all possible POS tags according to a tag dictionary compiled from the training data, following standard practice for POS-tagging (Ratnaparkhi 1996).
When an edge is expanded, it is combined with all edges in the chart in all possible ways to generate new edges. Two edges can be combined by concatenation of the surface strings in both orders and, in each case, constructing a dependency link between their heads in two ways (corresponding to the two options for the head of the new link). When there is a head constraint on the dependent word, a dependency link can be constructed only if it is consistent with the constraint. This algorithm implements abstract word ordering, partial-tree linearization, and full tree linearization—all generalized word ordering tasks—in a unified method.
Pseudocode for the decoder is shown as Algorithm 4. Many of the functions have the same definition as for Algorithm 1: INITAGENDA, INITCHART, TIMEOUT, POPBEST, GOALTEST, ADD. CANCOMBINE checks whether two edges do not contain an overlapping word (i.e., they do not contain a word more times than its count in the input); unlike the CCG case, all pairs of words are allowed to combine according to the dependency model. COMBINE creates a dependency link between two words, with the word order determined by the order in which the arguments are supplied to the function, and the head coming from either the first (HeadLeft) or second (HeadRight) argument (so there are four combinations considered and COMBINE is called four times in Algorithm 4). As for the CCG system, an online large-margin learning algorithm based on the decoding process is used to train the model. At each step, the expanded edge e is compared with the gold standard. If it is a gold edge, decoding continues; otherwise e is taken as a negative example e− and the lowest-scored gold edge in the agenda is taken
Algorithm 4 The decoding algorithm for partial-tree linearization. a← INITAGENDA( ) c← INITCHART( ) while not TIMEOUT( ) do
new← [] e← POPBEST(a) if GOALTEST(e) then
return e end if for ẽ ∈ c do
if CANCOMBINE(e, ẽ) then e′← COMBINE(e, ẽ, HeadLeft) APPEND(new, e′) e′← COMBINE(e, ẽ, HeadRight) APPEND(new, e′) end if if CANCOMBINE(ẽ, e) then
e′← COMBINE(ẽ, e, HeadLeft) APPEND(new, e′) e′← COMBINE(ẽ, e, HeadRight) APPEND(new, e′)
end if end for for e′ ∈ new do
ADD(a, e′) end for ADD(c, e)
end while
as a positive example e+, and a parameter update is executed (repeated here from Section 3.4):
θ⃗← θ⃗0 + f̂ (e−)− f̂ (e+) + 1 ∥ Φ(e+ )|e+| − Φ(e− ) |e−| ∥ 2
(Φ(e+) |e+| − Φ(e−)|e−| )
The training process is essentially the same as in Algorithm 3, but with the CCG grammar and model replaced with the dependency-based grammar and model.
In our conference paper describing the earlier version of the dependency-based system (Zhang 2013), the decoding step is finished immediately after the parameter update; in this article we expand the negative example, as in Algorithm3, putting it onto the chart and thereby exploring a larger part of the search space (in particular that part containing negative examples). Our later experiments show that this method yields improved results, consistent with the CCG system.","5 experiments :We use CCGBank (Hockenmaier and Steedman 2007) and the Penn Treebank (Marcus, Santorini, and Marcinkiewicz 1993) for CCG and dependency data, respectively. CCGbank is the CCG version of the Penn Treebank. Standard splits were used for both: Sections 02–21 for training, Section 00 for development, and Section 23 for the final test. Table 3 gives statistics for the Penn Treebank.
For the CCG experiments, original sentences from CCGBank are transformed into bags of words, with sequence information removed, and passed to our system as input data. The system outputs are compared to the original sentences for evaluation. Following Wan et al. (2009), we use the BLEU metric (Papineni et al. 2002) for string comparison. Although BLEU is not the perfect measure of fluency or grammaticality, being based on n-gram precision, it is currently widely used for automatic evaluation and allows us to compare directly with existing work (Wan et al. 2009). Note also that one criticism of BLEU for evaluating machine translation systems (i.e., that it can only register exact matches between the same words in the system and reference translation), does not apply here, because the system output always contains the same words as the original reference sentence. For the dependency-based experiments, gold-standard dependency trees were derived from bracketed sentences in the treebank using the Penn2Malt tool.2
For fair comparison with Wan et al. (2009), we keep base NPs as atomic units when preparing the input. Wan et al. used base NPs from the Penn Treebank annotation, and we follow this practice for the dependency-based experiments. For the CCG experiments we extract base NPs from CCGBbank by taking as base NPs those NPs that do not recursively contain other NPs. These base NPs mostly correspond to the base NPs from the Penn Treebank: In the training data, there are 242,813 Penn Treebank base NPs with an average size of 1.09, and 216,670 CCGBank base NPs with an average size of 1.19. The plots in Figure 4 show the development test scores of three CCG models by the number of training iterations. The three curves represent the scaled model of this article, the online large-margin model from Zhang, Blackwood, and Clark (2012), and the perceptron model from Zhang and Clark (2011), respectively. For each curve, the BLEU score generally increases as the number of training iterations increases, until it reaches its maximum at a particular iteration. We use the number of training iterations
2 http://w3.msi.vxu.se/~nivre/research/Penn2Malt.html.
that gives the best development test scores for the training of our model when testing on the test data.
Another way to observe the convergence of training is to measure the training times for each iteration at different numbers of iterations. The per-iteration training times for the large-margin and the scaled CCG models are shown in Figure 5. For each model, the training time for each iteration decreases as the number of training iterations increases, reflecting the convergence of learning-guided search. When the model gets better, fewer non-gold hypotheses are expanded before gold hypotheses, and hence it takes less time for the decoder to find the gold goal edge. Figure 6 shows the corresponding curve for dependency-based word ordering, with similar observations.
Because of the expanding of negative examples, the systems of this article took more time to train than those of our previous conference papers. However, the convergence
rate is also faster when negative training examples are expanded, as demonstrated by the rate of speed improvement as the number of training iterations increases. The training times of the perceptron algorithm are close to those of the large-margin algorithm, and hence are omitted from Figures 5 and 6. The new model gives the best development test scores, as shown in Figure 4. The next section investigates the effects of two of the innovations of this article: use of negative examples during training and the scaling of the model by hypothesis size. Table 4 shows a set of CCG development experiments to measure the effect of the scaled model and the expansion of negative examples during training. With the standard linear model (Zhang, Blackwood, and Clark 2012) and no expansions of negative examples, our system obtained a BLEU score of 39.04. The scaled model improved the BLEU score by 1.41 BLEU points to 40.45, and the expansion of negative examples gave a further improvement of 3.02 BLEU points.
These CCG development experiments show that the expansion of negative examples during training is an important factor in achieving good performance. When no negative examples are expanded, the higher score of the scaled linear model demonstrates the effectiveness of fair comparison between edges with different sizes. However,
it is a more important advantage of the scaled linear model that it allows the expansion of negative examples during training, which was not possible with the standard linear model. In the latter case, training failed to converge when negative examples were expanded, reflecting the limitations of the standard linear model in separating the training data. Similar results were found for dependency-based word ordering, where the best development BLEU score improved from 44.71 (Zhang 2013) to 46.44 with the expansion of negative training examples. Figure 7 shows the BLEU scores for the CCG system on the development data when the timeout limit for decoding a single sentence is set to 5 sec, 10 sec, 15 sec, 20 sec, 30 sec, 40 sec, 50 sec, and 60 sec, respectively. The timeout was applied during decoding at test time. The scaled model with negative training examples was used for this set of experiments, and the same model was used for all timeout settings. The results demonstrate that better outputs can be recovered given more search time, which is expected for a time-constrained best-first search framework. Recall that output is created greedily by combining the largest available edges, when the system times out. Similar results were obtained with the dependency-based system of Zhang (2013), where the development BLEU scores improved from 42.89 to 43.42, 43.58, and 43.72 when the timeout limit increased from 5 sec to 10 sec, 30 sec, and 60 sec, respectively. The scaled dependency-based model without expansion of negative examples was used in this set of experiments. Example output for sentences in the development set is shown in Tables 5 and 6, grouped by sentence length. The CCG systems of our previous conference papers and this article are compared, all with the timeout value set to 5 sec. All three systems perform relatively better with smaller sentences. For longer sentences, the fluency of
the output is significantly reduced. One source of errors is confusion between different noun phrases, and where they should be positioned, which becomes more severe with increased sentence length and adds to the difficulty in reading the outputs. The system of this article gave observably improved outputs compared with the two other systems. In the previous section, the same input settings were used for both training and testing, and the assumption was made that the input to the system would be a bag of words, with no constraints on the output structure. This somewhat artificial assumption allows a standardized evaluation but, as discussed previously, text generation applications are unlikely to satisfy this assumption and, in practice, the realization problem is likely to be easier compared with our previous set-up. In this section, we simulate practical situations in dependency-based pipelines by measuring the performance of our system using randomly chosen input POS tags and dependency relations. For maximum flexibility, so that the same system can be applied to different input scenarios, our system is trained without input POS tags or dependencies. However, if POS tags and dependencies are made available during testing, they will be used to provide hard constraints on the output (i.e., the output sentence with POS tags and dependencies must contain those in the input). From the perspective of search, input POS tags and dependencies greatly constrain the search space and lead to an easier search problem, with correspondingly improved outputs.
Table 7 shows a set of development results with varying amounts of POS and dependency information in the input. For each test, we randomly sampled a percentage of words for which the gold-standard POS tags or dependencies are given in the input. As can be seen from the table, increased amounts of POS and dependency information in the input lead to higher BLEU scores, and dependencies were more effective than POS tags in determining the word order in the output. When all POS tags and dependencies are given, our constraint-enabled system gave a BLEU score of 76.28.3
Table 8 shows the output of our system for the first nine development test sentences with different input settings. These examples illustrate the positive effect of input dependencies in specifying the outputs. Consider the second sentence as an example. When only input words are given, the output of the system is largely grammatical but nonsensical. With increasing amounts of dependency relations, the output begins to look more fluent, sometimes with the system reproducing the original sentence when all dependencies are given. Table 9 shows the test results of various systems. For the system of this article, we take the optimal setting from the development tests, using the scaled linear model and
3 When all POS tags and dependencies are also provided during training, the BLEU score is reduced to 74.79, showing the value in the system, which can adapt to varying amounts of POS and dependency information in the input at test time.
expansion of negative examples during training. For direct comparison with previous work, the timeout threshold was set to 5 sec. Our new system of this article significantly outperforms all previous systems and achieves the best published BLEU score on this task. It is worth noting that our systems without a language model outperform the system of our 2012 paper using a large-scale language model.
Interestingly, the dependency-based systems performed better than the CCG systems of this article. One of the main reasons is that the CCG systems generated shorter outputs by not finding full spanning derivations for a larger proportion of inputs. Because of the rigidity in combinatory rules, not all hypotheses in the chart can be combined with the hypothesis being expanded, leading to an increased likelihood of full spanning derivations being unreachable. Overall, the CCG system recovered 93.98% of the input words in the test set, and the dependency system recovered 97.71%. The previous sections report evaluations on the task of word ordering, an abstract yet fundamental problem in text generation. One question that is not addressed by these experiments is how the abstract task can be utilized to benefit full text generation, for which more considerations need to be taken into account in addition to word ordering. We investigate this question using the 2011 Generation Challenge shared task data, which provide a common-ground for the evaluation of text generation systems (Belz et al. 2011).
The data are based on the CoNLL 2008 shared task data (Surdeanu et al. 2008), which consist of selected sections of the Penn WSJ Treebank, converted to syntactic dependencies via the LTH tool (Johansson and Nugues 2007). Sections 2–21 are used for training, Section 24 for development, and Section 23 for testing. A small number of sentences from the original WSJ sections are not included in this set. The input format of the shared task is an unordered syntactic dependency tree, with nodes being lemmas, and dependency relations on the arcs. Named entities and hyphenated words are broken into individual nodes, and special dependency links are used to mark them. Information on coarse-grained POS, number, tense, and participle features is given to each node where applicable. The output is a fully ordered and inflected sentence.
We developed a full-text generation system according to this task specification, with the core component being the dependency-based word ordering system of Section 4. In addition to minor engineering details that were required to adapt the system to this new
task, one additional task that the generation system needs to carry out is morphological generation—finding the appropriate inflected form for each input lemma. Our approach is to perform joint word ordering and inflection using the learning-guided search framework, letting one statistical model decide the best order as well as the inflections of ambiguous lemmas. For a lemma, we generate one or more candidate inflections by using a lexicon and a set of inflection rules. Candidate inflections for an input lemma are generated according to the lemma itself and its input attributes, such as the number and tense. Some lemmas are unambiguous, which are inflected before being passed to the word ordering system. For the other lemmas, more than one candidate’s inflections are passed as input words to the word ordering system. To ensure that each lemma occurs only once in the output, a unique ID is given to all the inflections of the same lemma, making them mutually exclusive.
Four types of lemmas need morphological generation, including nouns, verbs, adjectives, and miscellaneous cases. The last category includes a (a or an) and not (not or n‘t), for which the best inflection can be decided only when n-gram information is available. For these lemmas, we pass all possible inflections to the search module. For nouns and adjectives, the inflection is relatively straightforward, since the number (e.g., singular, plural) of a lemma is given as an attribute of the input node, and comparative and superlative adjectives have specific parts of speech. For those cases where the necessary information is not available from the input, all possible inflections are handed over to the search module for further disambiguation. The most ambiguous lemma types are verbs, which can be further divided into be and other verbs. The uniqueness of be is that the inflections for the first and second person can be different. All verb inflections are disambiguated according to the tense and participle attributes of the input node. In addition, for verbs in the present tense, the subject needs to be determined in order to differentiate between third-person singular verbs and others. This can be straightforward when the subject is a noun or pronoun, but can be ambiguous when the subject is a wh-pronoun, in which case the real subject might not be directly identifiable from the dependency tree. We leave all possible inflections of be and other verbs to the word ordering system whenever the ambiguity is not directly solvable from the subject dependency link. Overall, the pre-processing step generates 1.15 inflections for each lemma on average.
For word ordering, the search procedure of Algorithm 4 is applied directly, and the feature templates of Table 2 are used with additional labeled dependency features described subsequently. The main reason that the dependency-based word ordering algorithm can perform joint morphological disambiguation is that it uses rich syntactic and n-gram features to score candidate hypotheses, which can also differentiate between correct and incorrect inflections under particular contexts. For example, an honest person and a honest person can be differentiated by n-gram features, while Tom and Sally is and Tom and Sally are can be differentiated by higher-order dependency features.
In addition to lemma-formed inputs, one other difference between the shared task and the word ordering problem solved by Algorithm 4 is that the former uses labeled dependencies whereas Algorithm 4 constructs unlabeled dependency trees. We address this issue by assigning dependency labels in the construction of dependency links, and applying an extra set of features. The new features are defined by making a duplicate of all the features from Table 2 that contain dir information, and associating each feature in the new copy with a dependency label.
The training of the word ordering system requires fully ordered dependency trees, while references in the shared task data are raw sentences. We perform a pre-processing
step to obtain gold-standard training data by matching the input lemmas to the reference sentence in order to obtain their gold-standard order. More specifically, given a training instance, we generate all candidate inflections for each lemma, resulting in an exponential set of possible mappings between the input tree and the reference sentence. We then prune these mappings bottom–up, assuming that the dependency tree is projective, and therefore that each word dominates a continuous span in the reference. After such pruning, only one correct mapping is found for the majority of the cases. For the cases where more than one mapping is found, we randomly choose one as the gold-standard. There are also instances for which no correct ordering can be found, and these are mostly due to non-projectivity in the shared task data, with a few cases being due to conflicts between our morphological generation system and the shared task data, or inconsistency in the data itself. Out of the 39K training instances, 2.8K conflicting instances are discarded, resulting in 36.2K gold-standard ordered dependency trees.
Table 10 shows the results of our system and the top two participating systems of the shared task. Our system outperforms the STUMABA system by 0.5 BLEU points, and the DCU system by 3.8 BLEU points. More evaluation of the system was published in Song et al. (2014).","6 related work :There is a recent line of research on text-to-text generation, which studies the linearization of dependency structures (Barzilay and McKeown 2005; Filippova and Strube 2007, 2009; He et al. 2009; Bohnet et al. 2010; Guo, Hogan, and van Genabith 2011). On the other hand, Wan et al. (2009) study the ordering of a bag of words without any dependency information given. We generalize the word ordering problem, and formulate it as a task of ordering a multi-set of words, regardless of input syntactic constraints.
Our bottom–up, chart-based generation algorithm is inspired by the line of work on chart-based realization (Kay 1996; Carroll et al. 1999; White 2004, 2006; Carroll and Oepen 2005). Kay (1996) first proposed the concept of chart realization, drawing analogies between realization and parsing of free order languages. He discussed efficiency issues and provided solutions to specific problems. For the task of realization, efficiency improvement has been further investigated (Carroll et al. 1999; Carroll and Oepen 2005). The inputs to these systems are logical forms, which form natural constraints on the interaction between edges. In our case, one constraint that has been leveraged in the dependency system is a projectivity assumption—we assume that the dependents of a word must all have been attached before the word is attached to its head word, and that spans do not cross during combination. In addition, we assume that the right dependents of a word must have been attached before a left dependent of the word is attached. This constraint avoids spurious ambiguities. The projectivity assumption
is an important basis for the feasibility of the dependency system; it is similar to the chunking constraints of White (2006) for CCG-based realization.
White (2004) describes a system that performs CCG realization using best-first search. The search process of our algorithm is similar to that work, but the input is different: logical forms in the case of White (2004) and bags of words in our case. Further along this line, Espinosa, White, and Mehay (2008) also describe a CCG-based realization system, applying “hypertagging”—a form of supertagging—to logical forms in order to make use of CCG lexical categories in the realization process. White and Rajkumar (2009) further use perceptron reranking on n-best realization output to improve the quality.
The use of perceptron learning to improve search has been proposed in guided learning for easy-first search (Shen, Satta, and Joshi 2007) and LaSO (Daumé and Marcu 2005). LaSO is a general framework for various search strategies. Our learning algorithm is similar to LaSO with best-first inference, but the parameter updates are different. In particular, LaSO updates parameters when all correct hypotheses are lost, but our algorithm makes an update as soon as the top item from the agenda is incorrect. Our algorithm updates the parameters using a stronger precondition, because of the large search space. Given an incorrect hypothesis, LaSO finds the corresponding gold hypothesis for a perceptron update by constructing its correct sibling. In contrast, our algorithm takes the lowest scored gold hypothesis currently in the agenda to avoid updating parameters for hypotheses that may have not been constructed.
Our parameter update strategy is closer to the guided learning mechanism for the easy-first algorithm of Shen, Satta, and Joshi (2007), which maintains a queue of hypotheses during search, and performs learning to ensure that the highest-scored hypothesis in the queue is correct. However, in easy-first search, hypotheses from the queue are ranked by the score of their next action, rather than the hypothesis score. Moreover, Shen, Satta, and Joshi use aggressive learning and regenerate the queue after each update, but we perform non-aggressive learning, which is faster and is more feasible for our large and complex search space. Similar methods to Shen, Satta, and Joshi (2007) have also been used in Shen and Joshi (2008) and Goldberg and Elhadad (2010).
Another framework that closely integrates learning and search is SEARN (Daumé, Langford, and Marcu 2009), which addresses structured prediction problems that can be transformed into a series of simple classification tasks. The transformation is akin to greedy search in the sense that the complex structure is constructed by sequential classification decisions. The key problem that SEARN addresses is how to learn the tth decision based on the previous t− 1 decisions, so that the overall loss in the resulting structure is minimized. Similar to our framework, SEARN allows arbitrary features. However, SEARN is more oriented to greedy search, optimizing local decisions. In contrast, our framework is oriented to best-first search, optimizing global structures.
Learning and search also interact with each other in a global discriminative learning and beam-search framework for incremental structured prediction (Zhang and Clark 2011). In this framework, an output is constructed incrementally by a sequence of transitions, while a beam is used to record the highest scored structures at each step. Online training is performed based on the search process, with the objective function being the margin between correct and incorrect structures. The method involves an earlyupdate strategy, which stops search and updates parameters immediately when the gold structure falls out of the beam during training. It was first proposed by Collins and Roark (2004) for incremental parsing, and later gained popularity in the investigations of many NLP tasks, including POS-tagging (Zhang and Clark 2010), transition-based dependency parsing (Zhang and Clark 2008; Huang and Sagae 2010), and machine
translation (Liu 2013). Huang, Fayong, and Guo (2012) propose a theoretical analysis to the early-update training strategy, pointing out that it is a type of training method that fixes score violations in inexact search. When the score of a gold-standard structure is lower than that of a non-gold structure, a violation exists. Our parameter udpate strategy in this article can also be treated as a mechanism for violation fixing.","7 conclusion :We investigated the general task of syntax-based word ordering, which is a fundamental problem for text generation, and a computationally very expensive search task. We provide a principled solution to this problem using learning-guided search, a framework that is applicable to other NLP problems with complex search spaces. We compared different methods for parameter updates, and showed that a scaled linear model gave the best results by allowing better comparisons between phrases of different sizes, increasing the separability of hypotheses and enabling the expansion of negative examples during training.
We formulate abstract word ordering as a spectrum of tasks with varying input specificity, from “pure” word ordering without any syntactic information to fullyinformed word ordering with a complete unordered dependency tree given. Experiments show that our proposed method can effectively use available input constraints in generating output sentences.
Evaluation on the NLG 2011 shared task data shows that our system can be successfully applied to a more realistic application scenario, in particular one where some dependency constraints are provided in the input and word inflection is required as well as word ordering. Additional tasks that may be required in a practical text generation scenario include word selection, including the determination of content words and generation of function words. The joint modeling solution that we have proposed for word ordering and inflection could also be adopted for word selection, although the search space is greatly increased when the words themselves need deciding, particularly content words.",,"Word ordering is a fundamental problem in text generation. In this article, we study word ordering using a syntax-based approach and a discriminative model. Two grammar formalisms are considered: Combinatory Categorial Grammar (CCG) and dependency grammar. Given the search for a likely string and syntactic analysis, the search space is massive, making discriminative training challenging. We develop a learning-guided search framework, based on best-first search, and investigate several alternative training algorithms. The framework we present is flexible in that it allows constraints to be imposed on output word orders. To demonstrate this flexibility, a variety of input conditions are considered. First, we investigate a “pure” word-ordering task in which the input is a multi-set of words, and the task is to order them into a grammatical and fluent sentence. This task has been tackled previously, and we report improved performance over existing systems on a standard Wall Street Journal test set. Second, we tackle the same reordering problem, but with a variety of input conditions, from the bare case with no dependencies or POS tags specified, to the extreme case where all POS tags and unordered, unlabeled dependencies are provided as input (and various conditions in between). When applied to the NLG 2011 shared task, our system gives competitive results compared with the best-performing systems, which provide a further demonstration of the practical utility of our system.","[{""affiliations"": [], ""name"": ""Yue Zhang""}, {""affiliations"": [], ""name"": ""Stephen Clark""}]",SP:b2d9b8b3028538d65ecf5e5ecf4f428224aff6b1,"[{""authors"": [""Auli"", ""Michael"", ""Adam Lopez.""], ""title"": ""Training a log-linear parser with loss functions via softmax-margin"", ""venue"": ""Proceedings of the 2011 Conference on Empirical Methods in Natural Language"", ""year"": 2011}, {""authors"": [""Barzilay"", ""Regina"", ""Kathleen McKeown.""], ""title"": ""Sentence fusion for multidocument news summarization"", ""venue"": ""Computational Linguistics, 31(3):297\u2013328."", ""year"": 2005}, {""authors"": [""Belz"", ""Anja"", ""Michael White"", ""Dominic Espinosa"", ""Eric Kow"", ""Deirdre Hogan"", ""Amanda Stent.""], ""title"": ""The first surface realisation shared task: Overview and evaluation results"", ""venue"": ""Proceedings of the 13th"", ""year"": 2011}, {""authors"": [""Blackwood"", ""Graeme"", ""Adri\u00e0 de Gispert"", ""William Byrne.""], ""title"": ""Fluency constraints for minimum Bayes-risk decoding of statistical machine translation lattices"", ""venue"": ""Proceedings of the 23rd International"", ""year"": 2010}, {""authors"": [""Bohnet"", ""Bernd"", ""Leo Wanner"", ""Simon Mill"", ""Alicia Burga.""], ""title"": ""Broad coverage multilingual deep sentence generation with a stochastic multi-level realizer"", ""venue"": ""Proceedings of the 23rd International"", ""year"": 2010}, {""authors"": [""Bos"", ""Johan"", ""Stephen Clark"", ""Mark Steedman"", ""James R. Curran"", ""Julia Hockenmaier.""], ""title"": ""Wide-coverage semantic representations from a CCG parser"", ""venue"": ""Proceedings of COLING-04,"", ""year"": 2004}, {""authors"": [""Caraballo"", ""Sharon A."", ""Eugene Charniak.""], ""title"": ""New figures of merit for best-first probabilistic chart parsing"", ""venue"": ""Computational Linguistics, 24:275\u2013298."", ""year"": 1998}, {""authors"": [""Carroll"", ""John"", ""Ann Copestake"", ""Dan Flickinger"", ""Victor Poznanski.""], ""title"": ""An efficient chart generator for (semi-) lexicalist grammars"", ""venue"": ""Proceedings of the 7th European Workshop on Natural Language"", ""year"": 1999}, {""authors"": [""Carroll"", ""John"", ""Stephan Oepen.""], ""title"": ""High efficiency realization for a wide-coverage unification grammar"", ""venue"": ""Proceedings of the Second International Joint Conference on Natural Language Processing, IJCNLP\u201905,"", ""year"": 2005}, {""authors"": [""Chang"", ""Pi-Chuan"", ""Kristina Toutanova.""], ""title"": ""A discriminative syntactic word order model for machine translation"", ""venue"": ""Proceedings of ACL, pages 9\u201316, Prague."", ""year"": 2007}, {""authors"": [""Chiang"", ""David.""], ""title"": ""Hierarchical phrase-based translation"", ""venue"": ""Computational Linguistics, 33(2):201\u2013228."", ""year"": 2007}, {""authors"": [""Clark"", ""Stephen"", ""James R. Curran.""], ""title"": ""Perceptron training for a wide-coverage lexicalized-grammar parser"", ""venue"": ""Proceedings of the ACL 2007 Workshop on Deep Linguistic Processing, pages 9\u201316, Prague."", ""year"": 2007}, {""authors"": [""Clark"", ""Stephen"", ""James R. Curran.""], ""title"": ""Wide-coverage efficient statistical parsing with CCG and log-linear models"", ""venue"": ""Computational Linguistics, 33(4):493\u2013552."", ""year"": 2007}, {""authors"": [""Collins"", ""Michael"", ""Brian Roark.""], ""title"": ""Incremental parsing with the perceptron algorithm"", ""venue"": ""Proceedings of ACL, pages 111\u2013118, Barcelona."", ""year"": 2004}, {""authors"": [""Crammer"", ""Koby"", ""Ofer Dekel"", ""Joseph Keshet"", ""Shai Shalev-Shwartz"", ""Yoram Singer.""], ""title"": ""Online passive-aggressive algorithms"", ""venue"": ""Journal of Machine Learning Research, 7:551\u2013585."", ""year"": 2006}, {""authors"": [""Daum\u00e9"", ""Hal"", ""John Langford"", ""Daniel Marcu.""], ""title"": ""Search-based structured prediction"", ""venue"": ""Machine Learning, 75(3):297\u2013325."", ""year"": 2009}, {""authors"": [""Daum\u00e9"", ""Hal"", ""Daniel Marcu.""], ""title"": ""Learning as search optimization: approximate large margin methods for structured prediction"", ""venue"": ""ICML, pages 169\u2013176, Bonn."", ""year"": 2005}, {""authors"": [""Espinosa"", ""Dominic"", ""Michael White"", ""Dennis Mehay.""], ""title"": ""Hypertagging: Supertagging for surface realization with CCG"", ""venue"": ""Proceedings of ACL-08: HLT, pages 183\u2013191, Columbus, OH."", ""year"": 2008}, {""authors"": [""Filippova"", ""Katja"", ""Michael Strube.""], ""title"": ""Generating constituent order in German clauses"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 320\u2013327, Prague."", ""year"": 2007}, {""authors"": [""Filippova"", ""Katja"", ""Michael Strube.""], ""title"": ""Tree linearization in English: Improving language model based approaches"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North"", ""year"": 2009}, {""authors"": [""Fowler"", ""Timothy A.D."", ""Gerald Penn.""], ""title"": ""Accurate context-free parsing with combinatory categorial grammar"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2010}, {""authors"": [""Goldberg"", ""Yoav"", ""Michael Elhadad.""], ""title"": ""An efficient algorithm for easy-first non-directional dependency parsing"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North"", ""year"": 2010}, {""authors"": [""Guo"", ""Yuqing"", ""Deirdre Hogan"", ""Josef van Genabith.""], ""title"": ""DCU at generation challenges 2011 surface realisation track"", ""venue"": ""Proceedings of the Generation Challenges Session at the 13th European Workshop"", ""year"": 2011}, {""authors"": [""He"", ""Wei"", ""Haifeng Wang"", ""Yuqing Guo"", ""Ting Liu.""], ""title"": ""Dependency based Chinese sentence realization"", ""venue"": ""Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint"", ""year"": 2009}, {""authors"": [""Hockenmaier"", ""Julia.""], ""title"": ""Data and Models for Statistical Parsing with Combinatory Categorial Grammar"", ""venue"": ""Ph.D. thesis, School of Informatics, University of Edinburgh."", ""year"": 2003}, {""authors"": [""Hockenmaier"", ""Julia"", ""Mark Steedman.""], ""title"": ""CCGbank: A corpus of CCG derivations and dependency structures extracted from the Penn Treebank"", ""venue"": ""Computational Linguistics, 33(3):355\u2013396."", ""year"": 2007}, {""authors"": [""Huang"", ""Liang"", ""Suphan Fayong"", ""Yang Guo.""], ""title"": ""Structured perceptron with inexact search"", ""venue"": ""Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics:"", ""year"": 2012}, {""authors"": [""Huang"", ""Liang"", ""Kenji Sagae.""], ""title"": ""Dynamic programming for linear-time incremental parsing"", ""venue"": ""Proceedings of ACL, pages 1077\u20131086, Uppsala."", ""year"": 2010}, {""authors"": [""Johansson"", ""Richard"", ""Pierre Nugues.""], ""title"": ""Extended constituent-to-dependency conversion for English"", ""venue"": ""16th Nordic Conference of Computational Linguistics, pages 105\u2013112, Tartu."", ""year"": 2007}, {""authors"": [""Kay"", ""Martin.""], ""title"": ""Chart generation"", ""venue"": ""Proceedings of ACL, pages 200\u2013204, Santa Cruz, CA."", ""year"": 1996}, {""authors"": [""Koehn"", ""Phillip.""], ""title"": ""Statistical Machine Translation"", ""venue"": ""Cambridge University Press."", ""year"": 2010}, {""authors"": [""Constantin"", ""Evan Herbst.""], ""title"": ""Moses: Open source toolkit for statistical machine translation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion"", ""year"": 2007}, {""authors"": [""Koehn"", ""Philipp"", ""Franz Josef Och"", ""Daniel Marcu.""], ""title"": ""Statistical phrase-based translation"", ""venue"": ""Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational"", ""year"": 2003}, {""authors"": [""Koo"", ""Terry"", ""Michael Collins.""], ""title"": ""Efficient third-order dependency parsers"", ""venue"": ""Proceedings of ACL, pages 1\u201311, Uppsala."", ""year"": 2010}, {""authors"": [""J. Lee"", ""S. Seneff.""], ""title"": ""Automatic grammar correction for second-language learners"", ""venue"": ""Proceedings of Interspeech, pages 1978\u20131981, Pittsburgh, PA."", ""year"": 2006}, {""authors"": [""Liu"", ""Yang.""], ""title"": ""A shift-reduce parsing algorithm for phrase-based string-to-dependency translation"", ""venue"": ""Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics"", ""year"": 2013}, {""authors"": [""Marcus"", ""Mitchell P."", ""Beatrice Santorini"", ""Mary Ann Marcinkiewicz.""], ""title"": ""Building a large annotated corpus of English: The Penn Treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1993}, {""authors"": [""Papineni"", ""Kishore"", ""Salim Roukos"", ""Todd Ward"", ""Wei-Jing Zhu.""], ""title"": ""BLEU: A method for automatic evaluation of machine translation"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for"", ""year"": 2002}, {""authors"": [""Ratnaparkhi"", ""Adwait.""], ""title"": ""A maximum entropy model for part-of-speech tagging"", ""venue"": ""Proceedings of EMNLP, pages 133\u2013142, Somerset, NJ."", ""year"": 1996}, {""authors"": [""Reiter"", ""Ehud"", ""Robert Dale.""], ""title"": ""Building applied natural language generation systems"", ""venue"": ""Natural Language Engineering, 3(1):57\u201387."", ""year"": 1997}, {""authors"": [""Shen"", ""Libin"", ""Aravind Joshi.""], ""title"": ""LTAG dependency parsing with bidirectional incremental construction"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing,"", ""year"": 2008}, {""authors"": [""Shen"", ""Libin"", ""Giorgio Satta"", ""Aravind Joshi.""], ""title"": ""Guided learning for bidirectional sequence classification"", ""venue"": ""Proceedings of ACL, pages 760\u2013767, Prague."", ""year"": 2007}, {""authors"": [""Song"", ""Linfeng"", ""Yue Zhang"", ""Kai Song"", ""Qun Liu.""], ""title"": ""Joint morphological generation and syntactic linearization"", ""venue"": ""Proceedings of the Twenty-Eighth AAAI Conference, Quebec."", ""year"": 2014}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""The Syntactic Process"", ""venue"": ""The MIT Press, Cambridge, MA."", ""year"": 2000}, {""authors"": [""Surdeanu"", ""Mihai"", ""Richard Johansson"", ""Adam Meyers"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Joakim Nivre""], ""title"": ""The CONLL-2008 shared task on joint parsing of syntactic and semantic dependencies"", ""venue"": ""In Proceedings of the Twelfth"", ""year"": 2008}, {""authors"": [""Wan"", ""Stephen"", ""Mark Dras"", ""Robert Dale"", ""C\u00e9cile Paris""], ""title"": ""Improving grammaticality in statistical sentence generation: Introducing a dependency spanning tree algorithm with an argument"", ""year"": 2009}, {""authors"": [""Weir"", ""David.""], ""title"": ""Characterizing Mildly Context-Sensitive Grammar Formalisms"", ""venue"": ""Ph.D. thesis, University of Pennsylviania."", ""year"": 1988}, {""authors"": [""White"", ""Michael.""], ""title"": ""Reining in CCG chart realization"", ""venue"": ""Proceedings of INLG-04, pages 182\u2013191, Brockenhurst."", ""year"": 2004}, {""authors"": [""White"", ""Michael.""], ""title"": ""Efficient Realization of Coordinate Structures in Combinatory Categorial Grammar"", ""venue"": ""Research on Language & Computation, 4(1):39\u201375."", ""year"": 2006}, {""authors"": [""White"", ""Michael"", ""Rajakrishnan Rajkumar.""], ""title"": ""Perceptron reranking for CCG realization"", ""venue"": ""Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 410\u2013419,"", ""year"": 2009}, {""authors"": [""Wu"", ""Dekai.""], ""title"": ""Stochastic inversion transduction grammars and bilingual parsing of parallel corpora"", ""venue"": ""Computational Linguistics, 23(3):377\u2013403."", ""year"": 1997}, {""authors"": [""Xu"", ""Peng"", ""Ciprian Chelba"", ""Frederick Jelinek.""], ""title"": ""A study on richer syntactic dependencies for structured language modeling"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for"", ""year"": 2002}, {""authors"": [""Zettlemoyer"", ""Luke S."", ""Michael Collins.""], ""title"": ""Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars"", ""venue"": ""Proceedings of the 21st Conference on"", ""year"": 2005}, {""authors"": [""Zhang"", ""Yue.""], ""title"": ""Partial-tree linearization: Generalized word ordering for text synthesis"", ""venue"": ""Proceedings of IJCAI, pages 2232\u20132238, Beijing."", ""year"": 2013}, {""authors"": [""Zhang"", ""Yue"", ""Graeme Blackwood"", ""Stephen Clark.""], ""title"": ""Syntax-based word ordering incorporating a large-scale language model"", ""venue"": ""Proceedings of the 13th Conference of the European Chapter of the"", ""year"": 2012}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""Joint word segmentation and POS tagging using a single perceptron"", ""venue"": ""Proceedings of ACL/HLT, pages 888\u2013896, Columbus, OH."", ""year"": 2008}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""A fast decoder for joint word segmentation and POS-tagging using a single discriminative model"", ""venue"": ""Proceedings of the 2010 Conference"", ""year"": 2010}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""Syntactic processing using the generalized perceptron and beam search"", ""venue"": ""Computational Linguistics, 37(1):105\u2013151."", ""year"": 2011}, {""authors"": [""Zhang"", ""Yue"", ""Joakim Nivre.""], ""title"": ""Transition-based dependency parsing with rich non-local features"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human"", ""year"": 2011}]","acknowledgments  :This work was carried out partly while Yue Zhang was a postdoctoral research associate at the University of Cambridge Computer Laboratory, where he and Stephen Clark were supported by the European Union Seventh Framework Programme (FP7-ICT-2009-4) under grant agreement no. 247762, and partly after Yue Zhang joined Singapore University of Technology and Design, where he was supported by the MOE grant T2-MOE-2013-01. We thank Bill Byrne, Marcus Tomalin, Adrià de Gispert, and Graeme Blackwood for numerous discussions; Anja Belz and Mike White for kindly providing the NLG 2011 shared task data; Kai Song for helping with tables and figures in the draft; Yijia Liu for helping with the bibliography; and the anonymous reviewers for the many
constructive comments that have greatly improved this article since the first draft.
References Auli, Michael and Adam Lopez. 2011.
Training a log-linear parser with loss functions via softmax-margin. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 333–343, Edinburgh.
Barzilay, Regina and Kathleen McKeown. 2005. Sentence fusion for multidocument news summarization. Computational Linguistics, 31(3):297–328. Belz, Anja, Michael White, Dominic Espinosa, Eric Kow, Deirdre Hogan, and Amanda Stent. 2011. The first surface realisation shared task: Overview and evaluation results. In Proceedings of the 13th European Workshop on Natural Language
Generation, ENLG ’11, pages 217–226, Stroudsburg, PA.
Blackwood, Graeme, Adrià de Gispert, and William Byrne. 2010. Fluency constraints for minimum Bayes-risk decoding of statistical machine translation lattices. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 71–79, Beijing. Bohnet, Bernd, Leo Wanner, Simon Mill, and Alicia Burga. 2010. Broad coverage multilingual deep sentence generation with a stochastic multi-level realizer. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 98–106, Beijing. Bos, Johan, Stephen Clark, Mark Steedman, James R. Curran, and Julia Hockenmaier. 2004. Wide-coverage semantic representations from a CCG parser. In Proceedings of COLING-04, pages 1240–1246, Geneva. Caraballo, Sharon A. and Eugene Charniak. 1998. New figures of merit for best-first probabilistic chart parsing. Computational Linguistics, 24:275–298. Carroll, John, Ann Copestake, Dan Flickinger, and Victor Poznanski. 1999. An efficient chart generator for (semi-) lexicalist grammars. In Proceedings of the 7th European Workshop on Natural Language Generation (EWNLG99), pages 86–95, Toulouse. Carroll, John and Stephan Oepen. 2005. High efficiency realization for a wide-coverage unification grammar. In Proceedings of the Second International Joint Conference on Natural Language Processing, IJCNLP’05, pages 165–176, Berlin. Chang, Pi-Chuan and Kristina Toutanova. 2007. A discriminative syntactic word order model for machine translation. In Proceedings of ACL, pages 9–16, Prague. Chiang, David. 2007. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228. Clark, Stephen and James R. Curran. 2007a. Perceptron training for a wide-coverage lexicalized-grammar parser. In Proceedings of the ACL 2007 Workshop on Deep Linguistic Processing, pages 9–16, Prague. Clark, Stephen and James R. Curran. 2007b. Wide-coverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics, 33(4):493–552. Collins, Michael and Brian Roark. 2004. Incremental parsing with the perceptron algorithm. In Proceedings of ACL, pages 111–118, Barcelona.
Crammer, Koby, Ofer Dekel, Joseph Keshet, Shai Shalev-Shwartz, and Yoram Singer. 2006. Online passive-aggressive algorithms. Journal of Machine Learning Research, 7:551–585. Daumé, Hal, John Langford, and Daniel Marcu. 2009. Search-based structured prediction. Machine Learning, 75(3):297–325. Daumé, Hal and Daniel Marcu. 2005. Learning as search optimization: approximate large margin methods for structured prediction. In ICML, pages 169–176, Bonn. Espinosa, Dominic, Michael White, and Dennis Mehay. 2008. Hypertagging: Supertagging for surface realization with CCG. In Proceedings of ACL-08: HLT, pages 183–191, Columbus, OH. Filippova, Katja and Michael Strube. 2007. Generating constituent order in German clauses. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 320–327, Prague. Filippova, Katja and Michael Strube. 2009. Tree linearization in English: Improving language model based approaches. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Short Papers, pages 225–228, Boulder, CO. Fowler, Timothy A. D. and Gerald Penn. 2010. Accurate context-free parsing with combinatory categorial grammar. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 335–344, Uppsala. Goldberg, Yoav and Michael Elhadad. 2010. An efficient algorithm for easy-first non-directional dependency parsing. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 742–750, Los Angeles, CA. Guo, Yuqing, Deirdre Hogan, and Josef van Genabith. 2011. DCU at generation challenges 2011 surface realisation track. In Proceedings of the Generation Challenges Session at the 13th European Workshop on Natural Language Generation, pages 227–229, Nancy. He, Wei, Haifeng Wang, Yuqing Guo, and Ting Liu. 2009. Dependency based Chinese sentence realization. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint
Conference on Natural Language Processing of the AFNLP, pages 809–816, Suntec. Hockenmaier, Julia. 2003. Data and Models for Statistical Parsing with Combinatory Categorial Grammar. Ph.D. thesis, School of Informatics, University of Edinburgh. Hockenmaier, Julia and Mark Steedman. 2007. CCGbank: A corpus of CCG derivations and dependency structures extracted from the Penn Treebank. Computational Linguistics, 33(3):355–396. Huang, Liang, Suphan Fayong, and Yang Guo. 2012. Structured perceptron with inexact search. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 142–151, Montréal. Huang, Liang and Kenji Sagae. 2010. Dynamic programming for linear-time incremental parsing. In Proceedings of ACL, pages 1077–1086, Uppsala. Johansson, Richard and Pierre Nugues. 2007. Extended constituent-to-dependency conversion for English. In 16th Nordic Conference of Computational Linguistics, pages 105–112, Tartu. Kay, Martin. 1996. Chart generation. In Proceedings of ACL, pages 200–204, Santa Cruz, CA. Koehn, Phillip. 2010. Statistical Machine Translation. Cambridge University Press. Koehn, Philipp, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180, Prague. Koehn, Philipp, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology - Volume 1, NAACL ’03, pages 48–54, Edmonton, Canada. Koo, Terry and Michael Collins. 2010. Efficient third-order dependency parsers. In Proceedings of ACL, pages 1–11, Uppsala. Lee, J. and S. Seneff. 2006. Automatic grammar correction for second-language learners. In Proceedings of Interspeech, pages 1978–1981, Pittsburgh, PA.
Liu, Yang. 2013. A shift-reduce parsing algorithm for phrase-based string-to-dependency translation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1–10, Sofia. Marcus, Mitchell P., Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. Papineni, Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, PA. Ratnaparkhi, Adwait. 1996. A maximum entropy model for part-of-speech tagging. In Proceedings of EMNLP, pages 133–142, Somerset, NJ. Reiter, Ehud and Robert Dale. 1997. Building applied natural language generation systems. Natural Language Engineering, 3(1):57–87. Shen, Libin and Aravind Joshi. 2008. LTAG dependency parsing with bidirectional incremental construction. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 495–504, Honolulu. HI. Shen, Libin, Giorgio Satta, and Aravind Joshi. 2007. Guided learning for bidirectional sequence classification. In Proceedings of ACL, pages 760–767, Prague. Song, Linfeng, Yue Zhang, Kai Song, and Qun Liu. 2014. Joint morphological generation and syntactic linearization. In Proceedings of the Twenty-Eighth AAAI Conference, Quebec. Steedman, Mark. 2000. The Syntactic Process. The MIT Press, Cambridge, MA. Surdeanu, Mihai, Richard Johansson, Adam Meyers, Lluı́s Màrquez, and Joakim Nivre. 2008. The CONLL-2008 shared task on joint parsing of syntactic and semantic dependencies. In Proceedings of the Twelfth Conference on Computational Natural Language Learning, pages 159–177, Manchester. Wan, Stephen, Mark Dras, Robert Dale, and Cécile Paris. 2009. Improving grammaticality in statistical sentence generation: Introducing a dependency spanning tree algorithm with an argument satisfaction model. In Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pages 852–860, Athens.
Weir, David. 1988. Characterizing Mildly Context-Sensitive Grammar Formalisms. Ph.D. thesis, University of Pennsylviania. White, Michael. 2004. Reining in CCG chart realization. In Proceedings of INLG-04, pages 182–191, Brockenhurst. White, Michael. 2006. Efficient Realization of Coordinate Structures in Combinatory Categorial Grammar. Research on Language & Computation, 4(1):39–75. White, Michael and Rajakrishnan Rajkumar. 2009. Perceptron reranking for CCG realization. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 410–419, Singapore. Wu, Dekai. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics, 23(3):377–403. Xu, Peng, Ciprian Chelba, and Frederick Jelinek. 2002. A study on richer syntactic dependencies for structured language modeling. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 191–198, Philadelphia, PA. Zettlemoyer, Luke S. and Michael Collins. 2005. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. In Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence, pages 658–666, Edinburgh.
Zhang, Yue. 2013. Partial-tree linearization: Generalized word ordering for text synthesis. In Proceedings of IJCAI, pages 2232–2238, Beijing. Zhang, Yue, Graeme Blackwood, and Stephen Clark. 2012. Syntax-based word ordering incorporating a large-scale language model. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 736–746, Avignon. Zhang, Yue and Stephen Clark. 2008. Joint word segmentation and POS tagging using a single perceptron. In Proceedings of ACL/HLT, pages 888–896, Columbus, OH. Zhang, Yue and Stephen Clark. 2010. A fast decoder for joint word segmentation and POS-tagging using a single discriminative model. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 843–852, Cambridge, MA. Zhang, Yue and Stephen Clark. 2011. Syntactic processing using the generalized perceptron and beam search. Computational Linguistics, 37(1):105–151. Zhang, Yue and Joakim Nivre. 2011. Transition-based dependency parsing with rich non-local features. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 188–193, Portland, OR.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,42 43.41 44.71 50.46 51.40 52.18 73.34 74.68 76.28 :,,,,,,,,"1 introduction :Word ordering is a fundamental problem in natural language generation (NLG, Reiter and Dale 1997). In this article we focus on text generation: Starting with a bag of words, or lemmas, as input, the task is to generate a fluent and grammatical sentence using
∗ Singapore University of Technology and Design. 20 Dover Drive, Singapore. E-mail: yue zhang@sutd.edu.sg.
∗∗ University of Cambridge Computer Laboratory, William Gates Building, 15 JJ Thomson Avenue, Cambridge, UK. E-mail: stephen.clark@cl.cam.ac.uk.
Submission received: 17 April 2013; revised version received: 23 April 2014; accepted for publication: 22 June 2014.
doi:10.1162/COLI a 00229
© 2015 Association for Computational Linguistics
those words. Additional annotation may also be provided with the input—for example, part-of-speech (POS) tags or syntactic dependencies. Applications that can benefit from better text generation algorithms include machine translation (Koehn 2010), abstractive text summarization (Barzilay and McKeown 2005), and grammar correction (Lee and Seneff 2006). Typically, statistical machine translation (SMT) systems (Chiang 2007; Koehn 2010) perform generation into the target language as part of an integrated system, which avoids the high computational complexity of independent word ordering. On the other hand, performing word ordering separately in a pipeline has many potential advantages. For SMT, it offers better modularity between adequacy (translation) and fluency (linearization), and can potentially improve target grammaticality for syntactically different languages (e.g., Chinese and English). More importantly, a standalone word ordering component can in principle be applied to a wide range of text generation tasks, including transfer-based machine translation (Chang and Toutanova 2007).
Most word ordering systems use an n-gram language model, which is effective at controling local fluency. Syntax-based language models, in particular dependency language models (Xu, Chelba, and Jelinek 2002), are sometimes used in an attempt to improve global fluency through the capturing of long-range dependencies. In this article, we take a syntax-based approach and consider two grammar formalisms: Combinatory Categorial Grammar (CCG) and dependency grammar. Our system also employs a discriminative model. Coupled with heuristic search, a strength of the model is that arbitrary features can be defined to capture complex syntactic patterns in output hypotheses. The discriminative model is trained using syntactically annotated data.
From the perspective of search, word ordering is a computationally difficult problem. Finding the best permutation for a set of words according to a bigram language model, for example, is NP-hard, which can be proved by linear reduction from the traveling salesman problem. In practice, exploring the whole search space of permutations is often prevented by adding constraints. In phrase-based machine translation (Koehn, Och, and Marcu 2003; Koehn et al. 2007), a distortion limit is used to constrain the position of output phrases. In syntax-based machine translation systems such as Wu (1997) and Chiang (2007), synchronous grammars limit the search space so that polynomialtime inference is feasible. In fluency improvement (Blackwood, de Gispert, and Byrne 2010), parts of translation hypotheses identified as having high local confidence are held fixed, so that word ordering elsewhere is strictly local.
In this article we begin by proposing a general system to solve the word ordering problem, which does not rely on constraints (which are typically task-specific). In particular, we treat syntax-based word ordering as a structured prediction problem, for which the input is a multi-set (bag) of words and the output is an ordered sentence, together with its syntactic analysis (either CCG derivation or dependency tree, depending on the grammar formalism being used). Given an input, our system searches for the highestscored output, according to a syntax-based discriminative model. One advantage of this formulation of the reordering problem, which can perhaps be thought of as a “pure” text realization task, is that systems for solving it are easily evaluated, because all that is required is a set of sentences for reordering and a standard evaluation metric such as BLEU (Papineni et al. 2002). However, one potential criticism of the “pure” problem is that it is unclear how it relates to real realization tasks, since in practice (e.g., in statistical machine translation systems) the input does provide constraints on the possible output orderings. Our general formulation still allows task-specific contraints to be added if appropriate. Hence as a test of the flexibility of our system,
and a demonstration of the applicability of the system to more realistic text generation scenarios, we consider two further tasks for the dependency-based realization system.
The first task considers a variety of input conditions for the dependency-based system, determined by two parameters. The first is whether POS information is provided for each word in the input multi-set. The second is whether syntactic dependencies between the words are provided. The extreme case is when all dependencies are provided, in which case the problem reduces to the tree linearization problem (Filippova and Strube 2009; He et al. 2009). However, the input can also lie between the two extremes of no- and full-dependency information.
The second task is the NLG 2011 shared task, which provides a further demonstration of the practical utility of our system. The shared task is closer to a real realization scenario, in that lemmas, rather than inflected words, are provided as input. Hence some modifications are required to our system in order that it can perform some word inflection, as well as deciding on the ordering. The shared task data also uses labeled, rather than unlabeled, syntactic dependencies, and so the system was modified to incorporate labels. The final result is that our system gives competitive BLEU scores, compared to the best-performing systems on the shared task.
The structured prediction problem we solve is a very hard problem. Due to the use of syntax, and the search for a sentence together with a single CCG derivation or dependency tree, the search space is exponentially larger than the n-gram word permutation problem. No efficient algorithm exists for finding the optimal solution. Kay (1996) recognized the computational difficulty of chart-based generation, which has many similarities to the problem we address in his seminal paper. We tackle the high complexity by using learning-guided best-first search, exploring a small path in the whole search space, which contains the most likely structures according to the discriminative model. One of the contributions of this article is to introduce, and provide a discriminative solution to, this difficult structured prediction problem, which is an interesting machine learning problem in its own right.
This article is based on, and significantly extends, three conference papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012; Zhang 2013). It includes a more detailed description and discussion of our guided-search approach to syntax-based word ordering, bringing together the CCG- and dependency-based systems under one unified framework. In addition, we discuss the limitations of our previous work, and show that a better model can be developed through scaling of the feature vectors. The resulting model allows fair comparison of constituents of different sizes, and enables the learning algorithms to expand negative examples during training, which leads to significantly improved results over our previous work. The competitive results on the NLG 2011 shared task data are new for this article, and demonstrate the applicability of our system to more realistic text realization scenarios.
The contributions of this article can be summarized as follows: r We address the problem of syntax-based word ordering, drawing attention to this challenging language modeling task and offering a general solution that does not rely on constraints to limit the search space.r We present a novel method for solving the word ordering problem that gives the best reported accuracies to date on the standard Wall Street Journal data.
r We show how our system can be used with two different grammar formalisms: Combinatory Categorial Grammar and dependency grammar.r We show how syntactic constraints can be easily incorporated into the system, presenting results for the dependency-based system with a range of input conditions.r We demonstrate the applicability of the system to more realistic text realization scenarios by obtaining competitive results on the NLG 2011 shared task data, including performing some word inflection as part of a joint system that also performs word reordering.r More generally, we propose a learning-guided, best-first search algorithm for application of discriminative models to extremely large search spaces containing structures of varying sizes. This method could be applied to other complex structured prediction tasks in NLP and machine learning. 2 overview of the search and training algorithms :In this section, the CCG-based system is used to describe the search and training algorithms. However, the same approach can be used for the dependency-based system, as described in Section 4: Instead of building hypotheses by applying CCG rules in a bottom-up manner, the dependency-based system creates dependency links between words.
Given a bag of words, the goal is to put them into an ordered sentence that has a plausible CCG derivation. The search space of the decoding problem consists of all possible CCG derivations for all possible word permutations, and the search goal is to find the highest-scored derivation in the search space. This is an NP-hard problem, as mentioned in the Introduction. We apply learning-guided search to address the high complexity. The intuition is that, because the whole search space cannot be exhausted in order to find the optimal solution, we choose to explore a small area in the search space. A statistical model is used to guide the search, so that only a small portion of the search space containing the most plausible hypotheses is explored.
One natural choice for the decoding algorithm is best-first search, which uses an agenda to order hypotheses, and expands the highest-scored hypothesis on the agenda at each step. The resulting hypotheses after each hypothesis expansion are put back on the agenda, and the process repeats until a goal hypothesis (a full sentence) is found. This search process is guided by the current scores of the hypotheses, and the search path will contain the most plausible hypotheses if they are scored higher than implausible ones. An alternative to best-first search is A∗ search, which makes use of a heuristic function to estimate future scores. A∗ can potentially be more efficient given an effective heuristic function; however, it is not straightforward to define an admissible and accurate estimate of future scores for our problem, and we leave this research question to future work.
In our formulation of the word ordering problem, a hypothesis is a phrase or sentence together with its CCG derivation. Hypotheses are constructed bottom–up: starting from single words, smaller phrases are combined into larger ones according to CCG rules. To allow the combination of hypotheses, we use an additional structure to store a set of hypotheses that have been expanded, which we call accepted
hypotheses. When a hypothesis from the agenda is expanded, it is combined with all accepted hypotheses in all possible ways to produce new hypotheses. The data structure for accepted hypotheses is similar to that used for best-first parsing (Caraballo and Charniak 1998), and we adopt the term chart for this structure. However, note there are important differences to the parsing problem. First, the parsing problem has a fixed word order and is considerably simpler than the word ordering problem we are tackling. Second, although we use the term chart, the structure for accepted hypotheses is not a dynamic programming chart in the same way as for the parsing problem. In our previous papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012), we applied a set of beams to this structure, which makes it similar to the data structure used for phrase-based MT decoding (Koehn 2010). However, we will show later that this structure is unnecessary when the model has more discriminative power, and a conceptually simpler single beam can be used. We will also investigate the possibility of applying dynamic-programming-style pruning to the chart.
We now give an overview of the training algorithm, which is crucial to both the speed and accuracy of the resulting decoder. CCGBank (Hockenmaier and Steedman 2007) is used to train the model. For each training sentence, the corresponding CCGBank derivation together with all its sub-derivations are treated as gold-standard hypotheses. All other hypotheses that can be constructed from the same bag of words are non-gold hypotheses. From the generation perspective this assumption is too strong, because sentences can have multiple orderings (with multiple derivations) that are both grammatical and fluent. Nevertheless, it is the most feasible choice given the training data available.
The efficiency of the decoding algorithm is dependent on the training algorithm because the agenda is ordered according to the hypothesis scores. Hence, a better model will lead to a goal hypothesis being found more quickly. In the ideal situation, all goldstandard hypotheses are scored higher than all non-gold hypotheses, and therefore only gold-standard hypotheses are expanded before the gold-standard goal hypothesis is found. In this case, the minimum number of hypotheses is expanded and the output is correct. The best-first search decoder is optimal not only with respect to accuracy but also speed. This ideal situation can hardly be met in practice, but it determines the goal of the training algorithm: to find a model that scores gold-standard hypotheses higher than non-gold ones.
Learning-guided search places more challenges on the training of a discriminative model than standard structured prediction problems, for example, CKY parsing for CCG (Clark and Curran 2007b). If we take gold-standard hypotheses as positive training examples, and non-gold hypotheses as negative examples, then the training goal is to find a large separating margin between the scores of all positive examples and all negative examples. For CKY parsing, the highest-scored negative example can be found via optimal Viterbi decoding, according to the current model, and this negative example can be used in place of all negative examples during the updating of parameters. In contrast, our best-first search algorithm cannot find an output in reasonable time unless a good model has already been trained, and therefore we cannot run the decoding algorithm in the standard way during training. In our previous papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012), we proposed an approximate online training algorithm, which forces positive examples to be kept in the hypothesis space without being discarded, and prevents the expansion of negative examples during the training process (so that the hypothesis space does not get too large). This algorithm ensures training efficiency, but greatly limits the space of negative examples that is explored during training (and hence fails to replicate the conditions experienced at test
time). In this article, we will show that, with an improved scoring model, it is possible to expand negative examples, which leads to improved performance.
A second and more subtle challenge for our training problem is that we need a stronger model for learning-guided search than for dynamic programming (DP)– based search, such as CKY decoding. For CKY decoding, the model is used to compare hypotheses within each chart cell, which cover the same input words. In contrast, for the best-first search decoder, the model is used to order hypotheses on the agenda, which can cover different numbers of words. It needs much stronger discriminating power, so that it can determine whether a single-word phrase is better than, say, a 40-word sentence. In this article we use scaling of the hypothesis scores by size, so that hypotheses of different sizes can be fairly compared. We also find that, with this new approach, negative examples can be expanded during training and a single beam applied to the chart, resulting in a conceptually simpler and more effective training algorithm and decoder. 3 ccg-based word ordering : We were motivated to use CCG as one of the grammar formalisms for our syntax-based realization system because of its successful application to a number of related tasks, such as wide-coverage parsing (Hockenmaier 2003; Clark and Curran 2007b; Auli and Lopez 2011), semantic parsing (Zettlemoyer and Collins 2005), wide-coverage semantic analysis (Bos et al. 2004), and generation itself (Espinosa, White, and Mehay 2008). The grammar formalism has been described in detail in those papers, and so here we provide only a short description.
CCG (Steedman 2000) is a lexicalized grammar formalism that associates words with lexical categories. Lexical categories are detailed grammatical labels, typically expressing subcategorization information. During CCG parsing, and during our search procedure, categories are combined using CCG’s combinatory rules. For example, a verb phrase in English (S\NP) can combine with an NP to its left, in this case using the combinatory rule of (backward) function application:
NP S\NP⇒ S
In addition to binary rule instances, such as this one, there are also unary rules that operate on a single category in order to change its type. For example, forward typeraising can change a subject NP into a complex category looking to the right for a verb phrase:
NP⇒ S/(S\NP)
Such a type-raised category can then combine with a transitive verb type using the rule of forward composition:
S/(S\NP) (S\NP)/NP⇒ S/NP
Following Fowler and Penn (2010), we extract the grammar by reading rule instances directly from the derivations in CCGbank (Hockenmaier and Steedman 2007), rather than defining the combinatory rule schema manually as in Clark and Curran
(2007b). Hence the grammar we use can be thought of as a context-free approximation to the mildly content sensitive grammar arising from the use of generalized composition rules (Weir 1988). Hockenmaier (2003) contains a detailed description of the grammar that is obtained in this way, including the various unary type-changing rules, as well as additional rules needed to deal with naturally occurring text, such as punctuation rules. For the rest of this article, the term edge is used to refer to a hypothesis in the decoding algorithm. An edge corresponds to a sentence or phrase with a CCG derivation. Edges are built bottom–up, starting from leaf edges, which are constructed by assigning possible lexical categories to input words. Each leaf edge corresponds to an input word with a particular lexical category. Two existing edges can be combined if there exists a CCG rule (extracted from CCGbank, as described earlier) that combines their category labels, and if they do not contain the same input word more times than its total count in the input. The resulting edge is assigned a category label according to the CCG rule, and covers the concatenated surface strings of the two sub-edges in their order of combination. New edges can also be built by applying unary rules to a single existing edge. We define a goal edge as an edge that covers all input words.
Two edges are equivalent if they have the same surface string and identical CCG derivations. Edge equivalence is used for comparison with gold-standard edges. Two edges are DP-equivalent when they have the same DP-signature. Based on the feature templates in Table 1, we define the DP-signature of an edge as the CCG category at the
root of its derivation, the head word associated with the root category, and the multiset of words it contains, together with the word and POS bigrams on either side of its surface string. Edges are built bottom–up from input words or existing edges. If we treat the assignment of lexical categories to input words and the application of unary and binary CCG rules to existing edges as edge-building actions, the structure of an edge can be defined recursively as the sub-structure resulting from its top action plus the structure of its sub-edges (if any), as shown in Figure 1. Here the top action of an edge refers to the most recent action that has been applied to build the edge.
In our previous papers we used a global linear model to score edges, where the score of an edge e is defined as:
f (e) = Φ(e) · θ⃗
Φ(e) represents the feature vector of e and θ⃗ is the parameter vector of the model. Similar to the structure of e, the feature vector Φ(e) can be defined recursively:
Φ(e) = (∑
es∈e Φ(es)
) +ϕ(e)
= ( ∑
ers∈re ϕ(ers)
) +ϕ(e)
In this equation, es ∈ e represents a sub-edge of e. Leaf edges do not have any subedges. Unary-branching edges have one sub-edge, and binary-branching edges have two sub-edges. ers ∈r e represents a (strictly) recursive sub-edge of e. The feature vector ϕ(e) represents the structure of the top action of e; it is extracted according to the feature templates in Table 1. Example instances of the feature templates are given according to the example string and CCG derivation in Figure 2. For leaf edges, ϕ(e) includes information about the lexical category label; for unary-branching edges, ϕ(e) includes information from the unary rule; for binary-branching edges, ϕ(e) includes information from the binary rule, and additionally the token, POS, and lexical category bigrams and trigrams that result from the surface string concatenation of its sub-edges.
By the given definition of Φ(e), f (e), the score of edge e, can be computed recursively as e is built during the decoding process:
f (e) = Φ(e) · θ⃗ = ((∑
es∈e Φ(es)
) +ϕ(e) ) · θ⃗
= (∑
es∈e Φ(es) · θ⃗
) +ϕ(e) · θ⃗
= (∑
es∈e f (es)
) +ϕ(e) · θ⃗
When the top action is applied, the score of f (e) is computed as the sum of f (es) (for all es ∈ e) plus ϕ(e) · θ⃗.
An important aspect of the scoring model is that it is used to compare edges with different sizes during decoding. The size of an edge can be measured in terms of the number of words it contains, or the number of syntax rules in its structure. We define the size of an edge as the number of recursive sub-edges in the edge plus one (e.g., the size of a leaf edge is 1), which is equivalent to the number of actions (i.e., lexical category assignment for leaf edges, and rule application for unary/binary edges) that have been applied to construct the edge. Edges with different sizes can have significantly different numbers of features, which can make the training of a discriminative linear model more difficult. Note that it is common in structured prediction problems for feature vectors to have slightly different sizes because of variant feature instantiation conditions. In CKY parsing, for example, constituents with different numbers of unary rules can be kept in the same chart cell and compared with each other, provided that they cover the same span in the input. In our case, however, the sizes of two feature vectors under comparison can be very different indeed, since a leaf edge with one word can be compared with an edge over the entire input sentence.
In our previous papers we observed empirical convergence of online learning using this linear model, and obtained competitive results. However, as explained in Section 2, only positive examples were expanded during training, and the expansion of negative examples led to non-convergence and made online training infeasible. In this article, in order to increase the discriminating power of the model and to make use of negative examples during training, we apply length normalization to the scoring function, so that the score of an edge is independent of its size. To achieve this, we scale the original
linear model score by the number of recursive sub-edges in the edge plus one. For a given edge e, the new score f̂ (e) is defined as:
f̂ (e) = f (e) |e|
= Φ(e) · θ⃗ |e|
=
∑ ers∈re ϕ(e r s) · θ⃗+ϕ(e) · θ⃗
|{ers ∈r e}+ 1|
In the equation, |e| represents the size of e, which is equal to the number of actions that have been applied when e is constructed. By dividing the score f (e) by the size of e, the score f̂ (e) represents an averaged value of ϕ(ers) · θ⃗ (ers ∈r e) and ϕ(e) · θ⃗, averaged by the number of recursive sub-edges plus one (i.e., the total actions), and is independent of the size of e. Given normalized feature vectors, the training of the parameter vector θ⃗ needs to be adjusted correspondingly, which will be discussed subsequently. The decoding algorithm takes a multi-set of input words, turns them into a set of leaf edges, and searches for a goal edge by repeated expansion of existing edges. For bestfirst decoding, an agenda and a chart are used. The agenda is a priority queue on which edges to be expanded are ordered according to their current scores. The chart is a fixedsize beam used to record a limited number of accepted edges. During initialization, leaf edges are generated by assigning all possible lexical categories to each input word, before they are put on the agenda. During each step in the decoding process, the highestscored edge on the agenda is popped off and expanded. If it is a goal edge, it is returned as the output, and the decoding finishes. Otherwise it is extended with unary rules, and combined with existing edges in the chart, using binary rules to produce new edges. The resulting edges are scored and put on the agenda, and the original edge is put into the chart. The process repeats until a goal edge is found, or a timeout limit is reached.
For the timeout case, a default output is produced by greedily combining existing edges in the chart in descending order of size. In particular, edges in the chart are sorted by size, and the largest is taken as the current default output. Then the sorted list is traversed, with an attempt to greedily concatenate the current edges in the list to the right of the current default output. If the combination is not allowed (i.e., the two edges contain some input words more times than its count in the input), the current edge is discarded. Otherwise, the current default output is updated.
In our previous papers we used a set of beams for the chart, each storing a certain number of highest-scored edges that cover a particular number of words. This structure is similar to the chart used for phrase-based SMT decoding. The main reason for the multiple beams is the non-comparability of edges in different beams, which can have feature vectors of significantly different sizes. In this article, however, our chart is a single beam structure containing the top-scored accepted edges. This simple data structure is enabled by the use of the scaled linear model, and leads to comparable accuracies to the multiple-beam chart. In addition to its simplicity, it also fits well with the use of agenda-based search, because edges of different sizes will ultimately be compared with each other on the agenda.
We apply DP-style pruning to the chart, keeping only the highest-scored edge among those that have the same DP-signature. During decoding, before a newly constructued edge e is put into the chart, the chart is examined to check whether it contains an existing edge e0 with the same DP-signature as e. If such an edge exists, it is popped off the chart and compared with the newly constructed edge e, with the higher scored edge ẽ being put into the chart and the lower scored edge e′ being discarded. If the newly constructed edge e is not discarded, then we expand e to generate new edges.
It is worth noting that, in this case, a new edge that results from the expansion of e can have DP-equivalent edges in the agenda or the chart, which had been generated by expansion of its DP-equivalent predecessor e′ = e0. Putting such new edges on the agenda will result in the system keeping multiple edges with the same signature. However, because applying DP-style pruning to the agenda requires updating the whole agenda, and is computationally expensive, we choose to tolerate such DP-equivalent duplications in the agenda.
Pseudocode for the decoder is shown as Algorithm 1. INITAGENDA returns an initialized agenda with all leaf edges. INITCHART returns a cleared chart. TIMEOUT
Algorithm 1 The decoding algorithm. a← INITAGENDA( ) c← INITCHART( ) while not TIMEOUT( ) do
new← [] e← POPBEST(a) if GOALTEST(e) then
return e end if (e′, ẽ)← DPCHARTPRUNE(c, e) if e′ is e then
continue end if for e′ ∈ UNARY(e, grammar) do
APPEND(new, e′) end for for ẽ ∈ c do
if CANCOMBINE(e, ẽ, grammar) then e′← BINARY(e, ẽ, grammar) APPEND(new, e′) end if if CANCOMBINE(ẽ, e, grammar) then
e′← BINARY(ẽ, e, grammar) APPEND(new, e′)
end if end for for e′ ∈ new do
ADD(a, e′) end for ADD(c, e)
end while
returns true if the timeout limit has been reached, and false otherwise. POPBEST pops the top edge from the agenda and returns the edge. GOALTEST takes an edge and returns true if and only if the edge is a goal edge. DPCHARTPRUNE takes an edge e and checks whether there exists in the chart an edge e0 that is DP-equivalent to e. In case e0 exists, it is popped off the chart and compared with e, with the lower scored edge e′ being discarded, and the higher scored edge ẽ being put into the chart. The function returns the pair e′ and ẽ. CANCOMBINE checks whether two edges can be combined in a given order. Two edges can be combined if they do not contain an overlapping word (i.e., they do not contain a word more times than its count in the input), and their categories can be combined according to the CCG grammar. ADD inserts an edge into the agenda or the chart. In the former case, it is placed into the priority queue according to its score, and, in the latter case, the lowest scored edge in the beam is pruned when the chart is full. We begin by introducing the training algorithm of our previous papers, shown in Algorithm 2, which has the same fundamental structure as the training algorithm of this article but is simpler. The algorithm is based on the decoder, where an agenda is used as a priority queue of edges to be expanded, and a set of accepted edges is kept in a fixed-size chart. The functions INITAGENDA, INITCHART, TIMEOUT, POPBEST, GOALTEST, DPCHARTPRUNE, UNARY, CANCOMBINE, and BINARY are identical to those used in the decoding algorithm. GOLDSTANDARD takes an edge and returns true if and only if it is a gold-standard edge. MINGOLD returns the lowest scored goldstandard edge in the agenda. UPDATEPARAMETERS represents the parameter update algorithm. RECOMPUTESCORES updates the scores of edges in the agenda and chart after the model is updated.
Similar to the decoding algorithm, the agenda is intialized using all possible leaf edges. During each step, the edge e on top of the agenda is popped off. If it is a goldstandard edge, it is expanded in exactly the same way as in the decoder, with the newly generated edges being put on the agenda, and e being inserted into the chart. If e is not a gold-standard edge, we take it as a negative example e−, and take the lowest scored gold-standard edge on the agenda e+ as a positive example, in order to make an update to the parameter vector θ⃗. Note that there must exist a gold-standard edge in the agenda, which can be proved by contradiction.1
The two edges e+ and e− used to perform a model update can be radically different. For example, they may not cover the same words, or even the same number of words. This is different from online training for CKY parsing, for which both positive and negative examples used to adjust parameter vectors reside in the same chart cell, and cover the same sequence of words. The training goal of a typical CKY parser (Clark and Curran 2007a, 2007b) is to find a large separation margin between feature vectors of different derivations of the same sentence, which have comparable sizes. Our goal is to score all gold-standard edges higher than all non-gold edges regardless of their size, which is a more challenging goal. After updating the parameters, the scores of the agenda edges above and including e−, together with all chart edges, are updated, and e− is discarded before the start of the next processing step.
1 An example proof can be based on induction, where the basis is that the agenda contains all gold leaf edges at first, and the induction step is based on edge combination.
Algorithm 2 The training algorithm of our previous papers. 1: a← INITAGENDA( ) 2: c← INITCHART( ) 3: while not TIMEOUT( ) do 4: new← [] 5: e← POPBEST(a) 6: if GOLDSTANDARD(e) and GOALTEST(e) then 7: return e 8: end if 9: if not GOLDSTANDARD(e) then 10: e−← e 11: e+←MINGOLD(a) 12: UPDATEPARAMETERS(e+, e−) 13: RECOMPUTESCORES(a, c) 14: continue 15: end if 16: (e′, ẽ)← DPCHARTPRUNE(c, e) 17: if e′ is e then 18: continue 19: end if 20: for e′ ∈ UNARY(e, grammar) do 21: APPEND(new, e′) 22: end for 23: for ẽ ∈ c do 24: if CANCOMBINE(e, ẽ, grammar) then 25: e′← BINARY(e, ẽ, grammar) 26: APPEND(new, e′) 27: end if 28: if CANCOMBINE(ẽ, e, grammar) then 29: e′← BINARY(ẽ, e, grammar) 30: APPEND(new, e′) 31: end if 32: end for 33: for e′ ∈ new do 34: ADD(a, e′) 35: end for 36: ADD(c, e) 37: end while
One way of viewing the training process is that it pushes gold-standard edges towards the top of the agenda, and, crucially, pushes them above non-gold edges (Zhang and Clark 2011). Given a positive example e+ and a negative example e−, a perceptronstyle update is used to penalize the score for ϕ(e−) and reward the score of ϕ(e+):
θ⃗← θ⃗0 +ϕ(e+)−ϕ(e−)
Here θ⃗0 and θ⃗ denote the parameter vectors before and after the update, respectively. This method proved effective empirically (Zhang and Clark 2011), but it did not
converge well when an n-gram language model was integrated into the system (Zhang, Blackwood, and Clark 2012).
Hence we applied an alternative method for score updates that proved more effective than the perceptron update and enabled the incorporation of a large-scale language model (Zhang, Blackwood, and Clark 2012). This method treats parameter update as finding a separation between gold-standard and non-gold edges. Given a positive example e+ and a negative example e−, we make a minimum update to the parameters so that the score of e+ is higher than that of e− by a margin of 1:
θ⃗← arg min θ⃗′ ∥ θ⃗′ − θ⃗0 ∥, such that Φ(e+)θ⃗′ − Φ(e−)θ⃗′ ≥ 1
The update is similar to the parameter update of online large-margin learning algorithms, such as 1-best MIRA (Crammer et al. 2006), and has a closed-form solution:
θ⃗← θ⃗0 + f (e−)− f (e+) + 1 ∥ Φ(e+)− Φ(e−) ∥2
( Φ(e+)− Φ(e−) ) This online learning method proved more effective than the perceptron algorithm empirically, but still has an important shortcoming in that it did not provide competitive results when allowing the expansion of negative examples during training, which can potentially improve the discriminative model (since expanding negative examples can result in a more representative sample of the search space). We address this issue by introducing a scaled linear model in this article, which, when combined with the expansion of negative examples, significantly improves performance. We apply the same online large-margin training principle; however, the parameter update has to be adjusted for the scaled linear model. In particular, the new goal is to find a separation between f̂ (e+) and f̂ (e−) instead of f (e+) and f (e−), for which the optimization corresponding to the parameter update becomes:
θ⃗← arg min θ⃗′
∥ θ⃗′ − θ⃗0 ∥, such that Φ(e+)θ⃗′ |e+| − Φ(e−)θ⃗ ′ |e−| ≥ 1
where θ⃗0 and θ⃗ represent the parameter vectors before and after the update, respectively. The equation has a closed-form solution:
θ⃗← θ⃗0 + f̂ (e−)− f̂ (e+) + 1 ∥ Φ(e+ )|e+| − Φ(e− ) |e−| ∥ 2
(Φ(e+) |e+| − Φ(e−)|e−| )
Pseudocode for the new training algorithm of this article is shown in Algorithm 3, where MAXNONGOLD returns the highest-scored non-gold edge in the chart. In addition to the aforementioned difference in parameter updates, new code is added to perform additional updates when gold-standard edges are removed from the chart. In our previous work, parameter updates happen only when the top edge from the agenda is not a gold-standard edge. In this article, the expansion of negative training examples will lead to negative examples being put into the chart during training, and hence the possibility of gold-standard edges being removed from the chart. There are two situations when this can happen. First, if a non-gold edge is inserted into the chart,
Algorithm 3 The training algorithm of this article. 1: a← INITAGENDA( ) 2: c← INITCHART( ) 3: while not TIMEOUT( ) do 4: new← [] 5: e← POPBEST(a) 6: if GOLDSTANDARD(e) and GOALTEST(e) then 7: return e 8: end if 9: if not GOLDSTANDARD(e) then 10: e−← e 11: e+←MINGOLD(a) 12: UPDATEPARAMETERS(e+, e−) 13: RECOMPUTESCORES(a, c) 14: end if 15: (e′, ẽ)← DPCHARTPRUNE(c, e) 16: if GOLDSTANDARD(e′) then 17: UPDATEPARAMETERS(e′, ẽ) 18: REMOVE(c, ẽ) 19: ADD(c, e′) 20: else 21: if e′ is e then 22: continue 23: end if 24: end if 25: for e′ ∈ UNARY(e, grammar) do 26: APPEND(new, e′) 27: end for 28: for ẽ ∈ c do 29: if CANCOMBINE(e, ẽ, grammar) then 30: e′← BINARY(e, ẽ, grammar) 31: APPEND(new, e′) 32: end if 33: if CANCOMBINE(ẽ, e, grammar) then 34: e′← BINARY(ẽ, e, grammar) 35: APPEND(new, e′) 36: end if 37: end for 38: for e′ ∈ new do 39: ADD(a, e′) 40: end for 41: e′← ADD(c, e) 42: if GOLDSTANDARD(e′) then 43: ẽ←MAXNONGOLD(c) 44: UPDATEPARAMETERS(e′, ẽ) 45: ADD(c, e′) 46: end if 47: end while
and there exists a gold-standard edge in the chart with the same DP-signature but a lower score, the gold-standard edge will be removed from the chart because of DP-style pruning (since only the highest-scored edge with the same DP-signature is kept in the chart).
Second, if the chart is full when a non-gold edge is put into the chart (recall that the chart is a fixed-size beam), then the lowest scored edge on the chart will be removed. This edge can be a gold-standard edge. In both the first and second case, a gold-standard edge is pruned as the result of the expansion of a negative example. On the other hand, in order for the gold-standard goal edge to be constructed, all gold-standard edges that have been expanded must remain in the chart. As a result, our training algorithm triggers a parameter update whenever a gold-standard edge is removed from the chart, the scores of all chart edges are updated, and the original pruned gold edge is returned to the chart. The original pruned gold-standard edge is treated as the positive example for the update. For the first situation, the newly inserted non-gold edge with the same DP-signature is taken as the negative example, and will be discarded after the parameter update (with a new score that is lower than the new score of the corresponding goldstandard). In the second situation, the highest-scored non-gold edge in the chart is taken as the negative example, and removed from the chart after the update.
In summary, there are two main differences between Algorithms 2 and 3. First, line 14 in Algorithm 2, which skips the expansion of negative examples, is removed in Algorithm 3. Second, lines 16–20 and 42–46 are added in Algorithm 3, which correspond to the updating of parameters when a gold-standard edge is removed from the chart. In addition, the definitions of UPDATEPARAMETERS are different for the perceptron training algorithm (Zhang and Clark 2011), the large-margin training algorithm (Zhang, Blackwood, and Clark 2012), and the large-margin algorithm of this article, as explained earlier. 4 dependency-based word ordering and tree linearization :As well as CCG, the same approach can be applied to the word ordering problem using other grammar formalisms. In this section, we present a dependency-based word ordering system, where the input is again a multi-set of words with gold-standard POS, and the output is an ordered sentence together with its dependency parse. Except for necessary changes to the edge data structure and edge expansion, the same algorithm can be applied to this task.
In addition to abstract word ordering, our framework can be used to solve a more informed, dependency-based word ordering task: tree linearization (Filippova and Strube 2009; He et al. 2009), a task that is very similar to abstract word ordering from a computational perspective. Both tasks involve the permutation of a set of input words, and are NP-hard. The only difference is that, for tree linearization, full unordered dependency trees are given as input. As a result, the output word permutations are more constrained (under the projectivity assumption), and more information is available for search disambiguation.
Tree linearization can be treated as a special case of word ordering, where a grammar constraint is applied such that the output sentence has to be consistent with the input tree. There is a spectrum of grammar constraints between abstract word ordering (no constraints) and tree linearization (full tree constraints). For example, one constraint might consist of a set of dependency relations between input words, but which do not form a complete unordered spanning tree. We call this word ordering task the partial-tree linearization problem, a task that is perhaps closer to NLP applications
than both the abstract word ordering task and the full tree linearization problem, in the sense that NLG pipelines might provide some syntactic relations between words for the linearization step, but not the full spanning tree.
The main content of this section is based on a conference paper (Zhang 2013), which we extend by using the technique of expanding negative training examples (one of the overall contributions of this article). Given a multi-set of input words W and a set of head-dependent relations H between the words in W, the task is to find an ordered sentence consisting of all the words in W and a dependency tree that contains all the relations in H. If each word in W is given a POS tag and H covers all words in W, then the task is (full-)tree linearization; if not then the task is partial-tree linearization. For partial-tree linearization, a subset of W is given fixed POS tags. In all cases, a word either has exactly one (gold) POS tag, or no POS tags. Similar to the CCG case, edge refers to the data structure for a hypothesis in the decoding algorithm. Here a leaf edge refers to an input word with a POS tag, and a non-leaf edge refers to a phrase or sentence with its dependency tree. Edges are constructed bottom–up, by recursively joining two existing edges and adding an unlabeled dependency link between their head words.
As for the CCG system, edges are scored by a global linear model:
f (e) = Φ(e) · θ⃗|e|
where Φ(e) represents the feature vector of e and θ⃗ is the parameter vector of the model. Table 2 shows the feature templates we use, which are inspired by the rich feature templates used for dependency parsing (Koo and Collins 2010; Zhang and Nivre 2011). In the table, h, m, s, hl, hr, ml, and mr are the indices of words in the newly constructed edge, where h and m refer to the head and dependent of the newly constructed arc, s refers to the nearest sibling of m (on the same side of h), and hl, hr, ml, and mr refer to the left and rightmost dependents of h and m, respectively. WORD, POS, LVAL, and RVAL are maps from indices to word forms, POS, left valencies, and right valencies of words, respectively. Example feature instances extracted from the sentence in Figure 3 are shown in the example column. Because of the non-local nature of some of the feature templates we define, we do not apply DP-style pruning for dependency-based treelinearization. The decoding algorithm is similar to that of the CCG system, where an agenda is a priority queue for edges to expand, and chart is a fixed-size beam for a list of accepted edges. During initialization, input words are assigned possible POS tags, resulting in a set of leaf edges that are put onto the agenda. For words with POS constraints, only
the allowed POS tag is assigned. For unconstrained words, we assign all possible POS tags according to a tag dictionary compiled from the training data, following standard practice for POS-tagging (Ratnaparkhi 1996).
When an edge is expanded, it is combined with all edges in the chart in all possible ways to generate new edges. Two edges can be combined by concatenation of the surface strings in both orders and, in each case, constructing a dependency link between their heads in two ways (corresponding to the two options for the head of the new link). When there is a head constraint on the dependent word, a dependency link can be constructed only if it is consistent with the constraint. This algorithm implements abstract word ordering, partial-tree linearization, and full tree linearization—all generalized word ordering tasks—in a unified method.
Pseudocode for the decoder is shown as Algorithm 4. Many of the functions have the same definition as for Algorithm 1: INITAGENDA, INITCHART, TIMEOUT, POPBEST, GOALTEST, ADD. CANCOMBINE checks whether two edges do not contain an overlapping word (i.e., they do not contain a word more times than its count in the input); unlike the CCG case, all pairs of words are allowed to combine according to the dependency model. COMBINE creates a dependency link between two words, with the word order determined by the order in which the arguments are supplied to the function, and the head coming from either the first (HeadLeft) or second (HeadRight) argument (so there are four combinations considered and COMBINE is called four times in Algorithm 4). As for the CCG system, an online large-margin learning algorithm based on the decoding process is used to train the model. At each step, the expanded edge e is compared with the gold standard. If it is a gold edge, decoding continues; otherwise e is taken as a negative example e− and the lowest-scored gold edge in the agenda is taken
Algorithm 4 The decoding algorithm for partial-tree linearization. a← INITAGENDA( ) c← INITCHART( ) while not TIMEOUT( ) do
new← [] e← POPBEST(a) if GOALTEST(e) then
return e end if for ẽ ∈ c do
if CANCOMBINE(e, ẽ) then e′← COMBINE(e, ẽ, HeadLeft) APPEND(new, e′) e′← COMBINE(e, ẽ, HeadRight) APPEND(new, e′) end if if CANCOMBINE(ẽ, e) then
e′← COMBINE(ẽ, e, HeadLeft) APPEND(new, e′) e′← COMBINE(ẽ, e, HeadRight) APPEND(new, e′)
end if end for for e′ ∈ new do
ADD(a, e′) end for ADD(c, e)
end while
as a positive example e+, and a parameter update is executed (repeated here from Section 3.4):
θ⃗← θ⃗0 + f̂ (e−)− f̂ (e+) + 1 ∥ Φ(e+ )|e+| − Φ(e− ) |e−| ∥ 2
(Φ(e+) |e+| − Φ(e−)|e−| )
The training process is essentially the same as in Algorithm 3, but with the CCG grammar and model replaced with the dependency-based grammar and model.
In our conference paper describing the earlier version of the dependency-based system (Zhang 2013), the decoding step is finished immediately after the parameter update; in this article we expand the negative example, as in Algorithm3, putting it onto the chart and thereby exploring a larger part of the search space (in particular that part containing negative examples). Our later experiments show that this method yields improved results, consistent with the CCG system. 5 experiments :We use CCGBank (Hockenmaier and Steedman 2007) and the Penn Treebank (Marcus, Santorini, and Marcinkiewicz 1993) for CCG and dependency data, respectively. CCGbank is the CCG version of the Penn Treebank. Standard splits were used for both: Sections 02–21 for training, Section 00 for development, and Section 23 for the final test. Table 3 gives statistics for the Penn Treebank.
For the CCG experiments, original sentences from CCGBank are transformed into bags of words, with sequence information removed, and passed to our system as input data. The system outputs are compared to the original sentences for evaluation. Following Wan et al. (2009), we use the BLEU metric (Papineni et al. 2002) for string comparison. Although BLEU is not the perfect measure of fluency or grammaticality, being based on n-gram precision, it is currently widely used for automatic evaluation and allows us to compare directly with existing work (Wan et al. 2009). Note also that one criticism of BLEU for evaluating machine translation systems (i.e., that it can only register exact matches between the same words in the system and reference translation), does not apply here, because the system output always contains the same words as the original reference sentence. For the dependency-based experiments, gold-standard dependency trees were derived from bracketed sentences in the treebank using the Penn2Malt tool.2
For fair comparison with Wan et al. (2009), we keep base NPs as atomic units when preparing the input. Wan et al. used base NPs from the Penn Treebank annotation, and we follow this practice for the dependency-based experiments. For the CCG experiments we extract base NPs from CCGBbank by taking as base NPs those NPs that do not recursively contain other NPs. These base NPs mostly correspond to the base NPs from the Penn Treebank: In the training data, there are 242,813 Penn Treebank base NPs with an average size of 1.09, and 216,670 CCGBank base NPs with an average size of 1.19. The plots in Figure 4 show the development test scores of three CCG models by the number of training iterations. The three curves represent the scaled model of this article, the online large-margin model from Zhang, Blackwood, and Clark (2012), and the perceptron model from Zhang and Clark (2011), respectively. For each curve, the BLEU score generally increases as the number of training iterations increases, until it reaches its maximum at a particular iteration. We use the number of training iterations
2 http://w3.msi.vxu.se/~nivre/research/Penn2Malt.html.
that gives the best development test scores for the training of our model when testing on the test data.
Another way to observe the convergence of training is to measure the training times for each iteration at different numbers of iterations. The per-iteration training times for the large-margin and the scaled CCG models are shown in Figure 5. For each model, the training time for each iteration decreases as the number of training iterations increases, reflecting the convergence of learning-guided search. When the model gets better, fewer non-gold hypotheses are expanded before gold hypotheses, and hence it takes less time for the decoder to find the gold goal edge. Figure 6 shows the corresponding curve for dependency-based word ordering, with similar observations.
Because of the expanding of negative examples, the systems of this article took more time to train than those of our previous conference papers. However, the convergence
rate is also faster when negative training examples are expanded, as demonstrated by the rate of speed improvement as the number of training iterations increases. The training times of the perceptron algorithm are close to those of the large-margin algorithm, and hence are omitted from Figures 5 and 6. The new model gives the best development test scores, as shown in Figure 4. The next section investigates the effects of two of the innovations of this article: use of negative examples during training and the scaling of the model by hypothesis size. Table 4 shows a set of CCG development experiments to measure the effect of the scaled model and the expansion of negative examples during training. With the standard linear model (Zhang, Blackwood, and Clark 2012) and no expansions of negative examples, our system obtained a BLEU score of 39.04. The scaled model improved the BLEU score by 1.41 BLEU points to 40.45, and the expansion of negative examples gave a further improvement of 3.02 BLEU points.
These CCG development experiments show that the expansion of negative examples during training is an important factor in achieving good performance. When no negative examples are expanded, the higher score of the scaled linear model demonstrates the effectiveness of fair comparison between edges with different sizes. However,
it is a more important advantage of the scaled linear model that it allows the expansion of negative examples during training, which was not possible with the standard linear model. In the latter case, training failed to converge when negative examples were expanded, reflecting the limitations of the standard linear model in separating the training data. Similar results were found for dependency-based word ordering, where the best development BLEU score improved from 44.71 (Zhang 2013) to 46.44 with the expansion of negative training examples. Figure 7 shows the BLEU scores for the CCG system on the development data when the timeout limit for decoding a single sentence is set to 5 sec, 10 sec, 15 sec, 20 sec, 30 sec, 40 sec, 50 sec, and 60 sec, respectively. The timeout was applied during decoding at test time. The scaled model with negative training examples was used for this set of experiments, and the same model was used for all timeout settings. The results demonstrate that better outputs can be recovered given more search time, which is expected for a time-constrained best-first search framework. Recall that output is created greedily by combining the largest available edges, when the system times out. Similar results were obtained with the dependency-based system of Zhang (2013), where the development BLEU scores improved from 42.89 to 43.42, 43.58, and 43.72 when the timeout limit increased from 5 sec to 10 sec, 30 sec, and 60 sec, respectively. The scaled dependency-based model without expansion of negative examples was used in this set of experiments. Example output for sentences in the development set is shown in Tables 5 and 6, grouped by sentence length. The CCG systems of our previous conference papers and this article are compared, all with the timeout value set to 5 sec. All three systems perform relatively better with smaller sentences. For longer sentences, the fluency of
the output is significantly reduced. One source of errors is confusion between different noun phrases, and where they should be positioned, which becomes more severe with increased sentence length and adds to the difficulty in reading the outputs. The system of this article gave observably improved outputs compared with the two other systems. In the previous section, the same input settings were used for both training and testing, and the assumption was made that the input to the system would be a bag of words, with no constraints on the output structure. This somewhat artificial assumption allows a standardized evaluation but, as discussed previously, text generation applications are unlikely to satisfy this assumption and, in practice, the realization problem is likely to be easier compared with our previous set-up. In this section, we simulate practical situations in dependency-based pipelines by measuring the performance of our system using randomly chosen input POS tags and dependency relations. For maximum flexibility, so that the same system can be applied to different input scenarios, our system is trained without input POS tags or dependencies. However, if POS tags and dependencies are made available during testing, they will be used to provide hard constraints on the output (i.e., the output sentence with POS tags and dependencies must contain those in the input). From the perspective of search, input POS tags and dependencies greatly constrain the search space and lead to an easier search problem, with correspondingly improved outputs.
Table 7 shows a set of development results with varying amounts of POS and dependency information in the input. For each test, we randomly sampled a percentage of words for which the gold-standard POS tags or dependencies are given in the input. As can be seen from the table, increased amounts of POS and dependency information in the input lead to higher BLEU scores, and dependencies were more effective than POS tags in determining the word order in the output. When all POS tags and dependencies are given, our constraint-enabled system gave a BLEU score of 76.28.3
Table 8 shows the output of our system for the first nine development test sentences with different input settings. These examples illustrate the positive effect of input dependencies in specifying the outputs. Consider the second sentence as an example. When only input words are given, the output of the system is largely grammatical but nonsensical. With increasing amounts of dependency relations, the output begins to look more fluent, sometimes with the system reproducing the original sentence when all dependencies are given. Table 9 shows the test results of various systems. For the system of this article, we take the optimal setting from the development tests, using the scaled linear model and
3 When all POS tags and dependencies are also provided during training, the BLEU score is reduced to 74.79, showing the value in the system, which can adapt to varying amounts of POS and dependency information in the input at test time.
expansion of negative examples during training. For direct comparison with previous work, the timeout threshold was set to 5 sec. Our new system of this article significantly outperforms all previous systems and achieves the best published BLEU score on this task. It is worth noting that our systems without a language model outperform the system of our 2012 paper using a large-scale language model.
Interestingly, the dependency-based systems performed better than the CCG systems of this article. One of the main reasons is that the CCG systems generated shorter outputs by not finding full spanning derivations for a larger proportion of inputs. Because of the rigidity in combinatory rules, not all hypotheses in the chart can be combined with the hypothesis being expanded, leading to an increased likelihood of full spanning derivations being unreachable. Overall, the CCG system recovered 93.98% of the input words in the test set, and the dependency system recovered 97.71%. The previous sections report evaluations on the task of word ordering, an abstract yet fundamental problem in text generation. One question that is not addressed by these experiments is how the abstract task can be utilized to benefit full text generation, for which more considerations need to be taken into account in addition to word ordering. We investigate this question using the 2011 Generation Challenge shared task data, which provide a common-ground for the evaluation of text generation systems (Belz et al. 2011).
The data are based on the CoNLL 2008 shared task data (Surdeanu et al. 2008), which consist of selected sections of the Penn WSJ Treebank, converted to syntactic dependencies via the LTH tool (Johansson and Nugues 2007). Sections 2–21 are used for training, Section 24 for development, and Section 23 for testing. A small number of sentences from the original WSJ sections are not included in this set. The input format of the shared task is an unordered syntactic dependency tree, with nodes being lemmas, and dependency relations on the arcs. Named entities and hyphenated words are broken into individual nodes, and special dependency links are used to mark them. Information on coarse-grained POS, number, tense, and participle features is given to each node where applicable. The output is a fully ordered and inflected sentence.
We developed a full-text generation system according to this task specification, with the core component being the dependency-based word ordering system of Section 4. In addition to minor engineering details that were required to adapt the system to this new
task, one additional task that the generation system needs to carry out is morphological generation—finding the appropriate inflected form for each input lemma. Our approach is to perform joint word ordering and inflection using the learning-guided search framework, letting one statistical model decide the best order as well as the inflections of ambiguous lemmas. For a lemma, we generate one or more candidate inflections by using a lexicon and a set of inflection rules. Candidate inflections for an input lemma are generated according to the lemma itself and its input attributes, such as the number and tense. Some lemmas are unambiguous, which are inflected before being passed to the word ordering system. For the other lemmas, more than one candidate’s inflections are passed as input words to the word ordering system. To ensure that each lemma occurs only once in the output, a unique ID is given to all the inflections of the same lemma, making them mutually exclusive.
Four types of lemmas need morphological generation, including nouns, verbs, adjectives, and miscellaneous cases. The last category includes a (a or an) and not (not or n‘t), for which the best inflection can be decided only when n-gram information is available. For these lemmas, we pass all possible inflections to the search module. For nouns and adjectives, the inflection is relatively straightforward, since the number (e.g., singular, plural) of a lemma is given as an attribute of the input node, and comparative and superlative adjectives have specific parts of speech. For those cases where the necessary information is not available from the input, all possible inflections are handed over to the search module for further disambiguation. The most ambiguous lemma types are verbs, which can be further divided into be and other verbs. The uniqueness of be is that the inflections for the first and second person can be different. All verb inflections are disambiguated according to the tense and participle attributes of the input node. In addition, for verbs in the present tense, the subject needs to be determined in order to differentiate between third-person singular verbs and others. This can be straightforward when the subject is a noun or pronoun, but can be ambiguous when the subject is a wh-pronoun, in which case the real subject might not be directly identifiable from the dependency tree. We leave all possible inflections of be and other verbs to the word ordering system whenever the ambiguity is not directly solvable from the subject dependency link. Overall, the pre-processing step generates 1.15 inflections for each lemma on average.
For word ordering, the search procedure of Algorithm 4 is applied directly, and the feature templates of Table 2 are used with additional labeled dependency features described subsequently. The main reason that the dependency-based word ordering algorithm can perform joint morphological disambiguation is that it uses rich syntactic and n-gram features to score candidate hypotheses, which can also differentiate between correct and incorrect inflections under particular contexts. For example, an honest person and a honest person can be differentiated by n-gram features, while Tom and Sally is and Tom and Sally are can be differentiated by higher-order dependency features.
In addition to lemma-formed inputs, one other difference between the shared task and the word ordering problem solved by Algorithm 4 is that the former uses labeled dependencies whereas Algorithm 4 constructs unlabeled dependency trees. We address this issue by assigning dependency labels in the construction of dependency links, and applying an extra set of features. The new features are defined by making a duplicate of all the features from Table 2 that contain dir information, and associating each feature in the new copy with a dependency label.
The training of the word ordering system requires fully ordered dependency trees, while references in the shared task data are raw sentences. We perform a pre-processing
step to obtain gold-standard training data by matching the input lemmas to the reference sentence in order to obtain their gold-standard order. More specifically, given a training instance, we generate all candidate inflections for each lemma, resulting in an exponential set of possible mappings between the input tree and the reference sentence. We then prune these mappings bottom–up, assuming that the dependency tree is projective, and therefore that each word dominates a continuous span in the reference. After such pruning, only one correct mapping is found for the majority of the cases. For the cases where more than one mapping is found, we randomly choose one as the gold-standard. There are also instances for which no correct ordering can be found, and these are mostly due to non-projectivity in the shared task data, with a few cases being due to conflicts between our morphological generation system and the shared task data, or inconsistency in the data itself. Out of the 39K training instances, 2.8K conflicting instances are discarded, resulting in 36.2K gold-standard ordered dependency trees.
Table 10 shows the results of our system and the top two participating systems of the shared task. Our system outperforms the STUMABA system by 0.5 BLEU points, and the DCU system by 3.8 BLEU points. More evaluation of the system was published in Song et al. (2014). 6 related work :There is a recent line of research on text-to-text generation, which studies the linearization of dependency structures (Barzilay and McKeown 2005; Filippova and Strube 2007, 2009; He et al. 2009; Bohnet et al. 2010; Guo, Hogan, and van Genabith 2011). On the other hand, Wan et al. (2009) study the ordering of a bag of words without any dependency information given. We generalize the word ordering problem, and formulate it as a task of ordering a multi-set of words, regardless of input syntactic constraints.
Our bottom–up, chart-based generation algorithm is inspired by the line of work on chart-based realization (Kay 1996; Carroll et al. 1999; White 2004, 2006; Carroll and Oepen 2005). Kay (1996) first proposed the concept of chart realization, drawing analogies between realization and parsing of free order languages. He discussed efficiency issues and provided solutions to specific problems. For the task of realization, efficiency improvement has been further investigated (Carroll et al. 1999; Carroll and Oepen 2005). The inputs to these systems are logical forms, which form natural constraints on the interaction between edges. In our case, one constraint that has been leveraged in the dependency system is a projectivity assumption—we assume that the dependents of a word must all have been attached before the word is attached to its head word, and that spans do not cross during combination. In addition, we assume that the right dependents of a word must have been attached before a left dependent of the word is attached. This constraint avoids spurious ambiguities. The projectivity assumption
is an important basis for the feasibility of the dependency system; it is similar to the chunking constraints of White (2006) for CCG-based realization.
White (2004) describes a system that performs CCG realization using best-first search. The search process of our algorithm is similar to that work, but the input is different: logical forms in the case of White (2004) and bags of words in our case. Further along this line, Espinosa, White, and Mehay (2008) also describe a CCG-based realization system, applying “hypertagging”—a form of supertagging—to logical forms in order to make use of CCG lexical categories in the realization process. White and Rajkumar (2009) further use perceptron reranking on n-best realization output to improve the quality.
The use of perceptron learning to improve search has been proposed in guided learning for easy-first search (Shen, Satta, and Joshi 2007) and LaSO (Daumé and Marcu 2005). LaSO is a general framework for various search strategies. Our learning algorithm is similar to LaSO with best-first inference, but the parameter updates are different. In particular, LaSO updates parameters when all correct hypotheses are lost, but our algorithm makes an update as soon as the top item from the agenda is incorrect. Our algorithm updates the parameters using a stronger precondition, because of the large search space. Given an incorrect hypothesis, LaSO finds the corresponding gold hypothesis for a perceptron update by constructing its correct sibling. In contrast, our algorithm takes the lowest scored gold hypothesis currently in the agenda to avoid updating parameters for hypotheses that may have not been constructed.
Our parameter update strategy is closer to the guided learning mechanism for the easy-first algorithm of Shen, Satta, and Joshi (2007), which maintains a queue of hypotheses during search, and performs learning to ensure that the highest-scored hypothesis in the queue is correct. However, in easy-first search, hypotheses from the queue are ranked by the score of their next action, rather than the hypothesis score. Moreover, Shen, Satta, and Joshi use aggressive learning and regenerate the queue after each update, but we perform non-aggressive learning, which is faster and is more feasible for our large and complex search space. Similar methods to Shen, Satta, and Joshi (2007) have also been used in Shen and Joshi (2008) and Goldberg and Elhadad (2010).
Another framework that closely integrates learning and search is SEARN (Daumé, Langford, and Marcu 2009), which addresses structured prediction problems that can be transformed into a series of simple classification tasks. The transformation is akin to greedy search in the sense that the complex structure is constructed by sequential classification decisions. The key problem that SEARN addresses is how to learn the tth decision based on the previous t− 1 decisions, so that the overall loss in the resulting structure is minimized. Similar to our framework, SEARN allows arbitrary features. However, SEARN is more oriented to greedy search, optimizing local decisions. In contrast, our framework is oriented to best-first search, optimizing global structures.
Learning and search also interact with each other in a global discriminative learning and beam-search framework for incremental structured prediction (Zhang and Clark 2011). In this framework, an output is constructed incrementally by a sequence of transitions, while a beam is used to record the highest scored structures at each step. Online training is performed based on the search process, with the objective function being the margin between correct and incorrect structures. The method involves an earlyupdate strategy, which stops search and updates parameters immediately when the gold structure falls out of the beam during training. It was first proposed by Collins and Roark (2004) for incremental parsing, and later gained popularity in the investigations of many NLP tasks, including POS-tagging (Zhang and Clark 2010), transition-based dependency parsing (Zhang and Clark 2008; Huang and Sagae 2010), and machine
translation (Liu 2013). Huang, Fayong, and Guo (2012) propose a theoretical analysis to the early-update training strategy, pointing out that it is a type of training method that fixes score violations in inexact search. When the score of a gold-standard structure is lower than that of a non-gold structure, a violation exists. Our parameter udpate strategy in this article can also be treated as a mechanism for violation fixing. 7 conclusion :We investigated the general task of syntax-based word ordering, which is a fundamental problem for text generation, and a computationally very expensive search task. We provide a principled solution to this problem using learning-guided search, a framework that is applicable to other NLP problems with complex search spaces. We compared different methods for parameter updates, and showed that a scaled linear model gave the best results by allowing better comparisons between phrases of different sizes, increasing the separability of hypotheses and enabling the expansion of negative examples during training.
We formulate abstract word ordering as a spectrum of tasks with varying input specificity, from “pure” word ordering without any syntactic information to fullyinformed word ordering with a complete unordered dependency tree given. Experiments show that our proposed method can effectively use available input constraints in generating output sentences.
Evaluation on the NLG 2011 shared task data shows that our system can be successfully applied to a more realistic application scenario, in particular one where some dependency constraints are provided in the input and word inflection is required as well as word ordering. Additional tasks that may be required in a practical text generation scenario include word selection, including the determination of content words and generation of function words. The joint modeling solution that we have proposed for word ordering and inflection could also be adopted for word selection, although the search space is greatly increased when the words themselves need deciding, particularly content words. Word ordering is a fundamental problem in text generation. In this article, we study word ordering using a syntax-based approach and a discriminative model. Two grammar formalisms are considered: Combinatory Categorial Grammar (CCG) and dependency grammar. Given the search for a likely string and syntactic analysis, the search space is massive, making discriminative training challenging. We develop a learning-guided search framework, based on best-first search, and investigate several alternative training algorithms. The framework we present is flexible in that it allows constraints to be imposed on output word orders. To demonstrate this flexibility, a variety of input conditions are considered. First, we investigate a “pure” word-ordering task in which the input is a multi-set of words, and the task is to order them into a grammatical and fluent sentence. This task has been tackled previously, and we report improved performance over existing systems on a standard Wall Street Journal test set. Second, we tackle the same reordering problem, but with a variety of input conditions, from the bare case with no dependencies or POS tags specified, to the extreme case where all POS tags and unordered, unlabeled dependencies are provided as input (and various conditions in between). When applied to the NLG 2011 shared task, our system gives competitive results compared with the best-performing systems, which provide a further demonstration of the practical utility of our system. [{""affiliations"": [], ""name"": ""Yue Zhang""}, {""affiliations"": [], ""name"": ""Stephen Clark""}] SP:b2d9b8b3028538d65ecf5e5ecf4f428224aff6b1 [{""authors"": [""Auli"", ""Michael"", ""Adam Lopez.""], ""title"": ""Training a log-linear parser with loss functions via softmax-margin"", ""venue"": ""Proceedings of the 2011 Conference on Empirical Methods in Natural Language"", ""year"": 2011}, {""authors"": [""Barzilay"", ""Regina"", ""Kathleen McKeown.""], ""title"": ""Sentence fusion for multidocument news summarization"", ""venue"": ""Computational Linguistics, 31(3):297\u2013328."", ""year"": 2005}, {""authors"": [""Belz"", ""Anja"", ""Michael White"", ""Dominic Espinosa"", ""Eric Kow"", ""Deirdre Hogan"", ""Amanda Stent.""], ""title"": ""The first surface realisation shared task: Overview and evaluation results"", ""venue"": ""Proceedings of the 13th"", ""year"": 2011}, {""authors"": [""Blackwood"", ""Graeme"", ""Adri\u00e0 de Gispert"", ""William Byrne.""], ""title"": ""Fluency constraints for minimum Bayes-risk decoding of statistical machine translation lattices"", ""venue"": ""Proceedings of the 23rd International"", ""year"": 2010}, {""authors"": [""Bohnet"", ""Bernd"", ""Leo Wanner"", ""Simon Mill"", ""Alicia Burga.""], ""title"": ""Broad coverage multilingual deep sentence generation with a stochastic multi-level realizer"", ""venue"": ""Proceedings of the 23rd International"", ""year"": 2010}, {""authors"": [""Bos"", ""Johan"", ""Stephen Clark"", ""Mark Steedman"", ""James R. Curran"", ""Julia Hockenmaier.""], ""title"": ""Wide-coverage semantic representations from a CCG parser"", ""venue"": ""Proceedings of COLING-04,"", ""year"": 2004}, {""authors"": [""Caraballo"", ""Sharon A."", ""Eugene Charniak.""], ""title"": ""New figures of merit for best-first probabilistic chart parsing"", ""venue"": ""Computational Linguistics, 24:275\u2013298."", ""year"": 1998}, {""authors"": [""Carroll"", ""John"", ""Ann Copestake"", ""Dan Flickinger"", ""Victor Poznanski.""], ""title"": ""An efficient chart generator for (semi-) lexicalist grammars"", ""venue"": ""Proceedings of the 7th European Workshop on Natural Language"", ""year"": 1999}, {""authors"": [""Carroll"", ""John"", ""Stephan Oepen.""], ""title"": ""High efficiency realization for a wide-coverage unification grammar"", ""venue"": ""Proceedings of the Second International Joint Conference on Natural Language Processing, IJCNLP\u201905,"", ""year"": 2005}, {""authors"": [""Chang"", ""Pi-Chuan"", ""Kristina Toutanova.""], ""title"": ""A discriminative syntactic word order model for machine translation"", ""venue"": ""Proceedings of ACL, pages 9\u201316, Prague."", ""year"": 2007}, {""authors"": [""Chiang"", ""David.""], ""title"": ""Hierarchical phrase-based translation"", ""venue"": ""Computational Linguistics, 33(2):201\u2013228."", ""year"": 2007}, {""authors"": [""Clark"", ""Stephen"", ""James R. Curran.""], ""title"": ""Perceptron training for a wide-coverage lexicalized-grammar parser"", ""venue"": ""Proceedings of the ACL 2007 Workshop on Deep Linguistic Processing, pages 9\u201316, Prague."", ""year"": 2007}, {""authors"": [""Clark"", ""Stephen"", ""James R. Curran.""], ""title"": ""Wide-coverage efficient statistical parsing with CCG and log-linear models"", ""venue"": ""Computational Linguistics, 33(4):493\u2013552."", ""year"": 2007}, {""authors"": [""Collins"", ""Michael"", ""Brian Roark.""], ""title"": ""Incremental parsing with the perceptron algorithm"", ""venue"": ""Proceedings of ACL, pages 111\u2013118, Barcelona."", ""year"": 2004}, {""authors"": [""Crammer"", ""Koby"", ""Ofer Dekel"", ""Joseph Keshet"", ""Shai Shalev-Shwartz"", ""Yoram Singer.""], ""title"": ""Online passive-aggressive algorithms"", ""venue"": ""Journal of Machine Learning Research, 7:551\u2013585."", ""year"": 2006}, {""authors"": [""Daum\u00e9"", ""Hal"", ""John Langford"", ""Daniel Marcu.""], ""title"": ""Search-based structured prediction"", ""venue"": ""Machine Learning, 75(3):297\u2013325."", ""year"": 2009}, {""authors"": [""Daum\u00e9"", ""Hal"", ""Daniel Marcu.""], ""title"": ""Learning as search optimization: approximate large margin methods for structured prediction"", ""venue"": ""ICML, pages 169\u2013176, Bonn."", ""year"": 2005}, {""authors"": [""Espinosa"", ""Dominic"", ""Michael White"", ""Dennis Mehay.""], ""title"": ""Hypertagging: Supertagging for surface realization with CCG"", ""venue"": ""Proceedings of ACL-08: HLT, pages 183\u2013191, Columbus, OH."", ""year"": 2008}, {""authors"": [""Filippova"", ""Katja"", ""Michael Strube.""], ""title"": ""Generating constituent order in German clauses"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 320\u2013327, Prague."", ""year"": 2007}, {""authors"": [""Filippova"", ""Katja"", ""Michael Strube.""], ""title"": ""Tree linearization in English: Improving language model based approaches"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North"", ""year"": 2009}, {""authors"": [""Fowler"", ""Timothy A.D."", ""Gerald Penn.""], ""title"": ""Accurate context-free parsing with combinatory categorial grammar"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2010}, {""authors"": [""Goldberg"", ""Yoav"", ""Michael Elhadad.""], ""title"": ""An efficient algorithm for easy-first non-directional dependency parsing"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North"", ""year"": 2010}, {""authors"": [""Guo"", ""Yuqing"", ""Deirdre Hogan"", ""Josef van Genabith.""], ""title"": ""DCU at generation challenges 2011 surface realisation track"", ""venue"": ""Proceedings of the Generation Challenges Session at the 13th European Workshop"", ""year"": 2011}, {""authors"": [""He"", ""Wei"", ""Haifeng Wang"", ""Yuqing Guo"", ""Ting Liu.""], ""title"": ""Dependency based Chinese sentence realization"", ""venue"": ""Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint"", ""year"": 2009}, {""authors"": [""Hockenmaier"", ""Julia.""], ""title"": ""Data and Models for Statistical Parsing with Combinatory Categorial Grammar"", ""venue"": ""Ph.D. thesis, School of Informatics, University of Edinburgh."", ""year"": 2003}, {""authors"": [""Hockenmaier"", ""Julia"", ""Mark Steedman.""], ""title"": ""CCGbank: A corpus of CCG derivations and dependency structures extracted from the Penn Treebank"", ""venue"": ""Computational Linguistics, 33(3):355\u2013396."", ""year"": 2007}, {""authors"": [""Huang"", ""Liang"", ""Suphan Fayong"", ""Yang Guo.""], ""title"": ""Structured perceptron with inexact search"", ""venue"": ""Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics:"", ""year"": 2012}, {""authors"": [""Huang"", ""Liang"", ""Kenji Sagae.""], ""title"": ""Dynamic programming for linear-time incremental parsing"", ""venue"": ""Proceedings of ACL, pages 1077\u20131086, Uppsala."", ""year"": 2010}, {""authors"": [""Johansson"", ""Richard"", ""Pierre Nugues.""], ""title"": ""Extended constituent-to-dependency conversion for English"", ""venue"": ""16th Nordic Conference of Computational Linguistics, pages 105\u2013112, Tartu."", ""year"": 2007}, {""authors"": [""Kay"", ""Martin.""], ""title"": ""Chart generation"", ""venue"": ""Proceedings of ACL, pages 200\u2013204, Santa Cruz, CA."", ""year"": 1996}, {""authors"": [""Koehn"", ""Phillip.""], ""title"": ""Statistical Machine Translation"", ""venue"": ""Cambridge University Press."", ""year"": 2010}, {""authors"": [""Constantin"", ""Evan Herbst.""], ""title"": ""Moses: Open source toolkit for statistical machine translation"", ""venue"": ""Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion"", ""year"": 2007}, {""authors"": [""Koehn"", ""Philipp"", ""Franz Josef Och"", ""Daniel Marcu.""], ""title"": ""Statistical phrase-based translation"", ""venue"": ""Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational"", ""year"": 2003}, {""authors"": [""Koo"", ""Terry"", ""Michael Collins.""], ""title"": ""Efficient third-order dependency parsers"", ""venue"": ""Proceedings of ACL, pages 1\u201311, Uppsala."", ""year"": 2010}, {""authors"": [""J. Lee"", ""S. Seneff.""], ""title"": ""Automatic grammar correction for second-language learners"", ""venue"": ""Proceedings of Interspeech, pages 1978\u20131981, Pittsburgh, PA."", ""year"": 2006}, {""authors"": [""Liu"", ""Yang.""], ""title"": ""A shift-reduce parsing algorithm for phrase-based string-to-dependency translation"", ""venue"": ""Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics"", ""year"": 2013}, {""authors"": [""Marcus"", ""Mitchell P."", ""Beatrice Santorini"", ""Mary Ann Marcinkiewicz.""], ""title"": ""Building a large annotated corpus of English: The Penn Treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1993}, {""authors"": [""Papineni"", ""Kishore"", ""Salim Roukos"", ""Todd Ward"", ""Wei-Jing Zhu.""], ""title"": ""BLEU: A method for automatic evaluation of machine translation"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for"", ""year"": 2002}, {""authors"": [""Ratnaparkhi"", ""Adwait.""], ""title"": ""A maximum entropy model for part-of-speech tagging"", ""venue"": ""Proceedings of EMNLP, pages 133\u2013142, Somerset, NJ."", ""year"": 1996}, {""authors"": [""Reiter"", ""Ehud"", ""Robert Dale.""], ""title"": ""Building applied natural language generation systems"", ""venue"": ""Natural Language Engineering, 3(1):57\u201387."", ""year"": 1997}, {""authors"": [""Shen"", ""Libin"", ""Aravind Joshi.""], ""title"": ""LTAG dependency parsing with bidirectional incremental construction"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing,"", ""year"": 2008}, {""authors"": [""Shen"", ""Libin"", ""Giorgio Satta"", ""Aravind Joshi.""], ""title"": ""Guided learning for bidirectional sequence classification"", ""venue"": ""Proceedings of ACL, pages 760\u2013767, Prague."", ""year"": 2007}, {""authors"": [""Song"", ""Linfeng"", ""Yue Zhang"", ""Kai Song"", ""Qun Liu.""], ""title"": ""Joint morphological generation and syntactic linearization"", ""venue"": ""Proceedings of the Twenty-Eighth AAAI Conference, Quebec."", ""year"": 2014}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""The Syntactic Process"", ""venue"": ""The MIT Press, Cambridge, MA."", ""year"": 2000}, {""authors"": [""Surdeanu"", ""Mihai"", ""Richard Johansson"", ""Adam Meyers"", ""Llu\u0131\u0301s M\u00e0rquez"", ""Joakim Nivre""], ""title"": ""The CONLL-2008 shared task on joint parsing of syntactic and semantic dependencies"", ""venue"": ""In Proceedings of the Twelfth"", ""year"": 2008}, {""authors"": [""Wan"", ""Stephen"", ""Mark Dras"", ""Robert Dale"", ""C\u00e9cile Paris""], ""title"": ""Improving grammaticality in statistical sentence generation: Introducing a dependency spanning tree algorithm with an argument"", ""year"": 2009}, {""authors"": [""Weir"", ""David.""], ""title"": ""Characterizing Mildly Context-Sensitive Grammar Formalisms"", ""venue"": ""Ph.D. thesis, University of Pennsylviania."", ""year"": 1988}, {""authors"": [""White"", ""Michael.""], ""title"": ""Reining in CCG chart realization"", ""venue"": ""Proceedings of INLG-04, pages 182\u2013191, Brockenhurst."", ""year"": 2004}, {""authors"": [""White"", ""Michael.""], ""title"": ""Efficient Realization of Coordinate Structures in Combinatory Categorial Grammar"", ""venue"": ""Research on Language & Computation, 4(1):39\u201375."", ""year"": 2006}, {""authors"": [""White"", ""Michael"", ""Rajakrishnan Rajkumar.""], ""title"": ""Perceptron reranking for CCG realization"", ""venue"": ""Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 410\u2013419,"", ""year"": 2009}, {""authors"": [""Wu"", ""Dekai.""], ""title"": ""Stochastic inversion transduction grammars and bilingual parsing of parallel corpora"", ""venue"": ""Computational Linguistics, 23(3):377\u2013403."", ""year"": 1997}, {""authors"": [""Xu"", ""Peng"", ""Ciprian Chelba"", ""Frederick Jelinek.""], ""title"": ""A study on richer syntactic dependencies for structured language modeling"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for"", ""year"": 2002}, {""authors"": [""Zettlemoyer"", ""Luke S."", ""Michael Collins.""], ""title"": ""Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars"", ""venue"": ""Proceedings of the 21st Conference on"", ""year"": 2005}, {""authors"": [""Zhang"", ""Yue.""], ""title"": ""Partial-tree linearization: Generalized word ordering for text synthesis"", ""venue"": ""Proceedings of IJCAI, pages 2232\u20132238, Beijing."", ""year"": 2013}, {""authors"": [""Zhang"", ""Yue"", ""Graeme Blackwood"", ""Stephen Clark.""], ""title"": ""Syntax-based word ordering incorporating a large-scale language model"", ""venue"": ""Proceedings of the 13th Conference of the European Chapter of the"", ""year"": 2012}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""Joint word segmentation and POS tagging using a single perceptron"", ""venue"": ""Proceedings of ACL/HLT, pages 888\u2013896, Columbus, OH."", ""year"": 2008}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""A fast decoder for joint word segmentation and POS-tagging using a single discriminative model"", ""venue"": ""Proceedings of the 2010 Conference"", ""year"": 2010}, {""authors"": [""Zhang"", ""Yue"", ""Stephen Clark.""], ""title"": ""Syntactic processing using the generalized perceptron and beam search"", ""venue"": ""Computational Linguistics, 37(1):105\u2013151."", ""year"": 2011}, {""authors"": [""Zhang"", ""Yue"", ""Joakim Nivre.""], ""title"": ""Transition-based dependency parsing with rich non-local features"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human"", ""year"": 2011}] acknowledgments  :This work was carried out partly while Yue Zhang was a postdoctoral research associate at the University of Cambridge Computer Laboratory, where he and Stephen Clark were supported by the European Union Seventh Framework Programme (FP7-ICT-2009-4) under grant agreement no. 247762, and partly after Yue Zhang joined Singapore University of Technology and Design, where he was supported by the MOE grant T2-MOE-2013-01. We thank Bill Byrne, Marcus Tomalin, Adrià de Gispert, and Graeme Blackwood for numerous discussions; Anja Belz and Mike White for kindly providing the NLG 2011 shared task data; Kai Song for helping with tables and figures in the draft; Yijia Liu for helping with the bibliography; and the anonymous reviewers for the many
constructive comments that have greatly improved this article since the first draft.
References Auli, Michael and Adam Lopez. 2011.
Training a log-linear parser with loss functions via softmax-margin. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 333–343, Edinburgh.
Barzilay, Regina and Kathleen McKeown. 2005. Sentence fusion for multidocument news summarization. Computational Linguistics, 31(3):297–328. Belz, Anja, Michael White, Dominic Espinosa, Eric Kow, Deirdre Hogan, and Amanda Stent. 2011. The first surface realisation shared task: Overview and evaluation results. In Proceedings of the 13th European Workshop on Natural Language
Generation, ENLG ’11, pages 217–226, Stroudsburg, PA.
Blackwood, Graeme, Adrià de Gispert, and William Byrne. 2010. Fluency constraints for minimum Bayes-risk decoding of statistical machine translation lattices. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 71–79, Beijing. Bohnet, Bernd, Leo Wanner, Simon Mill, and Alicia Burga. 2010. Broad coverage multilingual deep sentence generation with a stochastic multi-level realizer. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 98–106, Beijing. Bos, Johan, Stephen Clark, Mark Steedman, James R. Curran, and Julia Hockenmaier. 2004. Wide-coverage semantic representations from a CCG parser. In Proceedings of COLING-04, pages 1240–1246, Geneva. Caraballo, Sharon A. and Eugene Charniak. 1998. New figures of merit for best-first probabilistic chart parsing. Computational Linguistics, 24:275–298. Carroll, John, Ann Copestake, Dan Flickinger, and Victor Poznanski. 1999. An efficient chart generator for (semi-) lexicalist grammars. In Proceedings of the 7th European Workshop on Natural Language Generation (EWNLG99), pages 86–95, Toulouse. Carroll, John and Stephan Oepen. 2005. High efficiency realization for a wide-coverage unification grammar. In Proceedings of the Second International Joint Conference on Natural Language Processing, IJCNLP’05, pages 165–176, Berlin. Chang, Pi-Chuan and Kristina Toutanova. 2007. A discriminative syntactic word order model for machine translation. In Proceedings of ACL, pages 9–16, Prague. Chiang, David. 2007. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228. Clark, Stephen and James R. Curran. 2007a. Perceptron training for a wide-coverage lexicalized-grammar parser. In Proceedings of the ACL 2007 Workshop on Deep Linguistic Processing, pages 9–16, Prague. Clark, Stephen and James R. Curran. 2007b. Wide-coverage efficient statistical parsing with CCG and log-linear models. Computational Linguistics, 33(4):493–552. Collins, Michael and Brian Roark. 2004. Incremental parsing with the perceptron algorithm. In Proceedings of ACL, pages 111–118, Barcelona.
Crammer, Koby, Ofer Dekel, Joseph Keshet, Shai Shalev-Shwartz, and Yoram Singer. 2006. Online passive-aggressive algorithms. Journal of Machine Learning Research, 7:551–585. Daumé, Hal, John Langford, and Daniel Marcu. 2009. Search-based structured prediction. Machine Learning, 75(3):297–325. Daumé, Hal and Daniel Marcu. 2005. Learning as search optimization: approximate large margin methods for structured prediction. In ICML, pages 169–176, Bonn. Espinosa, Dominic, Michael White, and Dennis Mehay. 2008. Hypertagging: Supertagging for surface realization with CCG. In Proceedings of ACL-08: HLT, pages 183–191, Columbus, OH. Filippova, Katja and Michael Strube. 2007. Generating constituent order in German clauses. In Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 320–327, Prague. Filippova, Katja and Michael Strube. 2009. Tree linearization in English: Improving language model based approaches. In Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Short Papers, pages 225–228, Boulder, CO. Fowler, Timothy A. D. and Gerald Penn. 2010. Accurate context-free parsing with combinatory categorial grammar. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 335–344, Uppsala. Goldberg, Yoav and Michael Elhadad. 2010. An efficient algorithm for easy-first non-directional dependency parsing. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 742–750, Los Angeles, CA. Guo, Yuqing, Deirdre Hogan, and Josef van Genabith. 2011. DCU at generation challenges 2011 surface realisation track. In Proceedings of the Generation Challenges Session at the 13th European Workshop on Natural Language Generation, pages 227–229, Nancy. He, Wei, Haifeng Wang, Yuqing Guo, and Ting Liu. 2009. Dependency based Chinese sentence realization. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint
Conference on Natural Language Processing of the AFNLP, pages 809–816, Suntec. Hockenmaier, Julia. 2003. Data and Models for Statistical Parsing with Combinatory Categorial Grammar. Ph.D. thesis, School of Informatics, University of Edinburgh. Hockenmaier, Julia and Mark Steedman. 2007. CCGbank: A corpus of CCG derivations and dependency structures extracted from the Penn Treebank. Computational Linguistics, 33(3):355–396. Huang, Liang, Suphan Fayong, and Yang Guo. 2012. Structured perceptron with inexact search. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 142–151, Montréal. Huang, Liang and Kenji Sagae. 2010. Dynamic programming for linear-time incremental parsing. In Proceedings of ACL, pages 1077–1086, Uppsala. Johansson, Richard and Pierre Nugues. 2007. Extended constituent-to-dependency conversion for English. In 16th Nordic Conference of Computational Linguistics, pages 105–112, Tartu. Kay, Martin. 1996. Chart generation. In Proceedings of ACL, pages 200–204, Santa Cruz, CA. Koehn, Phillip. 2010. Statistical Machine Translation. Cambridge University Press. Koehn, Philipp, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180, Prague. Koehn, Philipp, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology - Volume 1, NAACL ’03, pages 48–54, Edmonton, Canada. Koo, Terry and Michael Collins. 2010. Efficient third-order dependency parsers. In Proceedings of ACL, pages 1–11, Uppsala. Lee, J. and S. Seneff. 2006. Automatic grammar correction for second-language learners. In Proceedings of Interspeech, pages 1978–1981, Pittsburgh, PA.
Liu, Yang. 2013. A shift-reduce parsing algorithm for phrase-based string-to-dependency translation. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1–10, Sofia. Marcus, Mitchell P., Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313–330. Papineni, Kishore, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, PA. Ratnaparkhi, Adwait. 1996. A maximum entropy model for part-of-speech tagging. In Proceedings of EMNLP, pages 133–142, Somerset, NJ. Reiter, Ehud and Robert Dale. 1997. Building applied natural language generation systems. Natural Language Engineering, 3(1):57–87. Shen, Libin and Aravind Joshi. 2008. LTAG dependency parsing with bidirectional incremental construction. In Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 495–504, Honolulu. HI. Shen, Libin, Giorgio Satta, and Aravind Joshi. 2007. Guided learning for bidirectional sequence classification. In Proceedings of ACL, pages 760–767, Prague. Song, Linfeng, Yue Zhang, Kai Song, and Qun Liu. 2014. Joint morphological generation and syntactic linearization. In Proceedings of the Twenty-Eighth AAAI Conference, Quebec. Steedman, Mark. 2000. The Syntactic Process. The MIT Press, Cambridge, MA. Surdeanu, Mihai, Richard Johansson, Adam Meyers, Lluı́s Màrquez, and Joakim Nivre. 2008. The CONLL-2008 shared task on joint parsing of syntactic and semantic dependencies. In Proceedings of the Twelfth Conference on Computational Natural Language Learning, pages 159–177, Manchester. Wan, Stephen, Mark Dras, Robert Dale, and Cécile Paris. 2009. Improving grammaticality in statistical sentence generation: Introducing a dependency spanning tree algorithm with an argument satisfaction model. In Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pages 852–860, Athens.
Weir, David. 1988. Characterizing Mildly Context-Sensitive Grammar Formalisms. Ph.D. thesis, University of Pennsylviania. White, Michael. 2004. Reining in CCG chart realization. In Proceedings of INLG-04, pages 182–191, Brockenhurst. White, Michael. 2006. Efficient Realization of Coordinate Structures in Combinatory Categorial Grammar. Research on Language & Computation, 4(1):39–75. White, Michael and Rajakrishnan Rajkumar. 2009. Perceptron reranking for CCG realization. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 410–419, Singapore. Wu, Dekai. 1997. Stochastic inversion transduction grammars and bilingual parsing of parallel corpora. Computational Linguistics, 23(3):377–403. Xu, Peng, Ciprian Chelba, and Frederick Jelinek. 2002. A study on richer syntactic dependencies for structured language modeling. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 191–198, Philadelphia, PA. Zettlemoyer, Luke S. and Michael Collins. 2005. Learning to map sentences to logical form: Structured classification with probabilistic categorial grammars. In Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence, pages 658–666, Edinburgh.
Zhang, Yue. 2013. Partial-tree linearization: Generalized word ordering for text synthesis. In Proceedings of IJCAI, pages 2232–2238, Beijing. Zhang, Yue, Graeme Blackwood, and Stephen Clark. 2012. Syntax-based word ordering incorporating a large-scale language model. In Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 736–746, Avignon. Zhang, Yue and Stephen Clark. 2008. Joint word segmentation and POS tagging using a single perceptron. In Proceedings of ACL/HLT, pages 888–896, Columbus, OH. Zhang, Yue and Stephen Clark. 2010. A fast decoder for joint word segmentation and POS-tagging using a single discriminative model. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 843–852, Cambridge, MA. Zhang, Yue and Stephen Clark. 2011. Syntactic processing using the generalized perceptron and beam search. Computational Linguistics, 37(1):105–151. Zhang, Yue and Joakim Nivre. 2011. Transition-based dependency parsing with rich non-local features. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pages 188–193, Portland, OR. 42 43.41 44.71 50.46 51.40 52.18 73.34 74.68 76.28 :","2 overview of the search and training algorithms :In this section, the CCG-based system is used to describe the search and training algorithms. However, the same approach can be used for the dependency-based system, as described in Section 4: Instead of building hypotheses by applying CCG rules in a bottom-up manner, the dependency-based system creates dependency links between words.
Given a bag of words, the goal is to put them into an ordered sentence that has a plausible CCG derivation. The search space of the decoding problem consists of all possible CCG derivations for all possible word permutations, and the search goal is to find the highest-scored derivation in the search space. This is an NP-hard problem, as mentioned in the Introduction. We apply learning-guided search to address the high complexity. The intuition is that, because the whole search space cannot be exhausted in order to find the optimal solution, we choose to explore a small area in the search space. A statistical model is used to guide the search, so that only a small portion of the search space containing the most plausible hypotheses is explored.
One natural choice for the decoding algorithm is best-first search, which uses an agenda to order hypotheses, and expands the highest-scored hypothesis on the agenda at each step. The resulting hypotheses after each hypothesis expansion are put back on the agenda, and the process repeats until a goal hypothesis (a full sentence) is found. This search process is guided by the current scores of the hypotheses, and the search path will contain the most plausible hypotheses if they are scored higher than implausible ones. An alternative to best-first search is A∗ search, which makes use of a heuristic function to estimate future scores. A∗ can potentially be more efficient given an effective heuristic function; however, it is not straightforward to define an admissible and accurate estimate of future scores for our problem, and we leave this research question to future work.
In our formulation of the word ordering problem, a hypothesis is a phrase or sentence together with its CCG derivation. Hypotheses are constructed bottom–up: starting from single words, smaller phrases are combined into larger ones according to CCG rules. To allow the combination of hypotheses, we use an additional structure to store a set of hypotheses that have been expanded, which we call accepted
hypotheses. When a hypothesis from the agenda is expanded, it is combined with all accepted hypotheses in all possible ways to produce new hypotheses. The data structure for accepted hypotheses is similar to that used for best-first parsing (Caraballo and Charniak 1998), and we adopt the term chart for this structure. However, note there are important differences to the parsing problem. First, the parsing problem has a fixed word order and is considerably simpler than the word ordering problem we are tackling. Second, although we use the term chart, the structure for accepted hypotheses is not a dynamic programming chart in the same way as for the parsing problem. In our previous papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012), we applied a set of beams to this structure, which makes it similar to the data structure used for phrase-based MT decoding (Koehn 2010). However, we will show later that this structure is unnecessary when the model has more discriminative power, and a conceptually simpler single beam can be used. We will also investigate the possibility of applying dynamic-programming-style pruning to the chart.
We now give an overview of the training algorithm, which is crucial to both the speed and accuracy of the resulting decoder. CCGBank (Hockenmaier and Steedman 2007) is used to train the model. For each training sentence, the corresponding CCGBank derivation together with all its sub-derivations are treated as gold-standard hypotheses. All other hypotheses that can be constructed from the same bag of words are non-gold hypotheses. From the generation perspective this assumption is too strong, because sentences can have multiple orderings (with multiple derivations) that are both grammatical and fluent. Nevertheless, it is the most feasible choice given the training data available.
The efficiency of the decoding algorithm is dependent on the training algorithm because the agenda is ordered according to the hypothesis scores. Hence, a better model will lead to a goal hypothesis being found more quickly. In the ideal situation, all goldstandard hypotheses are scored higher than all non-gold hypotheses, and therefore only gold-standard hypotheses are expanded before the gold-standard goal hypothesis is found. In this case, the minimum number of hypotheses is expanded and the output is correct. The best-first search decoder is optimal not only with respect to accuracy but also speed. This ideal situation can hardly be met in practice, but it determines the goal of the training algorithm: to find a model that scores gold-standard hypotheses higher than non-gold ones.
Learning-guided search places more challenges on the training of a discriminative model than standard structured prediction problems, for example, CKY parsing for CCG (Clark and Curran 2007b). If we take gold-standard hypotheses as positive training examples, and non-gold hypotheses as negative examples, then the training goal is to find a large separating margin between the scores of all positive examples and all negative examples. For CKY parsing, the highest-scored negative example can be found via optimal Viterbi decoding, according to the current model, and this negative example can be used in place of all negative examples during the updating of parameters. In contrast, our best-first search algorithm cannot find an output in reasonable time unless a good model has already been trained, and therefore we cannot run the decoding algorithm in the standard way during training. In our previous papers (Zhang and Clark 2011; Zhang, Blackwood, and Clark 2012), we proposed an approximate online training algorithm, which forces positive examples to be kept in the hypothesis space without being discarded, and prevents the expansion of negative examples during the training process (so that the hypothesis space does not get too large). This algorithm ensures training efficiency, but greatly limits the space of negative examples that is explored during training (and hence fails to replicate the conditions experienced at test
time). In this article, we will show that, with an improved scoring model, it is possible to expand negative examples, which leads to improved performance.
A second and more subtle challenge for our training problem is that we need a stronger model for learning-guided search than for dynamic programming (DP)– based search, such as CKY decoding. For CKY decoding, the model is used to compare hypotheses within each chart cell, which cover the same input words. In contrast, for the best-first search decoder, the model is used to order hypotheses on the agenda, which can cover different numbers of words. It needs much stronger discriminating power, so that it can determine whether a single-word phrase is better than, say, a 40-word sentence. In this article we use scaling of the hypothesis scores by size, so that hypotheses of different sizes can be fairly compared. We also find that, with this new approach, negative examples can be expanded during training and a single beam applied to the chart, resulting in a conceptually simpler and more effective training algorithm and decoder."
,,,,,,,,"A book aiming to build a bridge between two fields that share the subject of research but do not share the same views necessarily puts itself in a difficult position: The authors have either to strike a fair balance at peril of dissatisfying both sides or nail their colors to the mast and cater mainly to one of two communities. For semantic processing of natural language with either NLP methods or Semantic Web approaches, the authors clearly favor the latter and propose a strictly ontology-driven interpretation of natural language. The main contribution of the book, driving semantic processing from the ground up by a formal domain-specific ontology, is elaborated in ten well-structured chapters spanning 143 pages of content.","[{""affiliations"": [], ""name"": ""Philipp Cimiano""}, {""affiliations"": [], ""name"": ""Christina Unger""}]",SP:84be813d4c76e7279b5a4abf69a795823c01a776,"[{""authors"": [""J. Bos""], ""title"": ""Wide-coverage semantic analysis with boxer"", ""venue"": ""Proceedings of the 2008 Conference on Semantics in Text Processing STEP-08, pages 277\u2013286, Venice."", ""year"": 2008}, {""authors"": [""P. Cimiano""], ""title"": ""Flexible semantic composition with DUDES"", ""venue"": ""Proceedings of the 8th International Conference on Computational Semantics (IWCS\u201909), pages 272\u2013276, Tilburg."", ""year"": 2009}, {""authors"": [""Kamp H."", ""U. Reyle.""], ""title"": ""From Discourse to Logic"", ""venue"": ""Introduction to Modeltheoretic Semantics of Natural Language, Formal Logic and Discourse Representation Theory. Kluwer, Dordrecht."", ""year"": 1993}, {""authors"": [""Nirenburg S."", ""V. Raskin.""], ""title"": ""Ontological Semantics"", ""venue"": ""MIT Press, Cambridge, MA. Schabes Y. 1990. Mathematical and Computational Aspects of Lexicalized Grammars. Ph.D. thesis, University of"", ""year"": 2004}]",,,,,,,,,,,,,,,,book reviews :Ontology-Based Interpretation of Natural Language,,,reviewed by chris biemann :,,,,,,,,,,,,,,,,,,,,,,,,"philipp cimiano, christina unger, and john mccrae :(University of Arminia Bielefeld, Germany)
Morgan & Claypool, Synthesis Lectures on Human Language Technologies, March 2014, 178 pages, (doi:10.2200/S00561ED1V01Y201401HLT024) , $45.00","tu darmstadt, germany :A book aiming to build a bridge between two fields that share the subject of research but do not share the same views necessarily puts itself in a difficult position: The authors have either to strike a fair balance at peril of dissatisfying both sides or nail their colors to the mast and cater mainly to one of two communities. For semantic processing of natural language with either NLP methods or Semantic Web approaches, the authors clearly favor the latter and propose a strictly ontology-driven interpretation of natural language. The main contribution of the book, driving semantic processing from the ground up by a formal domain-specific ontology, is elaborated in ten well-structured chapters spanning 143 pages of content.
© 2015 Association for Computational Linguistics
Throughout the book, examples from the soccer domain excellently illustrate the main concepts in high machine-readable detail. The first chapter sets the scene with a motivating example: Humans have sufficient background knowledge to interpret a description of a soccer match and can grasp a lot of meaning beyond the literal content. For example, if team A scored two goals and team B eventually wins the match, team B scored at least three goals. To enable these and other kinds of inferences in machines, these need to be equipped with a domain ontology that formalizes domain knowledge and domain-specific reasoning, as well as with a mechanism to construct formal representations of natural language text. The key idea of this book is to place the ontology at the center of such an interpretation process: The domain and all its relevant semantic distinctions are defined in the ontology, thus the natural language semantic parser must only be aware of these. As opposed to generic tools such as Boxer (Bos 2008) that turn natural language into formal representations, driving the formalization of NL directly by its target ontology ensures that it can be directly consumed by further layers of interpretation, such as reasoners. The remainder of the first chapter gives a very short summary of the state of affairs in semantic interpretation and semantic parsing in NLP. Although this summary is sufficient to highlight what is missing for drawing inferences on top of natural language statements, it is—with a mere two pages— necessarily incomplete and omits even mainstream approaches such as frame semantic parsing or semantic role labeling. Further, the relation of the proposal to the Semantic Web is discussed, from which formats, interoperability, and description languages are leveraged. Primers on the RDF data format and on the soccer domain complete the chapter.
doi:10.1162/COLI r 00223
Chapter 2 defines the concept of ontology. Great care is taken to discuss the definition of ontologies and various ontology description languages such as OWL 2 DL, as well as their expressivity. In Chapter 3, linguistic formalisms for representing syntax and semantics of natural language are discussed and ways to connect these to the ontology are introduced. Although, in principle, there are many possible formalisms to serve the “linguistic side,” the authors decide on Lexicalized Tree Adjoining Grammar (LTA, Schabes 1990) for syntactic processing. In the LTAG lexicon, each lexical entry links to all elementary trees (subtrees that specify the valid contexts in terms of constituents) it anchors, whereas the corresponding ontology grammar links each ontological concept to the projections of all lexical items that verbalize this concept. Thus, each concept is associated with an exhaustive enumeration of patterns it is expressed in. On the semantic level, Discourse Representation Theory (DRT; Kamp and Reyle 1993) is chosen as a formalism for semantic operations such as coreference and quantification. DRT and LTAG are subsequently paired in a representation called DUDES (Dependency-based Underspecified Discourse Representation structures; Cimiano 2009). Here, the DRTinspired representations are again linked to ontology concepts. Alhough the choice of framework seems largely rooted in previous works of the first author, it is clearly stated that the connection to the ontology could also be realized for other syntactic (e.g., LFG, HPSG) and semantic (e.g., GLUE, MRS) frameworks.
Chapter 4 deals with the representation of the ontological lexicons, which specify the interpretation of lexical items with respect to the target ontology. The declarative LEMON lexicon model for ontologies, following closely the approach to the lexicon of Ontological Semantics (Nirenburg and Raskin 2004), is laid out and exemplified in great detail. In the fifth chapter, the authors describe how the ontology grammar for DUDES—at least, the ontology-specific parts—can be generated from the lexicon. While this might seem merely a syntactic manipulation as the patterns of concept manifestation have been encoded in the lexicon, the point is to keep the lexicon separate from DUDES to allow for the generation of other types of ontology grammars. Note that different domain ontologies generate different grammars that could yield different interpretations for the same sentence. In Chapter 6, the overall interplay of previously discussed ingredients for semantic interpretation is exemplified, and challenges with respect to structural ambiguities and underspecification are pointed out. These are supposed to be left underspecified until they can be resolved with ontological reasoning, which is the subject of Chapter 7. Subsequently, the formalization of time in the framework is declared at length in Chapter 8, building on time interval calculi from the literature. In Chapter 9, the application of question answering over structured data is discussed—not to be confused with question answering from unstructured sources in the flavor of TREC or IBM Watson. In such a system, the natural language query is translated into a SPARQL query that can be run against public endpoints. While the examples successfully illustrate the translation from DUDES representations to SPARQL queries, no qualitative or quantitative evaluation is provided despite the existence of public benchmarks.
The final chapter concludes with a very strong statement: The mainstream of computational linguistics research supposedly concentrates on domain-independent semantic representations, which is deemed ”wrong,” as it would not do justice to domain distinctions and would not enable the connection to domain-specific ontologies to allow for domain-specific reasoning. Regretting that a principled use of ontologies in NLP is not widespread, the authors finally paint a vision of an ontology-based NLP ecosystem modeled after their approach. While proposing the use of statistical methods to overcome brittleness, lack of coverage, and the limitations of logical
reasoning in the presence of uncertainty, it remains entirely unclear how this could be carried out.
Throughout, the book contains exercises that greatly help in internalizing the discussed concepts. Also, an array of available resources including a demo of the question answering system is announced. Some of these, however, were not yet available on the companion Web site at the time of review.
Even though the book states that its focus is “certainly not on robustness and coverage” (p. 6), I am still missing more concrete ideas on how robustness and coverage might be addressed than “by using machine learning techniques” (p. 142) without further elaboration. Thus, it is no coincidence that the implementation remains on the level of toy examples, however illustrating they might be. Also, while admitting to domain drift and the necessity of adaptation, it remains unclear how this adaptation, the integration of domain-specific and domain-independent processing, as well as the integration of several domains can be tractably attained. A deeper account on related approaches to computational semantics and a more recent bibliography on NLP/CL approaches could probably have avoided the impression that the interpretation of language as advocated here provides hardly more than other, same-old classic AI approaches.
Also, the pledge to the field of NLP to use more of Semantic Web technologies and services is more Semantic Web–centric than I would have expected from a book “attempting to provide a step towards the synergy between these two fields” (p. xv): True, it is unfortunate that these two fields interact as little as they do, as both are concerned with understanding the semantics of text, but the authors’ clearly voiced credo on how NLP should finally follow suit is not very helpful for creating such synergy. To bridge the gap, the Semantic Web community should probably start leveraging results and evaluation methodologies from CL/NLP instead of reinventing/ignoring them, and the NLP community probably has to understand that ontological grounding and resolving entities to URIs is in fact enabling applications that go well beyond dependency parsing and word sense disambiguation. The Semantic Web community should understand that statistical methods and unsupervised acquisition are not the enemy, but the key to scalable, domain-adaptive processing of natural language, which has been known to evade total formalization since Sapier’s “all grammars leak.” The NLP community, especially statistical semantics, in turn, should understand the need and the demand for linking concepts to manually curated taxonomies and controlled vocabularies, as these are pervasive in industrial knowledge management. NLP should see the benefit in standardization for interoperability, and Semantic Webbers have to finally figure out that from format does not follow function.
To sum up, a well-written book goes the extra mile to exemplify its core contribution, which is to consequently drive language processing by the ontology of a target application from the very start. The exercises, together with the online materials, make it a useful resource for teaching. Unlike a similar proposal by Nirenburg and Raskin (2004), it is constructed and exemplified with Semantic Web technology such as RDF, OWL, SPARQL, open data, standardization, reasoning, and domain ontology. For this reason, the book will supposedly be very well received in the Semantic Web community, but could be perceived controversially by the NLP community: While the idea to radically align natural language processing to a target-domain ontology is reasonable, wellmotivated from the application point of view, and a worthwhile contribution, the book unfortunately does not succeed in convincing most modern computational linguists that this approach can overcome known limitations of previous similar rule-based and knowledge-driven approaches, such as brittleness, poor scalability, and little adaptivity.",,,,,,"A book aiming to build a bridge between two fields that share the subject of research but do not share the same views necessarily puts itself in a difficult position: The authors have either to strike a fair balance at peril of dissatisfying both sides or nail their colors to the mast and cater mainly to one of two communities. For semantic processing of natural language with either NLP methods or Semantic Web approaches, the authors clearly favor the latter and propose a strictly ontology-driven interpretation of natural language. The main contribution of the book, driving semantic processing from the ground up by a formal domain-specific ontology, is elaborated in ten well-structured chapters spanning 143 pages of content. [{""affiliations"": [], ""name"": ""Philipp Cimiano""}, {""affiliations"": [], ""name"": ""Christina Unger""}] SP:84be813d4c76e7279b5a4abf69a795823c01a776 [{""authors"": [""J. Bos""], ""title"": ""Wide-coverage semantic analysis with boxer"", ""venue"": ""Proceedings of the 2008 Conference on Semantics in Text Processing STEP-08, pages 277\u2013286, Venice."", ""year"": 2008}, {""authors"": [""P. Cimiano""], ""title"": ""Flexible semantic composition with DUDES"", ""venue"": ""Proceedings of the 8th International Conference on Computational Semantics (IWCS\u201909), pages 272\u2013276, Tilburg."", ""year"": 2009}, {""authors"": [""Kamp H."", ""U. Reyle.""], ""title"": ""From Discourse to Logic"", ""venue"": ""Introduction to Modeltheoretic Semantics of Natural Language, Formal Logic and Discourse Representation Theory. Kluwer, Dordrecht."", ""year"": 1993}, {""authors"": [""Nirenburg S."", ""V. Raskin.""], ""title"": ""Ontological Semantics"", ""venue"": ""MIT Press, Cambridge, MA. Schabes Y. 1990. Mathematical and Computational Aspects of Lexicalized Grammars. Ph.D. thesis, University of"", ""year"": 2004}] book reviews :Ontology-Based Interpretation of Natural Language reviewed by chris biemann : philipp cimiano, christina unger, and john mccrae :(University of Arminia Bielefeld, Germany)
Morgan & Claypool, Synthesis Lectures on Human Language Technologies, March 2014, 178 pages, (doi:10.2200/S00561ED1V01Y201401HLT024) , $45.00 tu darmstadt, germany :A book aiming to build a bridge between two fields that share the subject of research but do not share the same views necessarily puts itself in a difficult position: The authors have either to strike a fair balance at peril of dissatisfying both sides or nail their colors to the mast and cater mainly to one of two communities. For semantic processing of natural language with either NLP methods or Semantic Web approaches, the authors clearly favor the latter and propose a strictly ontology-driven interpretation of natural language. The main contribution of the book, driving semantic processing from the ground up by a formal domain-specific ontology, is elaborated in ten well-structured chapters spanning 143 pages of content.
© 2015 Association for Computational Linguistics
Throughout the book, examples from the soccer domain excellently illustrate the main concepts in high machine-readable detail. The first chapter sets the scene with a motivating example: Humans have sufficient background knowledge to interpret a description of a soccer match and can grasp a lot of meaning beyond the literal content. For example, if team A scored two goals and team B eventually wins the match, team B scored at least three goals. To enable these and other kinds of inferences in machines, these need to be equipped with a domain ontology that formalizes domain knowledge and domain-specific reasoning, as well as with a mechanism to construct formal representations of natural language text. The key idea of this book is to place the ontology at the center of such an interpretation process: The domain and all its relevant semantic distinctions are defined in the ontology, thus the natural language semantic parser must only be aware of these. As opposed to generic tools such as Boxer (Bos 2008) that turn natural language into formal representations, driving the formalization of NL directly by its target ontology ensures that it can be directly consumed by further layers of interpretation, such as reasoners. The remainder of the first chapter gives a very short summary of the state of affairs in semantic interpretation and semantic parsing in NLP. Although this summary is sufficient to highlight what is missing for drawing inferences on top of natural language statements, it is—with a mere two pages— necessarily incomplete and omits even mainstream approaches such as frame semantic parsing or semantic role labeling. Further, the relation of the proposal to the Semantic Web is discussed, from which formats, interoperability, and description languages are leveraged. Primers on the RDF data format and on the soccer domain complete the chapter.
doi:10.1162/COLI r 00223
Chapter 2 defines the concept of ontology. Great care is taken to discuss the definition of ontologies and various ontology description languages such as OWL 2 DL, as well as their expressivity. In Chapter 3, linguistic formalisms for representing syntax and semantics of natural language are discussed and ways to connect these to the ontology are introduced. Although, in principle, there are many possible formalisms to serve the “linguistic side,” the authors decide on Lexicalized Tree Adjoining Grammar (LTA, Schabes 1990) for syntactic processing. In the LTAG lexicon, each lexical entry links to all elementary trees (subtrees that specify the valid contexts in terms of constituents) it anchors, whereas the corresponding ontology grammar links each ontological concept to the projections of all lexical items that verbalize this concept. Thus, each concept is associated with an exhaustive enumeration of patterns it is expressed in. On the semantic level, Discourse Representation Theory (DRT; Kamp and Reyle 1993) is chosen as a formalism for semantic operations such as coreference and quantification. DRT and LTAG are subsequently paired in a representation called DUDES (Dependency-based Underspecified Discourse Representation structures; Cimiano 2009). Here, the DRTinspired representations are again linked to ontology concepts. Alhough the choice of framework seems largely rooted in previous works of the first author, it is clearly stated that the connection to the ontology could also be realized for other syntactic (e.g., LFG, HPSG) and semantic (e.g., GLUE, MRS) frameworks.
Chapter 4 deals with the representation of the ontological lexicons, which specify the interpretation of lexical items with respect to the target ontology. The declarative LEMON lexicon model for ontologies, following closely the approach to the lexicon of Ontological Semantics (Nirenburg and Raskin 2004), is laid out and exemplified in great detail. In the fifth chapter, the authors describe how the ontology grammar for DUDES—at least, the ontology-specific parts—can be generated from the lexicon. While this might seem merely a syntactic manipulation as the patterns of concept manifestation have been encoded in the lexicon, the point is to keep the lexicon separate from DUDES to allow for the generation of other types of ontology grammars. Note that different domain ontologies generate different grammars that could yield different interpretations for the same sentence. In Chapter 6, the overall interplay of previously discussed ingredients for semantic interpretation is exemplified, and challenges with respect to structural ambiguities and underspecification are pointed out. These are supposed to be left underspecified until they can be resolved with ontological reasoning, which is the subject of Chapter 7. Subsequently, the formalization of time in the framework is declared at length in Chapter 8, building on time interval calculi from the literature. In Chapter 9, the application of question answering over structured data is discussed—not to be confused with question answering from unstructured sources in the flavor of TREC or IBM Watson. In such a system, the natural language query is translated into a SPARQL query that can be run against public endpoints. While the examples successfully illustrate the translation from DUDES representations to SPARQL queries, no qualitative or quantitative evaluation is provided despite the existence of public benchmarks.
The final chapter concludes with a very strong statement: The mainstream of computational linguistics research supposedly concentrates on domain-independent semantic representations, which is deemed ”wrong,” as it would not do justice to domain distinctions and would not enable the connection to domain-specific ontologies to allow for domain-specific reasoning. Regretting that a principled use of ontologies in NLP is not widespread, the authors finally paint a vision of an ontology-based NLP ecosystem modeled after their approach. While proposing the use of statistical methods to overcome brittleness, lack of coverage, and the limitations of logical
reasoning in the presence of uncertainty, it remains entirely unclear how this could be carried out.
Throughout, the book contains exercises that greatly help in internalizing the discussed concepts. Also, an array of available resources including a demo of the question answering system is announced. Some of these, however, were not yet available on the companion Web site at the time of review.
Even though the book states that its focus is “certainly not on robustness and coverage” (p. 6), I am still missing more concrete ideas on how robustness and coverage might be addressed than “by using machine learning techniques” (p. 142) without further elaboration. Thus, it is no coincidence that the implementation remains on the level of toy examples, however illustrating they might be. Also, while admitting to domain drift and the necessity of adaptation, it remains unclear how this adaptation, the integration of domain-specific and domain-independent processing, as well as the integration of several domains can be tractably attained. A deeper account on related approaches to computational semantics and a more recent bibliography on NLP/CL approaches could probably have avoided the impression that the interpretation of language as advocated here provides hardly more than other, same-old classic AI approaches.
Also, the pledge to the field of NLP to use more of Semantic Web technologies and services is more Semantic Web–centric than I would have expected from a book “attempting to provide a step towards the synergy between these two fields” (p. xv): True, it is unfortunate that these two fields interact as little as they do, as both are concerned with understanding the semantics of text, but the authors’ clearly voiced credo on how NLP should finally follow suit is not very helpful for creating such synergy. To bridge the gap, the Semantic Web community should probably start leveraging results and evaluation methodologies from CL/NLP instead of reinventing/ignoring them, and the NLP community probably has to understand that ontological grounding and resolving entities to URIs is in fact enabling applications that go well beyond dependency parsing and word sense disambiguation. The Semantic Web community should understand that statistical methods and unsupervised acquisition are not the enemy, but the key to scalable, domain-adaptive processing of natural language, which has been known to evade total formalization since Sapier’s “all grammars leak.” The NLP community, especially statistical semantics, in turn, should understand the need and the demand for linking concepts to manually curated taxonomies and controlled vocabularies, as these are pervasive in industrial knowledge management. NLP should see the benefit in standardization for interoperability, and Semantic Webbers have to finally figure out that from format does not follow function.
To sum up, a well-written book goes the extra mile to exemplify its core contribution, which is to consequently drive language processing by the ontology of a target application from the very start. The exercises, together with the online materials, make it a useful resource for teaching. Unlike a similar proposal by Nirenburg and Raskin (2004), it is constructed and exemplified with Semantic Web technology such as RDF, OWL, SPARQL, open data, standardization, reasoning, and domain ontology. For this reason, the book will supposedly be very well received in the Semantic Web community, but could be perceived controversially by the NLP community: While the idea to radically align natural language processing to a target-domain ontology is reasonable, wellmotivated from the application point of view, and a worthwhile contribution, the book unfortunately does not succeed in convincing most modern computational linguists that this approach can overcome known limitations of previous similar rule-based and knowledge-driven approaches, such as brittleness, poor scalability, and little adaptivity.",
"1 introduction :We live in an information age where all kinds of information is easily accessible through the Internet. The increasing demand for access to different types of information available online have interested researchers in a broad range of Information Retrieval–related areas, such as question answering, topic detection and tracking, summarization, multimedia retrieval, chemical and biological informatics, text structuring, and text mining. Although search engines do a remarkable job in searching through a heap of information, they have certain limitations, as they cannot satisfy the end users’ information need to have more direct access to relevant documents. For example, if we ask for the impact of the current global financial crisis in different parts of the world, we can expect to sift through thousands of results for the answer. This fact can be more understandable by the following scenario. When a user enters a query, they are served with a ranked list of relevant documents by the standard document retrieval systems (i.e., search engines),
∗ University of Lethbridge, 4401 University Drive West, Lethbridge, Alberta, T1K 3M4, Canada. E-mail: chali@cs.uleth.ca.
∗∗ Philips Research North America, 345 Scarborough Rd, Briarcliff Manor, New York, 10510, USA. E-mail: sadid.hasan@philips.com.
Submission received: 19 May, 2013; revised submission received: 26 May, 2014; accepted for publication: 22 June, 2014.
doi:10.1162/COLI a 00206
© 2015 Association for Computational Linguistics
and their search task is usually not over (Chali, Joty, and Hasan 2009). The next step for the user is to look into the documents themselves and search for the precise piece of information they were looking for. This method is time-consuming, and a correct answer could easily be missed by either an incorrect query resulting in missing documents or by careless reading. This is why Question Answering (QA) has received immense attention from the information retrieval, information extraction, machine learning, and natural language processing communities in the last 15 years (Hirschman and Gaizauskas 2001; Strzalkowski and Harabagiu 2008; Kotov and Zhai 2010).
The main goal of QA systems is to retrieve relevant answers to natural language questions from a collection of documents rather than using keyword matching techniques to extract documents. Automated QA research focuses on how to respond with exact answers to a wide variety of questions, including: factoid, list, definition, how, why, hypothetical, semantically constrained, and crosslingual questions (Simmons 1965; Kupiec 1993; Voorhees 1999; Hirschman and Gaizauskas 2001; Greenwood 2005; Wang 2006; Moldovan, Clark, and Bowden 2007). One of the main requirements of a QA system is that it must receive a well-formed question as input in order to come up with the best possible correct answer as output. Available studies revealed that humans are not very skilled in asking good questions about a topic of their interest. They are forgetful in nature; this often restricts them to properly express whatever that is peeking in their mind. Therefore, they would benefit from automated Question Generation (QG) systems that can assist in meeting their inquiry needs (Lauer, Peacock, and Graesser 1992; Graesser et al. 2001; Rus and Graesser 2009; Ali, Chali, and Hasan 2010; Kotov and Zhai 2010; Olney, Graesser, and Person 2012). Another benefit of QG is that it can be a good tool to help improve the quality of the QA systems (Graesser et al. 2001; Rus and Graesser 2009). These benefits of a QG system motivate us to address the important problem of topic-to-question generation, where the main goal is to generate all possible questions about a given topic. For example, given the topic Apple Inc. Logos, we would like to generate questions such as What is Apple Inc.?, Where is Apple Inc. located?, Who designed Apple’s Logo?, and so forth.
The problem of topic-to-question generation can be viewed as a generalization of the problem of answering complex questions. Complex questions are essentially broader information requests about a certain topic, whose answers could be obtained from pieces of information scattered in multiple documents. For example, consider the complex question:1 Describe steps taken and worldwide reaction prior to the introduction of the Euro on January 1, 1999. Include predictions and expectations reported in the press. This question is requesting an elaboration about the topic “Introduction of the Euro,” which can be answered by following complex procedures such as question decomposition or inferencing and synthesizing information from multiple documents (e.g., multi-document summarization). Answering complex questions is not easy as it is not always understandable to which direction one should move to search for the answer to a complex question. This situation arises because of the wider focus of the topic that is inherent in the complex question in consideration. For example, a complex question like Describe the tsunami disaster in Japan has a wider focus without a single or well-defined information need. To narrow down the focus, this question can be decomposed into a series of simple questions such as How many people were killed in the tsunami?, How many people became homeless?, Which cities were mostly damaged?, and so on.
1 The example complex questions have been provided according to the guidelines of the Document Understanding Conference (DUC, http://duc.nist.gov/) (2005–2007) tasks.
Decomposing a complex question automatically into simpler questions in this manner such that each of them can be answered individually by using the state-of-the-art QA systems, and then combining the individual answers to form a single answer to the original complex question, has proven effective to deal with the complex question answering problem (Harabagiu, Lacatusu, and Hickl 2006; Hickl et al. 2006; Chali, Hasan, and Imam 2012). Moreover, the generated simple questions can be used as the list of important aspects to act as a guide2 for selecting the most relevant sentences in producing more focused and more accurate summaries as the output of a summarization system (Chali, Hasan, and Imam 2011, 2012). From this discussion, it is obvious that the complex question decomposition problem can be generalized to the problem of topic-to-question generation to help improve the complex question answering systems.
In this article3, we consider the task of automatically generating questions from topics and assume that each topic is associated with a body of texts having useful information about the topic. This assumption has been inherited from the process of how a human asks questions based on their knowledge. For example, if a person knows that a university is an educational institution, then they can ask a question about its faculty and students. In this research, our main goal is to generate fact-based questions4 about a given topic from its associated content information. We generate questions by exploiting the named entity information and the predicate argument structures of the sentences (along with semantic roles) present in the given body of texts. The named entities and the semantic role labels are used to identify relevant parts of a sentence in order to form relevant questions about them. The importance of the generated questions is measured in two steps. In the first step, we identify whether the question is asking something about the topic or something that is very closely related to the topic. We call this the measure of topic relevance. For this purpose, we use Latent Dirichlet Allocation (LDA) (Blei, Ng, and Jordan 2003) to identify the subtopics (which are closely related to the original topic) in the given body of texts and apply the Extended String Subsequence Kernel (ESSK) (Hirao et al. 2003) to calculate their similarity with the questions. In the second step, we judge the syntactic correctness of each generated question. We apply the tree kernel functions (Collins and Duffy 2001) and re-implement the syntactic tree kernel model according to Moschitti et al. (2007) for computing the syntactic similarity of each question with the associated content information. We rank the questions by considering their topic relevance and syntactic correctness scores. Experimental results show the effectiveness of our approach for automatically generating topical questions. The remainder of the article is organized as follows. Section 2 describes the related work. Section 3 presents the description of our QG system. Section 4 explains the experiments and shows evaluation results; Section 5 concludes.","2 related work :Recently, question generation has received immense attention from researchers and different methods have been proposed to accomplish the task in different relevant fields (Andrenucci and Sneiders 2005). McGough et al. (2001) proposed an approach to build a Web-based testing system with the facility of dynamic QG. Wang et al. (2008) showed
2 http://www.nist.gov/tac/2011/Summarization/Guided-Summ.2011.guidelines.html. 3 This article is a longer version of our previously published work (Chali and Hasan 2012c). We provide
more theoretical descriptions and analyses, and conduct our experiments on a larger data set to report new results.
4 We mainly focus on generating who, what, where, which, when, why, and how questions in this research.
a method to automatically generate questions based on question templates (which are created from training on medical articles). Brown, Frishkoff, and Eskenazi (2005) described an approach to automatically generate questions to assess the user’s vocabulary knowledge. Chen, Aist, and Mostow (2009) developed a method to generate questions automatically from informational text to mimic the reader’s self-questioning strategy during reading. On the other hand, Agarwal, Shah, and Mannem (2011) considered the question generation problem beyond the sentence level and designed an approach that uses discourse connectives to generate questions from a given text. Several other QG models have been proposed over the years that deal with transforming answers to questions and utilizing question generation as an intermediate step in the question answering process (Echihabi and Marcu 2003; Hickl et al. 2005). There are some other researchers who have approached the task of generating questions for educational purposes (Mitkov and Ha 2003; Heilman and Smith 2010b).
Question asking and QG are important components in advanced learning technologies such as intelligent tutoring systems and inquiry-based environments (Graesser et al. 2001). A QG system is useful for building better question-asking facilities in intelligent tutoring systems. The Natural Language Processing (NLP), Natural Language Generation, Intelligent Tutoring System, and Information Retrieval communities have currently identified the Text-to-Question generation task as promising candidates for shared tasks5 (Rus and Graesser 2009; Boyer and Piwek 2010). In the Text-to-Question generation task, a QG system is given a text, and the goal is to generate a set of questions for which the text contains answers. The task of generating a question about a given text can be typically decomposed into three subtasks. First, given the source text, a content selection step is necessary to select a target to ask about, such as the desired answer. Second, given a target answer, an appropriate question type is selected (i.e., the form of question to ask is determined). Third, given the content and question type, the actual question is constructed. Based on this principle, several approaches have been described in Boyer and Piwek (2010) that use named entity information, syntactic knowledge, and semantic structures of the sentences to perform the task of generating questions from sentences and paragraphs (Heilman and Smith 2010a; Mannem, Prasad, and Joshi 2010). Inspired by these works, we perform the task of topic-to-question generation using named entity information and semantic structures of the sentences. A task that is similar to ours is the task of keywords-to-question generation that has been addressed recently in Zheng et al. (2011). They propose a user model for jointly generating keywords and questions. However, their approach is based on generating question templates from existing questions, which requires a large set of English questions as training data. In recent years, some other related researchers have proposed the tasks of high-quality question generation (Ignatova, Bernhard, and Gurevych 2008) and generating questions from queries (Lin 2008). Fact-based question generation has been accomplished previously (Rus, Cai, and Graesser 2007; Heilman and Smith 2010b). We also focus on generating fact-based questions in this research.
Besides grammaticality, an effective QG system should focus deeply on the importance of the generated questions (Vanderwende 2008). This motivates the use of a question-ranking module in a typical QG system. Over-generated questions can be ranked using different approaches, such as statistical ranking methods, dependency parsing, identification of the presence of pronouns and named entities, and topic scoring (Heilman and Smith 2010a; Mannem, Prasad, and Joshi 2010; McConnell et al. 2011).
5 http://www.questiongeneration.org/QGSTEC2010.
However, most of these automatic ranking approaches ignore the aspects of complex paraphrasing by not considering lexical semantic variations (e.g., synonymy) when measuring the importance of the questions. In our work, we use LDA (Blei, Ng, and Jordan 2003) to identify the subtopics (which are closely related to the original topic) in the given body of texts. We choose LDA because in recent years it has become one of the most popular topic modeling techniques and has been shown to be effective in several text-related tasks, such as document classification, information retrieval, and question answering (Wei and Croft 2006; Misra, Cappé, and Yvon 2008; Celikyilmaz, HakkaniTur, and Tur 2010).
Once we have the subtopics, we apply ESSK (Hirao et al. 2003) to calculate their similarity with the generated questions. The choice of ESSK is motivated by its successful use in different NLP tasks in recent years (Chali, Hasan, and Joty 2009, 2011; Chali and Hasan 2012a, 2012b). Hirao et al. (2003) introduced ESSK considering all possible senses of each word to perform their summarization task. Their method is effective. However, the fact that they do not disambiguate word senses cannot be disregarded. In our task, we apply ESSK to calculate the similarity between important topics (discovered using LDA) and the generated questions in order to measure the importance of each question. We use disambiguated word senses for this purpose.
Syntactic information has previously been used successfully in question answering (Zhang and Lee 2003; Moschitti and Basili 2006; Moschitti et al. 2007; Chali, Hasan, and Joty 2009, 2011). Pasca and Harabagiu (2001) argued that with the syntactic form of a sentence one can see which words depend on other words. We also feel that there should be a similarity between the words that are dependent in the sentences present in the associated body of texts and the dependency between words of the generated question. This motivates us to propose the use of syntactic kernels in judging the syntactic correctness of the generated questions automatically.
The main goal of our work is to generate as many questions as possible related to the topic. We use the named entity information and the predicate argument structures of the sentences to accomplish this goal. Our approach is different from the set-up in shared tasks (Rus and Graesser 2009; Boyer and Piwek 2010), as we generate a set of basic questions that are useful to add variety in the question space. A paragraph associated with each topic is used as the source of relevant information about the topic. We evaluate our systems in terms of topic relevance, which is different from prior research (Heilman and Smith 2010a; Mannem, Prasad, and Joshi 2010). Syntactic correctness is also an important property of a good question. For this reason, we evaluate our system in terms of syntactic correctness as well. The proposed system will be useful for generating topic-related questions from the associated content information, which can be used to incorporate a “question suggestions for a certain topic” facility in search systems (Kotov and Zhai 2010). For example, if a user searches for some information related to a certain topic, the search system could generate all possible topic-relevant questions from a preexistent related body of texts to provide suggestions. Kotov and Zhai (2010) approached a similar task by proposing a technique to augment the standard ranked list presentation of search results with a question-based interface to refine user-given queries.
The major contributions of our work can be summarized as follows:
We perform the task of topic-to-question generation, which can help users in expressing their information needs. Questions are generated using a set of general-purpose rules based on named entity information and the predicate argument structures of the sentences (along with semantic roles) present in the associated body of texts.
We identify the subtopics (which are closely related to the original topic) in the given body of texts by using LDA and calculate their similarity with the questions by applying ESSK (with disambiguated word senses). This helps us to measure the importance of each question.
We compute the syntactic similarity of each question with its associated content information by applying the tree kernel functions with the re-implementation of the syntactic tree kernel model. In this way, we judge the syntactic correctness of each generated question automatically.
We evaluate the ESSK similarity scores and the syntactic similarity scores in a ranking framework and show that the use of ESSK and syntactic kernels improve the relevance and the syntactic correctness of the top-ranked questions, respectively.
We identify circumstances in which our approach performs well and show that, using additional experiments by narrowing down the topic focus. Experiments with the topics about people (biographical focus) reveal improvements in the overall results.","3 topic-to-question generation :Our QG approach mainly builds on four steps. In the first step, complex sentences (from the given body of texts) related to a topic are simplified, as it is easier to generate questions from simple sentences. In the next step, named entity information and predicate argument structures of the sentences are extracted and are then used to generate questions. In the third step, LDA is used to identify important subtopics from the given body of texts, and then ESSK is applied to find their similarity with the generated questions. In the final step, a syntactic tree kernel is used and syntactic similarity between the generated questions and the sentences present in the body of texts determines the syntactic correctness of the questions. Questions are then ranked by considering the ESSK similarity scores and the syntactic similarity scores. We present an architectural diagram (Figure 1) to show the different components of our system and describe the overall procedure in the following subsections. Sentences may have complex grammatical structure with multiple embedded clauses. Therefore, the first step of our proposed system is to simplify the complex sentences with the intention of generating more accurate questions. We use the simplified factual statement extractor model6 of Heilman and Smith (2010a). Their model extracts the simpler forms of the complex source sentence by altering lexical items, syntactic structure, and semantics, as well as by removing phrase types such as leading conjunctions, sentence-level modifying phrases, and appositives. For example, given a complex sentence s, we get the corresponding simple sentences as follows:
Complex Sentence (s): Apple’s first logo, designed by Jobs and Wayne, depicts Sir Isaac Newton sitting under an apple tree. Simple Sentence (1): Apple’s first logo is designed by Jobs and Wayne.
6 Available at http://www.ark.cs.cmu.edu/mheilman/questions/.
Simple Sentence (2): Apple’s first logo depicts Sir Isaac Newton sitting under an apple tree. In the second step of our system, we at first process the simple sentences in order to generate all possible questions from them. We use the Illinois Named Entity Tagger,7 a state-of-the-art named entity (NE) tagger that tags plain text with named entities (people, organizations, locations, miscellaneous) (Ratinov and Roth 2009). Once we tag the topic under consideration and its associated body of texts, we use some general purpose rules to create some basic questions even though the answer is not present in the body of texts. For example, Apple Inc. is tagged as an organization, so we generate a question: Where is Apple Inc. located?. The main motivation behind generating such questions is to add variety to the generated question space. The basic questions are useful when there is very little or no knowledge available for a certain topic in consideration. This assumption is inherited from the scenario in the real world where a human can ask questions having very limited background knowledge about the topic. For example, if a person knows nothing about a university, they can ask What is it? or, if they at least know that a university is an institution, then they can ask the question, Where is it located?. In Table 1, we show some example rules for the basic questions generated in this work.
Our next task is to generate specific questions from the sentences present in the given body of texts. For this purpose, we parse the sentences semantically using a
7 Available at http://cogcomp.cs.illinois.edu/.
Semantic Role Labeling (SRL) system (Kingsbury and Palmer 2002; Hacioglu et al. 2003), ASSERT.8 ASSERT is an automatic statistical semantic role tagger that can annotate naturally occuring text with semantic arguments. When presented with a sentence, it performs a full syntactic analysis of the sentence, automatically identifies all the verb predicates in that sentence, extracts features for all constituents in the parse tree relative to the predicate, and identifies and tags the constituents with the appropriate semantic arguments. For example, the output of the SRL system for the sentence Apple’s first logo is designed by Jobs and Wayne is: [ARG1 Apple ’s first logo] is [TARGET designed ] [ARG0 by Jobs and Wayne]. The output contains one verb (predicate) with its arguments (i.e., semantic roles). These arguments are used to generate specific questions from the sentences. For example, we can replace [ARG1 ..] with What and generate a question as: What is designed by Jobs and Wayne?. Similarly, [ARG0 ..] can be replaced and the question: Who designed Apple’s first logo? can be generated. The semantic roles ARG0...ARG5 are called mandatory arguments. There are some additional arguments or semantic roles that can be tagged by ASSERT. They are called optional arguments and they start with the prefix ARGM. These are defined by the annotation guidelines set in Palmer, Gildea, and Kingsbury (2005). A set of about 350 general-purpose rules are used to transform the semantic-role labeled sentences into the questions. The rules were set up in a way that we could use the semantic role information to find the potential answer words in a sentence that would be replaced by suitable question words. In the case of a mandatory argument, the choice of question word depends on the argument’s named entity tag (who for a person, where for a location, etc.). Table 2 shows how different semantic roles can be replaced by possible question words in order to generate a question. In the third step of our proposed system, we pass the generated questions to the importance judgment module that uses LDA and ESSK to assign a topic relevance score to each question. The detailed procedure is discussed in the following subsections.
3.3.1 Latent Dirichlet Allocation (LDA). To measure the importance of the generated questions, we use LDA (Blei, Ng, and Jordan 2003) to identify the important subtopics9 from the given body of texts. LDA is a probabilistic topic modeling technique where the main principle is to view each document as a mixture of various topics. Here each topic is a probability distribution over words. LDA assumes that documents are made
8 Available at http://cemantix.org/assert.html. 9 The term sub-topic is used in the LDA topic modeling sense, which represents a probability distribution
over words.
up of words and word ordering is not important (“bag-of-words” assumption) (Misra, Cappé, and Yvon 2008). The main idea is to choose a distribution over topics while generating a new document. For each word in the new document, a topic is randomly chosen according to this distribution and a word is drawn from that topic. LDA uses a generative topic modeling approach to specify the following distribution over words within a document:
P(wi) = K∑
j=1
P(wi|zi = j)P(zi = j) (1)
where K is the number of topics, P(wi|zi = j) is the probability of word wi under topic j, and P(zi = j) is the sampling probability of topic j for the ith word. The multinomial distributions φ(j) = P(w|zi = j) and θ(d) = P(z) are termed as topic-word distribution and document-topic distribution, respectively (Blei, Ng, and Jordan 2003). A Dirichlet (α) prior is placed on θ and a Dirichlet (β) prior is set on φ to refine this basic model (Griffiths and Steyvers 2002; Blei, Ng, and Jordan 2003). Now the main goal is to estimate the two parameters: θ and φ. We apply this framework directly to solve our problem by considering each topic-related body of texts as a document. We use a GUIbased toolkit for topic modeling10 that uses the popular MALLET (McCallum 2002) toolkit for the back-end. The LDA model is built on the development set11 (Section 4.2). The process starts by removing a list of “stop words” from the document and runs 200 iterations of Gibbs sampling (Geman and Geman 1984) to estimate the parameters θ and φ. From each body of texts, we discover K topics and choose the most frequent words from the most likely unigrams as the desired subtopics. For example, from the associated body of texts of the topic Apple Inc. Logos, we get these subtopics: janoff, themes, logo, color, apple.
3.3.2 Extended String Subsequence Kernel (ESSK). Once we identify the subtopics, we apply ESSK (Hirao et al. 2003) to measure their similarity with the generated questions. In the general ESSK, each word in a sentence is considered an “alphabet,” and the alternative is all its possible senses. However, our ESSK implementation considers the
10 Available at http://code.google.com/p/topic-modeling-tool/. 11 The model was built and tested according to the guidelines of the topic modeling toolkit we used.
alternative of each word as its disambiguated sense. We use a dictionary-based Word Sense Disambiguation (WSD) system (Chali and Joty 2007) assuming one sense per discourse. We use WordNet (Fellbaum 1998) to find the semantic relations (such as repetition, synonym, hypernym and hyponym, holonym and meronym, and gloss) for all the words in a text. We assign a weight to each semantic relation based on heuristics and use all of them. Our WSD technique is decomposed into two steps: (1) building a representation of all possible senses of the words and (2) disambiguating the words based on the highest score. To be specific, each candidate word from the context is expanded to all of its senses. A disambiguation graph is constructed as the intermediate representation where the nodes denote word instances with their WordNet senses, and the weighted edges (connecting the senses of two different words) represent semantic relations. This graph is exploited to perform the WSD. We sum the weights of all edges, leaving the nodes under their different senses. The sense with the highest score is considered to be the most probable sense. In case of a tie between two or more senses, we select the sense that comes first in WordNet, because WordNet orders the senses of a word by decreasing order of their frequency. Our preliminary experiments suggested that WSD has a positive impact on the performance of our proposed system.
ESSK is used to measure the similarity between all possible subsequences of the question words/senses and topic words/senses. We calculate the similarity score Sim(Ti, Qj) using ESSK, where Ti denotes a topic/sub-topic word sequence and Qj stands for a generated question. Formally, ESSK is defined as follows:12
Kessk(T, Q) = d∑
m=1 ∑ ti∈T ∑ qj∈Q Km(ti, qj)
Km(ti, qj) = {
val(ti, qj) if m = 1 K
′ m−1(ti, qj) · val(ti, qj)
Here, K ′ m(ti, qj) is defined in the following. ti and qj are nodes of T and Q, respectively. The function val(t, q) returns the number of common attributes (i.e., the number of common words/senses) to the given nodes t and q.
K ′ m(ti, qj) = { 0 if j = 1 λK ′ m(ti, qj−1) + K ′′ m(ti, qj−1)
Here, λ is the decay parameter for the number of skipped words. K ′′ m(ti, qj) is defined as:
K ′′ m(ti, qj) = { 0 if i = 1 λK ′′ m(ti−1, qj) + Km(ti−1, qj)
12 The formulae denote a dynamic programming technique to compute the ESSK similarity score where d is the vector space dimension (i.e., the number of all possible subsequences of up to length d). More information about these formulae can be obtained from Hirao et al. (2003, 2004).
Finally, the similarity measure is defined after normalization:
simessk(T, Q) = Kessk(T, Q)√
Kessk(T, T)Kessk(Q, Q) The next step of our system is to judge the syntactic correctness of the generated questions. The generated questions might be syntactically incorrect due to the process of automatic question generation. It is time-consuming and considerable human intervention is necessary to check for the syntactically incorrect questions manually. We strongly believe that a question should have a similar syntactic structure to a sentence from which it is generated. For example, the sentence Apple’s first logo is designed by Jobs and Wayne., and the generated question What is designed by Jobs and Wayne? are syntactically similar. An example of an ungrammatical generated question that is not very similar to its source is: Janoff presented Jobs What?. To judge the syntactic correctness of each generated question automatically, we apply the tree kernel functions and re-implement the syntactic tree kernel model according to Moschitti et al. (2007) for computing the syntactic similarity of each question with the associated content information. We first parse the sentences and the questions into syntactic trees using the Charniak parser13 (Charniak 1999). Then, we calculate the similarity between the two corresponding trees using the tree kernel method (Collins and Duffy 2001). We convert each parenthetic representation generated by the Charniak parser into its corresponding tree and give the trees as input to the tree kernel functions for measuring the syntactic similarity.
The tree kernel function computes the number of common subtrees between two trees and gives the similarity score between each sentence in the given body of texts and the generated question based on the syntactic structure. Each sentence14 contributes a score to the questions and then the questions are ranked by considering the average of similarity scores.","4 experiments : We consider the task of automatically generating questions from topics where each topic is associated with a body of texts having a useful description about the topic. The question-ranking module of the proposed QG system ranks the questions by combining the topic relevance scores and the syntactic similarity scores of Section 3.3 and Section 3.4 using the following formula:
w ∗ ESSKscore + (1 − w) ∗ SYNscore (2)
13 Available at https://github.com/BLLIP/bllip-parser. 14 We consider that a question is syntactically fluent as well as relevant to the topic if it has similar syntactic
subtrees to those of the most sentences in the body of texts.
Here, w is the importance parameter, which holds a value in [0, 1]. We kept w = 0.5 to give equal importance15 to topic relevance and syntactic correctness. To run our experiments, we use the data set provided in the Question Generation Shared Task and Evaluation Challenge16 (2010) for the task of question generation from paragraphs. This data set consists of 60 paragraphs about 60 topics that were originally collected from several Wikipedia, OpenLearn, and Yahoo!Answers articles. The paragraphs contain around 5–7 sentences for a total of 100–200 tokens (including punctuation). This data set includes a diversity of topics of general interest. We consider these topics and treat the paragraphs as their associated useful content information in order to generate a set of questions using our proposed QG approach. We randomly select 10 topics and their associated paragraphs as the development data.17 A total of 2,186 questions are generated from the remaining 50 topics (test data) to be ranked. 4.3.1 Methodology. We use a methodology derived from Boyer and Piwek (2010) and Heilman and Smith (2010b) to evaluate the performance of our QG systems. Three native English-speaking university graduate students judge the quality of the top-ranked 20% questions using two criteria: topic relevance and syntactic correctness. For topic relevance, the given score is an integer between 1 (very poor) and 5 (very good) and is guided by the consideration of the following aspects: 1. Semantic correctness (i.e., the question is meaningful and related to the topic), 2. Correctness of question type (i.e., a correct question word is used), and 3. Referential clarity (i.e., it is clearly possible to understand what the question refers to). For syntactic correctness, the assigned score is also an integer between 1 (very poor) and 5 (very good). Whether a question is grammatically correct or not is checked here. The judges were asked to read the topics with their associated body of texts and then rate the top-ranked questions generated by different systems. For each question, we calculate the average of the judges’ scores. The judges were provided with an annotation guideline and sample judgments, according to the methodology derived from Boyer and Piwek (2010) and Heilman and Smith (2010b). The same judges evaluated all the system outputs and they were blind to the system identity when judging. No guidelines were provided on the relative importance of the various aspects that made the judgment task subjective. The inter-annotator agreement of Fleiss’s κ = 0.41, 0.45, 0.62, and 0.33 are computed for the three judges for the results in Tables 3–6, indicating moderate (for the first two tables), and substantial and fair agreement (Landis and Koch 1977) between the raters, respectively. These κ values were shown to be acceptable in the literature for the relevant NLP tasks (Dolan and Brockett 2005; Glickman, Dagan, and Koppel 2005; Heilman and Smith 2010b).
15 A syntactically incorrect question is not useful even if it is relevant to the topic. This motivated us to give equal importance to topic relevance and syntactic correctness. The parameter w can be tuned to investigate its impact on the system performance. 16 http://www.questiongeneration.org/mediawiki. 17 We use these data to build necessary general purpose rules for our QG model.
4.3.2 Systems for Comparison. We report the performance of the following systems in order to do a meaningful comparison with our proposed QG system:
(1) Baseline1: This is our QG system without any question-ranking method applied to it. Here, we randomly select top 20% questions and rate them. (2) Baseline2: For our second baseline, we build a QG system using an alternative topic modeling approach. Here, we use a topic signature model (instead of using LDA as discussed in Section 3.3.1) (Lin and Hovy 2000) to identify the important subtopics from the sentences present in the body of texts. The subtopics are the important words in the context that are closely related to the topic and have significantly greater probability of occurring in the given text compared with a large background corpus. We use a topic signature computation tool18 for this purpose. The background corpus that is used in this tool contains 5,000 documents from the English GigaWord Corpus. For example, from the given body of texts of the topic Apple Inc. Logos, we get these subtopics: jobs, logo, themes, rainbow, monochromatic. Then we use the same steps of Sections 3.3.2 and 3.4, and use Equation (2) to combine the scores. We evaluate the top-ranked 20% questions and show the results. (3) State-of-the-art: We choose a publicly available state-of-the-art QG system19 to generate questions from the sentences in the body of texts. This system was shown to achieve good performance in generating fact-based questions about the content of a given article (Heilman and Smith 2010b). Their method ranks the questions automatically using a logistic regression model. Given a paragraph as input, this system processes each sentence and generates a set of ranked questions for the entire paragraph. We evaluate the top-ranked 20% questions20 and report the results.
4.3.3 Results and Discussion. Table 3 shows the average topic relevance and syntactic correctness scores for all the systems. From these results, we can see that the proposed QG system improves the topic relevance and syntactic correctness scores over the Baseline1 system by 62% and 35%, respectively, and improves the topic relevance and syntactic correctness scores over the Baseline2 system by 7%, and 8%, respectively. On the other hand, the proposed QG system improves the topic relevance and syntactic correctness scores over the state-of-the-art system by 4% and 3%, respectively. From these results, we can clearly observe the effectiveness of our proposed QG system. The improvements in the results are statistically significant21 (p < 0.05).
The main goal of this work was to generate as many questions as possible related to the topic. For this reason, we considered generating the basic questions. These questions
18 Available at http://www.cis.upenn.edu/∼lannie/topicS.html. 19 Available at http://www.ark.cs.cmu.edu/mheilman/questions/. 20 We ignore the yes-no questions for our task. 21 We tested statistical significance using Student’s t test.
were also useful to provide variety in the question space. We generated these questions using the named entity information. As the performance of the NE taggers were unsatisfactory, we had a few of these questions generated. In most cases, these questions were outranked by other important questions, which included a combination of topics and subtopics to show higher topic relevance score measured by ESSK. Therefore, they do not have a considerable impact on the evaluation statistics. We claim that the overall performance of our systems could be further improved if the accuracy of the NE tagger and the semantic role labeler could be increased.
Acceptability Test. In another evaluation setting, the three annotators judge the questions for their overall acceptability as a good question. If a question shows no deficiency in terms of the criteria considered for topic relevance and syntactic correctness, it is termed as acceptable. We evaluate the top 15% and top 30% questions separately for each QG system and report the results indicating the percentage of questions rated as acceptable in Table 4. The results indicate that the percentage of the questions rated acceptable is reduced when we evaluate a greater number of questions—which proves the effectiveness of our QG system.
Narrowing Down the Focus. We run further experiments by narrowing down the topic focus. We consider only the topics about people (biographical focus). We choose 50 people as our topics from the list of the 20th century’s 100 most influential people, published in Time magazine in 1999 and obtained the paragraphs containing their biographical information from Wikipedia articles.22 We generate a total of 1, 845 questions from the 50 topics considered and rank them using different ranking schemes as discussed before. We evaluate the top 20% questions using the similar evaluation methodologies and report the results in Table 5. From these results, we can see that the proposed QG
22 http://en.wikipedia.org/wiki/Time 100.
system improves the topic relevance and syntactic correctness scores over the Baseline1 system by 33% and 35%, respectively, and improves the topic relevance and syntactic correctness scores over the Baseline2 system by 8% and 9%, respectively. Moreover, the proposed QG system improves both the topic relevance and syntactic correctness scores over the state-of-the-art system by 4%. From these results, we can clearly observe the effectiveness of our proposed QG system when we narrow down the topic focus. We also evaluate the top 15% and top 30% questions separately for each QG system and report the results, indicating the percentage of questions rated as acceptable in Table 6. From these tables, we can clearly see the improvements in all the scores for all the QG approaches. This is reasonable because the accuracy of the NE tagger and the semantic role labeler is increased for the biographical data.23 These results further demonstrate that the proposed system is significantly better (at p < 0.05) than the other considered systems. We plan to make our created resources available to other researchers.
4.3.4 An Input-Output Example. An input to our systems is, for instance,24 the topic Apple Inc. Logos with the associated content information (body of texts):
Apple’s first logo, designed by Jobs and Wayne, depicts Sir Isaac Newton sitting under an apple tree. Almost immediately, though, this was replaced by Rob Janoff’s “rainbow Apple”, the now-familiar rainbow-colored silhouette of an apple with a bite taken out of it. Janoff presented Jobs with several different monochromatic themes for the “bitten” logo, and Jobs immediately took a liking to it. While Jobs liked the logo, he insisted it be in color to humanize the company. The Apple logo was designed with a bite so that it would be recognized as an apple rather than a cherry. The colored stripes were conceived to make the logo more accessible, and to represent the fact the monitor could reproduce images in color. In 1998, with the roll-out of the new iMac, Apple discontinued the rainbow theme and began to use monochromatic themes, nearly identical in shape to its previous rainbow incarnation.
The output of our systems is the ranked lists of questions. We show an example output in Table 7. To provide a more detailed analysis of our results, the average output scores of the example questions are presented in Table 8. From this table, we can understand how different aspects of the evaluation criteria affected the performance of the different systems. For example, Q1 of the proposed system was given a very good score due to
23 Although a few basic questions were generated compared with other important questions containing topical words, we believe they did not have a considerable impact on the overall performance of our system. 24 The example input text is provided from the Question Generation Shared Task and Evaluation Challenge (QGSTEC 2010) data set that we used for our experiments (Section 4.2).
its relevance to the topic in consideration. On the other hand, Q3 of the state-of-the-art was assigned a lower score due to its lack of clarity with respect to the topic.","5 conclusion :In this article, we have considered the task of automatically generating questions from topics where each topic is associated with a body of texts containing useful information. The proposed method exploits named entity and semantic role labeling information to accomplish the task. A key aspect of our approach was the use of latent Dirichlet allocation (LDA) to automatically discover the hidden subtopics from the sentences. We have proposed a novel method to rank the generated questions by considering: (1) subtopical similarity determined using ESSK algorithm in combination with word sense disambiguation, and (2) syntactic similarity determined using the syntactic tree kernel based method. We have compared the proposed question generation (QG) system with two baseline systems and one state-of-the-art system. The evaluation results show that the proposed QG system significantly outperforms all other considered systems, as our top-ranked system generated questions were found to be better in topic-relevance and syntactic correctness than those of the other systems. Our results demonstrated that judging syntactic correctness of the generated questions using the syntactic tree kernel based model was suitable in our question generation setting. We would like to further
our research by using other available measures such as an n-gram language model or a parser confidence score (Wagner, Foster, and van Genabith 2009) in order to see how they would perform on the same task. In this article, we have also extended our experiments by narrowing down the topic focus. In this experiment, we have considered people as topics. A rigorous analysis of the evaluation results has revealed that the performance of our proposed QG system can be enhanced if we narrow down the topic focus. We hope to carry on these ideas and develop further mechanisms for question generation based on the dependency features of the answers and answer finding (Li and Roth 2006; Pinchak and Lin 2006).",,,,"This paper is concerned with automatic generation of all possible questions from a topic of interest. Specifically, we consider that each topic is associated with a body of texts containing useful information about the topic. Then, questions are generated by exploiting the named entity information and the predicate argument structures of the sentences present in the body of texts. The importance of the generated questions is measured using Latent Dirichlet Allocation by identifying the subtopics (which are closely related to the original topic) in the given body of texts and applying the Extended String Subsequence Kernel to calculate their similarity with the questions. We also propose the use of syntactic tree kernels for the automatic judgment of the syntactic correctness of the questions. The questions are ranked by considering both their importance (in the context of the given body of texts) and syntactic correctness. To the best of our knowledge, no previous study has accomplished this task in our setting. A series of experiments demonstrate that the proposed topic-to-question generation approach can significantly outperform the state-of-the-art results.","[{""affiliations"": [], ""name"": ""Yllias Chali""}, {""affiliations"": [], ""name"": ""Sadid A. Hasan""}]",SP:2ce1f080cb051186b15c05b770543e8bca73d2c6,"[{""authors"": [""M. Agarwal"", ""R. Shah"", ""P. Mannem.""], ""title"": ""Automatic question generation using discourse cues"", ""venue"": ""Proceedings of the 6th Workshop on Innovative Use of NLP for Building Educational Applications,"", ""year"": 2011}, {""authors"": [""H. Ali"", ""Y. Chali"", ""S.A. Hasan.""], ""title"": ""Automation of question generation from sentences"", ""venue"": ""Proceedings of QG2010: The Third Workshop on Question Generation, pages 58\u201367, Pittsburgh, PA."", ""year"": 2010}, {""authors"": [""A. Andrenucci"", ""E. Sneiders.""], ""title"": ""Automated question answering: Review of the main approaches"", ""venue"": ""Proceedings of the 3rd International Conference on Information Technology and Applications (ICITA\u201905),"", ""year"": 2005}, {""authors"": [""D.M. Blei"", ""A.Y. Ng"", ""M.I. Jordan.""], ""title"": ""Latent Dirichlet allocation"", ""venue"": ""Journal of Machine Learning Research, 3:993\u20131022."", ""year"": 2003}, {""authors"": [""J.C. Brown"", ""G.A. Frishkoff"", ""M. Eskenazi.""], ""title"": ""Automatic question generation for vocabulary assessment"", ""venue"": ""Proceedings of the Conference on Human Language Technology and Empirical"", ""year"": 2005}, {""authors"": [""A. Celikyilmaz"", ""D. Hakkani-Tur"", ""G. Tur""], ""title"": ""LDA based similarity modeling"", ""year"": 2010}, {""authors"": [""Y. Chali"", ""S.A. Hasan.""], ""title"": ""On the effectiveness of using sentence compression models for query-focused multi-document summarization"", ""venue"": ""Proceedings of the 24th International"", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan.""], ""title"": ""Query-focused multi-document summarization: Automatic data annotations and supervised learning approaches"", ""venue"": ""Journal of Natural Language"", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan.""], ""title"": ""Towards automatic topical question generation"", ""venue"": ""Proceedings of the 24th International Conference on Computational Linguistics (COLING 2012), pages 475\u2013492, Mumbai."", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""K. Imam.""], ""title"": ""An aspect-driven random walk model for topic-focused multi-document summarization"", ""venue"": ""Proceedings of the 7th Asian Information Retrieval Societies"", ""year"": 2011}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""K. Imam.""], ""title"": ""Learning good decompositions of complex questions"", ""venue"": ""Proceedings of the 17th International Conference on Applications of Natural Language Processing"", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""S.R. Joty""], ""title"": ""Do automatic annotation techniques have any impact on supervised complex question answering"", ""venue"": ""In Proceedings of the Joint Conference of the 47th Annual Meeting"", ""year"": 2009}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""S.R. Joty""], ""title"": ""Improving graph-based random walks for complex question answering"", ""year"": 2011}, {""authors"": [""Y. Chali"", ""S.R. Joty.""], ""title"": ""Word sense disambiguation using lexical cohesion"", ""venue"": ""Proceedings of the 4th International Conference on Semantic Evaluations, pages 476\u2013479, Prague."", ""year"": 2007}, {""authors"": [""Y. Chali"", ""S.R. Joty"", ""S.A. Hasan.""], ""title"": ""Complex question answering: Unsupervised learning approaches and experiments"", ""venue"": ""Journal of Artificial Intelligence Research, 35:1\u201347."", ""year"": 2009}, {""authors"": [""E. Charniak""], ""title"": ""A maximum-entropyinspired parser"", ""venue"": ""Technical Report CS-99-12, Brown University, Computer Science Department, Providence, RI."", ""year"": 1999}, {""authors"": [""W. Chen"", ""G. Aist"", ""J. Mostow.""], ""title"": ""Generating questions automatically from informational text"", ""venue"": ""Proceedings of the 2nd Workshop on Question Generation (AIED 2009), pages 17\u201324, Arlington, VA."", ""year"": 2009}, {""authors"": [""M. Collins"", ""N. Duffy.""], ""title"": ""Convolution Kernels for natural language"", ""venue"": ""Proceedings of Neural Information Processing Systems, pages 625\u2013632, Vancouver."", ""year"": 2001}, {""authors"": [""W.B. Dolan"", ""C. Brockett.""], ""title"": ""Automatically constructing a corpus of sentential paraphrases"", ""venue"": ""Proceedings of the 3rd International Workshop on Paraphrasing (IWP2005), pages 9\u201316, Jeju Island."", ""year"": 2005}, {""authors"": [""A. Echihabi"", ""D. Marcu.""], ""title"": ""A noisy-channel approach to question answering"", ""venue"": ""Proceedings of the 41st Annual Meeting on Association for Computational Linguistics - Volume 1, pages 16\u201323,"", ""year"": 2003}, {""authors"": [""C. Fellbaum""], ""title"": ""WordNet - An Electronic Lexical Database"", ""venue"": ""MIT Press."", ""year"": 1998}, {""authors"": [""S. Geman"", ""D. Geman.""], ""title"": ""Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images"", ""venue"": ""IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:721\u2013741."", ""year"": 1984}, {""authors"": [""O. Glickman"", ""I. Dagan"", ""M. Koppel.""], ""title"": ""A probabilistic classification approach for lexical textual entailment"", ""venue"": ""AAAI, pages 1,050\u20131,055, Pittsburgh, PA."", ""year"": 2005}, {""authors"": [""A.C. Graesser"", ""K. VanLehn"", ""C.P. Rose"", ""P.W. Jordan"", ""D. Harter.""], ""title"": ""Intelligent tutoring systems with conversational dialogue"", ""venue"": ""AI Magazine, 22(4):39\u201352."", ""year"": 2001}, {""authors"": [""M.A. Greenwood""], ""title"": ""Open-Domain Question Answering"", ""venue"": ""Ph.D. thesis, Department of Computer Science, University of Sheffield."", ""year"": 2005}, {""authors"": [""T.L. Griffiths"", ""M. Steyvers.""], ""title"": ""Prediction and semantic association"", ""venue"": ""NIPS\u201902, pages 11\u201318, Cambridge, MA."", ""year"": 2002}, {""authors"": [""K. Hacioglu"", ""S. Pradhan"", ""W. Ward"", ""J.H. Martin"", ""D. Jurafsky.""], ""title"": ""Shallow semantic parsing using support vector machines"", ""venue"": ""Technical Report TR-CSLR-2003-03, University of Colorado,"", ""year"": 2003}, {""authors"": [""S. Harabagiu"", ""F. Lacatusu"", ""A. Hickl.""], ""title"": ""Answering complex questions with random walk models"", ""venue"": ""Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development"", ""year"": 2006}, {""authors"": [""M. Heilman"", ""N.A. Smith.""], ""title"": ""Extracting simplified statements for factual question generation"", ""venue"": ""Proceedings of the Third Workshop on Question Generation, pages 11\u201320,"", ""year"": 2010}, {""authors"": [""M. Heilman"", ""N.A. Smith.""], ""title"": ""Good question! Statistical ranking for question generation"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the"", ""year"": 2010}, {""authors"": [""A. Hickl"", ""J. Lehmann"", ""D. Moldovan"", ""S. Harabagiu.""], ""title"": ""Experiments with interactive question-answering"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""A. Hickl"", ""P. Wang"", ""J. Lehmann"", ""Sanda Harabagiu.""], ""title"": ""Ferret: Interactive question-answering for real-world environments"", ""venue"": ""Proceedings of the COLING/ACL on Interactive Presentation"", ""year"": 2006}, {""authors"": [""T. Hirao"", ""J. Suzuki"", ""H. Isozaki"", ""E. Maeda.""], ""title"": ""NTT\u2019s multiple document summarization system for DUC2003"", ""venue"": ""Proceedings of the Document Understanding Conference, Edmonton."", ""year"": 2003}, {""authors"": [""T. Hirao"", ""J. Suzuki"", ""H. Isozaki"", ""E. Maeda.""], ""title"": ""Dependency-based sentence alignment for multiple document summarization"", ""venue"": ""Proceedings of COLING 2004, pages 446\u2013452, Geneva."", ""year"": 2004}, {""authors"": [""L. Hirschman"", ""R. Gaizauskas.""], ""title"": ""Natural language question answering: The view from here"", ""venue"": ""Natural Language Engineering, 7(4):275\u2013300."", ""year"": 2001}, {""authors"": [""K. Ignatova"", ""D. Bernhard"", ""I. Gurevych.""], ""title"": ""Generating high quality questions from low quality questions"", ""venue"": ""Proceedings of the Workshop on the Question Generation Shared Task and Evaluation Challenge,"", ""year"": 2008}, {""authors"": [""P. Kingsbury"", ""M. Palmer.""], ""title"": ""From Treebank to PropBank"", ""venue"": ""Proceedings of the International Conference on Language Resources and Evaluation, pages 1,989\u20131,993, Las Palmas."", ""year"": 2002}, {""authors"": [""A. Kotov"", ""C. Zhai.""], ""title"": ""Towards natural question guided search"", ""venue"": ""Proceedings of the 19th International Conference on the World Wide Web, WWW \u201910, pages 541\u2013550, Raleigh, NC."", ""year"": 2010}, {""authors"": [""J. Kupiec""], ""title"": ""MURAX: A robust linguistic approach for question answering using an on-line encyclopedia"", ""venue"": ""SIGIR, pages 181\u2013190, Pittsburgh, PA."", ""year"": 1993}, {""authors"": [""J.R. Landis"", ""G.G. Koch.""], ""title"": ""The measurement of observer agreement for categorical data"", ""venue"": ""Biometrics, 33(1):159\u2013174."", ""year"": 1977}, {""authors"": [""X. Li"", ""D. Roth.""], ""title"": ""Learning question classifiers: The role of semantic information"", ""venue"": ""Journal of Natural Language Engineering, 12(3):229\u2013249."", ""year"": 2006}, {""authors"": [""C.Y. Lin""], ""title"": ""Automatic question generation from queries"", ""venue"": ""Proceedings of the Workshop on the Question Generation Shared Task and Evaluation Challenge, Arlington, VA."", ""year"": 2008}, {""authors"": [""C.Y. Lin"", ""E.H. Hovy.""], ""title"": ""The automated acquisition of topic signatures for text summarization"", ""venue"": ""Proceedings of the 18th Conference on Computational Linguistics, pages 495\u2013501, Saarbr\u00fcken."", ""year"": 2000}, {""authors"": [""P. Mannem"", ""R. Prasad"", ""A. Joshi.""], ""title"": ""Question generation from paragraphs at UPenn: QGSTEC system description"", ""venue"": ""Proceedings of the Third Workshop on Question Generation, pages 84\u201391,"", ""year"": 2010}, {""authors"": [""A.K. McCallum""], ""title"": ""MALLET: A machine learning for language toolkit"", ""venue"": ""http://mallet.cs.umass.edu."", ""year"": 2002}, {""authors"": [""C.C. McConnell"", ""P. Mannem"", ""R. Prasad"", ""A. Joshi.""], ""title"": ""A new approach to ranking over-generated questions"", ""venue"": ""Proceedings of the AAAI Fall Symposium on Question Generation, pages 45\u201348,"", ""year"": 2011}, {""authors"": [""J. McGough"", ""J. Mortensen"", ""J. Johnson"", ""S. Fadali.""], ""title"": ""A Web-based testing system with dynamic question generation"", ""venue"": ""ASEE/IEEE Frontiers in Education Conference, pages S3C-23\u201328"", ""year"": 2001}, {""authors"": [""H. Misra"", ""O. Capp\u00e9"", ""F. Yvon.""], ""title"": ""Using LDA to detect semantically incoherent documents"", ""venue"": ""Proceedings of the Twelfth Conference on Computational"", ""year"": 2008}, {""authors"": [""R. Mitkov"", ""L.A. Ha.""], ""title"": ""Computer-aided generation of multiple-choice tests"", ""venue"": ""Proceedings of the HLT-NAACL 03 Workshop on Building Educational Applications Using Natural"", ""year"": 2003}, {""authors"": [""D. Moldovan"", ""C. Clark"", ""M. Bowden.""], ""title"": ""Lymba\u2019s PowerAnswer 4 in TREC 2007"", ""venue"": ""Proceedings of the 16th Text REtreival Conference, Gaithersburg, MD."", ""year"": 2007}, {""authors"": [""A. Moschitti"", ""R. Basili.""], ""title"": ""A tree kernel approach to question and answer classification in question answering systems"", ""venue"": ""Proceedings of the 5th International Conference on"", ""year"": 2006}, {""authors"": [""A. Moschitti"", ""S. Quarteroni"", ""R. Basili"", ""S. Manandhar.""], ""title"": ""Exploiting syntactic and shallow semantic kernels for question/answer classification"", ""venue"": ""Proceedings of the 45th Annual Meeting"", ""year"": 2007}, {""authors"": [""A.M. Olney"", ""A.C. Graesser"", ""N.K. Person.""], ""title"": ""Question generation from concept maps"", ""venue"": ""Dialogue and Discourse, 3(2):75\u201399."", ""year"": 2012}, {""authors"": [""M. Palmer"", ""D. Gildea"", ""P. Kingsbury.""], ""title"": ""The proposition bank: An annotated corpus of semantic roles"", ""venue"": ""Computational Linguistics, 31(1):71\u2013106."", ""year"": 2005}, {""authors"": [""M. Pasca"", ""S.M. Harabagiu.""], ""title"": ""Answer mining from on-line documents"", ""venue"": ""Proceedings of the Association for Computational Linguistics 39th Annual Meeting and 10th Conference of the European"", ""year"": 2001}, {""authors"": [""C. Pinchak"", ""D. Lin.""], ""title"": ""A probabilistic answer type model"", ""venue"": ""Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 393\u2013400, Trento."", ""year"": 2006}, {""authors"": [""L. Ratinov"", ""D. Roth.""], ""title"": ""Design challenges and misconceptions in named entity recognition"", ""venue"": ""Proceedings of the Thirteenth Conference on Computational Natural Language Learning, pages 147\u2013155,"", ""year"": 2009}, {""authors"": [""V. Rus"", ""Z. Cai"", ""A.C. Graesser.""], ""title"": ""Experiments on generating questions about facts"", ""venue"": ""Proceedings of the 8th International Conference on Computational Linguistics and Intelligent Text Processing,"", ""year"": 2007}, {""authors"": [""V. Rus"", ""A.C. Graesser.""], ""title"": ""The question generation shared task and evaluation challenge"", ""venue"": ""Workshop on the Question Generation Shared Task and Evaluation Challenge, Final Report,"", ""year"": 2009}, {""authors"": [""R.F. Simmons""], ""title"": ""Answering English questions by computer: A survey"", ""venue"": ""Communications of the ACM, 8(1):53\u201370."", ""year"": 1965}, {""authors"": [""T. Strzalkowski"", ""S. Harabagiu""], ""title"": ""Advances in Open Domain Question Answering"", ""year"": 2008}, {""authors"": [""L. Vanderwende""], ""title"": ""The importance of being important: Question generation"", ""venue"": ""Proceedings of the Workshop on the Question Generation Shared Task and Evaluation Challenge, Arlington, VA."", ""year"": 2008}, {""authors"": [""E.M. Voorhees""], ""title"": ""Overview of the TREC 1999 question answering track"", ""venue"": ""Proceedings of the 8th Text REtreival Conference, Gaithersburg, MD."", ""year"": 1999}, {""authors"": [""J. Wagner"", ""J. Foster"", ""J. van Genabith.""], ""title"": ""Judging grammaticality: Experiments in sentence classification"", ""venue"": ""CALICO Journal, 26(3):474\u2013490."", ""year"": 2009}, {""authors"": [""M. Wang""], ""title"": ""A survey of answer extraction techniques in factoid"", ""year"": 2006}, {""authors"": [""W. Wang"", ""H. Tianyong"", ""L. Wenyin.""], ""title"": ""Automatic question generation for learning evaluation in medicine"", ""venue"": ""6th International Conference on Advances in Web Based Learning,"", ""year"": 2008}, {""authors"": [""X. Wei"", ""W.B. Croft.""], ""title"": ""LDA-based document models for ad-hoc retrieval"", ""venue"": ""Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval,"", ""year"": 2006}, {""authors"": [""A. Zhang"", ""W. Lee.""], ""title"": ""Question classification using support vector machines"", ""venue"": ""Proceedings of the Special Interest Group on Information Retrieval, pages 26\u201332, Toronto."", ""year"": 2003}, {""authors"": [""Z. Zheng"", ""X. Si"", ""E.Y. Chang"", ""X. Zhu.""], ""title"": ""K2Q: Generating natural language questions from keywords with user refinements"", ""venue"": ""Proceedings of the 5th International Joint Conference on Natural"", ""year"": 2011}]",acknowledgments  :We would like to thank the anonymous reviewers for their useful comments. The research reported in this article was supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada – discovery grant and the University of Lethbridge. This work was done when the second author was at the University of Lethbridge.,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :We live in an information age where all kinds of information is easily accessible through the Internet. The increasing demand for access to different types of information available online have interested researchers in a broad range of Information Retrieval–related areas, such as question answering, topic detection and tracking, summarization, multimedia retrieval, chemical and biological informatics, text structuring, and text mining. Although search engines do a remarkable job in searching through a heap of information, they have certain limitations, as they cannot satisfy the end users’ information need to have more direct access to relevant documents. For example, if we ask for the impact of the current global financial crisis in different parts of the world, we can expect to sift through thousands of results for the answer. This fact can be more understandable by the following scenario. When a user enters a query, they are served with a ranked list of relevant documents by the standard document retrieval systems (i.e., search engines),
∗ University of Lethbridge, 4401 University Drive West, Lethbridge, Alberta, T1K 3M4, Canada. E-mail: chali@cs.uleth.ca.
∗∗ Philips Research North America, 345 Scarborough Rd, Briarcliff Manor, New York, 10510, USA. E-mail: sadid.hasan@philips.com.
Submission received: 19 May, 2013; revised submission received: 26 May, 2014; accepted for publication: 22 June, 2014.
doi:10.1162/COLI a 00206
© 2015 Association for Computational Linguistics
and their search task is usually not over (Chali, Joty, and Hasan 2009). The next step for the user is to look into the documents themselves and search for the precise piece of information they were looking for. This method is time-consuming, and a correct answer could easily be missed by either an incorrect query resulting in missing documents or by careless reading. This is why Question Answering (QA) has received immense attention from the information retrieval, information extraction, machine learning, and natural language processing communities in the last 15 years (Hirschman and Gaizauskas 2001; Strzalkowski and Harabagiu 2008; Kotov and Zhai 2010).
The main goal of QA systems is to retrieve relevant answers to natural language questions from a collection of documents rather than using keyword matching techniques to extract documents. Automated QA research focuses on how to respond with exact answers to a wide variety of questions, including: factoid, list, definition, how, why, hypothetical, semantically constrained, and crosslingual questions (Simmons 1965; Kupiec 1993; Voorhees 1999; Hirschman and Gaizauskas 2001; Greenwood 2005; Wang 2006; Moldovan, Clark, and Bowden 2007). One of the main requirements of a QA system is that it must receive a well-formed question as input in order to come up with the best possible correct answer as output. Available studies revealed that humans are not very skilled in asking good questions about a topic of their interest. They are forgetful in nature; this often restricts them to properly express whatever that is peeking in their mind. Therefore, they would benefit from automated Question Generation (QG) systems that can assist in meeting their inquiry needs (Lauer, Peacock, and Graesser 1992; Graesser et al. 2001; Rus and Graesser 2009; Ali, Chali, and Hasan 2010; Kotov and Zhai 2010; Olney, Graesser, and Person 2012). Another benefit of QG is that it can be a good tool to help improve the quality of the QA systems (Graesser et al. 2001; Rus and Graesser 2009). These benefits of a QG system motivate us to address the important problem of topic-to-question generation, where the main goal is to generate all possible questions about a given topic. For example, given the topic Apple Inc. Logos, we would like to generate questions such as What is Apple Inc.?, Where is Apple Inc. located?, Who designed Apple’s Logo?, and so forth.
The problem of topic-to-question generation can be viewed as a generalization of the problem of answering complex questions. Complex questions are essentially broader information requests about a certain topic, whose answers could be obtained from pieces of information scattered in multiple documents. For example, consider the complex question:1 Describe steps taken and worldwide reaction prior to the introduction of the Euro on January 1, 1999. Include predictions and expectations reported in the press. This question is requesting an elaboration about the topic “Introduction of the Euro,” which can be answered by following complex procedures such as question decomposition or inferencing and synthesizing information from multiple documents (e.g., multi-document summarization). Answering complex questions is not easy as it is not always understandable to which direction one should move to search for the answer to a complex question. This situation arises because of the wider focus of the topic that is inherent in the complex question in consideration. For example, a complex question like Describe the tsunami disaster in Japan has a wider focus without a single or well-defined information need. To narrow down the focus, this question can be decomposed into a series of simple questions such as How many people were killed in the tsunami?, How many people became homeless?, Which cities were mostly damaged?, and so on.
1 The example complex questions have been provided according to the guidelines of the Document Understanding Conference (DUC, http://duc.nist.gov/) (2005–2007) tasks.
Decomposing a complex question automatically into simpler questions in this manner such that each of them can be answered individually by using the state-of-the-art QA systems, and then combining the individual answers to form a single answer to the original complex question, has proven effective to deal with the complex question answering problem (Harabagiu, Lacatusu, and Hickl 2006; Hickl et al. 2006; Chali, Hasan, and Imam 2012). Moreover, the generated simple questions can be used as the list of important aspects to act as a guide2 for selecting the most relevant sentences in producing more focused and more accurate summaries as the output of a summarization system (Chali, Hasan, and Imam 2011, 2012). From this discussion, it is obvious that the complex question decomposition problem can be generalized to the problem of topic-to-question generation to help improve the complex question answering systems.
In this article3, we consider the task of automatically generating questions from topics and assume that each topic is associated with a body of texts having useful information about the topic. This assumption has been inherited from the process of how a human asks questions based on their knowledge. For example, if a person knows that a university is an educational institution, then they can ask a question about its faculty and students. In this research, our main goal is to generate fact-based questions4 about a given topic from its associated content information. We generate questions by exploiting the named entity information and the predicate argument structures of the sentences (along with semantic roles) present in the given body of texts. The named entities and the semantic role labels are used to identify relevant parts of a sentence in order to form relevant questions about them. The importance of the generated questions is measured in two steps. In the first step, we identify whether the question is asking something about the topic or something that is very closely related to the topic. We call this the measure of topic relevance. For this purpose, we use Latent Dirichlet Allocation (LDA) (Blei, Ng, and Jordan 2003) to identify the subtopics (which are closely related to the original topic) in the given body of texts and apply the Extended String Subsequence Kernel (ESSK) (Hirao et al. 2003) to calculate their similarity with the questions. In the second step, we judge the syntactic correctness of each generated question. We apply the tree kernel functions (Collins and Duffy 2001) and re-implement the syntactic tree kernel model according to Moschitti et al. (2007) for computing the syntactic similarity of each question with the associated content information. We rank the questions by considering their topic relevance and syntactic correctness scores. Experimental results show the effectiveness of our approach for automatically generating topical questions. The remainder of the article is organized as follows. Section 2 describes the related work. Section 3 presents the description of our QG system. Section 4 explains the experiments and shows evaluation results; Section 5 concludes. 2 related work :Recently, question generation has received immense attention from researchers and different methods have been proposed to accomplish the task in different relevant fields (Andrenucci and Sneiders 2005). McGough et al. (2001) proposed an approach to build a Web-based testing system with the facility of dynamic QG. Wang et al. (2008) showed
2 http://www.nist.gov/tac/2011/Summarization/Guided-Summ.2011.guidelines.html. 3 This article is a longer version of our previously published work (Chali and Hasan 2012c). We provide
more theoretical descriptions and analyses, and conduct our experiments on a larger data set to report new results.
4 We mainly focus on generating who, what, where, which, when, why, and how questions in this research.
a method to automatically generate questions based on question templates (which are created from training on medical articles). Brown, Frishkoff, and Eskenazi (2005) described an approach to automatically generate questions to assess the user’s vocabulary knowledge. Chen, Aist, and Mostow (2009) developed a method to generate questions automatically from informational text to mimic the reader’s self-questioning strategy during reading. On the other hand, Agarwal, Shah, and Mannem (2011) considered the question generation problem beyond the sentence level and designed an approach that uses discourse connectives to generate questions from a given text. Several other QG models have been proposed over the years that deal with transforming answers to questions and utilizing question generation as an intermediate step in the question answering process (Echihabi and Marcu 2003; Hickl et al. 2005). There are some other researchers who have approached the task of generating questions for educational purposes (Mitkov and Ha 2003; Heilman and Smith 2010b).
Question asking and QG are important components in advanced learning technologies such as intelligent tutoring systems and inquiry-based environments (Graesser et al. 2001). A QG system is useful for building better question-asking facilities in intelligent tutoring systems. The Natural Language Processing (NLP), Natural Language Generation, Intelligent Tutoring System, and Information Retrieval communities have currently identified the Text-to-Question generation task as promising candidates for shared tasks5 (Rus and Graesser 2009; Boyer and Piwek 2010). In the Text-to-Question generation task, a QG system is given a text, and the goal is to generate a set of questions for which the text contains answers. The task of generating a question about a given text can be typically decomposed into three subtasks. First, given the source text, a content selection step is necessary to select a target to ask about, such as the desired answer. Second, given a target answer, an appropriate question type is selected (i.e., the form of question to ask is determined). Third, given the content and question type, the actual question is constructed. Based on this principle, several approaches have been described in Boyer and Piwek (2010) that use named entity information, syntactic knowledge, and semantic structures of the sentences to perform the task of generating questions from sentences and paragraphs (Heilman and Smith 2010a; Mannem, Prasad, and Joshi 2010). Inspired by these works, we perform the task of topic-to-question generation using named entity information and semantic structures of the sentences. A task that is similar to ours is the task of keywords-to-question generation that has been addressed recently in Zheng et al. (2011). They propose a user model for jointly generating keywords and questions. However, their approach is based on generating question templates from existing questions, which requires a large set of English questions as training data. In recent years, some other related researchers have proposed the tasks of high-quality question generation (Ignatova, Bernhard, and Gurevych 2008) and generating questions from queries (Lin 2008). Fact-based question generation has been accomplished previously (Rus, Cai, and Graesser 2007; Heilman and Smith 2010b). We also focus on generating fact-based questions in this research.
Besides grammaticality, an effective QG system should focus deeply on the importance of the generated questions (Vanderwende 2008). This motivates the use of a question-ranking module in a typical QG system. Over-generated questions can be ranked using different approaches, such as statistical ranking methods, dependency parsing, identification of the presence of pronouns and named entities, and topic scoring (Heilman and Smith 2010a; Mannem, Prasad, and Joshi 2010; McConnell et al. 2011).
5 http://www.questiongeneration.org/QGSTEC2010.
However, most of these automatic ranking approaches ignore the aspects of complex paraphrasing by not considering lexical semantic variations (e.g., synonymy) when measuring the importance of the questions. In our work, we use LDA (Blei, Ng, and Jordan 2003) to identify the subtopics (which are closely related to the original topic) in the given body of texts. We choose LDA because in recent years it has become one of the most popular topic modeling techniques and has been shown to be effective in several text-related tasks, such as document classification, information retrieval, and question answering (Wei and Croft 2006; Misra, Cappé, and Yvon 2008; Celikyilmaz, HakkaniTur, and Tur 2010).
Once we have the subtopics, we apply ESSK (Hirao et al. 2003) to calculate their similarity with the generated questions. The choice of ESSK is motivated by its successful use in different NLP tasks in recent years (Chali, Hasan, and Joty 2009, 2011; Chali and Hasan 2012a, 2012b). Hirao et al. (2003) introduced ESSK considering all possible senses of each word to perform their summarization task. Their method is effective. However, the fact that they do not disambiguate word senses cannot be disregarded. In our task, we apply ESSK to calculate the similarity between important topics (discovered using LDA) and the generated questions in order to measure the importance of each question. We use disambiguated word senses for this purpose.
Syntactic information has previously been used successfully in question answering (Zhang and Lee 2003; Moschitti and Basili 2006; Moschitti et al. 2007; Chali, Hasan, and Joty 2009, 2011). Pasca and Harabagiu (2001) argued that with the syntactic form of a sentence one can see which words depend on other words. We also feel that there should be a similarity between the words that are dependent in the sentences present in the associated body of texts and the dependency between words of the generated question. This motivates us to propose the use of syntactic kernels in judging the syntactic correctness of the generated questions automatically.
The main goal of our work is to generate as many questions as possible related to the topic. We use the named entity information and the predicate argument structures of the sentences to accomplish this goal. Our approach is different from the set-up in shared tasks (Rus and Graesser 2009; Boyer and Piwek 2010), as we generate a set of basic questions that are useful to add variety in the question space. A paragraph associated with each topic is used as the source of relevant information about the topic. We evaluate our systems in terms of topic relevance, which is different from prior research (Heilman and Smith 2010a; Mannem, Prasad, and Joshi 2010). Syntactic correctness is also an important property of a good question. For this reason, we evaluate our system in terms of syntactic correctness as well. The proposed system will be useful for generating topic-related questions from the associated content information, which can be used to incorporate a “question suggestions for a certain topic” facility in search systems (Kotov and Zhai 2010). For example, if a user searches for some information related to a certain topic, the search system could generate all possible topic-relevant questions from a preexistent related body of texts to provide suggestions. Kotov and Zhai (2010) approached a similar task by proposing a technique to augment the standard ranked list presentation of search results with a question-based interface to refine user-given queries.
The major contributions of our work can be summarized as follows:
We perform the task of topic-to-question generation, which can help users in expressing their information needs. Questions are generated using a set of general-purpose rules based on named entity information and the predicate argument structures of the sentences (along with semantic roles) present in the associated body of texts.
We identify the subtopics (which are closely related to the original topic) in the given body of texts by using LDA and calculate their similarity with the questions by applying ESSK (with disambiguated word senses). This helps us to measure the importance of each question.
We compute the syntactic similarity of each question with its associated content information by applying the tree kernel functions with the re-implementation of the syntactic tree kernel model. In this way, we judge the syntactic correctness of each generated question automatically.
We evaluate the ESSK similarity scores and the syntactic similarity scores in a ranking framework and show that the use of ESSK and syntactic kernels improve the relevance and the syntactic correctness of the top-ranked questions, respectively.
We identify circumstances in which our approach performs well and show that, using additional experiments by narrowing down the topic focus. Experiments with the topics about people (biographical focus) reveal improvements in the overall results. 3 topic-to-question generation :Our QG approach mainly builds on four steps. In the first step, complex sentences (from the given body of texts) related to a topic are simplified, as it is easier to generate questions from simple sentences. In the next step, named entity information and predicate argument structures of the sentences are extracted and are then used to generate questions. In the third step, LDA is used to identify important subtopics from the given body of texts, and then ESSK is applied to find their similarity with the generated questions. In the final step, a syntactic tree kernel is used and syntactic similarity between the generated questions and the sentences present in the body of texts determines the syntactic correctness of the questions. Questions are then ranked by considering the ESSK similarity scores and the syntactic similarity scores. We present an architectural diagram (Figure 1) to show the different components of our system and describe the overall procedure in the following subsections. Sentences may have complex grammatical structure with multiple embedded clauses. Therefore, the first step of our proposed system is to simplify the complex sentences with the intention of generating more accurate questions. We use the simplified factual statement extractor model6 of Heilman and Smith (2010a). Their model extracts the simpler forms of the complex source sentence by altering lexical items, syntactic structure, and semantics, as well as by removing phrase types such as leading conjunctions, sentence-level modifying phrases, and appositives. For example, given a complex sentence s, we get the corresponding simple sentences as follows:
Complex Sentence (s): Apple’s first logo, designed by Jobs and Wayne, depicts Sir Isaac Newton sitting under an apple tree. Simple Sentence (1): Apple’s first logo is designed by Jobs and Wayne.
6 Available at http://www.ark.cs.cmu.edu/mheilman/questions/.
Simple Sentence (2): Apple’s first logo depicts Sir Isaac Newton sitting under an apple tree. In the second step of our system, we at first process the simple sentences in order to generate all possible questions from them. We use the Illinois Named Entity Tagger,7 a state-of-the-art named entity (NE) tagger that tags plain text with named entities (people, organizations, locations, miscellaneous) (Ratinov and Roth 2009). Once we tag the topic under consideration and its associated body of texts, we use some general purpose rules to create some basic questions even though the answer is not present in the body of texts. For example, Apple Inc. is tagged as an organization, so we generate a question: Where is Apple Inc. located?. The main motivation behind generating such questions is to add variety to the generated question space. The basic questions are useful when there is very little or no knowledge available for a certain topic in consideration. This assumption is inherited from the scenario in the real world where a human can ask questions having very limited background knowledge about the topic. For example, if a person knows nothing about a university, they can ask What is it? or, if they at least know that a university is an institution, then they can ask the question, Where is it located?. In Table 1, we show some example rules for the basic questions generated in this work.
Our next task is to generate specific questions from the sentences present in the given body of texts. For this purpose, we parse the sentences semantically using a
7 Available at http://cogcomp.cs.illinois.edu/.
Semantic Role Labeling (SRL) system (Kingsbury and Palmer 2002; Hacioglu et al. 2003), ASSERT.8 ASSERT is an automatic statistical semantic role tagger that can annotate naturally occuring text with semantic arguments. When presented with a sentence, it performs a full syntactic analysis of the sentence, automatically identifies all the verb predicates in that sentence, extracts features for all constituents in the parse tree relative to the predicate, and identifies and tags the constituents with the appropriate semantic arguments. For example, the output of the SRL system for the sentence Apple’s first logo is designed by Jobs and Wayne is: [ARG1 Apple ’s first logo] is [TARGET designed ] [ARG0 by Jobs and Wayne]. The output contains one verb (predicate) with its arguments (i.e., semantic roles). These arguments are used to generate specific questions from the sentences. For example, we can replace [ARG1 ..] with What and generate a question as: What is designed by Jobs and Wayne?. Similarly, [ARG0 ..] can be replaced and the question: Who designed Apple’s first logo? can be generated. The semantic roles ARG0...ARG5 are called mandatory arguments. There are some additional arguments or semantic roles that can be tagged by ASSERT. They are called optional arguments and they start with the prefix ARGM. These are defined by the annotation guidelines set in Palmer, Gildea, and Kingsbury (2005). A set of about 350 general-purpose rules are used to transform the semantic-role labeled sentences into the questions. The rules were set up in a way that we could use the semantic role information to find the potential answer words in a sentence that would be replaced by suitable question words. In the case of a mandatory argument, the choice of question word depends on the argument’s named entity tag (who for a person, where for a location, etc.). Table 2 shows how different semantic roles can be replaced by possible question words in order to generate a question. In the third step of our proposed system, we pass the generated questions to the importance judgment module that uses LDA and ESSK to assign a topic relevance score to each question. The detailed procedure is discussed in the following subsections.
3.3.1 Latent Dirichlet Allocation (LDA). To measure the importance of the generated questions, we use LDA (Blei, Ng, and Jordan 2003) to identify the important subtopics9 from the given body of texts. LDA is a probabilistic topic modeling technique where the main principle is to view each document as a mixture of various topics. Here each topic is a probability distribution over words. LDA assumes that documents are made
8 Available at http://cemantix.org/assert.html. 9 The term sub-topic is used in the LDA topic modeling sense, which represents a probability distribution
over words.
up of words and word ordering is not important (“bag-of-words” assumption) (Misra, Cappé, and Yvon 2008). The main idea is to choose a distribution over topics while generating a new document. For each word in the new document, a topic is randomly chosen according to this distribution and a word is drawn from that topic. LDA uses a generative topic modeling approach to specify the following distribution over words within a document:
P(wi) = K∑
j=1
P(wi|zi = j)P(zi = j) (1)
where K is the number of topics, P(wi|zi = j) is the probability of word wi under topic j, and P(zi = j) is the sampling probability of topic j for the ith word. The multinomial distributions φ(j) = P(w|zi = j) and θ(d) = P(z) are termed as topic-word distribution and document-topic distribution, respectively (Blei, Ng, and Jordan 2003). A Dirichlet (α) prior is placed on θ and a Dirichlet (β) prior is set on φ to refine this basic model (Griffiths and Steyvers 2002; Blei, Ng, and Jordan 2003). Now the main goal is to estimate the two parameters: θ and φ. We apply this framework directly to solve our problem by considering each topic-related body of texts as a document. We use a GUIbased toolkit for topic modeling10 that uses the popular MALLET (McCallum 2002) toolkit for the back-end. The LDA model is built on the development set11 (Section 4.2). The process starts by removing a list of “stop words” from the document and runs 200 iterations of Gibbs sampling (Geman and Geman 1984) to estimate the parameters θ and φ. From each body of texts, we discover K topics and choose the most frequent words from the most likely unigrams as the desired subtopics. For example, from the associated body of texts of the topic Apple Inc. Logos, we get these subtopics: janoff, themes, logo, color, apple.
3.3.2 Extended String Subsequence Kernel (ESSK). Once we identify the subtopics, we apply ESSK (Hirao et al. 2003) to measure their similarity with the generated questions. In the general ESSK, each word in a sentence is considered an “alphabet,” and the alternative is all its possible senses. However, our ESSK implementation considers the
10 Available at http://code.google.com/p/topic-modeling-tool/. 11 The model was built and tested according to the guidelines of the topic modeling toolkit we used.
alternative of each word as its disambiguated sense. We use a dictionary-based Word Sense Disambiguation (WSD) system (Chali and Joty 2007) assuming one sense per discourse. We use WordNet (Fellbaum 1998) to find the semantic relations (such as repetition, synonym, hypernym and hyponym, holonym and meronym, and gloss) for all the words in a text. We assign a weight to each semantic relation based on heuristics and use all of them. Our WSD technique is decomposed into two steps: (1) building a representation of all possible senses of the words and (2) disambiguating the words based on the highest score. To be specific, each candidate word from the context is expanded to all of its senses. A disambiguation graph is constructed as the intermediate representation where the nodes denote word instances with their WordNet senses, and the weighted edges (connecting the senses of two different words) represent semantic relations. This graph is exploited to perform the WSD. We sum the weights of all edges, leaving the nodes under their different senses. The sense with the highest score is considered to be the most probable sense. In case of a tie between two or more senses, we select the sense that comes first in WordNet, because WordNet orders the senses of a word by decreasing order of their frequency. Our preliminary experiments suggested that WSD has a positive impact on the performance of our proposed system.
ESSK is used to measure the similarity between all possible subsequences of the question words/senses and topic words/senses. We calculate the similarity score Sim(Ti, Qj) using ESSK, where Ti denotes a topic/sub-topic word sequence and Qj stands for a generated question. Formally, ESSK is defined as follows:12
Kessk(T, Q) = d∑
m=1 ∑ ti∈T ∑ qj∈Q Km(ti, qj)
Km(ti, qj) = {
val(ti, qj) if m = 1 K
′ m−1(ti, qj) · val(ti, qj)
Here, K ′ m(ti, qj) is defined in the following. ti and qj are nodes of T and Q, respectively. The function val(t, q) returns the number of common attributes (i.e., the number of common words/senses) to the given nodes t and q.
K ′ m(ti, qj) = { 0 if j = 1 λK ′ m(ti, qj−1) + K ′′ m(ti, qj−1)
Here, λ is the decay parameter for the number of skipped words. K ′′ m(ti, qj) is defined as:
K ′′ m(ti, qj) = { 0 if i = 1 λK ′′ m(ti−1, qj) + Km(ti−1, qj)
12 The formulae denote a dynamic programming technique to compute the ESSK similarity score where d is the vector space dimension (i.e., the number of all possible subsequences of up to length d). More information about these formulae can be obtained from Hirao et al. (2003, 2004).
Finally, the similarity measure is defined after normalization:
simessk(T, Q) = Kessk(T, Q)√
Kessk(T, T)Kessk(Q, Q) The next step of our system is to judge the syntactic correctness of the generated questions. The generated questions might be syntactically incorrect due to the process of automatic question generation. It is time-consuming and considerable human intervention is necessary to check for the syntactically incorrect questions manually. We strongly believe that a question should have a similar syntactic structure to a sentence from which it is generated. For example, the sentence Apple’s first logo is designed by Jobs and Wayne., and the generated question What is designed by Jobs and Wayne? are syntactically similar. An example of an ungrammatical generated question that is not very similar to its source is: Janoff presented Jobs What?. To judge the syntactic correctness of each generated question automatically, we apply the tree kernel functions and re-implement the syntactic tree kernel model according to Moschitti et al. (2007) for computing the syntactic similarity of each question with the associated content information. We first parse the sentences and the questions into syntactic trees using the Charniak parser13 (Charniak 1999). Then, we calculate the similarity between the two corresponding trees using the tree kernel method (Collins and Duffy 2001). We convert each parenthetic representation generated by the Charniak parser into its corresponding tree and give the trees as input to the tree kernel functions for measuring the syntactic similarity.
The tree kernel function computes the number of common subtrees between two trees and gives the similarity score between each sentence in the given body of texts and the generated question based on the syntactic structure. Each sentence14 contributes a score to the questions and then the questions are ranked by considering the average of similarity scores. 4 experiments : We consider the task of automatically generating questions from topics where each topic is associated with a body of texts having a useful description about the topic. The question-ranking module of the proposed QG system ranks the questions by combining the topic relevance scores and the syntactic similarity scores of Section 3.3 and Section 3.4 using the following formula:
w ∗ ESSKscore + (1 − w) ∗ SYNscore (2)
13 Available at https://github.com/BLLIP/bllip-parser. 14 We consider that a question is syntactically fluent as well as relevant to the topic if it has similar syntactic
subtrees to those of the most sentences in the body of texts.
Here, w is the importance parameter, which holds a value in [0, 1]. We kept w = 0.5 to give equal importance15 to topic relevance and syntactic correctness. To run our experiments, we use the data set provided in the Question Generation Shared Task and Evaluation Challenge16 (2010) for the task of question generation from paragraphs. This data set consists of 60 paragraphs about 60 topics that were originally collected from several Wikipedia, OpenLearn, and Yahoo!Answers articles. The paragraphs contain around 5–7 sentences for a total of 100–200 tokens (including punctuation). This data set includes a diversity of topics of general interest. We consider these topics and treat the paragraphs as their associated useful content information in order to generate a set of questions using our proposed QG approach. We randomly select 10 topics and their associated paragraphs as the development data.17 A total of 2,186 questions are generated from the remaining 50 topics (test data) to be ranked. 4.3.1 Methodology. We use a methodology derived from Boyer and Piwek (2010) and Heilman and Smith (2010b) to evaluate the performance of our QG systems. Three native English-speaking university graduate students judge the quality of the top-ranked 20% questions using two criteria: topic relevance and syntactic correctness. For topic relevance, the given score is an integer between 1 (very poor) and 5 (very good) and is guided by the consideration of the following aspects: 1. Semantic correctness (i.e., the question is meaningful and related to the topic), 2. Correctness of question type (i.e., a correct question word is used), and 3. Referential clarity (i.e., it is clearly possible to understand what the question refers to). For syntactic correctness, the assigned score is also an integer between 1 (very poor) and 5 (very good). Whether a question is grammatically correct or not is checked here. The judges were asked to read the topics with their associated body of texts and then rate the top-ranked questions generated by different systems. For each question, we calculate the average of the judges’ scores. The judges were provided with an annotation guideline and sample judgments, according to the methodology derived from Boyer and Piwek (2010) and Heilman and Smith (2010b). The same judges evaluated all the system outputs and they were blind to the system identity when judging. No guidelines were provided on the relative importance of the various aspects that made the judgment task subjective. The inter-annotator agreement of Fleiss’s κ = 0.41, 0.45, 0.62, and 0.33 are computed for the three judges for the results in Tables 3–6, indicating moderate (for the first two tables), and substantial and fair agreement (Landis and Koch 1977) between the raters, respectively. These κ values were shown to be acceptable in the literature for the relevant NLP tasks (Dolan and Brockett 2005; Glickman, Dagan, and Koppel 2005; Heilman and Smith 2010b).
15 A syntactically incorrect question is not useful even if it is relevant to the topic. This motivated us to give equal importance to topic relevance and syntactic correctness. The parameter w can be tuned to investigate its impact on the system performance. 16 http://www.questiongeneration.org/mediawiki. 17 We use these data to build necessary general purpose rules for our QG model.
4.3.2 Systems for Comparison. We report the performance of the following systems in order to do a meaningful comparison with our proposed QG system:
(1) Baseline1: This is our QG system without any question-ranking method applied to it. Here, we randomly select top 20% questions and rate them. (2) Baseline2: For our second baseline, we build a QG system using an alternative topic modeling approach. Here, we use a topic signature model (instead of using LDA as discussed in Section 3.3.1) (Lin and Hovy 2000) to identify the important subtopics from the sentences present in the body of texts. The subtopics are the important words in the context that are closely related to the topic and have significantly greater probability of occurring in the given text compared with a large background corpus. We use a topic signature computation tool18 for this purpose. The background corpus that is used in this tool contains 5,000 documents from the English GigaWord Corpus. For example, from the given body of texts of the topic Apple Inc. Logos, we get these subtopics: jobs, logo, themes, rainbow, monochromatic. Then we use the same steps of Sections 3.3.2 and 3.4, and use Equation (2) to combine the scores. We evaluate the top-ranked 20% questions and show the results. (3) State-of-the-art: We choose a publicly available state-of-the-art QG system19 to generate questions from the sentences in the body of texts. This system was shown to achieve good performance in generating fact-based questions about the content of a given article (Heilman and Smith 2010b). Their method ranks the questions automatically using a logistic regression model. Given a paragraph as input, this system processes each sentence and generates a set of ranked questions for the entire paragraph. We evaluate the top-ranked 20% questions20 and report the results.
4.3.3 Results and Discussion. Table 3 shows the average topic relevance and syntactic correctness scores for all the systems. From these results, we can see that the proposed QG system improves the topic relevance and syntactic correctness scores over the Baseline1 system by 62% and 35%, respectively, and improves the topic relevance and syntactic correctness scores over the Baseline2 system by 7%, and 8%, respectively. On the other hand, the proposed QG system improves the topic relevance and syntactic correctness scores over the state-of-the-art system by 4% and 3%, respectively. From these results, we can clearly observe the effectiveness of our proposed QG system. The improvements in the results are statistically significant21 (p < 0.05).
The main goal of this work was to generate as many questions as possible related to the topic. For this reason, we considered generating the basic questions. These questions
18 Available at http://www.cis.upenn.edu/∼lannie/topicS.html. 19 Available at http://www.ark.cs.cmu.edu/mheilman/questions/. 20 We ignore the yes-no questions for our task. 21 We tested statistical significance using Student’s t test.
were also useful to provide variety in the question space. We generated these questions using the named entity information. As the performance of the NE taggers were unsatisfactory, we had a few of these questions generated. In most cases, these questions were outranked by other important questions, which included a combination of topics and subtopics to show higher topic relevance score measured by ESSK. Therefore, they do not have a considerable impact on the evaluation statistics. We claim that the overall performance of our systems could be further improved if the accuracy of the NE tagger and the semantic role labeler could be increased.
Acceptability Test. In another evaluation setting, the three annotators judge the questions for their overall acceptability as a good question. If a question shows no deficiency in terms of the criteria considered for topic relevance and syntactic correctness, it is termed as acceptable. We evaluate the top 15% and top 30% questions separately for each QG system and report the results indicating the percentage of questions rated as acceptable in Table 4. The results indicate that the percentage of the questions rated acceptable is reduced when we evaluate a greater number of questions—which proves the effectiveness of our QG system.
Narrowing Down the Focus. We run further experiments by narrowing down the topic focus. We consider only the topics about people (biographical focus). We choose 50 people as our topics from the list of the 20th century’s 100 most influential people, published in Time magazine in 1999 and obtained the paragraphs containing their biographical information from Wikipedia articles.22 We generate a total of 1, 845 questions from the 50 topics considered and rank them using different ranking schemes as discussed before. We evaluate the top 20% questions using the similar evaluation methodologies and report the results in Table 5. From these results, we can see that the proposed QG
22 http://en.wikipedia.org/wiki/Time 100.
system improves the topic relevance and syntactic correctness scores over the Baseline1 system by 33% and 35%, respectively, and improves the topic relevance and syntactic correctness scores over the Baseline2 system by 8% and 9%, respectively. Moreover, the proposed QG system improves both the topic relevance and syntactic correctness scores over the state-of-the-art system by 4%. From these results, we can clearly observe the effectiveness of our proposed QG system when we narrow down the topic focus. We also evaluate the top 15% and top 30% questions separately for each QG system and report the results, indicating the percentage of questions rated as acceptable in Table 6. From these tables, we can clearly see the improvements in all the scores for all the QG approaches. This is reasonable because the accuracy of the NE tagger and the semantic role labeler is increased for the biographical data.23 These results further demonstrate that the proposed system is significantly better (at p < 0.05) than the other considered systems. We plan to make our created resources available to other researchers.
4.3.4 An Input-Output Example. An input to our systems is, for instance,24 the topic Apple Inc. Logos with the associated content information (body of texts):
Apple’s first logo, designed by Jobs and Wayne, depicts Sir Isaac Newton sitting under an apple tree. Almost immediately, though, this was replaced by Rob Janoff’s “rainbow Apple”, the now-familiar rainbow-colored silhouette of an apple with a bite taken out of it. Janoff presented Jobs with several different monochromatic themes for the “bitten” logo, and Jobs immediately took a liking to it. While Jobs liked the logo, he insisted it be in color to humanize the company. The Apple logo was designed with a bite so that it would be recognized as an apple rather than a cherry. The colored stripes were conceived to make the logo more accessible, and to represent the fact the monitor could reproduce images in color. In 1998, with the roll-out of the new iMac, Apple discontinued the rainbow theme and began to use monochromatic themes, nearly identical in shape to its previous rainbow incarnation.
The output of our systems is the ranked lists of questions. We show an example output in Table 7. To provide a more detailed analysis of our results, the average output scores of the example questions are presented in Table 8. From this table, we can understand how different aspects of the evaluation criteria affected the performance of the different systems. For example, Q1 of the proposed system was given a very good score due to
23 Although a few basic questions were generated compared with other important questions containing topical words, we believe they did not have a considerable impact on the overall performance of our system. 24 The example input text is provided from the Question Generation Shared Task and Evaluation Challenge (QGSTEC 2010) data set that we used for our experiments (Section 4.2).
its relevance to the topic in consideration. On the other hand, Q3 of the state-of-the-art was assigned a lower score due to its lack of clarity with respect to the topic. 5 conclusion :In this article, we have considered the task of automatically generating questions from topics where each topic is associated with a body of texts containing useful information. The proposed method exploits named entity and semantic role labeling information to accomplish the task. A key aspect of our approach was the use of latent Dirichlet allocation (LDA) to automatically discover the hidden subtopics from the sentences. We have proposed a novel method to rank the generated questions by considering: (1) subtopical similarity determined using ESSK algorithm in combination with word sense disambiguation, and (2) syntactic similarity determined using the syntactic tree kernel based method. We have compared the proposed question generation (QG) system with two baseline systems and one state-of-the-art system. The evaluation results show that the proposed QG system significantly outperforms all other considered systems, as our top-ranked system generated questions were found to be better in topic-relevance and syntactic correctness than those of the other systems. Our results demonstrated that judging syntactic correctness of the generated questions using the syntactic tree kernel based model was suitable in our question generation setting. We would like to further
our research by using other available measures such as an n-gram language model or a parser confidence score (Wagner, Foster, and van Genabith 2009) in order to see how they would perform on the same task. In this article, we have also extended our experiments by narrowing down the topic focus. In this experiment, we have considered people as topics. A rigorous analysis of the evaluation results has revealed that the performance of our proposed QG system can be enhanced if we narrow down the topic focus. We hope to carry on these ideas and develop further mechanisms for question generation based on the dependency features of the answers and answer finding (Li and Roth 2006; Pinchak and Lin 2006). This paper is concerned with automatic generation of all possible questions from a topic of interest. Specifically, we consider that each topic is associated with a body of texts containing useful information about the topic. Then, questions are generated by exploiting the named entity information and the predicate argument structures of the sentences present in the body of texts. The importance of the generated questions is measured using Latent Dirichlet Allocation by identifying the subtopics (which are closely related to the original topic) in the given body of texts and applying the Extended String Subsequence Kernel to calculate their similarity with the questions. We also propose the use of syntactic tree kernels for the automatic judgment of the syntactic correctness of the questions. The questions are ranked by considering both their importance (in the context of the given body of texts) and syntactic correctness. To the best of our knowledge, no previous study has accomplished this task in our setting. A series of experiments demonstrate that the proposed topic-to-question generation approach can significantly outperform the state-of-the-art results. [{""affiliations"": [], ""name"": ""Yllias Chali""}, {""affiliations"": [], ""name"": ""Sadid A. Hasan""}] SP:2ce1f080cb051186b15c05b770543e8bca73d2c6 [{""authors"": [""M. Agarwal"", ""R. Shah"", ""P. Mannem.""], ""title"": ""Automatic question generation using discourse cues"", ""venue"": ""Proceedings of the 6th Workshop on Innovative Use of NLP for Building Educational Applications,"", ""year"": 2011}, {""authors"": [""H. Ali"", ""Y. Chali"", ""S.A. Hasan.""], ""title"": ""Automation of question generation from sentences"", ""venue"": ""Proceedings of QG2010: The Third Workshop on Question Generation, pages 58\u201367, Pittsburgh, PA."", ""year"": 2010}, {""authors"": [""A. Andrenucci"", ""E. Sneiders.""], ""title"": ""Automated question answering: Review of the main approaches"", ""venue"": ""Proceedings of the 3rd International Conference on Information Technology and Applications (ICITA\u201905),"", ""year"": 2005}, {""authors"": [""D.M. Blei"", ""A.Y. Ng"", ""M.I. Jordan.""], ""title"": ""Latent Dirichlet allocation"", ""venue"": ""Journal of Machine Learning Research, 3:993\u20131022."", ""year"": 2003}, {""authors"": [""J.C. Brown"", ""G.A. Frishkoff"", ""M. Eskenazi.""], ""title"": ""Automatic question generation for vocabulary assessment"", ""venue"": ""Proceedings of the Conference on Human Language Technology and Empirical"", ""year"": 2005}, {""authors"": [""A. Celikyilmaz"", ""D. Hakkani-Tur"", ""G. Tur""], ""title"": ""LDA based similarity modeling"", ""year"": 2010}, {""authors"": [""Y. Chali"", ""S.A. Hasan.""], ""title"": ""On the effectiveness of using sentence compression models for query-focused multi-document summarization"", ""venue"": ""Proceedings of the 24th International"", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan.""], ""title"": ""Query-focused multi-document summarization: Automatic data annotations and supervised learning approaches"", ""venue"": ""Journal of Natural Language"", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan.""], ""title"": ""Towards automatic topical question generation"", ""venue"": ""Proceedings of the 24th International Conference on Computational Linguistics (COLING 2012), pages 475\u2013492, Mumbai."", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""K. Imam.""], ""title"": ""An aspect-driven random walk model for topic-focused multi-document summarization"", ""venue"": ""Proceedings of the 7th Asian Information Retrieval Societies"", ""year"": 2011}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""K. Imam.""], ""title"": ""Learning good decompositions of complex questions"", ""venue"": ""Proceedings of the 17th International Conference on Applications of Natural Language Processing"", ""year"": 2012}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""S.R. Joty""], ""title"": ""Do automatic annotation techniques have any impact on supervised complex question answering"", ""venue"": ""In Proceedings of the Joint Conference of the 47th Annual Meeting"", ""year"": 2009}, {""authors"": [""Y. Chali"", ""S.A. Hasan"", ""S.R. Joty""], ""title"": ""Improving graph-based random walks for complex question answering"", ""year"": 2011}, {""authors"": [""Y. Chali"", ""S.R. Joty.""], ""title"": ""Word sense disambiguation using lexical cohesion"", ""venue"": ""Proceedings of the 4th International Conference on Semantic Evaluations, pages 476\u2013479, Prague."", ""year"": 2007}, {""authors"": [""Y. Chali"", ""S.R. Joty"", ""S.A. Hasan.""], ""title"": ""Complex question answering: Unsupervised learning approaches and experiments"", ""venue"": ""Journal of Artificial Intelligence Research, 35:1\u201347."", ""year"": 2009}, {""authors"": [""E. Charniak""], ""title"": ""A maximum-entropyinspired parser"", ""venue"": ""Technical Report CS-99-12, Brown University, Computer Science Department, Providence, RI."", ""year"": 1999}, {""authors"": [""W. Chen"", ""G. Aist"", ""J. Mostow.""], ""title"": ""Generating questions automatically from informational text"", ""venue"": ""Proceedings of the 2nd Workshop on Question Generation (AIED 2009), pages 17\u201324, Arlington, VA."", ""year"": 2009}, {""authors"": [""M. Collins"", ""N. Duffy.""], ""title"": ""Convolution Kernels for natural language"", ""venue"": ""Proceedings of Neural Information Processing Systems, pages 625\u2013632, Vancouver."", ""year"": 2001}, {""authors"": [""W.B. Dolan"", ""C. Brockett.""], ""title"": ""Automatically constructing a corpus of sentential paraphrases"", ""venue"": ""Proceedings of the 3rd International Workshop on Paraphrasing (IWP2005), pages 9\u201316, Jeju Island."", ""year"": 2005}, {""authors"": [""A. Echihabi"", ""D. Marcu.""], ""title"": ""A noisy-channel approach to question answering"", ""venue"": ""Proceedings of the 41st Annual Meeting on Association for Computational Linguistics - Volume 1, pages 16\u201323,"", ""year"": 2003}, {""authors"": [""C. Fellbaum""], ""title"": ""WordNet - An Electronic Lexical Database"", ""venue"": ""MIT Press."", ""year"": 1998}, {""authors"": [""S. Geman"", ""D. Geman.""], ""title"": ""Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images"", ""venue"": ""IEEE Transactions on Pattern Analysis and Machine Intelligence, 6:721\u2013741."", ""year"": 1984}, {""authors"": [""O. Glickman"", ""I. Dagan"", ""M. Koppel.""], ""title"": ""A probabilistic classification approach for lexical textual entailment"", ""venue"": ""AAAI, pages 1,050\u20131,055, Pittsburgh, PA."", ""year"": 2005}, {""authors"": [""A.C. Graesser"", ""K. VanLehn"", ""C.P. Rose"", ""P.W. Jordan"", ""D. Harter.""], ""title"": ""Intelligent tutoring systems with conversational dialogue"", ""venue"": ""AI Magazine, 22(4):39\u201352."", ""year"": 2001}, {""authors"": [""M.A. Greenwood""], ""title"": ""Open-Domain Question Answering"", ""venue"": ""Ph.D. thesis, Department of Computer Science, University of Sheffield."", ""year"": 2005}, {""authors"": [""T.L. Griffiths"", ""M. Steyvers.""], ""title"": ""Prediction and semantic association"", ""venue"": ""NIPS\u201902, pages 11\u201318, Cambridge, MA."", ""year"": 2002}, {""authors"": [""K. Hacioglu"", ""S. Pradhan"", ""W. Ward"", ""J.H. Martin"", ""D. Jurafsky.""], ""title"": ""Shallow semantic parsing using support vector machines"", ""venue"": ""Technical Report TR-CSLR-2003-03, University of Colorado,"", ""year"": 2003}, {""authors"": [""S. Harabagiu"", ""F. Lacatusu"", ""A. Hickl.""], ""title"": ""Answering complex questions with random walk models"", ""venue"": ""Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development"", ""year"": 2006}, {""authors"": [""M. Heilman"", ""N.A. Smith.""], ""title"": ""Extracting simplified statements for factual question generation"", ""venue"": ""Proceedings of the Third Workshop on Question Generation, pages 11\u201320,"", ""year"": 2010}, {""authors"": [""M. Heilman"", ""N.A. Smith.""], ""title"": ""Good question! Statistical ranking for question generation"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the"", ""year"": 2010}, {""authors"": [""A. Hickl"", ""J. Lehmann"", ""D. Moldovan"", ""S. Harabagiu.""], ""title"": ""Experiments with interactive question-answering"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""A. Hickl"", ""P. Wang"", ""J. Lehmann"", ""Sanda Harabagiu.""], ""title"": ""Ferret: Interactive question-answering for real-world environments"", ""venue"": ""Proceedings of the COLING/ACL on Interactive Presentation"", ""year"": 2006}, {""authors"": [""T. Hirao"", ""J. Suzuki"", ""H. Isozaki"", ""E. Maeda.""], ""title"": ""NTT\u2019s multiple document summarization system for DUC2003"", ""venue"": ""Proceedings of the Document Understanding Conference, Edmonton."", ""year"": 2003}, {""authors"": [""T. Hirao"", ""J. Suzuki"", ""H. Isozaki"", ""E. Maeda.""], ""title"": ""Dependency-based sentence alignment for multiple document summarization"", ""venue"": ""Proceedings of COLING 2004, pages 446\u2013452, Geneva."", ""year"": 2004}, {""authors"": [""L. Hirschman"", ""R. Gaizauskas.""], ""title"": ""Natural language question answering: The view from here"", ""venue"": ""Natural Language Engineering, 7(4):275\u2013300."", ""year"": 2001}, {""authors"": [""K. Ignatova"", ""D. Bernhard"", ""I. Gurevych.""], ""title"": ""Generating high quality questions from low quality questions"", ""venue"": ""Proceedings of the Workshop on the Question Generation Shared Task and Evaluation Challenge,"", ""year"": 2008}, {""authors"": [""P. Kingsbury"", ""M. Palmer.""], ""title"": ""From Treebank to PropBank"", ""venue"": ""Proceedings of the International Conference on Language Resources and Evaluation, pages 1,989\u20131,993, Las Palmas."", ""year"": 2002}, {""authors"": [""A. Kotov"", ""C. Zhai.""], ""title"": ""Towards natural question guided search"", ""venue"": ""Proceedings of the 19th International Conference on the World Wide Web, WWW \u201910, pages 541\u2013550, Raleigh, NC."", ""year"": 2010}, {""authors"": [""J. Kupiec""], ""title"": ""MURAX: A robust linguistic approach for question answering using an on-line encyclopedia"", ""venue"": ""SIGIR, pages 181\u2013190, Pittsburgh, PA."", ""year"": 1993}, {""authors"": [""J.R. Landis"", ""G.G. Koch.""], ""title"": ""The measurement of observer agreement for categorical data"", ""venue"": ""Biometrics, 33(1):159\u2013174."", ""year"": 1977}, {""authors"": [""X. Li"", ""D. Roth.""], ""title"": ""Learning question classifiers: The role of semantic information"", ""venue"": ""Journal of Natural Language Engineering, 12(3):229\u2013249."", ""year"": 2006}, {""authors"": [""C.Y. Lin""], ""title"": ""Automatic question generation from queries"", ""venue"": ""Proceedings of the Workshop on the Question Generation Shared Task and Evaluation Challenge, Arlington, VA."", ""year"": 2008}, {""authors"": [""C.Y. Lin"", ""E.H. Hovy.""], ""title"": ""The automated acquisition of topic signatures for text summarization"", ""venue"": ""Proceedings of the 18th Conference on Computational Linguistics, pages 495\u2013501, Saarbr\u00fcken."", ""year"": 2000}, {""authors"": [""P. Mannem"", ""R. Prasad"", ""A. Joshi.""], ""title"": ""Question generation from paragraphs at UPenn: QGSTEC system description"", ""venue"": ""Proceedings of the Third Workshop on Question Generation, pages 84\u201391,"", ""year"": 2010}, {""authors"": [""A.K. McCallum""], ""title"": ""MALLET: A machine learning for language toolkit"", ""venue"": ""http://mallet.cs.umass.edu."", ""year"": 2002}, {""authors"": [""C.C. McConnell"", ""P. Mannem"", ""R. Prasad"", ""A. Joshi.""], ""title"": ""A new approach to ranking over-generated questions"", ""venue"": ""Proceedings of the AAAI Fall Symposium on Question Generation, pages 45\u201348,"", ""year"": 2011}, {""authors"": [""J. McGough"", ""J. Mortensen"", ""J. Johnson"", ""S. Fadali.""], ""title"": ""A Web-based testing system with dynamic question generation"", ""venue"": ""ASEE/IEEE Frontiers in Education Conference, pages S3C-23\u201328"", ""year"": 2001}, {""authors"": [""H. Misra"", ""O. Capp\u00e9"", ""F. Yvon.""], ""title"": ""Using LDA to detect semantically incoherent documents"", ""venue"": ""Proceedings of the Twelfth Conference on Computational"", ""year"": 2008}, {""authors"": [""R. Mitkov"", ""L.A. Ha.""], ""title"": ""Computer-aided generation of multiple-choice tests"", ""venue"": ""Proceedings of the HLT-NAACL 03 Workshop on Building Educational Applications Using Natural"", ""year"": 2003}, {""authors"": [""D. Moldovan"", ""C. Clark"", ""M. Bowden.""], ""title"": ""Lymba\u2019s PowerAnswer 4 in TREC 2007"", ""venue"": ""Proceedings of the 16th Text REtreival Conference, Gaithersburg, MD."", ""year"": 2007}, {""authors"": [""A. Moschitti"", ""R. Basili.""], ""title"": ""A tree kernel approach to question and answer classification in question answering systems"", ""venue"": ""Proceedings of the 5th International Conference on"", ""year"": 2006}, {""authors"": [""A. Moschitti"", ""S. Quarteroni"", ""R. Basili"", ""S. Manandhar.""], ""title"": ""Exploiting syntactic and shallow semantic kernels for question/answer classification"", ""venue"": ""Proceedings of the 45th Annual Meeting"", ""year"": 2007}, {""authors"": [""A.M. Olney"", ""A.C. Graesser"", ""N.K. Person.""], ""title"": ""Question generation from concept maps"", ""venue"": ""Dialogue and Discourse, 3(2):75\u201399."", ""year"": 2012}, {""authors"": [""M. Palmer"", ""D. Gildea"", ""P. Kingsbury.""], ""title"": ""The proposition bank: An annotated corpus of semantic roles"", ""venue"": ""Computational Linguistics, 31(1):71\u2013106."", ""year"": 2005}, {""authors"": [""M. Pasca"", ""S.M. Harabagiu.""], ""title"": ""Answer mining from on-line documents"", ""venue"": ""Proceedings of the Association for Computational Linguistics 39th Annual Meeting and 10th Conference of the European"", ""year"": 2001}, {""authors"": [""C. Pinchak"", ""D. Lin.""], ""title"": ""A probabilistic answer type model"", ""venue"": ""Proceedings of the 11th Conference of the European Chapter of the Association for Computational Linguistics, pages 393\u2013400, Trento."", ""year"": 2006}, {""authors"": [""L. Ratinov"", ""D. Roth.""], ""title"": ""Design challenges and misconceptions in named entity recognition"", ""venue"": ""Proceedings of the Thirteenth Conference on Computational Natural Language Learning, pages 147\u2013155,"", ""year"": 2009}, {""authors"": [""V. Rus"", ""Z. Cai"", ""A.C. Graesser.""], ""title"": ""Experiments on generating questions about facts"", ""venue"": ""Proceedings of the 8th International Conference on Computational Linguistics and Intelligent Text Processing,"", ""year"": 2007}, {""authors"": [""V. Rus"", ""A.C. Graesser.""], ""title"": ""The question generation shared task and evaluation challenge"", ""venue"": ""Workshop on the Question Generation Shared Task and Evaluation Challenge, Final Report,"", ""year"": 2009}, {""authors"": [""R.F. Simmons""], ""title"": ""Answering English questions by computer: A survey"", ""venue"": ""Communications of the ACM, 8(1):53\u201370."", ""year"": 1965}, {""authors"": [""T. Strzalkowski"", ""S. Harabagiu""], ""title"": ""Advances in Open Domain Question Answering"", ""year"": 2008}, {""authors"": [""L. Vanderwende""], ""title"": ""The importance of being important: Question generation"", ""venue"": ""Proceedings of the Workshop on the Question Generation Shared Task and Evaluation Challenge, Arlington, VA."", ""year"": 2008}, {""authors"": [""E.M. Voorhees""], ""title"": ""Overview of the TREC 1999 question answering track"", ""venue"": ""Proceedings of the 8th Text REtreival Conference, Gaithersburg, MD."", ""year"": 1999}, {""authors"": [""J. Wagner"", ""J. Foster"", ""J. van Genabith.""], ""title"": ""Judging grammaticality: Experiments in sentence classification"", ""venue"": ""CALICO Journal, 26(3):474\u2013490."", ""year"": 2009}, {""authors"": [""M. Wang""], ""title"": ""A survey of answer extraction techniques in factoid"", ""year"": 2006}, {""authors"": [""W. Wang"", ""H. Tianyong"", ""L. Wenyin.""], ""title"": ""Automatic question generation for learning evaluation in medicine"", ""venue"": ""6th International Conference on Advances in Web Based Learning,"", ""year"": 2008}, {""authors"": [""X. Wei"", ""W.B. Croft.""], ""title"": ""LDA-based document models for ad-hoc retrieval"", ""venue"": ""Proceedings of the 29th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval,"", ""year"": 2006}, {""authors"": [""A. Zhang"", ""W. Lee.""], ""title"": ""Question classification using support vector machines"", ""venue"": ""Proceedings of the Special Interest Group on Information Retrieval, pages 26\u201332, Toronto."", ""year"": 2003}, {""authors"": [""Z. Zheng"", ""X. Si"", ""E.Y. Chang"", ""X. Zhu.""], ""title"": ""K2Q: Generating natural language questions from keywords with user refinements"", ""venue"": ""Proceedings of the 5th International Joint Conference on Natural"", ""year"": 2011}] acknowledgments  :We would like to thank the anonymous reviewers for their useful comments. The research reported in this article was supported by the Natural Sciences and Engineering Research Council (NSERC) of Canada – discovery grant and the University of Lethbridge. This work was done when the second author was at the University of Lethbridge.",
,,,,,,,,"Research in automated grammatical error detection and correction has gained considerable momentum in the past few years. Although much progress has been made in this area, numerous challenges and opportunities exist for further research. For NLP researchers and students who are considering following or pursuing work in this area and for English language teaching practitioners and researchers who are interested in utilizing systems resulting from research in this area, this book by Leacock, Chodorow, Gamon, and Tetreault provides a timely, updated review of the state of the art in automatic learner error detection and correction.","[{""affiliations"": [], ""name"": ""Claudia Leacock""}, {""affiliations"": [], ""name"": ""Martin Chodorow""}, {""affiliations"": [], ""name"": ""Michael Gamon""}, {""affiliations"": [], ""name"": ""Joel Tetreault""}, {""affiliations"": [], ""name"": ""Xiaofei Lu""}]",SP:78276954020a4734870552efb95484ce65a50b09,[],,,,,,,,,,,,,,,,,,,"reviewed by xiaofei lu pennsylvania state university :Research in automated grammatical error detection and correction has gained considerable momentum in the past few years. Although much progress has been made in this area, numerous challenges and opportunities exist for further research. For NLP researchers and students who are considering following or pursuing work in this area and for English language teaching practitioners and researchers who are interested in utilizing systems resulting from research in this area, this book by Leacock, Chodorow, Gamon, and Tetreault provides a timely, updated review of the state of the art in automatic learner error detection and correction.
© 2015 Association for Computational Linguistics
The introductory chapter first highlights the growth in prominence of the field of grammatical error detection and correction since the publication of the first edition of the volume. It then orients the reader to the book by detailing the changes from the first edition, providing a working definition of grammatical error, justifying the book’s focus on English language learners, clarifying the use of the term “English language learner” to refer to learners of English as either a second or foreign language, putting forth a few claims about the relationship between NLP and Computer-Assisted Language Learning (CALL), specifying the intended audience of the book, and outlining the structure of the book.
Chapter 2 provides the historical context for automated grammatical error detection and correction. The chapter first discusses how varying degrees of tolerance of grammatical errors are achieved in computational grammars used in grammar-checking and proofreading tools. It then offers a quick overview of data-driven and hybrid approaches to error detection.
Chapter 3 summarizes the types of errors made by English language learners as reported in previous empirical and corpus-based learner error studies, touches upon the influence of L1 on learner English, and then delves into a detailed discussion of three specific problem areas for English language learners: prepositions, articles, and collocations.
Chapter 4 focuses on the evaluation of error detection systems. The chapter first introduces traditional evaluation measures (e.g., precision, recall, F-score, accuracy, and kappa) and the evaluation measures adopted in recent shared tasks on grammatical error correction. It then discusses evaluation using a corpus of correct usage, questioning the value of this approach for indicating how well the system may perform on actual learner data. The advantages and challenges of evaluating system performance on
doi:10.1162/COLI r 00211
Computational Linguistics Volume 41, Number 1
learner writing are then examined. The authors advocate the use of multiple annotators and crowdsourcing as a means to improve the reliability of manual evaluation. The chapter concludes with a discussion of how statistical significance testing of differences in system performance may be performed and provides a checklist for consistent reporting of system results.
Chapter 5 zooms in on data-driven approaches to detecting and correcting article and preposition errors. The chapter first looks at four types of information used by different systems, including lexical, syntactic, semantic, and source information. It then describes three prevailing types of data used to train statistical models of grammatical error correction, namely, well-formed text, artificial errors, and error-annotated learner corpora. Next, it discusses classification methods and n-gram statistics-related methods for error detection and correction. Finally, the chapter presents two end-to-end systems, Criterion and MSR ESL Assistant.
Chapter 6 focuses on collocation errors. Following a brief discussion of the properties of collocations, the chapter introduces a range of metrics commonly used to measure the association strength between pairs of words and reviews a number of recent collocation error detection and correction systems.
Chapter 7 moves beyond article, preposition, and collocation errors and examines rule-based and statistical methods for detecting and correcting verb-form, spelling, and punctuation errors and for identifying ungrammatical sentences. A small number of error detection systems for learners of languages other than English are also discussed.
Chapter 8 covers issues with learner error annotation, describes examples of comprehensive and targeted annotation schemes, and proposes three methods for improving the efficiency of large-scale annotation: sampling, crowdsourcing, and mining online revision logs.
Chapter 9 highlights several exciting emerging directions in the field of automated grammatical error correction. These include three recent shared tasks on grammatical error correction, the use of machine translation techniques for grammatical error correction, real-time crowdsourcing of grammatical error correction, and longitudinal studies on the efficacy of automated error correction systems for improving the writing of users of such systems.
Chapter 10 concludes the book with several suggestions for future research in the field, including annotation for evaluation, error detection for underrepresented languages, research on understudied error types, L1-specific error detection modules, collaboration with second language learning and education groups, and applications of grammatical error correction. The book also includes an Appendix that contains a list of textual learner corpora with at least some publicly accessible URLs or references.
I very strongly recommend this book to all NLP students and researchers who are interested in learning about, following, or pursuing research in automated grammatical error detection and correction. Not only does it offer a comprehensive systematic review of research in this field, but it also highlights real challenges and opportunities for fruitful future research. Practitioners and researchers in the language teaching and CALL community can also use this book to obtain a realistic understanding of the state of the art of the field. They will also most certainly welcome the volume’s repeated emphasis on the importance for the NLP community to join efforts with the language teaching and CALL community to assess the effect of automated grammatical correction systems on improving the quality of language learners’ writing.
I have just two minor quibbles with this book. First, it seems that the title may better reflect the content of the book and the goals of the research field in question with the words “and correction” added after “detection.” Second, in both Chapter 3
150
Book Reviews
and Chapter 6, native speakers’ preference of powerful computer over strong computer is cited as an example of the arbitrariness of collocations. I do not intend to delve into the debate over whether collocations are arbitrary or linguistically motivated, but I note that the preference in this particular case is not arbitrary but semantically motivated (consider the semantic difference between powerful man and strong man). In fact, this preference merely reflects the fact that we are generally more concerned about the functional power of computers than their physical build, and it would be a false positive if a system marks strong computer as an error in a sentence that talks about a computer that is well-built and not easily breakable.
This book review was edited by Pierre Isabelle.
Xiaofei Lu is Gil Watz Early Career Professor in Language and Linguistics, and Associate Professor of Applied Linguistics and Asian Studies at the Pennsylvania State University. His research interests are primarily in computational linguistics, corpus linguistics, and intelligent computerassisted language learning. Lu’s e-mail address is xxl13@psu.edu.
151",,,,,,,,,,,,,,,,,,,,,,,,,,"automated grammatical error detection for language learners, second edition :","claudia leacock1, martin chodorow2, michael gamon3, and joel tetreault4 :(1CTB McGraw-Hill; 2Hunter College and the Graduate Center, City University of New York; 3Microsoft Research; 4Yahoo! Labs)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 25), 2014, xv+154pp; paperbound, ISBN 978-1-62705-013-5",,,,"Research in automated grammatical error detection and correction has gained considerable momentum in the past few years. Although much progress has been made in this area, numerous challenges and opportunities exist for further research. For NLP researchers and students who are considering following or pursuing work in this area and for English language teaching practitioners and researchers who are interested in utilizing systems resulting from research in this area, this book by Leacock, Chodorow, Gamon, and Tetreault provides a timely, updated review of the state of the art in automatic learner error detection and correction. [{""affiliations"": [], ""name"": ""Claudia Leacock""}, {""affiliations"": [], ""name"": ""Martin Chodorow""}, {""affiliations"": [], ""name"": ""Michael Gamon""}, {""affiliations"": [], ""name"": ""Joel Tetreault""}, {""affiliations"": [], ""name"": ""Xiaofei Lu""}] SP:78276954020a4734870552efb95484ce65a50b09 [] reviewed by xiaofei lu pennsylvania state university :Research in automated grammatical error detection and correction has gained considerable momentum in the past few years. Although much progress has been made in this area, numerous challenges and opportunities exist for further research. For NLP researchers and students who are considering following or pursuing work in this area and for English language teaching practitioners and researchers who are interested in utilizing systems resulting from research in this area, this book by Leacock, Chodorow, Gamon, and Tetreault provides a timely, updated review of the state of the art in automatic learner error detection and correction.
© 2015 Association for Computational Linguistics
The introductory chapter first highlights the growth in prominence of the field of grammatical error detection and correction since the publication of the first edition of the volume. It then orients the reader to the book by detailing the changes from the first edition, providing a working definition of grammatical error, justifying the book’s focus on English language learners, clarifying the use of the term “English language learner” to refer to learners of English as either a second or foreign language, putting forth a few claims about the relationship between NLP and Computer-Assisted Language Learning (CALL), specifying the intended audience of the book, and outlining the structure of the book.
Chapter 2 provides the historical context for automated grammatical error detection and correction. The chapter first discusses how varying degrees of tolerance of grammatical errors are achieved in computational grammars used in grammar-checking and proofreading tools. It then offers a quick overview of data-driven and hybrid approaches to error detection.
Chapter 3 summarizes the types of errors made by English language learners as reported in previous empirical and corpus-based learner error studies, touches upon the influence of L1 on learner English, and then delves into a detailed discussion of three specific problem areas for English language learners: prepositions, articles, and collocations.
Chapter 4 focuses on the evaluation of error detection systems. The chapter first introduces traditional evaluation measures (e.g., precision, recall, F-score, accuracy, and kappa) and the evaluation measures adopted in recent shared tasks on grammatical error correction. It then discusses evaluation using a corpus of correct usage, questioning the value of this approach for indicating how well the system may perform on actual learner data. The advantages and challenges of evaluating system performance on
doi:10.1162/COLI r 00211
Computational Linguistics Volume 41, Number 1
learner writing are then examined. The authors advocate the use of multiple annotators and crowdsourcing as a means to improve the reliability of manual evaluation. The chapter concludes with a discussion of how statistical significance testing of differences in system performance may be performed and provides a checklist for consistent reporting of system results.
Chapter 5 zooms in on data-driven approaches to detecting and correcting article and preposition errors. The chapter first looks at four types of information used by different systems, including lexical, syntactic, semantic, and source information. It then describes three prevailing types of data used to train statistical models of grammatical error correction, namely, well-formed text, artificial errors, and error-annotated learner corpora. Next, it discusses classification methods and n-gram statistics-related methods for error detection and correction. Finally, the chapter presents two end-to-end systems, Criterion and MSR ESL Assistant.
Chapter 6 focuses on collocation errors. Following a brief discussion of the properties of collocations, the chapter introduces a range of metrics commonly used to measure the association strength between pairs of words and reviews a number of recent collocation error detection and correction systems.
Chapter 7 moves beyond article, preposition, and collocation errors and examines rule-based and statistical methods for detecting and correcting verb-form, spelling, and punctuation errors and for identifying ungrammatical sentences. A small number of error detection systems for learners of languages other than English are also discussed.
Chapter 8 covers issues with learner error annotation, describes examples of comprehensive and targeted annotation schemes, and proposes three methods for improving the efficiency of large-scale annotation: sampling, crowdsourcing, and mining online revision logs.
Chapter 9 highlights several exciting emerging directions in the field of automated grammatical error correction. These include three recent shared tasks on grammatical error correction, the use of machine translation techniques for grammatical error correction, real-time crowdsourcing of grammatical error correction, and longitudinal studies on the efficacy of automated error correction systems for improving the writing of users of such systems.
Chapter 10 concludes the book with several suggestions for future research in the field, including annotation for evaluation, error detection for underrepresented languages, research on understudied error types, L1-specific error detection modules, collaboration with second language learning and education groups, and applications of grammatical error correction. The book also includes an Appendix that contains a list of textual learner corpora with at least some publicly accessible URLs or references.
I very strongly recommend this book to all NLP students and researchers who are interested in learning about, following, or pursuing research in automated grammatical error detection and correction. Not only does it offer a comprehensive systematic review of research in this field, but it also highlights real challenges and opportunities for fruitful future research. Practitioners and researchers in the language teaching and CALL community can also use this book to obtain a realistic understanding of the state of the art of the field. They will also most certainly welcome the volume’s repeated emphasis on the importance for the NLP community to join efforts with the language teaching and CALL community to assess the effect of automated grammatical correction systems on improving the quality of language learners’ writing.
I have just two minor quibbles with this book. First, it seems that the title may better reflect the content of the book and the goals of the research field in question with the words “and correction” added after “detection.” Second, in both Chapter 3
150
Book Reviews
and Chapter 6, native speakers’ preference of powerful computer over strong computer is cited as an example of the arbitrariness of collocations. I do not intend to delve into the debate over whether collocations are arbitrary or linguistically motivated, but I note that the preference in this particular case is not arbitrary but semantically motivated (consider the semantic difference between powerful man and strong man). In fact, this preference merely reflects the fact that we are generally more concerned about the functional power of computers than their physical build, and it would be a false positive if a system marks strong computer as an error in a sentence that talks about a computer that is well-built and not easily breakable.
This book review was edited by Pierre Isabelle.
Xiaofei Lu is Gil Watz Early Career Professor in Language and Linguistics, and Associate Professor of Applied Linguistics and Asian Studies at the Pennsylvania State University. His research interests are primarily in computational linguistics, corpus linguistics, and intelligent computerassisted language learning. Lu’s e-mail address is xxl13@psu.edu.
151 automated grammatical error detection for language learners, second edition : claudia leacock1, martin chodorow2, michael gamon3, and joel tetreault4 :(1CTB McGraw-Hill; 2Hunter College and the Graduate Center, City University of New York; 3Microsoft Research; 4Yahoo! Labs)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 25), 2014, xv+154pp; paperbound, ISBN 978-1-62705-013-5",
"1 introduction :Sentiment analysis (Pang and Lee 2008; Liu 2012) has received much attention from both research and industry communities in recent years. Sentiment classification, which identifies sentiment polarity (positive or negative) from text (sentence or document), has been the most extensively studied task in sentiment analysis. Until now, there have been two mainstream approaches for sentiment classification. The lexicon-based approach (Turney 2002; Taboada et al. 2011) aims to aggregate the sentiment polarity of a sentence from the polarity of words or phrases found in the sentence, and the learning-based approach (Pang, Lee, and Vaithyanathan 2002) treats sentiment polarity identification as a special text classification task and focuses on building classifiers from a set of sentences (or documents) annotated with their corresponding sentiment polarity.
The lexicon-based sentiment classification approach is simple and interpretable, but suffers from scalability and is inevitably limited by sentiment lexicons that are commonly created manually by experts. It has been widely recognized that sentiment expressions are colloquial and evolve over time very frequently. Taking tweets from Twitter1 and movie reviews on IMDb2 as examples, people use very casual language as well as informal and new vocabulary to comment on general topics and movies. In practice, it is not feasible to create and maintain sentiment lexicons to capture sentiment expressions with high coverage. On the other hand, the learning-based approach relies on large annotated samples to overcome the vocabulary coverage and deals with variations of words in sentences. Human ratings in reviews (Maas et al. 2011) and emoticons in tweets (Davidov, Tsur, and Rappoport 2010; Zhao et al. 2012) are extensively used to collect a large number of training corpora to train the sentiment classifier. However, it is usually not easy to design effective features to build the classifier. Among others, unigrams have been reported as the most effective features (Pang, Lee, and Vaithyanathan 2002) in sentiment classification.
Handling complicated expressions delivering people’s opinions is one of the most challenging problems in sentiment analysis. Compositionalities such as negation, intensification, contrast, and their combinations are typical cases. We show some concrete examples here:
(1) The movie is not good. [negation]
(2) The movie is very good. [intensification]
(3) The movie is not funny at all. [negation + intensification]
(4) The movie is just so so, but i still like it. [contrast]
(5) The movie is not very good, but i still like it. [negation + intensification + contrast]
The negation expressions, intensification modifiers, and the contrastive conjunction can change the polarity (Examples (1), (3), (4), (5)), strength (Examples (2), (3), (5)), or both (Examples (3), (5)) of the sentiment of the sentences. We do not need any detailed explanations here as they can be commonly found and easily understood in people’s
1 http://twitter.com. 2 http://www.imdb.com.
daily lives. Existing works to address these issues usually rely on syntactic parsing results either used as features (Choi and Cardie 2008; Moilanen, Pulman, and Zhang 2010) in learning-based methods or hand-crafted rules (Moilanen and Pulman 2007; Jia, Yu, and Meng 2009; Klenner, Petrakis, and Fahrni 2009; Liu and Seneff 2009) in lexiconbased methods. However, even with the difficulty and feasibility of deriving the sentiment structure from syntactic parsing results put aside, it is an even more challenging task to generate stable and reliable parsing results for text that is ungrammatical in nature and has a high ratio of out-of-vocabulary words. The accuracy of the linguistic parsers trained on standard data sets (e.g., the Penn Treebank [Marcus, Marcinkiewicz, and Santorini 1993]) drops dramatically on user-generated-content (reviews, tweets, etc.), which is actually the prime focus of sentiment analysis algorithms. The error, unfortunately, will propagate downstream in the process of sentiment analysis methods building upon parsing results.
We therefore propose directly analyzing the sentiment structure of a sentence. The nested structure of sentiment expressions can be naturally modeled in a similar fashion as statistical syntactic parsing, which aims to find the linguistic structure of a sentence. This idea creates many opportunities for developing sentiment classifiers from a new perspective. The most challenging problem and barrier in building a statistical sentiment parser lies in the acquisition of training data. Ideally, we need examples of sentences annotated with polarity for the whole sentence as well as sentiment tags for constituents within a sentence, as with the Penn TreeBank for training traditional linguistic parsers. However, this is not practical as the annotations will be inevitably time-consuming and require laborious human efforts. Therefore, it is better to learn the sentiment parser only utilizing examples annotated with the polarity label of the whole sentence. For example, we can collect a huge number of publicly available reviews and rating scores on the Web. People may use the movie is gud (“gud” is a popular informal expression of “good”) to express a positive opinion towards a movie, and not a fan to express a negative opinion. Also, we can find review sentences such as The movie is gud, but I am still not a fan to indicate a negative opinion. We can then use these two fragments and the overall negative opinion of the sentence to deduce sentiment rules automatically from data. These sentiment fragments and rules can be used to analyze the sentiment structure for new sentences.
In this article, we propose a statistical parsing framework to directly analyze the structure of a sentence from the perspective of sentiment analysis. Specifically, we formulate a Context-Free Grammar (CFG)–based sentiment grammar. We then develop a statistical parser to derive the sentiment structure of a sentence. We leverage the CYK algorithm (Cocke 1969; Younger 1967; Kasami 1965) to conduct bottom–up parsing, and use dynamic programming to accelerate computation. Meanwhile, we propose using the polarity model to derive sentiment strength and polarity of a sentiment parse tree, and the ranking model to select the best one from the sentiment parsing results. We train the parser directly from examples of sentences annotated with sentiment polarity labels instead of syntactic annotations and polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, named s.parser, from a large number of review sentences with users’ ratings as rough sentiment polarity labels. The statistical parsing–based approach builds a principled and scalable framework to support the sentiment composition and inference which cannot be well handled by bag-of-words approaches. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way.
The major contributions of the work presented in this article are as follows.r We propose a statistical parsing framework for sentiment analysis that is capable of analyzing the sentiment structure for a sentence. This framework can naturally handle compositionality in a probabilistic way. It can be trained from sentences annotated with only sentiment polarity but without any syntactic annotations or polarity annotations of constituents within sentences.r We present the parsing model, polarity model, and ranking model in the proposed framework, which are formulated and can be improved independently. It provides a principled and flexible approach to sentiment classification.r We implement the statistical sentiment parsing framework, and conduct experiments on several benchmark data sets. The experimental results show that the proposed framework and algorithm can significantly outperform baseline methods.
The remainder of this article is organized as follows. We introduce related work in Section 2. We present the statistical sentiment parsing framework, including the parsing model, polarity model, and ranking model, in Section 3. Learning methods for our model are explained in Section 4. Experimental results are reported in Section 5. We conclude this article with future work in Section 6.","2 related work :In this section, we give a brief introduction to related work about sentiment classification (Section 2.1) and parsing (Section 2.2). We tackle the sentiment classification problem in a parsing manner, which is a significant departure from most previous research. Sentiment classification has been extensively studied in the past few years. In terms of text granularity, existing works can be divided into phrase-level, sentence-level, or document-level sentiment classification. We focus on sentence-level sentiment classification in this article. Regardless of what granularity the task is performed on, existing approaches deriving sentiment polarity from text fall into two major categories, namely, lexicon-based and learning-based approaches.
The lexicon-based sentiment analysis uses dictionary matching on a predefined sentiment lexicon to derive sentiment polarity. These methods often use a set of manually defined rules to deal with the negation of polarity. Turney (2002) proposed using the average sentiment orientation of phrases, which contains adjectives or adverbs, in a review to predict its sentiment orientation. Yu and Hatzivassiloglou (2003) calculated a modified log-likelihood ratio for every word by the co-occurrences with positive and negative seed words. To determine the polarity of a sentence, they compare the average log-likelihood value with threshold. Taboada et al. (2011) presented a lexicon-based approach for extracting sentiment from text. They used dictionaries of words with annotated sentiment orientation (polarity and strength) while incorporating intensification and negation. The lexicon-based methods often achieve high precisions and do not need
any labeled samples. But they suffer from coverage and domain adaption problems. Moreover, lexicons are often built and used without considering the context (Wilson, Wiebe, and Hoffmann 2009). Also, hand-crafted rules are often matched heuristically.
The sentiment dictionaries used for lexicon-based sentiment analysis can be created manually, or automatically using seed words to expand the list of words. Kamps et al. (2004) and Williams and Anand (2009) used various lexical relations (such as synonym and antonym relations) in WordNet to expend a set of seed words. Some other methods learn lexicons from data directly. Hatzivassiloglou and McKeown (1997) used a log-linear regression model with conjunction constraints to predict whether conjoined adjectives have similar or different polarities. Combining conjunction constraints across many adjectives, a clustering algorithm separated the adjectives into groups of different polarity. Finally, adjectives were labeled as positive or negative. Velikovich et al. (2010) constructed a term similarity graph using the cosine similarity of context vectors. They performed graph propagation from seeds on the graph, obtaining polarity words and phrases. Takamura, Inui, and Okumura (2005) regarded the polarity of words as spins of electrons, using the mean field approximation to compute the approximate probability function of the system instead of the intractable actual probability function. Kanayama and Nasukawa (2006) used tendencies for similar polarities to appear successively in contexts. They defined density and precision of coherency to filter neutral phrases and uncertain candidates. Choi and Cardie (2009a) and Lu et al. (2011) transformed the lexicon learning to an optimization problem, and used integer linear programming to solve it. Kaji and Kitsuregawa (2007) defined the χ2-based polarity value and PMI-based polarity value as a polarity strength to filter neutral phrases. de Marneffe, Manning, and Potts (2010) utilized review data to define polarity strength as the expected rating value. Mudinas, Zhang, and Levene (2012) used word count as a feature template and trained a classifier using Support Vector Machines with linear kernel. They then regarded the weights as polarity strengths. Krestel and Siersdorfer (2013) generated topicdependent lexicons from review articles by incorporating topic and rating probabilities and defined the polarity strength based on the results. In this article, the lexical relations defined in WordNet are not used because of its coverage. Furthermore, most of these methods define different criteria to propagate polarity information of seeds, or use optimization algorithms and sentence-level sentiment labels to learn polarity strength values. Their goal is to balance the precision and recall of learned lexicons. We also learn the polarity strength values of phrases from data. However, our primary objective is to obtain correct sentence-level polarity labels, and use them to form the sentiment grammar.
Learning-based sentiment analysis uses machine learning methods to classify sentences or documents into two (negative and positive) or three (negative, positive, and neutral) classes. Previous research has shown that sentiment classification is more difficult than traditional topic-based text classification, despite the fact that the number of classes in sentiment classification is smaller than that in topic-based text classification (Pang and Lee 2008). Pang, Lee, and Vaithyanathan (2002) investigated three machine learning methods to produce automated classifiers to generate class labels for movie reviews. They tested them on Naı̈ve Bayes, Maximum Entropy, and Support Vector Machine (SVM), and evaluated the contribution of different features including unigrams, bigrams, adjectives, and part-of-speech tags. Their experimental results suggested that a SVM classifier with unigram presence features outperforms other competitors. Pang and Lee (2004) separated subjective portions from the objective by finding minimum cuts in graphs to achieve better sentiment classification performance. Matsumoto, Takamura, and Okumura (2005) used text mining techniques to
extract frequent subsequences and dependency subtrees, and used them as features of SVM. McDonald et al. (2007) investigated a global structured model for jointly classifying polarity at different levels of granularity. This model allowed classification decisions from one level in the text to influence decisions at another. Yessenalina, Yue, and Cardie (2010) used sentence-level latent variables to improve document-level prediction. Täckström and McDonald (2011a) presented a latent variable model for only using document-level annotations to learn sentence-level sentiment labels, and Täckström and McDonald (2011b) improved it by using a semi-supervised latent variable model to utilize manually crafted sentence labels. Agarwal et al. (2011) and Tu et al. (2012) explored part-of-speech tag features and tree-kernel. Wang and Manning (2012) used SVM built over Naı̈ve Bayes log-count ratios as feature values to classify polarity. They showed that SVM was better at full-length reviews, and Multinomial Naı̈ve Bayes was better at short-length reviews. Liu, Agam, and Grossman (2012) proposed a set of heuristic rules based on dependency structure to detect negations and sentimentbearing expressions. Most of these methods are built on bag-of-words features, and sentiment compositions are handled by manually crafted rules. In contrast to these models, we derive polarity labels from tree structures parsed by the sentiment grammar.
There have been several attempts to assume that the problem of sentiment analysis is compositional. Sentiment classification can be solved by deriving the sentiment of a complex constituent (sentence) from the sentiment of small units (words and phrases) (Moilanen and Pulman 2007; Klenner, Petrakis, and Fahrni 2009; Choi and Cardie 2010; Nakagawa, Inui, and Kurohashi 2010). Moilanen and Pulman (2007) proposed using delicate written linguistic patterns as heuristic decision rules when computing the sentiment from individual words to phrases and finally to the sentence. The manually compiled rules were powerful enough to discriminate between the different sentiments in effective remedies (positive) / effective torture (negative), and in too colorful (negative) and too sad (negative). Nakagawa, Inui, and Kurohashi (2010) leveraged a conditional random field model to calculate the sentiment of all the parsed elements in the dependency tree and then generated the overall sentiment. It had an advantage over the rule-based approach (Moilanen and Pulman 2007) in that it did not explicitly denote any sentiment designation to words or phrases in parse trees. Instead, it modeled their sentiment polarity as latent variables with a certain probability of being positive or negative. Councill, McDonald, and Velikovich (2010) used a conditional random field model informed by a dependency parser to detect the scope of negation for sentiment analysis. Some other methods model sentiment compositionality in the vector space. They regard the composition operator as a matrix, and use matrix-vector multiplication to obtain the transformed vector representation. Socher et al. (2012) proposed a recursive neural network model that learned compositional vector representations for phrases and sentences. Their model assigned a vector and a matrix to every node in a parse tree. The vector captured the inherent meaning of the constituent, and the matrix captured how it changes the meaning of neighboring words or phrases. Socher et al. (2013) recently introduced a sentiment treebank based on the results of the Stanford parser (Klein and Manning 2003). The sentiment treebank included polarity labels of phrases that are annotated using Amazon Mechanical Turk. The authors trained recursive neural tensor networks on the sentiment treebank. For a new sentence, the model predicted polarity labels based on the syntactic parse tree, and used tensors to handle compositionality in the vector space. Dong et al. (2014) proposed utilizing multiple composition functions in recursive neural models and learning to select them adaptively. Most previous methods are either rigid in terms of handcrafted rules, or sensitive to the performance of existing syntactic parsers they use. This article addresses sentiment
compositions by defining sentiment grammar and borrowing some techniques in the parsing research field. Moreover, our method uses symbolic representations instead of vector spaces. The work presented in this article is close to traditional statistical parsing, as we borrow some algorithms to build the sentiment parser. Syntactic parsers are learned from the Treebank corpora, and find the most likely parse tree with the largest probability. In this article, we borrow some well-known techniques from syntactic parsing methods (Charniak 1997; Charniak and Johnson 2005; McDonald, Crammer, and Pereira 2005; Kübler, McDonald, and Nivre 2009), such as the CYK algorithm and Context-Free Grammar. These techniques are used to build the sentiment grammar and parsing model. They provide a natural way of defining the structure of sentiment trees and parse sentences to trees. The key difference lies in that our task is to calculate the polarity label of a sentence, instead of obtaining the parse tree. We only have sentencepolarity pairs as our training instances instead of annotated tree structures. Moreover, in the decoding process, our goal is to compute correct polarity labels by representing sentences as latent sentiment trees. Recently, Hall, Durrett, and Klein (2014) developed a discriminative constituency parser using rich surface features, adapting it to sentiment analysis. Besides extracting unigrams and bigrams as features, they learned interactions between tags and words located at the beginning or the end of spans. However, their method relies on phrase-level polarity annotations.
Semantic parsing is another body of work related to this article. A semantic parser is used to parse meaning representations for given sentences. Most existing semantic parsing works (Zelle and Mooney 1996; Kate and Mooney 2006; Raymond and Mooney 2006; Zettlemoyer and Collins 2007, 2009; Li, Liu, and Sun 2013) relied on fine-grained annotations of target logical forms, which required the supervision of experts and are relatively expensive. To balance the performance and the amount of human annotation, some works used only question-answer pairs or even binary correct/incorrect signals as their input. Clarke et al. (2010) used a binary correct/incorrect signal of a database query to map sentences to logical forms. It worked with FunQL language and transformed semantic parsing as an integer linear programming (ILP) problem. In each iteration, it solved ILP and updated the parameters of structural SVM. Liang, Jordan, and Klein (2013) learned a semantic parser from question-answer pairs, where the logical form was modeled as a latent tree-based semantic representation. Krishnamurthy and Mitchell (2012) presented a method for training a semantic parser using a knowledge base and an unlabeled text corpus, without any individually annotated sentences. Artzi and Zettlemoyer (2013) used various types of weak supervision to learn a grounded Combinatory Categorial Grammar semantic parser, which took context into consideration. Bao et al. (2014) presented a translation-based weakly supervised semantic parsing method to translate questions to answers based on CYK parsing. A log-linear model is defined to score derivations. All these weakly supervised semantic parsing methods learned to transform a natural language sentence to its semantic representation without annotated logical form. In this work, we build a sentiment parser. Specifically, we use a modified version of the CYK algorithm that parses sentences in a bottom–up fashion. We use the log-linear model to score candidates generated by beam search. Instead of using question-answer pairs, sentence-polarity pairs are used as our weak supervisions. We also use the parameter estimation algorithm proposed by Liang, Jordan, and Klein (2013).","3 statistical sentiment parsing :We present the statistical parsing framework for sentence-level sentiment classification in this section. The underlying idea is to model sentiment classification as a statistical parsing process. Figure 1 shows the overview of the statistical sentiment parsing framework. There are three major components. The input sentence s is transformed into and represented by sentiment trees derived from the parsing model (Section 3.2), using the sentiment grammar defined in Section 3.1. Trees are scored by the ranking model in Section 3.3. The sentiment tree with the highest ranking score is treated as the best derivation for s. Furthermore, the polarity model (Section 3.4) is used to compute polarity values for the sentiment trees.
Notably, the sentiment trees t are unobserved during training. We can only observe the sentence s and its polarity label y in training data. In other words, we train the model directly from the examples of sentences annotated only with sentiment polarity labels but without any syntactic annotations or polarity annotations of the constituents within sentences. To be specific, we first learn the sentiment grammar and the polarity model from data as described in Section 4.2. Then, given the sentence and polarity label pairs( s, y ) , we search the latent sentiment trees t and estimate the parameters of the ranking model as detailed in Section 4.1. To better illustrate the whole process, we describe the sentiment parsing procedure using an example sentence, The movie is not very good, but i still like it. The sentiment polarity label of the above sentence is “positive.” There is negation, intensification, and contrast in this example, which are difficult to capture using bag-of-words classification methods. This sentence is a complex case that demonstrates the capability of the proposed statistical sentiment parsing framework, which motivates the work in this article. The statistical sentiment parsing algorithm may generate a number of sentiment trees for the input sentence. Figure 2 shows the best sentiment parse tree. It shows that the statistical sentiment parsing framework can deal with the compositionality of sentiment in a natural way. In Table 1, we list the sentiment rules used during the parsing process. We show the generation process of the sentiment parse tree from the bottom–up and the calculation of sentiment strength and polarity for every text span in the parsing process.
the derivation process include {P→ the movie is; P→ good; P→ i still like it; P→ very P; N→ not P; N→ PN; N→ NE ; E → ,; P→ N but P; S→ P}.
In the following sections, we first provide a formal description of the sentiment grammar in Section 3.1. We then present the details of the parsing model in Section 3.2, the ranking model in Section 3.3, and the polarity model in Section 3.4. We develop the sentiment grammar upon CFG (Context-Free Grammar) (Chomsky 1956). Let G =< V, Σ, S, R > denote a CFG, where V is a finite set of non-terminals, Σ is a finite set of terminals (disjointed from V), S ∈ V is the start symbol, and R is a set of rewrite rules (or production rules) of the form A→ c where A ∈ V and c ∈ (V ∪ Σ)∗. We use Gs =< Vs, Σs, S, Rs > to denote the sentiment grammar in this article.
The non-terminal set is denoted as Vs = {N, P, S, E}, where S is the start symbol, the non-terminal N represents the negative polarity, and the non-terminal P represents the positive polarity. The rules in Rs are divided into the following six categories:
r Dictionary rules: X→ wk0, where X ∈ {N, P}, wk0 = w0 . . .wk−1, and wk0 ∈ Σ+s . These rules can be regarded as the sentiment dictionary used in traditional approaches. They are basic sentiment units assigned with polarity probabilities. For instance, P→ good is a dictionary rule.r Combination rules: X→ c, where c ∈ (Vs ∪ Σs)+, and two successive non-terminals are not allowed. There is at least one terminal in c. These rules combine terminals and non-terminals, such as N→ not P, and P→ N but P. They are used to handle negation, intensification, and contrast in sentiment analysis. The number of non-terminals in a combination rule is restricted to one and two.r Glue rules: X→ X1X2, where X, X1, X2 ∈ {N, P}. These rules combine two text spans that are derived into X1 and X2, respectively.r OOV rules: E → wk0, where wk0 ∈ Σ+. We use these rules to handle Out-Of-Vocabulary (OOV) text spans whose polarity probabilities are not learned from data.r Auxiliary rules: X→ EX1, X→ X1E , where X, X1 ∈ {N, P}. These rules combine a text span with polarity and an OOV text span.r Start rules: S→ Y, where Y ∈ {N, P, E}. The derivations begin with S, and S can be derived to N, P, and E .
Here, X represents the non-terminals N or P. The dictionary rules and combinations rules are automatically extracted from the data. We will describe the details in Section 4.2. By applying these rules, we can derive the polarity label of a sentence from the bottom–up. The glue rules are used to combine polarity information of two text spans together, and it treats the combined parts as independent. In order to tackle the OOV problem, we treat a text span that consists of OOV words as empty text span, and derive them to E . The OOV text spans are combined with other text spans without considering their sentiment information. Finally, each sentence is derived to the symbol S using the start rules that are the beginnings of derivations. We can use the sentiment grammar to compactly describe the derivation process of a sentence. We present the formal description of the statistical sentiment parsing model following deductive proof systems (Shieber, Schabes, and Pereira 1995; Goodman 1999) as used in traditional syntactic parsing. For a concrete example,
(A→ BC) [i, B, k] [k, C, j] [i, A, j] (6)
which represents if we have the rule A→ BC and B ∗⇒ wki and C ∗⇒ w jk ( ∗⇒ is used to represent the reflexive and transitive closure of immediate derivation), then we can obtain A ∗⇒ w ji . By adding a unary rule
(A→ w ji) [i, A, j] (7)
with the binary rule in Equation (6), we can express the standard CYK algorithm for CFG in Chomsky Normal Form (CNF). And the goal is [0, S, n], in which S is the start symbol and n is the length of the input sentence. In the given CYK example, the term in deductive rules can be one of the following two forms:r [i, X, j] is an item representing a subtree rooted in X spanning from i to j, orr (X→ γ) is a rule in the grammar.
Generally, we represent the form of an inference rule as:
(r) H1 . . . HK [i, X, j] (8)
where, if all the terms r and Hk are true, then we can infer [i, X, j] as true. Here, r denotes a sentiment rule, and Hk denotes an item. When we refer to both rules and items, we employ the word terms.
Theoretically, we can convert the sentiment rules to CNF versions, and then use the CYK algorithm to conduct parsing. Because the maximum number of non-terminal symbols in a rule is already restricted to two, we formulate the statistical sentiment parsing based on a customized CYK algorithm that is similar to the work of Chiang (2007). Let X, X1, X2 represent the non-terminals N or P; the inference rules for the statistical sentiment parsing are summarized in Figure 3. The parsing model generates many candidate parse trees T(s) for a sentence s. The goal of the ranking model is to score and rank these parse trees. The sentiment tree with the highest score is treated as the best representation for sentence s. We extract a feature vector φ(s, t) ∈ Rd for the specific sentence-tree pair (s, t), where t ∈ T(s) is the parse tree. Let ψ ∈ Rd be the parameter vector for the features. We use the log-linear model to calculate a probability p(t|s; T,ψ) for each parse tree t ∈ T(s). The probabilities indicate how likely the trees are to produce correct predictions. Given the sentence s and parameters ψ, the log-linear model defines a conditional probability:
p(t|s; T,ψ) = exp {φ(s, t)Tψ−A(ψ; s, T)} (9)
A(ψ; s, T) = log ∑
t∈T(s) exp {φ(s, t)Tψ} (10)
(X→ w ji) [i, X, j]
(X→ wi1i X1w j j1
) [i1, X1, j1] [i, X, j]
(X→ wi1i X1w i2 j1 X2w j j2 ) [i1, X1, j1] [i2, X2, j2]
[i, X, j]
(X→ X1X2) [i, X1, k] [k, X2, j] [i, X, j]
(E → w ji ) [i, E , j]
(X→ EX1) [i, E , k] [k, X1, j] [i, X, j]
(X→ X1E ) [i, X1, k] [k, E , j] [i, X, j]
where X, X1, X2 represent N or P.
Figure 3 Inference rules for the basic parsing model.
where A(ψ; s, T) is the log-partition function with respect to T(s). The log-linear model is a discriminative model, and it is widely used in natural language processing. We can use φ(s, t)Tψ as the score of the parse tree without normalization in the decoding process, because p(t|s; T,ψ) ∝ φ(s, t)Tψ, and this will not change the ranking order. The goal of the polarity model is to model the calculation of sentiment strength and polarity of a text span from its subspans in the parsing process. It is specified in terms of the rules used in the parsing process. We expand the notations in the inference rule (8) to incorporate the polarity model. The new form of inference rule is:
(r) H1Φ1 . . . HKΦK [i, X, j]Φ (11)
in which r, H1, . . . , HK are the terms described in Section 3.2. Every item Hk is assigned
polarity strength Φk : { P(N |w jkik ) P(P|w jkik ) for text span w jkik . For the item [i, X, j], the polarity model Φ(r, Φ1, . . . , ΦK ) is defined as a function that takes the rule r and polarity strength of subspans as input.
The polarity strength obtained by the polarity model should satisfy two constraints. First, the values calculated by the polarity model are non-negative, that is, P(X |w ji) ≥ 0, P(X |w j i) ≥ 0. Second, the positive and negative polarity values are normal-
ized to 1, namely, P(X |w ji) + P(X |w j i) = 1. Notably, X = { P , X = N N , X = P is the opposite
polarity of X . The inference rules with the polarity model are formally defined in Figure 4. In the following part, we define the polarity model for the different types of rules. If the rule is a dictionary rule X→ w ji, its sentiment strength is obtained as:
Φ :
{ P(X |w ji) = P̃(X |w j i)
P(X |w ji) = P̃(X |w j i)
(12)
where X ∈ {N ,P} denotes the sentiment polarity of the left hand side of the rule, X is the opposite polarity of X , and P̃(X |w ji), P̃(X |w j i) indicate the sentiment polarity values estimated from training data.
The glue rules X→ X1X2 combine two spans (w ki , w j k). The polarity value is calcu-
lated by their product, and normalized to 1.
Φ : P(X |w j i) =
P(X |w ki )P(X |w j k )
P(X |w ki )P(X |w j k )+P(X |w k i )P(X |w j k )
P(X |w ji) = 1− P(X |w j i)
(13)
For OOV text spans, the polarity model does not calculate the polarity values. When they are combined with in-vocabulary phrases by the auxiliary rules, the polarity values are determined by the text span with polarity and the OOV text span is ignored. More specifically,
Φ :
{ P(X |w ji) = P(X |w k i )
P(X |w ji) = P(X |w k i )
(14)
The combination rules are more complicated than other types of rules. In this article, we model the polarity probability calculation as the logistic regression. The logistic regression can be regarded as putting linear combination of the subspans’ polarity probabilities into a logistic function (or sigmoid function). We will show that the negation, intensification, and contrast can be well modeled by the regression-based method. It is formally shown as
P(X |w ji) = h ( θ0 + K∑ k=1 θkP(Xk|w jk ik ) )
= 1 1 + exp { − ( θ0 + ∑K k=1 θkP(Xk|w jk ik ) )} (15)
where h(x) = 11+exp{−x} is the logistic function, K is the number of non-terminals in a rule, and θ0, . . . ,θK are the parameters that are learned from data. As a concrete example, if the span w ji can match N→ not P and P
∗⇒ w ji+1, the inference rule with the polarity model is defined as
N→ not P [i + 1, P, j]Φ1
[i, N, j]
{ P(N |w ji) = h(θ0 + θ1P(P|w j i+1))
P(P|w ji) = 1− P(N |w j i)
(16)
where polarity probability is calculated by P(N |w ji) = h(θ0 + θ1P(P|w j i+1)).
To tackle negation, switch negation (Choi and Cardie 2008; Saurı́ 2008) simply reverses the sentiment polarity and corresponding sentiment strength. However, consider not great and not good; flipping polarity directly makes not good more positive than not great, which is unreasonable. Another potential problem of switch negation is that negative polarity items interact with intensifiers in undesirable ways (Kennedy and Inkpen 2006). For example, not very good turns out to be even more negative than not good, given the fact that very good is more positive than good. Therefore, Taboada et al. (2011) argue that shift negation is a better way to handle polarity negation. Instead of reversing polarity strength, shift negation shifts it toward the opposite polarity by
a fixed amount. This method can partially avoid the aforementioned two problems. However, they set the parameters manually, which might not be reliable and extensible enough to a new data set. Using the regression model, switch negation is captured by the negative scale item θk (k > 0), and shift negation is expressed by the shift item θ0.
The intensifiers are adjectives or adverbs that strengthen (amplifier) or decrease (downtoner) the semantic intensity of its neighboring item (Quirk 1985). For example, extremely good should obtain higher strength of positive polarity than good, because it is modified by the amplifier (extremely). Polanyi and Zaenen (2006) and Kennedy and Inkpen (2006) handle intensifiers by polarity addition and subtraction. This method, termed fixed intensification, increases a fixed amount of polarity for amplifiers and decreases for downtoners. Taboada et al. (2011) propose a method, called percentage intensification, to associate each intensification word with a percentage scale, which is larger than one for amplifiers, and less than one for downtoners. The regression model can capture these two methods to handle the intensification. The shift item θ0 represents the polarity addition and subtraction directly, and the scale item θk (k > 0) can scale the polarity by a percentage.
Table 2 illustrates how the regression based polarity model represents different negation and intensification methods. For a specific rule, the parameters and the compositional method are automatically learned from data (Section 4.2.3) instead of setting them manually as in previous work (Taboada et al. 2011). In a similar way, this method can handle the contrast. For example, the inference rule for N→ P but N is:
(N→ P but N) [i1, P, j1]Φ1 [i2, N, j2]Φ2
[i, N, j]
{ P(N |w ji) = h(θ0 + θ1P(P|w j1 i1 ) + θ2P(N |w j2 i2 ))
P(P|w ji) = 1− P(N |w j i)
(17)
where the polarity probability of the rule N→ P but N is computed by P(N |w ji) = h(θ0 + θ1P(P|w j1 i1 ) + θ2P(N |w j2 i2
)). It can express the contrast relation by specific parameters θ0, θ1, and θ2.
It should be noted that a linear regression model could turn out to be problematic, as it may produce unreasonable results. For example, if we do not add any constraint, we may get P(N |w ji) = −0.6 + P(P|w j i+1). When P(P|w j i+1) = 0.55, we will get P(N |w ji) = −0.6 + 0.55 = −0.05. This conflicts with the definition that the polarity probability ranges from zero to one. Figure 5 intuitively shows that the logistic function truncates polarity values to (0, 1) smoothly.
Parameter Negation Type P(X |w ji) = h(θ0 + θ1P(X |w j1 i1 )) Intensification Type P(X |w ji) = h(θ0 + θ1P(X |w j1 i1 ))
Shift Switch Percentage Fixed
θ0 (Shift item) X X θ1 (Scale item) X X We incorporate additional constraints into the parsing model. Those are used as pruning conditions in the derivation process not only to improve efficiency but also to force the derivation towards the correct direction. We expand the inference rules in Section 3.4 as,
(r) H1Φ1 . . . HKΦK [i, X, j]Φ C (18)
where C is a side condition. The constraints are interpreted in a Boolean manner. If the constraint C is satisfied, the rule can be used, otherwise, it cannot. We define two constraints in the parsing model.
First, in the parsing process, the polarity label of text span w ji obtained by the polarity model (Section 3.4) should be consistent with the non-terminal X (N or P) on the left hand side of the rule. To distinguish between the polarity labels and the non-terminals, we denote the corresponding polarity label of non-terminal X as X . Following this notation, we describe the first constraint as
C1 : P(X |w j i) > P(X |w j i) (19)
where X is the opposite polarity of X . For instance, if rule P→ not N matches the text span w ji, the polarity calculated by the polarity model should be consistent with P, i.e., the polarity obtained by the polarity model should be positive (P).
Second, when we apply the combination rules, the polarity strength of subspans needs to exceed a predefined threshold τ (≥ 0.5). Specifically, for combination rules X→ wi1i X1w i2 j1 X2w j j2 and X→ wi1i X1w j j1 , we define the second constraint as
C2 : P(Xk|w jk ik ) > τ, k = 1, . . . , K (20)
where K is the number of subspans in the rule, and Xk is the corresponding polarity label of non-terminal Xk in the right hand side. If P(Xk|w
jk ik ) is not larger than threshold
τ, we regard the polarity of phrase w jkik as neutral. For instance, we do not want to use the combination rule P→ a lot of P or N→ a lot of N for the phrase a lot of people. This constraint avoids improperly using the combination rules for neutral phrases. Notably, when τ is set as 0.5, this constraint is the same as the first one in Equation (19).
As shown in Figure 6, we add these two constraints to the inference rules. The OOV rules do not have any constraints, and the constraint C1 is applied for all the other rules. The constraint C2 is only applied for the combination rules. In this section, we summarize the decoding algorithm in Algorithm 1. For a sentence s, the CYK algorithm and dynamic programming are used to obtain the sentiment tree with the highest score. To be specific, the modified CYK parsing model parses the input sentence to sentiment trees in a bottom–up manner—that is, from short to long text spans. For every text span w ji, we match the rules in the sentiment grammar (Section 3.1) to generate the candidate set. Their polarity values are calculated using the polarity model described in Section 3.4. We also use the constraints described in Section 3.5 to prune search paths. The constraints improve the efficiency of the parsing algorithm and make derivations that meet our intuition.
Algorithm 1 Decoding Algorithm Input: wn0 : Sentence Output: Polarity of the input sentence
1: score[, , ]← {} 2: for l← 1 . . .n do . Modified CYK algorithm 3: for all i, j s.t. j− i = l do 4: for all inferable rule (r) H1...HK[i,X,j] for w j i do 5: Φ← calculate polarity value for r . Polarity model 6: if constraints are satisfied then . Constraint 7: sc← compute score for this derivation by ranking model . Ranking
model 8: if sc > score[i, j, X] then 9: score[i, j, X]← sc
10: return arg maxX∈{N ,P} score[0, X, n]
The features in the ranking model (Section 4.1.1) decompose along the structure of the sentiment tree. So the dynamic programming technique can be used to compute the derivation tree with the highest ranking score. For a span, the scores of its subspans are used to calculate the local scores of its derivations. For example, the score of the derivation (r) [i1,P,j1] [i2,N,j2][i,X,j] is score[i1, j1, P] + score[i2, j2, N] + score r, where score[i, j, X] is the highest score of text span w ji that is derived to the non-terminal X, and score r is the score of applying the rule r. As described in Section 3.3, the score of using rule r is scorer = φ(w ji, r) T ψ, where φ(w ji, r) is the feature vector of using the rule r for the span w ji, and ψ is the weight vector of the ranking model. The k highest score trees satisfying the constraints are stored in score[, , ] for decoding the longer text spans. After finishing the CYK parsing, arg maxX∈{N ,P} score[0, n, X] is regarded as the polarity label of input sentence. The time complexity is the same as the standard CYK’s.","4 model learning :We described the statistical sentiment parsing framework in Section 3. We present the model learning process in this section. The learning process consists of two steps. First, the sentiment grammar and the polarity model are learned from data. In other words, the rules and the parameters used to compute polarity values are learned. These basic sentiment building blocks are then used to build the parse trees. Second, we estimate the parameters of the ranking model using the sentence and polarity label pairs. At this stage, we concentrate on learning how to score the parse trees based on the learned sentiment grammar and polarity model.
Section 4.1 shows the features and the parameter estimation algorithm used in the ranking model. Section 4.2 illustrates how to learn the sentiment grammar and the polarity model. As shown in Section 3.3, we develop the ranking model on the log-linear model. In the following subsections, we first present the features used to rank sentiment tree
candidates. Then, we describe the objective function used in the optimization algorithm. Finally, we introduce the algorithm for parameter estimation using the gradient-based method.
4.1.1 Features. We extract a feature vector φ(s, t) ∈ Rd for each parse tree t of sentence s. The feature vector is used in the log-linear model. In Figure 7, we present the features extracted for the sentence The movie is not very good, but i still like it. The features are organized into feature templates. Each of them contains a set of features. These feature templates are shown as follows:r COMBHIT: This feature is the total number of combination rules used in t.r COMBRULE: It contains features {COMBRULE[r] : r is a combination rule},
each of which fires on the combination rule r appearing in t.r DICTHIT: This feature is the total number of dictionary rules used in t.r DICTRULE: It contains features {DICTRULE[r] : r is a dictionary rule}, each of which fires on the dictionary rule r appearing in t.
These features are generic local patterns that capture the properties of the sentiment tree. Another intuitive lexical feature template is [combination rule + word]. For instance, P→ very P(good) is a feature that lexicalizes the non-terminal P to good. However, if this feature is fired frequently, the phrase very good would be learned as a dictionary rule and can be used in the decoding process. So we do not use this feature template in order to reduce the feature size. It should be noted that these features
decompose along structures of sentiment trees, enabling us to use dynamic programming in the CYK algorithm.
4.1.2 Objective Function. We design the ranking model upon the log-linear model to score candidate sentiment trees. In the training dataD, we only have the input sentence s and its polarity label Ls. The forms of sentiment parse trees, which can obtain the correct sentiment polarity, are unobserved. So we work with the marginal log-likelihood of obtaining the correct polarity label Ls,
log p(Ls|s; T,ψ) = log p(t ∈ TLs (s)|s; T,ψ)
= A(ψ; s, TLs )−A(ψ; s, T) (21)
where TLs is the set of candidate trees whose prediction labels are Ls, and A(ψ; s, T) (Equation (10)) is the log-partition function with respect to T(s).
Based on the marginal log-likelihood function, the objective function O(ψ, T) consists of two terms. The first term is the sum of marginal log-likelihood over training instances that can obtain the correct polarity labels. The second term is a L2-norm regularization term on the parameters ψ. Formally,
O(ψ, T) = ∑
(s,Ls )∈D TLs (s)6=∅
log p(Ls|s; T,ψ)− λ2‖ψ‖ 2 2 (22)
To learn the parameters ψ, we use a gradient-based optimization method to maximize the objective function O(ψ, T). According to Wainwright and Jordan (2008), the derivative of the log-partition function is the expected feature vector
∂O(ψ, T) ∂ψ
= ∑
(s,Ls )∈D TLs (s)6=∅
(Ep(t|s;TLs ,ψ)[φ(s, t)]− Ep(t|s;T,ψ)[φ(s, t)])− λψ (23)
where Ep(x)[f (x)] = ∑ x p(x)f (x) for discrete x.
4.1.3 Parameter Estimation. The objective function O(ψ, T) is not concave (nor convex), hence the optimization potentially results in a local optimum. Stochastic Gradient Descent (SGD; Robbins and Monro 1951) is a widely used optimization method. The SGD algorithm picks up a training instance randomly, and updates the parameter vector ψ according to
ψj (t+1) = ψj (t) + α ( ∂O(ψ) ∂ψj |ψ=ψ(t) )
(24)
where α is the learning rate, and ∂O(ψ)∂ψj is the gradient of the objective function with respect to parameter ψj. The SGD is sensitive to α, and the learning rate is the same for all dimensions. As described in Section 4.1.1, we mix sparse features together with dense features. We want the learning rate to be different for each dimension. We use AdaGrad (Duchi, Hazan, and Singer 2011) to update the parameters, which sets an adaptive per-feature learning rate. The AdaGrad algorithm tends to use smaller update
steps when we meet a feature many times. In order to compute efficiently, a diagonal approximation version of AdaGrad is used. The update rule is
ψj (t+1) = ψj (t) + α 1√ G(t+1)j
( ∂O(ψ) ∂ψj |ψ=ψ(t) )
G(t+1)j = G (t) j + ( ∂O(ψ) ∂ψj |ψ=ψ(t) )2 (25)
where we introduce an adaptive term G(t)j . G (t) j becomes larger along with updating, and decreases the update step for dimension j. Compared with SGD, the only cost is to store and update G(t)j for each parameter.
To train the model, we use the method proposed by Liang, Jordan, and Klein (2013). With the candidate parse trees and objective function, the parameters ψ are updated to make the parsing model favor correct trees and give them a higher score. Because there are many parse trees for a sentence, we need to calculate Equation (23) efficiently. As indicated in Section 4.1.1, the features decompose along the structure of sentiment trees. So dynamic programming can be utilized to compute Ep(t|s;T,ψ)[φ(s, t)] of Equation (23). However, the first expectation term Ep(t|s;TLs ,ψ)[φ(s, t)] sums over the candidates that obtain the correct polarity labels. As this constraint does not decompose along the tree structure, there is no efficient dynamic program for this. Instead of searching all the parse trees spanning s, we use beam search to approximate this expectation. Beam search is a best-first search algorithm that explores at most K paths (K is the beam size). It keeps the local optimums to reduce the huge search space. Specifically, the beam search algorithm generates the K-best trees with the highest score φ(s, t)Tψ for each span. These local optimums are used recursively in the CYK process. The K-best trees for the whole span are regarded as the candidate set T̃. Then T̃ and T̃Ls are used to approximate Equation (23) as in Liang, Jordan, and Klein (2013).
The intuition behind this parameter estimation algorithm lies in: (1) if we have better parameters, we can obtain better candidate trees; (2) with better candidate trees, we can learn better parameters. Thus the optimization problem is solved in an iterative manner. We initialize the parameters as zeros. This leads to a random search and generates random candidate trees. With the initial candidates, the two steps in Algorithm 2 lead the parameters ψ towards the direction achieving better performance. In this section, we present the automatic learning of the sentiment grammar as defined in Section 3.1. We need to extract the dictionary rules and the combination rules from data. In traditional statistical parsing, grammar rules are induced from annotated parse trees (such as the Penn TreeBank), so ideally we need examples of sentiment structure trees, or sentences annotated with sentiment polarity for the whole sentence as well as those for constituents within sentences. However, this is not practical, if not unfeasible, as the annotations will be inevitably time consuming and require laborious human effort. We show that it is possible to induce the sentiment grammar directly from examples of sentences annotated with sentiment polarity labels without using any syntactic annotations or polarity annotations of constituents within sentences. The
Algorithm 2 Ranking Model Learning Algorithm Input: D: Training data {(s,Ls)}, S: Maximum number of iteration Output: ψ: Parameters of the ranking model
1: ψ(0)← (0, 0, . . . , 0)T 2: repeat 3: (s,Ls)← randomly select a training instance in D 4: T̃(t)← BEAMSEACH(s,ψ(t)) . Beam search to generate K-best candidates
5: G(t+1)j ← G (t) j + ( ∂O(ψ,T̃(t) ) ∂ψj |ψ=ψ(t) )2 6: ψ (t+1) j ← ψj
(t) + α 1√ G(t+1)j ( ∂O(ψ,T̃(t) ) ∂ψj |ψ=ψ(t) )
. Update parameters using
AdaGrad 7: t← t + 1 8: until t > S 9: return ψ(T)
sentences annotated with sentiment polarity labels are relatively easy to obtain, and we use them as our input to learn dictionary rules and combination rules.
We first present the basic idea behind the algorithm we propose. People are likely to express positive or negative opinions using very simple and straightforward sentiment expressions again and again in their reviews. Intuitively, we can mine dictionary rules from these massive review sentences by leveraging the redundancy characteristics. Furthermore, there are many complicated reviews that contain complex sentiment structures (e.g., negation, intensification, and contrast). If we already have dictionary rules on hand, we can use them to obtain basic sentiment information for the fragments within complicated reviews. We can then extract combination rules with the help of the dictionary rules and the sentiment polarity labels of complicated reviews. Because the simple and straightforward sentiment expressions are often coupled with complicated expressions, we need to conduct dictionary rule mining and the combination rule mining in an iterative way.
4.2.1 Dictionary Rule Learning. The dictionary rules GD are basic sentiment building blocks used in the parsing process. Each dictionary rule in GD is in the form X→ f , where f is a sentiment fragment. We use the polarity probabilities P(N | f ) and P(P| f ) in the polarity model. To build GD, we regard all the frequent fragments whose occurrence frequencies are larger than τf and lengths range from 1 to 7 as the sentiment fragments. We further filter the phrases formed by stop words and punctuations, which are not used to express sentiment.
For a balanced data set, the sentiment distribution of a candidate sentiment fragment f is calculated by
P(X | f ) = #( f,X ) + 1 #( f,N ) + #( f,P ) + 2 (26)
where X ∈ {N ,P}, and #( f,X ) denotes the number of reviews containing f with X being the polarity. It should be noted that Laplace smoothing is used in Equation (26) to deal with the zero frequency problem.
We do not learn the polarity probabilities P(N | f ) and P(P| f ) by directly counting occurrence frequency. For example, in the review sentence this movie is not good (negative), the naive counting method increases the count #(good,N ) in terms of the polarity of the whole sentence. Moreover, because of the common collocation not as good as (negative) in movie reviews, as good as is also regarded as negative if we count the frequency directly. The examples indicate why some polarity probabilities of phrases counting from data are different from our intuitions. These unreasonable polarity probabilities also make trouble for learning the polarity model. Consequently, in order to estimate more reasonable probabilities, we need to take the compositionality into consideration when learning sentiment fragments.
Following this motivation, we ignore the count #( f,X ), if the sentiment fragment f is covered by a negation rule r that negates the polarity of f . The word cover here means that f is derived within a non-terminal of the negation rule r. For instance, the negation rule N→ not P covers the sentiment fragment good in the sentence this is not a good movie (negative), that is, the good is derived from P of this negation rule. So we ignore the occurrence for #(good,N ) in this sentence. It should be noted that we still increase the count for #(not good,N ), because there is no negation rule covering the fragment not good.
As shown in Algorithm 3, we learn the dictionary rules and their polarity probabilities by counting the frequencies in negative and positive classes. Only the fragments whose occurrence numbers are larger than threshold τf are kept. Moreover, we take the combination rules into consideration to acquire more reasonable GD. Notably, a subsequence of a frequent fragment must also be frequent. This is similar to the key insight in the Apriori algorithm (Agrawal and Srikant 1994). When we learn the dictionary rules, we can count the sentiment fragments from short to long, and prune the infrequent fragments in the early stages if any subsequence is not frequent. This pruning method accelerates the dictionary rule learning process and makes the procedure fit in memory.
Algorithm 3 Dictionary Rule Learning Input: D: Data set, GC : Combination rules, τf : Frequency threshold Output: GD: Dictionary rules
1: function MINEDICTIONARYRULES(D,GC) 2: GD′← {} 3: for (s,Ls) in D do . s : w0w1 · · ·w|s|−1, Ls: Polarity label of s 4: for all i, j s.t. 0 ≤ i < j ≤ |s| do . w ji : wiwi+1 · · ·wj−1 5: if no negation rule in GC covers w j i then 6: #(w ji,Ls) ++ 7: add w ji to GD ′ 8: GD ← {} 9: for f in GD′ do 10: if #( f, ·) ≥ τf then 11: compute P(N | f ) and P(P| f ) using Equation (26) 12: add dictionary rule (Lf → f ) to GD . Lf = arg maxX∈{N,P} P(X | f ) 13: return GD
4.2.2 Combination Rule Learning. The combination rules GC are generalizations for the dictionary rules. They are used to handle the compositionality and process unseen phrases. The learning of combination rules is based on the learned dictionary rules and their polarity values. The sentiment fragments are generalized to combination rules by replacing the subsequences of dictionary rules with their polarity labels. For instance, as shown in Figure 8, the fragments is not (good/as expected/funny/well done) are all negative. After replacing the subspans good, as expected, funny, and well done with their polarity label P, we can learn the negation rule N→ is not P.
We present the combination rule learning approach in Algorithm 4. Specifically, the first step is to generate combination rule candidates. For every subsequence w ji of sentiment fragment f , we replace it with the corresponding non-terminal Lw ji
if P(Lw ji |w ji) is larger than the threshold τp, and we can get w i 0Lw ji
w| f |j . Next, we compare the polarity Lw ji with Lf . If Lf 6= Lw ji , we regard the rule Lf → w i0Lw ji
w| f |j as a negation rule. Otherwise, we further compare their polarity values. If this rule makes the polarity value become larger (or smaller), it will be treated as a strengthen (or weaken) rule. To obtain the contrast rules, we replace two subsequences with their polarity labels in a similar way. If the polarities of these two subsequences are different, we categorize this rule to the contrast type. Notably, these two non-terminals cannot be next to each other. After these steps, we get the rule candidate set GC ′ and the occurrence number of each rule. We then filter the rule candidates whose occurrence frequencies are too small, and assign the rule types (negation, strengthen, weaken, and contrast) according to their occurrence numbers.
4.2.3 Polarity Model Learning. As shown in Section 3.4, we define the polarity model to calculate the polarity probabilities using the sentiment grammar. In this section, we present how to learn the parameters of the polarity model for the combination rules.
,(
P(P|funny), P(N |is not funny)
)
, (P(P|well done), and P(N |is not well done)) are used to
learn the polarity model for N→ is not P.
Algorithm 4 Combination Rule Learning Input: D: Data set, GD: Dictionary rules, τp, τ∆, τr, τc: Thresholds Output: GC : Combination rules
1: function MINECOMBINATIONRULES(D,GD) 2: GC ′← {} 3: for (X→ f ) in GD do . f : w0w1 · · ·w| f |−1 4: for all i, j s.t. 0 ≤ i < j ≤ | f | do 5: if P(Lw ji |w ji) > τp then . Polarity label Lw ji = arg maxX∈{N ,P} P(X |w j i) 6: r: X→ wi0Lw ji w| f |j . Non-terminal Lw ji = arg maxX∈{N,P} P(X |w j i) 7: if X 6= Lw ji then 8: #(r, negation) ++ 9: else if P(X | f ) > P(Lw ji
|w ji) + τ∆ then 10: #(r, strengthen) ++ 11: else if P(X | f ) < P(Lw ji
|w ji)− τ∆ then 12: #(r, weaken) ++ 13: add r to GC ′ 14: for all i0, j0, i1, j1 s.t. 0 ≤ i0 < j0 < i1 < j1 ≤ | f | do 15: if P(L
w j0i0 |w j0i0 ) > τp and P(Lw j1i1 |w j1i1 ) > τp then
16: r: X→ wi00 Lw j0i0 wi1j0 Lw j1i1 w | f |j1 . Replace w j0 i0 , w j1i1 with the non-terminals 17: if L w j0i0 6= L w j1 i1 then 18: #(r, contrast) ++ 19: add r to GC ′ 20: GC ← {} 21: for r in GC ′ do 22: if #(r, ·) > τr and max
T
#(r,T) #(r) > τc then
23: add r to GC 24: return GC
As shown in Figure 8, we learn combination rules by replacing the subsequences of frequent sentiment fragments with their polarity labels. Both the replaced fragment and the whole fragment can be found in the dictionary rules, so their polarity probabilities have been estimated from data. We can use them as our training examples to figure out how context changes the polarity of replaced fragment, and learn parameters of the polarity model.
We describe the polarity model in Section 3.4. To further simplify the notation, we denote the input vector x = (1, P(X1|w j1 i1 ), . . . , P(XK|w jK iK
))T, and the response value as y. Then we can rewrite Equation (15) as
hθ(x) = 11 + exp{−θTx} (27)
where hθ(x) is the polarity probability calculated by the polarity model, and θ = (θ0,θ1, . . . ,θK )T is the parameter vector. Our goal is to estimate the parameter vector θ of the polarity model.
We fit the model to minimize the sum of squared residuals between the predicted polarity probabilities and the values computed from data. We define the cost function as
J (θ) = 12 ∑
m
(hθ(xm)− ym)2 (28)
where ( xm, ym ) is the m-th training instance.
The gradient descent algorithm is used to minimize the cost function J (θ). The partial derivative of J (θ) with respect to θj is
∂J (θ) ∂θj
= ∑
m
( hθ(xm)− ym ) ∂hθ(xm) ∂θj
= ∑
m
( hθ(xm)− ym ) hθ(xm) ( 1− hθ(xi) ) ∂θTxm ∂θj
= ∑
m
( hθ(xm)− ym ) hθ(xm) (1− hθ(xm)) xmj
(29)
We set the initial θ as zeros, and start with it. We use the Stochastic Gradient Descend algorithm to minimize the cost function. For the instance (x, y), the parameters are updated using
θj (t+1) = θj (t) − α ( ∂J (θ) ∂θj |θ=θ(t) )
= θj (t) − α(hθ(t) (x)− y)hθ(t) (x) ( 1− hθ(t) (x) ) xj
(30)
where α is the learning rate, and it is set to 0.01 in our experiments. We summarize the learning method in Algorithm 5. For each combination rule, we iteratively scan
Algorithm 5 Polarity Model Learning Algorithm Input: GC : Combination rules, ε: Stopping condition, α: Learning rate Output: θ: Parameters of the polarity model
1: function ESTIMATEPOLARITYMODEL(GC) 2: for all combination rule r ∈ GC do 3: θ(0)← (0, 0, ..., 0)T 4: repeat 5: ( x, y ) ← randomly select a training instance 6: θj (t+1)← θj(t) − α(hθ(t) (x)− y)hθ(t) (x) ( 1− hθ(t) (x) ) xj 7: t← t + 1 8: until
∥∥θ(t+1) − θ(t)∥∥22 < ε 9: assign θ(T) as the parameters of the polarity model for rule r
Algorithm 6 Sentiment Grammar Learning Input: D: Data set {(s,Ls)}, T: Maximum number of iteration . Ls: Polarity label of s Output: GD: Dictionary rules, GC : Combination rules
1: GC ← {} 2: repeat 3: GD ←MINEDICTIONARYRULES(D,GC) . Algorithm 3 4: GC ←MINECOMBINATIONRULES(D,GD) . Algorithm 4 5: until iteration number exceeds T 6: ESTIMATEPOLARITYMODEL(GC) . Algorithm 5 7: return GD,GC
through the training examples ( x, y )
in a random order, and update the parameters θ according to Equation (30). The stopping condition is ∥∥θ(t+1) − θ(t)∥∥22 < ε, which indicates the parameters become stable.
4.2.4 Summary of the Grammar Learning Algorithm. We summarize the grammar learning process in Algorithm 6, which learns the sentiment grammar in an iterative manner.
We first learn the dictionary rules and their polarity probabilities by counting the frequencies in negative and positive classes. Only the fragments whose occurrence numbers are larger than the threshold τf are kept. As mentioned in Section 4.2.1, the context can essentially change the distribution of sentiment fragments. We take the combination rules into consideration to acquire more reasonable GD. In the first iteration, the set of combination rules is empty. Therefore, we have no information about compositionality to improve dictionary rule learning. The initial GD contains some inaccurate sentiment distributions. Next, we replace the subsequences of dictionary rules to their polarity labels, and generalize these sentiment fragments to the combination rules GC as illustrated in Section 4.2.2. At the same time, we can obtain their compositional types and learn parameters of the polarity model. We iterate over these two steps to obtain refined GD and GC .","5 experimental studies :In this section, we describe experimental results on existing benchmark data sets with extensive comparisons with state-of-the-art sentiment classification methods. We also present the effects of different experimental settings in the proposed statistical sentiment parsing framework. We describe the data sets in Section 5.1.1, the experimental settings in Section 5.1.2, and the methods used for comparison in Section 5.1.3.
5.1.1 Data Sets. We conduct experiments on sentiment classification for sentence-level and phrase-level data. The sentence-level data sets contain user reviews and critic
reviews from Rotten Tomatoes3 and IMDb.4 We balance the positive and negative instances in the training data set to mitigate the problem of data imbalance. Moreover, the Stanford Sentiment Treebank5 contains polarity labels of all syntactically plausible phrases. In addition, we use the MPQA6 data set for the phrase-level task. We describe these data sets as follows.
RT-C: 436,000 critic reviews from Rotten Tomatoes. It consists of 218,000 negative and 218,000 positive critic reviews. The average review length is 23.2 words. Critic reviews from Rotten Tomatoes contain a label (Rotten: Negative, Fresh: Positive) to indicate the polarity, which we use directly as the polarity label of corresponding reviews.
PL05-C: The sentence polarity data set v1.0 (Pang and Lee 2005) contains 5,331 positive and 5,331 negative snippets written by critics from Rotten Tomatoes. This data set is widely used as the benchmark data set in the sentence-level polarity classification task. The data source is the same as RT-C.
SST: The Stanford Sentiment Treebank (Socher et al. 2013) is built upon PL05-C. The sentences are parsed to parse trees. Then, 215,154 syntactically plausible phrases are extracted and annotated by workers from Amazon Mechanical Turk. The experimental settings of positive/negative classification for sentences are the same as in Socher et al. (2013).
RT-U: 737,806 user reviews from Rotten Tomatoes. Because we focus on sentencelevel sentiment classification, we filter out user reviews that are longer than 200 characters. The average length of these short user reviews from Rotten Tomatoes is 15.4 words. Following previous work on polarity classification, we use the review score to select highly polarized reviews. For the user reviews from Rotten Tomatoes, a negative review has a score <2.5 out of 5, and a positive review has a score >3.5 out of 5.
IMDB-U: 600,000 user reviews from IMDb. The user reviews in IMDb contain comments and short summaries (usually a sentence) to summarize the overall sentiment expressed in the reviews. We use the review summaries as the sentence-level reviews. The average length is 6.6 words. For user reviews of IMDb, a negative review has a score <4 out of 10, and a positive review has a score >7 out of 10.
C-TEST: 2,000 labeled critic reviews sampled from RT-C. We use C-TEST as the testing data set for RT-C. Note that we exclude these from the training data set (i.e., RT-C).
U-TEST: 2,000 manually labeled user reviews sampled from RT-U. User reviews often contain some noisy ratings compared with critic reviews. To eliminate the effect of noise, we sample 2,000 user reviews from RT-U, and annotate their polarity labels manually. We use U-TEST as a testing data set for RT-U and IMDB-U, which are both user reviews. Note that we exclude them from the training data set (i.e., RT-U).
MPQA: The opinion polarity subtask of the MPQA data set (Wiebe, Wilson, and Cardie 2005). The authors manually annotate sentiment polarity labels for the expressions (i.e., sub-sentences) within a sentence. We regard the expressions as short sentences in our experiments. There are 7,308 negative examples and 3,316 positive examples in this data set. The average number of words per example is 3.1.
3 http://www.rottentomatoes.com. 4 http://www.imdb.com. 5 http://nlp.stanford.edu/sentiment/treebank.html. 6 http://mpqa.cs.pitt.edu/corpora/mpqa corpus.
Table 3 shows the summary of these data sets, and all of them are publicly available at http://goo.gl/WxTdPf.
5.1.2 Settings. To compare with other published results for PL05-C and MPQA, the training and testing regime (10-fold cross-validation) is the same as in Pang and Lee (2005), Nakagawa, Inui, and Kurohashi (2010) and Socher et al. (2011). For SST, the regime is the same as in Socher et al. (2013). We use C-TEST as the testing data for RT-C, and U-TEST as the testing data for RT-U and IMDB-U. There are a number of settings that have trade-offs in performance, computation, and the generalization power of our model. The best settings are chosen by a portion of training split data that serves as the validation set. We provide the performance comparisons using different experimental settings in Section 5.4.
Number of training examples: The size of training data has been widely recognized as one of the most important factors in machine learning-based methods. Generally, using more data leads to better performance. By default, all the training data is used in our experiments. We use the same size of training data in different methods for fair comparisons.
Number of training iterations (T): We use AdaGrad (Duchi, Hazan, and Singer 2011) as the optimization algorithm in the learning process. The algorithm starts with randomly initialized parameters, and alternates between searching candidate sentiment trees and updating parameters of the ranking model. We treat one-pass scan of training data as an iteration.
Beam size (K): The beam size is used to make a trade-off between the search space and the computation cost. Moreover, an appropriate beam size can prune unfavorable candidates. We set K = 30 in our experiments.
Regularization (λ): The regularization parameter λ in Equation (22) is used to avoid over-fitting. The value used in the experiments is 0.01.
Minimum fragment frequency: It is difficult to estimate reliable polarity probabilities when the fragment appears very few times. Hence, a minimum fragment frequency that is too small will introduce noise in the fragment learning process. On the other hand, a large threshold will lose much useful information. The minimum fragment frequency is chosen according to the size of the training data set and the validation performance. To be specific, we set this parameter as 4 for RT-C, SST, RT-U, and IMDB-U, and 2 for PL05-C and MPQA.
Maximum fragment length: High order n-grams are more precise and deterministic expressions than unigrams and bigrams. So it would be useful to use long fragments to capture polarity information. According to the experimental results, as the maximum fragment length increases, the accuracy of sentiment classification increases. The maximum fragment length is set to 7 words in our experiments.
5.1.3 Sentiment Classification Methods for Comparison. We evaluate the proposed statistical sentiment parsing framework on the different data sets, and compare the results with some baselines and state-of-the-art sentiment classification methods described as follows.
SVM-m: Support Vector Machine (SVM) achieves good performance in the sentiment classification task (Pang and Lee 2005). Though unigrams and bigrams are reported as the most effective features in existing work (Pang and Lee 2005), we use high-order n-gram (1 ≤ n ≤ m) features to conduct fair comparisons. Hereafter, m has the same meaning. We use LIBLINEAR (Fan et al. 2008) in our experiments because it can handle well the high feature dimension and a large number of training examples. We try different hyper-parameters C ∈ {10−2, 10−1, 1, 5, 10, 20} for SVM, and select C on the validation set.
MNB-m: As indicated in Wang and Manning (2012), Multinomial Naı̈ve Bayes (MNB) often outperforms SVM for sentence-level sentiment classification. We utilize Laplace smoothing (Manning, Raghavan, and Schütze 2008) to tackle the zero probability problem. High order n-gram (1 ≤ n ≤ m) features are considered in the experiments.
LM-m: Language Model (LM) is a generative model calculating the probability of word sequences. It is used for sentiment analysis in Cui, Mittal, and Datar (2006). Probability of generating sentence s is calculated by P(s) = ∏|s|−1 i=0 P ( wi|wi−10 ) , where wi−10 denotes the word sequence w0 . . .wi−1. We use Good-Turing smoothing (Good 1953) to overcome sparsity when estimating the probability of high-order n-gram. We train language models on negative and positive sentences separately. For a sentence, its polarity is determined by comparing the probabilities calculated from the positive and negative language models. The unknown-word token is treated as a regular word (denoted by<UNK>). The SRI Language Modeling Toolkit (Stolcke 2002) is used in our experiment.
Voting-w/Rev: This approach is proposed by Choi and Cardie (2009b), and is used as a baseline in Nakagawa, Inui, and Kurohashi (2010). The polarity of a subjective sentence is decided by the voting of each phrase’s prior polarity. The polarity of phrases that have odd numbers of negation phrases in their ancestors is reversed. The results are reported by Nakagawa, Inui, and Kurohashi (2010).
HardRule: This baseline method is compared by Nakagawa, Inui, and Kurohashi (2010). The polarity of a subjective sentence is deterministically decided based on rules, by considering the sentiment polarity of dependency subtrees. The polarity of a modifier is reversed if its head phrase has a negation word. The decision rules are applied from the leaf nodes to the root node in a dependency tree. We use the results reported by Nakagawa, Inui, and Kurohashi (2010).
Tree-CRF: Nakagawa, Inui, and Kurohashi (2010) present a dependency tree-based method using conditional random fields with hidden variables. In this model, the polarity of each dependency subtree is represented by a hidden variable. The value of the hidden variable of the root node is identified as the polarity of the whole sentence. The experimental results are reported by Nakagawa, Inui, and Kurohashi (2010).
RAE-pretrain: Socher et al. (2011) introduce a framework based on recursive autoencoders to learn vector space representations for multi-word phrases and predict sentiment distributions for sentences. We use the results with pre-trained word vectors learned on Wikipedia, which leads to better results compared with randomized word vectors. We directly compare the results with those in Socher et al. (2011).
MV-RNN: Socher et al. (2012) try to capture the compositional meaning of long phrases through matrix-vector recursive neural networks. This model assigns a vector and a matrix to every node in the parse tree. Matrices are regarded as operators, and vectors capture the meaning of phrases. The results are reported by Socher et al. (2012, 2013).
s.parser-LongMatch: The longest matching rules are utilized in the decoding process. In other words, the derivations that contain the fewest rules are used for all text spans. In addition, the dictionary rules are preferred to the combination rules if both of them match the same text span. The dynamic programming algorithm is used in the implementation.
s.parser-w/oComb: This is our method without using the combination rules (such as N→ not P) learned from data. We present the experimental results of the sentiment classification methods on the different data sets in Table 4. The top three methods on each data set are in bold, and the best methods are also underlined. The experimental results show that s.parser achieves better performance than other methods on most data sets.
The data sets RT-C, PL05-C, and SST are critic reviews. On RT-C, the accuracy of s.parser increases by 2 percentage points, 2.9 percentage points, and 7.1 percentage points from the best results of SVM, MNB, and LM, respectively. On PL05-C, the accuracy of s.parser also rises by 2.1 percentage points, 0.7 percentage points, and 4.4 percentage points from the best results of SVM, MNB, and LM, respectively. Compared to Voting-w/Rev and HardRule, s.parser outperforms them by 16.4 percentage points and 16.6 percentage points. The results indicate that our method significantly outperforms the baselines that use manual rules, as rule-based methods lack a probabilistic way to model the compositionality of context. Furthermore, s.parser achieves an accuracy improvement rate of 2.2 percentage points, 1.8 percentage points, and 0.5 percentage points over Tree-CRF, RAE-pretrain, and MV-RNN, respectively. On SST, s.parser outperforms SVM, MNB, and LM by 3.4 percentage points, 1.4 percentage points, and 3.8 percentage points, respectively. The performance is better than MV-RNN with an improvement rate of 1.8 percentage points. Moreover, the result is comparable to the 85.4% obtained by recursive neural tensor networks (Socher et al. 2013) without depending on syntactic parsing results.
On the user review data sets RT-U and IMDB-U, our method also achieves the best results. More specifically, on the data set RT-U, s.parser outperforms the best results of SVM, MNB, and LM by 1.7 percentage points, 2.9 percentage points, and 1.5 percentage points, respectively. On the data set IMDB-U, our method brings an improved accuracy rate of 2.1 percentage points, 3.7 percentage points, and 2.2 percentage points over SVM, MNB, and LM, respectively. We find that MNB performs better than SVM and LM on the critics review data sets RT-C and PL05-C. Also, SVM and LM achieve better results on the user review data sets RT-U and IMDB-U. The s.parser is more robust for the different genres of data sets.
On the data set MPQA, the accuracy of s.parser increases by 0.5 percentage points, 1.1 percentage points, and 14.8 percentage points from the best results of SVM, MNB, and LM, respectively. Compared with Voting-w/Rev and HardRule, s.parser achieves 4.5 percentage point and 4.4 percentage point improvements over them. As illustrated in Table 3, the size and length of sentences in MPQA are much smaller than those in the other four data sets. The RAE-pretrain achieves better results than other methods on this data set, because the word embeddings pre-trained on Wikipedia can leverage smoothing to relieve the sparsity problem in MPQA. If we do not use any external resources (i.e., Wikipedia), the accuracy of RAE on MPQA is 85.7%, which is lower than Tree-CRF and s.parser. The results indicate that s.parser achieves the best result if no external resource is used.
In addition, we compare the results of s.parser-LongMatch and s.parser-w/oComb. The s.parser-LongMatch utilizes the dictionary rules and combination rules in the longest matching manner, whereas s.parser-w/oComb removes the combination rules in the parsing process. Compared with the results of s.parser, we find that both the ranking model and the combination rules play a positive role in the model. The ranking model learns to score parse trees by assigning larger weights to the rules that tend to obtain correct labels. Also, the combination rules generalize these dictionary rules to deal with the sentiment compositionality in a symbolic way, which enables the model to process unseen phrases. Furthermore, s.parser-LongMatch achieves better results than s.parser-w/oComb. This indicates that the effects of the combination rules are more pronounced than the ranking model.
The bag-of-words classifiers work well for long documents relying on sentiment words that appear many times in a document. The redundancy characteristics provide strong evidence for sentiment classification. Even though some phrases of a document are not estimated accurately, it can still result in a correct polarity label. However, for short text, such as a sentence, the compositionality plays an important role in sentiment classification. Tree-CRF, MV-RNN, and s.parser take compositionality into consideration in different ways, and they achieve significant improvements over SVM, MNB, and LM. We also find that the high order n-grams contribute to classification accuracy on most of the data sets, but they harm the accuracy of LM on PL05-C. The high-order n-grams can partially solve compositionality in a brute-force way. We further investigate the effect of the size of training data for different sentiment classification methods. This is meaningful as the number of the publicly available reviews is increasing dramatically nowadays. The methods that can take advantage of more training data will be even more useful in practice.
We report the results of s.parser compared with SVM, MNB, and LM on the data set RT-C using different training data size. In order to make the figure clear, we only present the results of SVM/MNB/LM-1/5 here. As shown in Figure 9, we find that the size of training data plays an important role for all these sentiment classification methods. The basic conclusion is that the performance of all the methods rise as the data size increases, especially when the data size is smaller than a certain number. It meets our intuition that the size of data is the key factor when the size is relatively small. When the size of data is larger, the growth of accuracy becomes slower. The performance of the baseline methods starts to converge after the data size is larger than 200,000. The comparisons illustrate that s.parser significantly outperforms these baselines. And the performance of s.parser becomes even better when the data size increases. The convergence of s.parser’s performance is slower than the others. It indicates that s.parser leverages data more effectively and benefits more from a larger data set. With more training data, s.parser learns more dictionary rules and combination rules. These rules enhance the generalization ability of our model. Furthermore, it estimates more reliable parameters for the polarity model and ranking model. In contrast, the bag-of-words based approaches (such as SVM, MNB, and LM) cannot make full use of high-order information in the data set. The generalization ability of the combination rules of s.parser leads to better performance, and take advantage of larger data. It should be noted that there are similar trends with the other data sets. In this section, we investigate the effects of different experimental settings. We show the results on the data set RT-C by only changing one factor and fixing the others.
Figure 10 shows the effect of minimum fragment frequency, and maximum fragment length. Specifically, Figure 10a indicates that a minimum fragment frequency that is too small will introduce noise, and it is difficult to estimate reliable polarity probabilities for infrequent fragments. However, a minimum fragment frequency that is too large will discard too much useful information. As shown in Figure 10b, we find that accuracy increases as the maximum fragment length increases. The results illustrate that the large maximum fragment length is helpful for s.parser. We can learn more
combination rules with a larger maximum fragment length, and long dictionary rules capture more precise expressions than unigrams. This conclusion is the same as that in Section 5.2.
As shown in Figure 11, we also investigate how the training iteration, regularization, and beam size affect the results. As shown in Figure 11a, we try a wide range of regularization parameters λ in Equation (22). The results indicate that it is insensitive to the choice of λ. Figure 11b shows the effects of different beam size K in the search process. When beam size K = 1, the optimization algorithm cannot learn the weights. In this case, the decoding process is to select one search path randomly, and compute its polarity probabilities. The results become better as the beam size K increases. On the other hand, the computation costs increase. The proper beam size K can prune some candidates to speed up the search procedure. It should be noted that the sentence length also effects the run time. The sentiment grammar plays a central role in the statistical sentiment parsing framework. It is obvious that the accuracy of s.parser relies on the quality of the automatically learned sentiment grammar. The quality can be implicitly evaluated by the accuracy of sentiment classification results, as we have shown in previous sections. However, there is no straightforward way to explicitly evaluate the quality of the learned grammar. In this section, we provide several case studies of the learned dictionary rules and combination rules to further illustrate the results of the sentiment grammar learning process as detailed in Section 4.2.
To start with, we report the total number of dictionary rules and combination rules learned from the data sets. As shown in Table 5, the results indicate that we can learn more dictionary rules and combination rules from the larger data sets. Although we learn more dictionary rules from RT-C than from IMDB-U, the number of combination rules learned from RT-C is less than from IMDB-U. It indicates that the language usage of RT-C is more diverse than that of IMDB-U. For SST, more rules are learned because of its constituent-level annotations.
Furthermore, we explore how the minimum fragment frequency τf affects the number of dictionary rules, and present the distribution of dictionary rule length. As illustrated in Figure 12a, we find that the relation between total number of dictionary rules |GD| and minimum fragment frequency τf obeys the power law, that is, the log10(|GD|)− log2(τf ) graph takes a linear form. It indicates that most of the fragments appear few times, and only some of them appear frequently. Notably, all the syntactically plausible phrases of SST are annotated, so its distribution is different from the other sentence-level data sets. Figure 12b shows the cumulative distribution of dictionary rule length l. It presents most dictionary rules as short ones. For all data sets except SST, more than 80% of dictionary rules are shorter than five words. The length distributions of data sets RT-C and IMDB-U are similar, whereas we obtain more high order n-grams from RT-U and SST.
We further investigate the effect of context for dictionary rule learning. Table 6 shows some dictionary rules with polarity probabilities learned by our method and
naive counting on RT-C. We notice that if we count the fragment occurrence number directly, some polarities of fragments are learned incorrectly. This is caused by the effect of context as described in Section 4.2.1. By taking the context into consideration, we obtain more reasonable polarity probabilities of dictionary rules. Our dictionary rule learning method takes compositionality into consideration, namely, we skip the count if there exist some negation indicators outside the phrase. This constraint tries to ensure that the polarity of fragments is the same as the whole sentence. As shown in the results, the polarity probabilities learned by our method are more reasonable and meet people’s intuitions. However, there are also some negative examples caused by “false subjective.” For instance, the neutral phrase to pay it tends to appear in negative sentences, and it is learned as a negative phrase. This makes sense for the data distribution, but it may lead to the mismatch for the combination rules.
In Figure 13, we show the polarity model of some combination rules learned from the data set RT-C. The first two examples are negation rules. We find that both switch negation and shift negation exist in data, instead of using only one negation type in previous work (Choi and Cardie 2008; Saurı́ 2008; Taboada et al. 2011). For the rule “N→ i do not P,” we find that it is a switch negation rule. This rule reverses the polarity and the corresponding polarity strength. For instance, the i do not like it very much is more negative than the i do not like it. As shown in Figure 13b, the “N→ is not P” is a shift negation that reduces a fixed polarity strength to reverse the original polarity. Specifically, the is not good is more negative than the is not great, as described in Section 3.4. We have a similar conclusion for the next two weaken rules. As illustrated in Figure 13c, the “P→ P actress” describes one aspect of a movie, hence it is more likely to decrease the polarity intensity. We find that this rule is a fixed intensification rule that reduces the polarity probability by a fixed value. The “N→ a bit of N” is a percentage intensification rule, which scales polarity intensity by a percentage. It reduces more strength for stronger polarity. The last two rules in Figure 13e and Figure 13f are strengthen rules. Both “P→ lot of P” and “N→ N terribly” increase the polarity strength of the sub-fragments. These cases indicate that it is necessary to learn how the context performs compositionality from data. In order to capture the compositionality for different rules, we define the polarity model and learn parameters for each rule. This also agrees with the models of Socher et al. (2012) and Dong et al. (2014),
which use multiple composition matrices to make compositions specific and is an improvement over the recursive neural network that employs one composition matrix.","6 conclusion and future work :In this article, we propose a statistical parsing framework for sentence-level sentiment classification that provides a novel approach to designing sentiment classifiers from a new perspective. It directly analyzes the sentiment structure of a sentence other than relying on syntactic parsing results, as in existing literature. We show that complicated phenomena in sentiment analysis, such as negation, intensification, and contrast, can be handled in a similar manner to simple and straightforward sentiment expressions in a unified and probabilistic way. We provide a formal model to represent the sentiment grammar built upon Context-Free Grammars. The framework consists of: (1) a parsing model to analyze the sentiment structure of a sentence; (2) a polarity model to calculate sentiment strength and polarity for each text span in the parsing process; and (3) a ranking model to select the best parsing result from a list of candidate sentiment parse trees. We show that the sentiment parser can be trained from the examples of sentences annotated only with sentiment polarity labels but without using any syntactic or sentiment annotations within sentences. We evaluate the proposed framework on standard sentiment classification data sets. The experimental results show that the statistical sentiment parsing notably outperforms the baseline sentiment classification approaches.
We believe the work on statistical sentiment parsing can be advanced from many different perspectives. First, statistical parsing has been a well-established research field, in which many different grammars and parsing algorithms have been proposed in previously published literature. It will be an interesting direction to apply and adjust more advanced models and algorithms from the syntactic parsing and the semantic parsing to our framework. We leave it as a line of future work. Second, we can incorporate target and aspect information in the statistical sentiment parsing framework to facilitate the target-dependent and aspect-based sentiment analysis. Intuitively, this can be done by introducing semantic tags of targets and aspects as new non-terminals in the sentiment grammar and revising grammar rules accordingly. However, acquiring training data will be an even more challenging task, as we need more fine-grained information. Third, as the statistical sentiment parsing produces more fine-grained information (e.g., the basic sentiment expressions from the dictionary rules as well as the sentiment structure trees), we will have more opportunities to generate better opinion summaries. Moreover, we are interested in jointly learning parameters of the polarity model and the parsing model from data. Last but not the least, we are interested in investigating the domain adaptation, which is a very important and challenging problem in sentiment analysis. Generally, we may need to learn domain-specific dictionary rules for different domains, whereas we believe combination rules are mostly generic across different domains. This is also worth consideration for further study.",,,"We present a statistical parsing framework for sentence-level sentiment classification in this article. Unlike previous works that use syntactic parsing results for sentiment analysis, we develop a statistical parser to directly analyze the sentiment structure of a sentence. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way. We formulate the sentiment grammar upon Context-Free Grammars (CFGs), and provide a formal description of the sentiment parsing framework. We develop the parsing model to obtain possible sentiment parse trees for a sentence, from which the polarity model is proposed to derive the sentiment strength and polarity, and the ranking model is dedicated to selecting the best sentiment tree. We train the parser directly from examples of sentences annotated only with sentiment polarity labels but without any syntactic annotations or polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, s.parser, from a large amount of review sentences with users’ ratings as rough sentiment polarity labels. Extensive experiments on existing benchmark data sets show significant improvements over baseline sentiment classification approaches.","[{""affiliations"": [], ""name"": ""Li Dong""}, {""affiliations"": [], ""name"": ""Furu Wei""}, {""affiliations"": [], ""name"": ""Shujie Liu""}, {""affiliations"": [], ""name"": ""Ming Zhou""}, {""affiliations"": [], ""name"": ""Ke Xu""}]",SP:bfc3292de4a5066ae9bf2d380c6ef61c6c870698,"[{""authors"": [""Agarwal"", ""Apoorv"", ""Boyi Xie"", ""Ilia Vovsha"", ""Owen Rambow"", ""Rebecca Passonneau.""], ""title"": ""Sentiment analysis of twitter data"", ""venue"": ""Proceedings of the Workshop on Languages in Social Media,"", ""year"": 2011}, {""authors"": [""Agrawal"", ""Rakesh"", ""Ramakrishnan Srikant.""], ""title"": ""Fast algorithms for mining association rules in large databases"", ""venue"": ""Proceedings of the 20th International Conference on Very Large Data Bases,"", ""year"": 1994}, {""authors"": [""Artzi"", ""Yoav"", ""Luke Zettlemoyer.""], ""title"": ""Weakly supervised learning of semantic parsers for mapping instructions to actions"", ""venue"": ""Transactions of the Association for Computational Linguistics,"", ""year"": 2013}, {""authors"": [""Bao"", ""Junwei"", ""Nan Duan"", ""Ming Zhou"", ""Tiejun Zhao.""], ""title"": ""Knowledge-based question answering as machine translation"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association"", ""year"": 2014}, {""authors"": [""Charniak"", ""Eugene.""], ""title"": ""Statistical parsing with a context-free grammar and word statistics"", ""venue"": ""Proceedings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative"", ""year"": 1997}, {""authors"": [""Charniak"", ""Eugene"", ""Mark Johnson.""], ""title"": ""Coarse-to-fine n-best parsing and maxent discriminative reranking"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL \u201905,"", ""year"": 2005}, {""authors"": [""Chiang"", ""David.""], ""title"": ""Hierarchical phrase-based translation"", ""venue"": ""Computational Linguistics, 33(2):201\u2013228."", ""year"": 2007}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Learning with compositional semantics as structural inference for subsentential sentiment analysis"", ""venue"": ""Proceedings of the Conference on Empirical Methods in Natural Language"", ""year"": 2008}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification"", ""venue"": ""Proceedings of the 2009 Conference on Empirical Methods in"", ""year"": 2009}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification"", ""venue"": ""Proceedings of the 2009 Conference on Empirical"", ""year"": 2009}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Hierarchical sequential learning for extracting opinions and their attributes"", ""venue"": ""Proceedings of the ACL 2010 Conference Short Papers, ACLShort \u201910, pages 269\u2013274,"", ""year"": 2010}, {""authors"": [""Chomsky"", ""Noam.""], ""title"": ""Three models for the description of language"", ""venue"": ""IRE Transactions on Information Theory, 2(3):113\u2013124."", ""year"": 1956}, {""authors"": [""Clarke"", ""James"", ""Dan Goldwasser"", ""Ming-Wei Chang"", ""Dan Roth.""], ""title"": ""Driving semantic parsing from the world\u2019s response"", ""venue"": ""Proceedings of the Fourteenth Conference on Computational Natural"", ""year"": 2010}, {""authors"": [""Cocke"", ""John.""], ""title"": ""Programming Languages and Their Compilers: Preliminary Notes"", ""venue"": ""Courant Institute of Mathematical Sciences, New York University."", ""year"": 1969}, {""authors"": [""Councill"", ""Isaac G."", ""Ryan McDonald"", ""Leonid Velikovich.""], ""title"": ""What\u2019s great and what\u2019s not: Learning to classify the scope of negation for improved sentiment analysis"", ""venue"": ""Proceedings of the Workshop on"", ""year"": 2010}, {""authors"": [""Cui"", ""Hang"", ""Vibhu Mittal"", ""Mayur Datar.""], ""title"": ""Comparative experiments on sentiment classification for online product reviews"", ""venue"": ""Proceedings of the 21st National Conference on Artificial Intelligence -"", ""year"": 2006}, {""authors"": [""Davidov"", ""Dmitry"", ""Oren Tsur"", ""Ari Rappoport.""], ""title"": ""Enhanced sentiment learning using twitter hashtags and smileys"", ""venue"": ""Proceedings of the 23rd International Conference on Computational"", ""year"": 2010}, {""authors"": [""de Marneffe"", ""Marie-Catherine"", ""Christopher D. Manning"", ""Christopher Potts""], ""title"": ""was it good? it was provocative.\u201d Learning the meaning of scalar adjectives"", ""year"": 2010}, {""authors"": [""Dong"", ""Li"", ""Furu Wei"", ""Ming Zhou"", ""Ke Xu.""], ""title"": ""Adaptive multi-compositionality for recursive neural models with applications to sentiment analysis"", ""venue"": ""AAAI Conference on Artificial Intelligence, pages 1,537\u20131,543,"", ""year"": 2014}, {""authors"": [""Duchi"", ""John"", ""Elad Hazan"", ""Yoram Singer.""], ""title"": ""Adaptive subgradient methods for online learning and stochastic optimization"", ""venue"": ""Journal of Machine Learning Research, 12:2121\u20132159."", ""year"": 2011}, {""authors"": [""Fan"", ""Rong-En"", ""Kai-Wei Chang"", ""Cho-Jui Hsieh"", ""Xiang-Rui Wang"", ""Chih-Jen Lin.""], ""title"": ""Liblinear: A library for large linear classification"", ""venue"": ""Journal of Machine Learning Research, 9:1871\u20131874."", ""year"": 2008}, {""authors"": [""Good"", ""Irving John.""], ""title"": ""The population frequencies of species and the estimation of population parameters"", ""venue"": ""Biometrika, 40(3-4):237\u2013264."", ""year"": 1953}, {""authors"": [""Goodman"", ""Joshua.""], ""title"": ""Semiring parsing"", ""venue"": ""Computational Linguistics, 25(4):573\u2013605."", ""year"": 1999}, {""authors"": [""Hall"", ""David"", ""Greg Durrett"", ""Dan Klein.""], ""title"": ""Less grammar, more features"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 228\u2013237,"", ""year"": 2014}, {""authors"": [""Hatzivassiloglou"", ""Vasileios"", ""Kathleen R. McKeown.""], ""title"": ""Predicting the semantic orientation of adjectives"", ""venue"": ""Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and Eighth"", ""year"": 1997}, {""authors"": [""Jia"", ""Lifeng"", ""Clement Yu"", ""Weiyi Meng.""], ""title"": ""The effect of negation on sentiment analysis and retrieval effectiveness"", ""venue"": ""Conference on Information and Knowledge Management, pages 1,827\u20131,830,"", ""year"": 2009}, {""authors"": [""Kaji"", ""Nobuhiro"", ""Masaru Kitsuregawa.""], ""title"": ""Building lexicon for sentiment analysis from massive collection of HTML documents"", ""venue"": ""Proceedings of the 2007 Joint Conference on Empirical Methods in Natural"", ""year"": 2007}, {""authors"": [""Kamps"", ""Jaap"", ""Robert J. Mokken"", ""Maarten Marx"", ""Maarten de Rijke.""], ""title"": ""Using WordNet to measure semantic orientation of adjectives"", ""venue"": ""Proceedings of the 4th International Conference on Language"", ""year"": 2004}, {""authors"": [""Kanayama"", ""Hiroshi"", ""Tetsuya Nasukawa.""], ""title"": ""Fully automatic lexicon expansion for domain-oriented sentiment analysis"", ""venue"": ""Proceedings of the 2006 Conference on Empirical Methods in Natural Language"", ""year"": 2006}, {""authors"": [""Kasami"", ""Tadao.""], ""title"": ""An efficient recognition and syntax-analysis algorithm for context-free languages"", ""venue"": ""Technical report AFCRL-65-758, Air Force Cambridge Research Lab, Bedford, MA."", ""year"": 1965}, {""authors"": [""Kate"", ""Rohit J."", ""Raymond J. Mooney.""], ""title"": ""Using string-kernels for learning semantic parsers"", ""venue"": ""ACL 2006: Proceedings of the 21st International Conference on Computational Linguistics and the 44th"", ""year"": 2006}, {""authors"": [""Kennedy"", ""Alistair"", ""Diana Inkpen.""], ""title"": ""Sentiment classification of movie reviews using contextual valence shifters"", ""venue"": ""Computational Intelligence, 22:110\u2013125."", ""year"": 2006}, {""authors"": [""Klein"", ""Dan"", ""Christopher D. Manning.""], ""title"": ""Accurate unlexicalized parsing"", ""venue"": ""Proceedings of the 41st Annual Meeting on Association for Computational Linguistics Volume 1, ACL \u201903, pages 423\u2013430,"", ""year"": 2003}, {""authors"": [""Klenner"", ""Manfred"", ""Stefanos Petrakis"", ""Angela Fahrni.""], ""title"": ""Robust compositional polarity classification"", ""venue"": ""Proceedings of the International Conference RANLP-2009, pages 180\u2013184, Borovets."", ""year"": 2009}, {""authors"": [""Krestel"", ""Ralf"", ""Stefan Siersdorfer.""], ""title"": ""Generating contextualized sentiment lexica based on latent topics and user ratings"", ""venue"": ""Proceedings of the 24th ACM Conference on Hypertext and Social Media,"", ""year"": 2013}, {""authors"": [""Krishnamurthy"", ""Jayant"", ""Tom M. Mitchell.""], ""title"": ""Weakly supervised training of semantic parsers"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and"", ""year"": 2012}, {""authors"": [""K\u00fcbler"", ""Sandra"", ""Ryan McDonald"", ""Joakim Nivre.""], ""title"": ""Dependency parsing"", ""venue"": ""Synthesis Lectures on Human Language Technologies, 1(1):1\u2013127."", ""year"": 2009}, {""authors"": [""Li"", ""Peng"", ""Yang Liu"", ""Maosong Sun.""], ""title"": ""An extended ghkm algorithm for inducing Lambda-SCFG"", ""venue"": ""Conference on Artificial Intelligence, pages 605\u2013611, Bellevue, WA."", ""year"": 2013}, {""authors"": [""Liang"", ""Percy"", ""Michael I. Jordan"", ""Dan Klein.""], ""title"": ""Learning dependency-based compositional semantics"", ""venue"": ""Computational Linguistics, 39(2):389\u2013446."", ""year"": 2013}, {""authors"": [""Liu"", ""Bing.""], ""title"": ""Sentiment Analysis and Opinion Mining"", ""venue"": ""Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers."", ""year"": 2012}, {""authors"": [""Liu"", ""Jingjing"", ""Stephanie Seneff""], ""title"": ""Review sentiment scoring via a 305"", ""year"": 2009}, {""authors"": [""Liu"", ""Shizhu"", ""Gady Agam"", ""David A. Grossman.""], ""title"": ""Generalized sentiment-bearing expression features for sentiment analysis"", ""venue"": ""International Conference on Computational Linguistics,"", ""year"": 2012}, {""authors"": [""Lu"", ""Yue"", ""Mal\u00fa Castellanos"", ""Umeshwar Dayal"", ""ChengXiang Zhai.""], ""title"": ""Automatic construction of a context-aware sentiment lexicon: an optimization approach"", ""venue"": ""International World Wide Web Conference,"", ""year"": 2011}, {""authors"": [""Maas"", ""Andrew L."", ""Raymond E. Daly"", ""Peter T. Pham"", ""Dan Huang"", ""Andrew Y. Ng"", ""Christopher Potts.""], ""title"": ""Learning word vectors for sentiment analysis"", ""venue"": ""Proceedings of the 49th Annual Meeting of the"", ""year"": 2011}, {""authors"": [""Manning"", ""Christopher D."", ""Prabhakar Raghavan"", ""Hinrich Sch\u00fctze.""], ""title"": ""Introduction to Information Retrieval"", ""venue"": ""Cambridge University Press, New York."", ""year"": 2008}, {""authors"": [""Marcus"", ""Mitchell P."", ""Mary Ann Marcinkiewicz"", ""Beatrice Santorini.""], ""title"": ""Building a large annotated corpus of English: the Penn Treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1993}, {""authors"": [""Matsumoto"", ""Shotaro"", ""Hiroya Takamura"", ""Manabu Okumura.""], ""title"": ""Sentiment classification using word sub-sequences and dependency sub-trees"", ""venue"": ""Proceedings of the 9th Pacific-Asia Conference on Advances"", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Koby Crammer"", ""Fernando Pereira.""], ""title"": ""Online large-margin training of dependency parsers"", ""venue"": ""Proceedings of the 43rd Annual Meeting on Association for Computational"", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Kerry Hannan"", ""Tyler Neylon"", ""Mike Wells"", ""Jeff Reynar.""], ""title"": ""Structured models for fine-to-coarse sentiment analysis"", ""venue"": ""Proceedings of the Association for Computational Linguistics"", ""year"": 2007}, {""authors"": [""Moilanen"", ""Karo"", ""Stephen Pulman.""], ""title"": ""Sentiment composition"", ""venue"": ""Proceedings of Recent Advances in Natural Language"", ""year"": 2007}, {""authors"": [""Moilanen"", ""Karo"", ""Stephen Pulman"", ""Yue Zhang.""], ""title"": ""Packed feelings and ordered sentiments: Sentiment parsing with quasi-compositional polarity sequencing and compression"", ""venue"": ""Proceedings of the 1st"", ""year"": 2010}, {""authors"": [""Mudinas"", ""Andrius"", ""Dell Zhang"", ""Mark Levene.""], ""title"": ""Combining lexicon and learning based approaches for concept-level sentiment analysis"", ""venue"": ""Proceedings of the First International"", ""year"": 2012}, {""authors"": [""Nakagawa"", ""Tetsuji"", ""Kentaro Inui"", ""Sadao Kurohashi.""], ""title"": ""Dependency tree-based sentiment classification using CRFS with hidden variables"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of"", ""year"": 2010}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts"", ""venue"": ""Proceedings of the 42nd Meeting of the Association for Computational"", ""year"": 2004}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales"", ""venue"": ""Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics,"", ""year"": 2005}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""Opinion mining and sentiment analysis"", ""venue"": ""Foundations and Trends in Information Retrieval, 2(1-2):1\u2013135."", ""year"": 2008}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee"", ""Shivakumar Vaithyanathan.""], ""title"": ""Thumbs up?: Sentiment classification using machine learning techniques"", ""venue"": ""Proceedings of the ACL-02 Conference on Empirical"", ""year"": 2002}, {""authors"": [""Polanyi"", ""Livia"", ""Annie Zaenen.""], ""title"": ""Contextual valence shifters"", ""venue"": ""J. G. Shanahan, Y. Qu, and J. Wiebe, editors, Computing Attitude and Affect in Text: Theory and Applications. Springer"", ""year"": 2006}, {""authors"": [""R. Quirk""], ""title"": ""A Comprehensive Grammar of the English Language"", ""venue"": ""Longman."", ""year"": 1985}, {""authors"": [""Raymond"", ""Ruifang Ge"", ""J. Mooney.""], ""title"": ""Discriminative reranking for semantic parsing"", ""venue"": ""Proceedings of the COLING/ACL on Main Conference Poster Sessions, COLING-ACL \u201906, pages 263\u2013270,"", ""year"": 2006}, {""authors"": [""H. Robbins"", ""S. Monro.""], ""title"": ""A stochastic approximation method"", ""venue"": ""Annals of Mathematical Statistics, 22:400\u2013407."", ""year"": 1951}, {""authors"": [""Saur\u0131"", ""Roser""], ""title"": ""A Factuality Profiler for Eventualities in Text"", ""year"": 2008}, {""authors"": [""Shieber"", ""Stuart M."", ""Yves Schabes"", ""Fernando C.N. Pereira.""], ""title"": ""Principles and implementation of deductive parsing"", ""venue"": ""Journal of Logic Programming, 24(1\u20132):3\u201336."", ""year"": 1995}, {""authors"": [""Socher"", ""Richard"", ""Brody Huval"", ""Christopher D. Manning"", ""Andrew Y. Ng.""], ""title"": ""Semantic compositionality through recursive matrix-vector spaces"", ""venue"": ""Proceedings of the 2012 Joint Conference on"", ""year"": 2012}, {""authors"": [""Socher"", ""Richard"", ""Jeffrey Pennington"", ""Eric H. Huang"", ""Andrew Y. Ng"", ""Christopher D. Manning.""], ""title"": ""Semi-supervised recursive autoencoders for predicting sentiment distributions"", ""venue"": ""Proceedings of the"", ""year"": 2011}, {""authors"": [""Socher"", ""Richard"", ""Alex Perelygin"", ""Jean Y. Wu"", ""Jason Chuang"", ""Christopher D. Manning"", ""Andrew Y. Ng"", ""Christopher Potts""], ""title"": ""Recursive deep models for semantic compositionality over a sentiment"", ""year"": 2013}, {""authors"": [""Stolcke"", ""Andreas.""], ""title"": ""SRILM: An extensible language modeling toolkit"", ""venue"": ""Proceedings of the 7th International Conference on Spoken Language Processing (ICSLP 2002, pages 901\u2013904, Denver, CO."", ""year"": 2002}, {""authors"": [""Taboada"", ""Maite"", ""Julian Brooke"", ""Milan Tofiloski"", ""Kimberly Voll"", ""Manfred Stede.""], ""title"": ""Lexicon-based methods for sentiment analysis"", ""venue"": ""Computational Linguistics, 37(2):267\u2013307."", ""year"": 2011}, {""authors"": [""T\u00e4ckstr\u00f6m"", ""Oscar"", ""Ryan McDonald.""], ""title"": ""Discovering fine-grained sentiment with latent variable structured prediction models"", ""venue"": ""Proceedings of the 33rd European Conference on Advances in Information"", ""year"": 2011}, {""authors"": [""T\u00e4ckstr\u00f6m"", ""Oscar"", ""Ryan McDonald.""], ""title"": ""Semi-supervised latent variable models for sentence-level sentiment analysis"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational"", ""year"": 2011}, {""authors"": [""Takamura"", ""Hiroya"", ""Takashi Inui"", ""Manabu Okumura.""], ""title"": ""Extracting semantic orientations of words using spin model"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""Tu"", ""Zhaopeng"", ""Yifan He"", ""Jennifer Foster"", ""Josef van Genabith"", ""Qun Liu"", ""Shouxun Lin.""], ""title"": ""Identifying high-impact sub-structures for convolution kernels in document-level sentiment classification"", ""venue"": ""In"", ""year"": 2012}, {""authors"": [""Turney"", ""Peter D.""], ""title"": ""Thumbs up or thumbs down?: Semantic orientation applied to unsupervised classification of reviews"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2002}, {""authors"": [""Velikovich"", ""Leonid"", ""Sasha Blair-Goldensohn"", ""Kerry Hannan"", ""Ryan McDonald.""], ""title"": ""The viability of Web-derived polarity lexicons"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North"", ""year"": 2010}, {""authors"": [""Wainwright"", ""Martin J."", ""Michael I. Jordan.""], ""title"": ""Graphical models, exponential families, and variational inference"", ""venue"": ""Foundations and Trends in Machine Learning, 1(1-2):1\u2013305."", ""year"": 2008}, {""authors"": [""Wang"", ""Sida"", ""Christopher Manning.""], ""title"": ""Baselines and bigrams: Simple, good sentiment and topic classification"", ""venue"": ""Proceedings of the 50th Annual Meeting of the Association for Computational"", ""year"": 2012}, {""authors"": [""Wiebe"", ""Janyce"", ""Theresa Wilson"", ""Claire Cardie.""], ""title"": ""Annotating expressions of opinions and emotions in language"", ""venue"": ""Language Resources and Evaluation, 39(2-3):165\u2013210."", ""year"": 2005}, {""authors"": [""Williams"", ""Gbolahan K"", ""Sarabjot Singh Anand""], ""title"": ""Predicting the polarity"", ""year"": 2009}, {""authors"": [""Wilson"", ""Theresa"", ""Janyce Wiebe"", ""Paul Hoffmann.""], ""title"": ""Recognizing contextual polarity: An exploration of features for phrase-level sentiment analysis"", ""venue"": ""Computational Linguistics, 35:399\u2013433."", ""year"": 2009}, {""authors"": [""Yessenalina"", ""Ainur"", ""Yisong Yue"", ""Claire Cardie.""], ""title"": ""Multi-level structured models for document-level sentiment classification"", ""venue"": ""Proceedings of the 2010 Conference on Empirical Methods in Natural"", ""year"": 2010}, {""authors"": [""Younger"", ""Daniel H.""], ""title"": ""Recognition and parsing of context-free languages in time n3"", ""venue"": ""Information and Control, 10(2):189\u2013208."", ""year"": 1967}, {""authors"": [""Yu"", ""Hong"", ""Vasileios Hatzivassiloglou.""], ""title"": ""Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences"", ""venue"": ""Proceedings of the 2003"", ""year"": 2003}, {""authors"": [""Zelle"", ""John M"", ""Raymond J. Mooney""], ""title"": ""Learning to parse database queries"", ""year"": 1996}, {""authors"": [""Zettlemoyer"", ""Luke S."", ""Michael Collins.""], ""title"": ""Online learning of relaxed CCG grammars for parsing to logical form"", ""venue"": ""Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language"", ""year"": 2007}, {""authors"": [""Zettlemoyer"", ""Luke S."", ""Michael Collins.""], ""title"": ""Learning context-dependent mappings from sentences to logical form"", ""venue"": ""Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the"", ""year"": 2009}, {""authors"": [""Zhao"", ""Jichang"", ""Li Dong"", ""Junjie Wu"", ""Ke Xu.""], ""title"": ""Moodlens: An emoticon-based sentiment analysis system for Chinese tweets"", ""venue"": ""Proceedings of the 18th ACM SIGKDD International Conference"", ""year"": 2012}]","acknowledgments  :This research was partly supported by NSFC (grant no. 61421003) and the fund of the State Key Lab of Software Development Environment (grant no. SKLSDE-2015ZX-05). We gratefully acknowledge helpful discussions with Dr. Nan Yang and the anonymous reviewers.
References Agarwal, Apoorv, Boyi Xie, Ilia Vovsha,
Owen Rambow, and Rebecca Passonneau. 2011. Sentiment analysis of twitter data. In Proceedings of the Workshop on Languages in Social Media, LSM ’11, pages 30–38, Stroudsburg, PA.
Agrawal, Rakesh and Ramakrishnan Srikant. 1994. Fast algorithms for mining association rules in large databases. In Proceedings of the 20th International Conference on Very Large Data Bases, VLDB ’94, pages 487–499, San Francisco, CA. Artzi, Yoav and Luke Zettlemoyer. 2013. Weakly supervised learning of semantic parsers for mapping instructions to actions. Transactions of the Association for Computational Linguistics, 1(1):49–62. Bao, Junwei, Nan Duan, Ming Zhou, and Tiejun Zhao. 2014. Knowledge-based question answering as machine translation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 967–976, Baltimore, MD. Charniak, Eugene. 1997. Statistical parsing with a context-free grammar and word statistics. In Proceedings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative Applications of Artificial Intelligence, AAAI’97/IAAI’97, pages 598–603, Providence, RI. Charniak, Eugene and Mark Johnson. 2005. Coarse-to-fine n-best parsing and maxent discriminative reranking. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL ’05, pages 173–180, Stroudsburg, PA. Chiang, David. 2007. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228. Choi, Yejin and Claire Cardie. 2008. Learning with compositional semantics as structural inference for subsentential sentiment analysis. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’08, pages 793–801, Stroudsburg, PA. Choi, Yejin and Claire Cardie. 2009a. Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 2 – Volume 2, EMNLP ’09, pages 590–598, Stroudsburg, PA. Choi, Yejin and Claire Cardie. 2009b. Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing:
Volume 2 – Volume 2, pages 590–598, Singapore.
Choi, Yejin and Claire Cardie. 2010. Hierarchical sequential learning for extracting opinions and their attributes. In Proceedings of the ACL 2010 Conference Short Papers, ACLShort ’10, pages 269–274, Stroudsburg, PA. Chomsky, Noam. 1956. Three models for the description of language. IRE Transactions on Information Theory, 2(3):113–124. Clarke, James, Dan Goldwasser, Ming-Wei Chang, and Dan Roth. 2010. Driving semantic parsing from the world’s response. In Proceedings of the Fourteenth Conference on Computational Natural Language Learning, CoNLL ’10, pages 18–27, Stroudsburg, PA. Cocke, John. 1969. Programming Languages and Their Compilers: Preliminary Notes. Courant Institute of Mathematical Sciences, New York University. Councill, Isaac G., Ryan McDonald, and Leonid Velikovich. 2010. What’s great and what’s not: Learning to classify the scope of negation for improved sentiment analysis. In Proceedings of the Workshop on Negation and Speculation in Natural Language Processing, NeSp-NLP ’10, pages 51–59, Stroudsburg, PA. Cui, Hang, Vibhu Mittal, and Mayur Datar. 2006. Comparative experiments on sentiment classification for online product reviews. In Proceedings of the 21st National Conference on Artificial Intelligence - Volume 2, AAAI’06, pages 1,265–1,270, Boston, MA. Davidov, Dmitry, Oren Tsur, and Ari Rappoport. 2010. Enhanced sentiment learning using twitter hashtags and smileys. In Proceedings of the 23rd International Conference on Computational Linguistics: Posters, COLING ’10, pages 241–249, Stroudsburg, PA. de Marneffe, Marie-Catherine, Christopher D. Manning, and Christopher Potts. 2010. “was it good? it was provocative.” Learning the meaning of scalar adjectives. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, ACL ’10, pages 167–176, Stroudsburg, PA. Dong, Li, Furu Wei, Ming Zhou, and Ke Xu. 2014. Adaptive multi-compositionality for recursive neural models with applications to sentiment analysis. In AAAI Conference on Artificial Intelligence, pages 1,537–1,543, Quebec.
Duchi, John, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159. Fan, Rong-En, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui Wang, and Chih-Jen Lin. 2008. Liblinear: A library for large linear classification. Journal of Machine Learning Research, 9:1871–1874. Good, Irving John. 1953. The population frequencies of species and the estimation of population parameters. Biometrika, 40(3-4):237–264. Goodman, Joshua. 1999. Semiring parsing. Computational Linguistics, 25(4):573–605. Hall, David, Greg Durrett, and Dan Klein. 2014. Less grammar, more features. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 228–237, Baltimore, MD. Hatzivassiloglou, Vasileios and Kathleen R. McKeown. 1997. Predicting the semantic orientation of adjectives. In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and Eighth Conference of the European Chapter of the Association for Computational Linguistics, ACL ’98, pages 174–181, Stroudsburg, PA. Jia, Lifeng, Clement Yu, and Weiyi Meng. 2009. The effect of negation on sentiment analysis and retrieval effectiveness. In Conference on Information and Knowledge Management, pages 1,827–1,830, Hong Kong. Kaji, Nobuhiro and Masaru Kitsuregawa. 2007. Building lexicon for sentiment analysis from massive collection of HTML documents. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 1,075–1,083, Prague. Kamps, Jaap, Robert J. Mokken, Maarten Marx, and Maarten de Rijke. 2004. Using WordNet to measure semantic orientation of adjectives. In Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC 2004), volume IV, pages 1,115–1,118, Paris. Kanayama, Hiroshi and Tetsuya Nasukawa. 2006. Fully automatic lexicon expansion for domain-oriented sentiment analysis. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, EMNLP ’06, pages 355–363, Stroudsburg, PA.
Kasami, Tadao. 1965. An efficient recognition and syntax-analysis algorithm for context-free languages. Technical report AFCRL-65-758, Air Force Cambridge Research Lab, Bedford, MA. Kate, Rohit J. and Raymond J. Mooney. 2006. Using string-kernels for learning semantic parsers. In ACL 2006: Proceedings of the 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the ACL, pages 913–920, Morristown, NJ. Kennedy, Alistair and Diana Inkpen. 2006. Sentiment classification of movie reviews using contextual valence shifters. Computational Intelligence, 22:110–125. Klein, Dan and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics - Volume 1, ACL ’03, pages 423–430, Stroudsburg, PA. Klenner, Manfred, Stefanos Petrakis, and Angela Fahrni. 2009. Robust compositional polarity classification. In Proceedings of the International Conference RANLP-2009, pages 180–184, Borovets. Krestel, Ralf and Stefan Siersdorfer. 2013. Generating contextualized sentiment lexica based on latent topics and user ratings. In Proceedings of the 24th ACM Conference on Hypertext and Social Media, HT ’13, pages 129–138, New York, NY. Krishnamurthy, Jayant and Tom M. Mitchell. 2012. Weakly supervised training of semantic parsers. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL ’12, pages 754–765, Stroudsburg, PA. Kübler, Sandra, Ryan McDonald, and Joakim Nivre. 2009. Dependency parsing. Synthesis Lectures on Human Language Technologies, 1(1):1–127. Li, Peng, Yang Liu, and Maosong Sun. 2013. An extended ghkm algorithm for inducing Lambda-SCFG. In Conference on Artificial Intelligence, pages 605–611, Bellevue, WA. Liang, Percy, Michael I. Jordan, and Dan Klein. 2013. Learning dependency-based compositional semantics. Computational Linguistics, 39(2):389–446. Liu, Bing. 2012. Sentiment Analysis and Opinion Mining. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers. Liu, Jingjing and Stephanie Seneff. 2009. Review sentiment scoring via a
parse-and-paraphrase paradigm. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1 - Volume 1, EMNLP ’09, pages 161–169, Stroudsburg, PA.
Liu, Shizhu, Gady Agam, and David A. Grossman. 2012. Generalized sentiment-bearing expression features for sentiment analysis. In International Conference on Computational Linguistics, pages 733–744, Mumbai. Lu, Yue, Malú Castellanos, Umeshwar Dayal, and ChengXiang Zhai. 2011. Automatic construction of a context-aware sentiment lexicon: an optimization approach. In International World Wide Web Conference, pages 347–356, Hyderabad. Maas, Andrew L., Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies - Volume 1, HLT ’11, pages 142–150, Stroudsburg, PA. Manning, Christopher D., Prabhakar Raghavan, and Hinrich Schütze. 2008. Introduction to Information Retrieval. Cambridge University Press, New York. Marcus, Mitchell P., Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of English: the Penn Treebank. Computational Linguistics, 19(2):313–330. Matsumoto, Shotaro, Hiroya Takamura, and Manabu Okumura. 2005. Sentiment classification using word sub-sequences and dependency sub-trees. In Proceedings of the 9th Pacific-Asia Conference on Advances in Knowledge Discovery and Data Mining, PAKDD’05, pages 301–311, Hanoi. McDonald, Ryan, Koby Crammer, and Fernando Pereira. 2005. Online large-margin training of dependency parsers. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL ’05, pages 91–98, Stroudsburg, PA. McDonald, Ryan, Kerry Hannan, Tyler Neylon, Mike Wells, and Jeff Reynar. 2007. Structured models for fine-to-coarse sentiment analysis. In Proceedings of the Association for Computational Linguistics (ACL), pages 432–439, Prague. Moilanen, Karo and Stephen Pulman. 2007. Sentiment composition. In Proceedings of Recent Advances in Natural Language
Processing (RANLP 2007), pages 378–382, Borovets.
Moilanen, Karo, Stephen Pulman, and Yue Zhang. 2010. Packed feelings and ordered sentiments: Sentiment parsing with quasi-compositional polarity sequencing and compression. In Proceedings of the 1st Workshop on Computational Approaches to Subjectivity and Sentiment Analysis (WASSA 2010) at the 19th European Conference on Artificial Intelligence (ECAI 2010), pages 36–43, Lisbon. Mudinas, Andrius, Dell Zhang, and Mark Levene. 2012. Combining lexicon and learning based approaches for concept-level sentiment analysis. In Proceedings of the First International Workshop on Issues of Sentiment Discovery and Opinion Mining, WISDOM ’12, pages 5:1–5:8, New York, NY. Nakagawa, Tetsuji, Kentaro Inui, and Sadao Kurohashi. 2010. Dependency tree-based sentiment classification using CRFS with hidden variables. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, HLT ’10, pages 786–794, Stroudsburg, PA. Pang, Bo and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Meeting of the Association for Computational Linguistics (ACL’04), Main Volume, pages 271–278, Barcelona. Pang, Bo and Lillian Lee. 2005. Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL ’05, pages 115–124, Stroudsburg, PA. Pang, Bo and Lillian Lee. 2008. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1-2):1–135. Pang, Bo, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up?: Sentiment classification using machine learning techniques. In Proceedings of the ACL-02 Conference on Empirical Methods in Natural Language Processing - Volume 10, EMNLP ’02, pages 79–86, Stroudsburg, PA. Polanyi, Livia and Annie Zaenen. 2006. Contextual valence shifters. In J. G. Shanahan, Y. Qu, and J. Wiebe, editors, Computing Attitude and Affect in Text: Theory and Applications. Springer Netherlands, pages 1–10.
Quirk, R. 1985. A Comprehensive Grammar of the English Language. Longman. Raymond, Ruifang Ge and J. Mooney. 2006. Discriminative reranking for semantic parsing. In Proceedings of the COLING/ACL on Main Conference Poster Sessions, COLING-ACL ’06, pages 263–270, Stroudsburg, PA. Robbins, H. and S. Monro. 1951. A stochastic approximation method. Annals of Mathematical Statistics, 22:400–407. Saurı́, Roser. 2008. A Factuality Profiler for Eventualities in Text. Ph.D. thesis, Brandeis University. Shieber, Stuart M., Yves Schabes, and Fernando C. N. Pereira. 1995. Principles and implementation of deductive parsing. Journal of Logic Programming, 24(1–2):3–36. Socher, Richard, Brody Huval, Christopher D. Manning, and Andrew Y. Ng. 2012. Semantic compositionality through recursive matrix-vector spaces. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL ’12, pages 1,201–1,211, Stroudsburg, PA. Socher, Richard, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning. 2011. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 151–161, Stroudsburg, PA. Socher, Richard, Alex Perelygin, Jean Y. Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1,631–1,642, Seattle, WA. Stolcke, Andreas. 2002. SRILM: An extensible language modeling toolkit. In Proceedings of the 7th International Conference on Spoken Language Processing (ICSLP 2002, pages 901–904, Denver, CO. Taboada, Maite, Julian Brooke, Milan Tofiloski, Kimberly Voll, and Manfred Stede. 2011. Lexicon-based methods for sentiment analysis. Computational Linguistics, 37(2):267–307. Täckström, Oscar and Ryan McDonald. 2011a. Discovering fine-grained sentiment with latent variable structured prediction models. In Proceedings of the 33rd European Conference on Advances in Information
Retrieval, ECIR’11, pages 368–374, Berlin.
Täckström, Oscar and Ryan McDonald. 2011b. Semi-supervised latent variable models for sentence-level sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers - Volume 2, HLT ’11, pages 569–574, Stroudsburg, PA. Takamura, Hiroya, Takashi Inui, and Manabu Okumura. 2005. Extracting semantic orientations of words using spin model. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL ’05, pages 133–140, Stroudsburg, PA. Tu, Zhaopeng, Yifan He, Jennifer Foster, Josef van Genabith, Qun Liu, and Shouxun Lin. 2012. Identifying high-impact sub-structures for convolution kernels in document-level sentiment classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Short Papers - Volume 2, ACL ’12, pages 338–343, Stroudsburg, PA. Turney, Peter D. 2002. Thumbs up or thumbs down?: Semantic orientation applied to unsupervised classification of reviews. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, ACL ’02, pages 417–424, Stroudsburg, PA. Velikovich, Leonid, Sasha Blair-Goldensohn, Kerry Hannan, and Ryan McDonald. 2010. The viability of Web-derived polarity lexicons. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, HLT ’10, pages 777–785, Stroudsburg, PA. Wainwright, Martin J. and Michael I. Jordan. 2008. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 1(1-2):1–305. Wang, Sida and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (ACL 2012), pages 90–94, Jeju Island. Wiebe, Janyce, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39(2-3):165–210. Williams, Gbolahan K. and Sarabjot Singh Anand. 2009. Predicting the polarity
strength of adjectives using Wordnet. In ICWSM, pages 346–349, San Jose, CA.
Wilson, Theresa, Janyce Wiebe, and Paul Hoffmann. 2009. Recognizing contextual polarity: An exploration of features for phrase-level sentiment analysis. Computational Linguistics, 35:399–433. Yessenalina, Ainur, Yisong Yue, and Claire Cardie. 2010. Multi-level structured models for document-level sentiment classification. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1,046–1,056, Cambridge, MA. Younger, Daniel H. 1967. Recognition and parsing of context-free languages in time n3. Information and Control, 10(2):189–208. Yu, Hong and Vasileios Hatzivassiloglou. 2003. Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, EMNLP ’03, pages 129–136, Stroudsburg, PA. Zelle, John M. and Raymond J. Mooney. 1996. Learning to parse database queries
using inductive logic programming. In AAAI/IAAI, pages 1,050–1,055, Portland, OR.
Zettlemoyer, Luke S. and Michael Collins. 2007. Online learning of relaxed CCG grammars for parsing to logical form. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL-2007), pages 678–687, Prague. Zettlemoyer, Luke S. and Michael Collins. 2009. Learning context-dependent mappings from sentences to logical form. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 2 - Volume 2, ACL ’09, pages 976–984, Stroudsburg, PA. Zhao, Jichang, Li Dong, Junjie Wu, and Ke Xu. 2012. Moodlens: An emoticon-based sentiment analysis system for Chinese tweets. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’12, pages 1,528–1,531, New York, NY.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"a statistical parsing framework for :Sentiment Classification
Li Dong∗, ∗∗ Beihang University
Furu Wei†, ‡ Microsoft Research
Shujie Liu† Microsoft Research
Ming Zhou† Microsoft Research
Ke Xu∗ Beihang University
We present a statistical parsing framework for sentence-level sentiment classification in this article. Unlike previous works that use syntactic parsing results for sentiment analysis, we develop a statistical parser to directly analyze the sentiment structure of a sentence. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way. We formulate the sentiment grammar upon Context-Free Grammars (CFGs), and provide a formal description of the sentiment parsing framework. We develop the parsing model to obtain possible sentiment parse trees for a sentence, from which the polarity model is proposed to derive the sentiment strength and polarity, and the ranking model is dedicated to selecting the best sentiment tree. We train the parser directly from examples of sentences annotated only with sentiment polarity labels but without any syntactic annotations or polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, s.parser, from a large amount of review sentences with users’ ratings as rough sentiment polarity labels. Extensive experiments on existing benchmark data sets show significant improvements over baseline sentiment classification approaches.
∗ State Key Laboratory of Software Development Environment, Beihang University, XueYuan Road No.37, HaiDian District, Beijing, P.R. China 100191. E-mail: donglixp@gmail.com; kexu@nlsde.buaa.edu.cn.
∗∗ Contribution during internship at Microsoft Research. † Natural Language Computing Group, Microsoft Research Asia, Building 2, No. 5 Danling Street, Haidian
District, Beijing, P.R. China 100080. E-mail: {fuwei, shujliu, mingzhou}@microsoft.com. ‡ Corresponding author.
Submission received: 10 December 2013; revised version received: 26 July 2014; accepted for publication: 28 January 2015.
doi:10.1162/COLI a 00221
© 2015 Association for Computational Linguistics",,,,,,,,,,,,,,,,,"1 introduction :Sentiment analysis (Pang and Lee 2008; Liu 2012) has received much attention from both research and industry communities in recent years. Sentiment classification, which identifies sentiment polarity (positive or negative) from text (sentence or document), has been the most extensively studied task in sentiment analysis. Until now, there have been two mainstream approaches for sentiment classification. The lexicon-based approach (Turney 2002; Taboada et al. 2011) aims to aggregate the sentiment polarity of a sentence from the polarity of words or phrases found in the sentence, and the learning-based approach (Pang, Lee, and Vaithyanathan 2002) treats sentiment polarity identification as a special text classification task and focuses on building classifiers from a set of sentences (or documents) annotated with their corresponding sentiment polarity.
The lexicon-based sentiment classification approach is simple and interpretable, but suffers from scalability and is inevitably limited by sentiment lexicons that are commonly created manually by experts. It has been widely recognized that sentiment expressions are colloquial and evolve over time very frequently. Taking tweets from Twitter1 and movie reviews on IMDb2 as examples, people use very casual language as well as informal and new vocabulary to comment on general topics and movies. In practice, it is not feasible to create and maintain sentiment lexicons to capture sentiment expressions with high coverage. On the other hand, the learning-based approach relies on large annotated samples to overcome the vocabulary coverage and deals with variations of words in sentences. Human ratings in reviews (Maas et al. 2011) and emoticons in tweets (Davidov, Tsur, and Rappoport 2010; Zhao et al. 2012) are extensively used to collect a large number of training corpora to train the sentiment classifier. However, it is usually not easy to design effective features to build the classifier. Among others, unigrams have been reported as the most effective features (Pang, Lee, and Vaithyanathan 2002) in sentiment classification.
Handling complicated expressions delivering people’s opinions is one of the most challenging problems in sentiment analysis. Compositionalities such as negation, intensification, contrast, and their combinations are typical cases. We show some concrete examples here:
(1) The movie is not good. [negation]
(2) The movie is very good. [intensification]
(3) The movie is not funny at all. [negation + intensification]
(4) The movie is just so so, but i still like it. [contrast]
(5) The movie is not very good, but i still like it. [negation + intensification + contrast]
The negation expressions, intensification modifiers, and the contrastive conjunction can change the polarity (Examples (1), (3), (4), (5)), strength (Examples (2), (3), (5)), or both (Examples (3), (5)) of the sentiment of the sentences. We do not need any detailed explanations here as they can be commonly found and easily understood in people’s
1 http://twitter.com. 2 http://www.imdb.com.
daily lives. Existing works to address these issues usually rely on syntactic parsing results either used as features (Choi and Cardie 2008; Moilanen, Pulman, and Zhang 2010) in learning-based methods or hand-crafted rules (Moilanen and Pulman 2007; Jia, Yu, and Meng 2009; Klenner, Petrakis, and Fahrni 2009; Liu and Seneff 2009) in lexiconbased methods. However, even with the difficulty and feasibility of deriving the sentiment structure from syntactic parsing results put aside, it is an even more challenging task to generate stable and reliable parsing results for text that is ungrammatical in nature and has a high ratio of out-of-vocabulary words. The accuracy of the linguistic parsers trained on standard data sets (e.g., the Penn Treebank [Marcus, Marcinkiewicz, and Santorini 1993]) drops dramatically on user-generated-content (reviews, tweets, etc.), which is actually the prime focus of sentiment analysis algorithms. The error, unfortunately, will propagate downstream in the process of sentiment analysis methods building upon parsing results.
We therefore propose directly analyzing the sentiment structure of a sentence. The nested structure of sentiment expressions can be naturally modeled in a similar fashion as statistical syntactic parsing, which aims to find the linguistic structure of a sentence. This idea creates many opportunities for developing sentiment classifiers from a new perspective. The most challenging problem and barrier in building a statistical sentiment parser lies in the acquisition of training data. Ideally, we need examples of sentences annotated with polarity for the whole sentence as well as sentiment tags for constituents within a sentence, as with the Penn TreeBank for training traditional linguistic parsers. However, this is not practical as the annotations will be inevitably time-consuming and require laborious human efforts. Therefore, it is better to learn the sentiment parser only utilizing examples annotated with the polarity label of the whole sentence. For example, we can collect a huge number of publicly available reviews and rating scores on the Web. People may use the movie is gud (“gud” is a popular informal expression of “good”) to express a positive opinion towards a movie, and not a fan to express a negative opinion. Also, we can find review sentences such as The movie is gud, but I am still not a fan to indicate a negative opinion. We can then use these two fragments and the overall negative opinion of the sentence to deduce sentiment rules automatically from data. These sentiment fragments and rules can be used to analyze the sentiment structure for new sentences.
In this article, we propose a statistical parsing framework to directly analyze the structure of a sentence from the perspective of sentiment analysis. Specifically, we formulate a Context-Free Grammar (CFG)–based sentiment grammar. We then develop a statistical parser to derive the sentiment structure of a sentence. We leverage the CYK algorithm (Cocke 1969; Younger 1967; Kasami 1965) to conduct bottom–up parsing, and use dynamic programming to accelerate computation. Meanwhile, we propose using the polarity model to derive sentiment strength and polarity of a sentiment parse tree, and the ranking model to select the best one from the sentiment parsing results. We train the parser directly from examples of sentences annotated with sentiment polarity labels instead of syntactic annotations and polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, named s.parser, from a large number of review sentences with users’ ratings as rough sentiment polarity labels. The statistical parsing–based approach builds a principled and scalable framework to support the sentiment composition and inference which cannot be well handled by bag-of-words approaches. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way.
The major contributions of the work presented in this article are as follows.r We propose a statistical parsing framework for sentiment analysis that is capable of analyzing the sentiment structure for a sentence. This framework can naturally handle compositionality in a probabilistic way. It can be trained from sentences annotated with only sentiment polarity but without any syntactic annotations or polarity annotations of constituents within sentences.r We present the parsing model, polarity model, and ranking model in the proposed framework, which are formulated and can be improved independently. It provides a principled and flexible approach to sentiment classification.r We implement the statistical sentiment parsing framework, and conduct experiments on several benchmark data sets. The experimental results show that the proposed framework and algorithm can significantly outperform baseline methods.
The remainder of this article is organized as follows. We introduce related work in Section 2. We present the statistical sentiment parsing framework, including the parsing model, polarity model, and ranking model, in Section 3. Learning methods for our model are explained in Section 4. Experimental results are reported in Section 5. We conclude this article with future work in Section 6. 2 related work :In this section, we give a brief introduction to related work about sentiment classification (Section 2.1) and parsing (Section 2.2). We tackle the sentiment classification problem in a parsing manner, which is a significant departure from most previous research. Sentiment classification has been extensively studied in the past few years. In terms of text granularity, existing works can be divided into phrase-level, sentence-level, or document-level sentiment classification. We focus on sentence-level sentiment classification in this article. Regardless of what granularity the task is performed on, existing approaches deriving sentiment polarity from text fall into two major categories, namely, lexicon-based and learning-based approaches.
The lexicon-based sentiment analysis uses dictionary matching on a predefined sentiment lexicon to derive sentiment polarity. These methods often use a set of manually defined rules to deal with the negation of polarity. Turney (2002) proposed using the average sentiment orientation of phrases, which contains adjectives or adverbs, in a review to predict its sentiment orientation. Yu and Hatzivassiloglou (2003) calculated a modified log-likelihood ratio for every word by the co-occurrences with positive and negative seed words. To determine the polarity of a sentence, they compare the average log-likelihood value with threshold. Taboada et al. (2011) presented a lexicon-based approach for extracting sentiment from text. They used dictionaries of words with annotated sentiment orientation (polarity and strength) while incorporating intensification and negation. The lexicon-based methods often achieve high precisions and do not need
any labeled samples. But they suffer from coverage and domain adaption problems. Moreover, lexicons are often built and used without considering the context (Wilson, Wiebe, and Hoffmann 2009). Also, hand-crafted rules are often matched heuristically.
The sentiment dictionaries used for lexicon-based sentiment analysis can be created manually, or automatically using seed words to expand the list of words. Kamps et al. (2004) and Williams and Anand (2009) used various lexical relations (such as synonym and antonym relations) in WordNet to expend a set of seed words. Some other methods learn lexicons from data directly. Hatzivassiloglou and McKeown (1997) used a log-linear regression model with conjunction constraints to predict whether conjoined adjectives have similar or different polarities. Combining conjunction constraints across many adjectives, a clustering algorithm separated the adjectives into groups of different polarity. Finally, adjectives were labeled as positive or negative. Velikovich et al. (2010) constructed a term similarity graph using the cosine similarity of context vectors. They performed graph propagation from seeds on the graph, obtaining polarity words and phrases. Takamura, Inui, and Okumura (2005) regarded the polarity of words as spins of electrons, using the mean field approximation to compute the approximate probability function of the system instead of the intractable actual probability function. Kanayama and Nasukawa (2006) used tendencies for similar polarities to appear successively in contexts. They defined density and precision of coherency to filter neutral phrases and uncertain candidates. Choi and Cardie (2009a) and Lu et al. (2011) transformed the lexicon learning to an optimization problem, and used integer linear programming to solve it. Kaji and Kitsuregawa (2007) defined the χ2-based polarity value and PMI-based polarity value as a polarity strength to filter neutral phrases. de Marneffe, Manning, and Potts (2010) utilized review data to define polarity strength as the expected rating value. Mudinas, Zhang, and Levene (2012) used word count as a feature template and trained a classifier using Support Vector Machines with linear kernel. They then regarded the weights as polarity strengths. Krestel and Siersdorfer (2013) generated topicdependent lexicons from review articles by incorporating topic and rating probabilities and defined the polarity strength based on the results. In this article, the lexical relations defined in WordNet are not used because of its coverage. Furthermore, most of these methods define different criteria to propagate polarity information of seeds, or use optimization algorithms and sentence-level sentiment labels to learn polarity strength values. Their goal is to balance the precision and recall of learned lexicons. We also learn the polarity strength values of phrases from data. However, our primary objective is to obtain correct sentence-level polarity labels, and use them to form the sentiment grammar.
Learning-based sentiment analysis uses machine learning methods to classify sentences or documents into two (negative and positive) or three (negative, positive, and neutral) classes. Previous research has shown that sentiment classification is more difficult than traditional topic-based text classification, despite the fact that the number of classes in sentiment classification is smaller than that in topic-based text classification (Pang and Lee 2008). Pang, Lee, and Vaithyanathan (2002) investigated three machine learning methods to produce automated classifiers to generate class labels for movie reviews. They tested them on Naı̈ve Bayes, Maximum Entropy, and Support Vector Machine (SVM), and evaluated the contribution of different features including unigrams, bigrams, adjectives, and part-of-speech tags. Their experimental results suggested that a SVM classifier with unigram presence features outperforms other competitors. Pang and Lee (2004) separated subjective portions from the objective by finding minimum cuts in graphs to achieve better sentiment classification performance. Matsumoto, Takamura, and Okumura (2005) used text mining techniques to
extract frequent subsequences and dependency subtrees, and used them as features of SVM. McDonald et al. (2007) investigated a global structured model for jointly classifying polarity at different levels of granularity. This model allowed classification decisions from one level in the text to influence decisions at another. Yessenalina, Yue, and Cardie (2010) used sentence-level latent variables to improve document-level prediction. Täckström and McDonald (2011a) presented a latent variable model for only using document-level annotations to learn sentence-level sentiment labels, and Täckström and McDonald (2011b) improved it by using a semi-supervised latent variable model to utilize manually crafted sentence labels. Agarwal et al. (2011) and Tu et al. (2012) explored part-of-speech tag features and tree-kernel. Wang and Manning (2012) used SVM built over Naı̈ve Bayes log-count ratios as feature values to classify polarity. They showed that SVM was better at full-length reviews, and Multinomial Naı̈ve Bayes was better at short-length reviews. Liu, Agam, and Grossman (2012) proposed a set of heuristic rules based on dependency structure to detect negations and sentimentbearing expressions. Most of these methods are built on bag-of-words features, and sentiment compositions are handled by manually crafted rules. In contrast to these models, we derive polarity labels from tree structures parsed by the sentiment grammar.
There have been several attempts to assume that the problem of sentiment analysis is compositional. Sentiment classification can be solved by deriving the sentiment of a complex constituent (sentence) from the sentiment of small units (words and phrases) (Moilanen and Pulman 2007; Klenner, Petrakis, and Fahrni 2009; Choi and Cardie 2010; Nakagawa, Inui, and Kurohashi 2010). Moilanen and Pulman (2007) proposed using delicate written linguistic patterns as heuristic decision rules when computing the sentiment from individual words to phrases and finally to the sentence. The manually compiled rules were powerful enough to discriminate between the different sentiments in effective remedies (positive) / effective torture (negative), and in too colorful (negative) and too sad (negative). Nakagawa, Inui, and Kurohashi (2010) leveraged a conditional random field model to calculate the sentiment of all the parsed elements in the dependency tree and then generated the overall sentiment. It had an advantage over the rule-based approach (Moilanen and Pulman 2007) in that it did not explicitly denote any sentiment designation to words or phrases in parse trees. Instead, it modeled their sentiment polarity as latent variables with a certain probability of being positive or negative. Councill, McDonald, and Velikovich (2010) used a conditional random field model informed by a dependency parser to detect the scope of negation for sentiment analysis. Some other methods model sentiment compositionality in the vector space. They regard the composition operator as a matrix, and use matrix-vector multiplication to obtain the transformed vector representation. Socher et al. (2012) proposed a recursive neural network model that learned compositional vector representations for phrases and sentences. Their model assigned a vector and a matrix to every node in a parse tree. The vector captured the inherent meaning of the constituent, and the matrix captured how it changes the meaning of neighboring words or phrases. Socher et al. (2013) recently introduced a sentiment treebank based on the results of the Stanford parser (Klein and Manning 2003). The sentiment treebank included polarity labels of phrases that are annotated using Amazon Mechanical Turk. The authors trained recursive neural tensor networks on the sentiment treebank. For a new sentence, the model predicted polarity labels based on the syntactic parse tree, and used tensors to handle compositionality in the vector space. Dong et al. (2014) proposed utilizing multiple composition functions in recursive neural models and learning to select them adaptively. Most previous methods are either rigid in terms of handcrafted rules, or sensitive to the performance of existing syntactic parsers they use. This article addresses sentiment
compositions by defining sentiment grammar and borrowing some techniques in the parsing research field. Moreover, our method uses symbolic representations instead of vector spaces. The work presented in this article is close to traditional statistical parsing, as we borrow some algorithms to build the sentiment parser. Syntactic parsers are learned from the Treebank corpora, and find the most likely parse tree with the largest probability. In this article, we borrow some well-known techniques from syntactic parsing methods (Charniak 1997; Charniak and Johnson 2005; McDonald, Crammer, and Pereira 2005; Kübler, McDonald, and Nivre 2009), such as the CYK algorithm and Context-Free Grammar. These techniques are used to build the sentiment grammar and parsing model. They provide a natural way of defining the structure of sentiment trees and parse sentences to trees. The key difference lies in that our task is to calculate the polarity label of a sentence, instead of obtaining the parse tree. We only have sentencepolarity pairs as our training instances instead of annotated tree structures. Moreover, in the decoding process, our goal is to compute correct polarity labels by representing sentences as latent sentiment trees. Recently, Hall, Durrett, and Klein (2014) developed a discriminative constituency parser using rich surface features, adapting it to sentiment analysis. Besides extracting unigrams and bigrams as features, they learned interactions between tags and words located at the beginning or the end of spans. However, their method relies on phrase-level polarity annotations.
Semantic parsing is another body of work related to this article. A semantic parser is used to parse meaning representations for given sentences. Most existing semantic parsing works (Zelle and Mooney 1996; Kate and Mooney 2006; Raymond and Mooney 2006; Zettlemoyer and Collins 2007, 2009; Li, Liu, and Sun 2013) relied on fine-grained annotations of target logical forms, which required the supervision of experts and are relatively expensive. To balance the performance and the amount of human annotation, some works used only question-answer pairs or even binary correct/incorrect signals as their input. Clarke et al. (2010) used a binary correct/incorrect signal of a database query to map sentences to logical forms. It worked with FunQL language and transformed semantic parsing as an integer linear programming (ILP) problem. In each iteration, it solved ILP and updated the parameters of structural SVM. Liang, Jordan, and Klein (2013) learned a semantic parser from question-answer pairs, where the logical form was modeled as a latent tree-based semantic representation. Krishnamurthy and Mitchell (2012) presented a method for training a semantic parser using a knowledge base and an unlabeled text corpus, without any individually annotated sentences. Artzi and Zettlemoyer (2013) used various types of weak supervision to learn a grounded Combinatory Categorial Grammar semantic parser, which took context into consideration. Bao et al. (2014) presented a translation-based weakly supervised semantic parsing method to translate questions to answers based on CYK parsing. A log-linear model is defined to score derivations. All these weakly supervised semantic parsing methods learned to transform a natural language sentence to its semantic representation without annotated logical form. In this work, we build a sentiment parser. Specifically, we use a modified version of the CYK algorithm that parses sentences in a bottom–up fashion. We use the log-linear model to score candidates generated by beam search. Instead of using question-answer pairs, sentence-polarity pairs are used as our weak supervisions. We also use the parameter estimation algorithm proposed by Liang, Jordan, and Klein (2013). 3 statistical sentiment parsing :We present the statistical parsing framework for sentence-level sentiment classification in this section. The underlying idea is to model sentiment classification as a statistical parsing process. Figure 1 shows the overview of the statistical sentiment parsing framework. There are three major components. The input sentence s is transformed into and represented by sentiment trees derived from the parsing model (Section 3.2), using the sentiment grammar defined in Section 3.1. Trees are scored by the ranking model in Section 3.3. The sentiment tree with the highest ranking score is treated as the best derivation for s. Furthermore, the polarity model (Section 3.4) is used to compute polarity values for the sentiment trees.
Notably, the sentiment trees t are unobserved during training. We can only observe the sentence s and its polarity label y in training data. In other words, we train the model directly from the examples of sentences annotated only with sentiment polarity labels but without any syntactic annotations or polarity annotations of the constituents within sentences. To be specific, we first learn the sentiment grammar and the polarity model from data as described in Section 4.2. Then, given the sentence and polarity label pairs( s, y ) , we search the latent sentiment trees t and estimate the parameters of the ranking model as detailed in Section 4.1. To better illustrate the whole process, we describe the sentiment parsing procedure using an example sentence, The movie is not very good, but i still like it. The sentiment polarity label of the above sentence is “positive.” There is negation, intensification, and contrast in this example, which are difficult to capture using bag-of-words classification methods. This sentence is a complex case that demonstrates the capability of the proposed statistical sentiment parsing framework, which motivates the work in this article. The statistical sentiment parsing algorithm may generate a number of sentiment trees for the input sentence. Figure 2 shows the best sentiment parse tree. It shows that the statistical sentiment parsing framework can deal with the compositionality of sentiment in a natural way. In Table 1, we list the sentiment rules used during the parsing process. We show the generation process of the sentiment parse tree from the bottom–up and the calculation of sentiment strength and polarity for every text span in the parsing process.
the derivation process include {P→ the movie is; P→ good; P→ i still like it; P→ very P; N→ not P; N→ PN; N→ NE ; E → ,; P→ N but P; S→ P}.
In the following sections, we first provide a formal description of the sentiment grammar in Section 3.1. We then present the details of the parsing model in Section 3.2, the ranking model in Section 3.3, and the polarity model in Section 3.4. We develop the sentiment grammar upon CFG (Context-Free Grammar) (Chomsky 1956). Let G =< V, Σ, S, R > denote a CFG, where V is a finite set of non-terminals, Σ is a finite set of terminals (disjointed from V), S ∈ V is the start symbol, and R is a set of rewrite rules (or production rules) of the form A→ c where A ∈ V and c ∈ (V ∪ Σ)∗. We use Gs =< Vs, Σs, S, Rs > to denote the sentiment grammar in this article.
The non-terminal set is denoted as Vs = {N, P, S, E}, where S is the start symbol, the non-terminal N represents the negative polarity, and the non-terminal P represents the positive polarity. The rules in Rs are divided into the following six categories:
r Dictionary rules: X→ wk0, where X ∈ {N, P}, wk0 = w0 . . .wk−1, and wk0 ∈ Σ+s . These rules can be regarded as the sentiment dictionary used in traditional approaches. They are basic sentiment units assigned with polarity probabilities. For instance, P→ good is a dictionary rule.r Combination rules: X→ c, where c ∈ (Vs ∪ Σs)+, and two successive non-terminals are not allowed. There is at least one terminal in c. These rules combine terminals and non-terminals, such as N→ not P, and P→ N but P. They are used to handle negation, intensification, and contrast in sentiment analysis. The number of non-terminals in a combination rule is restricted to one and two.r Glue rules: X→ X1X2, where X, X1, X2 ∈ {N, P}. These rules combine two text spans that are derived into X1 and X2, respectively.r OOV rules: E → wk0, where wk0 ∈ Σ+. We use these rules to handle Out-Of-Vocabulary (OOV) text spans whose polarity probabilities are not learned from data.r Auxiliary rules: X→ EX1, X→ X1E , where X, X1 ∈ {N, P}. These rules combine a text span with polarity and an OOV text span.r Start rules: S→ Y, where Y ∈ {N, P, E}. The derivations begin with S, and S can be derived to N, P, and E .
Here, X represents the non-terminals N or P. The dictionary rules and combinations rules are automatically extracted from the data. We will describe the details in Section 4.2. By applying these rules, we can derive the polarity label of a sentence from the bottom–up. The glue rules are used to combine polarity information of two text spans together, and it treats the combined parts as independent. In order to tackle the OOV problem, we treat a text span that consists of OOV words as empty text span, and derive them to E . The OOV text spans are combined with other text spans without considering their sentiment information. Finally, each sentence is derived to the symbol S using the start rules that are the beginnings of derivations. We can use the sentiment grammar to compactly describe the derivation process of a sentence. We present the formal description of the statistical sentiment parsing model following deductive proof systems (Shieber, Schabes, and Pereira 1995; Goodman 1999) as used in traditional syntactic parsing. For a concrete example,
(A→ BC) [i, B, k] [k, C, j] [i, A, j] (6)
which represents if we have the rule A→ BC and B ∗⇒ wki and C ∗⇒ w jk ( ∗⇒ is used to represent the reflexive and transitive closure of immediate derivation), then we can obtain A ∗⇒ w ji . By adding a unary rule
(A→ w ji) [i, A, j] (7)
with the binary rule in Equation (6), we can express the standard CYK algorithm for CFG in Chomsky Normal Form (CNF). And the goal is [0, S, n], in which S is the start symbol and n is the length of the input sentence. In the given CYK example, the term in deductive rules can be one of the following two forms:r [i, X, j] is an item representing a subtree rooted in X spanning from i to j, orr (X→ γ) is a rule in the grammar.
Generally, we represent the form of an inference rule as:
(r) H1 . . . HK [i, X, j] (8)
where, if all the terms r and Hk are true, then we can infer [i, X, j] as true. Here, r denotes a sentiment rule, and Hk denotes an item. When we refer to both rules and items, we employ the word terms.
Theoretically, we can convert the sentiment rules to CNF versions, and then use the CYK algorithm to conduct parsing. Because the maximum number of non-terminal symbols in a rule is already restricted to two, we formulate the statistical sentiment parsing based on a customized CYK algorithm that is similar to the work of Chiang (2007). Let X, X1, X2 represent the non-terminals N or P; the inference rules for the statistical sentiment parsing are summarized in Figure 3. The parsing model generates many candidate parse trees T(s) for a sentence s. The goal of the ranking model is to score and rank these parse trees. The sentiment tree with the highest score is treated as the best representation for sentence s. We extract a feature vector φ(s, t) ∈ Rd for the specific sentence-tree pair (s, t), where t ∈ T(s) is the parse tree. Let ψ ∈ Rd be the parameter vector for the features. We use the log-linear model to calculate a probability p(t|s; T,ψ) for each parse tree t ∈ T(s). The probabilities indicate how likely the trees are to produce correct predictions. Given the sentence s and parameters ψ, the log-linear model defines a conditional probability:
p(t|s; T,ψ) = exp {φ(s, t)Tψ−A(ψ; s, T)} (9)
A(ψ; s, T) = log ∑
t∈T(s) exp {φ(s, t)Tψ} (10)
(X→ w ji) [i, X, j]
(X→ wi1i X1w j j1
) [i1, X1, j1] [i, X, j]
(X→ wi1i X1w i2 j1 X2w j j2 ) [i1, X1, j1] [i2, X2, j2]
[i, X, j]
(X→ X1X2) [i, X1, k] [k, X2, j] [i, X, j]
(E → w ji ) [i, E , j]
(X→ EX1) [i, E , k] [k, X1, j] [i, X, j]
(X→ X1E ) [i, X1, k] [k, E , j] [i, X, j]
where X, X1, X2 represent N or P.
Figure 3 Inference rules for the basic parsing model.
where A(ψ; s, T) is the log-partition function with respect to T(s). The log-linear model is a discriminative model, and it is widely used in natural language processing. We can use φ(s, t)Tψ as the score of the parse tree without normalization in the decoding process, because p(t|s; T,ψ) ∝ φ(s, t)Tψ, and this will not change the ranking order. The goal of the polarity model is to model the calculation of sentiment strength and polarity of a text span from its subspans in the parsing process. It is specified in terms of the rules used in the parsing process. We expand the notations in the inference rule (8) to incorporate the polarity model. The new form of inference rule is:
(r) H1Φ1 . . . HKΦK [i, X, j]Φ (11)
in which r, H1, . . . , HK are the terms described in Section 3.2. Every item Hk is assigned
polarity strength Φk : { P(N |w jkik ) P(P|w jkik ) for text span w jkik . For the item [i, X, j], the polarity model Φ(r, Φ1, . . . , ΦK ) is defined as a function that takes the rule r and polarity strength of subspans as input.
The polarity strength obtained by the polarity model should satisfy two constraints. First, the values calculated by the polarity model are non-negative, that is, P(X |w ji) ≥ 0, P(X |w j i) ≥ 0. Second, the positive and negative polarity values are normal-
ized to 1, namely, P(X |w ji) + P(X |w j i) = 1. Notably, X = { P , X = N N , X = P is the opposite
polarity of X . The inference rules with the polarity model are formally defined in Figure 4. In the following part, we define the polarity model for the different types of rules. If the rule is a dictionary rule X→ w ji, its sentiment strength is obtained as:
Φ :
{ P(X |w ji) = P̃(X |w j i)
P(X |w ji) = P̃(X |w j i)
(12)
where X ∈ {N ,P} denotes the sentiment polarity of the left hand side of the rule, X is the opposite polarity of X , and P̃(X |w ji), P̃(X |w j i) indicate the sentiment polarity values estimated from training data.
The glue rules X→ X1X2 combine two spans (w ki , w j k). The polarity value is calcu-
lated by their product, and normalized to 1.
Φ : P(X |w j i) =
P(X |w ki )P(X |w j k )
P(X |w ki )P(X |w j k )+P(X |w k i )P(X |w j k )
P(X |w ji) = 1− P(X |w j i)
(13)
For OOV text spans, the polarity model does not calculate the polarity values. When they are combined with in-vocabulary phrases by the auxiliary rules, the polarity values are determined by the text span with polarity and the OOV text span is ignored. More specifically,
Φ :
{ P(X |w ji) = P(X |w k i )
P(X |w ji) = P(X |w k i )
(14)
The combination rules are more complicated than other types of rules. In this article, we model the polarity probability calculation as the logistic regression. The logistic regression can be regarded as putting linear combination of the subspans’ polarity probabilities into a logistic function (or sigmoid function). We will show that the negation, intensification, and contrast can be well modeled by the regression-based method. It is formally shown as
P(X |w ji) = h ( θ0 + K∑ k=1 θkP(Xk|w jk ik ) )
= 1 1 + exp { − ( θ0 + ∑K k=1 θkP(Xk|w jk ik ) )} (15)
where h(x) = 11+exp{−x} is the logistic function, K is the number of non-terminals in a rule, and θ0, . . . ,θK are the parameters that are learned from data. As a concrete example, if the span w ji can match N→ not P and P
∗⇒ w ji+1, the inference rule with the polarity model is defined as
N→ not P [i + 1, P, j]Φ1
[i, N, j]
{ P(N |w ji) = h(θ0 + θ1P(P|w j i+1))
P(P|w ji) = 1− P(N |w j i)
(16)
where polarity probability is calculated by P(N |w ji) = h(θ0 + θ1P(P|w j i+1)).
To tackle negation, switch negation (Choi and Cardie 2008; Saurı́ 2008) simply reverses the sentiment polarity and corresponding sentiment strength. However, consider not great and not good; flipping polarity directly makes not good more positive than not great, which is unreasonable. Another potential problem of switch negation is that negative polarity items interact with intensifiers in undesirable ways (Kennedy and Inkpen 2006). For example, not very good turns out to be even more negative than not good, given the fact that very good is more positive than good. Therefore, Taboada et al. (2011) argue that shift negation is a better way to handle polarity negation. Instead of reversing polarity strength, shift negation shifts it toward the opposite polarity by
a fixed amount. This method can partially avoid the aforementioned two problems. However, they set the parameters manually, which might not be reliable and extensible enough to a new data set. Using the regression model, switch negation is captured by the negative scale item θk (k > 0), and shift negation is expressed by the shift item θ0.
The intensifiers are adjectives or adverbs that strengthen (amplifier) or decrease (downtoner) the semantic intensity of its neighboring item (Quirk 1985). For example, extremely good should obtain higher strength of positive polarity than good, because it is modified by the amplifier (extremely). Polanyi and Zaenen (2006) and Kennedy and Inkpen (2006) handle intensifiers by polarity addition and subtraction. This method, termed fixed intensification, increases a fixed amount of polarity for amplifiers and decreases for downtoners. Taboada et al. (2011) propose a method, called percentage intensification, to associate each intensification word with a percentage scale, which is larger than one for amplifiers, and less than one for downtoners. The regression model can capture these two methods to handle the intensification. The shift item θ0 represents the polarity addition and subtraction directly, and the scale item θk (k > 0) can scale the polarity by a percentage.
Table 2 illustrates how the regression based polarity model represents different negation and intensification methods. For a specific rule, the parameters and the compositional method are automatically learned from data (Section 4.2.3) instead of setting them manually as in previous work (Taboada et al. 2011). In a similar way, this method can handle the contrast. For example, the inference rule for N→ P but N is:
(N→ P but N) [i1, P, j1]Φ1 [i2, N, j2]Φ2
[i, N, j]
{ P(N |w ji) = h(θ0 + θ1P(P|w j1 i1 ) + θ2P(N |w j2 i2 ))
P(P|w ji) = 1− P(N |w j i)
(17)
where the polarity probability of the rule N→ P but N is computed by P(N |w ji) = h(θ0 + θ1P(P|w j1 i1 ) + θ2P(N |w j2 i2
)). It can express the contrast relation by specific parameters θ0, θ1, and θ2.
It should be noted that a linear regression model could turn out to be problematic, as it may produce unreasonable results. For example, if we do not add any constraint, we may get P(N |w ji) = −0.6 + P(P|w j i+1). When P(P|w j i+1) = 0.55, we will get P(N |w ji) = −0.6 + 0.55 = −0.05. This conflicts with the definition that the polarity probability ranges from zero to one. Figure 5 intuitively shows that the logistic function truncates polarity values to (0, 1) smoothly.
Parameter Negation Type P(X |w ji) = h(θ0 + θ1P(X |w j1 i1 )) Intensification Type P(X |w ji) = h(θ0 + θ1P(X |w j1 i1 ))
Shift Switch Percentage Fixed
θ0 (Shift item) X X θ1 (Scale item) X X We incorporate additional constraints into the parsing model. Those are used as pruning conditions in the derivation process not only to improve efficiency but also to force the derivation towards the correct direction. We expand the inference rules in Section 3.4 as,
(r) H1Φ1 . . . HKΦK [i, X, j]Φ C (18)
where C is a side condition. The constraints are interpreted in a Boolean manner. If the constraint C is satisfied, the rule can be used, otherwise, it cannot. We define two constraints in the parsing model.
First, in the parsing process, the polarity label of text span w ji obtained by the polarity model (Section 3.4) should be consistent with the non-terminal X (N or P) on the left hand side of the rule. To distinguish between the polarity labels and the non-terminals, we denote the corresponding polarity label of non-terminal X as X . Following this notation, we describe the first constraint as
C1 : P(X |w j i) > P(X |w j i) (19)
where X is the opposite polarity of X . For instance, if rule P→ not N matches the text span w ji, the polarity calculated by the polarity model should be consistent with P, i.e., the polarity obtained by the polarity model should be positive (P).
Second, when we apply the combination rules, the polarity strength of subspans needs to exceed a predefined threshold τ (≥ 0.5). Specifically, for combination rules X→ wi1i X1w i2 j1 X2w j j2 and X→ wi1i X1w j j1 , we define the second constraint as
C2 : P(Xk|w jk ik ) > τ, k = 1, . . . , K (20)
where K is the number of subspans in the rule, and Xk is the corresponding polarity label of non-terminal Xk in the right hand side. If P(Xk|w
jk ik ) is not larger than threshold
τ, we regard the polarity of phrase w jkik as neutral. For instance, we do not want to use the combination rule P→ a lot of P or N→ a lot of N for the phrase a lot of people. This constraint avoids improperly using the combination rules for neutral phrases. Notably, when τ is set as 0.5, this constraint is the same as the first one in Equation (19).
As shown in Figure 6, we add these two constraints to the inference rules. The OOV rules do not have any constraints, and the constraint C1 is applied for all the other rules. The constraint C2 is only applied for the combination rules. In this section, we summarize the decoding algorithm in Algorithm 1. For a sentence s, the CYK algorithm and dynamic programming are used to obtain the sentiment tree with the highest score. To be specific, the modified CYK parsing model parses the input sentence to sentiment trees in a bottom–up manner—that is, from short to long text spans. For every text span w ji, we match the rules in the sentiment grammar (Section 3.1) to generate the candidate set. Their polarity values are calculated using the polarity model described in Section 3.4. We also use the constraints described in Section 3.5 to prune search paths. The constraints improve the efficiency of the parsing algorithm and make derivations that meet our intuition.
Algorithm 1 Decoding Algorithm Input: wn0 : Sentence Output: Polarity of the input sentence
1: score[, , ]← {} 2: for l← 1 . . .n do . Modified CYK algorithm 3: for all i, j s.t. j− i = l do 4: for all inferable rule (r) H1...HK[i,X,j] for w j i do 5: Φ← calculate polarity value for r . Polarity model 6: if constraints are satisfied then . Constraint 7: sc← compute score for this derivation by ranking model . Ranking
model 8: if sc > score[i, j, X] then 9: score[i, j, X]← sc
10: return arg maxX∈{N ,P} score[0, X, n]
The features in the ranking model (Section 4.1.1) decompose along the structure of the sentiment tree. So the dynamic programming technique can be used to compute the derivation tree with the highest ranking score. For a span, the scores of its subspans are used to calculate the local scores of its derivations. For example, the score of the derivation (r) [i1,P,j1] [i2,N,j2][i,X,j] is score[i1, j1, P] + score[i2, j2, N] + score r, where score[i, j, X] is the highest score of text span w ji that is derived to the non-terminal X, and score r is the score of applying the rule r. As described in Section 3.3, the score of using rule r is scorer = φ(w ji, r) T ψ, where φ(w ji, r) is the feature vector of using the rule r for the span w ji, and ψ is the weight vector of the ranking model. The k highest score trees satisfying the constraints are stored in score[, , ] for decoding the longer text spans. After finishing the CYK parsing, arg maxX∈{N ,P} score[0, n, X] is regarded as the polarity label of input sentence. The time complexity is the same as the standard CYK’s. 4 model learning :We described the statistical sentiment parsing framework in Section 3. We present the model learning process in this section. The learning process consists of two steps. First, the sentiment grammar and the polarity model are learned from data. In other words, the rules and the parameters used to compute polarity values are learned. These basic sentiment building blocks are then used to build the parse trees. Second, we estimate the parameters of the ranking model using the sentence and polarity label pairs. At this stage, we concentrate on learning how to score the parse trees based on the learned sentiment grammar and polarity model.
Section 4.1 shows the features and the parameter estimation algorithm used in the ranking model. Section 4.2 illustrates how to learn the sentiment grammar and the polarity model. As shown in Section 3.3, we develop the ranking model on the log-linear model. In the following subsections, we first present the features used to rank sentiment tree
candidates. Then, we describe the objective function used in the optimization algorithm. Finally, we introduce the algorithm for parameter estimation using the gradient-based method.
4.1.1 Features. We extract a feature vector φ(s, t) ∈ Rd for each parse tree t of sentence s. The feature vector is used in the log-linear model. In Figure 7, we present the features extracted for the sentence The movie is not very good, but i still like it. The features are organized into feature templates. Each of them contains a set of features. These feature templates are shown as follows:r COMBHIT: This feature is the total number of combination rules used in t.r COMBRULE: It contains features {COMBRULE[r] : r is a combination rule},
each of which fires on the combination rule r appearing in t.r DICTHIT: This feature is the total number of dictionary rules used in t.r DICTRULE: It contains features {DICTRULE[r] : r is a dictionary rule}, each of which fires on the dictionary rule r appearing in t.
These features are generic local patterns that capture the properties of the sentiment tree. Another intuitive lexical feature template is [combination rule + word]. For instance, P→ very P(good) is a feature that lexicalizes the non-terminal P to good. However, if this feature is fired frequently, the phrase very good would be learned as a dictionary rule and can be used in the decoding process. So we do not use this feature template in order to reduce the feature size. It should be noted that these features
decompose along structures of sentiment trees, enabling us to use dynamic programming in the CYK algorithm.
4.1.2 Objective Function. We design the ranking model upon the log-linear model to score candidate sentiment trees. In the training dataD, we only have the input sentence s and its polarity label Ls. The forms of sentiment parse trees, which can obtain the correct sentiment polarity, are unobserved. So we work with the marginal log-likelihood of obtaining the correct polarity label Ls,
log p(Ls|s; T,ψ) = log p(t ∈ TLs (s)|s; T,ψ)
= A(ψ; s, TLs )−A(ψ; s, T) (21)
where TLs is the set of candidate trees whose prediction labels are Ls, and A(ψ; s, T) (Equation (10)) is the log-partition function with respect to T(s).
Based on the marginal log-likelihood function, the objective function O(ψ, T) consists of two terms. The first term is the sum of marginal log-likelihood over training instances that can obtain the correct polarity labels. The second term is a L2-norm regularization term on the parameters ψ. Formally,
O(ψ, T) = ∑
(s,Ls )∈D TLs (s)6=∅
log p(Ls|s; T,ψ)− λ2‖ψ‖ 2 2 (22)
To learn the parameters ψ, we use a gradient-based optimization method to maximize the objective function O(ψ, T). According to Wainwright and Jordan (2008), the derivative of the log-partition function is the expected feature vector
∂O(ψ, T) ∂ψ
= ∑
(s,Ls )∈D TLs (s)6=∅
(Ep(t|s;TLs ,ψ)[φ(s, t)]− Ep(t|s;T,ψ)[φ(s, t)])− λψ (23)
where Ep(x)[f (x)] = ∑ x p(x)f (x) for discrete x.
4.1.3 Parameter Estimation. The objective function O(ψ, T) is not concave (nor convex), hence the optimization potentially results in a local optimum. Stochastic Gradient Descent (SGD; Robbins and Monro 1951) is a widely used optimization method. The SGD algorithm picks up a training instance randomly, and updates the parameter vector ψ according to
ψj (t+1) = ψj (t) + α ( ∂O(ψ) ∂ψj |ψ=ψ(t) )
(24)
where α is the learning rate, and ∂O(ψ)∂ψj is the gradient of the objective function with respect to parameter ψj. The SGD is sensitive to α, and the learning rate is the same for all dimensions. As described in Section 4.1.1, we mix sparse features together with dense features. We want the learning rate to be different for each dimension. We use AdaGrad (Duchi, Hazan, and Singer 2011) to update the parameters, which sets an adaptive per-feature learning rate. The AdaGrad algorithm tends to use smaller update
steps when we meet a feature many times. In order to compute efficiently, a diagonal approximation version of AdaGrad is used. The update rule is
ψj (t+1) = ψj (t) + α 1√ G(t+1)j
( ∂O(ψ) ∂ψj |ψ=ψ(t) )
G(t+1)j = G (t) j + ( ∂O(ψ) ∂ψj |ψ=ψ(t) )2 (25)
where we introduce an adaptive term G(t)j . G (t) j becomes larger along with updating, and decreases the update step for dimension j. Compared with SGD, the only cost is to store and update G(t)j for each parameter.
To train the model, we use the method proposed by Liang, Jordan, and Klein (2013). With the candidate parse trees and objective function, the parameters ψ are updated to make the parsing model favor correct trees and give them a higher score. Because there are many parse trees for a sentence, we need to calculate Equation (23) efficiently. As indicated in Section 4.1.1, the features decompose along the structure of sentiment trees. So dynamic programming can be utilized to compute Ep(t|s;T,ψ)[φ(s, t)] of Equation (23). However, the first expectation term Ep(t|s;TLs ,ψ)[φ(s, t)] sums over the candidates that obtain the correct polarity labels. As this constraint does not decompose along the tree structure, there is no efficient dynamic program for this. Instead of searching all the parse trees spanning s, we use beam search to approximate this expectation. Beam search is a best-first search algorithm that explores at most K paths (K is the beam size). It keeps the local optimums to reduce the huge search space. Specifically, the beam search algorithm generates the K-best trees with the highest score φ(s, t)Tψ for each span. These local optimums are used recursively in the CYK process. The K-best trees for the whole span are regarded as the candidate set T̃. Then T̃ and T̃Ls are used to approximate Equation (23) as in Liang, Jordan, and Klein (2013).
The intuition behind this parameter estimation algorithm lies in: (1) if we have better parameters, we can obtain better candidate trees; (2) with better candidate trees, we can learn better parameters. Thus the optimization problem is solved in an iterative manner. We initialize the parameters as zeros. This leads to a random search and generates random candidate trees. With the initial candidates, the two steps in Algorithm 2 lead the parameters ψ towards the direction achieving better performance. In this section, we present the automatic learning of the sentiment grammar as defined in Section 3.1. We need to extract the dictionary rules and the combination rules from data. In traditional statistical parsing, grammar rules are induced from annotated parse trees (such as the Penn TreeBank), so ideally we need examples of sentiment structure trees, or sentences annotated with sentiment polarity for the whole sentence as well as those for constituents within sentences. However, this is not practical, if not unfeasible, as the annotations will be inevitably time consuming and require laborious human effort. We show that it is possible to induce the sentiment grammar directly from examples of sentences annotated with sentiment polarity labels without using any syntactic annotations or polarity annotations of constituents within sentences. The
Algorithm 2 Ranking Model Learning Algorithm Input: D: Training data {(s,Ls)}, S: Maximum number of iteration Output: ψ: Parameters of the ranking model
1: ψ(0)← (0, 0, . . . , 0)T 2: repeat 3: (s,Ls)← randomly select a training instance in D 4: T̃(t)← BEAMSEACH(s,ψ(t)) . Beam search to generate K-best candidates
5: G(t+1)j ← G (t) j + ( ∂O(ψ,T̃(t) ) ∂ψj |ψ=ψ(t) )2 6: ψ (t+1) j ← ψj
(t) + α 1√ G(t+1)j ( ∂O(ψ,T̃(t) ) ∂ψj |ψ=ψ(t) )
. Update parameters using
AdaGrad 7: t← t + 1 8: until t > S 9: return ψ(T)
sentences annotated with sentiment polarity labels are relatively easy to obtain, and we use them as our input to learn dictionary rules and combination rules.
We first present the basic idea behind the algorithm we propose. People are likely to express positive or negative opinions using very simple and straightforward sentiment expressions again and again in their reviews. Intuitively, we can mine dictionary rules from these massive review sentences by leveraging the redundancy characteristics. Furthermore, there are many complicated reviews that contain complex sentiment structures (e.g., negation, intensification, and contrast). If we already have dictionary rules on hand, we can use them to obtain basic sentiment information for the fragments within complicated reviews. We can then extract combination rules with the help of the dictionary rules and the sentiment polarity labels of complicated reviews. Because the simple and straightforward sentiment expressions are often coupled with complicated expressions, we need to conduct dictionary rule mining and the combination rule mining in an iterative way.
4.2.1 Dictionary Rule Learning. The dictionary rules GD are basic sentiment building blocks used in the parsing process. Each dictionary rule in GD is in the form X→ f , where f is a sentiment fragment. We use the polarity probabilities P(N | f ) and P(P| f ) in the polarity model. To build GD, we regard all the frequent fragments whose occurrence frequencies are larger than τf and lengths range from 1 to 7 as the sentiment fragments. We further filter the phrases formed by stop words and punctuations, which are not used to express sentiment.
For a balanced data set, the sentiment distribution of a candidate sentiment fragment f is calculated by
P(X | f ) = #( f,X ) + 1 #( f,N ) + #( f,P ) + 2 (26)
where X ∈ {N ,P}, and #( f,X ) denotes the number of reviews containing f with X being the polarity. It should be noted that Laplace smoothing is used in Equation (26) to deal with the zero frequency problem.
We do not learn the polarity probabilities P(N | f ) and P(P| f ) by directly counting occurrence frequency. For example, in the review sentence this movie is not good (negative), the naive counting method increases the count #(good,N ) in terms of the polarity of the whole sentence. Moreover, because of the common collocation not as good as (negative) in movie reviews, as good as is also regarded as negative if we count the frequency directly. The examples indicate why some polarity probabilities of phrases counting from data are different from our intuitions. These unreasonable polarity probabilities also make trouble for learning the polarity model. Consequently, in order to estimate more reasonable probabilities, we need to take the compositionality into consideration when learning sentiment fragments.
Following this motivation, we ignore the count #( f,X ), if the sentiment fragment f is covered by a negation rule r that negates the polarity of f . The word cover here means that f is derived within a non-terminal of the negation rule r. For instance, the negation rule N→ not P covers the sentiment fragment good in the sentence this is not a good movie (negative), that is, the good is derived from P of this negation rule. So we ignore the occurrence for #(good,N ) in this sentence. It should be noted that we still increase the count for #(not good,N ), because there is no negation rule covering the fragment not good.
As shown in Algorithm 3, we learn the dictionary rules and their polarity probabilities by counting the frequencies in negative and positive classes. Only the fragments whose occurrence numbers are larger than threshold τf are kept. Moreover, we take the combination rules into consideration to acquire more reasonable GD. Notably, a subsequence of a frequent fragment must also be frequent. This is similar to the key insight in the Apriori algorithm (Agrawal and Srikant 1994). When we learn the dictionary rules, we can count the sentiment fragments from short to long, and prune the infrequent fragments in the early stages if any subsequence is not frequent. This pruning method accelerates the dictionary rule learning process and makes the procedure fit in memory.
Algorithm 3 Dictionary Rule Learning Input: D: Data set, GC : Combination rules, τf : Frequency threshold Output: GD: Dictionary rules
1: function MINEDICTIONARYRULES(D,GC) 2: GD′← {} 3: for (s,Ls) in D do . s : w0w1 · · ·w|s|−1, Ls: Polarity label of s 4: for all i, j s.t. 0 ≤ i < j ≤ |s| do . w ji : wiwi+1 · · ·wj−1 5: if no negation rule in GC covers w j i then 6: #(w ji,Ls) ++ 7: add w ji to GD ′ 8: GD ← {} 9: for f in GD′ do 10: if #( f, ·) ≥ τf then 11: compute P(N | f ) and P(P| f ) using Equation (26) 12: add dictionary rule (Lf → f ) to GD . Lf = arg maxX∈{N,P} P(X | f ) 13: return GD
4.2.2 Combination Rule Learning. The combination rules GC are generalizations for the dictionary rules. They are used to handle the compositionality and process unseen phrases. The learning of combination rules is based on the learned dictionary rules and their polarity values. The sentiment fragments are generalized to combination rules by replacing the subsequences of dictionary rules with their polarity labels. For instance, as shown in Figure 8, the fragments is not (good/as expected/funny/well done) are all negative. After replacing the subspans good, as expected, funny, and well done with their polarity label P, we can learn the negation rule N→ is not P.
We present the combination rule learning approach in Algorithm 4. Specifically, the first step is to generate combination rule candidates. For every subsequence w ji of sentiment fragment f , we replace it with the corresponding non-terminal Lw ji
if P(Lw ji |w ji) is larger than the threshold τp, and we can get w i 0Lw ji
w| f |j . Next, we compare the polarity Lw ji with Lf . If Lf 6= Lw ji , we regard the rule Lf → w i0Lw ji
w| f |j as a negation rule. Otherwise, we further compare their polarity values. If this rule makes the polarity value become larger (or smaller), it will be treated as a strengthen (or weaken) rule. To obtain the contrast rules, we replace two subsequences with their polarity labels in a similar way. If the polarities of these two subsequences are different, we categorize this rule to the contrast type. Notably, these two non-terminals cannot be next to each other. After these steps, we get the rule candidate set GC ′ and the occurrence number of each rule. We then filter the rule candidates whose occurrence frequencies are too small, and assign the rule types (negation, strengthen, weaken, and contrast) according to their occurrence numbers.
4.2.3 Polarity Model Learning. As shown in Section 3.4, we define the polarity model to calculate the polarity probabilities using the sentiment grammar. In this section, we present how to learn the parameters of the polarity model for the combination rules.
,(
P(P|funny), P(N |is not funny)
)
, (P(P|well done), and P(N |is not well done)) are used to
learn the polarity model for N→ is not P.
Algorithm 4 Combination Rule Learning Input: D: Data set, GD: Dictionary rules, τp, τ∆, τr, τc: Thresholds Output: GC : Combination rules
1: function MINECOMBINATIONRULES(D,GD) 2: GC ′← {} 3: for (X→ f ) in GD do . f : w0w1 · · ·w| f |−1 4: for all i, j s.t. 0 ≤ i < j ≤ | f | do 5: if P(Lw ji |w ji) > τp then . Polarity label Lw ji = arg maxX∈{N ,P} P(X |w j i) 6: r: X→ wi0Lw ji w| f |j . Non-terminal Lw ji = arg maxX∈{N,P} P(X |w j i) 7: if X 6= Lw ji then 8: #(r, negation) ++ 9: else if P(X | f ) > P(Lw ji
|w ji) + τ∆ then 10: #(r, strengthen) ++ 11: else if P(X | f ) < P(Lw ji
|w ji)− τ∆ then 12: #(r, weaken) ++ 13: add r to GC ′ 14: for all i0, j0, i1, j1 s.t. 0 ≤ i0 < j0 < i1 < j1 ≤ | f | do 15: if P(L
w j0i0 |w j0i0 ) > τp and P(Lw j1i1 |w j1i1 ) > τp then
16: r: X→ wi00 Lw j0i0 wi1j0 Lw j1i1 w | f |j1 . Replace w j0 i0 , w j1i1 with the non-terminals 17: if L w j0i0 6= L w j1 i1 then 18: #(r, contrast) ++ 19: add r to GC ′ 20: GC ← {} 21: for r in GC ′ do 22: if #(r, ·) > τr and max
T
#(r,T) #(r) > τc then
23: add r to GC 24: return GC
As shown in Figure 8, we learn combination rules by replacing the subsequences of frequent sentiment fragments with their polarity labels. Both the replaced fragment and the whole fragment can be found in the dictionary rules, so their polarity probabilities have been estimated from data. We can use them as our training examples to figure out how context changes the polarity of replaced fragment, and learn parameters of the polarity model.
We describe the polarity model in Section 3.4. To further simplify the notation, we denote the input vector x = (1, P(X1|w j1 i1 ), . . . , P(XK|w jK iK
))T, and the response value as y. Then we can rewrite Equation (15) as
hθ(x) = 11 + exp{−θTx} (27)
where hθ(x) is the polarity probability calculated by the polarity model, and θ = (θ0,θ1, . . . ,θK )T is the parameter vector. Our goal is to estimate the parameter vector θ of the polarity model.
We fit the model to minimize the sum of squared residuals between the predicted polarity probabilities and the values computed from data. We define the cost function as
J (θ) = 12 ∑
m
(hθ(xm)− ym)2 (28)
where ( xm, ym ) is the m-th training instance.
The gradient descent algorithm is used to minimize the cost function J (θ). The partial derivative of J (θ) with respect to θj is
∂J (θ) ∂θj
= ∑
m
( hθ(xm)− ym ) ∂hθ(xm) ∂θj
= ∑
m
( hθ(xm)− ym ) hθ(xm) ( 1− hθ(xi) ) ∂θTxm ∂θj
= ∑
m
( hθ(xm)− ym ) hθ(xm) (1− hθ(xm)) xmj
(29)
We set the initial θ as zeros, and start with it. We use the Stochastic Gradient Descend algorithm to minimize the cost function. For the instance (x, y), the parameters are updated using
θj (t+1) = θj (t) − α ( ∂J (θ) ∂θj |θ=θ(t) )
= θj (t) − α(hθ(t) (x)− y)hθ(t) (x) ( 1− hθ(t) (x) ) xj
(30)
where α is the learning rate, and it is set to 0.01 in our experiments. We summarize the learning method in Algorithm 5. For each combination rule, we iteratively scan
Algorithm 5 Polarity Model Learning Algorithm Input: GC : Combination rules, ε: Stopping condition, α: Learning rate Output: θ: Parameters of the polarity model
1: function ESTIMATEPOLARITYMODEL(GC) 2: for all combination rule r ∈ GC do 3: θ(0)← (0, 0, ..., 0)T 4: repeat 5: ( x, y ) ← randomly select a training instance 6: θj (t+1)← θj(t) − α(hθ(t) (x)− y)hθ(t) (x) ( 1− hθ(t) (x) ) xj 7: t← t + 1 8: until
∥∥θ(t+1) − θ(t)∥∥22 < ε 9: assign θ(T) as the parameters of the polarity model for rule r
Algorithm 6 Sentiment Grammar Learning Input: D: Data set {(s,Ls)}, T: Maximum number of iteration . Ls: Polarity label of s Output: GD: Dictionary rules, GC : Combination rules
1: GC ← {} 2: repeat 3: GD ←MINEDICTIONARYRULES(D,GC) . Algorithm 3 4: GC ←MINECOMBINATIONRULES(D,GD) . Algorithm 4 5: until iteration number exceeds T 6: ESTIMATEPOLARITYMODEL(GC) . Algorithm 5 7: return GD,GC
through the training examples ( x, y )
in a random order, and update the parameters θ according to Equation (30). The stopping condition is ∥∥θ(t+1) − θ(t)∥∥22 < ε, which indicates the parameters become stable.
4.2.4 Summary of the Grammar Learning Algorithm. We summarize the grammar learning process in Algorithm 6, which learns the sentiment grammar in an iterative manner.
We first learn the dictionary rules and their polarity probabilities by counting the frequencies in negative and positive classes. Only the fragments whose occurrence numbers are larger than the threshold τf are kept. As mentioned in Section 4.2.1, the context can essentially change the distribution of sentiment fragments. We take the combination rules into consideration to acquire more reasonable GD. In the first iteration, the set of combination rules is empty. Therefore, we have no information about compositionality to improve dictionary rule learning. The initial GD contains some inaccurate sentiment distributions. Next, we replace the subsequences of dictionary rules to their polarity labels, and generalize these sentiment fragments to the combination rules GC as illustrated in Section 4.2.2. At the same time, we can obtain their compositional types and learn parameters of the polarity model. We iterate over these two steps to obtain refined GD and GC . 5 experimental studies :In this section, we describe experimental results on existing benchmark data sets with extensive comparisons with state-of-the-art sentiment classification methods. We also present the effects of different experimental settings in the proposed statistical sentiment parsing framework. We describe the data sets in Section 5.1.1, the experimental settings in Section 5.1.2, and the methods used for comparison in Section 5.1.3.
5.1.1 Data Sets. We conduct experiments on sentiment classification for sentence-level and phrase-level data. The sentence-level data sets contain user reviews and critic
reviews from Rotten Tomatoes3 and IMDb.4 We balance the positive and negative instances in the training data set to mitigate the problem of data imbalance. Moreover, the Stanford Sentiment Treebank5 contains polarity labels of all syntactically plausible phrases. In addition, we use the MPQA6 data set for the phrase-level task. We describe these data sets as follows.
RT-C: 436,000 critic reviews from Rotten Tomatoes. It consists of 218,000 negative and 218,000 positive critic reviews. The average review length is 23.2 words. Critic reviews from Rotten Tomatoes contain a label (Rotten: Negative, Fresh: Positive) to indicate the polarity, which we use directly as the polarity label of corresponding reviews.
PL05-C: The sentence polarity data set v1.0 (Pang and Lee 2005) contains 5,331 positive and 5,331 negative snippets written by critics from Rotten Tomatoes. This data set is widely used as the benchmark data set in the sentence-level polarity classification task. The data source is the same as RT-C.
SST: The Stanford Sentiment Treebank (Socher et al. 2013) is built upon PL05-C. The sentences are parsed to parse trees. Then, 215,154 syntactically plausible phrases are extracted and annotated by workers from Amazon Mechanical Turk. The experimental settings of positive/negative classification for sentences are the same as in Socher et al. (2013).
RT-U: 737,806 user reviews from Rotten Tomatoes. Because we focus on sentencelevel sentiment classification, we filter out user reviews that are longer than 200 characters. The average length of these short user reviews from Rotten Tomatoes is 15.4 words. Following previous work on polarity classification, we use the review score to select highly polarized reviews. For the user reviews from Rotten Tomatoes, a negative review has a score <2.5 out of 5, and a positive review has a score >3.5 out of 5.
IMDB-U: 600,000 user reviews from IMDb. The user reviews in IMDb contain comments and short summaries (usually a sentence) to summarize the overall sentiment expressed in the reviews. We use the review summaries as the sentence-level reviews. The average length is 6.6 words. For user reviews of IMDb, a negative review has a score <4 out of 10, and a positive review has a score >7 out of 10.
C-TEST: 2,000 labeled critic reviews sampled from RT-C. We use C-TEST as the testing data set for RT-C. Note that we exclude these from the training data set (i.e., RT-C).
U-TEST: 2,000 manually labeled user reviews sampled from RT-U. User reviews often contain some noisy ratings compared with critic reviews. To eliminate the effect of noise, we sample 2,000 user reviews from RT-U, and annotate their polarity labels manually. We use U-TEST as a testing data set for RT-U and IMDB-U, which are both user reviews. Note that we exclude them from the training data set (i.e., RT-U).
MPQA: The opinion polarity subtask of the MPQA data set (Wiebe, Wilson, and Cardie 2005). The authors manually annotate sentiment polarity labels for the expressions (i.e., sub-sentences) within a sentence. We regard the expressions as short sentences in our experiments. There are 7,308 negative examples and 3,316 positive examples in this data set. The average number of words per example is 3.1.
3 http://www.rottentomatoes.com. 4 http://www.imdb.com. 5 http://nlp.stanford.edu/sentiment/treebank.html. 6 http://mpqa.cs.pitt.edu/corpora/mpqa corpus.
Table 3 shows the summary of these data sets, and all of them are publicly available at http://goo.gl/WxTdPf.
5.1.2 Settings. To compare with other published results for PL05-C and MPQA, the training and testing regime (10-fold cross-validation) is the same as in Pang and Lee (2005), Nakagawa, Inui, and Kurohashi (2010) and Socher et al. (2011). For SST, the regime is the same as in Socher et al. (2013). We use C-TEST as the testing data for RT-C, and U-TEST as the testing data for RT-U and IMDB-U. There are a number of settings that have trade-offs in performance, computation, and the generalization power of our model. The best settings are chosen by a portion of training split data that serves as the validation set. We provide the performance comparisons using different experimental settings in Section 5.4.
Number of training examples: The size of training data has been widely recognized as one of the most important factors in machine learning-based methods. Generally, using more data leads to better performance. By default, all the training data is used in our experiments. We use the same size of training data in different methods for fair comparisons.
Number of training iterations (T): We use AdaGrad (Duchi, Hazan, and Singer 2011) as the optimization algorithm in the learning process. The algorithm starts with randomly initialized parameters, and alternates between searching candidate sentiment trees and updating parameters of the ranking model. We treat one-pass scan of training data as an iteration.
Beam size (K): The beam size is used to make a trade-off between the search space and the computation cost. Moreover, an appropriate beam size can prune unfavorable candidates. We set K = 30 in our experiments.
Regularization (λ): The regularization parameter λ in Equation (22) is used to avoid over-fitting. The value used in the experiments is 0.01.
Minimum fragment frequency: It is difficult to estimate reliable polarity probabilities when the fragment appears very few times. Hence, a minimum fragment frequency that is too small will introduce noise in the fragment learning process. On the other hand, a large threshold will lose much useful information. The minimum fragment frequency is chosen according to the size of the training data set and the validation performance. To be specific, we set this parameter as 4 for RT-C, SST, RT-U, and IMDB-U, and 2 for PL05-C and MPQA.
Maximum fragment length: High order n-grams are more precise and deterministic expressions than unigrams and bigrams. So it would be useful to use long fragments to capture polarity information. According to the experimental results, as the maximum fragment length increases, the accuracy of sentiment classification increases. The maximum fragment length is set to 7 words in our experiments.
5.1.3 Sentiment Classification Methods for Comparison. We evaluate the proposed statistical sentiment parsing framework on the different data sets, and compare the results with some baselines and state-of-the-art sentiment classification methods described as follows.
SVM-m: Support Vector Machine (SVM) achieves good performance in the sentiment classification task (Pang and Lee 2005). Though unigrams and bigrams are reported as the most effective features in existing work (Pang and Lee 2005), we use high-order n-gram (1 ≤ n ≤ m) features to conduct fair comparisons. Hereafter, m has the same meaning. We use LIBLINEAR (Fan et al. 2008) in our experiments because it can handle well the high feature dimension and a large number of training examples. We try different hyper-parameters C ∈ {10−2, 10−1, 1, 5, 10, 20} for SVM, and select C on the validation set.
MNB-m: As indicated in Wang and Manning (2012), Multinomial Naı̈ve Bayes (MNB) often outperforms SVM for sentence-level sentiment classification. We utilize Laplace smoothing (Manning, Raghavan, and Schütze 2008) to tackle the zero probability problem. High order n-gram (1 ≤ n ≤ m) features are considered in the experiments.
LM-m: Language Model (LM) is a generative model calculating the probability of word sequences. It is used for sentiment analysis in Cui, Mittal, and Datar (2006). Probability of generating sentence s is calculated by P(s) = ∏|s|−1 i=0 P ( wi|wi−10 ) , where wi−10 denotes the word sequence w0 . . .wi−1. We use Good-Turing smoothing (Good 1953) to overcome sparsity when estimating the probability of high-order n-gram. We train language models on negative and positive sentences separately. For a sentence, its polarity is determined by comparing the probabilities calculated from the positive and negative language models. The unknown-word token is treated as a regular word (denoted by<UNK>). The SRI Language Modeling Toolkit (Stolcke 2002) is used in our experiment.
Voting-w/Rev: This approach is proposed by Choi and Cardie (2009b), and is used as a baseline in Nakagawa, Inui, and Kurohashi (2010). The polarity of a subjective sentence is decided by the voting of each phrase’s prior polarity. The polarity of phrases that have odd numbers of negation phrases in their ancestors is reversed. The results are reported by Nakagawa, Inui, and Kurohashi (2010).
HardRule: This baseline method is compared by Nakagawa, Inui, and Kurohashi (2010). The polarity of a subjective sentence is deterministically decided based on rules, by considering the sentiment polarity of dependency subtrees. The polarity of a modifier is reversed if its head phrase has a negation word. The decision rules are applied from the leaf nodes to the root node in a dependency tree. We use the results reported by Nakagawa, Inui, and Kurohashi (2010).
Tree-CRF: Nakagawa, Inui, and Kurohashi (2010) present a dependency tree-based method using conditional random fields with hidden variables. In this model, the polarity of each dependency subtree is represented by a hidden variable. The value of the hidden variable of the root node is identified as the polarity of the whole sentence. The experimental results are reported by Nakagawa, Inui, and Kurohashi (2010).
RAE-pretrain: Socher et al. (2011) introduce a framework based on recursive autoencoders to learn vector space representations for multi-word phrases and predict sentiment distributions for sentences. We use the results with pre-trained word vectors learned on Wikipedia, which leads to better results compared with randomized word vectors. We directly compare the results with those in Socher et al. (2011).
MV-RNN: Socher et al. (2012) try to capture the compositional meaning of long phrases through matrix-vector recursive neural networks. This model assigns a vector and a matrix to every node in the parse tree. Matrices are regarded as operators, and vectors capture the meaning of phrases. The results are reported by Socher et al. (2012, 2013).
s.parser-LongMatch: The longest matching rules are utilized in the decoding process. In other words, the derivations that contain the fewest rules are used for all text spans. In addition, the dictionary rules are preferred to the combination rules if both of them match the same text span. The dynamic programming algorithm is used in the implementation.
s.parser-w/oComb: This is our method without using the combination rules (such as N→ not P) learned from data. We present the experimental results of the sentiment classification methods on the different data sets in Table 4. The top three methods on each data set are in bold, and the best methods are also underlined. The experimental results show that s.parser achieves better performance than other methods on most data sets.
The data sets RT-C, PL05-C, and SST are critic reviews. On RT-C, the accuracy of s.parser increases by 2 percentage points, 2.9 percentage points, and 7.1 percentage points from the best results of SVM, MNB, and LM, respectively. On PL05-C, the accuracy of s.parser also rises by 2.1 percentage points, 0.7 percentage points, and 4.4 percentage points from the best results of SVM, MNB, and LM, respectively. Compared to Voting-w/Rev and HardRule, s.parser outperforms them by 16.4 percentage points and 16.6 percentage points. The results indicate that our method significantly outperforms the baselines that use manual rules, as rule-based methods lack a probabilistic way to model the compositionality of context. Furthermore, s.parser achieves an accuracy improvement rate of 2.2 percentage points, 1.8 percentage points, and 0.5 percentage points over Tree-CRF, RAE-pretrain, and MV-RNN, respectively. On SST, s.parser outperforms SVM, MNB, and LM by 3.4 percentage points, 1.4 percentage points, and 3.8 percentage points, respectively. The performance is better than MV-RNN with an improvement rate of 1.8 percentage points. Moreover, the result is comparable to the 85.4% obtained by recursive neural tensor networks (Socher et al. 2013) without depending on syntactic parsing results.
On the user review data sets RT-U and IMDB-U, our method also achieves the best results. More specifically, on the data set RT-U, s.parser outperforms the best results of SVM, MNB, and LM by 1.7 percentage points, 2.9 percentage points, and 1.5 percentage points, respectively. On the data set IMDB-U, our method brings an improved accuracy rate of 2.1 percentage points, 3.7 percentage points, and 2.2 percentage points over SVM, MNB, and LM, respectively. We find that MNB performs better than SVM and LM on the critics review data sets RT-C and PL05-C. Also, SVM and LM achieve better results on the user review data sets RT-U and IMDB-U. The s.parser is more robust for the different genres of data sets.
On the data set MPQA, the accuracy of s.parser increases by 0.5 percentage points, 1.1 percentage points, and 14.8 percentage points from the best results of SVM, MNB, and LM, respectively. Compared with Voting-w/Rev and HardRule, s.parser achieves 4.5 percentage point and 4.4 percentage point improvements over them. As illustrated in Table 3, the size and length of sentences in MPQA are much smaller than those in the other four data sets. The RAE-pretrain achieves better results than other methods on this data set, because the word embeddings pre-trained on Wikipedia can leverage smoothing to relieve the sparsity problem in MPQA. If we do not use any external resources (i.e., Wikipedia), the accuracy of RAE on MPQA is 85.7%, which is lower than Tree-CRF and s.parser. The results indicate that s.parser achieves the best result if no external resource is used.
In addition, we compare the results of s.parser-LongMatch and s.parser-w/oComb. The s.parser-LongMatch utilizes the dictionary rules and combination rules in the longest matching manner, whereas s.parser-w/oComb removes the combination rules in the parsing process. Compared with the results of s.parser, we find that both the ranking model and the combination rules play a positive role in the model. The ranking model learns to score parse trees by assigning larger weights to the rules that tend to obtain correct labels. Also, the combination rules generalize these dictionary rules to deal with the sentiment compositionality in a symbolic way, which enables the model to process unseen phrases. Furthermore, s.parser-LongMatch achieves better results than s.parser-w/oComb. This indicates that the effects of the combination rules are more pronounced than the ranking model.
The bag-of-words classifiers work well for long documents relying on sentiment words that appear many times in a document. The redundancy characteristics provide strong evidence for sentiment classification. Even though some phrases of a document are not estimated accurately, it can still result in a correct polarity label. However, for short text, such as a sentence, the compositionality plays an important role in sentiment classification. Tree-CRF, MV-RNN, and s.parser take compositionality into consideration in different ways, and they achieve significant improvements over SVM, MNB, and LM. We also find that the high order n-grams contribute to classification accuracy on most of the data sets, but they harm the accuracy of LM on PL05-C. The high-order n-grams can partially solve compositionality in a brute-force way. We further investigate the effect of the size of training data for different sentiment classification methods. This is meaningful as the number of the publicly available reviews is increasing dramatically nowadays. The methods that can take advantage of more training data will be even more useful in practice.
We report the results of s.parser compared with SVM, MNB, and LM on the data set RT-C using different training data size. In order to make the figure clear, we only present the results of SVM/MNB/LM-1/5 here. As shown in Figure 9, we find that the size of training data plays an important role for all these sentiment classification methods. The basic conclusion is that the performance of all the methods rise as the data size increases, especially when the data size is smaller than a certain number. It meets our intuition that the size of data is the key factor when the size is relatively small. When the size of data is larger, the growth of accuracy becomes slower. The performance of the baseline methods starts to converge after the data size is larger than 200,000. The comparisons illustrate that s.parser significantly outperforms these baselines. And the performance of s.parser becomes even better when the data size increases. The convergence of s.parser’s performance is slower than the others. It indicates that s.parser leverages data more effectively and benefits more from a larger data set. With more training data, s.parser learns more dictionary rules and combination rules. These rules enhance the generalization ability of our model. Furthermore, it estimates more reliable parameters for the polarity model and ranking model. In contrast, the bag-of-words based approaches (such as SVM, MNB, and LM) cannot make full use of high-order information in the data set. The generalization ability of the combination rules of s.parser leads to better performance, and take advantage of larger data. It should be noted that there are similar trends with the other data sets. In this section, we investigate the effects of different experimental settings. We show the results on the data set RT-C by only changing one factor and fixing the others.
Figure 10 shows the effect of minimum fragment frequency, and maximum fragment length. Specifically, Figure 10a indicates that a minimum fragment frequency that is too small will introduce noise, and it is difficult to estimate reliable polarity probabilities for infrequent fragments. However, a minimum fragment frequency that is too large will discard too much useful information. As shown in Figure 10b, we find that accuracy increases as the maximum fragment length increases. The results illustrate that the large maximum fragment length is helpful for s.parser. We can learn more
combination rules with a larger maximum fragment length, and long dictionary rules capture more precise expressions than unigrams. This conclusion is the same as that in Section 5.2.
As shown in Figure 11, we also investigate how the training iteration, regularization, and beam size affect the results. As shown in Figure 11a, we try a wide range of regularization parameters λ in Equation (22). The results indicate that it is insensitive to the choice of λ. Figure 11b shows the effects of different beam size K in the search process. When beam size K = 1, the optimization algorithm cannot learn the weights. In this case, the decoding process is to select one search path randomly, and compute its polarity probabilities. The results become better as the beam size K increases. On the other hand, the computation costs increase. The proper beam size K can prune some candidates to speed up the search procedure. It should be noted that the sentence length also effects the run time. The sentiment grammar plays a central role in the statistical sentiment parsing framework. It is obvious that the accuracy of s.parser relies on the quality of the automatically learned sentiment grammar. The quality can be implicitly evaluated by the accuracy of sentiment classification results, as we have shown in previous sections. However, there is no straightforward way to explicitly evaluate the quality of the learned grammar. In this section, we provide several case studies of the learned dictionary rules and combination rules to further illustrate the results of the sentiment grammar learning process as detailed in Section 4.2.
To start with, we report the total number of dictionary rules and combination rules learned from the data sets. As shown in Table 5, the results indicate that we can learn more dictionary rules and combination rules from the larger data sets. Although we learn more dictionary rules from RT-C than from IMDB-U, the number of combination rules learned from RT-C is less than from IMDB-U. It indicates that the language usage of RT-C is more diverse than that of IMDB-U. For SST, more rules are learned because of its constituent-level annotations.
Furthermore, we explore how the minimum fragment frequency τf affects the number of dictionary rules, and present the distribution of dictionary rule length. As illustrated in Figure 12a, we find that the relation between total number of dictionary rules |GD| and minimum fragment frequency τf obeys the power law, that is, the log10(|GD|)− log2(τf ) graph takes a linear form. It indicates that most of the fragments appear few times, and only some of them appear frequently. Notably, all the syntactically plausible phrases of SST are annotated, so its distribution is different from the other sentence-level data sets. Figure 12b shows the cumulative distribution of dictionary rule length l. It presents most dictionary rules as short ones. For all data sets except SST, more than 80% of dictionary rules are shorter than five words. The length distributions of data sets RT-C and IMDB-U are similar, whereas we obtain more high order n-grams from RT-U and SST.
We further investigate the effect of context for dictionary rule learning. Table 6 shows some dictionary rules with polarity probabilities learned by our method and
naive counting on RT-C. We notice that if we count the fragment occurrence number directly, some polarities of fragments are learned incorrectly. This is caused by the effect of context as described in Section 4.2.1. By taking the context into consideration, we obtain more reasonable polarity probabilities of dictionary rules. Our dictionary rule learning method takes compositionality into consideration, namely, we skip the count if there exist some negation indicators outside the phrase. This constraint tries to ensure that the polarity of fragments is the same as the whole sentence. As shown in the results, the polarity probabilities learned by our method are more reasonable and meet people’s intuitions. However, there are also some negative examples caused by “false subjective.” For instance, the neutral phrase to pay it tends to appear in negative sentences, and it is learned as a negative phrase. This makes sense for the data distribution, but it may lead to the mismatch for the combination rules.
In Figure 13, we show the polarity model of some combination rules learned from the data set RT-C. The first two examples are negation rules. We find that both switch negation and shift negation exist in data, instead of using only one negation type in previous work (Choi and Cardie 2008; Saurı́ 2008; Taboada et al. 2011). For the rule “N→ i do not P,” we find that it is a switch negation rule. This rule reverses the polarity and the corresponding polarity strength. For instance, the i do not like it very much is more negative than the i do not like it. As shown in Figure 13b, the “N→ is not P” is a shift negation that reduces a fixed polarity strength to reverse the original polarity. Specifically, the is not good is more negative than the is not great, as described in Section 3.4. We have a similar conclusion for the next two weaken rules. As illustrated in Figure 13c, the “P→ P actress” describes one aspect of a movie, hence it is more likely to decrease the polarity intensity. We find that this rule is a fixed intensification rule that reduces the polarity probability by a fixed value. The “N→ a bit of N” is a percentage intensification rule, which scales polarity intensity by a percentage. It reduces more strength for stronger polarity. The last two rules in Figure 13e and Figure 13f are strengthen rules. Both “P→ lot of P” and “N→ N terribly” increase the polarity strength of the sub-fragments. These cases indicate that it is necessary to learn how the context performs compositionality from data. In order to capture the compositionality for different rules, we define the polarity model and learn parameters for each rule. This also agrees with the models of Socher et al. (2012) and Dong et al. (2014),
which use multiple composition matrices to make compositions specific and is an improvement over the recursive neural network that employs one composition matrix. 6 conclusion and future work :In this article, we propose a statistical parsing framework for sentence-level sentiment classification that provides a novel approach to designing sentiment classifiers from a new perspective. It directly analyzes the sentiment structure of a sentence other than relying on syntactic parsing results, as in existing literature. We show that complicated phenomena in sentiment analysis, such as negation, intensification, and contrast, can be handled in a similar manner to simple and straightforward sentiment expressions in a unified and probabilistic way. We provide a formal model to represent the sentiment grammar built upon Context-Free Grammars. The framework consists of: (1) a parsing model to analyze the sentiment structure of a sentence; (2) a polarity model to calculate sentiment strength and polarity for each text span in the parsing process; and (3) a ranking model to select the best parsing result from a list of candidate sentiment parse trees. We show that the sentiment parser can be trained from the examples of sentences annotated only with sentiment polarity labels but without using any syntactic or sentiment annotations within sentences. We evaluate the proposed framework on standard sentiment classification data sets. The experimental results show that the statistical sentiment parsing notably outperforms the baseline sentiment classification approaches.
We believe the work on statistical sentiment parsing can be advanced from many different perspectives. First, statistical parsing has been a well-established research field, in which many different grammars and parsing algorithms have been proposed in previously published literature. It will be an interesting direction to apply and adjust more advanced models and algorithms from the syntactic parsing and the semantic parsing to our framework. We leave it as a line of future work. Second, we can incorporate target and aspect information in the statistical sentiment parsing framework to facilitate the target-dependent and aspect-based sentiment analysis. Intuitively, this can be done by introducing semantic tags of targets and aspects as new non-terminals in the sentiment grammar and revising grammar rules accordingly. However, acquiring training data will be an even more challenging task, as we need more fine-grained information. Third, as the statistical sentiment parsing produces more fine-grained information (e.g., the basic sentiment expressions from the dictionary rules as well as the sentiment structure trees), we will have more opportunities to generate better opinion summaries. Moreover, we are interested in jointly learning parameters of the polarity model and the parsing model from data. Last but not the least, we are interested in investigating the domain adaptation, which is a very important and challenging problem in sentiment analysis. Generally, we may need to learn domain-specific dictionary rules for different domains, whereas we believe combination rules are mostly generic across different domains. This is also worth consideration for further study. We present a statistical parsing framework for sentence-level sentiment classification in this article. Unlike previous works that use syntactic parsing results for sentiment analysis, we develop a statistical parser to directly analyze the sentiment structure of a sentence. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way. We formulate the sentiment grammar upon Context-Free Grammars (CFGs), and provide a formal description of the sentiment parsing framework. We develop the parsing model to obtain possible sentiment parse trees for a sentence, from which the polarity model is proposed to derive the sentiment strength and polarity, and the ranking model is dedicated to selecting the best sentiment tree. We train the parser directly from examples of sentences annotated only with sentiment polarity labels but without any syntactic annotations or polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, s.parser, from a large amount of review sentences with users’ ratings as rough sentiment polarity labels. Extensive experiments on existing benchmark data sets show significant improvements over baseline sentiment classification approaches. [{""affiliations"": [], ""name"": ""Li Dong""}, {""affiliations"": [], ""name"": ""Furu Wei""}, {""affiliations"": [], ""name"": ""Shujie Liu""}, {""affiliations"": [], ""name"": ""Ming Zhou""}, {""affiliations"": [], ""name"": ""Ke Xu""}] SP:bfc3292de4a5066ae9bf2d380c6ef61c6c870698 [{""authors"": [""Agarwal"", ""Apoorv"", ""Boyi Xie"", ""Ilia Vovsha"", ""Owen Rambow"", ""Rebecca Passonneau.""], ""title"": ""Sentiment analysis of twitter data"", ""venue"": ""Proceedings of the Workshop on Languages in Social Media,"", ""year"": 2011}, {""authors"": [""Agrawal"", ""Rakesh"", ""Ramakrishnan Srikant.""], ""title"": ""Fast algorithms for mining association rules in large databases"", ""venue"": ""Proceedings of the 20th International Conference on Very Large Data Bases,"", ""year"": 1994}, {""authors"": [""Artzi"", ""Yoav"", ""Luke Zettlemoyer.""], ""title"": ""Weakly supervised learning of semantic parsers for mapping instructions to actions"", ""venue"": ""Transactions of the Association for Computational Linguistics,"", ""year"": 2013}, {""authors"": [""Bao"", ""Junwei"", ""Nan Duan"", ""Ming Zhou"", ""Tiejun Zhao.""], ""title"": ""Knowledge-based question answering as machine translation"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association"", ""year"": 2014}, {""authors"": [""Charniak"", ""Eugene.""], ""title"": ""Statistical parsing with a context-free grammar and word statistics"", ""venue"": ""Proceedings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative"", ""year"": 1997}, {""authors"": [""Charniak"", ""Eugene"", ""Mark Johnson.""], ""title"": ""Coarse-to-fine n-best parsing and maxent discriminative reranking"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL \u201905,"", ""year"": 2005}, {""authors"": [""Chiang"", ""David.""], ""title"": ""Hierarchical phrase-based translation"", ""venue"": ""Computational Linguistics, 33(2):201\u2013228."", ""year"": 2007}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Learning with compositional semantics as structural inference for subsentential sentiment analysis"", ""venue"": ""Proceedings of the Conference on Empirical Methods in Natural Language"", ""year"": 2008}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification"", ""venue"": ""Proceedings of the 2009 Conference on Empirical Methods in"", ""year"": 2009}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification"", ""venue"": ""Proceedings of the 2009 Conference on Empirical"", ""year"": 2009}, {""authors"": [""Choi"", ""Yejin"", ""Claire Cardie.""], ""title"": ""Hierarchical sequential learning for extracting opinions and their attributes"", ""venue"": ""Proceedings of the ACL 2010 Conference Short Papers, ACLShort \u201910, pages 269\u2013274,"", ""year"": 2010}, {""authors"": [""Chomsky"", ""Noam.""], ""title"": ""Three models for the description of language"", ""venue"": ""IRE Transactions on Information Theory, 2(3):113\u2013124."", ""year"": 1956}, {""authors"": [""Clarke"", ""James"", ""Dan Goldwasser"", ""Ming-Wei Chang"", ""Dan Roth.""], ""title"": ""Driving semantic parsing from the world\u2019s response"", ""venue"": ""Proceedings of the Fourteenth Conference on Computational Natural"", ""year"": 2010}, {""authors"": [""Cocke"", ""John.""], ""title"": ""Programming Languages and Their Compilers: Preliminary Notes"", ""venue"": ""Courant Institute of Mathematical Sciences, New York University."", ""year"": 1969}, {""authors"": [""Councill"", ""Isaac G."", ""Ryan McDonald"", ""Leonid Velikovich.""], ""title"": ""What\u2019s great and what\u2019s not: Learning to classify the scope of negation for improved sentiment analysis"", ""venue"": ""Proceedings of the Workshop on"", ""year"": 2010}, {""authors"": [""Cui"", ""Hang"", ""Vibhu Mittal"", ""Mayur Datar.""], ""title"": ""Comparative experiments on sentiment classification for online product reviews"", ""venue"": ""Proceedings of the 21st National Conference on Artificial Intelligence -"", ""year"": 2006}, {""authors"": [""Davidov"", ""Dmitry"", ""Oren Tsur"", ""Ari Rappoport.""], ""title"": ""Enhanced sentiment learning using twitter hashtags and smileys"", ""venue"": ""Proceedings of the 23rd International Conference on Computational"", ""year"": 2010}, {""authors"": [""de Marneffe"", ""Marie-Catherine"", ""Christopher D. Manning"", ""Christopher Potts""], ""title"": ""was it good? it was provocative.\u201d Learning the meaning of scalar adjectives"", ""year"": 2010}, {""authors"": [""Dong"", ""Li"", ""Furu Wei"", ""Ming Zhou"", ""Ke Xu.""], ""title"": ""Adaptive multi-compositionality for recursive neural models with applications to sentiment analysis"", ""venue"": ""AAAI Conference on Artificial Intelligence, pages 1,537\u20131,543,"", ""year"": 2014}, {""authors"": [""Duchi"", ""John"", ""Elad Hazan"", ""Yoram Singer.""], ""title"": ""Adaptive subgradient methods for online learning and stochastic optimization"", ""venue"": ""Journal of Machine Learning Research, 12:2121\u20132159."", ""year"": 2011}, {""authors"": [""Fan"", ""Rong-En"", ""Kai-Wei Chang"", ""Cho-Jui Hsieh"", ""Xiang-Rui Wang"", ""Chih-Jen Lin.""], ""title"": ""Liblinear: A library for large linear classification"", ""venue"": ""Journal of Machine Learning Research, 9:1871\u20131874."", ""year"": 2008}, {""authors"": [""Good"", ""Irving John.""], ""title"": ""The population frequencies of species and the estimation of population parameters"", ""venue"": ""Biometrika, 40(3-4):237\u2013264."", ""year"": 1953}, {""authors"": [""Goodman"", ""Joshua.""], ""title"": ""Semiring parsing"", ""venue"": ""Computational Linguistics, 25(4):573\u2013605."", ""year"": 1999}, {""authors"": [""Hall"", ""David"", ""Greg Durrett"", ""Dan Klein.""], ""title"": ""Less grammar, more features"", ""venue"": ""Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 228\u2013237,"", ""year"": 2014}, {""authors"": [""Hatzivassiloglou"", ""Vasileios"", ""Kathleen R. McKeown.""], ""title"": ""Predicting the semantic orientation of adjectives"", ""venue"": ""Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and Eighth"", ""year"": 1997}, {""authors"": [""Jia"", ""Lifeng"", ""Clement Yu"", ""Weiyi Meng.""], ""title"": ""The effect of negation on sentiment analysis and retrieval effectiveness"", ""venue"": ""Conference on Information and Knowledge Management, pages 1,827\u20131,830,"", ""year"": 2009}, {""authors"": [""Kaji"", ""Nobuhiro"", ""Masaru Kitsuregawa.""], ""title"": ""Building lexicon for sentiment analysis from massive collection of HTML documents"", ""venue"": ""Proceedings of the 2007 Joint Conference on Empirical Methods in Natural"", ""year"": 2007}, {""authors"": [""Kamps"", ""Jaap"", ""Robert J. Mokken"", ""Maarten Marx"", ""Maarten de Rijke.""], ""title"": ""Using WordNet to measure semantic orientation of adjectives"", ""venue"": ""Proceedings of the 4th International Conference on Language"", ""year"": 2004}, {""authors"": [""Kanayama"", ""Hiroshi"", ""Tetsuya Nasukawa.""], ""title"": ""Fully automatic lexicon expansion for domain-oriented sentiment analysis"", ""venue"": ""Proceedings of the 2006 Conference on Empirical Methods in Natural Language"", ""year"": 2006}, {""authors"": [""Kasami"", ""Tadao.""], ""title"": ""An efficient recognition and syntax-analysis algorithm for context-free languages"", ""venue"": ""Technical report AFCRL-65-758, Air Force Cambridge Research Lab, Bedford, MA."", ""year"": 1965}, {""authors"": [""Kate"", ""Rohit J."", ""Raymond J. Mooney.""], ""title"": ""Using string-kernels for learning semantic parsers"", ""venue"": ""ACL 2006: Proceedings of the 21st International Conference on Computational Linguistics and the 44th"", ""year"": 2006}, {""authors"": [""Kennedy"", ""Alistair"", ""Diana Inkpen.""], ""title"": ""Sentiment classification of movie reviews using contextual valence shifters"", ""venue"": ""Computational Intelligence, 22:110\u2013125."", ""year"": 2006}, {""authors"": [""Klein"", ""Dan"", ""Christopher D. Manning.""], ""title"": ""Accurate unlexicalized parsing"", ""venue"": ""Proceedings of the 41st Annual Meeting on Association for Computational Linguistics Volume 1, ACL \u201903, pages 423\u2013430,"", ""year"": 2003}, {""authors"": [""Klenner"", ""Manfred"", ""Stefanos Petrakis"", ""Angela Fahrni.""], ""title"": ""Robust compositional polarity classification"", ""venue"": ""Proceedings of the International Conference RANLP-2009, pages 180\u2013184, Borovets."", ""year"": 2009}, {""authors"": [""Krestel"", ""Ralf"", ""Stefan Siersdorfer.""], ""title"": ""Generating contextualized sentiment lexica based on latent topics and user ratings"", ""venue"": ""Proceedings of the 24th ACM Conference on Hypertext and Social Media,"", ""year"": 2013}, {""authors"": [""Krishnamurthy"", ""Jayant"", ""Tom M. Mitchell.""], ""title"": ""Weakly supervised training of semantic parsers"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and"", ""year"": 2012}, {""authors"": [""K\u00fcbler"", ""Sandra"", ""Ryan McDonald"", ""Joakim Nivre.""], ""title"": ""Dependency parsing"", ""venue"": ""Synthesis Lectures on Human Language Technologies, 1(1):1\u2013127."", ""year"": 2009}, {""authors"": [""Li"", ""Peng"", ""Yang Liu"", ""Maosong Sun.""], ""title"": ""An extended ghkm algorithm for inducing Lambda-SCFG"", ""venue"": ""Conference on Artificial Intelligence, pages 605\u2013611, Bellevue, WA."", ""year"": 2013}, {""authors"": [""Liang"", ""Percy"", ""Michael I. Jordan"", ""Dan Klein.""], ""title"": ""Learning dependency-based compositional semantics"", ""venue"": ""Computational Linguistics, 39(2):389\u2013446."", ""year"": 2013}, {""authors"": [""Liu"", ""Bing.""], ""title"": ""Sentiment Analysis and Opinion Mining"", ""venue"": ""Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers."", ""year"": 2012}, {""authors"": [""Liu"", ""Jingjing"", ""Stephanie Seneff""], ""title"": ""Review sentiment scoring via a 305"", ""year"": 2009}, {""authors"": [""Liu"", ""Shizhu"", ""Gady Agam"", ""David A. Grossman.""], ""title"": ""Generalized sentiment-bearing expression features for sentiment analysis"", ""venue"": ""International Conference on Computational Linguistics,"", ""year"": 2012}, {""authors"": [""Lu"", ""Yue"", ""Mal\u00fa Castellanos"", ""Umeshwar Dayal"", ""ChengXiang Zhai.""], ""title"": ""Automatic construction of a context-aware sentiment lexicon: an optimization approach"", ""venue"": ""International World Wide Web Conference,"", ""year"": 2011}, {""authors"": [""Maas"", ""Andrew L."", ""Raymond E. Daly"", ""Peter T. Pham"", ""Dan Huang"", ""Andrew Y. Ng"", ""Christopher Potts.""], ""title"": ""Learning word vectors for sentiment analysis"", ""venue"": ""Proceedings of the 49th Annual Meeting of the"", ""year"": 2011}, {""authors"": [""Manning"", ""Christopher D."", ""Prabhakar Raghavan"", ""Hinrich Sch\u00fctze.""], ""title"": ""Introduction to Information Retrieval"", ""venue"": ""Cambridge University Press, New York."", ""year"": 2008}, {""authors"": [""Marcus"", ""Mitchell P."", ""Mary Ann Marcinkiewicz"", ""Beatrice Santorini.""], ""title"": ""Building a large annotated corpus of English: the Penn Treebank"", ""venue"": ""Computational Linguistics, 19(2):313\u2013330."", ""year"": 1993}, {""authors"": [""Matsumoto"", ""Shotaro"", ""Hiroya Takamura"", ""Manabu Okumura.""], ""title"": ""Sentiment classification using word sub-sequences and dependency sub-trees"", ""venue"": ""Proceedings of the 9th Pacific-Asia Conference on Advances"", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Koby Crammer"", ""Fernando Pereira.""], ""title"": ""Online large-margin training of dependency parsers"", ""venue"": ""Proceedings of the 43rd Annual Meeting on Association for Computational"", ""year"": 2005}, {""authors"": [""McDonald"", ""Ryan"", ""Kerry Hannan"", ""Tyler Neylon"", ""Mike Wells"", ""Jeff Reynar.""], ""title"": ""Structured models for fine-to-coarse sentiment analysis"", ""venue"": ""Proceedings of the Association for Computational Linguistics"", ""year"": 2007}, {""authors"": [""Moilanen"", ""Karo"", ""Stephen Pulman.""], ""title"": ""Sentiment composition"", ""venue"": ""Proceedings of Recent Advances in Natural Language"", ""year"": 2007}, {""authors"": [""Moilanen"", ""Karo"", ""Stephen Pulman"", ""Yue Zhang.""], ""title"": ""Packed feelings and ordered sentiments: Sentiment parsing with quasi-compositional polarity sequencing and compression"", ""venue"": ""Proceedings of the 1st"", ""year"": 2010}, {""authors"": [""Mudinas"", ""Andrius"", ""Dell Zhang"", ""Mark Levene.""], ""title"": ""Combining lexicon and learning based approaches for concept-level sentiment analysis"", ""venue"": ""Proceedings of the First International"", ""year"": 2012}, {""authors"": [""Nakagawa"", ""Tetsuji"", ""Kentaro Inui"", ""Sadao Kurohashi.""], ""title"": ""Dependency tree-based sentiment classification using CRFS with hidden variables"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of"", ""year"": 2010}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts"", ""venue"": ""Proceedings of the 42nd Meeting of the Association for Computational"", ""year"": 2004}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales"", ""venue"": ""Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics,"", ""year"": 2005}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee.""], ""title"": ""Opinion mining and sentiment analysis"", ""venue"": ""Foundations and Trends in Information Retrieval, 2(1-2):1\u2013135."", ""year"": 2008}, {""authors"": [""Pang"", ""Bo"", ""Lillian Lee"", ""Shivakumar Vaithyanathan.""], ""title"": ""Thumbs up?: Sentiment classification using machine learning techniques"", ""venue"": ""Proceedings of the ACL-02 Conference on Empirical"", ""year"": 2002}, {""authors"": [""Polanyi"", ""Livia"", ""Annie Zaenen.""], ""title"": ""Contextual valence shifters"", ""venue"": ""J. G. Shanahan, Y. Qu, and J. Wiebe, editors, Computing Attitude and Affect in Text: Theory and Applications. Springer"", ""year"": 2006}, {""authors"": [""R. Quirk""], ""title"": ""A Comprehensive Grammar of the English Language"", ""venue"": ""Longman."", ""year"": 1985}, {""authors"": [""Raymond"", ""Ruifang Ge"", ""J. Mooney.""], ""title"": ""Discriminative reranking for semantic parsing"", ""venue"": ""Proceedings of the COLING/ACL on Main Conference Poster Sessions, COLING-ACL \u201906, pages 263\u2013270,"", ""year"": 2006}, {""authors"": [""H. Robbins"", ""S. Monro.""], ""title"": ""A stochastic approximation method"", ""venue"": ""Annals of Mathematical Statistics, 22:400\u2013407."", ""year"": 1951}, {""authors"": [""Saur\u0131"", ""Roser""], ""title"": ""A Factuality Profiler for Eventualities in Text"", ""year"": 2008}, {""authors"": [""Shieber"", ""Stuart M."", ""Yves Schabes"", ""Fernando C.N. Pereira.""], ""title"": ""Principles and implementation of deductive parsing"", ""venue"": ""Journal of Logic Programming, 24(1\u20132):3\u201336."", ""year"": 1995}, {""authors"": [""Socher"", ""Richard"", ""Brody Huval"", ""Christopher D. Manning"", ""Andrew Y. Ng.""], ""title"": ""Semantic compositionality through recursive matrix-vector spaces"", ""venue"": ""Proceedings of the 2012 Joint Conference on"", ""year"": 2012}, {""authors"": [""Socher"", ""Richard"", ""Jeffrey Pennington"", ""Eric H. Huang"", ""Andrew Y. Ng"", ""Christopher D. Manning.""], ""title"": ""Semi-supervised recursive autoencoders for predicting sentiment distributions"", ""venue"": ""Proceedings of the"", ""year"": 2011}, {""authors"": [""Socher"", ""Richard"", ""Alex Perelygin"", ""Jean Y. Wu"", ""Jason Chuang"", ""Christopher D. Manning"", ""Andrew Y. Ng"", ""Christopher Potts""], ""title"": ""Recursive deep models for semantic compositionality over a sentiment"", ""year"": 2013}, {""authors"": [""Stolcke"", ""Andreas.""], ""title"": ""SRILM: An extensible language modeling toolkit"", ""venue"": ""Proceedings of the 7th International Conference on Spoken Language Processing (ICSLP 2002, pages 901\u2013904, Denver, CO."", ""year"": 2002}, {""authors"": [""Taboada"", ""Maite"", ""Julian Brooke"", ""Milan Tofiloski"", ""Kimberly Voll"", ""Manfred Stede.""], ""title"": ""Lexicon-based methods for sentiment analysis"", ""venue"": ""Computational Linguistics, 37(2):267\u2013307."", ""year"": 2011}, {""authors"": [""T\u00e4ckstr\u00f6m"", ""Oscar"", ""Ryan McDonald.""], ""title"": ""Discovering fine-grained sentiment with latent variable structured prediction models"", ""venue"": ""Proceedings of the 33rd European Conference on Advances in Information"", ""year"": 2011}, {""authors"": [""T\u00e4ckstr\u00f6m"", ""Oscar"", ""Ryan McDonald.""], ""title"": ""Semi-supervised latent variable models for sentence-level sentiment analysis"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational"", ""year"": 2011}, {""authors"": [""Takamura"", ""Hiroya"", ""Takashi Inui"", ""Manabu Okumura.""], ""title"": ""Extracting semantic orientations of words using spin model"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational"", ""year"": 2005}, {""authors"": [""Tu"", ""Zhaopeng"", ""Yifan He"", ""Jennifer Foster"", ""Josef van Genabith"", ""Qun Liu"", ""Shouxun Lin.""], ""title"": ""Identifying high-impact sub-structures for convolution kernels in document-level sentiment classification"", ""venue"": ""In"", ""year"": 2012}, {""authors"": [""Turney"", ""Peter D.""], ""title"": ""Thumbs up or thumbs down?: Semantic orientation applied to unsupervised classification of reviews"", ""venue"": ""Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics,"", ""year"": 2002}, {""authors"": [""Velikovich"", ""Leonid"", ""Sasha Blair-Goldensohn"", ""Kerry Hannan"", ""Ryan McDonald.""], ""title"": ""The viability of Web-derived polarity lexicons"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North"", ""year"": 2010}, {""authors"": [""Wainwright"", ""Martin J."", ""Michael I. Jordan.""], ""title"": ""Graphical models, exponential families, and variational inference"", ""venue"": ""Foundations and Trends in Machine Learning, 1(1-2):1\u2013305."", ""year"": 2008}, {""authors"": [""Wang"", ""Sida"", ""Christopher Manning.""], ""title"": ""Baselines and bigrams: Simple, good sentiment and topic classification"", ""venue"": ""Proceedings of the 50th Annual Meeting of the Association for Computational"", ""year"": 2012}, {""authors"": [""Wiebe"", ""Janyce"", ""Theresa Wilson"", ""Claire Cardie.""], ""title"": ""Annotating expressions of opinions and emotions in language"", ""venue"": ""Language Resources and Evaluation, 39(2-3):165\u2013210."", ""year"": 2005}, {""authors"": [""Williams"", ""Gbolahan K"", ""Sarabjot Singh Anand""], ""title"": ""Predicting the polarity"", ""year"": 2009}, {""authors"": [""Wilson"", ""Theresa"", ""Janyce Wiebe"", ""Paul Hoffmann.""], ""title"": ""Recognizing contextual polarity: An exploration of features for phrase-level sentiment analysis"", ""venue"": ""Computational Linguistics, 35:399\u2013433."", ""year"": 2009}, {""authors"": [""Yessenalina"", ""Ainur"", ""Yisong Yue"", ""Claire Cardie.""], ""title"": ""Multi-level structured models for document-level sentiment classification"", ""venue"": ""Proceedings of the 2010 Conference on Empirical Methods in Natural"", ""year"": 2010}, {""authors"": [""Younger"", ""Daniel H.""], ""title"": ""Recognition and parsing of context-free languages in time n3"", ""venue"": ""Information and Control, 10(2):189\u2013208."", ""year"": 1967}, {""authors"": [""Yu"", ""Hong"", ""Vasileios Hatzivassiloglou.""], ""title"": ""Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences"", ""venue"": ""Proceedings of the 2003"", ""year"": 2003}, {""authors"": [""Zelle"", ""John M"", ""Raymond J. Mooney""], ""title"": ""Learning to parse database queries"", ""year"": 1996}, {""authors"": [""Zettlemoyer"", ""Luke S."", ""Michael Collins.""], ""title"": ""Online learning of relaxed CCG grammars for parsing to logical form"", ""venue"": ""Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language"", ""year"": 2007}, {""authors"": [""Zettlemoyer"", ""Luke S."", ""Michael Collins.""], ""title"": ""Learning context-dependent mappings from sentences to logical form"", ""venue"": ""Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the"", ""year"": 2009}, {""authors"": [""Zhao"", ""Jichang"", ""Li Dong"", ""Junjie Wu"", ""Ke Xu.""], ""title"": ""Moodlens: An emoticon-based sentiment analysis system for Chinese tweets"", ""venue"": ""Proceedings of the 18th ACM SIGKDD International Conference"", ""year"": 2012}] acknowledgments  :This research was partly supported by NSFC (grant no. 61421003) and the fund of the State Key Lab of Software Development Environment (grant no. SKLSDE-2015ZX-05). We gratefully acknowledge helpful discussions with Dr. Nan Yang and the anonymous reviewers.
References Agarwal, Apoorv, Boyi Xie, Ilia Vovsha,
Owen Rambow, and Rebecca Passonneau. 2011. Sentiment analysis of twitter data. In Proceedings of the Workshop on Languages in Social Media, LSM ’11, pages 30–38, Stroudsburg, PA.
Agrawal, Rakesh and Ramakrishnan Srikant. 1994. Fast algorithms for mining association rules in large databases. In Proceedings of the 20th International Conference on Very Large Data Bases, VLDB ’94, pages 487–499, San Francisco, CA. Artzi, Yoav and Luke Zettlemoyer. 2013. Weakly supervised learning of semantic parsers for mapping instructions to actions. Transactions of the Association for Computational Linguistics, 1(1):49–62. Bao, Junwei, Nan Duan, Ming Zhou, and Tiejun Zhao. 2014. Knowledge-based question answering as machine translation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 967–976, Baltimore, MD. Charniak, Eugene. 1997. Statistical parsing with a context-free grammar and word statistics. In Proceedings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative Applications of Artificial Intelligence, AAAI’97/IAAI’97, pages 598–603, Providence, RI. Charniak, Eugene and Mark Johnson. 2005. Coarse-to-fine n-best parsing and maxent discriminative reranking. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL ’05, pages 173–180, Stroudsburg, PA. Chiang, David. 2007. Hierarchical phrase-based translation. Computational Linguistics, 33(2):201–228. Choi, Yejin and Claire Cardie. 2008. Learning with compositional semantics as structural inference for subsentential sentiment analysis. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’08, pages 793–801, Stroudsburg, PA. Choi, Yejin and Claire Cardie. 2009a. Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 2 – Volume 2, EMNLP ’09, pages 590–598, Stroudsburg, PA. Choi, Yejin and Claire Cardie. 2009b. Adapting a polarity lexicon using integer linear programming for domain-specific sentiment classification. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing:
Volume 2 – Volume 2, pages 590–598, Singapore.
Choi, Yejin and Claire Cardie. 2010. Hierarchical sequential learning for extracting opinions and their attributes. In Proceedings of the ACL 2010 Conference Short Papers, ACLShort ’10, pages 269–274, Stroudsburg, PA. Chomsky, Noam. 1956. Three models for the description of language. IRE Transactions on Information Theory, 2(3):113–124. Clarke, James, Dan Goldwasser, Ming-Wei Chang, and Dan Roth. 2010. Driving semantic parsing from the world’s response. In Proceedings of the Fourteenth Conference on Computational Natural Language Learning, CoNLL ’10, pages 18–27, Stroudsburg, PA. Cocke, John. 1969. Programming Languages and Their Compilers: Preliminary Notes. Courant Institute of Mathematical Sciences, New York University. Councill, Isaac G., Ryan McDonald, and Leonid Velikovich. 2010. What’s great and what’s not: Learning to classify the scope of negation for improved sentiment analysis. In Proceedings of the Workshop on Negation and Speculation in Natural Language Processing, NeSp-NLP ’10, pages 51–59, Stroudsburg, PA. Cui, Hang, Vibhu Mittal, and Mayur Datar. 2006. Comparative experiments on sentiment classification for online product reviews. In Proceedings of the 21st National Conference on Artificial Intelligence - Volume 2, AAAI’06, pages 1,265–1,270, Boston, MA. Davidov, Dmitry, Oren Tsur, and Ari Rappoport. 2010. Enhanced sentiment learning using twitter hashtags and smileys. In Proceedings of the 23rd International Conference on Computational Linguistics: Posters, COLING ’10, pages 241–249, Stroudsburg, PA. de Marneffe, Marie-Catherine, Christopher D. Manning, and Christopher Potts. 2010. “was it good? it was provocative.” Learning the meaning of scalar adjectives. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, ACL ’10, pages 167–176, Stroudsburg, PA. Dong, Li, Furu Wei, Ming Zhou, and Ke Xu. 2014. Adaptive multi-compositionality for recursive neural models with applications to sentiment analysis. In AAAI Conference on Artificial Intelligence, pages 1,537–1,543, Quebec.
Duchi, John, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12:2121–2159. Fan, Rong-En, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui Wang, and Chih-Jen Lin. 2008. Liblinear: A library for large linear classification. Journal of Machine Learning Research, 9:1871–1874. Good, Irving John. 1953. The population frequencies of species and the estimation of population parameters. Biometrika, 40(3-4):237–264. Goodman, Joshua. 1999. Semiring parsing. Computational Linguistics, 25(4):573–605. Hall, David, Greg Durrett, and Dan Klein. 2014. Less grammar, more features. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 228–237, Baltimore, MD. Hatzivassiloglou, Vasileios and Kathleen R. McKeown. 1997. Predicting the semantic orientation of adjectives. In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics and Eighth Conference of the European Chapter of the Association for Computational Linguistics, ACL ’98, pages 174–181, Stroudsburg, PA. Jia, Lifeng, Clement Yu, and Weiyi Meng. 2009. The effect of negation on sentiment analysis and retrieval effectiveness. In Conference on Information and Knowledge Management, pages 1,827–1,830, Hong Kong. Kaji, Nobuhiro and Masaru Kitsuregawa. 2007. Building lexicon for sentiment analysis from massive collection of HTML documents. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 1,075–1,083, Prague. Kamps, Jaap, Robert J. Mokken, Maarten Marx, and Maarten de Rijke. 2004. Using WordNet to measure semantic orientation of adjectives. In Proceedings of the 4th International Conference on Language Resources and Evaluation (LREC 2004), volume IV, pages 1,115–1,118, Paris. Kanayama, Hiroshi and Tetsuya Nasukawa. 2006. Fully automatic lexicon expansion for domain-oriented sentiment analysis. In Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing, EMNLP ’06, pages 355–363, Stroudsburg, PA.
Kasami, Tadao. 1965. An efficient recognition and syntax-analysis algorithm for context-free languages. Technical report AFCRL-65-758, Air Force Cambridge Research Lab, Bedford, MA. Kate, Rohit J. and Raymond J. Mooney. 2006. Using string-kernels for learning semantic parsers. In ACL 2006: Proceedings of the 21st International Conference on Computational Linguistics and the 44th Annual Meeting of the ACL, pages 913–920, Morristown, NJ. Kennedy, Alistair and Diana Inkpen. 2006. Sentiment classification of movie reviews using contextual valence shifters. Computational Intelligence, 22:110–125. Klein, Dan and Christopher D. Manning. 2003. Accurate unlexicalized parsing. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics - Volume 1, ACL ’03, pages 423–430, Stroudsburg, PA. Klenner, Manfred, Stefanos Petrakis, and Angela Fahrni. 2009. Robust compositional polarity classification. In Proceedings of the International Conference RANLP-2009, pages 180–184, Borovets. Krestel, Ralf and Stefan Siersdorfer. 2013. Generating contextualized sentiment lexica based on latent topics and user ratings. In Proceedings of the 24th ACM Conference on Hypertext and Social Media, HT ’13, pages 129–138, New York, NY. Krishnamurthy, Jayant and Tom M. Mitchell. 2012. Weakly supervised training of semantic parsers. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL ’12, pages 754–765, Stroudsburg, PA. Kübler, Sandra, Ryan McDonald, and Joakim Nivre. 2009. Dependency parsing. Synthesis Lectures on Human Language Technologies, 1(1):1–127. Li, Peng, Yang Liu, and Maosong Sun. 2013. An extended ghkm algorithm for inducing Lambda-SCFG. In Conference on Artificial Intelligence, pages 605–611, Bellevue, WA. Liang, Percy, Michael I. Jordan, and Dan Klein. 2013. Learning dependency-based compositional semantics. Computational Linguistics, 39(2):389–446. Liu, Bing. 2012. Sentiment Analysis and Opinion Mining. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers. Liu, Jingjing and Stephanie Seneff. 2009. Review sentiment scoring via a
parse-and-paraphrase paradigm. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1 - Volume 1, EMNLP ’09, pages 161–169, Stroudsburg, PA.
Liu, Shizhu, Gady Agam, and David A. Grossman. 2012. Generalized sentiment-bearing expression features for sentiment analysis. In International Conference on Computational Linguistics, pages 733–744, Mumbai. Lu, Yue, Malú Castellanos, Umeshwar Dayal, and ChengXiang Zhai. 2011. Automatic construction of a context-aware sentiment lexicon: an optimization approach. In International World Wide Web Conference, pages 347–356, Hyderabad. Maas, Andrew L., Raymond E. Daly, Peter T. Pham, Dan Huang, Andrew Y. Ng, and Christopher Potts. 2011. Learning word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies - Volume 1, HLT ’11, pages 142–150, Stroudsburg, PA. Manning, Christopher D., Prabhakar Raghavan, and Hinrich Schütze. 2008. Introduction to Information Retrieval. Cambridge University Press, New York. Marcus, Mitchell P., Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of English: the Penn Treebank. Computational Linguistics, 19(2):313–330. Matsumoto, Shotaro, Hiroya Takamura, and Manabu Okumura. 2005. Sentiment classification using word sub-sequences and dependency sub-trees. In Proceedings of the 9th Pacific-Asia Conference on Advances in Knowledge Discovery and Data Mining, PAKDD’05, pages 301–311, Hanoi. McDonald, Ryan, Koby Crammer, and Fernando Pereira. 2005. Online large-margin training of dependency parsers. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL ’05, pages 91–98, Stroudsburg, PA. McDonald, Ryan, Kerry Hannan, Tyler Neylon, Mike Wells, and Jeff Reynar. 2007. Structured models for fine-to-coarse sentiment analysis. In Proceedings of the Association for Computational Linguistics (ACL), pages 432–439, Prague. Moilanen, Karo and Stephen Pulman. 2007. Sentiment composition. In Proceedings of Recent Advances in Natural Language
Processing (RANLP 2007), pages 378–382, Borovets.
Moilanen, Karo, Stephen Pulman, and Yue Zhang. 2010. Packed feelings and ordered sentiments: Sentiment parsing with quasi-compositional polarity sequencing and compression. In Proceedings of the 1st Workshop on Computational Approaches to Subjectivity and Sentiment Analysis (WASSA 2010) at the 19th European Conference on Artificial Intelligence (ECAI 2010), pages 36–43, Lisbon. Mudinas, Andrius, Dell Zhang, and Mark Levene. 2012. Combining lexicon and learning based approaches for concept-level sentiment analysis. In Proceedings of the First International Workshop on Issues of Sentiment Discovery and Opinion Mining, WISDOM ’12, pages 5:1–5:8, New York, NY. Nakagawa, Tetsuji, Kentaro Inui, and Sadao Kurohashi. 2010. Dependency tree-based sentiment classification using CRFS with hidden variables. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, HLT ’10, pages 786–794, Stroudsburg, PA. Pang, Bo and Lillian Lee. 2004. A sentimental education: Sentiment analysis using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Meeting of the Association for Computational Linguistics (ACL’04), Main Volume, pages 271–278, Barcelona. Pang, Bo and Lillian Lee. 2005. Seeing stars: exploiting class relationships for sentiment categorization with respect to rating scales. In Proceedings of the 43rd Annual Meeting on Association for Computational Linguistics, ACL ’05, pages 115–124, Stroudsburg, PA. Pang, Bo and Lillian Lee. 2008. Opinion mining and sentiment analysis. Foundations and Trends in Information Retrieval, 2(1-2):1–135. Pang, Bo, Lillian Lee, and Shivakumar Vaithyanathan. 2002. Thumbs up?: Sentiment classification using machine learning techniques. In Proceedings of the ACL-02 Conference on Empirical Methods in Natural Language Processing - Volume 10, EMNLP ’02, pages 79–86, Stroudsburg, PA. Polanyi, Livia and Annie Zaenen. 2006. Contextual valence shifters. In J. G. Shanahan, Y. Qu, and J. Wiebe, editors, Computing Attitude and Affect in Text: Theory and Applications. Springer Netherlands, pages 1–10.
Quirk, R. 1985. A Comprehensive Grammar of the English Language. Longman. Raymond, Ruifang Ge and J. Mooney. 2006. Discriminative reranking for semantic parsing. In Proceedings of the COLING/ACL on Main Conference Poster Sessions, COLING-ACL ’06, pages 263–270, Stroudsburg, PA. Robbins, H. and S. Monro. 1951. A stochastic approximation method. Annals of Mathematical Statistics, 22:400–407. Saurı́, Roser. 2008. A Factuality Profiler for Eventualities in Text. Ph.D. thesis, Brandeis University. Shieber, Stuart M., Yves Schabes, and Fernando C. N. Pereira. 1995. Principles and implementation of deductive parsing. Journal of Logic Programming, 24(1–2):3–36. Socher, Richard, Brody Huval, Christopher D. Manning, and Andrew Y. Ng. 2012. Semantic compositionality through recursive matrix-vector spaces. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, EMNLP-CoNLL ’12, pages 1,201–1,211, Stroudsburg, PA. Socher, Richard, Jeffrey Pennington, Eric H. Huang, Andrew Y. Ng, and Christopher D. Manning. 2011. Semi-supervised recursive autoencoders for predicting sentiment distributions. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, EMNLP ’11, pages 151–161, Stroudsburg, PA. Socher, Richard, Alex Perelygin, Jean Y. Wu, Jason Chuang, Christopher D. Manning, Andrew Y. Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1,631–1,642, Seattle, WA. Stolcke, Andreas. 2002. SRILM: An extensible language modeling toolkit. In Proceedings of the 7th International Conference on Spoken Language Processing (ICSLP 2002, pages 901–904, Denver, CO. Taboada, Maite, Julian Brooke, Milan Tofiloski, Kimberly Voll, and Manfred Stede. 2011. Lexicon-based methods for sentiment analysis. Computational Linguistics, 37(2):267–307. Täckström, Oscar and Ryan McDonald. 2011a. Discovering fine-grained sentiment with latent variable structured prediction models. In Proceedings of the 33rd European Conference on Advances in Information
Retrieval, ECIR’11, pages 368–374, Berlin.
Täckström, Oscar and Ryan McDonald. 2011b. Semi-supervised latent variable models for sentence-level sentiment analysis. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies: Short Papers - Volume 2, HLT ’11, pages 569–574, Stroudsburg, PA. Takamura, Hiroya, Takashi Inui, and Manabu Okumura. 2005. Extracting semantic orientations of words using spin model. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, ACL ’05, pages 133–140, Stroudsburg, PA. Tu, Zhaopeng, Yifan He, Jennifer Foster, Josef van Genabith, Qun Liu, and Shouxun Lin. 2012. Identifying high-impact sub-structures for convolution kernels in document-level sentiment classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Short Papers - Volume 2, ACL ’12, pages 338–343, Stroudsburg, PA. Turney, Peter D. 2002. Thumbs up or thumbs down?: Semantic orientation applied to unsupervised classification of reviews. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, ACL ’02, pages 417–424, Stroudsburg, PA. Velikovich, Leonid, Sasha Blair-Goldensohn, Kerry Hannan, and Ryan McDonald. 2010. The viability of Web-derived polarity lexicons. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, HLT ’10, pages 777–785, Stroudsburg, PA. Wainwright, Martin J. and Michael I. Jordan. 2008. Graphical models, exponential families, and variational inference. Foundations and Trends in Machine Learning, 1(1-2):1–305. Wang, Sida and Christopher Manning. 2012. Baselines and bigrams: Simple, good sentiment and topic classification. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (ACL 2012), pages 90–94, Jeju Island. Wiebe, Janyce, Theresa Wilson, and Claire Cardie. 2005. Annotating expressions of opinions and emotions in language. Language Resources and Evaluation, 39(2-3):165–210. Williams, Gbolahan K. and Sarabjot Singh Anand. 2009. Predicting the polarity
strength of adjectives using Wordnet. In ICWSM, pages 346–349, San Jose, CA.
Wilson, Theresa, Janyce Wiebe, and Paul Hoffmann. 2009. Recognizing contextual polarity: An exploration of features for phrase-level sentiment analysis. Computational Linguistics, 35:399–433. Yessenalina, Ainur, Yisong Yue, and Claire Cardie. 2010. Multi-level structured models for document-level sentiment classification. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1,046–1,056, Cambridge, MA. Younger, Daniel H. 1967. Recognition and parsing of context-free languages in time n3. Information and Control, 10(2):189–208. Yu, Hong and Vasileios Hatzivassiloglou. 2003. Towards answering opinion questions: Separating facts from opinions and identifying the polarity of opinion sentences. In Proceedings of the 2003 Conference on Empirical Methods in Natural Language Processing, EMNLP ’03, pages 129–136, Stroudsburg, PA. Zelle, John M. and Raymond J. Mooney. 1996. Learning to parse database queries
using inductive logic programming. In AAAI/IAAI, pages 1,050–1,055, Portland, OR.
Zettlemoyer, Luke S. and Michael Collins. 2007. Online learning of relaxed CCG grammars for parsing to logical form. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL-2007), pages 678–687, Prague. Zettlemoyer, Luke S. and Michael Collins. 2009. Learning context-dependent mappings from sentences to logical form. In Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP: Volume 2 - Volume 2, ACL ’09, pages 976–984, Stroudsburg, PA. Zhao, Jichang, Li Dong, Junjie Wu, and Ke Xu. 2012. Moodlens: An emoticon-based sentiment analysis system for Chinese tweets. In Proceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’12, pages 1,528–1,531, New York, NY. a statistical parsing framework for :Sentiment Classification
Li Dong∗, ∗∗ Beihang University
Furu Wei†, ‡ Microsoft Research
Shujie Liu† Microsoft Research
Ming Zhou† Microsoft Research
Ke Xu∗ Beihang University
We present a statistical parsing framework for sentence-level sentiment classification in this article. Unlike previous works that use syntactic parsing results for sentiment analysis, we develop a statistical parser to directly analyze the sentiment structure of a sentence. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way. We formulate the sentiment grammar upon Context-Free Grammars (CFGs), and provide a formal description of the sentiment parsing framework. We develop the parsing model to obtain possible sentiment parse trees for a sentence, from which the polarity model is proposed to derive the sentiment strength and polarity, and the ranking model is dedicated to selecting the best sentiment tree. We train the parser directly from examples of sentences annotated only with sentiment polarity labels but without any syntactic annotations or polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, s.parser, from a large amount of review sentences with users’ ratings as rough sentiment polarity labels. Extensive experiments on existing benchmark data sets show significant improvements over baseline sentiment classification approaches.
∗ State Key Laboratory of Software Development Environment, Beihang University, XueYuan Road No.37, HaiDian District, Beijing, P.R. China 100191. E-mail: donglixp@gmail.com; kexu@nlsde.buaa.edu.cn.
∗∗ Contribution during internship at Microsoft Research. † Natural Language Computing Group, Microsoft Research Asia, Building 2, No. 5 Danling Street, Haidian
District, Beijing, P.R. China 100080. E-mail: {fuwei, shujliu, mingzhou}@microsoft.com. ‡ Corresponding author.
Submission received: 10 December 2013; revised version received: 26 July 2014; accepted for publication: 28 January 2015.
doi:10.1162/COLI a 00221
© 2015 Association for Computational Linguistics","1 introduction :Sentiment analysis (Pang and Lee 2008; Liu 2012) has received much attention from both research and industry communities in recent years. Sentiment classification, which identifies sentiment polarity (positive or negative) from text (sentence or document), has been the most extensively studied task in sentiment analysis. Until now, there have been two mainstream approaches for sentiment classification. The lexicon-based approach (Turney 2002; Taboada et al. 2011) aims to aggregate the sentiment polarity of a sentence from the polarity of words or phrases found in the sentence, and the learning-based approach (Pang, Lee, and Vaithyanathan 2002) treats sentiment polarity identification as a special text classification task and focuses on building classifiers from a set of sentences (or documents) annotated with their corresponding sentiment polarity.
The lexicon-based sentiment classification approach is simple and interpretable, but suffers from scalability and is inevitably limited by sentiment lexicons that are commonly created manually by experts. It has been widely recognized that sentiment expressions are colloquial and evolve over time very frequently. Taking tweets from Twitter1 and movie reviews on IMDb2 as examples, people use very casual language as well as informal and new vocabulary to comment on general topics and movies. In practice, it is not feasible to create and maintain sentiment lexicons to capture sentiment expressions with high coverage. On the other hand, the learning-based approach relies on large annotated samples to overcome the vocabulary coverage and deals with variations of words in sentences. Human ratings in reviews (Maas et al. 2011) and emoticons in tweets (Davidov, Tsur, and Rappoport 2010; Zhao et al. 2012) are extensively used to collect a large number of training corpora to train the sentiment classifier. However, it is usually not easy to design effective features to build the classifier. Among others, unigrams have been reported as the most effective features (Pang, Lee, and Vaithyanathan 2002) in sentiment classification.
Handling complicated expressions delivering people’s opinions is one of the most challenging problems in sentiment analysis. Compositionalities such as negation, intensification, contrast, and their combinations are typical cases. We show some concrete examples here:
(1) The movie is not good. [negation]
(2) The movie is very good. [intensification]
(3) The movie is not funny at all. [negation + intensification]
(4) The movie is just so so, but i still like it. [contrast]
(5) The movie is not very good, but i still like it. [negation + intensification + contrast]
The negation expressions, intensification modifiers, and the contrastive conjunction can change the polarity (Examples (1), (3), (4), (5)), strength (Examples (2), (3), (5)), or both (Examples (3), (5)) of the sentiment of the sentences. We do not need any detailed explanations here as they can be commonly found and easily understood in people’s
1 http://twitter.com. 2 http://www.imdb.com.
daily lives. Existing works to address these issues usually rely on syntactic parsing results either used as features (Choi and Cardie 2008; Moilanen, Pulman, and Zhang 2010) in learning-based methods or hand-crafted rules (Moilanen and Pulman 2007; Jia, Yu, and Meng 2009; Klenner, Petrakis, and Fahrni 2009; Liu and Seneff 2009) in lexiconbased methods. However, even with the difficulty and feasibility of deriving the sentiment structure from syntactic parsing results put aside, it is an even more challenging task to generate stable and reliable parsing results for text that is ungrammatical in nature and has a high ratio of out-of-vocabulary words. The accuracy of the linguistic parsers trained on standard data sets (e.g., the Penn Treebank [Marcus, Marcinkiewicz, and Santorini 1993]) drops dramatically on user-generated-content (reviews, tweets, etc.), which is actually the prime focus of sentiment analysis algorithms. The error, unfortunately, will propagate downstream in the process of sentiment analysis methods building upon parsing results.
We therefore propose directly analyzing the sentiment structure of a sentence. The nested structure of sentiment expressions can be naturally modeled in a similar fashion as statistical syntactic parsing, which aims to find the linguistic structure of a sentence. This idea creates many opportunities for developing sentiment classifiers from a new perspective. The most challenging problem and barrier in building a statistical sentiment parser lies in the acquisition of training data. Ideally, we need examples of sentences annotated with polarity for the whole sentence as well as sentiment tags for constituents within a sentence, as with the Penn TreeBank for training traditional linguistic parsers. However, this is not practical as the annotations will be inevitably time-consuming and require laborious human efforts. Therefore, it is better to learn the sentiment parser only utilizing examples annotated with the polarity label of the whole sentence. For example, we can collect a huge number of publicly available reviews and rating scores on the Web. People may use the movie is gud (“gud” is a popular informal expression of “good”) to express a positive opinion towards a movie, and not a fan to express a negative opinion. Also, we can find review sentences such as The movie is gud, but I am still not a fan to indicate a negative opinion. We can then use these two fragments and the overall negative opinion of the sentence to deduce sentiment rules automatically from data. These sentiment fragments and rules can be used to analyze the sentiment structure for new sentences.
In this article, we propose a statistical parsing framework to directly analyze the structure of a sentence from the perspective of sentiment analysis. Specifically, we formulate a Context-Free Grammar (CFG)–based sentiment grammar. We then develop a statistical parser to derive the sentiment structure of a sentence. We leverage the CYK algorithm (Cocke 1969; Younger 1967; Kasami 1965) to conduct bottom–up parsing, and use dynamic programming to accelerate computation. Meanwhile, we propose using the polarity model to derive sentiment strength and polarity of a sentiment parse tree, and the ranking model to select the best one from the sentiment parsing results. We train the parser directly from examples of sentences annotated with sentiment polarity labels instead of syntactic annotations and polarity annotations of constituents within sentences. Therefore we can obtain training data easily. In particular, we train a sentiment parser, named s.parser, from a large number of review sentences with users’ ratings as rough sentiment polarity labels. The statistical parsing–based approach builds a principled and scalable framework to support the sentiment composition and inference which cannot be well handled by bag-of-words approaches. We show that complicated phenomena in sentiment analysis (e.g., negation, intensification, and contrast) can be handled the same way as simple and straightforward sentiment expressions in a unified and probabilistic way.
The major contributions of the work presented in this article are as follows.r We propose a statistical parsing framework for sentiment analysis that is capable of analyzing the sentiment structure for a sentence. This framework can naturally handle compositionality in a probabilistic way. It can be trained from sentences annotated with only sentiment polarity but without any syntactic annotations or polarity annotations of constituents within sentences.r We present the parsing model, polarity model, and ranking model in the proposed framework, which are formulated and can be improved independently. It provides a principled and flexible approach to sentiment classification.r We implement the statistical sentiment parsing framework, and conduct experiments on several benchmark data sets. The experimental results show that the proposed framework and algorithm can significantly outperform baseline methods.
The remainder of this article is organized as follows. We introduce related work in Section 2. We present the statistical sentiment parsing framework, including the parsing model, polarity model, and ranking model, in Section 3. Learning methods for our model are explained in Section 4. Experimental results are reported in Section 5. We conclude this article with future work in Section 6. 6 conclusion and future work :In this article, we propose a statistical parsing framework for sentence-level sentiment classification that provides a novel approach to designing sentiment classifiers from a new perspective. It directly analyzes the sentiment structure of a sentence other than relying on syntactic parsing results, as in existing literature. We show that complicated phenomena in sentiment analysis, such as negation, intensification, and contrast, can be handled in a similar manner to simple and straightforward sentiment expressions in a unified and probabilistic way. We provide a formal model to represent the sentiment grammar built upon Context-Free Grammars. The framework consists of: (1) a parsing model to analyze the sentiment structure of a sentence; (2) a polarity model to calculate sentiment strength and polarity for each text span in the parsing process; and (3) a ranking model to select the best parsing result from a list of candidate sentiment parse trees. We show that the sentiment parser can be trained from the examples of sentences annotated only with sentiment polarity labels but without using any syntactic or sentiment annotations within sentences. We evaluate the proposed framework on standard sentiment classification data sets. The experimental results show that the statistical sentiment parsing notably outperforms the baseline sentiment classification approaches.
We believe the work on statistical sentiment parsing can be advanced from many different perspectives. First, statistical parsing has been a well-established research field, in which many different grammars and parsing algorithms have been proposed in previously published literature. It will be an interesting direction to apply and adjust more advanced models and algorithms from the syntactic parsing and the semantic parsing to our framework. We leave it as a line of future work. Second, we can incorporate target and aspect information in the statistical sentiment parsing framework to facilitate the target-dependent and aspect-based sentiment analysis. Intuitively, this can be done by introducing semantic tags of targets and aspects as new non-terminals in the sentiment grammar and revising grammar rules accordingly. However, acquiring training data will be an even more challenging task, as we need more fine-grained information. Third, as the statistical sentiment parsing produces more fine-grained information (e.g., the basic sentiment expressions from the dictionary rules as well as the sentiment structure trees), we will have more opportunities to generate better opinion summaries. Moreover, we are interested in jointly learning parameters of the polarity model and the parsing model from data. Last but not the least, we are interested in investigating the domain adaptation, which is a very important and challenging problem in sentiment analysis. Generally, we may need to learn domain-specific dictionary rules for different domains, whereas we believe combination rules are mostly generic across different domains. This is also worth consideration for further study."
"1 introduction :A Tree Adjoining Grammar (TAG; Joshi and Schabes 1997) consists of a set of elementary trees and two combining operations, substitution and adjunction. Consequently, a TAG derivation can be described by a tree (called a derivation tree) specifying which elementary TAG trees were combined using which operations to yield that derivation. In this tree, each vertex is labeled with a tree name and each edge with a description of the operation (node address and operation type) used to combine the trees labeling its end vertices. As we shall see in Section 3.2, in TAG, each derivation tree specifies a unique parse tree, also called derived tree.
In previous work, it has been argued that TAG derivation trees provide a good approximation of semantic dependencies between the words of a sentence (Kroch 1989; Rambow, Vijay-Shanker, and Weir 1995; Candito and Kahane 1998; Kallmeyer and Kuhlmann 2012). As shown by Schabes and Shieber (1994), however, there are several possible ways of defining TAG derivation trees, depending on how multiple adjunction of several auxiliary trees at the same tree node is handled. The standard notion of derivation proposed by Vijay-Shanker (1987) forbids multiple adjunction, thus enforcing dependent derivations. In contrast, the extended notion of derivation proposed by Schabes
∗ UMR 7503, Campus Scientifique, BP 239, F-54506 Vandoeuvre-lès-Nancy Cedex, France. E-mail:{claire.gardent,shashi.narayan}@loria.fr.
Submission received: 2 August 2013; revised version received: 16 June 2014; accepted for publication: 21 September 2014.
doi:10.1162/COLI a 00217
© 2015 Association for Computational Linguistics
and Shieber (1992, 1994) allows multiple adjunction at a single node, thereby yielding so-called independent derivations (i.e., derivations where the relation between the adjoining trees is left unspecified). The difference between the two types of derivations is illustrated in Figure 1. While in the standard (dependent) derivation, one adjective tree is adjoined to the other adjective tree, which itself is adjoined to the noun tree for pepper; in the extended (independent) derivation, both adjective trees adjoin to the noun tree.
Schabes and Shieber (1994) argue that allowing both for dependent and independent derivations better reflects linguistic dependencies. Making use of the distinction introduced in TAG between predicative and modifier auxiliary trees (Schabes and Shieber (1994), Section 3.1), they define a parsing algorithm that assigns dependent derivations to predicative auxiliary trees but independent derivations to multiple modifier auxiliary trees adjoining to the same node. In case both predicative and modifier auxiliary trees adjoin to the same node, their parsing algorithm ensures that predicative trees appear above the modifier trees in the derived tree.
This parsing algorithm is defined for featureless variants of TAG. In contrast, in implemented TAGs (e.g., XTAG [The XTAG Research Group 2001], SemXTAG [Gardent 2008], or XXTAG1 [Alahverdzhieva 2008]) feature structures and feature unification are central. They are used to minimize the size of the grammar; to model linguistic phenomena such as verb/subject agreement; and to encode a unification-based syntax/semantics interface (e.g., Gardent and Kallmeyer 2003).
In this article, we extend Schabes and Shieber’s proposal to Feature-Based TAG (FB-TAG); and we show that the resulting parsing algorithm naturally accounts for the interplay of dependent vs. independent derivation structures with syntactic constraints, linear ordering, and scopal vs. nonscopal semantic dependencies.
The article is organized as follows. In Section 2, we recap the motivations for independent derivations put forward by Schabes and Shieber (1994) and we briefly discuss the interactions that may arise between dependent and independent derivations. Section 3 summarizes their approach. In Section 4, we present the intuitions and motivations underlying our proposal and we highlight the differences with Schabes and Shieber’s approach. Section 5 presents our proposal. Section 6 concludes.","2 why are independent derivations desirable? :We start by summarizing Schabes and Shieber’s motivations for independent derivations. We then discuss the interactions between dependent and independent derivations.
1 XXTAG stands for XMG (Crabbé et al. 2013)-based XTAG. Schabes and Shieber (1994) give three main motivations for independent derivations. The first motivation concerns the interaction of verbs with multiple modifiers. Consider sentences2 in Examples (1) and (2).
(1) a. Richard Parker and Pi wandered the Algae Island yesterday through the meerkats.
b. Richard Parker and Pi wandered the Algae Island yesterday.
c. Richard Parker and Pi wandered the Algae Island through the meerkats.
(2) a. ⊗
The Orangutan reminded Pi of his mother yesterday through the meerkats.
b. The Orangutan reminded Pi of his mother yesterday.
c. ⊗ The Orangutan reminded Pi of his mother through the meerkats.
Movement verbs such as to wander allow for directional modifiers such as through the meerkats, whereas verbs such as to remind do not. In TAG, such restrictions can be modeled using selective adjoining constraints to specify which modifier tree may or may not be adjoined at a particular node in a given tree. Therefore, it is possible to license (1) and to rule out (2c). In Example (2a), however, under the dependent notion of adjunction, the tree for the directional adverbial through the meerkats will adjoin to the modifier tree for yesterday, which itself will adjoin to the tree selected by reminded. Thus, constraints placed by the verb on its modifiers must be passed through by modifier trees (here, the tree for yesterday) to also rule out sentences such as Example (2a). Propagating selective adjunction constraints in TAG would lead to a formalism for which derivation trees are no longer context-free (Schabes and Shieber 1994).
The second motivation for independent adjunction stems from probabilistic approaches. Stochastic lexicalized TAG specifies the probability of an adjunction of a given auxiliary tree at a given node in another elementary tree (Resnik 1992; Schabes 1992). Thus, under the standard notion of derivation, the overall probability of the string roasted red pepper would be determined by the probability of red adjoining to pepper and the probability of roasted adjoining to red. In contrast, independent adjunction would result in a derivation such that the overall probability of the string roasted red pepper would be determined by the probability of both red and roasted adjoining to pepper. Schabes and Shieber (1994, page 97) argue that it is plausible that “the most important relationships to characterize statistically are those between modifier and modified, rather than between two modifiers.”
A third motivation comes from semantics and, more particularly, from scope ambiguities involving modifiers. Given a sentence such as Example (3), where the relative scope of the modifiers twice and intentionally is ambiguous,3 Shieber (1994) shows that, under the extended definition of adjunction, a synchronous TAG modeling the relation between syntactic trees and logical formulae can account for both readings.
(3) John blinked twice intentionally.
2 The characters in these sentences are borrowed from Yann Martel’s book Life of Pi. 3 The sentence can describe either a single intentional act of blinking twice or two intentional acts each of single
blinking (Shieber 1994).
The account crucially relies on multiple independent adjunction of the two modifier trees to the tree for blink: Depending on which order the auxiliary trees for twice and intentionally adjoins to blink, the logical formula built will be either intentionally(twice(blink)) or twice(intentionally(blink)), thus capturing the ambiguity. To capture the different types of semantic dependencies and morpho-syntactic constraints that may hold between multiple auxiliary trees adjoining to the same entity, both dependent and independent derivations are needed.
As argued earlier, because there are no constraints or semantic relation holding between each of them, multiple intersective modifiers applying to the same entity (Example (4)) are best modeled using an independent derivation.
(4) The tall black meerkat slept. (Independent derivation)
In contrast, because they may involve strong scopal and morpho-syntactic constraints, stacked predicative verbs (i.e., verbs taking a sentential complement, Example (5a)) and non-intersective modifiers (Example (5c)) require dependent derivations. Consider sentences (5a–b), for instance. If predicative trees were assigned an independent derivation, sentence (5a) would be judged ungrammatical (because want requires an infinitival complement but would adjoin to the finite verb slept) and conversely, sentence (5b) would incorrectly be judged grammatical (because both want and try require an infinitival complement). Similarly, in Example (5c), the church is Syrian Orthodox, not Syrian and Orthodox. Assigning a dependent rather than an independent derivation to such cases straightforwardly captures the distinction between intersective and non-intersective modifiers.
(5) a. XJohn wanted to assume that Peter slept. (Dependent derivation)
b. ⊗ John wanted Peter tries to walk.
c. The meerkat admired the Syrian Orthodox church. (Dependent derivation)
Finally, some multiple adjunctions may involve both dependent and independent derivations, for example, when multiple modifiers and predicative verbs adjoin to the same verb (Example (6a)) or in the case of a derivation (Example (6b)) involving both intersective (old) and non-intersective (i.e., Syrian in Syrian Orthodox) modifiers.
(6) a. Yann said that John knows that Richard Parker and Pi wandered the Algae Island yesterday through the meerkats. (Mixed derivation)
b. The meerkat admired the old Syrian Orthodox church. (Mixed derivation)
As we shall see in Section 5.3, the parsing algorithm we propose licenses dependent, independent, and mixed derivations but is restricted to appropriately distinguish between various types of modifiers. Moreover, the feature information encoded in the grammar further restricts the derivation structures produced, thereby accounting for the interactions between adjunction, linear ordering, and morpho-syntactic constraints.","3 multiple adjunction in tree adjoining grammars :Vijay-Shanker and Weir (1991) introduce a compilation of TAG to Linear Indexed Grammars (LIGs; Gazdar 1988) that makes the derivation process explicit. Schabes and
Shieber (1994) modify this compilation to allow for both dependent and independent derivations. The resulting LIG is further exploited to specify a parsing algorithm that recovers those derivations.
In this section, we summarize Schabes and Shieber’s proposal. We start (Section 3.1) with an informal description of their approach. In Section 3.2, we introduce ordered derivation trees. Section 3.3 provides a brief introduction to LIG. Section 3.4 summarizes the TAG-to-LIG compilation proposed by Vijay-Shanker and Weir (1991). Finally, Section 3.5 describes the modifications introduced by Schabes and Shieber (1994) to allow both for dependent and for independent derivations. Tree Adjoining Grammar distinguishes between two types of auxiliary trees, namely, modifier vs. predicative auxiliary trees (Joshi and Vijay-Shanker 2001). Whereas predicative trees are assigned to verbs taking a sentential argument, modifier trees are assigned to all other auxiliary trees (e.g., verbal auxiliaries, adjectives, adverbs, prepositions, and determiners). More generally, the difference between a predicative and a modifier tree is that in a predicative tree, the foot node, like the substitution nodes, corresponds to an argument node selected by its lexical anchor (i.e., the word that selects that tree) whereas in a modifier auxiliary tree, the foot node is an open slot corresponding to the phrase being modified. When associating semantic entities with tree nodes (as proposed, for example, by Joshi and Vijay-Shanker [2001] and Gardent and Kallmeyer [2003]), this difference can be seen by noting the entities associated with root and foot nodes: These are distinct in a predicative tree but identical in modifier trees.
In their approach, Schabes and Shieber specify a TAG-to-LIG conversion that systematically associates dependent derivations with predicative auxiliary trees and independent derivations with modifier auxiliary trees. In addition, they introduce two mechanisms to ensure that each derivation tree unambiguously specifies a linguistically plausible derived tree.
First, they enforce ordering constraints between modifier trees adjoining at the same node (which are thus ambiguous with respect to the derived tree they describe) by assuming that derivation trees are ordered and that linear precedence (LP) statements can be used to constrain the order of siblings in a derivation tree. For instance, given the independent derivation shown in Figure 1, an LP statement stating that βred must occur before βroasted in the derivation tree will ensure that βroasted appears above βred in the derived tree and therefore that the resulting derived tree yields the phrase roasted red pepper rather than red roasted pepper.
Second, when both predicative and modifier trees adjoin at the same address, predicative trees always occur above all modifier trees in the derived tree (“outermost predication”). This ensures, for instance, that under the reading where yesterday refers to the arriving rather than the saying (i.e., when both say and yesterday adjoin to arrive), Example (7a) is derived but not Example (7b).
(7) a. XPeter says that yesterday John arrived late.
b. ⊗ Yesterday Peter says that John arrived late. In the standard version of TAG, each derivation tree describes a unique derived tree. In the case of a dependent derivation, unicity follows from the fact that dependent
derivations specify the order in which adjunction takes place (e.g., β2 adjoins to β1 and the result to α). As a result, if β2 adjoins to β1, there is only one possible derived tree—namely, a tree where β2 appears above β1.
When allowing for independent derivations, however, several derived trees are possible, depending on the order in which the auxiliary trees are adjoined. To ensure a unique mapping from derivation to derived tree, Schabes and Shieber (1994) therefore introduce the notion of ordered derivation trees. Ordered derivation trees differ from standard TAG derivation trees in that (i) they may contain sibling edges labeled with the same address, and (ii) they specify a total order on such siblings.
Figure 2 shows an example ordered derivation tree and associated derived tree. As indicated by the shared g address on their parent edge, auxiliary trees β1, . . . , βn adjoin to the same node—namely, the node with address g in the elementary tree τ. Because the derivation tree is ordered, β1 will appear below β2 in the derived tree, which in turn will be below β3, and so on. In short, given a set of auxiliary trees all adjoining to the same tree node, the derived tree produced from an ordered derivation tree following an independent derivation will be identical to the derived tree produced with the corresponding dependent derivation—that is, the dependent derivation where β1, . . . , βn appear in increasing index order from top to bottom. Like context-free grammars, LIGs (Gazdar 1988) are string rewriting systems where strings are composed of terminals and nonterminals. In a LIG, however, nonterminal symbols may be associated with a stack of symbols, called indices. A LIG rule can thus be represented as follows:
N[..µ] → N1[µ1] . . .Ni−1[µi−1]Ni[..µi]Ni+1[µi+1] . . .Nn[µn] (8)
N and Ni are nonterminals whereas µ and µi are strings of stack symbols. The symbol .. stands for the remainder of the stack symbols. Note that the remainder of the stack symbols associated with the left-hand side is associated with only one of the nonterminal (namely, Ni) on the right-hand side.
LIGs have been used in the literature (Weir and Joshi 1988; Vijay-Shanker and Weir 1991) to provide a common framework for the extensions of context-free grammars. In particular, Vijay-Shanker and Weir (1991, 1993) showed a weak equivalence between LIGs, TAGs, and combinatory categorial grammars (Steedman 2000) and proposed a LIG-based polynomial-time CYK recognition algorithm for TAGs and combinatory categorical grammars. In what follows, we show how Schabes and Shieber (1994) use a LIG variant of TAGs to license both dependent and independent derivations. The TAG-to-LIG compilation proposed by Vijay-Shanker and Weir (1991) produces LIG rules that simulate a traversal of the derived tree produced by the original TAG grammar. In these LIG rules, each node η of a TAG elementary tree is viewed as having both a top t[..η] and a bottom b[..η] component to account for the possibility of an adjunction. Figure 3 illustrates the traversal of the TAG-derived trees specified by the LIG resulting fromVijay-Shanker and Weir (1991) TAG-to-LIG compilation.
Figure 4 lists the LIG rules resulting from the TAG to LIG compilation process. Each nonterminal (t[..η] or b[..η]) with the top of the stack symbol in a LIG rule corresponds to a unique node in some elementary tree of the grammar. The inner stack symbols are used to keep track of the nodes higher in the derived tree where an auxiliary tree has been adjoined.
Rules of Types 1 and 2 capture immediate dominance between the bottom of a node η and the top of its immediate daughters in two configurations, depending on whether η dominates the foot node (Type 1) or not (Type 2). Rules of Type 3 handle nodes that require neither substitution nor adjunction. This rule handles cases where no adjunction occurs at a node by rewriting the top of this node to its bottom. Rules of Type 6 model substitution. Finally, rules of Types 4 and 5 handle adjunction. They specify that, for any given node η and any auxiliary tree β that may adjoin to η, the top of η rewrites to the top of the root node of β; and the bottom of the foot of β to the bottom of η. It follows that there can be no multiple adjunction in this LIG version of TAG. To associate predicative tree adjunctions with dependent derivations and multiple modifier adjunctions with independent derivations, Schabes and Shieber (1994) modify the compilation of TAG to LIG, proposed by Vijay-Shanker and Weir (1991) as sketched in Figure 5. Type 4(a) rules apply to adjunctions involving predicative trees. They are identical to Type 4 rules in the Vijay-Shanker and Weir’s approach and therefore enforce a standard (dependent) derivation for predicative trees. In contrast, Type 4(b) rules apply to adjunctions involving modifiers and result in an independent derivation.
Note also that the Outermost Predication constraint (i.e., predicative trees always occur above modifier trees adjoined at the same node) alluded to in Section 3.1 follows from the interactions between the Type 4(a) and Type 4(b) LIG rules.
Schabes and Shieber prove the weak-generative equivalence of TAGs under both Standard and Extended derivation using the LIG compilation. They also propose a recognition and a parsing algorithm with complexity of O(n6) in the length of the string.","4 multiple adjunction in feature-based tag :In this section, we explain why a straightforward extension of Schabes and Shieber’s proposal to FB-TAG would not work and we outline the intuitions and motivations underlying our approach. Section 5 will then introduce the details of our proposal. We start by a brief description of FB-TAG and of the unifications performed during derivation. FB-TAG was introduced by Vijay-Shanker (1987) and Vijay-Shanker and Joshi (1988, 1991) to support the use of feature structures in TAG. Figure 6 shows a toy FB-TAG for illustration.
An FB-TAG differs from a TAG in that tree nodes are decorated with feature structures. Nonterminal and foot nodes are decorated with two feature structures called top (T) and bottom (B), and substitution nodes are decorated with a single top feature structure. During derivation, feature structure unification constrains tree combination, as illustrated in Figure 7. Substitution unifies the top feature structure of a substitution node with the top feature structure of the root node of the tree being substituted. The adjunction of an auxiliary tree β to a tree node ηo unifies the top and bottom feature structures of ηo with the top feature structure of the root node of β and the bottom feature structure of its foot node, respectively. Finally, at the end of the derivation, the top and bottom feature structures of all nodes in the derived tree are unified.
Figure 8 shows the standard derivation for the phrase all the meerkats, using the grammar shown in Figure 6, and Figure 9 shows the corresponding derived and derivation trees. As can be seen, the feature constraints encoded in the grammar correctly ensure that all the meerkats can be derived (leftmost derivation tree in Figure 9) but not the all meerkats (rightmost in Figure 9). The incorrect derivation is blocked by the feature structure [det : nil] on the foot of the auxiliary tree βthe, which leads to a unification failure if βthe is adjoined at the root of βall with bottom feature structure [det : the]. To motivate our approach, we start by considering a simple extension of Schabes and Shieber’s LIG framework to FB-TAG, where each LIG rule enforces unifications mimicking those applied in FB-TAG. In particular, let us assume that Type 3 rules
(“No adjunction”) unify the top and the bottom feature structures of nodes where no adjunction occurs, and Type 4(b) rules (“Start root of adjunction”) unify the top feature (η.T) of the node (η) being adjoined to with the top feature structure (ηr.T) of the root node (ηr) of the auxiliary tree being adjoined: 4
b[..η] → t[..ηηr] η.T ∪ ηr.T (Type 4b) (9)
t[..η] → b[..η] η.T ∪ η.B (Type 3) (10)
As shown in Figure 10, this approach can incorrectly lead to derivation failures in the case of an independent multiple adjunction. Intuitively, the reason for this is that, in Schabes and Shieber’s approach, multiple adjunction starts and ends from the bottom component of the node being adjoined to. This is fine when no features are involved because the category of the node being adjoined to is always identical to the root and foot node of the auxiliary trees being adjoined. When nodes carry feature structures, however, a unification clash can occur that makes derivation fail. Thus, in our example, derivation incorrectly fails because the bottom feature structures of the root node of the auxiliary tree for all and the bottom feature structure of the root node of the auxiliary tree for the should unify but have conflicting value. As shown by the dependent derivation for all the meerkats depicted in Figure 8, this is incorrect. As we just saw, in the case of multiple independent adjunctions, a straightforward extension of Schabes and Shieber’s LIG framework to FB-TAG fails to correctly capture the unification constraints encoded in the grammar. More generally, when extending multiple independent adjunction to FB-TAG, it is crucial that the feature constraints encoded by the linguist describe the same set of derived trees no matter
4 We associate η.T ∪ ηr.T with the Type 4(b) rules to mimic the adjunction in FB-TAG, as shown in Figure 7.
which derivation tree is produced. We therefore propose a parsing algorithm that, given several auxiliary trees β1, . . . ,βn adjoining at the same node ηo, performs the same unifications independently of whether the derivation is dependent, independent, or mixed dependent/independent.
Figure 11 shows the unifications resulting from the multiple adjunction of β1 and β2 to a single node ηo. While it depicts the unifications enforced by our parsing algorithm for the derivation tree shown on the right hand side (i.e., for the independent adjunction of β1 and β2 to ηo), these unifications are in fact exactly the same as those that would be enforced by a dependent adjunction of β2 into β1 into ηo.
One key point illustrated by Figure 11 is that whereas multiple adjunction operates on a single node (here ηo), the unification constraints of FB-TAG require that the bottom feature structure of the foot of an auxiliary tree which appears higher in the derived tree (here, β2) unifies with the bottom feature structure of the root of the auxiliary tree appearing immediately below it in the derived tree (here β1)—not with that of the root of the node to which it adjoins (here ηo). In other words, while a multiple adjunction on ηo operates on ηo only, a correct implementation of FB-TAG unification constraints requires keeping track of the feature structures associated with the auxiliary trees successively adjoining to ηo.
In our proposal, we capture this bookkeeping requirement by associating tree nodes not with feature structures but with reference variables pointing to feature structures. The parsing algorithm is then specified so as to support dependent, independent, and mixed derivations while enforcing the same unifications as would be performed under a dependent adjunction. Before giving the technical details of our parsing algorithm (Section 5), we first highlight some differences between our and Schabes and Shieber’s approach. In particular, we show that (i) whereas Schabes and Shieber resort to three distinct mechanisms to account for word order constraints (i.e., selective adjoining constraints, linear
precedence statements on derivation trees, and a constraint on parsing), the FBTAG approach supports a uniform treatment of word order and (ii) our approach straightforwardly accounts for mixed dependent/independent derivations that would require some additional stipulation in Schabes and Shieber’s approach.
4.4.1 Ordering Constraints among Modifier Auxiliary Trees. In TAG, determiners and verbal auxiliaries are modifier rather than predicative auxiliary trees (cf. Section 3.1). Because Schabes and Shieber’s definitions systematically associate modifiers with independent derivations, all examples in (11a–d) undergo an independent derivation and constraints must therefore be provided to determine the order of the sibling nodes in the resulting derivation tree.
(11) a. XThe sonatas should have been being played by Sarah.
b. ⊗ The sonatas have should been being played by Sarah.
c. XAll the meerkats
d. ⊗ The all meerkats
To specify these constraints on similar cases (soft ordering constraints on adjectives and strict ordering constraints on temporal and spatial adverbial phrases in German), Schabes and Shieber (1994) suggest the use of LP constraints on derivation tree siblings. As illustrated by the derivation of Example (11c–d) in Figures 8 and 9, in the FBTAG approach, such additional constraints are unnecessary: They simply fall out of the feature constraints encoded in the grammar.
Note that even if determiners and auxiliary verbs were to be handled using dependent adjunction, the word ordering constraints used by Schabes and Shieber would fail to account for cases such as Example (12), where auxiliary verbs are interleaved with adverbs.
(12) John has often been selected for nomination.
In this case, if the auxiliary verbs has and been were treated as predicative trees, Schabes and Shieber’s constraint that predicative trees adjoin above modifier trees would preclude the derivation of Example (12) and incorrectly predict the derived sentence to be John has been often selected for nomination.
4.4.2 Ordering Constraints among Predicative Trees. As discussed by Schabes and Shieber (1994), auxiliary predicative trees may impose different constraints on the type of sentential complement they accept. Thus Example (13a) is correct but not Example (13b) because want expects an infinitival complement (previously shown in Example (5)).
(13) a. XJohn wanted to assume that Peter slept.
b. ⊗ John wanted Peter tries to walk.
Although in the Schabes and Shieber’s approach, selective adjoining constraints are used to license Example (13a) and rule out Example (13b), in the FB-TAG approach, this can be achieved using feature constraints.
4.4.3 Ordering Constraints between Predicative and Modifier Auxiliary Trees. In sentences such as Example (14a) where both modifier and predicative auxiliary trees adjoin to the same address, the predicative trees should generally adjoin above any modifier trees so that the predicative verb precedes the modifier in the derived string.
(14) a. XJohn promised that Peter will leave tomorrow. b. ⊗ Tomorrow John promised that Peter will leave.
To ensure the appropriate linearization, the Schabes and Shieber’s approach introduces the outermost-predication rule, which stipulates that predicative trees adjoin above modifier auxiliary trees. In contrast, the FB-TAG approach allows both orders and lets feature constraints rule out ungrammatical sentences such as Example (14b). This allows the approach to directly extend to a counter-example discussed by Schabes and Shieber (1994), where a modifier (here, At what time) must in fact adjoin above a predicative tree.
(15) At what time did Brockway say Harrison arrived?
Figure 12 shows two possible derivation trees for the sentence (15) under the interpretation where it is the time of arriving (rather than the time of saying) which is questioned. These derivation trees show the two possible relative orderings of the (predicative) auxiliary tree for say and the (modifier) auxiliary tree at what time. Because the Outermost-Predication rule requires that predicative trees adjoin above modifier trees (and thus occur outermost in the derivation tree), in Schabes and Shieber’s approach, only the right-hand side derivation is possible, thus failing to derive sentence (15). In contrast, because our approach does not explicitly constrain the relative ordering of predicative and modifier auxiliary trees adjoining to the same node, both derivations are possible, thereby licensing both Example (15) and the sentence Did Brockway say at what time Harrison arrived?
4.4.4 Mixed Dependent and Independent Multiple Adjunctions. In Schabes and Shieber’s approach, all modifier auxiliary trees undergo independent derivation. As shown in Section 2.2, however, non-intersective modifiers arguably license a dependent derivation while some cases of multiple adjunction may involve both a dependent and an independent derivation. As we shall see in Section 5, our FB-TAG approach accounts for such cases by allowing both for independent and dependent derivations, by ruling out dependent derivations for intersective modifiers and by using feature constraints to regulate the interactions between multiply adjoining auxiliary trees.","5 extending schabes and shieber’s lig framework for fb-tags :We now propose a compilation of FB-TAG-to-LIG that makes both dependent and independent derivations in FB-TAG explicit. We use this resulting LIG to specify an
Earley algorithm for recovering multiple adjunctions in FB-TAG. This compilation differs in two main ways from that proposed by Schabes and Shieber (1994). First, tree nodes are associated with reference variables pointing to feature structures. Second, the LIG rules are modified and extended with unification operations. To account for FB-TAG unifications while allowing for independent derivations, we replace the feature structures of FB-TAG with reference variables pointing to those feature structures. Each node in the elementary trees is decorated with two reference variables: The top reference variable PT contains the reference to the top feature structure T and the bottom reference variable PB contains the reference to the bottom feature structure B. The top and the bottom feature structures of a node η can be traced by val(η.PT) and val(η.PB), respectively, where PT and PB are the top and the bottom reference variables decorating the node η, and the function val(P) returns the feature structures referred to by the reference variable P.
When specifying the parsing algorithm, we use reference variables to ensure the appropriate unifications, as follows. In an independent derivation where the node ηo is adjoined to, first by β1 and second by β2, the bottom feature structure ηo.B of ηo (i) unifies with the bottom feature structure ηf 1.B of the foot of β1 and (ii) is reassigned (:=) to the bottom reference variable ηr1.PB of the root of β1. When β2 is adjoined, its foot node will therefore correctly be unified, not with the bottom feature structure of ηo but with that of ηr1. To support both dependent and independent derivations while enforcing the correct unifications, we modify the TAG-to-LIG compilation in such a way that the resulting LIG rules capture the tree traversal depicted in Figure 13. Independent derivations are accounted for by the fact that adjunction starts and ends at the bottom component of the node being adjoined to (Type 4 and 5 rules). Our LIG compilation automatically supports dependent derivations by allowing sequential adjunctions at the roots of auxiliary trees.
Type 4: Start root of adjunction. For each elementary tree node η that allows the adjunction of the auxiliary tree with the root node ηr, the following LIG production rule is generated.
b[..η] → t[..ηηr] val(η.PT) ∪ val(ηr.PT),η.PB := ηr.PB
Type 5: Start foot of adjunction. For each elementary tree node η that allows the adjunction of the auxiliary tree with the foot node ηf , the following LIG production rule is generated.
b[..ηηf ] → b[..η] val(η.PB) ∪ val(ηf .PB)
Type 6: Start substitution. For each elementary tree node η that allows the substitution of the initial tree with the root node ηr, the following LIG production rule is generated (not shown in Figure 13).
t[η] → t[ηr] val(η.PT) ∪ val(ηr.PT)
To perform multiple adjunction while enforcing the appropriate feature unifications (as depicted in Figure 11), we split Type 3 rules into two subtypes. Type 3(a) rules apply to the root of auxiliary trees and perform no unification. By no unification, they ensure that feature structures are not blocked for the possibility of the adjunction of the following auxiliary tree and allow for the correct unifications to be carried out for independent derivations. Type 3(b) rules function as termination of multiple adjunction by unifying the top and bottom feature structures of the node. It is applicable to all tree nodes except roots of auxiliary trees.
Type 3(a): Terminating adjunction at the root of the auxiliary tree. For each root node η of the auxiliary trees, the following LIG production rule is generated.
t[..η] → b[..η] ∅
Type 3(b): Terminating adjunction at any other node. For each node η that is not a root node of some auxiliary tree and is not marked for substitution, the following LIG production rule is generated.
t[..η] → b[..η] val(η.PT) ∪ val(η.PB)
Given this set of rules, both dependent and independent derivations are possible. For example, given two auxiliary trees β1 and β2 adjoining at the node η in an elementary tree τ, a dependent derivation will occur whenever the Type 4 rule applies to predict the adjunction of, for example, β2 at the root of β1. Conversely, if the Type 3(a) rule applies at the root of β1, recognition will move from the top of the root of β1 to its bottom, allowing for Type 5 rule to complete the adjunction of β1 at the node η; and the Type 4 rule applies to predict the adjunction of β2 at the node η of τ, registering an independent derivation. In this section, we present our parsing algorithm for FB-TAGs with dependent and independent derivations. We start with an informal description of how the algorithm handles the interactions between unification and independent derivations. We then go on to specify the inference rules making up the algorithm. We do this in two steps. First, we present a basic set of rules allowing for both dependent and independent derivations. Second, we show how to constrain this algorithm to minimize spurious ambiguity.
5.3.1 Independent Derivations and Feature-Structure Unification. Before specifying the parsing algorithm, we illustrate by means of an example the interplay between multiple independent adjunction and feature structure unifications.
Figure 14 displays the feature unifications and reassignment performed during the recognition process of a multiple independent adjunction. The linear ordering of the equations reflects the order of the parsing completion operations.
Given the auxiliary tree β1 and the adjunction site ηo, the picture shows that unifying the bottom feature structure of the foot node of β1 with the bottom feature structure of ηo (Step 1: Type 5,ηo.B ∪ ηf 1.B) occurs before the bottom reference variable of ηo is reassigned to the bottom feature structure of the root of β1 (Step 4: Type 4,ηo.PB → ηr1.B). Also, the reassignment ensures that the follow-up adjunction of β2 at the node ηo has access to the bottom feature of the root of the previous auxiliary tree β1 (Step 5: Type 5,ηr1.B ∪ ηf 2.B). At the end of the adjunction (Step 9), the Type 3(b) rule ensures that the top and the bottom features of the root of the last auxiliary tree (here, β2) adjoined are unified (ηr2.T ∪ ηr2.B).
t b ηo ηo.PT → ηo.T ηo.PB → ηo.B
t b
ηr1 ηr1.PT → ηr1.T
ηr1.PB → ηr1.B
t b
ηf 1 ηf 1.PT → ηf 1.T
ηf 1.PB → ηf 1.B
β1
t b
ηr2 ηr2.PT → ηr2.T
ηr2.PB → ηr2.B
t b
ηf 2 ηf 2.PT → ηf 2.T
ηf 2.PB → ηf 2.B
β2
(1) (2)
(5)
(6)
(4)
(3)
(8)
(7)
(9)
(9). Type 3(b),ηo.PT → F7,ηo.PB → F7, F7 = ηo.T ∪ ηr1.T ∪ ηr2.T ∪ ηr2.B
(8). Type 4,ηo.PT → F6,ηr2.PT → F6, F6 = ηo.T ∪ ηr1.T ∪ ηr2.T,ηo.PB → ηr2.B
(7). Type 3(a),ηr2.PT → ηr2.T,ηr2.PB → ηr2.B
(6). Type 3(b),ηf 2.PT → F5,ηf 2.PB → F5, F5 = ηr1.B ∪ ηf 2.T ∪ ηf 2.B
(5). Type 5,ηo.PB → F4,ηf 2.PB → F4, F4 = ηr1.B ∪ ηf 2.B (4). Type 4,ηo.PT → F3,ηr1.PT → F3, F3 = ηo.T ∪ ηr1.T,ηo.PB → ηr1.B
(3). Type 3(a),ηr1.PT → ηr1.T,ηr1.PB → ηr1.B
(2). Type 3(b),ηf 1.PT → F2,ηf 1.PB → F2 , F2 = ηo.B ∪ ηf 1.T ∪ ηf 1.B
(1). Type 5,ηo.PB → F1,ηf 1.PB → F1, F1 = ηo.B ∪ ηf1.B
Figure 14 Multiple independent adjunction of β1 and β2 to ηo. The unifications and reassignments are listed in the order in which they are performed during the recognition process.
As we shall see subsequently, this correct ordering between unification and reassignment follows from the proposed Earley algorithm. Type 4 completor rules complete the prediction triggered at the root of an auxiliary tree (“Start root of adjunction”) and Type 5 completor rules complete the prediction triggered at the foot node of an auxiliary tree (“Start foot of adjunction”). Because completion operates bottom–up, it follows that Type 5 rules apply before Type 4 rules. Thus, when adjoining an auxiliary tree β1 to a node ηo, the Type 5 completor rules, unifying the bottom feature structure of the foot node of β1 with the bottom feature structure of the node ηo, occurs before the Type 4 completor rules, which reassign the bottom reference variable of ηo to the bottom feature structure of the root of β1.
5.3.2 Inference Rules. The parsing algorithm for FB-TAG is a modification of the algorithm presented by Schabes and Shieber (1994). It is a chart-based parsing method based on the Earley type deduction system. Each item in the chart is of the format 〈N[..η] → Γ •∆, i, j, k, l〉, where N is some LIG nonterminal (i.e., t or b), and Γ and ∆ present the sequences of LIG nonterminals associated with stacks of node indices. The indices i, j, k, and l are markers in the input string, showing the recognized portion:5 The recognized item starts in position i, ends in position l, and if η dominates a foot node, the tree dominated by the foot node starts in j and ends in k. If the foot node is not dominated by the recognized nonterminal sequence Γ, the values for j and k are taken to be the dummy value ’−’. As in Earley algorithms, the • separates the nonterminal sequence Γ which was parsed from the nonterminal sequence ∆ yet to be parsed.
The first three types of rules (Scanner, Predictor, and Type 1/2 Completor) are identical to those introduced by Schabes and Shieber (1994) and do not involve any unification operations.
The Type 3(b) completor rule enforces top and bottom unification on all nodes that are not the root of an auxiliary tree, and the Type 3(a) completor rule prevents top and bottom unification at the root of auxiliary trees.
The Type 4 completor rule unifies the top feature structure of the root of the auxiliary tree with the top feature structure of the adjunction site. In addition, it ensures that on completion of an adjunction at node η, the bottom feature structure of η is reassigned to the bottom feature structure labeling the root of the auxiliary tree. In this way, the unifications occurring in an independent derivation will mirror those occurring in a dependent one in that any following adjunction will induce unifications as if it were happening at the root node ηr of the preceding auxiliary tree (not at η).
On completion of a foot node prediction (the tree dominated by the foot of the auxiliary tree has been recognized), the Type 5 completor rule unifies the bottom feature structure of the foot of the auxiliary tree with the bottom feature structure of the adjunction site.
Finally, the Type 6 completor unifies the top feature structure of a substitution node with the top feature structure of the root of the tree being substituted.
r Scanner:
〈b[..η] → Γ • w∆, i, j, k, l〉
〈b[..η] → Γw •∆, i, j, k, l + 1〉 , w = wl+1, ∅
5 The indices 〈i, j, k, l〉 have been used in previous parsing algorithms for tree-adjoining grammars (Vijay-Shankar and Joshi 1985; Schabes and Joshi 1988; Schabes 1991). They deliver the same functionality here.
If w (a terminal symbol) occurs at position l+1, the scanner rule creates a new item whose span extends to l+1.
r Predictor:
〈N[..η] → Γ • N′[µ]∆, i, j, k, l〉
〈N′[µ] → •Θ, l,−,−, l〉 , ∅
Predictor rules are produced for all types of production rules. N and N′ are LIG variables taking the value t or b. Γ, ∆, and Θ are the sequences of LIG nonterminals associated with stacks of node indices. µ is a sequence of node indices.
r Type 1 and 2 Completor:
〈b[..η] → Γ • t[η1]∆, m, j ′, k′, i〉 〈t[η1] → Θ•, i, j, k, l〉
〈b[..η] → Γt[η1] •∆, m, j ⊕ j′, k ⊕ k′, l〉 ,
∅,
η1 not a root node
Types 1 and 2 Completor rules permit completing Rules 1 and 2 whenever the top of a child node is fully recognized. Here, t[η1] has been fully recognized as the substring between i and l (i.e., wi+1 . . .wl). Therefore, t[η1] can be completed in b[..η]. If one of t[η1] or b[..η] dominates the foot node of the tree, the final b[..η] will have indices associated with the substring recognized by the foot subtree. The operation ⊕ is defined as follows:
x ⊕ y =

  
  
x, if y = −
y, if x = −
x, if x = y
undefined, otherwise
r Type 3(a) Completor:
〈t[..η] → •b[..η], i,−,−, i〉 〈b[..η] → Θ•, i, j, k, l〉
〈t[..η] → b[..η]•, i, j, k, l〉 ,
∅,
η an auxiliary tree root node
The Type 3(a) Completor rule is used to complete the prediction of an auxiliary tree rooted in η. Once the auxiliary tree dominated by b[..η] has been recognized, the auxiliary tree itself is completely recognized. As explained earlier, there is in this case no feature unification between the top and the bottom of the root of the auxiliary tree.
r Type 3(b) Completor:
〈t[..η] → •b[..η], i,−,−, i〉 〈b[..η] → Θ•, i, j, k, l〉
〈t[..η] → b[..η]•, i, j, k, l〉 ,
val(η.PT ) ∪ val(η.PB),
η not an auxiliary tree root node
The Type 3(b) Completor rule ensures the unification of the top and bottom feature structures for all nodes that are not the root node of an auxiliary tree.
r Type 4 Completor:
〈b[..η] → •t[..ηηr], i,−,−, i〉
〈t[..ηηr] → Θ•, i, j, k, l〉
〈b[..η] → ∆•, j, p, q, k〉
〈b[..η] → t[..ηηr]•, i, p, q, l〉 ,
val(η.PT) ∪ val(ηr.PT)
η.PB := ηr.PB
The auxiliary tree associated with the predicted adjunction (t[..ηηr]) at the node η and the subtree dominated by the node η (below b[..η]) are completed, hence b[..η] can be completely recognized with this adjunction. The associated feature unification unifies the content of the top reference variable of the adjoining node site η with the content of the top reference variable of the root node ηr of the adjoined auxiliary tree. After the successful adjunction of this adjoining tree, the bottom reference variable of the adjoining node site η is reassigned to the content of the bottom reference variable of the root node ηr of the adjoined auxiliary tree.
r Type 5 Completor:
〈b[..ηηf ] → •b[..η], i,−,−, i〉 〈b[..η] → Θ•, i, j, k, l〉
〈b[..ηηf ] → b[..η]•, i, i, l, l〉 , val(η.PB) ∪ val(ηf .PB)
The foot node prediction can be completed when the adjunction has been performed and the bottom part of the adjoining node site η has been recognized. The associated feature unification unifies the content of the bottom reference variable of the adjoining node site η with the content of the bottom reference variable of the foot node ηf of the auxiliary tree being adjoined.
r Type 6 Completor:
〈t[η] → •t[ηr], i,−,−, i〉 〈t[ηr] → Θ•, i,−,−, l〉 〈t[η] → t[ηr]•, i,−,−, l〉 , val(η.PT) ∪ val(ηr.PT)
The Type 6 Completor rule completes the substitution at the node η. The associated feature unification unifies the content of the top reference variable of the node η with the content of the top reference variable of the root node ηr of the initial tree.
Given these inference rules, the recognition process is initialized using axioms of the form 〈t[ηs] → •Γ, 0,−,−, 0〉 for each rule t[ηs] → Γ where ηs is the root node of an initial tree labeled with the start symbol. Given an input string w1 . . .wn to be recognized, the goal items in the chart are of the form 〈S → t[ηs]•, 0,−,−, n〉. Once at least one goal item is found in the chart, the recognition process succeeds and the string is successfully accepted by the grammar; otherwise it is rejected. We refer the reader to Appendix A (cf. Figure A.2) for a detailed example of the recognition of the sentence all the meerkats using the proposed inference system.
Note also that although the recognition algorithm we described uses unreduced rules (i.e., generated grammar rules maintaining the full information of nonterminals and the associated index stacks), it is possible to define a more efficient algorithm by having reduced LIG rules and chart items listing only the single top stack element for each constituent (Vijay-Shanker and Weir 1991, 1993). The resulting recognition
algorithm is still complete because the proposed TAG-to-LIG compilation maintains a one-to-one correspondence between the generated rules and their reduced forms (Schabes and Shieber 1994).
As mentioned by Schabes and Shieber (1994), this recognition algorithm can be turned into a parsing algorithm by associating a set of operations with each chart item to build up associated derived trees.
Note also that the derivation trees built as a side effect of the parsing process are the (dependent and/or independent) derivation trees of an FB-LTAG and are therefore context-free.
5.3.3 Handling Spurious Parses. As explained at the end of Section 5.2, the parsing algorithm presented in the previous section systematically allows for dependent and independent adjunction. For example, the recognition of the sentence all the meerkats (Figure A.2) produces both dependent and independent derivations that are not rejected by the unification constraints. In Section 2.2, however, we argued that different types of auxiliary trees license different types of derivations. To capture these distinctions, we modify the recognition algorithm so that it associates scopal auxiliary trees (Example (16a–b)) with dependent derivations only and multiple intersective modifier auxiliary trees (Example (16c)) with only an independent derivation.
(16) a. John thinks that Peter said that the meerkat left. b. The meerkat admired the Syrian orthodox church. c. The tall black meerkat slept.
To block dependent adjunctions between intersective modifiers, we modify the TAG-to-LIG transformation so that, given two intersective modifier trees β1 and β2, no Type 4 or Type 5 rule is produced.
Type 4: Start root of adjunction. For each elementary tree node η in tree β1 that allows the adjunction of the auxiliary tree β2 with root node ηr, the following LIG production rule is generated if and only if β1 and β2 are not intersective modifier auxiliary trees.
b[..η] → t[..ηηr] val(η.PT) ∪ val(ηr.PT)
Type 5: Start foot of adjunction. For each elementary tree node η that allows the adjunction of the auxiliary tree β2 with the foot node ηf , the following LIG production rule is generated if and only if β1 and β2 are not intersective modifier auxiliary trees.
b[..ηηf ] → b[..η] val(η.PB) ∪ val(ηf .PB)
Thus, for instance, in the derivation of all the meerkats depicted in Appendix Figure A.2, the following rules will not be produced, thereby blocking the production of the dependent derivation.
b[.. NPther ] → t[.. NP the r NP all r ] b[.. NPallr ] → t[.. NP all r NP the r ]
} Type4 b[.. NPallr NP the f ] → b[.. NP all r ]
b[.. NPther NP all f ] → b[.. NP the r ]
}
Type5
Similarly, to block independent adjunctions between scopal auxiliary trees, we add a flag scopal? to states in the parsing algorithm. The Type 4 Completor rules associated
with scopal modifiers are modified to mark the progress of a scopal adjunction and to block the independent adjunction of another scopal modifier at the same node.
Type 4 Completor:
〈b[..η] → •t[..ηηr], i,−,−, i, scopal?〉
〈t[..ηηr] → Θ•, i, j, k, l, scopal?〉
〈b[..η] → ∆•, j, p, q, k, scopal?〉
〈b[..η] → t[..ηηr]•, i, p, q, l, True〉 ,
val(η.PT) ∪ val(ηr.PT)
η.PB := ηr.PB
In the Type 4 Completor rule, once a scopal auxiliary treeβ with root node ηr adjoins at some node η, the bottom component of the node η is marked with True, recording that a scopal adjunction has occurred at node η and that it therefore should not accept any further scopal adjunction.
Thus, for instance, the derivation of Syrian orthodox churches will proceed in a similar manner as the derivation of all the meerkats depicted in Appendix Figure A.2, but it will fail to produce the chart items (40, 42, ..., 52)associated with the independent adjunction. Therefore, only the dependent derivation will be produced.
Note that this modification does not block modifier adjunction above a predicative adjunction. Therefore, it successfully recognizes the sentence At what time did Brockway say Harrison arrived?, shown in Example (15), where a wh-modifier needs to be adjoined above a predicative adjunction. Figure 15 shows the complete recognition algorithm modified to rule out spurious parses in the case of multiple scopal auxiliary trees and intersective modifier auxiliary trees.
5.3.4 Weak Generative Equivalence. The weak-generative equivalence refers to the set of strings characterized by the formal system. In contrast, the strong-generative equivalence relates to the set of structural descriptions (such as derivation trees, dags, proof trees, etc.) assigned by a formal system to the strings that it specifies (Vijay-Shankar and Joshi 1985; Joshi 2000).
Using an argument similar to that put forward by Schabes and Shieber (1994), we can prove the weak-generative equivalence of TAGs under the dependent and our independent derivations. We call the set of languages generated by the standard derivation in TAG, TALstd; the set of languages generated by Schabes and Shieber’s extended derivation in TAG, TALextss ; the set of languages generated with our modifications for FB-TAG, TALext; and the set of languages generated by the LIG, LIL. Our derivation allows both dependent and independent derivations; therefore, our derivation will recognize all the strings recognized by the standard (dependent) derivation. More precisely, our derivation can mimic the standard derivation by not allowing more than one adjunction on a tree node by treating all auxiliary trees as scopal auxiliary trees (cf. Section 5.3.3), henceforth, TALstd ⊆ TALext. The proposed compilation from TAGs to LIGs for the independent derivation concluded TALext ⊆ LIL. Finally, LIL ⊆ TALstd has been proven by Vijay-Shanker (1987). Combining these three inclusions, we can conclude that TALstd = TALext. In addition, Schabes and Shieber (1994) have shown that TALstd = TALextss . Hence, we can conclude the weak-generative equivalence of all three derivations in TAGs, TALstd = TALextss = TALext. Feature structures enhance TAGs’ descriptive ability without affecting their generative capacity (Vijay-Shanker and Joshi 1988). The proposed algorithm simulates the established unification mechanism in FBTAG without affecting the representation and the stipulations (e.g., null adjunction at
the foot node and the bounded feature structures) of the grammar itself. Therefore, the association with feature structures will not affect this equivalence.","6 conclusion :Although independent derivations have been shown by Schabes and Shieber (1994) to be essential for correctly supporting syntactic analysis, semantic interpretation, and statistical language modeling, the parsing algorithm they propose is restricted to TAG and is therefore not directly applicable to large scale implemented Feature-Based TAGs. We have provided a recognition algorithm for FB-TAGs that supports both dependent and independent derivations under certain restrictions enforced jointly by feature constraints and by side conditions in the parsing algorithm. The resulting algorithm combines the benefits of independent derivations with those of Feature-Based Grammars. In particular, we showed that it accounts for a range of interactions between dependent vs. independent derivation on the one hand, and syntactic constraints, linear ordering, and scopal vs. nonscopal semantic dependencies on the other hand.
Appendix A: Recognition of the String all the meerkats
We show the recognition of the string all the meerkats as per the inference rules described in Section 5.3.2.
Figure A.1 shows the grammar used for the derivation. To support multiple adjunction in FB-TAG, it implements two main modifications. First, to facilitate the TAG to LIG compilation, each tree node in the grammar is marked with a unique identifier. For example, in Figure A.1, NPmk, NPther , Det the, NPthef *, NP all r , Det all, and NPthef * are unique node identifiers in the grammar. Second, to implement the reassignment mechanism in
NPmk P
Tmk →[Tmk :[]]
P Bmk
→[Bmk:[]]
meerkat
αmeerkat
NPther P Tther →[Tther :[]]
P Bther
→[Bther :[det:the]]
Detthe P Tthe det
→[Tthedet :[]]
P Bthe det →[Bthedet:[]]
NPthef * P Tthe f →
[ Tthef :[det:nil] ]
P Bthe
f →
[ Bthef :[] ]
the
βthe
NPallr P Tallr →[Tallr :[]]
P Ballr
→[Ballr :[det:all]]
Detall P Tall det
→[Talldet:[]]
P Ball det →[Balldet:[]]
NPallf * P Tall f →
[ Tallf :[] ]
P Ball
f →
[ Ballf :[] ]
all
βall
Figure A.1 A feature-based Tree-Adjoining Grammar.
the parsing algorithm, the top (T) and the bottom (B) feature structures of each node are assigned reference variables PT and PB, respectively.
Figure A.2 shows the LIG production rules for the FB-TAG shown in Figure A.1. Table A.1 shows the step-wise recognition of the string all the meerkats. Recall Section 5.3.2: The LIG rules shown in Figure A.1 does not deal with spurious parses and produces all valid derivations, dependent or independent, that are not blocked by feature unification constraints. Hence, for the string all the meerkats, it generates both independent (Step 52) and dependent (Step 53) derivations. As explained in Figures A.3 and A.4, both derivations undergo the identical set of feature unifications. In both figures, prediction rules are abbreviated because they do not enforce any feature unification.",,,"In parsing with Tree Adjoining Grammar (TAG), independent derivations have been shown by Schabes and Shieber (1994) to be essential for correctly supporting syntactic analysis, semantic interpretation, and statistical language modeling. However, the parsing algorithm they propose is not directly applicable to Feature-Based TAGs (FB-TAG). We provide a recognition algorithm for FB-TAG that supports both dependent and independent derivations. The resulting algorithm combines the benefits of independent derivations with those of Feature-Based grammars. In particular, we show that it accounts for a range of interactions between dependent vs. independent derivation on the one hand, and syntactic constraints, linear ordering, and scopal vs. nonscopal semantic dependencies on the other hand.","[{""affiliations"": [], ""name"": ""Claire Gardent""}, {""affiliations"": [], ""name"": ""Shashi Narayan""}]",SP:0296a131d413b037f2c2d5e31e8c13c042469177,"[{""authors"": [""Alahverdzhieva"", ""Katya.""], ""title"": ""XTAG using XMG"", ""venue"": ""Master\u2019s thesis, Universit\u00e9 de Nancy."", ""year"": 2008}, {""authors"": [""Candito"", ""Marie-H\u00e9l\u00e8ne"", ""Sylvain Kahane.""], ""title"": ""Can the TAG derivation tree represent a semantic graph? An answer in the light of Meaning-Text Theory"", ""venue"": ""Proceedings of the Fourth Workshop"", ""year"": 1998}, {""authors"": [""Crabb\u00e9"", ""Beno\u0131\u0302t"", ""Denys Duchier"", ""Claire Gardent"", ""Joseph Le Roux"", ""Yannick Parmentier""], ""title"": ""XMG: eXtensible MetaGrammar"", ""venue"": ""Computational Linguistics,"", ""year"": 2013}, {""authors"": [""Gardent"", ""Claire.""], ""title"": ""Integrating a unification-based semantics in a large scale Lexicalised Tree Adjoining Grammar for French"", ""venue"": ""Proceedings of the 22nd International Conference on"", ""year"": 2008}, {""authors"": [""Gardent"", ""Claire"", ""Laura Kallmeyer.""], ""title"": ""Semantic construction in Feature-Based TAG"", ""venue"": ""Proceedings of the 10th Conference of the European Chapter of the Association for Computational Linguistics (EACL),"", ""year"": 2003}, {""authors"": [""Gazdar"", ""Gerald.""], ""title"": ""Applicability of indexed grammars to natural languages"", ""venue"": ""Uwe Reyle and Christian Rohrer, editors, Natural Language Parsing and Linguistic Theories, volume 35 of Studies"", ""year"": 1988}, {""authors"": [""Joshi"", ""Aravind K.""], ""title"": ""Relationship between strong and weak generative power of formal systems"", ""venue"": ""Proceedings of the 5th International Workshop on Tree Adjoining Grammars and Related Formalisms (TAG+5),"", ""year"": 2000}, {""authors"": [""Joshi"", ""Aravind K."", ""Yves Schabes.""], ""title"": ""Tree-Adjoining Grammars"", ""venue"": ""Handbook of Formal Languages and Automata, 3:69\u2013124."", ""year"": 1997}, {""authors"": [""Joshi"", ""Aravind K"", ""K. Vijay-Shanker""], ""title"": ""Compositional semantics with lexicalized tree-adjoining grammar (LTAG): How"", ""year"": 2001}, {""authors"": [""Kallmeyer"", ""Laura"", ""Marco Kuhlmann.""], ""title"": ""A formal model for plausible dependencies in lexicalized tree adjoining grammar"", ""venue"": ""Proceedings of the 11th International Workshop on"", ""year"": 2012}, {""authors"": [""Kroch"", ""Anthony.""], ""title"": ""Asymmetries in long distance extraction in a tree adjoining grammar"", ""venue"": ""M. Baltin and A. Kroch, editors, Alternative conceptions of phrase structure, University of Chicago Press,"", ""year"": 1989}, {""authors"": [""Rambow"", ""Owen"", ""K. Vijay-Shanker"", ""David Weir.""], ""title"": ""D-Tree Grammars"", ""venue"": ""Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics (ACL), pages 151\u2013158, Cambridge, MA."", ""year"": 1995}, {""authors"": [""Resnik"", ""Philip.""], ""title"": ""Probabilistic tree-adjoining grammar as a framework for statistical natural language processing"", ""venue"": ""Proceedings of the 14th Conference on Computational linguistics (COLING),"", ""year"": 1992}, {""authors"": [""Schabes"", ""Yves.""], ""title"": ""The valid prefix property and left to right parsing of tree-adjoining grammar"", ""venue"": ""Proceedings of the 2nd International Workshop on Parsing Technologies (IWPT), pages 21\u201330, Cancun."", ""year"": 1991}, {""authors"": [""Schabes"", ""Yves.""], ""title"": ""Stochastic lexicalized tree-adjoining grammars"", ""venue"": ""Proceedings of the 14th Conference on Computational Linguistics (COLING), volume 2, pages 425\u2013432, Nantes."", ""year"": 1992}, {""authors"": [""Schabes"", ""Yves"", ""Aravind K. Joshi.""], ""title"": ""An Earley-type parsing algorithm for tree adjoining grammars"", ""venue"": ""Proceedings of the 26th Annual Meeting of the Association for Computational Linguistics (ACL),"", ""year"": 1988}, {""authors"": [""Schabes"", ""Yves"", ""Stuart M. Shieber.""], ""title"": ""An alternative conception of tree-adjoining derivation"", ""venue"": ""Proceedings of the 30th Annual Meeting of the Association for Computational Linguistics (ACL), pages 167\u2013176,"", ""year"": 1992}, {""authors"": [""Schabes"", ""Yves"", ""Stuart M. Shieber.""], ""title"": ""An alternative conception of tree-adjoining derivation"", ""venue"": ""Computational Linguistics, 20(1):91\u2013124."", ""year"": 1994}, {""authors"": [""Shieber"", ""Stuart M""], ""title"": ""Restricting the weak-generative capacity of synchronous"", ""year"": 1994}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""The syntactic process"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 2000}, {""authors"": [""The XTAG Research Group.""], ""title"": ""A lexicalized tree adjoining grammar for English"", ""venue"": ""Technical report ICRS-01-03, Institute for Research in Cognitive Science, University of Pennsylvannia."", ""year"": 2001}, {""authors"": [""K. Vijay-Shankar"", ""Aravind K. Joshi.""], ""title"": ""Some computational properties of tree adjoining grammars"", ""venue"": ""Proceedings of the 23rd Annual Meeting of the Association for Computational Linguistics (ACL),"", ""year"": 1985}, {""authors"": [""K. Vijay-Shanker""], ""title"": ""A Study of Tree Adjoining Grammars"", ""venue"": ""Ph.D. thesis, Department of Computer and Information Science, University of Pennsylvania, Philadelphia."", ""year"": 1987}, {""authors"": [""K. Vijay-Shanker"", ""Arvind K. Joshi.""], ""title"": ""Feature structures based tree adjoining grammars"", ""venue"": ""Proceedings of the 12th International Conference on Computational Linguistics (COLING),"", ""year"": 1988}, {""authors"": [""K. Vijay-Shanker"", ""Arvind K. Joshi.""], ""title"": ""Unification-based tree adjoining grammar"", ""venue"": ""Technical report MS-CIS-91-25, School of Engineering and Applied Science, Department of Computer and Information"", ""year"": 1991}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""Polynomial parsing of extensions of Context-Free Grammars"", ""venue"": ""Masaru Tomita, editor, Current Issues in Parsing Technology, volume 126 of The Springer"", ""year"": 1991}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""Parsing some constrained grammar formalisms"", ""venue"": ""Computational Linguistics, 19(4):591\u2013636."", ""year"": 1993}, {""authors"": [""Weir"", ""David J."", ""Aravind K. Joshi.""], ""title"": ""Combinatory Categorial Grammars: Generative power and relationship to linear context-free rewriting systems"", ""venue"": ""Proceedings of the 26th Annual Meeting"", ""year"": 1988}]","acknowledgments  :We would like to thank Laura Perez-Beltrachini for useful discussion on the related topic. We are grateful to our anonymous reviewers for their insightful comments, which helped us improve the quality of the article in terms of both presentation and content.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"1 introduction :A Tree Adjoining Grammar (TAG; Joshi and Schabes 1997) consists of a set of elementary trees and two combining operations, substitution and adjunction. Consequently, a TAG derivation can be described by a tree (called a derivation tree) specifying which elementary TAG trees were combined using which operations to yield that derivation. In this tree, each vertex is labeled with a tree name and each edge with a description of the operation (node address and operation type) used to combine the trees labeling its end vertices. As we shall see in Section 3.2, in TAG, each derivation tree specifies a unique parse tree, also called derived tree.
In previous work, it has been argued that TAG derivation trees provide a good approximation of semantic dependencies between the words of a sentence (Kroch 1989; Rambow, Vijay-Shanker, and Weir 1995; Candito and Kahane 1998; Kallmeyer and Kuhlmann 2012). As shown by Schabes and Shieber (1994), however, there are several possible ways of defining TAG derivation trees, depending on how multiple adjunction of several auxiliary trees at the same tree node is handled. The standard notion of derivation proposed by Vijay-Shanker (1987) forbids multiple adjunction, thus enforcing dependent derivations. In contrast, the extended notion of derivation proposed by Schabes
∗ UMR 7503, Campus Scientifique, BP 239, F-54506 Vandoeuvre-lès-Nancy Cedex, France. E-mail:{claire.gardent,shashi.narayan}@loria.fr.
Submission received: 2 August 2013; revised version received: 16 June 2014; accepted for publication: 21 September 2014.
doi:10.1162/COLI a 00217
© 2015 Association for Computational Linguistics
and Shieber (1992, 1994) allows multiple adjunction at a single node, thereby yielding so-called independent derivations (i.e., derivations where the relation between the adjoining trees is left unspecified). The difference between the two types of derivations is illustrated in Figure 1. While in the standard (dependent) derivation, one adjective tree is adjoined to the other adjective tree, which itself is adjoined to the noun tree for pepper; in the extended (independent) derivation, both adjective trees adjoin to the noun tree.
Schabes and Shieber (1994) argue that allowing both for dependent and independent derivations better reflects linguistic dependencies. Making use of the distinction introduced in TAG between predicative and modifier auxiliary trees (Schabes and Shieber (1994), Section 3.1), they define a parsing algorithm that assigns dependent derivations to predicative auxiliary trees but independent derivations to multiple modifier auxiliary trees adjoining to the same node. In case both predicative and modifier auxiliary trees adjoin to the same node, their parsing algorithm ensures that predicative trees appear above the modifier trees in the derived tree.
This parsing algorithm is defined for featureless variants of TAG. In contrast, in implemented TAGs (e.g., XTAG [The XTAG Research Group 2001], SemXTAG [Gardent 2008], or XXTAG1 [Alahverdzhieva 2008]) feature structures and feature unification are central. They are used to minimize the size of the grammar; to model linguistic phenomena such as verb/subject agreement; and to encode a unification-based syntax/semantics interface (e.g., Gardent and Kallmeyer 2003).
In this article, we extend Schabes and Shieber’s proposal to Feature-Based TAG (FB-TAG); and we show that the resulting parsing algorithm naturally accounts for the interplay of dependent vs. independent derivation structures with syntactic constraints, linear ordering, and scopal vs. nonscopal semantic dependencies.
The article is organized as follows. In Section 2, we recap the motivations for independent derivations put forward by Schabes and Shieber (1994) and we briefly discuss the interactions that may arise between dependent and independent derivations. Section 3 summarizes their approach. In Section 4, we present the intuitions and motivations underlying our proposal and we highlight the differences with Schabes and Shieber’s approach. Section 5 presents our proposal. Section 6 concludes. 2 why are independent derivations desirable? :We start by summarizing Schabes and Shieber’s motivations for independent derivations. We then discuss the interactions between dependent and independent derivations.
1 XXTAG stands for XMG (Crabbé et al. 2013)-based XTAG. Schabes and Shieber (1994) give three main motivations for independent derivations. The first motivation concerns the interaction of verbs with multiple modifiers. Consider sentences2 in Examples (1) and (2).
(1) a. Richard Parker and Pi wandered the Algae Island yesterday through the meerkats.
b. Richard Parker and Pi wandered the Algae Island yesterday.
c. Richard Parker and Pi wandered the Algae Island through the meerkats.
(2) a. ⊗
The Orangutan reminded Pi of his mother yesterday through the meerkats.
b. The Orangutan reminded Pi of his mother yesterday.
c. ⊗ The Orangutan reminded Pi of his mother through the meerkats.
Movement verbs such as to wander allow for directional modifiers such as through the meerkats, whereas verbs such as to remind do not. In TAG, such restrictions can be modeled using selective adjoining constraints to specify which modifier tree may or may not be adjoined at a particular node in a given tree. Therefore, it is possible to license (1) and to rule out (2c). In Example (2a), however, under the dependent notion of adjunction, the tree for the directional adverbial through the meerkats will adjoin to the modifier tree for yesterday, which itself will adjoin to the tree selected by reminded. Thus, constraints placed by the verb on its modifiers must be passed through by modifier trees (here, the tree for yesterday) to also rule out sentences such as Example (2a). Propagating selective adjunction constraints in TAG would lead to a formalism for which derivation trees are no longer context-free (Schabes and Shieber 1994).
The second motivation for independent adjunction stems from probabilistic approaches. Stochastic lexicalized TAG specifies the probability of an adjunction of a given auxiliary tree at a given node in another elementary tree (Resnik 1992; Schabes 1992). Thus, under the standard notion of derivation, the overall probability of the string roasted red pepper would be determined by the probability of red adjoining to pepper and the probability of roasted adjoining to red. In contrast, independent adjunction would result in a derivation such that the overall probability of the string roasted red pepper would be determined by the probability of both red and roasted adjoining to pepper. Schabes and Shieber (1994, page 97) argue that it is plausible that “the most important relationships to characterize statistically are those between modifier and modified, rather than between two modifiers.”
A third motivation comes from semantics and, more particularly, from scope ambiguities involving modifiers. Given a sentence such as Example (3), where the relative scope of the modifiers twice and intentionally is ambiguous,3 Shieber (1994) shows that, under the extended definition of adjunction, a synchronous TAG modeling the relation between syntactic trees and logical formulae can account for both readings.
(3) John blinked twice intentionally.
2 The characters in these sentences are borrowed from Yann Martel’s book Life of Pi. 3 The sentence can describe either a single intentional act of blinking twice or two intentional acts each of single
blinking (Shieber 1994).
The account crucially relies on multiple independent adjunction of the two modifier trees to the tree for blink: Depending on which order the auxiliary trees for twice and intentionally adjoins to blink, the logical formula built will be either intentionally(twice(blink)) or twice(intentionally(blink)), thus capturing the ambiguity. To capture the different types of semantic dependencies and morpho-syntactic constraints that may hold between multiple auxiliary trees adjoining to the same entity, both dependent and independent derivations are needed.
As argued earlier, because there are no constraints or semantic relation holding between each of them, multiple intersective modifiers applying to the same entity (Example (4)) are best modeled using an independent derivation.
(4) The tall black meerkat slept. (Independent derivation)
In contrast, because they may involve strong scopal and morpho-syntactic constraints, stacked predicative verbs (i.e., verbs taking a sentential complement, Example (5a)) and non-intersective modifiers (Example (5c)) require dependent derivations. Consider sentences (5a–b), for instance. If predicative trees were assigned an independent derivation, sentence (5a) would be judged ungrammatical (because want requires an infinitival complement but would adjoin to the finite verb slept) and conversely, sentence (5b) would incorrectly be judged grammatical (because both want and try require an infinitival complement). Similarly, in Example (5c), the church is Syrian Orthodox, not Syrian and Orthodox. Assigning a dependent rather than an independent derivation to such cases straightforwardly captures the distinction between intersective and non-intersective modifiers.
(5) a. XJohn wanted to assume that Peter slept. (Dependent derivation)
b. ⊗ John wanted Peter tries to walk.
c. The meerkat admired the Syrian Orthodox church. (Dependent derivation)
Finally, some multiple adjunctions may involve both dependent and independent derivations, for example, when multiple modifiers and predicative verbs adjoin to the same verb (Example (6a)) or in the case of a derivation (Example (6b)) involving both intersective (old) and non-intersective (i.e., Syrian in Syrian Orthodox) modifiers.
(6) a. Yann said that John knows that Richard Parker and Pi wandered the Algae Island yesterday through the meerkats. (Mixed derivation)
b. The meerkat admired the old Syrian Orthodox church. (Mixed derivation)
As we shall see in Section 5.3, the parsing algorithm we propose licenses dependent, independent, and mixed derivations but is restricted to appropriately distinguish between various types of modifiers. Moreover, the feature information encoded in the grammar further restricts the derivation structures produced, thereby accounting for the interactions between adjunction, linear ordering, and morpho-syntactic constraints. 3 multiple adjunction in tree adjoining grammars :Vijay-Shanker and Weir (1991) introduce a compilation of TAG to Linear Indexed Grammars (LIGs; Gazdar 1988) that makes the derivation process explicit. Schabes and
Shieber (1994) modify this compilation to allow for both dependent and independent derivations. The resulting LIG is further exploited to specify a parsing algorithm that recovers those derivations.
In this section, we summarize Schabes and Shieber’s proposal. We start (Section 3.1) with an informal description of their approach. In Section 3.2, we introduce ordered derivation trees. Section 3.3 provides a brief introduction to LIG. Section 3.4 summarizes the TAG-to-LIG compilation proposed by Vijay-Shanker and Weir (1991). Finally, Section 3.5 describes the modifications introduced by Schabes and Shieber (1994) to allow both for dependent and for independent derivations. Tree Adjoining Grammar distinguishes between two types of auxiliary trees, namely, modifier vs. predicative auxiliary trees (Joshi and Vijay-Shanker 2001). Whereas predicative trees are assigned to verbs taking a sentential argument, modifier trees are assigned to all other auxiliary trees (e.g., verbal auxiliaries, adjectives, adverbs, prepositions, and determiners). More generally, the difference between a predicative and a modifier tree is that in a predicative tree, the foot node, like the substitution nodes, corresponds to an argument node selected by its lexical anchor (i.e., the word that selects that tree) whereas in a modifier auxiliary tree, the foot node is an open slot corresponding to the phrase being modified. When associating semantic entities with tree nodes (as proposed, for example, by Joshi and Vijay-Shanker [2001] and Gardent and Kallmeyer [2003]), this difference can be seen by noting the entities associated with root and foot nodes: These are distinct in a predicative tree but identical in modifier trees.
In their approach, Schabes and Shieber specify a TAG-to-LIG conversion that systematically associates dependent derivations with predicative auxiliary trees and independent derivations with modifier auxiliary trees. In addition, they introduce two mechanisms to ensure that each derivation tree unambiguously specifies a linguistically plausible derived tree.
First, they enforce ordering constraints between modifier trees adjoining at the same node (which are thus ambiguous with respect to the derived tree they describe) by assuming that derivation trees are ordered and that linear precedence (LP) statements can be used to constrain the order of siblings in a derivation tree. For instance, given the independent derivation shown in Figure 1, an LP statement stating that βred must occur before βroasted in the derivation tree will ensure that βroasted appears above βred in the derived tree and therefore that the resulting derived tree yields the phrase roasted red pepper rather than red roasted pepper.
Second, when both predicative and modifier trees adjoin at the same address, predicative trees always occur above all modifier trees in the derived tree (“outermost predication”). This ensures, for instance, that under the reading where yesterday refers to the arriving rather than the saying (i.e., when both say and yesterday adjoin to arrive), Example (7a) is derived but not Example (7b).
(7) a. XPeter says that yesterday John arrived late.
b. ⊗ Yesterday Peter says that John arrived late. In the standard version of TAG, each derivation tree describes a unique derived tree. In the case of a dependent derivation, unicity follows from the fact that dependent
derivations specify the order in which adjunction takes place (e.g., β2 adjoins to β1 and the result to α). As a result, if β2 adjoins to β1, there is only one possible derived tree—namely, a tree where β2 appears above β1.
When allowing for independent derivations, however, several derived trees are possible, depending on the order in which the auxiliary trees are adjoined. To ensure a unique mapping from derivation to derived tree, Schabes and Shieber (1994) therefore introduce the notion of ordered derivation trees. Ordered derivation trees differ from standard TAG derivation trees in that (i) they may contain sibling edges labeled with the same address, and (ii) they specify a total order on such siblings.
Figure 2 shows an example ordered derivation tree and associated derived tree. As indicated by the shared g address on their parent edge, auxiliary trees β1, . . . , βn adjoin to the same node—namely, the node with address g in the elementary tree τ. Because the derivation tree is ordered, β1 will appear below β2 in the derived tree, which in turn will be below β3, and so on. In short, given a set of auxiliary trees all adjoining to the same tree node, the derived tree produced from an ordered derivation tree following an independent derivation will be identical to the derived tree produced with the corresponding dependent derivation—that is, the dependent derivation where β1, . . . , βn appear in increasing index order from top to bottom. Like context-free grammars, LIGs (Gazdar 1988) are string rewriting systems where strings are composed of terminals and nonterminals. In a LIG, however, nonterminal symbols may be associated with a stack of symbols, called indices. A LIG rule can thus be represented as follows:
N[..µ] → N1[µ1] . . .Ni−1[µi−1]Ni[..µi]Ni+1[µi+1] . . .Nn[µn] (8)
N and Ni are nonterminals whereas µ and µi are strings of stack symbols. The symbol .. stands for the remainder of the stack symbols. Note that the remainder of the stack symbols associated with the left-hand side is associated with only one of the nonterminal (namely, Ni) on the right-hand side.
LIGs have been used in the literature (Weir and Joshi 1988; Vijay-Shanker and Weir 1991) to provide a common framework for the extensions of context-free grammars. In particular, Vijay-Shanker and Weir (1991, 1993) showed a weak equivalence between LIGs, TAGs, and combinatory categorial grammars (Steedman 2000) and proposed a LIG-based polynomial-time CYK recognition algorithm for TAGs and combinatory categorical grammars. In what follows, we show how Schabes and Shieber (1994) use a LIG variant of TAGs to license both dependent and independent derivations. The TAG-to-LIG compilation proposed by Vijay-Shanker and Weir (1991) produces LIG rules that simulate a traversal of the derived tree produced by the original TAG grammar. In these LIG rules, each node η of a TAG elementary tree is viewed as having both a top t[..η] and a bottom b[..η] component to account for the possibility of an adjunction. Figure 3 illustrates the traversal of the TAG-derived trees specified by the LIG resulting fromVijay-Shanker and Weir (1991) TAG-to-LIG compilation.
Figure 4 lists the LIG rules resulting from the TAG to LIG compilation process. Each nonterminal (t[..η] or b[..η]) with the top of the stack symbol in a LIG rule corresponds to a unique node in some elementary tree of the grammar. The inner stack symbols are used to keep track of the nodes higher in the derived tree where an auxiliary tree has been adjoined.
Rules of Types 1 and 2 capture immediate dominance between the bottom of a node η and the top of its immediate daughters in two configurations, depending on whether η dominates the foot node (Type 1) or not (Type 2). Rules of Type 3 handle nodes that require neither substitution nor adjunction. This rule handles cases where no adjunction occurs at a node by rewriting the top of this node to its bottom. Rules of Type 6 model substitution. Finally, rules of Types 4 and 5 handle adjunction. They specify that, for any given node η and any auxiliary tree β that may adjoin to η, the top of η rewrites to the top of the root node of β; and the bottom of the foot of β to the bottom of η. It follows that there can be no multiple adjunction in this LIG version of TAG. To associate predicative tree adjunctions with dependent derivations and multiple modifier adjunctions with independent derivations, Schabes and Shieber (1994) modify the compilation of TAG to LIG, proposed by Vijay-Shanker and Weir (1991) as sketched in Figure 5. Type 4(a) rules apply to adjunctions involving predicative trees. They are identical to Type 4 rules in the Vijay-Shanker and Weir’s approach and therefore enforce a standard (dependent) derivation for predicative trees. In contrast, Type 4(b) rules apply to adjunctions involving modifiers and result in an independent derivation.
Note also that the Outermost Predication constraint (i.e., predicative trees always occur above modifier trees adjoined at the same node) alluded to in Section 3.1 follows from the interactions between the Type 4(a) and Type 4(b) LIG rules.
Schabes and Shieber prove the weak-generative equivalence of TAGs under both Standard and Extended derivation using the LIG compilation. They also propose a recognition and a parsing algorithm with complexity of O(n6) in the length of the string. 4 multiple adjunction in feature-based tag :In this section, we explain why a straightforward extension of Schabes and Shieber’s proposal to FB-TAG would not work and we outline the intuitions and motivations underlying our approach. Section 5 will then introduce the details of our proposal. We start by a brief description of FB-TAG and of the unifications performed during derivation. FB-TAG was introduced by Vijay-Shanker (1987) and Vijay-Shanker and Joshi (1988, 1991) to support the use of feature structures in TAG. Figure 6 shows a toy FB-TAG for illustration.
An FB-TAG differs from a TAG in that tree nodes are decorated with feature structures. Nonterminal and foot nodes are decorated with two feature structures called top (T) and bottom (B), and substitution nodes are decorated with a single top feature structure. During derivation, feature structure unification constrains tree combination, as illustrated in Figure 7. Substitution unifies the top feature structure of a substitution node with the top feature structure of the root node of the tree being substituted. The adjunction of an auxiliary tree β to a tree node ηo unifies the top and bottom feature structures of ηo with the top feature structure of the root node of β and the bottom feature structure of its foot node, respectively. Finally, at the end of the derivation, the top and bottom feature structures of all nodes in the derived tree are unified.
Figure 8 shows the standard derivation for the phrase all the meerkats, using the grammar shown in Figure 6, and Figure 9 shows the corresponding derived and derivation trees. As can be seen, the feature constraints encoded in the grammar correctly ensure that all the meerkats can be derived (leftmost derivation tree in Figure 9) but not the all meerkats (rightmost in Figure 9). The incorrect derivation is blocked by the feature structure [det : nil] on the foot of the auxiliary tree βthe, which leads to a unification failure if βthe is adjoined at the root of βall with bottom feature structure [det : the]. To motivate our approach, we start by considering a simple extension of Schabes and Shieber’s LIG framework to FB-TAG, where each LIG rule enforces unifications mimicking those applied in FB-TAG. In particular, let us assume that Type 3 rules
(“No adjunction”) unify the top and the bottom feature structures of nodes where no adjunction occurs, and Type 4(b) rules (“Start root of adjunction”) unify the top feature (η.T) of the node (η) being adjoined to with the top feature structure (ηr.T) of the root node (ηr) of the auxiliary tree being adjoined: 4
b[..η] → t[..ηηr] η.T ∪ ηr.T (Type 4b) (9)
t[..η] → b[..η] η.T ∪ η.B (Type 3) (10)
As shown in Figure 10, this approach can incorrectly lead to derivation failures in the case of an independent multiple adjunction. Intuitively, the reason for this is that, in Schabes and Shieber’s approach, multiple adjunction starts and ends from the bottom component of the node being adjoined to. This is fine when no features are involved because the category of the node being adjoined to is always identical to the root and foot node of the auxiliary trees being adjoined. When nodes carry feature structures, however, a unification clash can occur that makes derivation fail. Thus, in our example, derivation incorrectly fails because the bottom feature structures of the root node of the auxiliary tree for all and the bottom feature structure of the root node of the auxiliary tree for the should unify but have conflicting value. As shown by the dependent derivation for all the meerkats depicted in Figure 8, this is incorrect. As we just saw, in the case of multiple independent adjunctions, a straightforward extension of Schabes and Shieber’s LIG framework to FB-TAG fails to correctly capture the unification constraints encoded in the grammar. More generally, when extending multiple independent adjunction to FB-TAG, it is crucial that the feature constraints encoded by the linguist describe the same set of derived trees no matter
4 We associate η.T ∪ ηr.T with the Type 4(b) rules to mimic the adjunction in FB-TAG, as shown in Figure 7.
which derivation tree is produced. We therefore propose a parsing algorithm that, given several auxiliary trees β1, . . . ,βn adjoining at the same node ηo, performs the same unifications independently of whether the derivation is dependent, independent, or mixed dependent/independent.
Figure 11 shows the unifications resulting from the multiple adjunction of β1 and β2 to a single node ηo. While it depicts the unifications enforced by our parsing algorithm for the derivation tree shown on the right hand side (i.e., for the independent adjunction of β1 and β2 to ηo), these unifications are in fact exactly the same as those that would be enforced by a dependent adjunction of β2 into β1 into ηo.
One key point illustrated by Figure 11 is that whereas multiple adjunction operates on a single node (here ηo), the unification constraints of FB-TAG require that the bottom feature structure of the foot of an auxiliary tree which appears higher in the derived tree (here, β2) unifies with the bottom feature structure of the root of the auxiliary tree appearing immediately below it in the derived tree (here β1)—not with that of the root of the node to which it adjoins (here ηo). In other words, while a multiple adjunction on ηo operates on ηo only, a correct implementation of FB-TAG unification constraints requires keeping track of the feature structures associated with the auxiliary trees successively adjoining to ηo.
In our proposal, we capture this bookkeeping requirement by associating tree nodes not with feature structures but with reference variables pointing to feature structures. The parsing algorithm is then specified so as to support dependent, independent, and mixed derivations while enforcing the same unifications as would be performed under a dependent adjunction. Before giving the technical details of our parsing algorithm (Section 5), we first highlight some differences between our and Schabes and Shieber’s approach. In particular, we show that (i) whereas Schabes and Shieber resort to three distinct mechanisms to account for word order constraints (i.e., selective adjoining constraints, linear
precedence statements on derivation trees, and a constraint on parsing), the FBTAG approach supports a uniform treatment of word order and (ii) our approach straightforwardly accounts for mixed dependent/independent derivations that would require some additional stipulation in Schabes and Shieber’s approach.
4.4.1 Ordering Constraints among Modifier Auxiliary Trees. In TAG, determiners and verbal auxiliaries are modifier rather than predicative auxiliary trees (cf. Section 3.1). Because Schabes and Shieber’s definitions systematically associate modifiers with independent derivations, all examples in (11a–d) undergo an independent derivation and constraints must therefore be provided to determine the order of the sibling nodes in the resulting derivation tree.
(11) a. XThe sonatas should have been being played by Sarah.
b. ⊗ The sonatas have should been being played by Sarah.
c. XAll the meerkats
d. ⊗ The all meerkats
To specify these constraints on similar cases (soft ordering constraints on adjectives and strict ordering constraints on temporal and spatial adverbial phrases in German), Schabes and Shieber (1994) suggest the use of LP constraints on derivation tree siblings. As illustrated by the derivation of Example (11c–d) in Figures 8 and 9, in the FBTAG approach, such additional constraints are unnecessary: They simply fall out of the feature constraints encoded in the grammar.
Note that even if determiners and auxiliary verbs were to be handled using dependent adjunction, the word ordering constraints used by Schabes and Shieber would fail to account for cases such as Example (12), where auxiliary verbs are interleaved with adverbs.
(12) John has often been selected for nomination.
In this case, if the auxiliary verbs has and been were treated as predicative trees, Schabes and Shieber’s constraint that predicative trees adjoin above modifier trees would preclude the derivation of Example (12) and incorrectly predict the derived sentence to be John has been often selected for nomination.
4.4.2 Ordering Constraints among Predicative Trees. As discussed by Schabes and Shieber (1994), auxiliary predicative trees may impose different constraints on the type of sentential complement they accept. Thus Example (13a) is correct but not Example (13b) because want expects an infinitival complement (previously shown in Example (5)).
(13) a. XJohn wanted to assume that Peter slept.
b. ⊗ John wanted Peter tries to walk.
Although in the Schabes and Shieber’s approach, selective adjoining constraints are used to license Example (13a) and rule out Example (13b), in the FB-TAG approach, this can be achieved using feature constraints.
4.4.3 Ordering Constraints between Predicative and Modifier Auxiliary Trees. In sentences such as Example (14a) where both modifier and predicative auxiliary trees adjoin to the same address, the predicative trees should generally adjoin above any modifier trees so that the predicative verb precedes the modifier in the derived string.
(14) a. XJohn promised that Peter will leave tomorrow. b. ⊗ Tomorrow John promised that Peter will leave.
To ensure the appropriate linearization, the Schabes and Shieber’s approach introduces the outermost-predication rule, which stipulates that predicative trees adjoin above modifier auxiliary trees. In contrast, the FB-TAG approach allows both orders and lets feature constraints rule out ungrammatical sentences such as Example (14b). This allows the approach to directly extend to a counter-example discussed by Schabes and Shieber (1994), where a modifier (here, At what time) must in fact adjoin above a predicative tree.
(15) At what time did Brockway say Harrison arrived?
Figure 12 shows two possible derivation trees for the sentence (15) under the interpretation where it is the time of arriving (rather than the time of saying) which is questioned. These derivation trees show the two possible relative orderings of the (predicative) auxiliary tree for say and the (modifier) auxiliary tree at what time. Because the Outermost-Predication rule requires that predicative trees adjoin above modifier trees (and thus occur outermost in the derivation tree), in Schabes and Shieber’s approach, only the right-hand side derivation is possible, thus failing to derive sentence (15). In contrast, because our approach does not explicitly constrain the relative ordering of predicative and modifier auxiliary trees adjoining to the same node, both derivations are possible, thereby licensing both Example (15) and the sentence Did Brockway say at what time Harrison arrived?
4.4.4 Mixed Dependent and Independent Multiple Adjunctions. In Schabes and Shieber’s approach, all modifier auxiliary trees undergo independent derivation. As shown in Section 2.2, however, non-intersective modifiers arguably license a dependent derivation while some cases of multiple adjunction may involve both a dependent and an independent derivation. As we shall see in Section 5, our FB-TAG approach accounts for such cases by allowing both for independent and dependent derivations, by ruling out dependent derivations for intersective modifiers and by using feature constraints to regulate the interactions between multiply adjoining auxiliary trees. 5 extending schabes and shieber’s lig framework for fb-tags :We now propose a compilation of FB-TAG-to-LIG that makes both dependent and independent derivations in FB-TAG explicit. We use this resulting LIG to specify an
Earley algorithm for recovering multiple adjunctions in FB-TAG. This compilation differs in two main ways from that proposed by Schabes and Shieber (1994). First, tree nodes are associated with reference variables pointing to feature structures. Second, the LIG rules are modified and extended with unification operations. To account for FB-TAG unifications while allowing for independent derivations, we replace the feature structures of FB-TAG with reference variables pointing to those feature structures. Each node in the elementary trees is decorated with two reference variables: The top reference variable PT contains the reference to the top feature structure T and the bottom reference variable PB contains the reference to the bottom feature structure B. The top and the bottom feature structures of a node η can be traced by val(η.PT) and val(η.PB), respectively, where PT and PB are the top and the bottom reference variables decorating the node η, and the function val(P) returns the feature structures referred to by the reference variable P.
When specifying the parsing algorithm, we use reference variables to ensure the appropriate unifications, as follows. In an independent derivation where the node ηo is adjoined to, first by β1 and second by β2, the bottom feature structure ηo.B of ηo (i) unifies with the bottom feature structure ηf 1.B of the foot of β1 and (ii) is reassigned (:=) to the bottom reference variable ηr1.PB of the root of β1. When β2 is adjoined, its foot node will therefore correctly be unified, not with the bottom feature structure of ηo but with that of ηr1. To support both dependent and independent derivations while enforcing the correct unifications, we modify the TAG-to-LIG compilation in such a way that the resulting LIG rules capture the tree traversal depicted in Figure 13. Independent derivations are accounted for by the fact that adjunction starts and ends at the bottom component of the node being adjoined to (Type 4 and 5 rules). Our LIG compilation automatically supports dependent derivations by allowing sequential adjunctions at the roots of auxiliary trees.
Type 4: Start root of adjunction. For each elementary tree node η that allows the adjunction of the auxiliary tree with the root node ηr, the following LIG production rule is generated.
b[..η] → t[..ηηr] val(η.PT) ∪ val(ηr.PT),η.PB := ηr.PB
Type 5: Start foot of adjunction. For each elementary tree node η that allows the adjunction of the auxiliary tree with the foot node ηf , the following LIG production rule is generated.
b[..ηηf ] → b[..η] val(η.PB) ∪ val(ηf .PB)
Type 6: Start substitution. For each elementary tree node η that allows the substitution of the initial tree with the root node ηr, the following LIG production rule is generated (not shown in Figure 13).
t[η] → t[ηr] val(η.PT) ∪ val(ηr.PT)
To perform multiple adjunction while enforcing the appropriate feature unifications (as depicted in Figure 11), we split Type 3 rules into two subtypes. Type 3(a) rules apply to the root of auxiliary trees and perform no unification. By no unification, they ensure that feature structures are not blocked for the possibility of the adjunction of the following auxiliary tree and allow for the correct unifications to be carried out for independent derivations. Type 3(b) rules function as termination of multiple adjunction by unifying the top and bottom feature structures of the node. It is applicable to all tree nodes except roots of auxiliary trees.
Type 3(a): Terminating adjunction at the root of the auxiliary tree. For each root node η of the auxiliary trees, the following LIG production rule is generated.
t[..η] → b[..η] ∅
Type 3(b): Terminating adjunction at any other node. For each node η that is not a root node of some auxiliary tree and is not marked for substitution, the following LIG production rule is generated.
t[..η] → b[..η] val(η.PT) ∪ val(η.PB)
Given this set of rules, both dependent and independent derivations are possible. For example, given two auxiliary trees β1 and β2 adjoining at the node η in an elementary tree τ, a dependent derivation will occur whenever the Type 4 rule applies to predict the adjunction of, for example, β2 at the root of β1. Conversely, if the Type 3(a) rule applies at the root of β1, recognition will move from the top of the root of β1 to its bottom, allowing for Type 5 rule to complete the adjunction of β1 at the node η; and the Type 4 rule applies to predict the adjunction of β2 at the node η of τ, registering an independent derivation. In this section, we present our parsing algorithm for FB-TAGs with dependent and independent derivations. We start with an informal description of how the algorithm handles the interactions between unification and independent derivations. We then go on to specify the inference rules making up the algorithm. We do this in two steps. First, we present a basic set of rules allowing for both dependent and independent derivations. Second, we show how to constrain this algorithm to minimize spurious ambiguity.
5.3.1 Independent Derivations and Feature-Structure Unification. Before specifying the parsing algorithm, we illustrate by means of an example the interplay between multiple independent adjunction and feature structure unifications.
Figure 14 displays the feature unifications and reassignment performed during the recognition process of a multiple independent adjunction. The linear ordering of the equations reflects the order of the parsing completion operations.
Given the auxiliary tree β1 and the adjunction site ηo, the picture shows that unifying the bottom feature structure of the foot node of β1 with the bottom feature structure of ηo (Step 1: Type 5,ηo.B ∪ ηf 1.B) occurs before the bottom reference variable of ηo is reassigned to the bottom feature structure of the root of β1 (Step 4: Type 4,ηo.PB → ηr1.B). Also, the reassignment ensures that the follow-up adjunction of β2 at the node ηo has access to the bottom feature of the root of the previous auxiliary tree β1 (Step 5: Type 5,ηr1.B ∪ ηf 2.B). At the end of the adjunction (Step 9), the Type 3(b) rule ensures that the top and the bottom features of the root of the last auxiliary tree (here, β2) adjoined are unified (ηr2.T ∪ ηr2.B).
t b ηo ηo.PT → ηo.T ηo.PB → ηo.B
t b
ηr1 ηr1.PT → ηr1.T
ηr1.PB → ηr1.B
t b
ηf 1 ηf 1.PT → ηf 1.T
ηf 1.PB → ηf 1.B
β1
t b
ηr2 ηr2.PT → ηr2.T
ηr2.PB → ηr2.B
t b
ηf 2 ηf 2.PT → ηf 2.T
ηf 2.PB → ηf 2.B
β2
(1) (2)
(5)
(6)
(4)
(3)
(8)
(7)
(9)
(9). Type 3(b),ηo.PT → F7,ηo.PB → F7, F7 = ηo.T ∪ ηr1.T ∪ ηr2.T ∪ ηr2.B
(8). Type 4,ηo.PT → F6,ηr2.PT → F6, F6 = ηo.T ∪ ηr1.T ∪ ηr2.T,ηo.PB → ηr2.B
(7). Type 3(a),ηr2.PT → ηr2.T,ηr2.PB → ηr2.B
(6). Type 3(b),ηf 2.PT → F5,ηf 2.PB → F5, F5 = ηr1.B ∪ ηf 2.T ∪ ηf 2.B
(5). Type 5,ηo.PB → F4,ηf 2.PB → F4, F4 = ηr1.B ∪ ηf 2.B (4). Type 4,ηo.PT → F3,ηr1.PT → F3, F3 = ηo.T ∪ ηr1.T,ηo.PB → ηr1.B
(3). Type 3(a),ηr1.PT → ηr1.T,ηr1.PB → ηr1.B
(2). Type 3(b),ηf 1.PT → F2,ηf 1.PB → F2 , F2 = ηo.B ∪ ηf 1.T ∪ ηf 1.B
(1). Type 5,ηo.PB → F1,ηf 1.PB → F1, F1 = ηo.B ∪ ηf1.B
Figure 14 Multiple independent adjunction of β1 and β2 to ηo. The unifications and reassignments are listed in the order in which they are performed during the recognition process.
As we shall see subsequently, this correct ordering between unification and reassignment follows from the proposed Earley algorithm. Type 4 completor rules complete the prediction triggered at the root of an auxiliary tree (“Start root of adjunction”) and Type 5 completor rules complete the prediction triggered at the foot node of an auxiliary tree (“Start foot of adjunction”). Because completion operates bottom–up, it follows that Type 5 rules apply before Type 4 rules. Thus, when adjoining an auxiliary tree β1 to a node ηo, the Type 5 completor rules, unifying the bottom feature structure of the foot node of β1 with the bottom feature structure of the node ηo, occurs before the Type 4 completor rules, which reassign the bottom reference variable of ηo to the bottom feature structure of the root of β1.
5.3.2 Inference Rules. The parsing algorithm for FB-TAG is a modification of the algorithm presented by Schabes and Shieber (1994). It is a chart-based parsing method based on the Earley type deduction system. Each item in the chart is of the format 〈N[..η] → Γ •∆, i, j, k, l〉, where N is some LIG nonterminal (i.e., t or b), and Γ and ∆ present the sequences of LIG nonterminals associated with stacks of node indices. The indices i, j, k, and l are markers in the input string, showing the recognized portion:5 The recognized item starts in position i, ends in position l, and if η dominates a foot node, the tree dominated by the foot node starts in j and ends in k. If the foot node is not dominated by the recognized nonterminal sequence Γ, the values for j and k are taken to be the dummy value ’−’. As in Earley algorithms, the • separates the nonterminal sequence Γ which was parsed from the nonterminal sequence ∆ yet to be parsed.
The first three types of rules (Scanner, Predictor, and Type 1/2 Completor) are identical to those introduced by Schabes and Shieber (1994) and do not involve any unification operations.
The Type 3(b) completor rule enforces top and bottom unification on all nodes that are not the root of an auxiliary tree, and the Type 3(a) completor rule prevents top and bottom unification at the root of auxiliary trees.
The Type 4 completor rule unifies the top feature structure of the root of the auxiliary tree with the top feature structure of the adjunction site. In addition, it ensures that on completion of an adjunction at node η, the bottom feature structure of η is reassigned to the bottom feature structure labeling the root of the auxiliary tree. In this way, the unifications occurring in an independent derivation will mirror those occurring in a dependent one in that any following adjunction will induce unifications as if it were happening at the root node ηr of the preceding auxiliary tree (not at η).
On completion of a foot node prediction (the tree dominated by the foot of the auxiliary tree has been recognized), the Type 5 completor rule unifies the bottom feature structure of the foot of the auxiliary tree with the bottom feature structure of the adjunction site.
Finally, the Type 6 completor unifies the top feature structure of a substitution node with the top feature structure of the root of the tree being substituted.
r Scanner:
〈b[..η] → Γ • w∆, i, j, k, l〉
〈b[..η] → Γw •∆, i, j, k, l + 1〉 , w = wl+1, ∅
5 The indices 〈i, j, k, l〉 have been used in previous parsing algorithms for tree-adjoining grammars (Vijay-Shankar and Joshi 1985; Schabes and Joshi 1988; Schabes 1991). They deliver the same functionality here.
If w (a terminal symbol) occurs at position l+1, the scanner rule creates a new item whose span extends to l+1.
r Predictor:
〈N[..η] → Γ • N′[µ]∆, i, j, k, l〉
〈N′[µ] → •Θ, l,−,−, l〉 , ∅
Predictor rules are produced for all types of production rules. N and N′ are LIG variables taking the value t or b. Γ, ∆, and Θ are the sequences of LIG nonterminals associated with stacks of node indices. µ is a sequence of node indices.
r Type 1 and 2 Completor:
〈b[..η] → Γ • t[η1]∆, m, j ′, k′, i〉 〈t[η1] → Θ•, i, j, k, l〉
〈b[..η] → Γt[η1] •∆, m, j ⊕ j′, k ⊕ k′, l〉 ,
∅,
η1 not a root node
Types 1 and 2 Completor rules permit completing Rules 1 and 2 whenever the top of a child node is fully recognized. Here, t[η1] has been fully recognized as the substring between i and l (i.e., wi+1 . . .wl). Therefore, t[η1] can be completed in b[..η]. If one of t[η1] or b[..η] dominates the foot node of the tree, the final b[..η] will have indices associated with the substring recognized by the foot subtree. The operation ⊕ is defined as follows:
x ⊕ y =

  
  
x, if y = −
y, if x = −
x, if x = y
undefined, otherwise
r Type 3(a) Completor:
〈t[..η] → •b[..η], i,−,−, i〉 〈b[..η] → Θ•, i, j, k, l〉
〈t[..η] → b[..η]•, i, j, k, l〉 ,
∅,
η an auxiliary tree root node
The Type 3(a) Completor rule is used to complete the prediction of an auxiliary tree rooted in η. Once the auxiliary tree dominated by b[..η] has been recognized, the auxiliary tree itself is completely recognized. As explained earlier, there is in this case no feature unification between the top and the bottom of the root of the auxiliary tree.
r Type 3(b) Completor:
〈t[..η] → •b[..η], i,−,−, i〉 〈b[..η] → Θ•, i, j, k, l〉
〈t[..η] → b[..η]•, i, j, k, l〉 ,
val(η.PT ) ∪ val(η.PB),
η not an auxiliary tree root node
The Type 3(b) Completor rule ensures the unification of the top and bottom feature structures for all nodes that are not the root node of an auxiliary tree.
r Type 4 Completor:
〈b[..η] → •t[..ηηr], i,−,−, i〉
〈t[..ηηr] → Θ•, i, j, k, l〉
〈b[..η] → ∆•, j, p, q, k〉
〈b[..η] → t[..ηηr]•, i, p, q, l〉 ,
val(η.PT) ∪ val(ηr.PT)
η.PB := ηr.PB
The auxiliary tree associated with the predicted adjunction (t[..ηηr]) at the node η and the subtree dominated by the node η (below b[..η]) are completed, hence b[..η] can be completely recognized with this adjunction. The associated feature unification unifies the content of the top reference variable of the adjoining node site η with the content of the top reference variable of the root node ηr of the adjoined auxiliary tree. After the successful adjunction of this adjoining tree, the bottom reference variable of the adjoining node site η is reassigned to the content of the bottom reference variable of the root node ηr of the adjoined auxiliary tree.
r Type 5 Completor:
〈b[..ηηf ] → •b[..η], i,−,−, i〉 〈b[..η] → Θ•, i, j, k, l〉
〈b[..ηηf ] → b[..η]•, i, i, l, l〉 , val(η.PB) ∪ val(ηf .PB)
The foot node prediction can be completed when the adjunction has been performed and the bottom part of the adjoining node site η has been recognized. The associated feature unification unifies the content of the bottom reference variable of the adjoining node site η with the content of the bottom reference variable of the foot node ηf of the auxiliary tree being adjoined.
r Type 6 Completor:
〈t[η] → •t[ηr], i,−,−, i〉 〈t[ηr] → Θ•, i,−,−, l〉 〈t[η] → t[ηr]•, i,−,−, l〉 , val(η.PT) ∪ val(ηr.PT)
The Type 6 Completor rule completes the substitution at the node η. The associated feature unification unifies the content of the top reference variable of the node η with the content of the top reference variable of the root node ηr of the initial tree.
Given these inference rules, the recognition process is initialized using axioms of the form 〈t[ηs] → •Γ, 0,−,−, 0〉 for each rule t[ηs] → Γ where ηs is the root node of an initial tree labeled with the start symbol. Given an input string w1 . . .wn to be recognized, the goal items in the chart are of the form 〈S → t[ηs]•, 0,−,−, n〉. Once at least one goal item is found in the chart, the recognition process succeeds and the string is successfully accepted by the grammar; otherwise it is rejected. We refer the reader to Appendix A (cf. Figure A.2) for a detailed example of the recognition of the sentence all the meerkats using the proposed inference system.
Note also that although the recognition algorithm we described uses unreduced rules (i.e., generated grammar rules maintaining the full information of nonterminals and the associated index stacks), it is possible to define a more efficient algorithm by having reduced LIG rules and chart items listing only the single top stack element for each constituent (Vijay-Shanker and Weir 1991, 1993). The resulting recognition
algorithm is still complete because the proposed TAG-to-LIG compilation maintains a one-to-one correspondence between the generated rules and their reduced forms (Schabes and Shieber 1994).
As mentioned by Schabes and Shieber (1994), this recognition algorithm can be turned into a parsing algorithm by associating a set of operations with each chart item to build up associated derived trees.
Note also that the derivation trees built as a side effect of the parsing process are the (dependent and/or independent) derivation trees of an FB-LTAG and are therefore context-free.
5.3.3 Handling Spurious Parses. As explained at the end of Section 5.2, the parsing algorithm presented in the previous section systematically allows for dependent and independent adjunction. For example, the recognition of the sentence all the meerkats (Figure A.2) produces both dependent and independent derivations that are not rejected by the unification constraints. In Section 2.2, however, we argued that different types of auxiliary trees license different types of derivations. To capture these distinctions, we modify the recognition algorithm so that it associates scopal auxiliary trees (Example (16a–b)) with dependent derivations only and multiple intersective modifier auxiliary trees (Example (16c)) with only an independent derivation.
(16) a. John thinks that Peter said that the meerkat left. b. The meerkat admired the Syrian orthodox church. c. The tall black meerkat slept.
To block dependent adjunctions between intersective modifiers, we modify the TAG-to-LIG transformation so that, given two intersective modifier trees β1 and β2, no Type 4 or Type 5 rule is produced.
Type 4: Start root of adjunction. For each elementary tree node η in tree β1 that allows the adjunction of the auxiliary tree β2 with root node ηr, the following LIG production rule is generated if and only if β1 and β2 are not intersective modifier auxiliary trees.
b[..η] → t[..ηηr] val(η.PT) ∪ val(ηr.PT)
Type 5: Start foot of adjunction. For each elementary tree node η that allows the adjunction of the auxiliary tree β2 with the foot node ηf , the following LIG production rule is generated if and only if β1 and β2 are not intersective modifier auxiliary trees.
b[..ηηf ] → b[..η] val(η.PB) ∪ val(ηf .PB)
Thus, for instance, in the derivation of all the meerkats depicted in Appendix Figure A.2, the following rules will not be produced, thereby blocking the production of the dependent derivation.
b[.. NPther ] → t[.. NP the r NP all r ] b[.. NPallr ] → t[.. NP all r NP the r ]
} Type4 b[.. NPallr NP the f ] → b[.. NP all r ]
b[.. NPther NP all f ] → b[.. NP the r ]
}
Type5
Similarly, to block independent adjunctions between scopal auxiliary trees, we add a flag scopal? to states in the parsing algorithm. The Type 4 Completor rules associated
with scopal modifiers are modified to mark the progress of a scopal adjunction and to block the independent adjunction of another scopal modifier at the same node.
Type 4 Completor:
〈b[..η] → •t[..ηηr], i,−,−, i, scopal?〉
〈t[..ηηr] → Θ•, i, j, k, l, scopal?〉
〈b[..η] → ∆•, j, p, q, k, scopal?〉
〈b[..η] → t[..ηηr]•, i, p, q, l, True〉 ,
val(η.PT) ∪ val(ηr.PT)
η.PB := ηr.PB
In the Type 4 Completor rule, once a scopal auxiliary treeβ with root node ηr adjoins at some node η, the bottom component of the node η is marked with True, recording that a scopal adjunction has occurred at node η and that it therefore should not accept any further scopal adjunction.
Thus, for instance, the derivation of Syrian orthodox churches will proceed in a similar manner as the derivation of all the meerkats depicted in Appendix Figure A.2, but it will fail to produce the chart items (40, 42, ..., 52)associated with the independent adjunction. Therefore, only the dependent derivation will be produced.
Note that this modification does not block modifier adjunction above a predicative adjunction. Therefore, it successfully recognizes the sentence At what time did Brockway say Harrison arrived?, shown in Example (15), where a wh-modifier needs to be adjoined above a predicative adjunction. Figure 15 shows the complete recognition algorithm modified to rule out spurious parses in the case of multiple scopal auxiliary trees and intersective modifier auxiliary trees.
5.3.4 Weak Generative Equivalence. The weak-generative equivalence refers to the set of strings characterized by the formal system. In contrast, the strong-generative equivalence relates to the set of structural descriptions (such as derivation trees, dags, proof trees, etc.) assigned by a formal system to the strings that it specifies (Vijay-Shankar and Joshi 1985; Joshi 2000).
Using an argument similar to that put forward by Schabes and Shieber (1994), we can prove the weak-generative equivalence of TAGs under the dependent and our independent derivations. We call the set of languages generated by the standard derivation in TAG, TALstd; the set of languages generated by Schabes and Shieber’s extended derivation in TAG, TALextss ; the set of languages generated with our modifications for FB-TAG, TALext; and the set of languages generated by the LIG, LIL. Our derivation allows both dependent and independent derivations; therefore, our derivation will recognize all the strings recognized by the standard (dependent) derivation. More precisely, our derivation can mimic the standard derivation by not allowing more than one adjunction on a tree node by treating all auxiliary trees as scopal auxiliary trees (cf. Section 5.3.3), henceforth, TALstd ⊆ TALext. The proposed compilation from TAGs to LIGs for the independent derivation concluded TALext ⊆ LIL. Finally, LIL ⊆ TALstd has been proven by Vijay-Shanker (1987). Combining these three inclusions, we can conclude that TALstd = TALext. In addition, Schabes and Shieber (1994) have shown that TALstd = TALextss . Hence, we can conclude the weak-generative equivalence of all three derivations in TAGs, TALstd = TALextss = TALext. Feature structures enhance TAGs’ descriptive ability without affecting their generative capacity (Vijay-Shanker and Joshi 1988). The proposed algorithm simulates the established unification mechanism in FBTAG without affecting the representation and the stipulations (e.g., null adjunction at
the foot node and the bounded feature structures) of the grammar itself. Therefore, the association with feature structures will not affect this equivalence. 6 conclusion :Although independent derivations have been shown by Schabes and Shieber (1994) to be essential for correctly supporting syntactic analysis, semantic interpretation, and statistical language modeling, the parsing algorithm they propose is restricted to TAG and is therefore not directly applicable to large scale implemented Feature-Based TAGs. We have provided a recognition algorithm for FB-TAGs that supports both dependent and independent derivations under certain restrictions enforced jointly by feature constraints and by side conditions in the parsing algorithm. The resulting algorithm combines the benefits of independent derivations with those of Feature-Based Grammars. In particular, we showed that it accounts for a range of interactions between dependent vs. independent derivation on the one hand, and syntactic constraints, linear ordering, and scopal vs. nonscopal semantic dependencies on the other hand.
Appendix A: Recognition of the String all the meerkats
We show the recognition of the string all the meerkats as per the inference rules described in Section 5.3.2.
Figure A.1 shows the grammar used for the derivation. To support multiple adjunction in FB-TAG, it implements two main modifications. First, to facilitate the TAG to LIG compilation, each tree node in the grammar is marked with a unique identifier. For example, in Figure A.1, NPmk, NPther , Det the, NPthef *, NP all r , Det all, and NPthef * are unique node identifiers in the grammar. Second, to implement the reassignment mechanism in
NPmk P
Tmk →[Tmk :[]]
P Bmk
→[Bmk:[]]
meerkat
αmeerkat
NPther P Tther →[Tther :[]]
P Bther
→[Bther :[det:the]]
Detthe P Tthe det
→[Tthedet :[]]
P Bthe det →[Bthedet:[]]
NPthef * P Tthe f →
[ Tthef :[det:nil] ]
P Bthe
f →
[ Bthef :[] ]
the
βthe
NPallr P Tallr →[Tallr :[]]
P Ballr
→[Ballr :[det:all]]
Detall P Tall det
→[Talldet:[]]
P Ball det →[Balldet:[]]
NPallf * P Tall f →
[ Tallf :[] ]
P Ball
f →
[ Ballf :[] ]
all
βall
Figure A.1 A feature-based Tree-Adjoining Grammar.
the parsing algorithm, the top (T) and the bottom (B) feature structures of each node are assigned reference variables PT and PB, respectively.
Figure A.2 shows the LIG production rules for the FB-TAG shown in Figure A.1. Table A.1 shows the step-wise recognition of the string all the meerkats. Recall Section 5.3.2: The LIG rules shown in Figure A.1 does not deal with spurious parses and produces all valid derivations, dependent or independent, that are not blocked by feature unification constraints. Hence, for the string all the meerkats, it generates both independent (Step 52) and dependent (Step 53) derivations. As explained in Figures A.3 and A.4, both derivations undergo the identical set of feature unifications. In both figures, prediction rules are abbreviated because they do not enforce any feature unification. In parsing with Tree Adjoining Grammar (TAG), independent derivations have been shown by Schabes and Shieber (1994) to be essential for correctly supporting syntactic analysis, semantic interpretation, and statistical language modeling. However, the parsing algorithm they propose is not directly applicable to Feature-Based TAGs (FB-TAG). We provide a recognition algorithm for FB-TAG that supports both dependent and independent derivations. The resulting algorithm combines the benefits of independent derivations with those of Feature-Based grammars. In particular, we show that it accounts for a range of interactions between dependent vs. independent derivation on the one hand, and syntactic constraints, linear ordering, and scopal vs. nonscopal semantic dependencies on the other hand. [{""affiliations"": [], ""name"": ""Claire Gardent""}, {""affiliations"": [], ""name"": ""Shashi Narayan""}] SP:0296a131d413b037f2c2d5e31e8c13c042469177 [{""authors"": [""Alahverdzhieva"", ""Katya.""], ""title"": ""XTAG using XMG"", ""venue"": ""Master\u2019s thesis, Universit\u00e9 de Nancy."", ""year"": 2008}, {""authors"": [""Candito"", ""Marie-H\u00e9l\u00e8ne"", ""Sylvain Kahane.""], ""title"": ""Can the TAG derivation tree represent a semantic graph? An answer in the light of Meaning-Text Theory"", ""venue"": ""Proceedings of the Fourth Workshop"", ""year"": 1998}, {""authors"": [""Crabb\u00e9"", ""Beno\u0131\u0302t"", ""Denys Duchier"", ""Claire Gardent"", ""Joseph Le Roux"", ""Yannick Parmentier""], ""title"": ""XMG: eXtensible MetaGrammar"", ""venue"": ""Computational Linguistics,"", ""year"": 2013}, {""authors"": [""Gardent"", ""Claire.""], ""title"": ""Integrating a unification-based semantics in a large scale Lexicalised Tree Adjoining Grammar for French"", ""venue"": ""Proceedings of the 22nd International Conference on"", ""year"": 2008}, {""authors"": [""Gardent"", ""Claire"", ""Laura Kallmeyer.""], ""title"": ""Semantic construction in Feature-Based TAG"", ""venue"": ""Proceedings of the 10th Conference of the European Chapter of the Association for Computational Linguistics (EACL),"", ""year"": 2003}, {""authors"": [""Gazdar"", ""Gerald.""], ""title"": ""Applicability of indexed grammars to natural languages"", ""venue"": ""Uwe Reyle and Christian Rohrer, editors, Natural Language Parsing and Linguistic Theories, volume 35 of Studies"", ""year"": 1988}, {""authors"": [""Joshi"", ""Aravind K.""], ""title"": ""Relationship between strong and weak generative power of formal systems"", ""venue"": ""Proceedings of the 5th International Workshop on Tree Adjoining Grammars and Related Formalisms (TAG+5),"", ""year"": 2000}, {""authors"": [""Joshi"", ""Aravind K."", ""Yves Schabes.""], ""title"": ""Tree-Adjoining Grammars"", ""venue"": ""Handbook of Formal Languages and Automata, 3:69\u2013124."", ""year"": 1997}, {""authors"": [""Joshi"", ""Aravind K"", ""K. Vijay-Shanker""], ""title"": ""Compositional semantics with lexicalized tree-adjoining grammar (LTAG): How"", ""year"": 2001}, {""authors"": [""Kallmeyer"", ""Laura"", ""Marco Kuhlmann.""], ""title"": ""A formal model for plausible dependencies in lexicalized tree adjoining grammar"", ""venue"": ""Proceedings of the 11th International Workshop on"", ""year"": 2012}, {""authors"": [""Kroch"", ""Anthony.""], ""title"": ""Asymmetries in long distance extraction in a tree adjoining grammar"", ""venue"": ""M. Baltin and A. Kroch, editors, Alternative conceptions of phrase structure, University of Chicago Press,"", ""year"": 1989}, {""authors"": [""Rambow"", ""Owen"", ""K. Vijay-Shanker"", ""David Weir.""], ""title"": ""D-Tree Grammars"", ""venue"": ""Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics (ACL), pages 151\u2013158, Cambridge, MA."", ""year"": 1995}, {""authors"": [""Resnik"", ""Philip.""], ""title"": ""Probabilistic tree-adjoining grammar as a framework for statistical natural language processing"", ""venue"": ""Proceedings of the 14th Conference on Computational linguistics (COLING),"", ""year"": 1992}, {""authors"": [""Schabes"", ""Yves.""], ""title"": ""The valid prefix property and left to right parsing of tree-adjoining grammar"", ""venue"": ""Proceedings of the 2nd International Workshop on Parsing Technologies (IWPT), pages 21\u201330, Cancun."", ""year"": 1991}, {""authors"": [""Schabes"", ""Yves.""], ""title"": ""Stochastic lexicalized tree-adjoining grammars"", ""venue"": ""Proceedings of the 14th Conference on Computational Linguistics (COLING), volume 2, pages 425\u2013432, Nantes."", ""year"": 1992}, {""authors"": [""Schabes"", ""Yves"", ""Aravind K. Joshi.""], ""title"": ""An Earley-type parsing algorithm for tree adjoining grammars"", ""venue"": ""Proceedings of the 26th Annual Meeting of the Association for Computational Linguistics (ACL),"", ""year"": 1988}, {""authors"": [""Schabes"", ""Yves"", ""Stuart M. Shieber.""], ""title"": ""An alternative conception of tree-adjoining derivation"", ""venue"": ""Proceedings of the 30th Annual Meeting of the Association for Computational Linguistics (ACL), pages 167\u2013176,"", ""year"": 1992}, {""authors"": [""Schabes"", ""Yves"", ""Stuart M. Shieber.""], ""title"": ""An alternative conception of tree-adjoining derivation"", ""venue"": ""Computational Linguistics, 20(1):91\u2013124."", ""year"": 1994}, {""authors"": [""Shieber"", ""Stuart M""], ""title"": ""Restricting the weak-generative capacity of synchronous"", ""year"": 1994}, {""authors"": [""Steedman"", ""Mark.""], ""title"": ""The syntactic process"", ""venue"": ""MIT Press, Cambridge, MA."", ""year"": 2000}, {""authors"": [""The XTAG Research Group.""], ""title"": ""A lexicalized tree adjoining grammar for English"", ""venue"": ""Technical report ICRS-01-03, Institute for Research in Cognitive Science, University of Pennsylvannia."", ""year"": 2001}, {""authors"": [""K. Vijay-Shankar"", ""Aravind K. Joshi.""], ""title"": ""Some computational properties of tree adjoining grammars"", ""venue"": ""Proceedings of the 23rd Annual Meeting of the Association for Computational Linguistics (ACL),"", ""year"": 1985}, {""authors"": [""K. Vijay-Shanker""], ""title"": ""A Study of Tree Adjoining Grammars"", ""venue"": ""Ph.D. thesis, Department of Computer and Information Science, University of Pennsylvania, Philadelphia."", ""year"": 1987}, {""authors"": [""K. Vijay-Shanker"", ""Arvind K. Joshi.""], ""title"": ""Feature structures based tree adjoining grammars"", ""venue"": ""Proceedings of the 12th International Conference on Computational Linguistics (COLING),"", ""year"": 1988}, {""authors"": [""K. Vijay-Shanker"", ""Arvind K. Joshi.""], ""title"": ""Unification-based tree adjoining grammar"", ""venue"": ""Technical report MS-CIS-91-25, School of Engineering and Applied Science, Department of Computer and Information"", ""year"": 1991}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""Polynomial parsing of extensions of Context-Free Grammars"", ""venue"": ""Masaru Tomita, editor, Current Issues in Parsing Technology, volume 126 of The Springer"", ""year"": 1991}, {""authors"": [""K. Vijay-Shanker"", ""David J. Weir.""], ""title"": ""Parsing some constrained grammar formalisms"", ""venue"": ""Computational Linguistics, 19(4):591\u2013636."", ""year"": 1993}, {""authors"": [""Weir"", ""David J."", ""Aravind K. Joshi.""], ""title"": ""Combinatory Categorial Grammars: Generative power and relationship to linear context-free rewriting systems"", ""venue"": ""Proceedings of the 26th Annual Meeting"", ""year"": 1988}] acknowledgments  :We would like to thank Laura Perez-Beltrachini for useful discussion on the related topic. We are grateful to our anonymous reviewers for their insightful comments, which helped us improve the quality of the article in terms of both presentation and content.",
,,,,,,,,"Jane Robinson, a pioneering computational linguist, made major contributions to machine translation, natural language, and speech systems research programs at the RAND Corporation, IBM, and in the AI Center at SRI International. She served as ACL president in 1982. Jane became a computational linguist accidentally. She had a Ph.D. in history from UCLA, but could not obtain a faculty position in that field because those were reserved for men. Instead, she took positions teaching English, first at UCLA and then at California State College, Los Angeles. While at LA State, where she was tasked with teaching engineers how to write, Jane noticed an announcement for a talk on Chomsky’s transformational grammar. She went to the talk thinking this work on grammar might help her teach better. Although its subject matter did not match her expectations, the talk marked a turning point in her career. In the late 1950s, Jane became a consultant to the RAND Corporation group working on machine translation under Dave Hays (ACL president, 1964). From the beginning, Jane was concerned with identifying connections between different traditions in formal grammars and their corresponding detailed linguistic realizations. Her 1965 International Conference on Computational Linguistics (COLING) paper, “Endocentric constructions and the Cocke parsing logic” (Robinson 1965), is a beautiful example of connecting specific linguistic phenomena to parsing strategies in a way that preserves the nature of the linguistic phenomena, endocentric constructions. While at RAND, Jane became colleague and friend to many in the machine translation and emerging computational linguistics world, including Susumo Kuno (ACL president 1967), Martin Kay (ACL president, 1969), Joyce Friedman (ACL president, 1971), and Karen Sparck Jones (ACL president, 1994). In the late 1960s, Jane moved to the Automata Theory and Computability Group at the IBM Thomas J. Watson Research Center in Yorktown Heights, NY. She used her knowledge of formal work on grammars and parsing to draw correspondences between Dependency Grammars and Phrase Structure Grammars. Although Jane came","[{""affiliations"": [], ""name"": ""Jane J. Robinson""}, {""affiliations"": [], ""name"": ""Barbara J. Grosz""}, {""affiliations"": [], ""name"": ""Eva Hajicova""}, {""affiliations"": [], ""name"": ""Aravind Joshi""}]",SP:9767d3c59a3ce828f5d7052c19d89becb4c46b7d,"[{""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Endocentric constructions and the cocke parsing logic"", ""venue"": ""Proceedings of the 1965 Conference on Computational Linguistics, COLING \u201965, pages 1\u201323, Stroudsburg, PA."", ""year"": 1965}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Methods for obtaining corresponding phrase structure and dependency grammars"", ""venue"": ""Proceedings of the 1967 Conference on Computational Linguistics, COLING \u201967, pages 1\u201325,"", ""year"": 1967}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Case, category, and configuration"", ""venue"": ""Journal of Linguistics, 6(1):57\u201380."", ""year"": 1970}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Dependency structures and transformational rules"", ""venue"": ""Language, 46(2):259\u2013285."", ""year"": 1970}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Time to go"", ""venue"": ""http://poemsofjanerobinson.blogspot. com/2008/06/time-to-go.html. 726"", ""year"": 2008}]","acknowledgments  :This remembrance was a composite of the reminiscences of a number of people, including Tom Garvey, Marguerite Hays, Peter Hart, Gary Hendrix, Jerry Hobbs, David Israel, Ray Perrault, Candy Sidner, and Marty Tenenbaum.",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,obituary  :,"jane j. robinson :Barbara J. Grosz∗
Harvard University
Eva Hajicova Charles University in Prague
Aravind Joshi University of Pennsylvania
Jane Robinson, a pioneering computational linguist, made major contributions to machine translation, natural language, and speech systems research programs at the RAND Corporation, IBM, and in the AI Center at SRI International. She served as ACL president in 1982.
Jane became a computational linguist accidentally. She had a Ph.D. in history from UCLA, but could not obtain a faculty position in that field because those were reserved for men. Instead, she took positions teaching English, first at UCLA and then at California State College, Los Angeles. While at LA State, where she was tasked with teaching engineers how to write, Jane noticed an announcement for a talk on Chomsky’s transformational grammar. She went to the talk thinking this work on grammar might help her teach better. Although its subject matter did not match her expectations, the talk marked a turning point in her career.
In the late 1950s, Jane became a consultant to the RAND Corporation group working on machine translation under Dave Hays (ACL president, 1964). From the beginning, Jane was concerned with identifying connections between different traditions in formal grammars and their corresponding detailed linguistic realizations. Her 1965 International Conference on Computational Linguistics (COLING) paper, “Endocentric constructions and the Cocke parsing logic” (Robinson 1965), is a beautiful example of connecting specific linguistic phenomena to parsing strategies in a way that preserves the nature of the linguistic phenomena, endocentric constructions. While at RAND, Jane became colleague and friend to many in the machine translation and emerging computational linguistics world, including Susumo Kuno (ACL president 1967), Martin Kay (ACL president, 1969), Joyce Friedman (ACL president, 1971), and Karen Sparck Jones (ACL president, 1994).
In the late 1960s, Jane moved to the Automata Theory and Computability Group at the IBM Thomas J. Watson Research Center in Yorktown Heights, NY. She used her knowledge of formal work on grammars and parsing to draw correspondences between Dependency Grammars and Phrase Structure Grammars. Although Jane came
∗ E-mail: grosz@eecs.harvard.edu.
doi:10.1162/COLI a 00235
© 2015 Association for Computational Linguistics
from the Dependency Grammar tradition, her balanced, careful analysis of tradeoffs enabled others to bridge the approaches. Her 1967 COLING paper, “Methods for obtaining corresponding phrase structure and dependency grammars” (Robinson 1967), is a wonderful example of her understanding of the seminal issues underlying these different systems. She subsequently published the classic paper connecting dependency structure and transformational rules, “Dependency structures and transformational rules” (Robinson 1970b). This paper exemplifies Jane’s scholarship and her deftness in dealing with the formal and computational issues of language processing in a very fair and informative manner. Her 1970 paper, “Case, category and configuration” (Robinson 1970a), demonstrated in a very convincing way the possibility of formally interpreting Fillmore’s case grammar in terms of dependencies, in a more economic fashion and without any loss of information.
In 1973, Don Walker (ACL secretary-treasurer, 1976–1993) recruited Jane to the speech group in the AI Center (AIC) at SRI International. Jane remained a key member of the AIC’s natural language group until she retired in the mid 1980s. She made major contributions to a wide range of research, ranging from grammars for speech understanding systems and dialogues to such discourse issues as codes and clues in contexts. For several of the NLP systems SRI developed in the 1970s and 1980s, the grammar was the coordinating point for all knowledge about the language, so Jane interacted with everyone developing any component of the system, from the architecture through semantics and discourse. Speech processing was a bit “deaf” in those days, and Jane frequently remarked that she had to write not only a grammar for English, but also its dual (one for non-English to rule out bad parsings). During her time at SRI, Jane wrote some of the most comprehensive grammars for NLP systems.
One of us (Barbara Grosz) notes that Jane’s contributions to the AIC’s natural language group went far beyond her official grammar writing responsibilities. She served as mentor (before that word was widely used in academia) for a large group of “young Turks,” as she referred to those of us in the younger cohort involved in building NLP systems; she was our in-house expert in linguistics; provided critiques of drafts of papers, making them shorter, clearer, and more scholarly; debugged our ideas across the full spectrum of system components; and introduced us to the most senior people in linguistics and computational linguistics.
Another of us (Aravind Joshi, ACL president, 1975) recalls meeting Jane in September 1975 at an NLP workshop at the University of Michigan. At that time, he and his students were working on two separate areas of computational linguistics. One of them concerned the minimal formal machinery needed for representing structural aspects of language and the other one dealt with some aspects of cooperation in natural language interfaces to databases. After just a brief discussion with Jane, when she asked what he was doing, it became clear to him, “that I was in the presence of someone who had already worked on such diverse areas. From thereon, whenever I had an opportunity to meet with Jane, I took advantage of her deep understanding.”
And the third of us (Eva Hajicova, ACL president, 1998) recalls the important ties Jane formed with Praguian linguists interested in formal grammar starting in 1965, when the founder of the Prague group, Petr Sgall, first visited the United States and met Jane at RAND. Their common research interests in computational linguistics, particularly Dependency Grammar, forged deep personal relationships between Jane and Sgall’s linguistics group in Prague. Jane visited Prague twice, once before and once after the change of the communist regime. During difficult political times, Jane provided the Prague group with linguistics literature published in the West, and she introduced them to her colleagues and students, yielding additional important connections, which
have continued to this day. Eva notes that, “only those who have the same historical experience as we in Prague have had can appreciate fully how important such activities were for our research and for our students.”
Jane read broadly and her training as a historian made her a careful and deep scholar. Throughout her life she would go to talks that seemed a bit far afield and come back with new ideas. For those who worked with her she was an invaluable source of out-of-the-box thinking as well as the go-to person for what to read in linguistics. Jane often said that had she been born in a later generation, she would have become an astronomer. Given her love of exploration, she might have been an astronaut. (You can see this bent in the poem she wrote for her poetry class friend’s funeral, “Time To Go” [Robinson 2008]). Lucky for all of us, she wound up in computational linguistics.
Jane was the mother of four children and the proud grandmother of two grandsons, one now a lawyer, the other an actor. She extended her family to encompass her colleagues, building camaraderie through dinners at her home—lamb stew, mostly— and picnics at Foothills Park in Palo Alto, activities which drew the families of AIC researchers together and yielded many lifelong friendships. Jane built such friendships throughout her career. In responding to our questions about Jane’s time at RAND, Dave Hays’s wife Rita noted that Jane remained close friends with her and Dave even after Dave went to SUNY Buffalo and Jane to IBM Yorktown Heights and then SRI. Rita and Jane were traveling and hiking companions into Jane’s nineties.
To celebrate her sixtieth birthday, Jane “got in shape” to hike in the Himalayas around Annapurna. She came back with beautiful photos and the desire to join the young Turks who went backpacking in Yosemite. (Although she had been a regular visitor to Yosemite since her forties, she had not backpacked before.) She offered to drive everyone in the huge Chrysler Imperial she had gotten to feel safe driving in New York. Jane backpacked into her late seventies, then switched to the luxury of the High Sierra Camp tent cabins and later to a small cabin with an inside shower. For those lucky enough to visit Yosemite with her, she was as much a guide to the mountains as she had been a guide to linguistics.
For people who worked with Jane at SRI, she was a towering figure in the field, a wonderful colleague who imparted deep wisdom as well as linguistic facts, and a dear friend. She was senior to most of the members of the AIC. Looking back on her arrival, Peter Hart (a director of the AIC) noted that Jane’s presence changed the tone of the early SRI AI Center. She brought not only a keen intellect and depth of knowledge, but “also a gentleness, openness, and generosity of spirit.” Gary Hendrix (who led the NLP group at SRI for several years) remembers that for many who worked with her at SRI, “Jane was like a second mother, loving and giving and nurturing.” Jerry Hobbs (ACL president, 1990) recalls that, “one of the things I learned from her, though imperfectly, was how to be tough with grace.” Several remember Jane as one of a handful of elder statesmen that their generation could look up to.
Jane was a colleague and friend of Ray Perrault (ACL president, 1983, and current AIC director) from the time he was a Ph.D. student at the University of Michigan. In the early 1970s, Jane frequently visited Joyce Friedman (ACL president, 1971), her old friend and his advisor. Ray remembers Jane was a gentle but firm critic of his thesis, and he fondly recalls her passing through in her huge Chrysler on her move from IBM to SRI. Ray was ACL vice president when Jane was president, and she was delighted when he decided to join the young Turks at SRI. Subsequently, she “even tolerated me as her manager until her retirement and became doting godmother to my son.”
Jane made a difference in people’s lives, not just their research. Her death marks the end of an era and the passing of an icon. We will miss her greatly.",,"Jane Robinson, a pioneering computational linguist, made major contributions to machine translation, natural language, and speech systems research programs at the RAND Corporation, IBM, and in the AI Center at SRI International. She served as ACL president in 1982. Jane became a computational linguist accidentally. She had a Ph.D. in history from UCLA, but could not obtain a faculty position in that field because those were reserved for men. Instead, she took positions teaching English, first at UCLA and then at California State College, Los Angeles. While at LA State, where she was tasked with teaching engineers how to write, Jane noticed an announcement for a talk on Chomsky’s transformational grammar. She went to the talk thinking this work on grammar might help her teach better. Although its subject matter did not match her expectations, the talk marked a turning point in her career. In the late 1950s, Jane became a consultant to the RAND Corporation group working on machine translation under Dave Hays (ACL president, 1964). From the beginning, Jane was concerned with identifying connections between different traditions in formal grammars and their corresponding detailed linguistic realizations. Her 1965 International Conference on Computational Linguistics (COLING) paper, “Endocentric constructions and the Cocke parsing logic” (Robinson 1965), is a beautiful example of connecting specific linguistic phenomena to parsing strategies in a way that preserves the nature of the linguistic phenomena, endocentric constructions. While at RAND, Jane became colleague and friend to many in the machine translation and emerging computational linguistics world, including Susumo Kuno (ACL president 1967), Martin Kay (ACL president, 1969), Joyce Friedman (ACL president, 1971), and Karen Sparck Jones (ACL president, 1994). In the late 1960s, Jane moved to the Automata Theory and Computability Group at the IBM Thomas J. Watson Research Center in Yorktown Heights, NY. She used her knowledge of formal work on grammars and parsing to draw correspondences between Dependency Grammars and Phrase Structure Grammars. Although Jane came [{""affiliations"": [], ""name"": ""Jane J. Robinson""}, {""affiliations"": [], ""name"": ""Barbara J. Grosz""}, {""affiliations"": [], ""name"": ""Eva Hajicova""}, {""affiliations"": [], ""name"": ""Aravind Joshi""}] SP:9767d3c59a3ce828f5d7052c19d89becb4c46b7d [{""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Endocentric constructions and the cocke parsing logic"", ""venue"": ""Proceedings of the 1965 Conference on Computational Linguistics, COLING \u201965, pages 1\u201323, Stroudsburg, PA."", ""year"": 1965}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Methods for obtaining corresponding phrase structure and dependency grammars"", ""venue"": ""Proceedings of the 1967 Conference on Computational Linguistics, COLING \u201967, pages 1\u201325,"", ""year"": 1967}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Case, category, and configuration"", ""venue"": ""Journal of Linguistics, 6(1):57\u201380."", ""year"": 1970}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Dependency structures and transformational rules"", ""venue"": ""Language, 46(2):259\u2013285."", ""year"": 1970}, {""authors"": [""Robinson"", ""Jane J.""], ""title"": ""Time to go"", ""venue"": ""http://poemsofjanerobinson.blogspot. com/2008/06/time-to-go.html. 726"", ""year"": 2008}] acknowledgments  :This remembrance was a composite of the reminiscences of a number of people, including Tom Garvey, Marguerite Hays, Peter Hart, Gary Hendrix, Jerry Hobbs, David Israel, Ray Perrault, Candy Sidner, and Marty Tenenbaum. obituary  : jane j. robinson :Barbara J. Grosz∗
Harvard University
Eva Hajicova Charles University in Prague
Aravind Joshi University of Pennsylvania
Jane Robinson, a pioneering computational linguist, made major contributions to machine translation, natural language, and speech systems research programs at the RAND Corporation, IBM, and in the AI Center at SRI International. She served as ACL president in 1982.
Jane became a computational linguist accidentally. She had a Ph.D. in history from UCLA, but could not obtain a faculty position in that field because those were reserved for men. Instead, she took positions teaching English, first at UCLA and then at California State College, Los Angeles. While at LA State, where she was tasked with teaching engineers how to write, Jane noticed an announcement for a talk on Chomsky’s transformational grammar. She went to the talk thinking this work on grammar might help her teach better. Although its subject matter did not match her expectations, the talk marked a turning point in her career.
In the late 1950s, Jane became a consultant to the RAND Corporation group working on machine translation under Dave Hays (ACL president, 1964). From the beginning, Jane was concerned with identifying connections between different traditions in formal grammars and their corresponding detailed linguistic realizations. Her 1965 International Conference on Computational Linguistics (COLING) paper, “Endocentric constructions and the Cocke parsing logic” (Robinson 1965), is a beautiful example of connecting specific linguistic phenomena to parsing strategies in a way that preserves the nature of the linguistic phenomena, endocentric constructions. While at RAND, Jane became colleague and friend to many in the machine translation and emerging computational linguistics world, including Susumo Kuno (ACL president 1967), Martin Kay (ACL president, 1969), Joyce Friedman (ACL president, 1971), and Karen Sparck Jones (ACL president, 1994).
In the late 1960s, Jane moved to the Automata Theory and Computability Group at the IBM Thomas J. Watson Research Center in Yorktown Heights, NY. She used her knowledge of formal work on grammars and parsing to draw correspondences between Dependency Grammars and Phrase Structure Grammars. Although Jane came
∗ E-mail: grosz@eecs.harvard.edu.
doi:10.1162/COLI a 00235
© 2015 Association for Computational Linguistics
from the Dependency Grammar tradition, her balanced, careful analysis of tradeoffs enabled others to bridge the approaches. Her 1967 COLING paper, “Methods for obtaining corresponding phrase structure and dependency grammars” (Robinson 1967), is a wonderful example of her understanding of the seminal issues underlying these different systems. She subsequently published the classic paper connecting dependency structure and transformational rules, “Dependency structures and transformational rules” (Robinson 1970b). This paper exemplifies Jane’s scholarship and her deftness in dealing with the formal and computational issues of language processing in a very fair and informative manner. Her 1970 paper, “Case, category and configuration” (Robinson 1970a), demonstrated in a very convincing way the possibility of formally interpreting Fillmore’s case grammar in terms of dependencies, in a more economic fashion and without any loss of information.
In 1973, Don Walker (ACL secretary-treasurer, 1976–1993) recruited Jane to the speech group in the AI Center (AIC) at SRI International. Jane remained a key member of the AIC’s natural language group until she retired in the mid 1980s. She made major contributions to a wide range of research, ranging from grammars for speech understanding systems and dialogues to such discourse issues as codes and clues in contexts. For several of the NLP systems SRI developed in the 1970s and 1980s, the grammar was the coordinating point for all knowledge about the language, so Jane interacted with everyone developing any component of the system, from the architecture through semantics and discourse. Speech processing was a bit “deaf” in those days, and Jane frequently remarked that she had to write not only a grammar for English, but also its dual (one for non-English to rule out bad parsings). During her time at SRI, Jane wrote some of the most comprehensive grammars for NLP systems.
One of us (Barbara Grosz) notes that Jane’s contributions to the AIC’s natural language group went far beyond her official grammar writing responsibilities. She served as mentor (before that word was widely used in academia) for a large group of “young Turks,” as she referred to those of us in the younger cohort involved in building NLP systems; she was our in-house expert in linguistics; provided critiques of drafts of papers, making them shorter, clearer, and more scholarly; debugged our ideas across the full spectrum of system components; and introduced us to the most senior people in linguistics and computational linguistics.
Another of us (Aravind Joshi, ACL president, 1975) recalls meeting Jane in September 1975 at an NLP workshop at the University of Michigan. At that time, he and his students were working on two separate areas of computational linguistics. One of them concerned the minimal formal machinery needed for representing structural aspects of language and the other one dealt with some aspects of cooperation in natural language interfaces to databases. After just a brief discussion with Jane, when she asked what he was doing, it became clear to him, “that I was in the presence of someone who had already worked on such diverse areas. From thereon, whenever I had an opportunity to meet with Jane, I took advantage of her deep understanding.”
And the third of us (Eva Hajicova, ACL president, 1998) recalls the important ties Jane formed with Praguian linguists interested in formal grammar starting in 1965, when the founder of the Prague group, Petr Sgall, first visited the United States and met Jane at RAND. Their common research interests in computational linguistics, particularly Dependency Grammar, forged deep personal relationships between Jane and Sgall’s linguistics group in Prague. Jane visited Prague twice, once before and once after the change of the communist regime. During difficult political times, Jane provided the Prague group with linguistics literature published in the West, and she introduced them to her colleagues and students, yielding additional important connections, which
have continued to this day. Eva notes that, “only those who have the same historical experience as we in Prague have had can appreciate fully how important such activities were for our research and for our students.”
Jane read broadly and her training as a historian made her a careful and deep scholar. Throughout her life she would go to talks that seemed a bit far afield and come back with new ideas. For those who worked with her she was an invaluable source of out-of-the-box thinking as well as the go-to person for what to read in linguistics. Jane often said that had she been born in a later generation, she would have become an astronomer. Given her love of exploration, she might have been an astronaut. (You can see this bent in the poem she wrote for her poetry class friend’s funeral, “Time To Go” [Robinson 2008]). Lucky for all of us, she wound up in computational linguistics.
Jane was the mother of four children and the proud grandmother of two grandsons, one now a lawyer, the other an actor. She extended her family to encompass her colleagues, building camaraderie through dinners at her home—lamb stew, mostly— and picnics at Foothills Park in Palo Alto, activities which drew the families of AIC researchers together and yielded many lifelong friendships. Jane built such friendships throughout her career. In responding to our questions about Jane’s time at RAND, Dave Hays’s wife Rita noted that Jane remained close friends with her and Dave even after Dave went to SUNY Buffalo and Jane to IBM Yorktown Heights and then SRI. Rita and Jane were traveling and hiking companions into Jane’s nineties.
To celebrate her sixtieth birthday, Jane “got in shape” to hike in the Himalayas around Annapurna. She came back with beautiful photos and the desire to join the young Turks who went backpacking in Yosemite. (Although she had been a regular visitor to Yosemite since her forties, she had not backpacked before.) She offered to drive everyone in the huge Chrysler Imperial she had gotten to feel safe driving in New York. Jane backpacked into her late seventies, then switched to the luxury of the High Sierra Camp tent cabins and later to a small cabin with an inside shower. For those lucky enough to visit Yosemite with her, she was as much a guide to the mountains as she had been a guide to linguistics.
For people who worked with Jane at SRI, she was a towering figure in the field, a wonderful colleague who imparted deep wisdom as well as linguistic facts, and a dear friend. She was senior to most of the members of the AIC. Looking back on her arrival, Peter Hart (a director of the AIC) noted that Jane’s presence changed the tone of the early SRI AI Center. She brought not only a keen intellect and depth of knowledge, but “also a gentleness, openness, and generosity of spirit.” Gary Hendrix (who led the NLP group at SRI for several years) remembers that for many who worked with her at SRI, “Jane was like a second mother, loving and giving and nurturing.” Jerry Hobbs (ACL president, 1990) recalls that, “one of the things I learned from her, though imperfectly, was how to be tough with grace.” Several remember Jane as one of a handful of elder statesmen that their generation could look up to.
Jane was a colleague and friend of Ray Perrault (ACL president, 1983, and current AIC director) from the time he was a Ph.D. student at the University of Michigan. In the early 1970s, Jane frequently visited Joyce Friedman (ACL president, 1971), her old friend and his advisor. Ray remembers Jane was a gentle but firm critic of his thesis, and he fondly recalls her passing through in her huge Chrysler on her move from IBM to SRI. Ray was ACL vice president when Jane was president, and she was delighted when he decided to join the young Turks at SRI. Subsequently, she “even tolerated me as her manager until her retirement and became doting godmother to my son.”
Jane made a difference in people’s lives, not just their research. Her death marks the end of an era and the passing of an icon. We will miss her greatly.",
,,,,,,,,"Recognizing textual entailment (RTE) has been proposed as a task in computational linguistics under a successful series of annual evaluation campaigns started in 2005 with the Pascal RTE-1 shared task. RTE is defined as the capability of a system to recognize that the meaning of a portion of text (usually one or few sentences) entails the meaning of another portion of text. Subsequently, the task has also been extended to recognizing specific cases of non-entailment, as when the meaning of the first text contradicts the meaning of the second text. Although the study of entailment phenomena in natural language was addressed much earlier, the novelty of the RTE evaluation was to propose a simple text-to-text task to compare human and system judgments, making it possible to build data sets and to experiment with a variety of approaches. Two main reasons likely contributed to the success of the initiative: First, the possibility to address, for the first time, the complexity of entailment phenomena under a data-driven perspective; second, the text-to-text approach allows one to easily incorporate a textual entailment engine into applications (e.g., question answering, summarization, information extraction) as a core inferential component.","[{""affiliations"": [], ""name"": ""Ido Dagan""}, {""affiliations"": [], ""name"": ""Dan Roth""}, {""affiliations"": [], ""name"": ""Mark Sammons""}, {""affiliations"": [], ""name"": ""Fabio Massimo Zanzotto""}]",SP:3fe56745c946a1869bfa630b62b21325f9dff869,[],,,,,,,,,,,,,,,,,,,"reviewed by bernardo magnini fondazione bruno kessler :Recognizing textual entailment (RTE) has been proposed as a task in computational linguistics under a successful series of annual evaluation campaigns started in 2005 with the Pascal RTE-1 shared task. RTE is defined as the capability of a system to recognize that the meaning of a portion of text (usually one or few sentences) entails the meaning of another portion of text. Subsequently, the task has also been extended to recognizing specific cases of non-entailment, as when the meaning of the first text contradicts the meaning of the second text. Although the study of entailment phenomena in natural language was addressed much earlier, the novelty of the RTE evaluation was to propose a simple text-to-text task to compare human and system judgments, making it possible to build data sets and to experiment with a variety of approaches. Two main reasons likely contributed to the success of the initiative: First, the possibility to address, for the first time, the complexity of entailment phenomena under a data-driven perspective; second, the text-to-text approach allows one to easily incorporate a textual entailment engine into applications (e.g., question answering, summarization, information extraction) as a core inferential component.
© 2015 Association for Computational Linguistics
Given the large interest in RTE, a book on the topic is very much welcome. The book satisfies at least two needs in the community. On the one hand, it was time to collect and organize a large amount of research and publications produced in the last decade or so. Under this perspective a particular merit of the book is to set a homogeneous and general descriptive framework under which different approaches can be reported and compared with a uniform conceptual schema. On the other hand, after a decade of research in the field, it is the right time to open the discussion both to the results that have been achieved and to future challenges in RTE. With this perspective the book targets not only people interested in textual entailment but people interested in semantic inference and semantic processing at large, intended as a core capability of natural language understanding. Researchers working in related areas such as textual similarity, causality, and temporal reasoning, just to mention a few, may benefit from this monograph.
The book is organized into six chapters, covering RTE motivations, task definitions, approaches and data-driven methodologies, relevant case studies, knowledge acquisition for RTE and, finally, future research perspectives.
Chapter 1 provides the necessary context on textual entailment and RTE. Even a reader not familiar with the field can find in this chapter motivations, definitions, many
doi:10.1162/COLI r 00213
Computational Linguistics Volume 41, Number 1
examples, a complete report on the RTE evaluations with their data sets, and pointers to a large variety of applications. The chapter also tries to state both commonalities and differences between the “textual” and the “logical” notion of entailment. I found this part somehow less elaborated and less convincing, particularly because the authors seem to consider logics only for deductive argumentation, where the conclusion from one or more premises is true in every circumstance. Actually, logics have elaborated theories of probabilistic argumentation where the conclusion has different degrees of plausibility, as the textual entailment definition suggests.
Chapter 2 introduces architecture and approaches proposed for recognizing textual entailment. The goal of the chapter is to show the complexity of the RTE task and to provide the main options that have been investigated to date. Chapter 3 is really the core of the book, as it provides a general conceptualization of the RTE task based on alignments between the two texts at different levels (token, lexical, syntactic, etc.). The chapter considers three steps in the process: first, candidate alignments are established between the two portions of text; then, the more promising alignments are selected; and, finally, an entailment decision is taken based on classification, possibly using a machine learning algorithm. Very useful, this chapter provides enough background on machine learning that even the non-familiar reader can get oriented in the technical description.
Chapter 4 presents several case studies about RTE systems. This is an interesting and complete overview, which includes, among others, systems exploiting edit distance, logical representation and inference, several kinds of transformations between the two portions of text, alignment-focused approaches (where alignment is seen as a part of the inference process), and, finally, ensemble systems (i.e., systems that combine other entailment systems). Chapter 5 focuses on knowledge resources for textual entailment, providing an exhaustive overview both on the use of existing lexical resources (e.g., WordNet) and on methodologies for the automatic acquisition of entailment rules, including recent work based on distributional similarity methods and acquisition from parallel and comparable corpora. The chapter does not avoid a final discussion on problematic aspects of the use of knowledge resources for textual entailment, particularly considering that, despite the fact that a huge number of entailment rules has been acquired and managed by different systems, their impact on performance is still debated in the community. The last chapter of the book discusses research directions in RTE, pointing out the fundamental role of high-quality pre-processing (e.g., parsing, semantic role labeling) in natural language processing, as well as the need for a shared software environment (e.g., an open-source platform for textual inferences) in order to increase replicability and comparability of approaches.
Overall, the book provides a very good overview of research on textual entailment, where state-of-the-art approaches and systems are clearly presented and compared with the help of a common conceptual schema, with many examples and bibliographical references and, when necessary, providing technical details. The book indubitably covers a need of the RTE community, and will be a reference for students and researchers approaching this field. Although the book addresses almost all of the RTE work so far, there are a few aspects for which, stimulated by the book itself, the reader more familiar with the topic would have expected more discussion. The first aspect concerns the relation between text similarity and textual entailment, which is somehow latent throughout the book (Chapter 3 presents similarity features for the task), although not explicitly addressed. As semantic text similarity (STS) in the last years has been quite successful, and as there are evident overlaps between RTE and STS systems, a discussion about the potential interactions among these communities would have been useful. There is a second aspect that might have required a deeper discussion. Although
158
Book Reviews
recognizing textual entailment, at least in the perspective suggested by the book, requires the capability to establish meaningful alignments between two portions of text, the book devotes little space to opposite phenomena, namely, discovering reasons for non-entailment, or for strong dissimilarities. The impression is that contradiction detection is not yet completely integrated within the RTE data-driven conceptual framework (not surprisingly, the topic is mentioned in the case study on natural logic in Chapter 4) and that further investigation is necessary in the future.
It is evident that the main source of inspiration for the book is the activity around RTE evaluations in the last ten years, yet this book goes well beyond being simply a summary of what has happened. The reader interested in potential future research directions in textual entailment will certainly appreciate the generic architecture presented in Chapter 2, the alignment-based conceptualization of the RTE task presented in Chapter 3, the open perspectives of Chapter 6, and, finally, the huge bibliographical selection. Overall, this book is very well positioned to measure both the maturity and the future ambitions of an important research area in computational linguistics.
This book review was edited by Pierre Isabelle.
Bernardo Magnini is senior researcher at FBK in Trento, Italy, where he is Co-Responsible of the Research Unit on Human Language Technology. His research interests include question answering and textual entailment. He is currently the scientific coordinator of the EXCITEMENT project (EXploring Customer Interactions through Textual EntailMENT), funded by the European Commission. Magnini’s address is FBK, Via Sommarive 18, 38123, Trento, Italy; e-mail: magnini@fbk.eu.
159",,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"recognizing textual entailment: models and applications :Ido Dagan1, Dan Roth2, Mark Sammons2, and Fabio Massimo Zanzotto3 (1Bar-Ilan University, Israel, 2University of Illinois, Urbana, IL, and 3University of Rome “Tor Vergata,” Italy)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 23), 2013, xx+200 pp; paperbound, ISBN 978-1-59829-834-5; e-book, ISBN 978-1-59829-835-2","Recognizing textual entailment (RTE) has been proposed as a task in computational linguistics under a successful series of annual evaluation campaigns started in 2005 with the Pascal RTE-1 shared task. RTE is defined as the capability of a system to recognize that the meaning of a portion of text (usually one or few sentences) entails the meaning of another portion of text. Subsequently, the task has also been extended to recognizing specific cases of non-entailment, as when the meaning of the first text contradicts the meaning of the second text. Although the study of entailment phenomena in natural language was addressed much earlier, the novelty of the RTE evaluation was to propose a simple text-to-text task to compare human and system judgments, making it possible to build data sets and to experiment with a variety of approaches. Two main reasons likely contributed to the success of the initiative: First, the possibility to address, for the first time, the complexity of entailment phenomena under a data-driven perspective; second, the text-to-text approach allows one to easily incorporate a textual entailment engine into applications (e.g., question answering, summarization, information extraction) as a core inferential component. [{""affiliations"": [], ""name"": ""Ido Dagan""}, {""affiliations"": [], ""name"": ""Dan Roth""}, {""affiliations"": [], ""name"": ""Mark Sammons""}, {""affiliations"": [], ""name"": ""Fabio Massimo Zanzotto""}] SP:3fe56745c946a1869bfa630b62b21325f9dff869 [] reviewed by bernardo magnini fondazione bruno kessler :Recognizing textual entailment (RTE) has been proposed as a task in computational linguistics under a successful series of annual evaluation campaigns started in 2005 with the Pascal RTE-1 shared task. RTE is defined as the capability of a system to recognize that the meaning of a portion of text (usually one or few sentences) entails the meaning of another portion of text. Subsequently, the task has also been extended to recognizing specific cases of non-entailment, as when the meaning of the first text contradicts the meaning of the second text. Although the study of entailment phenomena in natural language was addressed much earlier, the novelty of the RTE evaluation was to propose a simple text-to-text task to compare human and system judgments, making it possible to build data sets and to experiment with a variety of approaches. Two main reasons likely contributed to the success of the initiative: First, the possibility to address, for the first time, the complexity of entailment phenomena under a data-driven perspective; second, the text-to-text approach allows one to easily incorporate a textual entailment engine into applications (e.g., question answering, summarization, information extraction) as a core inferential component.
© 2015 Association for Computational Linguistics
Given the large interest in RTE, a book on the topic is very much welcome. The book satisfies at least two needs in the community. On the one hand, it was time to collect and organize a large amount of research and publications produced in the last decade or so. Under this perspective a particular merit of the book is to set a homogeneous and general descriptive framework under which different approaches can be reported and compared with a uniform conceptual schema. On the other hand, after a decade of research in the field, it is the right time to open the discussion both to the results that have been achieved and to future challenges in RTE. With this perspective the book targets not only people interested in textual entailment but people interested in semantic inference and semantic processing at large, intended as a core capability of natural language understanding. Researchers working in related areas such as textual similarity, causality, and temporal reasoning, just to mention a few, may benefit from this monograph.
The book is organized into six chapters, covering RTE motivations, task definitions, approaches and data-driven methodologies, relevant case studies, knowledge acquisition for RTE and, finally, future research perspectives.
Chapter 1 provides the necessary context on textual entailment and RTE. Even a reader not familiar with the field can find in this chapter motivations, definitions, many
doi:10.1162/COLI r 00213
Computational Linguistics Volume 41, Number 1
examples, a complete report on the RTE evaluations with their data sets, and pointers to a large variety of applications. The chapter also tries to state both commonalities and differences between the “textual” and the “logical” notion of entailment. I found this part somehow less elaborated and less convincing, particularly because the authors seem to consider logics only for deductive argumentation, where the conclusion from one or more premises is true in every circumstance. Actually, logics have elaborated theories of probabilistic argumentation where the conclusion has different degrees of plausibility, as the textual entailment definition suggests.
Chapter 2 introduces architecture and approaches proposed for recognizing textual entailment. The goal of the chapter is to show the complexity of the RTE task and to provide the main options that have been investigated to date. Chapter 3 is really the core of the book, as it provides a general conceptualization of the RTE task based on alignments between the two texts at different levels (token, lexical, syntactic, etc.). The chapter considers three steps in the process: first, candidate alignments are established between the two portions of text; then, the more promising alignments are selected; and, finally, an entailment decision is taken based on classification, possibly using a machine learning algorithm. Very useful, this chapter provides enough background on machine learning that even the non-familiar reader can get oriented in the technical description.
Chapter 4 presents several case studies about RTE systems. This is an interesting and complete overview, which includes, among others, systems exploiting edit distance, logical representation and inference, several kinds of transformations between the two portions of text, alignment-focused approaches (where alignment is seen as a part of the inference process), and, finally, ensemble systems (i.e., systems that combine other entailment systems). Chapter 5 focuses on knowledge resources for textual entailment, providing an exhaustive overview both on the use of existing lexical resources (e.g., WordNet) and on methodologies for the automatic acquisition of entailment rules, including recent work based on distributional similarity methods and acquisition from parallel and comparable corpora. The chapter does not avoid a final discussion on problematic aspects of the use of knowledge resources for textual entailment, particularly considering that, despite the fact that a huge number of entailment rules has been acquired and managed by different systems, their impact on performance is still debated in the community. The last chapter of the book discusses research directions in RTE, pointing out the fundamental role of high-quality pre-processing (e.g., parsing, semantic role labeling) in natural language processing, as well as the need for a shared software environment (e.g., an open-source platform for textual inferences) in order to increase replicability and comparability of approaches.
Overall, the book provides a very good overview of research on textual entailment, where state-of-the-art approaches and systems are clearly presented and compared with the help of a common conceptual schema, with many examples and bibliographical references and, when necessary, providing technical details. The book indubitably covers a need of the RTE community, and will be a reference for students and researchers approaching this field. Although the book addresses almost all of the RTE work so far, there are a few aspects for which, stimulated by the book itself, the reader more familiar with the topic would have expected more discussion. The first aspect concerns the relation between text similarity and textual entailment, which is somehow latent throughout the book (Chapter 3 presents similarity features for the task), although not explicitly addressed. As semantic text similarity (STS) in the last years has been quite successful, and as there are evident overlaps between RTE and STS systems, a discussion about the potential interactions among these communities would have been useful. There is a second aspect that might have required a deeper discussion. Although
158
Book Reviews
recognizing textual entailment, at least in the perspective suggested by the book, requires the capability to establish meaningful alignments between two portions of text, the book devotes little space to opposite phenomena, namely, discovering reasons for non-entailment, or for strong dissimilarities. The impression is that contradiction detection is not yet completely integrated within the RTE data-driven conceptual framework (not surprisingly, the topic is mentioned in the case study on natural logic in Chapter 4) and that further investigation is necessary in the future.
It is evident that the main source of inspiration for the book is the activity around RTE evaluations in the last ten years, yet this book goes well beyond being simply a summary of what has happened. The reader interested in potential future research directions in textual entailment will certainly appreciate the generic architecture presented in Chapter 2, the alignment-based conceptualization of the RTE task presented in Chapter 3, the open perspectives of Chapter 6, and, finally, the huge bibliographical selection. Overall, this book is very well positioned to measure both the maturity and the future ambitions of an important research area in computational linguistics.
This book review was edited by Pierre Isabelle.
Bernardo Magnini is senior researcher at FBK in Trento, Italy, where he is Co-Responsible of the Research Unit on Human Language Technology. His research interests include question answering and textual entailment. He is currently the scientific coordinator of the EXCITEMENT project (EXploring Customer Interactions through Textual EntailMENT), funded by the European Commission. Magnini’s address is FBK, Via Sommarive 18, 38123, Trento, Italy; e-mail: magnini@fbk.eu.
159 recognizing textual entailment: models and applications :Ido Dagan1, Dan Roth2, Mark Sammons2, and Fabio Massimo Zanzotto3 (1Bar-Ilan University, Israel, 2University of Illinois, Urbana, IL, and 3University of Rome “Tor Vergata,” Italy)
Morgan & Claypool (Synthesis Lectures on Human Language Technologies, edited by Graeme Hirst, volume 23), 2013, xx+200 pp; paperbound, ISBN 978-1-59829-834-5; e-book, ISBN 978-1-59829-835-2",
"1 introduction :The goal of semantic parsing is to automatically process natural language text and map the underlying meaning of text to appropriate representations. Semantic role labeling
∗ School of Informatics, University of Edinburgh, EH8 9AB Edinburgh, United Kingdom. E-mail: mroth@inf.ed.ac.uk.
∗∗ Department of Computational Linguistics, Heidelberg University, 69120 Heidelberg, Germany. E-mail: frank@cl.uni-heidelberg.de.
Submission received: 5 June 2014; revised version received: 13 April 2015; accepted for publication: 28 August 2015.
doi:10.1162/COLI a 00236
© 2015 Association for Computational Linguistics
induces shallow semantic representations, so-called predicate–argument structures, by processing sentences and mapping them to predicates and associated arguments. Arguments of these structures can, however, be non-local in natural language text, as shown in Example 1.
Example 1 (a) El Salvador is the only Latin American country which has troops in Iraq.
(b) Nicaragua withdrew its troops last month.1
Applying a semantic role labeling system on sentence (1b) produces a representation that consists of the predicate withdraw, a temporal modifier (last month) and two associated arguments: the entity withdrawing (Nicaragua) and the thing being withdrawn (its troops). From the previous sentence (1a), we can additionally infer a third argument: namely, the source from which Nicaragua withdrew its troops (Iraq). By leaving this piece of information implicit, the text fragment in sentence (1b) illustrates a typical case of non-local, or implicit, role realization (Gerber and Chai 2012). In this article, we view implicit arguments as a discourse-level phenomenon and treat corresponding instances as implicit references to discourse entities.
Taking this perspective, we build upon previous work on discourse analysis. Following Sidner (1979) and Joshi and Kuhn (1979), utterances in discourse typically focus on a set of salient entities, which are also called the foci or centers. Using the notion of centers, Grosz, Joshi, and Weinstein (1995) defined the Centering framework, which relates the salience of an entity in discourse to linguistic factors such as choice of referring expression and syntactic form.2 Both extremes of salience, that is, contexts of referential continuity and irrelevance, can also be reflected by the non-realization of an entity (Brown 1983). Specific instances of this phenomenon, so-called zero anaphora, have been well-studied in pro-drop languages such as Japanese (Kameyama 1985), Turkish (Turan 1995), and Italian (Di Eugenio 1990). For English, only a few studies exist that explicitly investigated the effect of non-realizations on coherence. Existing work suggests, however, that indirect references and non-realizations are important for modeling and measuring coherence (Poesio et al. 2004; Karamanis et al. 2009), respectively, and that such phenomena need to be taken into consideration to explain local coherence where adjacent sentences are neither connected by discourse relations nor in terms of coreference (Louis and Nenkova 2010).
In this work, we propose a new model to predict whether realizing an argument contributes to local coherence in a given position in discourse. Example 1 illustrated a text fragment, in which argument realization is necessary in the first sentence but redundant in the second. That is, mentioning Iraq in the second sentence is not necessary (for a human being) to understand the meaning of the text. In contrast, making both references explicit, as shown in Example 2, would be redundant and could lead to the perception that the text is merely a concatenation of two independent sentences — rather than a set of adjacent sentences that form a meaningful, or coherent, discourse.
1 In the remainder of this article, we use the following notation for predicate–argument structures: underlined words refer to predicates and [possibly empty] constituents in [square brackets] refer to arguments. 2 We note that the relation between coherence and choice of referring expressions has also been studied outside of the Centering framework. A comprehensive overview of research in linguistics on this subject can be found in Arnold (1998).
Example 2 (a) El Salvador is now the only Latin American country which has troops in Iraq.
(b) [Nicaragua] withdrew [its troops] [from Iraq] last month.
The phenomenon of implicit arguments has neither been extensively studied in the context of semantic role labeling nor in coherence modeling. One of the main reasons for this lies in the fact that models for these tasks are typically developed on the basis of annotated corpora. In contrast, implicit arguments are not overt in text and are difficult to uncover, hence there exist only few and small data sets that contain respective annotations explicitly. In this article, we present a novel framework for computationally inducing a corpus with automatic annotations of implicit arguments and their respective antecedents in discourse. To achieve this goal, our framework exploits pairs of comparable texts, which convey information about the same events, states, and entities but potentially differ in terms of depth and perspective. These differences can also affect argument realization, making it possible to identify instances of co-referring but only partially overlapping argument structures. Given automatically induced instances of implicit arguments and their document contexts, we show how these can be utilized to enhance current models of local coherence and implicit argument linking. The resulting models and research insights will be of importance for many applications that involve the understanding or generation of natural language text beyond the sentence level.","2 overview :The goal of this work is to provide data and methods to better capture the phenomenon of implicit arguments in semantic role labeling and coherence modeling. To accomplish this goal, we propose inducing suitable training data automatically. For semantic role labeling, this kind of training data will contain explicit links between implicit arguments and their respective discourse antecedents; for modeling coherent argument realization, the data will also provide discourse contexts of explicit and implicit references to the same entity. We propose inducing such information by exploiting pairs of comparable texts. Our motivation for this lies in the fact that comparable texts convey by and large the same information but differ in depth and presentation. This means that while we expect two comparable texts to contain references to the same events and participants, an entity might be understood implicitly at a specific point in one text (because it can be inferred from context), whereas an explicit mention might be necessary in a comparable text (given the different context). Based on this assumption, our method aims to find complementary (explicit) information in pairs of texts, which can be aligned and subsequently be projected to identify missing (implicit) pieces in one another. Hence, we separate our task into two parts. In the first part, described in Section 4, we ask the question:
Q1: How can we detect predicate–argument structures that are shared across two discourses?
Based on a corpus of comparable texts, we propose a new task of aligning predicate– argument structures (PASs) across complete discourses. We define this task on the basis of semantic annotations, which we compute automatically using a state-of-the-art semantic role labeler. For the alignment task itself, we develop a graph-based clustering
model. We show that by taking into account features on the levels of predicate, arguments, and discourse context, this model is able to predict accurate alignments across comparable texts, without relying on preprocessing methods that identify parallel text fragments. Based on automatically aligned structures, the second part of our approach aims to identify and resolve implicit arguments by projecting explicit information from one text to another. The question for this part of our approach (cf. Section 5) is:
Q2: How can we find implicit arguments and their antecedents in discourse?
Our approach to answering this question relies solely on information that can be computed automatically: We first identify potential instances of implicit arguments by comparing two aligned predicate–argument structures; for each entity that is mentioned in one PAS (explicit argument) but not in an aligned structure (implicit argument), we then apply a cross-document coreference resolution technique to also find co-referring entity mentions in the document in which an implicit reference was found.
Finally, we demonstrate the utility of automatically induced instances of implicit arguments in two task-based settings: linking arguments in discourse and modeling local coherence. The question here is:
Q3: How can induced argument instances and their respective discourse contexts be utilized to enhance NLP models?
We address both tasks by applying our automatically induced data set as training material to improve the performance of supervised models. For the first task, described in Section 6, we apply our data to enhance training of an existing system that identifies and links implicit arguments in discourse. To evaluate the impact of our induced data set, we test the modified model on a standard evaluation data set, on which we compare our results with those of previous work. For the second task, we develop a new coherence model that predicts whether an argument realization or non-realization in context improves the perceived coherence of the affected segment in discourse. We evaluate this coherence model in an intrinsic evaluation scenario, in which we compare model predictions to human judgments on argument use (cf. Section 7).","3 background :This article draws on insights from computational linguistic research on implicit arguments and coherence modeling as well as from previous work on inducing semantic resources. It is further related to recent work in paraphrasing and event coreference. The goal of this work is to induce instances of implicit arguments, together with their discourse antecedents, and to utilize them in semantic processing and coherence modeling. This section summarizes previous work on implicit arguments and coherence modeling, and provides an outlook on how instances of implicit arguments can be of use in a novel entity-based model of local coherence.
Implicit Arguments. The role of implicit arguments was studied early on in the context of semantic processing (Fillmore 1986; Palmer et al. 1986), although most semantic
role labeling systems nowadays operate solely within local syntactic structures and do not perform any additional inference regarding missing information. First data sets that focus on implicit arguments have only recently become available: Ruppenhofer et al. (2010) organized a SemEval shared task on “linking events and participants in discourse,” Gerber and Chai (2012) made available implicit argument annotations for NomBank (Meyers, Reeves, and Macleod 2008), and Moor, Roth, and Frank (2013) provided annotations for parts of the OntoNotes corpus (Weischedel et al. 2011). All three resources are, however, severely limited: Annotations in the latter two studies are restricted to 10 and 5 predicate types, respectively; the training set of the SemEval task, in contrast, consists of full-text annotations for all occurring predicates but contains only 245 instances of resolved implicit arguments in total. All groups working on the shared task identified data sparsity as one of the main issues (Chen et al. 2010; Ruppenhofer et al. 2012; Laparra and Rigau 2013). Silberer and Frank (2012) point out that additional training data can be heuristically created by treating anaphoric pronoun mentions as implicit arguments. Their experimental results confirmed that artificial training data can indeed improve results, but only when obtained from corpora with manual semantic role annotations (on the sentence level) and gold coreference chains. As observed in related work on classifying discourse relations, information learned from artificial training data might not always generalize well to naturally occurring examples (Sporleder and Lascarides 2008). To automatically create data that is linguistically more similar to manually labeled implicit arguments, we introduce an alternative method that induces instances of implicit arguments from a raw corpus of comparable texts.
Coherence Modeling. In the context of coherence modeling, much previous work has focused on entity-based approaches, with the most prominent model being the entity grid by Barzilay and Lapata (2005). This model has originally been proposed for automatic sentence ordering but has since also been applied in coherence evaluation and readability assessment (Barzilay and Lapata 2008; Pitler and Nenkova 2008), story generation (McIntyre and Lapata 2010), and authorship attribution (Feng and Hirst 2014). Based on the original model, several extensions have been proposed: For example, Filippova and Strube (2007) and Elsner and Charniak (2011b) suggested additional features to characterize semantic relatedness between entities and features specific to single entities, respectively. Other entity-based approaches to coherence modeling include the pronoun model by Charniak and Elsner (2009) and the discourse-new model by Elsner and Charniak (2008). All of these approaches are, however, solely based on explicit entity mentions, resulting in insufficient representations when dealing with inferable references. Example 3 illustrates this shortcoming.
Example 3 implicit roles (a) 27 tons of cigarettes were picked up in Le Havre. agent
(b) The containers had arrived yesterday. content, destination
The two sentences in the given example do not share any explicit references to the same entity. Hence, none of the aforementioned models is able to correctly predict that both sentences cohere in the presented order—but not in the reversed order.
As discussed in more detail by Poesio et al. (2004), such cases can still be resolved when taking into account indirect realizations, that is, associative references to a discourse entity. Accordingly, recent work by Hou, Markert, and Strube (2013) proposed
considering bridging resolution to fill specific gaps in the entity grid model. Following this argument, the containers in sentence (3b) could be understood as a bridging anaphor that refers back to the 27 tons of cigarettes in sentence (3a). As mentioned in the NomBank annotation guidelines (Meyers 2007), some cases of nominal predicates and their arguments do coincide with bridging relations and we treat them as implicit arguments in this article. For example, when interpreting container as a nominalization of the verbal predicate CONTAIN, the cigarettes from the previous sentence can be understood as one of its arguments: namely, its content. Not all cases of implicit arguments involve bridging anaphora and vice versa, however. The main difference lies in the fact that bridging typically describes a relation between entities. Hence, bridging instances only coincide with those of implicit arguments if the bridging anaphor is linguistically realized as a (nominal) predicate (container in Example 3b). In contrast, implicit arguments frequently form parts of predicate–argument structures that refer to events, states, and entities that are not bridging anaphora. For example, the destination of the predicate ARRIVE in sentence (3b) can also be interpreted as an implicit argument, namely, Le Havre from the preceding context.
Taking implicit arguments into consideration, we can see that the second sentence refers back to two previously introduced entities, Le Havre and cigarettes, reflecting the fact that the sentences cohere in the presented order. In Section 7, we present a novel model of local coherence that aims to capture the effect of argument realizations on perceived coherence. As an overall goal, this model should be able to predict whether an entity reference should be realized explicitly to establish (local) coherence—or whether the entity can already be understood from context. Based on such predictions, the model can be applied in text generation to ensure that necessary references are explicit and that redundant repetitions are avoided. The methods applied in this article are based on ideas from previous work on inducing semantic resources from parallel and comparable texts. Most work in this direction has been done in the context of cross-lingual settings, including the learning of translations of words and phrases using statistical word alignments (Kay and Röscheisen 1993; DeNero, Bouchard-Côté, and Klein 2008, inter alia) and approaches to projecting annotations from one language to another (Yarowsky and Ngai 2001; Kozhevnikov and Titov 2013, inter alia). In the following, we discuss previous approaches to annotation projection as well as related work in paraphrasing and event coreference.
Annotation Projection. A widespread method for the induction of semantic resources is the so-called annotation projection approach. The rationale of this approach is to induce annotated data in one language, given already-annotated instances in another language. As an example, semantic role annotations of a text in English can be transferred to a parallel text in order to induce annotated instances for a lexicon in another language (Padó and Lapata 2009). In previous work, this method has been applied on various levels of linguistic analysis: from syntactic information in the form of part-of-speech tags and dependencies (Yarowsky and Ngai 2001; Hwa et al. 2005), through annotation of temporal expressions and semantic roles (Spreyer and Frank 2008; van der Plas, Merlo, and Henderson 2011), to discourse-level phenomena such as coreference (Postolache, Cristea, and Orasan 2006). All of the aforementioned instances of the projection approach make use of the same underlying technique: Firstly, words are
aligned in a parallel corpus using statistical word alignment; secondly, annotations on a single word or between multiple words in one text are transferred to the corresponding aligned word(s) in the parallel text. This procedure typically assumes that two parallel sentences express the same meaning. A notable exception is the work by Fürstenau and Lapata (2012), which utilizes alignments between syntactic structures to “project” semantic role information from a role-annotated corpus to unseen sentences that are selected from a corpus in the same language.
In our work, we apply annotation projection to monolingual comparable texts. In comparison to parallel texts, we have to account for potential differences in perspective and detail that make our task—in particular, the alignment sub-task—considerably more difficult. In contrast to Fürstenau and Lapata’s setting, which involves incomparable texts, we assume that text pairs in our setting still convey information on the same events and participants. This means that in addition to aligning predicate–structures across texts, we can also merge complementary details realized in each structure. We achieve this by making use of the same principle that underlies the projection approach: Given annotation that is available in one text (explicit argument), we project this information to a text in which the corresponding annotation is not available (implicit argument). Because our task is based in a monolingual setting, we can make use of the same preprocessing tools across texts.
Paraphrasing and Event Coreference. We overcome the difficulty of inducing word alignments across comparable texts by computing alignments on the basis of predicate– argument structures. Using predicate–argument structures as targets makes our setting related to previous work on paraphrase detection and coreference resolution of event mentions. Each of these tasks focuses, however, on a different level of linguistic analysis from ours: Following the definitions embraced by Recasens and Vila (2010), “paraphrasing” is a relation between two lexical units that have the same meaning, whereas “coreference” indicates that two referential expressions point to the same referent in discourse.3 In contrast to work on paraphrasing, we are specifically interested in pairs of text fragments that involve implicit arguments, which can only be resolved in context.
In line with our goal of inducing implicit arguments, we define the units or expressions to be aligned in our task as the predicate–argument structures that can (automatically) be identified in text. This task definition further makes our task distinct from event coreference, where coreference is established based on a pre-specified set of events, reference types, or definitions of event identity (Walker et al. 2006; Pradhan et al. 2007; Lee et al. 2012, inter alia). Although corresponding annotations can certainly overlap with those in our task, we emphasize that the focus of our work is not to find all occurrences of co-referring events. Instead, our goal is to align predicate–argument structures that have a common meaning in context across different discourses. Hence, we neither consider intra-document coreference nor pronominal event references here. As alignable units in our work are not restricted to a pre-specified definition of event or event identity, the task addressed here involves any kind of event, state, or entity that is linguistically realized as a predicate–argument structure. Examples that go beyond traditional event coreference include in particular noun phrases such as “a [rival] of the company” and “the [finance] director [of IBM]”(cf. Meyers, Reeves, and Macleod 2008).
3 Note that in the remainder of this article, we use the term “coreference” in a wider sense to also encompass referents to events and entities in the real world, rather than just referents grounded in discourse.","4 induction framework step 1: aligning predicate–argument structures :In the first step of our induction framework, we align predicate–argument structures (PASs) across pairs of comparable text. To execute this task properly, we define annotation guidelines and construct a gold standard for the development and evaluation of alignment models. All annotations are performed on pairs of documents from a corpus of comparable texts. We introduce this corpus and describe our annotations in Section 4.1. We describe a graph-based model that we developed for the automatic alignment in Section 4.2. Finally, we present an experimental evaluation of our model, together with another recently proposed model and several baselines, in Section 4.3. As a basis for aligning predicate–argument structures across texts, we make use of a data set of comparable texts extracted from the English Gigaword corpus (Parker et al. 2011). The Gigaword corpus is one of the largest English corpora available in the news domain and contains over 9.8 million articles from seven newswire agencies that report on (the same) real-world incidents. The data set of comparable texts used in this work contains 167,728 pairs of articles that were extracted by matching the headlines of texts published within the same time frame (Roth and Frank 2012a). A set of such document headlines is given in Example 4:
Example 4 India fires tested anti-ship cruise missile (Xinhua News Agency, 29 October 2003)
India tests supersonic cruise anti-ship missile (Agence France Presse, 29 October 2003)
URGENT: India tests anti-ship cruise missile (Associated Press Worldstream, 29 October 2003)
We preprocess each article in the set of 167,728 pairs using the MATE tools (Björkelund et al. 2010; Bohnet 2010), including a state-of-the-art semantic role labeler that identifies PropBank/NomBank-style predicate–argument structures (Palmer, Gildea, and Kingsbury 2005; Meyers, Reeves, and Macleod 2008). Based on the acquired PAS, we perform manual alignments. In Section 4.1.1, we summarize the annotation guidelines for this step. An overview of the resulting development and evaluation data set is provided in Section 4.1.2.
4.1.1 Manual Annotation. We selected 70 pairs of comparable texts and asked two annotators to manually align predicate–argument structures obtained from preprocessing. Both annotators were students in computational linguistics, one undergraduate and one postgraduate. The texts were selected with the constraint that each text consists of 100 to 300 words. We chose this constraint as longer text pairs seemed to contain a higher number of unrelated predicates, making the alignment tasks difficult to manage for the annotators. Both annotators received detailed guidelines that describe alignment requirements and the overall procedure.4 We summarize essentials in the following.
4 cf. http://projects.cl.uni-heidelberg.de/india/files/guidelines.pdf.
Sure and Possible Links. Following standard practice in word alignment (cf. Cohn, Callison-Burch and Lapata 2008, inter alia) the annotators were instructed to distinguish between sure (S) and possible (P) alignments, depending on how certainly, in their opinion, two predicates (including their arguments) describe the same event, state, or entity. Examples 5 and 6 show a sure and possible predicate pairing, respectively.
Example 5 The regulator ruled on September 27 that Nasdaq was qualified to bid. The authority had already approved a similar application by Nasdaq.
Example 6 Myanmar’s government said it has released some 220 political prisoners. The government has been regularly releasing members of Suu Kyi’s party.
Replaceability. As a rule of thumb for deciding whether to align two structures, annotators were told to check how well the affected predicate–argument structures could be replaced by one another in their given context.
Missing Context. In case one text does not provide enough context to decide whether two predicates in the paired documents refer to the same event, an alignment should not be marked as sure.
Similar Predicates. Annotators were told explicitly that sure links can be used even if two predicates are semantically different but have the same meaning in context. Example 7 illustrates such a case.
Example 7 The volcano roared back to life two weeks ago. It began erupting last month.
1-to-1 vs. n-to-m. We asked the annotators to find as many 1-to-1 correspondences as possible and to prefer 1-to-1 matches over n-to-m alignments. In case of multiple mentions of the same event, we further asked the annotators to provide only one sure link per predicate and mark remaining cases as possible links. As an additional guideline, annotators were asked to only label the PAS pair with the highest information overlap as a sure link. If there is no difference in information overlap, the predicate pair that occurs first in both texts should be marked as a sure alignment. The intuition behind this guideline is that the first mention introduces the actual event whereas later mentions just (co-)refer or add further information.
4.1.2 Resulting Data Set. In total, the annotators (A/B) aligned 487/451 sure and 221/180 possible alignments. Following Brockett (2007), we computed agreement on labeled annotations, including unaligned predicate pairs as an additional null category. We computed κ following Fleiss, Levin, and Paik (1981) and observed an overall score of 0.62, with κ values per category of 0.74 and 0.19 for sure and possible alignments, respectively. The numbers show that both annotators substantially agree on which pairs of predicate–argument structures “surely” express the same proposition. Identifying further references to the same event or state, in contrast, can only be achieved with
fairly low agreement. For the construction of a gold standard, we take the intersection of all sure alignments by both annotators and the union of all possible alignments.5 We further resolved cases that involved a sure alignment on which the annotators disagreed in a group discussion and added them to our gold standard accordingly. We split the final corpus into a development set of 10 document pairs and an evaluation set of 60 document pairs.
Table 1 summarizes information about the resulting annotations in the development and evaluation set. As can be seen, the documents in the development set contain a smaller number of predicates (39.5 vs. 57.6) and alignments (8.7 vs. 13.4) on average. The fraction of aligned predicates is, however, about the same (22.0% vs. 23.3%). Across both data sets, the average numbers of observed predicates is approximately 55, of which 31 are verbs and 24 are nouns. In the development and evaluation sets, the average number of sure alignments are 3.5 and 7.4. From all aligned predicate pairs in both data sets, 82.6% are the same part of speech (30.0% both nouns, 52.6% both verbs). In total, 48.0% of all alignments are between predicates of identical lemmata. As a rough indicator for diverging argument structures captured in the annotated alignments, we analyzed the number of aligned predicates that involve a different number of realized arguments. In both data sets together, this criterion applied in 344 cases (38.9% of all alignments). For the automatic alignment of predicate–argument structure alignments across texts, we opt for an unsupervised graph-based method. That is, we represent pairs of documents as bipartite graphs and subsequently aim to separate a graph into subgraphs that represent corresponding predicate–argument structures across texts. We define our general graph representation for this task in Section 4.2.1. In Section 4.2.2, we introduce
5 In our evaluation, only sure alignments need to be predicted by a system, whereas possible alignments are optional and not counted towards the attested recall (cf. Section 4.3).
a range of similarity measures that are used to weight edges in the graph representation. In Section 4.2.3, we introduce our algorithm to separate graphs, representing pairs of documents, into subgraphs of corresponding predicates. The overall model is an extended variant of the clustering approach described in Roth and Frank (2012b) and uses an enhanced set of similarity measures.
4.2.1 Graph Representation. We build a bipartite graph representation for each pair of texts, using as vertices the predicate–argument structures assigned during preprocessing (cf. Section 4.1). We represent each predicate as a node and integrate information about arguments implicitly. Given the sets of predicates P1 and P2 of two comparable texts T1 and T2, respectively, we formally define an undirected graph GP1,P2 following Equation 7.
GP1,P2 = 〈V, E〉 where V = P1 ∪ P2 E = P1 × P2
(7)
Edge Weights. We specify the edge weight between two nodes representing predicates p1 ∈ P1 and p2 ∈ P2 as a weighted linear combination of the similarity measures S described in the next section.
wp1p2 = |S|∑ i λi ∗ simi(p1, p2) (8)
Initially we set all weighting parameters λi to have uniform weights by default. We describe a tuning routine to find an optimized weighting scheme for the individual measures in the experimental evaluation of our approach (Section 4.3).
4.2.2 Similarity Measures. We use a number of similarity measures that make use of complementary information on the predicates, arguments, and discourse context of two predicate–argument structures. Given two lemmatized predicates p1, p2 and their sets of arguments A1 = args(p1), A2 = args(p2), we define seven measures in total. Three of them are specific to the predicates themselves, two take into account information on associated arguments, and two measures capture discourse-level properties. The predicate-specific measures as well as one argument-specific measure correspond to measures previously described in Roth and Frank (2012b). This article extends previous work by considering discourse-level information and an additional argument-specific measure that takes into account argument labels. We demonstrate the benefits of these measures in practice in Section 4.3.
Similarity in WordNet. Given all synsets that contain the two predicates p1, p2, we compute their similarity in WordNet (Fellbaum 1998) as the maximal pairwise score calculated using the information content based measure proposed by Lin (1998). We rely on the WordNet hierarchy to find the least common subsumer (lcs) of two synsets and use the pre-computed Information Content (IC) files from Pedersen (2010) to compute this measure as defined in Equation (9).
simWN(p1, p2) = max 〈s1,s2〉:si∈synsets(pi ) IC(lcs(s1, s2)) IC(s1) ∗ IC(s2)
(9)
Similarity in VerbNet. We additionally make use of VerbNet (Kipper et al. 2008) to compute similarities between verb pairs that cannot be captured by WordNet relations. Verbs in VerbNet are categorized into classes according to their meaning as well as syntactic behavior. A verb class C can recursively embed sub-classes Cs ∈ sub(C) that represent finer semantic and syntactic distinctions. In Equation (10), we define a simple similarity function that assigns fixed scores to pairs of predicates p1, p2 depending on their relatedness within VerbNet.6
simVN(p1, p2) =  1.0 if ∃C : p1, p2 ∈ C 0.8 if ∃C, Cs : Cs ∈ sub(C)
∧ ((p1 ∈ C, p2 ∈ Cs) ∨ (p1 ∈ Cs, p2 ∈ C)) default else
(10)
We empirically set the default value to the average VerbNet similarity (with unrelated pairs counted as 0.0) computed over one million random pairs of predicates in our corpus.
Similarity in a Semantic Space. As predicates can be absent from WordNet and VerbNet, or distributed over separate hierarchies due to different parts-of-speech (verbal vs. nominal predicates), we additionally calculate similarity based on distributional meaning in a semantic space (Landauer and Dumais 1997). This measure is based on the similarity of contexts of two given predicates over all their instances in a corpus. To compute this measure, we first calculate the Pointwise Mutual Information (PMI) for each predicate p ∈ {p1, p2} and the n most frequent context words c ∈ C following Equation (11).
pmi(p, c) = freq(p, c)
freq(p) ∗ freq(c) (11)
As we are dealing with predicates of different parts-of-speech, we calculate joint frequencies in terms of context windows instead of relying on syntactic dependencies as proposed in more recent approaches to distributional semantics (Padó and Lapata 2007; Erk and Padó 2008; Baroni and Lenci 2010). More precisely, we extract context windows of five words to the left and to the right from the Gigaword corpus (Parker et al. 2011), and compute the PMI for the 2,000 most frequent context words c1 . . . c2,000 ∈ C. The same setting has been successfully applied in related tasks, including word sense disambiguation (Guo and Diab 2011) and measuring phrase similarity (Mitchell and Lapata 2010). Vector representations are computed following Equation (12), and similarities are calculated as the cosine function of the angle between two vectors, as defined in Equation (13).
~p = (pmi(p, c1), pmi(p, c2), . . . , pmi(p, c2,000)) (12)
simDist(p1, p2) = ~p1 · ~p2
||~p1|| ∗ ||~p2|| (13)
6 Note that the weight of 0.8 was set in an ad hoc manner (instead of being optimized) in order to avoid overfitting on our small development corpus.
Bag-of-Words Similarity. As a simple argument-specific measure, we compute the overlap of word tokens over all (concatenated) arguments of each predicate–argument structure (PAS). Formally, this similarity measure considers all arguments a1 ∈ A1 and a2 ∈ A2 associated with the predicates p1 and p2, respectively, and calculates overlap as defined in Equation (14). In order to control the impact of frequently occurring words, we weight each word by its Inverse Document Frequency (IDF), which we calculate over all documents d in our corpus D:
simABoW(p1, p2) =
∑ w∈A1∩A2 idf(w)∑
w∈A1 idf(w) + ∑ w∈A2 idf(w) (14)
idf(w) = log |D|
|{d ∈ D|w ∈ words(d)}| (15)
Head of Arguments Similarity. We further define an argument-specific measure that only compares the semantic heads of arguments that have the same argument label. For example, given two PASs that consist of predicates p1, p2 and arguments labeled A0 and A1, we compute the similarity of the two arguments labeled A0 (also denoted as label(a) = ‘A0’) and the similarity of the two arguments labeled A1 (label(a) = ‘A1’). Each similarity between argument heads is computed using the WordNet-based measure described earlier. We extract the semantic head of each argument by considering the dependency tree, as predicted by MATE tools, in which we look for the noun or verb on the highest level within the argument span. Finally, we collapse all pairwise argument similarities into one measure by taking the average following Equation (16).
simAheads(p1, p2) =
∑ {a1,a2|label(a1 )==label(a2 )} simWN(head(a1), head(a2))
|{a1, a2|label(a1) == label(a2)}| (16)
Relative Discourse Position. On the discourse level, we measure the distance of two predicate–argument structures with respect to their relative positions in the text. Given two predicates, we compute this measure based on the absolute difference between their relative positions. The relative position in discourse is computed as the sentence index in which the predicate p1 or p2 occurs, divided by the total number of sentences in the affected document (d1 or d2, respectively). The measure, as defined in Equation (17), ranges from 1.0 (relative positions are exactly the same) to 0.0 (one predicate at the beginning, the other at the end of a text).
simDPos(p1, p2) = 1− (∣∣∣sentence index(p1)length(d1) − sentence index(p2)length(d2) ∣∣∣) (17)
Context Similarity. We further consider occurrences of (shared) predicates in the immediate discourse context of two predicates. Our measure for this type of similarity is
computed as the relative number of overlapping predicate types within the preceding and succeeding n neighboring predicate instances as defined in Equation (18).
simDCon(p1, p2) = context(p1 )∩context(p2 ) context(p1 )∪context(p2 ) , with
context(p) = {p′|index(p′) ∈ [index(p)− n : index(p) + n]} (18)
We compute the index of a predicate as the number of preceding predicates within the same text, that is, starting with zero. Based on preliminary experiments on our development set, we empirically set n to 5.
4.2.3 Alignment via Clustering. The goal of this step of our induction framework is to identify pairs of PASs that describe the same event, state, or entity. As a means to achieve this goal, we represent pairs of comparable texts as graphs and aim to find those edges in a graph that represent connections between predicates that need to be aligned. Although our aim is to find edges between nodes, we note that the majority of predicates (nodes) in our data set are not aligned and hence a crucial prerequisite to generate precise alignments is to filter out those nodes that are unlikely to be good alignment candidates. To achieve the filtering and alignment goals at the same time, we rely on graph clustering techniques that have successfully been applied in the NLP literature (Su and Markert 2009; Cai and Strube 2010; Chen and Ji 2010, inter alia) and that can be used to partition a graph into singleton nodes and smaller subgraphs.
The clustering method applied in our model relies on so-called minimum cuts (henceforth also called mincuts) in order to partition a bipartite graph, representing pairs of texts, into clusters of alignable predicate–argument structures. A mincut operation divides a given graph into two disjoint subgraphs. Each cut is performed between some source node s and some target node t, such that (1) each of the two nodes will be in a different subgraph and (2) the sum of weights of all removed edges will be as small as possible. We implement basic graph operations using the freely available Java library JGraphT7 and determine each mincut using the method described in Goldberg and Tarhan (1986).
Given a constructed input graph G, our algorithm recursively applies mincuts in three steps as follows:
1. Identify edge e with the lowest weight in the current (sub)graph G.
2. Perform a mincut such that the nodes connected by e will be in two different subgraphs G′ and G′′.
3. Recursively apply Steps 1 and 2 to subgraphs G′ and G′′.
As our goal is to induce clusters that correspond to pairs of corresponding structures, we apply Step 3 of the clustering approach outlined above only to subgraphs that contain more than two nodes. Algorithm 1 provides a pseudocode implementation of this procedure. An example of an input graph and the applied minimum cuts is illustrated in Figure 1. As shown in the illustration, we only use edges in our initial graph representation that represent alignment candidates with a similarity above a
7 http://jgrapht.org/.
Algorithm 1. Pseudo code of our clustering algorithm. function CLUSTER(G)
clusters← ∅ E← GETEDGES(G) . Step 1 e← GETEDGEWITHLOWESTWEIGHT(E) s← GETSOURCENODE(e) t← GETTARGETNODE(e) G′ ← MINCUT(G, s, t) . Step 2 C ← GETCONNECTEDCOMPONENTS(G′) for all Gs ∈ C do . Step 3
if SIZE(Gs) <= 2 then clusters← clusters ∪ Gs else clusters← clusters ∪ CLUSTER(Gs)
end if end for return clusters;
end function
threshold determined on the development part of our manually aligned data. Based on the initial representation, the first cut (Cut 1) is performed between the nodes connected by the edge with the lowest weight in the overall graph (13.0). This cut separates the nodes representing the “earnings NNS” predicates from the rest of the graph. Similarly, Cut 2 separates another cluster of two nodes. Finally, Cuts 3 and 4 remove a single node from the only remaining cluster that had more than two nodes.
The main benefit of our method compared with off-the-shelf clustering techniques is that we can define the termination criterion in line with the goal of our task, namely,
to align pairs of structures across comparable texts, instead of having to optimize additional parameters that require careful fine-tuning (such as the number of clusters or a clustering threshold). In the next section, we provide empirical evidence for the advantage of this approach. This section describes the evaluation of our graph-based clustering model on the task of aligning predicate–argument structures across comparable texts. We define two variants of our graph-based model: Full makes use of all similarity measures described in Section 4.2.2, and EMNLP’12, the model introduced in Roth and Frank (2012b), only uses the predicate-specific measures and the bag-of-words measure for argument similarity. Both models, Full and EMNLP’12, make use of the clustering algorithm introduced in Section 4.2.3.
For evaluation, we compare our model against two baselines and a model from the literature that has recently been proposed for this task (Wolfe et al. 2013). Similar to our approach, Wolfe et al. use various resources to calculate the similarity of two predicate– argument structures. Differences to our model lie in the kind of utilized resources, the use of additional data to learn feature weights, and the fact that each alignment decision is made using a binary classifier. We evaluate the predictions of each model on the manually annotated test set described in Section 4.1.2.8
As evaluation measures, we use precision, recall, and F1-score. Following previous work aligning monolingual texts (Cohn, Callison-Burch, and Lapata 2008), we measure precision as the number of predicted alignments also annotated in the gold standard divided by the total number of predictions, and recall as the number of correctly predicted sure alignments divided by the total number of sure alignments in the gold standard. F1-score is computed as as the harmonic mean between precision and recall.
4.3.1 Baselines. A simple baseline for predicate alignment is to simply cluster all predicates that have identical lemmata (henceforth called LemmaId). To assess the benefits of the clustering step, we propose a second baseline that uses the same similarity measures as our Full model but does not use the mincut clustering described in Section 4.2.3. Instead, it greedily merges as many 1-to-1 alignments as possible, starting with the highest similarity (Greedy). As a more sophisticated baseline, we make use of alignment tools commonly used in statistical machine translation. We train our own word alignment model using the state-of-the-art word alignment tool Berkeley Aligner (Liang, Taskar, and Klein 2006). As word alignment tools require pairs of sentences as input, we first extract paraphrases using a re-implementation of a previously proposed paraphrase detection system based on lemma and n-gram overlap (Wan et al. 2006). In the following section, we abbreviate the alignment based model as WordAlign.
4.3.2 Results. The results for the alignment task are presented in Table 2. From all approaches, Greedy and WordAlign yield the lowest performance. For WordAlign, we observe two main reasons. On the one hand, sentence paraphrase detection does not perform perfectly. Hence, the extracted sentence pairs do not always contain gold
8 We also performed an evaluation on sentence-level predicate alignment, but skipped the discussion here as this task is not relevant for our induction framework. As the additional discourse-level measures of the Full model are not needed for alignment within sentences, we refer the interested reader to the EMNLP’12 model and its evaluation in Roth and Frank (2012b).
alignments. On the other hand, even sentence pairs that contain gold alignments can differ considerably in content and length, making them hard to align for statistical word alignment tools. The difficulty of this task can also be seen in the results for the Greedy model, which only achieves an F1-score of 17.2%. In contrast, we observe that the majority of all sure alignments can be retrieved by applying the LemmaId model (60.3% recall).
The Full model achieves a recall of 48.9% but significantly outperforms all baselines (p < 0.01) in terms of precision (71.8%). This is an important factor for us as we plan to use the alignments in subsequent steps of our framework. With 58.2%, Full also achieves the best overall F1-score. By comparing the results with those of the EMNLP’12 model, we can see that the discourse-level similarity measures provide a significant improvement in terms of precision without a considerable loss in recall. This advantage in precision can also be seen in comparison to Wolfe et al. In contrast, their system outperforms our model with respect to recall. There are two main reasons for this: On the one hand, their model makes use of much larger resources to compute alignments, including a paraphrasing database that contains over 7 million rewriting rules; on the other hand, their model is supervised and makes use of additional data to learn weights for each of their features. In contrast, Full and EMNLP’12 only make use of a small development data set to determine a threshold for graph construction. Though the difference is not significant, it is worth noting that our model outperforms that by Wolfe et al. by 0.6 percentage points in F1-score, despite not making use of any additional training data.
Ablating Similarity Measures. All aforementioned results were conducted in experiments with a uniform weighting scheme of similarity measures as introduced in Section 4.2.2. Table 2 additionally shows the performance impact of individual similarity measures by removing them completely (i.e., setting their weight to 0.0). The numbers indicate that almost all measures contribute positively to the overall performance when using equal weights. Except for the argument head similarity, all ablation tests revealed significant
drops in performance, either with respect to precision or recall. This result highlights the importance of incorporating predicate-specific, argument-specific, and discoursespecific information regarding individual predications in this task.
Tuning Weights for High Precision. Subsequently, we tested various combinations of weights on our development set in order to estimate a better weighting scheme. This tuning procedure is implemented as a grid search technique, in which random weights between 0.0 and 1.0 are assigned to each measure. For graph construction, all weights are normalized again to sum to 1.0. We additionally try different thresholds for adding edges in the graph representation. To achieve high precision, we weight precision three times higher than recall while evaluating different parameters. In total, we tested 2,000 different parameter assignments on our development set. Following this process, we found the best result to be achieved with a threshold of 0.85 and the following weights:r 0.11, 0.14, and 0.21 for simWN, simVN, and simDist, respectively,
(i.e., 46% of the total weight for predicate-specific measures)r 0.21 and 0.05 for simABoW and simAheads, respectively, (i.e., 26% of the total weight for argument-specific measures)r 0.21 and 0.07 for simDPos and simDCon, respectively (i.e., 28% of the total weight for discourse-specific measures)
The weighting scheme shows that information from all categories is considered. When applying the tuned model on our evaluation data set, we note that results in recall drop to 29.1% (−19.8 percentage points). Precision, on the other hand, increases to 86.2% (+14.4 percentage points). In this section, we introduced the task of aligning predicate–argument structures across monolingual comparable texts. We designed annotation guidelines and created a data set of gold-standard alignments. Based on this data set, we developed and evaluated a novel clustering-based alignment model that uses a combination of various similarity measures and a graph-based clustering algorithm that we specifically designed for this task. In an intrinsic evaluation, we showed that our novel model outperforms a range of baselines as well as previous approaches to this particular task. As an additional contribution, we defined a tuning routine that can be utilized to train a high precision model for the discourse-level alignment task. Our results show that, by using this tuning step, corresponding structures in our evaluation set can be identified with a precision of 86.2%. This intermediate result is essential for the success of our overall framework. In the next section, we present Step 2 of our implicit argument induction technique, in which we examine pairs of automatically aligned predicate–argument structures as a means to identify and link implicit arguments.","5 induction framework step 2: extracting implicit arguments :In the second step of our induction framework, we rely on alignments between PAS to detect implicit arguments. That is, we aim to identify argument instances that are present, or explicit, in one PAS but absent, or implicit, from the aligned PAS. Based on the identified instances, our model tries to find antecedents of implicit arguments
within their respective discourse contexts. To perform this task automatically, we rely on several preprocessing tools, which we apply on the full corpus of over 160,000 document pairs introduced in Roth and Frank (2012a).
We describe the applied preparatory steps and the used preprocessing tools in Section 5.1. Using automatically computed annotations as input, we describe a heuristic method to detect implicit arguments and their discourse antecedents in Section 5.2. As a basis for the actual induction, we rely on several preparatory steps that identify information two documents have in common (cf. Figure 2). In particular, we compute and align PAS using the graph-based model described in Section 4, and determine co-referring entities across pairs of texts using coreference resolution techniques on concatenated document pairs (Lee et al. 2012). In theory, arguments implicit in one structure can straightforwardly be induced based on this information by looking for co-referring mentions of the argument explicit in the aligned structure. In practice, we make use of additional checks and filters to ensure that only reliable information is being used. We describe the preprocessing steps in the following paragraphs and provide additional details on our implementation of the induction procedure in Section 5.2.
Single Document Preprocessing. We apply several preprocessing steps to each document in our data set. First, we use the Stanford CoreNLP package (Manning et al. 2014) for tokenization and sentence splitting. We then apply the MATE tools (Björkelund et al. 2010; Bohnet 2010), including the integrated PropBank/NomBank-based semantic parser, to determine local PAS. Finally, we resolve pronouns that occur in a PAS using the coreference resolution system by Martschat et al. (2012), which placed second for English in the CoNLL-2012 Shared Task (Pradhan et al. 2012).
High Precision Alignments. Once all single documents are preprocessed, we align PAS across pairs of comparable texts. We want to induce reliable instances of implicit arguments based on aligned PASs pairs and hence apply our graph-based clustering technique using the high-precision tuning step described in Section 4.3. We run
the high-precision model on all pairs of texts in the corpus. As a result, we extract a total number of 283,588 aligned pairs of PASs. An overview of properties of this data set is given in Table 3. We observe that most alignments involve the same part-of-speech (98.4%) and the same predicate lemma (96.6%). This fact simplifies the task of inducing implicit arguments as it implies that in most cases the PropBank argument labels in a pair of aligned structures correspond to each other and do not have to be mapped via a higher-level set of roles.
Cross-Document Coreference. For each argument that is explicit in one PAS but implicit in an aligned PAS, we want to determine a suitable antecedent within the discourse context of the implicit instance. We solve this task by viewing the aligned explicit argument as an entity reference and identify co-referring mentions in both texts by applying coreference resolution techniques across pairs of documents. In practice, we follow the methodology of Lee et al. (2012), who propose applying standard coreference methods on pairs of texts by simply concatenating two documents and providing them as a single input document. As merging two different discourses can lead to problems for a coreference system that is based on features and feature weights trained on single documents, we follow Lee et al. and apply a rule-based system. Like them, we use the Stanford Coreference system (Lee et al. 2013), which applies a sequence of coreference “sieves” to the input, ordered according to their precision. To obtain a highly accurate and reliable output, we consider only the most precise resolution sieves:r “String Match”r “Relaxed String Match”r “Precise Constructs”r “Strict Head Match A,” “Strict Head Match B,” “Strict Head Match C”r “Proper Head Noun Match” Note that none of these sieves involve pronoun resolution. Instead, we decided to use the resolved pronouns from the single-document coreference step. This decision
is based on the fact that the system by Martschat et al. (2012) outperforms the Stanford system with all sieves on the CoNLL’11 test set by an average F1-score of 3.0 absolute points. The high-precision sieves are, however, better suited for the cross-document task as we plan to rely on the resulting coreference chains for identifying potential antecedents of implicit arguments. That is, we prefer fewer but more reliable chains in order to minimize the impact of possible preprocessing errors. Given a pair of aligned predicates from two comparable texts, we examine the output of the semantic parser (cf. Section 5.1) to identify arguments in each PAS. We compare the set of labels assigned to the arguments in each structure to determine whether one PAS contains an argument (explicit) that has not been realized in the other PAS (implicit). For each implicit argument, we identify appropriate antecedents by considering the cross-document coreference chain of its explicit counterpart. As our goal is to link implicit arguments within discourse, we require candidate antecedents to be mentions that occur in the same document. We impose a number of restrictions on the resulting pairs of implicit arguments and antecedents to reduce the impact of different types of preprocessing errors:
Mislabeled Arguments. In some cases, the parser annotated the same argument in two texts using different labels. To ensure that mislabeled arguments are not recognized as implicit, we require that pairs of aligned PAS contain a different number of arguments.
Missed Arguments. Depending on sentence structure, the semantic parser is sometimes unable to determine all local arguments. This often leads to the identification of erroneous implicit arguments. To intercept some of these cases, we require that all antecedents from the cross-document coreference chain must be outside of the sentence that contains the affected PAS. We apply the outlined identification and linking approach to all text pairs in our corpus of comparable texts. As a result, we induce a total of 698 implicit argument and antecedent pairs. A summary of properties of the obtained pairs can be found in Table 4. The full data set involves predicates of 294 different lemmata. Each pair was found in a separate document. Note that 698 implicit arguments from 283,588 pairs of PAS seem to represent a fairly low recall. The reason for this is that most PAS pairs in the highprecision data set consist of identically labeled argument sets (84.5%) and in most of the remaining cases an antecedent in discourse cannot be found using the high-precision coreference sieves. This does not mean that implicit arguments are rare in general. As discussed in Section 4.1, 38.9% of all manually aligned PAS pairs involve a different number of arguments.
We manually evaluated a subset of 90 induced implicit arguments and found 80 discourse antecedents to be correct (89%). Examples are displayed in Table 5. A closer analysis revealed that some incorrectly linked instances still result from preprocessing
errors. In particular, combinations of errors can lead to incorrectly identified instances as showcased in Example 19:
Example 19 “The Guatemalan Congress on Thursday ratified 126-12 [a Central America-US]A0 [free trade]A1 agreement, lawmakers said.” Induced missing argument and discourse antecedent: [goods]A2/co-agent
Instead of recognizing Central America and US as two separate arguments (agent and co-agent) of the predicate AGREE, the semantic parser labels both entities as one argument (A0, agent); our system hence tries to determine a discourse antecedent for an argument that is predicted to be missing despite being actually realized (A2, co-agent). In the aligned PAS, the co-agent is realized as a prepositional phrase: “[with the United States]A2”. The cross-document coreference tool incorrectly predicts the United States in this phrase to be coreferent with the phrase U.S. goods and services; our system hence detects goods as the antecedent for the erroneously predicted implicit argument.
Further error sources are incorrectly extracted document pairs and alignments between PAS that do not correspond to each other. Example 20 shows two PAS from a pair of texts that describe different sets of changes in economic activity.
Example 20 “[Production]A1 rose [3.9 percent from October in the capital good sector]A2” “[Industrial production excluding energy, food and construction . . . ]A1 rose [2.6 percent]A2 [in November]TMP from the previous month”
Example 21 shows a pair of aligned structures in a pair of comparable texts that both report on two planned trips by President Obama. Here, the alignment model erroneously aligned a reference to one trip with a reference to both trips.
Example 21 “The postponement ... marks the second time [the president’s]A0 trip [to Indonesia]A1 has been postponed” “Obama was to depart on a [weeklong]TMP trip [to both countries . . . ]A1 on June 13” In this section, we introduced a computational implementation of Step 2 of our rulebased induction method for identifying implicit arguments and their discourse antecedents. Our approach depends on automatic annotations from semantic role labeling, PAS alignment, and coreference resolution. We implement two particular types of measures to minimize the impact of preprocessing errors: (1) we avoid imprecise input by applying high-precision tools instead of methods that are tuned for balanced precision and recall; and (2) we circumvent some common preprocessing errors by formulating
constraints on resulting instances of implicit arguments and discourse antecedents. An intrinsic analysis reveals that we cannot eliminate all error sources this way, although we found the induced data set to be of high precision.
The next sections present extrinsic evaluations in which we demonstrate the utility of the induced instances empirically. As discussed in Section 1, we assume that a proper treatment of implicit arguments can improve the coverage of semantic role labeling systems and entity-based models of coherence. We show how our data set can be utilized in both of these areas: In Section 6, we use our data set of implicit arguments to enhance the training process for models of implicit argument linking; in Sections 7, we demonstrate that implicit arguments can affect the perceived coherence of a text and that instances of aligned (explicit and implicit) arguments in our data set can be used to successfully predict this impact adequately.","6 task setting 1: linking implicit arguments in discourse :Our first extrinsic evaluation assesses the utility of automatically induced pairs of implicit arguments and antecedents for the task of implicit argument linking. For this scenario, we use the data sets from the SemEval 2010 task on “Linking Events and their Participants in Discourse” (Ruppenhofer et al. 2010, henceforth just SemEval task). For direct comparison with previous results and heuristic acquisition techniques, we apply the implicit argument identification and linking model by Silberer and Frank (2012, henceforth S&F) for training and testing. We briefly describe the SemEval task data and the model by S & F in the next sections. Both the training and test sets of the SemEval task are text corpora extracted from Sherlock Holmes novels, with manual frame semantic annotations including implicit arguments. In the actual linking task (“NI-only”), gold labels are provided for local arguments and participating systems have to perform the following three sub-tasks: firstly, non-instantiated roles have to be identified; secondly, they have to be classified as being “accessible to the speaker and hearer” (definite null instantiation, DNI) or as being “only existentially bound within discourse” (indefinite null instantiations, INI); and finally, all resolvable null instantiations have to be linked to discourse antecedents.
The task organizers provide two versions of their data set: one based on FrameNet annotations and one based on PropBank/NomBank annotations. We found, however, that the latter only contains a subset of the implicit argument annotations from the FrameNet-based version. As all previous results in this task have been reported on the FrameNet data set, we adopt the same setting. Note that our automatically induced data set, which we want to apply as additional training data, is labeled with a PropBank/NomBank-style parser. Consequently, we need to map our annotations to FrameNet in order to make use of them in this task. The organizers of the SemEval task provide a manual mapping dictionary for predicates in the annotated data set. We use this manual mapping and additionally use SemLink (Palmer 2009) for mapping predicates and arguments not covered by the dictionary. We make use of the system by S&F to train a new model for the NI-only task. As mentioned in the previous subsection, this task consists of three steps: In Step (1), implicit
arguments are identified as unfilled and non-redundant FrameNet core roles; in Step (2), an SVM classifier is used to predict whether implicit arguments are definite based on a small number of features—semantic type of the affected Frame Element, the relative frequency of its realization type in the SemEval training corpus, and a Boolean feature that indicates whether the affected sentence is in passive voice and does not contain a (deep) subject. In Step (3), we apply the same features and classifier as S&F to find appropriate antecedents for (predicted) definite arguments. S&F report that their best results were obtained considering the following target antecedents: all entity mentions that are syntactic constituents from the present and the past two sentences and all entity mentions that occurred at least five times in the previous discourse (“Chains+Win” setting). In their evaluation, the latter of these two restrictions crucially depended on gold coreference chains. As the automatic coreference chains in our data are rather sparse (and noisy), we only consider syntactic constituents from the present and the past two sentences as antecedents (“SentWin” setting).
Before training and testing a new model with our own data, we perform feature selection using 10-fold cross validation. To find the best set of features, we run the feature selection on a combination of the SemEval training data and our full additional data set.9 The only features that were selected in this process concern the “prominence” of the candidate antecedent, its semantic agreement with the selectional preferences of the predicate, the part-of-speech-tags used in each reference to the candidate entity, and the semantic types of all roles that the entity fills according to local role annotations. These features are a subset of the best features described in Silberer and Frank (2012). Evaluation measures. For direct comparison in the full task, both with S&F’s model and other models, we adopt the precision, recall, and F1 measures as defined in Ruppenhofer et al. (2010).
Baselines. We compare our results with those previously reported on the SemEval task (see Table 6 for a summary): The best performing system in the actual task in 2010 was developed by Chen et al. (2010) and is an adaptation of the semantic role labeling system SEMAFOR (Das et al. 2010). In 2011, Tonelli and Delmonte (2011) presented a revised version of their SemEval system (Tonelli and Delmonte 2010), which outperforms SEMAFOR in terms of recall (6%) and F1-score (8%). The best results in terms of recall and F1-score to date have been reported by Laparra and Rigau (2012), with 25% and 19%, respectively. Our model outperforms their state-of-the-art system in terms of precision (21%) but achieves a lower recall (8%). Two influencing factors for their high recall are probably (1) their improved method for identifying (resolvable) implicit arguments, and (2) their addition of lexicalized and ontological features.
Our Results. Comparison of our results with those reported by S&F, whose system we use, shows that our additional data improves precision (from 6% to 21%) and F1-score (from 7% to 12%). The loss of one percentage point in recall is marginal given the size of the test set (only 259 implicit arguments have an annotated antecedent). Our
9 Note that this feature selection procedure is the same as applied by Silberer and Frank. Hence, the evaluated models are directly comparable and all differences in results can directly be traced back to the use of additional data.
result in precision is the second highest score reported on this task. Interestingly, the improvements are higher than those achieved in the original study by Silberer and Frank (2012), even though their best additional training set is three times bigger than ours and contains manual semantic annotations.
We conjecture that their low gain in precision could be a side effect triggered by two different factors. Firstly, the heuristically created training instances, induced by treating anaphoric pronouns as implicit argument instances, might not reflect the same properties as actual implicit arguments. For example, pronouns can occur in syntactic constructions in which an actual argument cannot be omitted in practice, leading an incorrect overgeneralization. An additional negative factor might be that their model relies on coreference chains, which are automatically generated for the test set and hence rather noisy. In contrast, our automatically induced data does not contain manual annotations of semantic roles and coreference chains, hence we do not rely on gold information during training and testing. The results show that, despite this limitation, our new model outperforms previous models trained using the same system, indicating the utility and high reliability of our automatically induced data.
Impact of training data size. To assess the impact of training data size, we perform an additional experiment with subsets of automatically induced implicit arguments. Specifically, we train different classifiers using the full SemEval training data and varying amounts of our automatically induced training data (random samples of 1%, 10%, and 25%). The model uses the best feature set determined on the combination of SemEval and our full additional data set in all settings. For each setting, we report average results obtained over four runs. The outcomes of this experiment are summarized in Table 7. The numbers reveal that using only 1% of the additional data for training already leads to a classification performance of 0.13 in F1-score. The large improvement over the S&F model without additional data, which achieves an F1-score of 0.07, can be explained by the fact that the features selected by our model generalize better than those selected on the SemEval training data only. The improvements are highest when using between 10% and 25% of the additional data, indicating that the use of additional induced instances indeed increases performance but that utilizing more out-of-domain than in-domain data for training seems to be harmful.
10 Results as reported by Tonelli and Delmonte (2011). 11 Results computed as an average over the scores given for both test files; rounded towards the number
given for the test file that contained more instances. In this section, we presented a NLP application of the automatically induced data set of implicit arguments that we introduced in Section 5. We found that automatically induced implicit arguments can successfully be used as training data to improve a system for linking implicit arguments in discourse. Although the presented model cannot compete with state-of-the-art systems, the addition of our data led to an enhanced performance compared with the same system with different and without additional training data. Compared with the model without additional training data, our induced data set increased results in terms of precision and F1-score by 15 and 5 percentage points, respectively.","7 task setting 2: modeling local coherence :In our second experiment, we examine whether argument realization decisions affect the perceived coherence of a text and investigate how and which factors related to their impact can be used to model realization decisions computationally. We approach this question in the following way: We exploit PAS alignments across comparable documents to identify contexts with implicit and explicit arguments; we then make use of these automatically induced contexts in order to train a discourse coherence model that is able to predict whether—in a given context—an argument should be realized or remain implicit.
Induction of such a model and its evaluation will be approached as follows: First, we assemble a data set of document pairs that differ only with respect to a single realization decision; given each pair in this data set, we ask human annotators to indicate their preference for the implicit or explicit argument instance in the prespecified context (Section 7.1); secondly, we attempt to emulate the decision process computationally using a discriminative model based on discourse and entity-specific features (Section 7.2). To assess the performance of the new model, we train it on automatically induced training data and evaluate it, in comparison with previous models of local coherence, against human annotations (Section 7.3). We use the data set of automatically induced implicit arguments (henceforth source data), described in Section 5, as a starting point for composing a set of document pairs that involve implicit and explicit arguments. To make sure that each document pair in this data set only differs with respect to a single realization decision, we first create
two copies of each document from the source data: One copy remains in its original form, and the other copy will be modified with respect to a single argument realization. Example (22) illustrates an original and modified sentence.
Example 22 (a) [The Dalai Lama’s]A0 visit [to France]A1 ends on Tuesday.
(a‘) [The Dalai Lama’s]A0 visit ends on Tuesday.
Note that adding or removing arguments at random can lead to structures that are semantically implausible. Hence, we consider two restrictions. First, we ensure that the remaining context is still understandable after an argument is removed by only considering texts in which follow-up references can still be resolved based on earlier antecedents. Second, we only modify arguments of PAS that actually occur and are aligned across two texts. Given a pair of PAS that differ with respect to an argument realization, we create modifications by replacing the specific implicit or explicit argument in one text with the corresponding argument in the paired text. Examples (22) and (23) show two such comparable sentences. The original PAS in Example (22a) contains an explicit argument that is implicit in the aligned PAS and hence removed in the modified version. Similarly, the original text in (23a) involves an implicit argument that is made explicit in the modified version (23a‘).
Example 23 (a) [The Dalai Lama’s]A0 visit coincides with the Beijing Olympics.
(a‘) [The Dalai Lama’s]A0 visit [to France]A1 coincides with the Beijing Olympics.
We ensure that the modified structure fits into the given context grammatically by only considering pairs of PASs with identical predicate form and constituent order. We found that this restriction constrains affected arguments to be modifiers, prepositional phrases, and direct objects. We argue that this is actually a desirable property because more complicated alternations could affect coherence by themselves. In other words, resulting interplays would make it difficult to distinguish between the isolated effect of argument realization itself and other effects, triggered for example by sentence order (Gordon, Grosz, and Gilliom 1993).
Annotation. We set up a Web experiment using the evaluation toolkit by Belz and Kow (2011) to collect ratings of local coherence for implicit and explicit arguments. For this experiment, we compiled a data set of 150 document pairs. Each text in such a pair consists of the same content, with the only difference being one argument realization.
We presented all 150 pairs to two annotators12 and asked them to indicate their preference for one alternative over the other using a continuous slider scale. The annotators got to see the full texts, with the alternatives presented next to each other. To make texts easier to read and differences easier to spot, we collapsed all identical sentences into one column and highlighted the aligned predicate (in both texts) and the affected argument (in the explicit case). An example is shown in Figure 3. To avoid any bias in the annotation process, we shuffled the sequence of text pairs and randomly assigned the side of display (left/right) of each realization type (explicit/implicit). Instead of
12 Both annotators are native speakers of English and undergraduate students in literature and linguistics.
providing a definition of local coherence ourselves, we asked annotators to rate how “natural” a realization reads given the discourse context. This procedure is in line with previous work by Pitler and Nenkova (2008), who “view text readability and text coherence as equivalent properties,” given that the annotators are “competent language users.”
We found that annotators made use of the full rating scale, with the extremes indicating either a strong preference for the text on the left-hand side or the righthand side, respectively. However, most ratings were concentrated more towards the center of the scale (i.e., around zero). This seems to imply that the use of implicit or explicit arguments did not make a considerable difference most of the time. The annotators affirmed that some cases do not read naturally when a specific argument is omitted or redundantly realized at a given position in discourse. For example, the text fragment in Example (24) shows two sentences in which an argument has been realized twice, leading to a perceived redundancy in the second sentence (A4, destination); conversely, Example (25) showcases an excerpt in which a non-redundant argument (A2, co-signer) has been omitted.
Example 24 The remaining contraband was picked up at Le Havre. The containers had arrived [in Le Havre] from China.
Example 25 Lt.-Gen. Mohamed Lamari (. . . ) denied his country wanted South African weapons to fight Muslim rebels fighting the government. “We are not going to fight a flea with a hammer,” Lamari told reporters after signing the agreement of intent [∅].
We computed correlation between ratings of both annotators using Spearman’s ρ (Spearman 1904) and found a low but significant correlation (ρ = 0.22, p < 0.01). To
construct a test set with gold annotations, we mapped continuous ratings to discrete labels (implicit, explicit, neutral) and selected a subset of 70 instances for which clear13 preferences were observed. Table 8 provides some statistics on the collected data. Even though correlation with respect to continuous ratings is relatively low, we find that both annotators have the same general preference in most cases, with a raw agreement of 75% overall. Our test set contains 47 (31%) cases of explicit arguments and 23 (15%) cases of implicit arguments. Those cases in which annotators disagreed (25%) or had no preference (28%) were discarded. We model the decision process that underlies the (non-)realization of arguments using an SVM-based model and a range of discourse features. The features are based on the following three factors: the affected predicate–argument structure (Parg), the (automatic) coreference chain of the affected entity (Coref), and the discourse context (Disc).
Parg. The first group of features is concerned with the characteristics of the affected PAS: This includes the absolute and relative number of explicitly realized arguments in the structure, the number of modifiers in it, and the total length of the structure as well as of the complete sentence (in words).
Coref. The coreference-specific features include transition patterns as inspired by the entity grid model (cf. Section 3), the absolute number of previous and follow-up mentions of the (non-)realized argument, their POS tags, and the distance between the current PAS to the closest previous and follow-up mention (in number of words and sentences). In contrast to previous work on the entity grid model, we do not type transition features with respect to the grammatical function of explicit realizations. The reason for skipping this information lies in the insignificant amount of relevant samples in our training data (see the following).
Disc. On the discourse level, we define a small set of additional features that include the total number of coreference chains in the text, the occurrence of pronouns in the current sentence, lexical repetitions in the previous and follow-up sentence, the current position in discourse (begin, middle, end), and a feature indicating whether the affected argument occurred in the first sentence.
13 Absolute continuous rating of >1 standard deviation from zero.
Most of these features overlap with those successfully applied in previous work on modeling coherence. For example, the transition patterns are inspired by the entity grid model. In addition to entity-grid like features, Pitler and Nenkova (2008) also use text length, word overlap, and pronoun occurrences as features for predicting readability. Our own contribution lies in the definition of PAS-specific features and the adaptation of all features to the task of predicting (non-)realization of arguments in a PAS. In the evaluation (cf. Section 7.3), we report results for two models: A simplified model that only makes use of entity-grid like features and a full model that uses all features described here. To learn feature weights, we make use of the training data described in the following section. The goal of our model is to correctly predict the realization type (implicit or explicit) of an argument that maximizes the perceived coherence of the document. As a proxy for coherence, we use the naturalness ratings given by our annotators. We evaluate classification performance on the 70 data points in our annotated test set for which clear preferences have been established. We report results in terms of precision, recall, and F1-score per class as well as micro- and macro-averaged F1-score across classes. We compute precision as the fraction of correct classifier decisions divided by the total number of classifications made for a specific class label; recall as the fraction of correct classifier decisions divided by the total number of test items with the specific label; and F1 as the harmonic mean between precision and recall.
7.3.1 Baselines. For comparison, we apply a couple of coherence models proposed in previous work: the original entity grid model by Barzilay and Lapata (2005), a modified version that uses topic models (Elsner and Charniak 2011a), and an extended version that includes entity-specific features (Elsner and Charniak 2011b); we further apply the discourse-new model by Elsner and Charniak (2008), and the pronoun-based model by Charniak and Elsner (2009). For all of the aforementioned models, we use their respective implementation provided with the Brown Coherence Toolkit.14 Note that the toolkit only returns one coherence score for each document. To use the model scores for predicting argument realization, we use two documents per data point—one that contains the affected argument explicitly and one that does not (implicit argument)— and treat the higher scoring variant as classification output. If both documents achieve the same score, we count the test item neither as correctly nor as incorrectly classified.
7.3.2 Our Models. Like the applied baseline models, our models do not make use of any manually labeled data for training. Instead, we utilize the automatically identified instances of explicit and implicit arguments from pairs of comparable texts, which we described in Section 5. To train our own model, we prepare this data set as follows: Firstly, we remove all data points that were selected for the test set; secondly, we split all pairs of texts into two groups—texts that contain a PAS in which an implicit argument has been identified (IA), and their comparable counterparts, which contain the aligned PAS with an explicit argument (EA). All texts are labeled according to their group. For all texts in group EA, we remove the explicit argument from the aligned PAS. This way, the feature extractor always gets to see the text and automatic annotations as if
14 http://www.ling.ohio-state.edu/%7Emelsner/.
the realization decision had not been performed and can thus extract unbiased feature values for the affected entity and argument position. Given each feature representation, we train a classifier using the default parameters of the LIBSVM package (Chang and Lin 2011).15
We apply our own model on each data point in the small annotated test set, where we always treat the affected argument, regardless of its actual annotation, as implicit to extract unbiased feature values for classification. Based on the features described in Section 7.2 and trained on the automatically constructed PAS alignments, our model predicts the realization type of each argument in the given context. We note that our model has an advantage here because it is specifically designed for this task and trained on corresponding data. All models compute local coherence ratings based on entity occurrences, however, and should thus be able to predict which realization type coheres best with the given discourse context. That is, because the input document pairs are identical except for the affected argument position, the coherence scores assigned by each model to pairs of text only differ with respect to the affected entity realization.
7.3.3 Results. The results are summarized in Table 9. We begin our analysis with a discussion on results in terms of micro-averaged F1-scores. As observable from the last column in Table 9, each of the baseline models achieves an F1-score between 31% and 41%, whereas our (full) model achieves 74%. As our data set is imbalanced and biased towards explicit realizations, we also provide macro-averaged F1-scores that show average performance over both classes without taking into account their respective sizes. Here, we find that baseline results lie between 27% and 39%, whereas our (full) model achieves 69%. Before discussing the results of our model in more detail, we investigate baselines performances.
We observe that the original entity grid model exhibits a preference for the implicit realization type: It predicts this class in 61 (87%) cases, resulting in only 9% of all explicit arguments being correctly classified. Overall, the entity grid achieves a
15 The default settings in our LIBSVM version are: equal costs of both classes and use of a sigmoid kernel.
(micro-averaged) F1-score of 31%. Taking a closer look at the features of the model reveals that this is an expected outcome: In its original setting, the entity grid learns realization patterns in the form of sentence-to-sentence transitions. As discussed in the paper that introduced the entity grid model (Barzilay and Lapata 2008), most entities are, however, only mentioned a few times in a text. Consequently, non-realizations constitute the “most frequent” class, independently of whether they are relevant in a given context or not. The models by Charniak and Elsner (2009) and Elsner and Charniak (2011a), which are not based on an entity grid, suffer less from such an effect and achieve better results, with F1-scores up to 35% and 41%, respectively. The extended entity grid model also alleviates the bias towards non-realizations, resulting in a slightly improved F1-score of 33%. To abstract away from this issue, we train a simplified version of our own model that only uses features that involve entity transition patterns. The main difference between this simplified model and the original entity grid model lies in the different use of training data: whereas entity grid models treat all non-realized items equally, our model gets to “see” actual examples of entities that are implicit. In other words, our simplified model takes into account implicit mentions of entities, not only explicit ones. The results confirm that this extra information has a significant impact (p < 0.01, using a randomization test; Yeh 2000) on test set performance, and raises the ratio of correctly classified explicit arguments to 73%. Yet, the simplified model only provides a correct label for 30% of instances in the test data that are marked as implicit arguments, with a class-specific F1-score of only 42%. As demonstrated by the performance of our full model, a combination of all features is needed to achieve the best overall results of 69% and 74% in macro and micro-averaged F1-scores, respectively. Applied to the two classes separately, our model achieves an F1-score of 53% on arguments that are annotated as implicit and 82% on explicit arguments. Both of these scores are the best across all tested models.
To determine the impact of the three different feature groups, we derive the weight of each feature from the model learned by LIBSVM. Table 10 gives an overview over the ten highest weights for implicit and explicit realization classification. We use the following terminology in the feature description: “the entity” refers to the entity that is referred to by the to-be-classified argument, “next/previous mention” denotes a co-referring mention to the same entity, “the PAS” refers to the predicate–argument structure that contains the affected argument (implicitly), and “the sentence” refers to the sentence in which this PAS is realized. All “distances” refer to the number of tokens that appear between the predicate that heads the PAS and the previous or next mention of the entity. As can be seen at the top and bottom ends of Table 10, the strongest feature for classifying an argument as implicit is whether the entity is also realized in the preceding or following two sentences. The strongest feature for classifying an argument as explicit is whether the next mention is a pronoun. The trained weights indicate that the model is learning some interesting patterns that reflect rules such as “avoid close repetitions,” “keep sentences short,” and “pronouns and proper names can more often be dropped than definite noun phrases.” In this chapter, we presented a computational linguistic application of the automatically induced data set of implicit arguments that we introduced in Section 5. This data set has been induced from pairs of comparable text and is a unique resource in that it contains automatic annotations of implicit arguments, aligned explicit arguments, and discourse antecedents. In our experiments on perceived coherence, we found that the
use of implicit vs. explicit arguments, although often being a subtle difference, can have a clear impact on readability ratings by human annotators. We showed that our novel coherence model, which is solely trained on automatically induced data, is able to predict this difference in newswire articles with an F1-score of up to 74%.","8 discussion and conclusions :In this article, we introduced a framework for inducing instances of implicit arguments and their discourse antecedents from pairs of comparable texts, and showcased applications of these instances in natural language processing. As described in Section 2, our framework consists of two steps: aligning predicate–argument structures across comparable texts and identifying implicit arguments and antecedents. In the following paragraphs, we summarize our framework and highlight specific contributions.
Predicate–Argument Structure Alignments. With the goal of inducing instances of implicit arguments, we proposed a novel task that aims to align pairs of PASs across pairs of comparable texts. In Section 4, we introduced a manually annotated data set for the development and evaluation of models for this particular task. We found that pairs of PASs can be aligned across documents with good inter-annotator agreement given appropriate annotation guidelines. Based on the development part of our corpus, we designed and fine-tuned a novel graph-based clustering model. To apply this model, we represent PAS in pairs of documents as bipartite graphs and recursively divide this graph into subgraphs. All clustering decisions by the model are based on pairwise similarities between PASs, combining information on predicates, associated arguments,
and their respective discourse contexts. We empirically evaluated our model against various baselines and a competitive model that has recently been proposed in the literature. The results of our evaluation show that our model outperforms all other current models on the PAS alignment task by a margin of at least 0.6 percentage points in F1-score, despite only a single threshold parameter being adjusted on our development set. As an additional contribution, we defined a tuning procedure, in which we adjust our method for high precision. Following this tuning routine, our model is capable of aligning PAS pairs with a precision of 86.2%.
Heuristic Induction Method. Based on aligned pairs of PASs, the second step in our induction framework is to identify instances of implicit arguments. In Section 5, we described a computational implementation of this step in which aligned argument structures are automatically compared and discourse antecedents for implicit arguments are found by means of entity coreference chains across documents. To reduce the effect of preprocessing errors, our implementation makes use of precise preprocessing methods and a small set of restrictions that exclude instances whose automatic annotations are likely to be erroneous. We found that by combining information from different preprocessing modules, we can induce instances of implicit arguments and discourse antecedents with high precision.
Coherence Modeling and Implicit Argument Linking. To examine the utility and reliability of our data set, we additionally performed extrinsic evaluations in task-based settings. We described two particular applications of our data: linking implicit arguments to discourse antecedents and predicting coherent argument realizations. In the first application, described in Section 6, we used our data set as additional training data to enhance a pre-existing system for identifying and linking implicit arguments in discourse. Experimental results showed that the addition of our training data can improve performance in terms of precision (+15 percentage points) and F1-score (+5 percentage points).
For the second application, we developed a novel model of local coherence, described in Section 7, which predicts whether a specific argument should be explicitly realized at a given point in discourse or whether it can already be inferred from context. Our experiments revealed that this model, when trained on automatically induced data, can predict human judgments on argument realization in newswire text with F1-scores between 53% and 82%. In comparison, we found that entity-based coherence models from previous work only achieve results below 50%, reflecting the fact that they do not capture this phenomenon appropriately.
In conclusion, a considerable amount of work still needs to be done to enhance models for handling implicit arguments in discourse. In the long run, however, this research direction will be beneficial for many applications that involve the understanding or generation of text beyond the sentence level. In this article, we provided several research contributions that form a reliable basis for future work. In particular, we developed a framework for automatically inducing instances of implicit arguments, and we designed a novel coherence model that predicts the effect of argument realizations on perceived textual coherence. From a theoretical perspective, we validated that both explicit and implicit arguments can affect coherence and that automatically induced training data can be utilized to model this phenomenon appropriately. We further showed that our induced data set, which contains instances of implicit arguments and discourse antecedents, can be applied to enhance current models for implicit argument
linking. Future work will be able to build on these insights, further enhance existing models, and apply them to improve current state-of-the-art NLP systems.
The resources described in this article are available for download at http:// projects.cl.uni-heidelberg.de/india/.","In this article, we investigate aspects of sentential meaning that are not expressed in local predicate–argument structures. In particular, we examine instances of semantic arguments that are only inferable from discourse context. The goal of this work is to automatically acquire and process such instances, which we also refer to as implicit arguments, to improve computational models of language. As contributions towards this goal, we establish an effective framework for the difficult task of inducing implicit arguments and their antecedents in discourse and empirically demonstrate the importance of modeling this phenomenon in discourse-level tasks. Our framework builds upon a novel projection approach that allows for the accurate detection of implicit arguments by aligning and comparing predicate–argument structures across pairs of comparable texts. As part of this framework, we develop a graph-based model for predicate alignment that significantly outperforms previous approaches. Based on such alignments, we show that implicit argument instances can be automatically induced and applied to improve a current model of linking implicit arguments in discourse. We further validate that decisions on argument realization, although being a subtle phenomenon most of the time, can considerably affect the perceived coherence of a text. Our experiments reveal that previous models of coherence are not able to predict this impact. Consequently, we develop a novel coherence model, which learns to accurately predict argument realization based on automatically aligned pairs of implicit and explicit arguments.","[{""affiliations"": [], ""name"": ""Michael Roth""}, {""affiliations"": [], ""name"": ""Anette Frank""}]",SP:e9e9e7f46c40792a25f778c0c791118d43613b2e,"[{""authors"": [""Arnold"", ""Jennifer E.""], ""title"": ""Reference Form and Discourse Patterns"", ""venue"": ""Ph.D. thesis, Stanford University."", ""year"": 1998}, {""authors"": [""Baroni"", ""Marco"", ""Alessandro Lenci.""], ""title"": ""Distributional memory: A general framework for corpus-based semantics"", ""venue"": ""Computational Linguistics, 36(4):673\u2013721."", ""year"": 2010}, {""authors"": [""Barzilay"", ""Regina"", ""Mirella Lapata.""], ""title"": ""Modeling local coherence: An entity-based approach"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, pages 141\u2013148. Ann Arbor, MI."", ""year"": 2005}, {""authors"": [""Barzilay"", ""Regina"", ""Mirella Lapata.""], ""title"": ""Modeling local coherence: An entity-based approach"", ""venue"": ""Computational Linguistics, 34(1):1\u201334."", ""year"": 2008}, {""authors"": [""Belz"", ""Anja"", ""Eric Kow.""], ""title"": ""Discrete vs"", ""venue"": ""continuous rating scales for language evaluation in NLP. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language"", ""year"": 2011}, {""authors"": [""Bj\u00f6rkelund"", ""Anders"", ""Bernd Bohnet"", ""Love Hafdell"", ""Pierre Nugues.""], ""title"": ""A high-performance syntactic and semantic dependency parser"", ""venue"": ""COLING 2010: Demonstration Volume, pages 33\u201336, Beijing."", ""year"": 2010}, {""authors"": [""Bohnet"", ""Bernd.""], ""title"": ""Top accuracy and fast dependency parsing is not a contradiction"", ""venue"": ""Proceedings of the 23rd International Conference on Computational Linguistics, pages 89\u201397, Beijing."", ""year"": 2010}, {""authors"": [""Brockett"", ""Chris.""], ""title"": ""Aligning the RTE 2006 corpus"", ""venue"": ""Technical Report MSR-TR-2007-77, Microsoft Research."", ""year"": 2007}, {""authors"": [""Brown"", ""Cheryl.""], ""title"": ""Topic continuity in written English narrative"", ""venue"": ""Talmy Givon, editor, Topic Continuity in Discourse: A"", ""year"": 1983}, {""authors"": [""Cai"", ""Jie"", ""Michael Strube.""], ""title"": ""End-to-end coreference resolution via hypergraph partitioning"", ""venue"": ""Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010), pages 143\u2013151,"", ""year"": 2010}, {""authors"": [""Chang"", ""Chih-Chung"", ""Chih-Jen Lin.""], ""title"": ""LIBSVM: A library for support vector machines"", ""venue"": ""ACM Transactions on Intelligent Systems and Technology, 2(27):1\u201327."", ""year"": 2011}, {""authors"": [""Charniak"", ""Eugene"", ""Micha Elsner.""], ""title"": ""EM works for pronoun anaphora resolution"", ""venue"": ""Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pages 148\u2013156, Athens."", ""year"": 2009}, {""authors"": [""Chen"", ""Desai"", ""Nathan Schneider"", ""Dipanjan Das"", ""Noah A. Smith.""], ""title"": ""SEMAFOR: Frame argument resolution with log-linear models"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic"", ""year"": 2010}, {""authors"": [""Chen"", ""Zheng"", ""Heng Ji.""], ""title"": ""Graph-based clustering for computational linguistics: A survey"", ""venue"": ""Proceedings of TextGraphs-5 - 2010 Workshop on Graph-based Methods for Natural Language Processing, pages 1\u20139,"", ""year"": 2010}, {""authors"": [""Cohn"", ""Trevor"", ""Chris Callison-Burch"", ""Mirella Lapata.""], ""title"": ""Constructing corpora for development and evaluation of paraphrase systems"", ""venue"": ""Computational Linguistics, 34(4):597\u2013614."", ""year"": 2008}, {""authors"": [""Das"", ""Dipanjan"", ""Nathan Schneider"", ""Desai Chen"", ""Noah A. Smith.""], ""title"": ""Probabilistic frame-semantic parsing"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American"", ""year"": 2010}, {""authors"": [""DeNero"", ""John"", ""Alexandre Bouchard-C\u00f4t\u00e9"", ""Dan Klein.""], ""title"": ""Sampling alignment structure under a Bayesian translation model"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language"", ""year"": 2008}, {""authors"": [""Di Eugenio"", ""Barbara.""], ""title"": ""Centering theory and the Italian pronominal system"", ""venue"": ""Proceedings of the 13th Conference on Computational Linguistics, pages 270\u2013275, Helsinki."", ""year"": 1990}, {""authors"": [""Elsner"", ""Micha"", ""Eugene Charniak.""], ""title"": ""Coreference-inspired coherence modeling"", ""venue"": ""Proceedings of ACL-08: HLT, Short Papers, pages 41\u201344, Columbus, OH."", ""year"": 2008}, {""authors"": [""Elsner"", ""Micha"", ""Eugene Charniak.""], ""title"": ""Disentangling chat with local coherence models"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies,"", ""year"": 2011}, {""authors"": [""Elsner"", ""Micha"", ""Eugene Charniak.""], ""title"": ""Extending the entity grid with entity-specific features"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human"", ""year"": 2011}, {""authors"": [""Erk"", ""Katrin"", ""Sebastian Pad\u00f3.""], ""title"": ""A structured vector space model for word meaning in context"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, Waikiki, HI."", ""year"": 2008}, {""authors"": [""Fellbaum"", ""Christiane""], ""title"": ""WordNet: An Electronic Lexical Database"", ""year"": 1998}, {""authors"": [""Feng"", ""Vanessa Wei"", ""Graeme Hirst.""], ""title"": ""Patterns of local discourse coherence as a feature for authorship attribution"", ""venue"": ""Literary and Linguistic Computing, 29:191\u2013198."", ""year"": 2014}, {""authors"": [""Filippova"", ""Katja"", ""Michael Strube.""], ""title"": ""Extending the entity-grid coherence model to semantically related entities"", ""venue"": ""Proceedings of the 11th European Workshop on Natural Language Generation,"", ""year"": 2007}, {""authors"": [""Fillmore"", ""Charles J.""], ""title"": ""Pragmatically controlled zero anaphora"", ""venue"": ""Proceedings of the Twelfth Annual Meeting of the Berkeley Linguistics Society, pages 95\u2013107, Berkeley, CA."", ""year"": 1986}, {""authors"": [""Fleiss"", ""Joseph L"", ""Bruce Levin"", ""Myunghee Cho Paik.""], ""title"": ""The measurement of interrater agreement"", ""venue"": ""Statistical Methods for Rates and Proportions, 2:212\u2013236."", ""year"": 1981}, {""authors"": [""F\u00fcrstenau"", ""Hagen"", ""Mirella Lapata.""], ""title"": ""Semi-supervised semantic role labeling via structural alignment"", ""venue"": ""Computational Linguistics, 38(1):135\u2013171."", ""year"": 2012}, {""authors"": [""Gerber"", ""Matthew"", ""Joyce Chai.""], ""title"": ""Semantic role labeling of implicit arguments for nominal predicates"", ""venue"": ""Computational Linguistics, 38(4):755\u2013798."", ""year"": 2012}, {""authors"": [""Goldberg"", ""Andrew V."", ""Robert E. Tarjan.""], ""title"": ""A new approach to the maximum flow problem"", ""venue"": ""Proceedings of the Eighteenth Annual ACM Symposium on Theory of Computing, pages 136\u2013146,"", ""year"": 1986}, {""authors"": [""Gordon"", ""Peter C."", ""Barbara J. Grosz"", ""Laura A. Gilliom.""], ""title"": ""Pronouns, names, and the centering of attention in discourse"", ""venue"": ""Cognitive Science, 17:311\u2013347."", ""year"": 1993}, {""authors"": [""Gorinski"", ""Philip"", ""Josef Ruppenhofer"", ""Caroline Sporleder.""], ""title"": ""Towards weakly supervised resolution of null instantiations"", ""venue"": ""Proceedings of the 10th International Conference on Computational"", ""year"": 2013}, {""authors"": [""Grosz"", ""Barbara J."", ""Aravind K. Joshi"", ""Scott Weinstein.""], ""title"": ""Centering: A framework for modeling the local coherence of discourse"", ""venue"": ""Computational Linguistics, 21(2):203\u2013225."", ""year"": 1995}, {""authors"": [""Guo"", ""Weiwei"", ""Mona Diab.""], ""title"": ""Semantic topic models: Combining word distributional statistics and dictionary definitions"", ""venue"": ""Proceedings of the 2011 Conference on Empirical Methods in Natural"", ""year"": 2011}, {""authors"": [""Hou"", ""Yufang"", ""Katja Markert"", ""Michael Strube.""], ""title"": ""Global inference for bridging anaphora resolution"", ""venue"": ""Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational"", ""year"": 2013}, {""authors"": [""Hwa"", ""Rebecca"", ""Philip Resnik"", ""Amy Weinberg"", ""Clara Cabezas"", ""Okan Kolak.""], ""title"": ""Bootstrapping parsers via syntactic projection across parallel texts"", ""venue"": ""Natural Language Engineering, 11(3):311\u2013325."", ""year"": 2005}, {""authors"": [""Joshi"", ""Aravind K."", ""Steve Kuhn.""], ""title"": ""Centered logic: The role of entity centered sentence representation in natural language inferencing"", ""venue"": ""Proceedings of the 6th International Joint Conference on Artificial"", ""year"": 1979}, {""authors"": [""Kameyama"", ""Megumi.""], ""title"": ""Zero Anaphora: The Case of Japanese"", ""venue"": ""Ph.D. thesis, Stanford University."", ""year"": 1985}, {""authors"": [""Karamanis"", ""Nikiforos"", ""Chris Mellish"", ""Massimo Poesio"", ""Jon Oberlander.""], ""title"": ""Evaluating centering for information ordering using corpora"", ""venue"": ""Computational Linguistics, 35(1):29\u201346."", ""year"": 2009}, {""authors"": [""Kay"", ""Martin"", ""Martin R\u00f6scheisen.""], ""title"": ""Text-translation alignment"", ""venue"": ""Computational Linguistics, 19(1):121\u2013142."", ""year"": 1993}, {""authors"": [""Kipper"", ""Karin"", ""Anna Korhonen"", ""Neville Ryant"", ""Martha Palmer.""], ""title"": ""A large-scale classification of English verbs"", ""venue"": ""Language Resources and Evaluation Journal, 42(1):21\u201340."", ""year"": 2008}, {""authors"": [""Kozhevnikov"", ""Mikhail"", ""Ivan Titov""], ""title"": ""Crosslingual transfer of semantic role"", ""year"": 2013}, {""authors"": [""T.K. Landauer"", ""S.T. Dumais.""], ""title"": ""A solution to Plato\u2019s problem: The latent semantic analysis theory of the acquisition, induction, and representation of knowledge"", ""venue"": ""Psychological Review,"", ""year"": 1997}, {""authors"": [""Laparra"", ""Egoitz"", ""German Rigau.""], ""title"": ""Exploiting explicit annotations and semantic types for implicit argument resolution"", ""venue"": ""Proceedings of the Sixth IEEE International Conference on Semantic"", ""year"": 2012}, {""authors"": [""Laparra"", ""Egoitz"", ""German Rigau.""], ""title"": ""Sources of evidence for implicit argument resolution"", ""venue"": ""Proceedings of the 10th International Conference on Computational Semantics (IWCS 2013) \u2013 Long Papers,"", ""year"": 2013}, {""authors"": [""Lee"", ""Heeyoung"", ""Angel Chang"", ""Yves Peirsman"", ""Nathanael Chambers"", ""Mihai Surdeanu"", ""Dan Jurafsky""], ""title"": ""Deterministic coreference resolution based on entity-centric, precision-ranked rules"", ""year"": 2013}, {""authors"": [""Lee"", ""Heeyoung"", ""Marta Recasens"", ""Angel Chang"", ""Mihai Surdeanu"", ""Dan Jurafsky.""], ""title"": ""Joint entity and event coreference resolution across documents"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical"", ""year"": 2012}, {""authors"": [""Liang"", ""Percy"", ""Benjamin Taskar"", ""Dan Klein.""], ""title"": ""Alignment by agreement"", ""venue"": ""North American Association for Computational Linguistics (NAACL), pages 104\u2013111, New York, NY."", ""year"": 2006}, {""authors"": [""Lin"", ""Dekang.""], ""title"": ""An information-theoretic definition of similarity"", ""venue"": ""Proceedings of the 15th International Conference on Machine Learning, pages 296\u2013304, Madison, WI."", ""year"": 1998}, {""authors"": [""Louis"", ""Annie"", ""Ani Nenkova.""], ""title"": ""Creating local coherence: An empirical assessment"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association"", ""year"": 2010}, {""authors"": [""Manning"", ""Christopher"", ""Mihai Surdeanu"", ""John Bauer"", ""Jenny Finkel"", ""Steven Bethard"", ""David McClosky.""], ""title"": ""The Stanford corenlp natural language processing toolkit"", ""venue"": ""Proceedings of the 52nd Annual"", ""year"": 2014}, {""authors"": [""Martschat"", ""Sebastian"", ""Jie Cai"", ""Samuel Broscheit"", ""\u00c9va M\u00fajdricza-Maydt"", ""Michael Strube.""], ""title"": ""A multigraph model for coreference resolution"", ""venue"": ""Joint Conference on EMNLP and CoNLL - Shared"", ""year"": 2012}, {""authors"": [""McIntyre"", ""Neil"", ""Mirella Lapata.""], ""title"": ""Plot induction and evolutionary search for story generation"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1562\u20131572,"", ""year"": 2010}, {""authors"": [""Meyers"", ""Adam.""], ""title"": ""Annotation guidelines for NomBank\u2013noun argument structure for PropBank"", ""venue"": ""Technical report, New York University. Available from http://nlp. cs.nyu.edu/meyers/nombank/nombank-"", ""year"": 2007}, {""authors"": [""Mitchell"", ""Jeff"", ""Mirella Lapata.""], ""title"": ""Composition in distributional models of semantics"", ""venue"": ""Cognitive Science, 34(8):1388\u20131429."", ""year"": 2010}, {""authors"": [""Moor"", ""Tatjana"", ""Michael Roth"", ""Anette Frank.""], ""title"": ""Predicate-specific annotations for implicit role binding: Corpus annotation, data analysis and evaluation experiments"", ""venue"": ""Proceedings of the 10th"", ""year"": 2013}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Mirella Lapata.""], ""title"": ""Dependency-based construction of semantic space models"", ""venue"": ""Computational Linguistics, 33(2):161\u2013199."", ""year"": 2007}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Mirella Lapata.""], ""title"": ""Cross-lingual annotation projection of semantic roles"", ""venue"": ""Journal of Artificial Intelligence Research, 36:307\u2013340."", ""year"": 2009}, {""authors"": [""Palmer"", ""Martha.""], ""title"": ""Semlink: Linking PropBank, VerbNet and FrameNet"", ""venue"": ""Proceedings of the Generative Lexicon Conference, pages 9\u201315, Pisa."", ""year"": 2009}, {""authors"": [""Palmer"", ""Martha"", ""Daniel Gildea"", ""Paul Kingsbury.""], ""title"": ""The Proposition Bank: An annotated corpus of semantic roles"", ""venue"": ""Computational Linguistics, 31(1):71\u2013106."", ""year"": 2005}, {""authors"": [""Palmer"", ""Martha S."", ""Deborah A. Dahl"", ""Rebecca J. Schiffman"", ""Lynette Hirschman"", ""Marcia Linebarger"", ""John Dowding.""], ""title"": ""Recovering implicit information"", ""venue"": ""Proceedings of the 24th Annual Meeting"", ""year"": 1986}, {""authors"": [""Parker"", ""Robert"", ""David Graff"", ""Jumbo Kong"", ""Ke Chen"", ""Kazuaki Maeda.""], ""title"": ""English Gigaword Fifth Edition"", ""venue"": ""Linguistic Data Consortium, Philadelphia, PA."", ""year"": 2011}, {""authors"": [""Pedersen"", ""Ted.""], ""title"": ""Information content measures of semantic similarity perform better without sense-tagged text"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of"", ""year"": 2010}, {""authors"": [""Pitler"", ""Emily"", ""Ani Nenkova.""], ""title"": ""Revisiting readability: A unified framework for predicting text quality"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language"", ""year"": 2008}, {""authors"": [""Poesio"", ""Massimo"", ""Rosemary Stevenson"", ""Barbara Di Eugenio"", ""Janet Hitzeman.""], ""title"": ""Centering: A parametric theory and its instantiations"", ""venue"": ""Computational Linguistics, 30(3):309\u2013363."", ""year"": 2004}, {""authors"": [""Postolache"", ""Oana"", ""Dan Cristea"", ""Constantin Orasan.""], ""title"": ""Transferring coreference chains through word alignment"", ""venue"": ""Proceedings of the Fifth International Conference on Language"", ""year"": 2006}, {""authors"": [""Pradhan"", ""Sameer"", ""Alessandro Moschitti"", ""Nianwen Xue"", ""Olga Uryupina"", ""Yuchen Zhang.""], ""title"": ""CoNLL-2012 shared task: Modeling multilingual unrestricted coreference in OntoNotes"", ""venue"": ""Joint"", ""year"": 2012}, {""authors"": [""Pradhan"", ""Sameer S."", ""Lance Ramshaw"", ""Ralph Weischedel"", ""Jessica MacBride"", ""Linnea Micciulla.""], ""title"": ""Unrestricted coreference: Identifying entities and events in OntoNotes"", ""venue"": ""Proceedings of the First IEEE"", ""year"": 2007}, {""authors"": [""Recasens"", ""Marta"", ""Marta Vila.""], ""title"": ""Squibs: On paraphrase and coreference"", ""venue"": ""Computational Linguistics, 36(4):639\u2013647."", ""year"": 2010}, {""authors"": [""Roth"", ""Michael"", ""Anette Frank.""], ""title"": ""Aligning predicate argument structures in monolingual comparable texts: A new corpus for a new task"", ""venue"": ""Proceedings of the First Joint Conference on Lexical and"", ""year"": 2012}, {""authors"": [""Roth"", ""Michael"", ""Anette Frank.""], ""title"": ""Aligning predicates across monolingual comparable texts using graph-based clustering"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods in Natural"", ""year"": 2012}, {""authors"": [""Ruppenhofer"", ""Josef"", ""Russell Lee-Goldman"", ""Caroline Sporleder"", ""Roser Morante.""], ""title"": ""Beyond sentence-level semantic role labeling: linking argument structures in discourse"", ""venue"": ""Language Resources and"", ""year"": 2012}, {""authors"": [""Ruppenhofer"", ""Josef"", ""Caroline Sporleder"", ""Roser Morante"", ""Collin Baker"", ""Martha Palmer.""], ""title"": ""SemEval-2010 Task 10: Linking events and their participants in discourse"", ""venue"": ""Proceedings of the 5th"", ""year"": 2010}, {""authors"": [""Sidner"", ""Candace L.""], ""title"": ""Towards a computational theory of definite anaphora comprehension in English"", ""venue"": ""Technical Report AI-Memo 537, Massachusetts Institute of Technology, AI Lab,"", ""year"": 1979}, {""authors"": [""Silberer"", ""Carina"", ""Anette Frank.""], ""title"": ""Casting implicit role linking as an anaphora resolution task"", ""venue"": ""Proceedings of the First Joint Conference on Lexical and Computational Semantics (*SEM 2012),"", ""year"": 2012}, {""authors"": [""Spearman"", ""Charles.""], ""title"": ""The proof and measurement of association between two things"", ""venue"": ""The American Journal of Psychology, 15(1):72\u2013101."", ""year"": 1904}, {""authors"": [""Sporleder"", ""Caroline"", ""Alex Lascarides.""], ""title"": ""Using automatically labelled examples to classify rhetorical relations: A critical assessment"", ""venue"": ""Natural Language Engineering, 14(3):369\u2013416."", ""year"": 2008}, {""authors"": [""Spreyer"", ""Kathrin"", ""Anette Frank.""], ""title"": ""Projection-based acquisition of a temporal labeller"", ""venue"": ""Proceedings of the Third International Joint Conference on Natural Language Processing, pages 489\u2013496,"", ""year"": 2008}, {""authors"": [""Su"", ""Fangzhong"", ""Katja Markert.""], ""title"": ""Subjectivity recognition on word senses via semi-supervised mincuts"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North"", ""year"": 2009}, {""authors"": [""Tonelli"", ""Sara"", ""Rodolfo Delmonte.""], ""title"": ""VENSES++: Adapting a deep semantic processing system to the identification of null instantiations"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic"", ""year"": 2010}, {""authors"": [""Tonelli"", ""Sara"", ""Rodolfo Delmonte""], ""title"": ""Desperately seeking implicit arguments"", ""year"": 2011}, {""authors"": [""Turan"", ""\u00dcmit Deniz.""], ""title"": ""Null vs"", ""venue"": ""Overt Subjects in Turkish: A Centering Approach. Ph.D. thesis, University of Pennsylvania."", ""year"": 1995}, {""authors"": [""van der Plas"", ""Lonneke"", ""Paola Merlo"", ""James Henderson""], ""title"": ""Scaling up automatic cross-lingual semantic role annotation"", ""venue"": ""In Proceedings of the 49th Annual Meeting of the Association"", ""year"": 2011}, {""authors"": [""Walker"", ""Christopher"", ""Stephanie Strassel"", ""Julie Medero"", ""Kazuaki Maeda.""], ""title"": ""ACE 2005 Multilingual Training Corpus"", ""venue"": ""Linguistic Data Consortium, Philadelphia, PA."", ""year"": 2006}, {""authors"": [""Wan"", ""Stephen"", ""Mark Dras"", ""Robert Dale"", ""Cecile Paris.""], ""title"": ""Using dependency-based features to take the \u201cPara-farce\u201d out of paraphrase"", ""venue"": ""Proceedings of the Australasian Language"", ""year"": 2006}, {""authors"": [""Xuchen Yao""], ""title"": ""PARMA: A predicate"", ""year"": 2013}]",acknowledgments  :We thank our annotators in Heidelberg and Edinburgh. We are grateful to the anonymous reviewers for helpful feedback and suggestions. The research leading to these results has received funding by the Landes-graduiertenförderung BadenWürttemberg within the research initiative “Coherence in language processing” at Heidelberg University. Work by M. R. at the University of Edinburgh was funded by a DFG Research Fellowship (RO 4848/1-1).,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,"a framework and its applications :Michael Roth∗ University of Edinburgh
Anette Frank∗∗ Heidelberg University
In this article, we investigate aspects of sentential meaning that are not expressed in local predicate–argument structures. In particular, we examine instances of semantic arguments that are only inferable from discourse context. The goal of this work is to automatically acquire and process such instances, which we also refer to as implicit arguments, to improve computational models of language. As contributions towards this goal, we establish an effective framework for the difficult task of inducing implicit arguments and their antecedents in discourse and empirically demonstrate the importance of modeling this phenomenon in discourse-level tasks.
Our framework builds upon a novel projection approach that allows for the accurate detection of implicit arguments by aligning and comparing predicate–argument structures across pairs of comparable texts. As part of this framework, we develop a graph-based model for predicate alignment that significantly outperforms previous approaches. Based on such alignments, we show that implicit argument instances can be automatically induced and applied to improve a current model of linking implicit arguments in discourse. We further validate that decisions on argument realization, although being a subtle phenomenon most of the time, can considerably affect the perceived coherence of a text. Our experiments reveal that previous models of coherence are not able to predict this impact. Consequently, we develop a novel coherence model, which learns to accurately predict argument realization based on automatically aligned pairs of implicit and explicit arguments.",,,,,,,,,,,,,,,,,"1 introduction :The goal of semantic parsing is to automatically process natural language text and map the underlying meaning of text to appropriate representations. Semantic role labeling
∗ School of Informatics, University of Edinburgh, EH8 9AB Edinburgh, United Kingdom. E-mail: mroth@inf.ed.ac.uk.
∗∗ Department of Computational Linguistics, Heidelberg University, 69120 Heidelberg, Germany. E-mail: frank@cl.uni-heidelberg.de.
Submission received: 5 June 2014; revised version received: 13 April 2015; accepted for publication: 28 August 2015.
doi:10.1162/COLI a 00236
© 2015 Association for Computational Linguistics
induces shallow semantic representations, so-called predicate–argument structures, by processing sentences and mapping them to predicates and associated arguments. Arguments of these structures can, however, be non-local in natural language text, as shown in Example 1.
Example 1 (a) El Salvador is the only Latin American country which has troops in Iraq.
(b) Nicaragua withdrew its troops last month.1
Applying a semantic role labeling system on sentence (1b) produces a representation that consists of the predicate withdraw, a temporal modifier (last month) and two associated arguments: the entity withdrawing (Nicaragua) and the thing being withdrawn (its troops). From the previous sentence (1a), we can additionally infer a third argument: namely, the source from which Nicaragua withdrew its troops (Iraq). By leaving this piece of information implicit, the text fragment in sentence (1b) illustrates a typical case of non-local, or implicit, role realization (Gerber and Chai 2012). In this article, we view implicit arguments as a discourse-level phenomenon and treat corresponding instances as implicit references to discourse entities.
Taking this perspective, we build upon previous work on discourse analysis. Following Sidner (1979) and Joshi and Kuhn (1979), utterances in discourse typically focus on a set of salient entities, which are also called the foci or centers. Using the notion of centers, Grosz, Joshi, and Weinstein (1995) defined the Centering framework, which relates the salience of an entity in discourse to linguistic factors such as choice of referring expression and syntactic form.2 Both extremes of salience, that is, contexts of referential continuity and irrelevance, can also be reflected by the non-realization of an entity (Brown 1983). Specific instances of this phenomenon, so-called zero anaphora, have been well-studied in pro-drop languages such as Japanese (Kameyama 1985), Turkish (Turan 1995), and Italian (Di Eugenio 1990). For English, only a few studies exist that explicitly investigated the effect of non-realizations on coherence. Existing work suggests, however, that indirect references and non-realizations are important for modeling and measuring coherence (Poesio et al. 2004; Karamanis et al. 2009), respectively, and that such phenomena need to be taken into consideration to explain local coherence where adjacent sentences are neither connected by discourse relations nor in terms of coreference (Louis and Nenkova 2010).
In this work, we propose a new model to predict whether realizing an argument contributes to local coherence in a given position in discourse. Example 1 illustrated a text fragment, in which argument realization is necessary in the first sentence but redundant in the second. That is, mentioning Iraq in the second sentence is not necessary (for a human being) to understand the meaning of the text. In contrast, making both references explicit, as shown in Example 2, would be redundant and could lead to the perception that the text is merely a concatenation of two independent sentences — rather than a set of adjacent sentences that form a meaningful, or coherent, discourse.
1 In the remainder of this article, we use the following notation for predicate–argument structures: underlined words refer to predicates and [possibly empty] constituents in [square brackets] refer to arguments. 2 We note that the relation between coherence and choice of referring expressions has also been studied outside of the Centering framework. A comprehensive overview of research in linguistics on this subject can be found in Arnold (1998).
Example 2 (a) El Salvador is now the only Latin American country which has troops in Iraq.
(b) [Nicaragua] withdrew [its troops] [from Iraq] last month.
The phenomenon of implicit arguments has neither been extensively studied in the context of semantic role labeling nor in coherence modeling. One of the main reasons for this lies in the fact that models for these tasks are typically developed on the basis of annotated corpora. In contrast, implicit arguments are not overt in text and are difficult to uncover, hence there exist only few and small data sets that contain respective annotations explicitly. In this article, we present a novel framework for computationally inducing a corpus with automatic annotations of implicit arguments and their respective antecedents in discourse. To achieve this goal, our framework exploits pairs of comparable texts, which convey information about the same events, states, and entities but potentially differ in terms of depth and perspective. These differences can also affect argument realization, making it possible to identify instances of co-referring but only partially overlapping argument structures. Given automatically induced instances of implicit arguments and their document contexts, we show how these can be utilized to enhance current models of local coherence and implicit argument linking. The resulting models and research insights will be of importance for many applications that involve the understanding or generation of natural language text beyond the sentence level. 2 overview :The goal of this work is to provide data and methods to better capture the phenomenon of implicit arguments in semantic role labeling and coherence modeling. To accomplish this goal, we propose inducing suitable training data automatically. For semantic role labeling, this kind of training data will contain explicit links between implicit arguments and their respective discourse antecedents; for modeling coherent argument realization, the data will also provide discourse contexts of explicit and implicit references to the same entity. We propose inducing such information by exploiting pairs of comparable texts. Our motivation for this lies in the fact that comparable texts convey by and large the same information but differ in depth and presentation. This means that while we expect two comparable texts to contain references to the same events and participants, an entity might be understood implicitly at a specific point in one text (because it can be inferred from context), whereas an explicit mention might be necessary in a comparable text (given the different context). Based on this assumption, our method aims to find complementary (explicit) information in pairs of texts, which can be aligned and subsequently be projected to identify missing (implicit) pieces in one another. Hence, we separate our task into two parts. In the first part, described in Section 4, we ask the question:
Q1: How can we detect predicate–argument structures that are shared across two discourses?
Based on a corpus of comparable texts, we propose a new task of aligning predicate– argument structures (PASs) across complete discourses. We define this task on the basis of semantic annotations, which we compute automatically using a state-of-the-art semantic role labeler. For the alignment task itself, we develop a graph-based clustering
model. We show that by taking into account features on the levels of predicate, arguments, and discourse context, this model is able to predict accurate alignments across comparable texts, without relying on preprocessing methods that identify parallel text fragments. Based on automatically aligned structures, the second part of our approach aims to identify and resolve implicit arguments by projecting explicit information from one text to another. The question for this part of our approach (cf. Section 5) is:
Q2: How can we find implicit arguments and their antecedents in discourse?
Our approach to answering this question relies solely on information that can be computed automatically: We first identify potential instances of implicit arguments by comparing two aligned predicate–argument structures; for each entity that is mentioned in one PAS (explicit argument) but not in an aligned structure (implicit argument), we then apply a cross-document coreference resolution technique to also find co-referring entity mentions in the document in which an implicit reference was found.
Finally, we demonstrate the utility of automatically induced instances of implicit arguments in two task-based settings: linking arguments in discourse and modeling local coherence. The question here is:
Q3: How can induced argument instances and their respective discourse contexts be utilized to enhance NLP models?
We address both tasks by applying our automatically induced data set as training material to improve the performance of supervised models. For the first task, described in Section 6, we apply our data to enhance training of an existing system that identifies and links implicit arguments in discourse. To evaluate the impact of our induced data set, we test the modified model on a standard evaluation data set, on which we compare our results with those of previous work. For the second task, we develop a new coherence model that predicts whether an argument realization or non-realization in context improves the perceived coherence of the affected segment in discourse. We evaluate this coherence model in an intrinsic evaluation scenario, in which we compare model predictions to human judgments on argument use (cf. Section 7). 3 background :This article draws on insights from computational linguistic research on implicit arguments and coherence modeling as well as from previous work on inducing semantic resources. It is further related to recent work in paraphrasing and event coreference. The goal of this work is to induce instances of implicit arguments, together with their discourse antecedents, and to utilize them in semantic processing and coherence modeling. This section summarizes previous work on implicit arguments and coherence modeling, and provides an outlook on how instances of implicit arguments can be of use in a novel entity-based model of local coherence.
Implicit Arguments. The role of implicit arguments was studied early on in the context of semantic processing (Fillmore 1986; Palmer et al. 1986), although most semantic
role labeling systems nowadays operate solely within local syntactic structures and do not perform any additional inference regarding missing information. First data sets that focus on implicit arguments have only recently become available: Ruppenhofer et al. (2010) organized a SemEval shared task on “linking events and participants in discourse,” Gerber and Chai (2012) made available implicit argument annotations for NomBank (Meyers, Reeves, and Macleod 2008), and Moor, Roth, and Frank (2013) provided annotations for parts of the OntoNotes corpus (Weischedel et al. 2011). All three resources are, however, severely limited: Annotations in the latter two studies are restricted to 10 and 5 predicate types, respectively; the training set of the SemEval task, in contrast, consists of full-text annotations for all occurring predicates but contains only 245 instances of resolved implicit arguments in total. All groups working on the shared task identified data sparsity as one of the main issues (Chen et al. 2010; Ruppenhofer et al. 2012; Laparra and Rigau 2013). Silberer and Frank (2012) point out that additional training data can be heuristically created by treating anaphoric pronoun mentions as implicit arguments. Their experimental results confirmed that artificial training data can indeed improve results, but only when obtained from corpora with manual semantic role annotations (on the sentence level) and gold coreference chains. As observed in related work on classifying discourse relations, information learned from artificial training data might not always generalize well to naturally occurring examples (Sporleder and Lascarides 2008). To automatically create data that is linguistically more similar to manually labeled implicit arguments, we introduce an alternative method that induces instances of implicit arguments from a raw corpus of comparable texts.
Coherence Modeling. In the context of coherence modeling, much previous work has focused on entity-based approaches, with the most prominent model being the entity grid by Barzilay and Lapata (2005). This model has originally been proposed for automatic sentence ordering but has since also been applied in coherence evaluation and readability assessment (Barzilay and Lapata 2008; Pitler and Nenkova 2008), story generation (McIntyre and Lapata 2010), and authorship attribution (Feng and Hirst 2014). Based on the original model, several extensions have been proposed: For example, Filippova and Strube (2007) and Elsner and Charniak (2011b) suggested additional features to characterize semantic relatedness between entities and features specific to single entities, respectively. Other entity-based approaches to coherence modeling include the pronoun model by Charniak and Elsner (2009) and the discourse-new model by Elsner and Charniak (2008). All of these approaches are, however, solely based on explicit entity mentions, resulting in insufficient representations when dealing with inferable references. Example 3 illustrates this shortcoming.
Example 3 implicit roles (a) 27 tons of cigarettes were picked up in Le Havre. agent
(b) The containers had arrived yesterday. content, destination
The two sentences in the given example do not share any explicit references to the same entity. Hence, none of the aforementioned models is able to correctly predict that both sentences cohere in the presented order—but not in the reversed order.
As discussed in more detail by Poesio et al. (2004), such cases can still be resolved when taking into account indirect realizations, that is, associative references to a discourse entity. Accordingly, recent work by Hou, Markert, and Strube (2013) proposed
considering bridging resolution to fill specific gaps in the entity grid model. Following this argument, the containers in sentence (3b) could be understood as a bridging anaphor that refers back to the 27 tons of cigarettes in sentence (3a). As mentioned in the NomBank annotation guidelines (Meyers 2007), some cases of nominal predicates and their arguments do coincide with bridging relations and we treat them as implicit arguments in this article. For example, when interpreting container as a nominalization of the verbal predicate CONTAIN, the cigarettes from the previous sentence can be understood as one of its arguments: namely, its content. Not all cases of implicit arguments involve bridging anaphora and vice versa, however. The main difference lies in the fact that bridging typically describes a relation between entities. Hence, bridging instances only coincide with those of implicit arguments if the bridging anaphor is linguistically realized as a (nominal) predicate (container in Example 3b). In contrast, implicit arguments frequently form parts of predicate–argument structures that refer to events, states, and entities that are not bridging anaphora. For example, the destination of the predicate ARRIVE in sentence (3b) can also be interpreted as an implicit argument, namely, Le Havre from the preceding context.
Taking implicit arguments into consideration, we can see that the second sentence refers back to two previously introduced entities, Le Havre and cigarettes, reflecting the fact that the sentences cohere in the presented order. In Section 7, we present a novel model of local coherence that aims to capture the effect of argument realizations on perceived coherence. As an overall goal, this model should be able to predict whether an entity reference should be realized explicitly to establish (local) coherence—or whether the entity can already be understood from context. Based on such predictions, the model can be applied in text generation to ensure that necessary references are explicit and that redundant repetitions are avoided. The methods applied in this article are based on ideas from previous work on inducing semantic resources from parallel and comparable texts. Most work in this direction has been done in the context of cross-lingual settings, including the learning of translations of words and phrases using statistical word alignments (Kay and Röscheisen 1993; DeNero, Bouchard-Côté, and Klein 2008, inter alia) and approaches to projecting annotations from one language to another (Yarowsky and Ngai 2001; Kozhevnikov and Titov 2013, inter alia). In the following, we discuss previous approaches to annotation projection as well as related work in paraphrasing and event coreference.
Annotation Projection. A widespread method for the induction of semantic resources is the so-called annotation projection approach. The rationale of this approach is to induce annotated data in one language, given already-annotated instances in another language. As an example, semantic role annotations of a text in English can be transferred to a parallel text in order to induce annotated instances for a lexicon in another language (Padó and Lapata 2009). In previous work, this method has been applied on various levels of linguistic analysis: from syntactic information in the form of part-of-speech tags and dependencies (Yarowsky and Ngai 2001; Hwa et al. 2005), through annotation of temporal expressions and semantic roles (Spreyer and Frank 2008; van der Plas, Merlo, and Henderson 2011), to discourse-level phenomena such as coreference (Postolache, Cristea, and Orasan 2006). All of the aforementioned instances of the projection approach make use of the same underlying technique: Firstly, words are
aligned in a parallel corpus using statistical word alignment; secondly, annotations on a single word or between multiple words in one text are transferred to the corresponding aligned word(s) in the parallel text. This procedure typically assumes that two parallel sentences express the same meaning. A notable exception is the work by Fürstenau and Lapata (2012), which utilizes alignments between syntactic structures to “project” semantic role information from a role-annotated corpus to unseen sentences that are selected from a corpus in the same language.
In our work, we apply annotation projection to monolingual comparable texts. In comparison to parallel texts, we have to account for potential differences in perspective and detail that make our task—in particular, the alignment sub-task—considerably more difficult. In contrast to Fürstenau and Lapata’s setting, which involves incomparable texts, we assume that text pairs in our setting still convey information on the same events and participants. This means that in addition to aligning predicate–structures across texts, we can also merge complementary details realized in each structure. We achieve this by making use of the same principle that underlies the projection approach: Given annotation that is available in one text (explicit argument), we project this information to a text in which the corresponding annotation is not available (implicit argument). Because our task is based in a monolingual setting, we can make use of the same preprocessing tools across texts.
Paraphrasing and Event Coreference. We overcome the difficulty of inducing word alignments across comparable texts by computing alignments on the basis of predicate– argument structures. Using predicate–argument structures as targets makes our setting related to previous work on paraphrase detection and coreference resolution of event mentions. Each of these tasks focuses, however, on a different level of linguistic analysis from ours: Following the definitions embraced by Recasens and Vila (2010), “paraphrasing” is a relation between two lexical units that have the same meaning, whereas “coreference” indicates that two referential expressions point to the same referent in discourse.3 In contrast to work on paraphrasing, we are specifically interested in pairs of text fragments that involve implicit arguments, which can only be resolved in context.
In line with our goal of inducing implicit arguments, we define the units or expressions to be aligned in our task as the predicate–argument structures that can (automatically) be identified in text. This task definition further makes our task distinct from event coreference, where coreference is established based on a pre-specified set of events, reference types, or definitions of event identity (Walker et al. 2006; Pradhan et al. 2007; Lee et al. 2012, inter alia). Although corresponding annotations can certainly overlap with those in our task, we emphasize that the focus of our work is not to find all occurrences of co-referring events. Instead, our goal is to align predicate–argument structures that have a common meaning in context across different discourses. Hence, we neither consider intra-document coreference nor pronominal event references here. As alignable units in our work are not restricted to a pre-specified definition of event or event identity, the task addressed here involves any kind of event, state, or entity that is linguistically realized as a predicate–argument structure. Examples that go beyond traditional event coreference include in particular noun phrases such as “a [rival] of the company” and “the [finance] director [of IBM]”(cf. Meyers, Reeves, and Macleod 2008).
3 Note that in the remainder of this article, we use the term “coreference” in a wider sense to also encompass referents to events and entities in the real world, rather than just referents grounded in discourse. 4 induction framework step 1: aligning predicate–argument structures :In the first step of our induction framework, we align predicate–argument structures (PASs) across pairs of comparable text. To execute this task properly, we define annotation guidelines and construct a gold standard for the development and evaluation of alignment models. All annotations are performed on pairs of documents from a corpus of comparable texts. We introduce this corpus and describe our annotations in Section 4.1. We describe a graph-based model that we developed for the automatic alignment in Section 4.2. Finally, we present an experimental evaluation of our model, together with another recently proposed model and several baselines, in Section 4.3. As a basis for aligning predicate–argument structures across texts, we make use of a data set of comparable texts extracted from the English Gigaword corpus (Parker et al. 2011). The Gigaword corpus is one of the largest English corpora available in the news domain and contains over 9.8 million articles from seven newswire agencies that report on (the same) real-world incidents. The data set of comparable texts used in this work contains 167,728 pairs of articles that were extracted by matching the headlines of texts published within the same time frame (Roth and Frank 2012a). A set of such document headlines is given in Example 4:
Example 4 India fires tested anti-ship cruise missile (Xinhua News Agency, 29 October 2003)
India tests supersonic cruise anti-ship missile (Agence France Presse, 29 October 2003)
URGENT: India tests anti-ship cruise missile (Associated Press Worldstream, 29 October 2003)
We preprocess each article in the set of 167,728 pairs using the MATE tools (Björkelund et al. 2010; Bohnet 2010), including a state-of-the-art semantic role labeler that identifies PropBank/NomBank-style predicate–argument structures (Palmer, Gildea, and Kingsbury 2005; Meyers, Reeves, and Macleod 2008). Based on the acquired PAS, we perform manual alignments. In Section 4.1.1, we summarize the annotation guidelines for this step. An overview of the resulting development and evaluation data set is provided in Section 4.1.2.
4.1.1 Manual Annotation. We selected 70 pairs of comparable texts and asked two annotators to manually align predicate–argument structures obtained from preprocessing. Both annotators were students in computational linguistics, one undergraduate and one postgraduate. The texts were selected with the constraint that each text consists of 100 to 300 words. We chose this constraint as longer text pairs seemed to contain a higher number of unrelated predicates, making the alignment tasks difficult to manage for the annotators. Both annotators received detailed guidelines that describe alignment requirements and the overall procedure.4 We summarize essentials in the following.
4 cf. http://projects.cl.uni-heidelberg.de/india/files/guidelines.pdf.
Sure and Possible Links. Following standard practice in word alignment (cf. Cohn, Callison-Burch and Lapata 2008, inter alia) the annotators were instructed to distinguish between sure (S) and possible (P) alignments, depending on how certainly, in their opinion, two predicates (including their arguments) describe the same event, state, or entity. Examples 5 and 6 show a sure and possible predicate pairing, respectively.
Example 5 The regulator ruled on September 27 that Nasdaq was qualified to bid. The authority had already approved a similar application by Nasdaq.
Example 6 Myanmar’s government said it has released some 220 political prisoners. The government has been regularly releasing members of Suu Kyi’s party.
Replaceability. As a rule of thumb for deciding whether to align two structures, annotators were told to check how well the affected predicate–argument structures could be replaced by one another in their given context.
Missing Context. In case one text does not provide enough context to decide whether two predicates in the paired documents refer to the same event, an alignment should not be marked as sure.
Similar Predicates. Annotators were told explicitly that sure links can be used even if two predicates are semantically different but have the same meaning in context. Example 7 illustrates such a case.
Example 7 The volcano roared back to life two weeks ago. It began erupting last month.
1-to-1 vs. n-to-m. We asked the annotators to find as many 1-to-1 correspondences as possible and to prefer 1-to-1 matches over n-to-m alignments. In case of multiple mentions of the same event, we further asked the annotators to provide only one sure link per predicate and mark remaining cases as possible links. As an additional guideline, annotators were asked to only label the PAS pair with the highest information overlap as a sure link. If there is no difference in information overlap, the predicate pair that occurs first in both texts should be marked as a sure alignment. The intuition behind this guideline is that the first mention introduces the actual event whereas later mentions just (co-)refer or add further information.
4.1.2 Resulting Data Set. In total, the annotators (A/B) aligned 487/451 sure and 221/180 possible alignments. Following Brockett (2007), we computed agreement on labeled annotations, including unaligned predicate pairs as an additional null category. We computed κ following Fleiss, Levin, and Paik (1981) and observed an overall score of 0.62, with κ values per category of 0.74 and 0.19 for sure and possible alignments, respectively. The numbers show that both annotators substantially agree on which pairs of predicate–argument structures “surely” express the same proposition. Identifying further references to the same event or state, in contrast, can only be achieved with
fairly low agreement. For the construction of a gold standard, we take the intersection of all sure alignments by both annotators and the union of all possible alignments.5 We further resolved cases that involved a sure alignment on which the annotators disagreed in a group discussion and added them to our gold standard accordingly. We split the final corpus into a development set of 10 document pairs and an evaluation set of 60 document pairs.
Table 1 summarizes information about the resulting annotations in the development and evaluation set. As can be seen, the documents in the development set contain a smaller number of predicates (39.5 vs. 57.6) and alignments (8.7 vs. 13.4) on average. The fraction of aligned predicates is, however, about the same (22.0% vs. 23.3%). Across both data sets, the average numbers of observed predicates is approximately 55, of which 31 are verbs and 24 are nouns. In the development and evaluation sets, the average number of sure alignments are 3.5 and 7.4. From all aligned predicate pairs in both data sets, 82.6% are the same part of speech (30.0% both nouns, 52.6% both verbs). In total, 48.0% of all alignments are between predicates of identical lemmata. As a rough indicator for diverging argument structures captured in the annotated alignments, we analyzed the number of aligned predicates that involve a different number of realized arguments. In both data sets together, this criterion applied in 344 cases (38.9% of all alignments). For the automatic alignment of predicate–argument structure alignments across texts, we opt for an unsupervised graph-based method. That is, we represent pairs of documents as bipartite graphs and subsequently aim to separate a graph into subgraphs that represent corresponding predicate–argument structures across texts. We define our general graph representation for this task in Section 4.2.1. In Section 4.2.2, we introduce
5 In our evaluation, only sure alignments need to be predicted by a system, whereas possible alignments are optional and not counted towards the attested recall (cf. Section 4.3).
a range of similarity measures that are used to weight edges in the graph representation. In Section 4.2.3, we introduce our algorithm to separate graphs, representing pairs of documents, into subgraphs of corresponding predicates. The overall model is an extended variant of the clustering approach described in Roth and Frank (2012b) and uses an enhanced set of similarity measures.
4.2.1 Graph Representation. We build a bipartite graph representation for each pair of texts, using as vertices the predicate–argument structures assigned during preprocessing (cf. Section 4.1). We represent each predicate as a node and integrate information about arguments implicitly. Given the sets of predicates P1 and P2 of two comparable texts T1 and T2, respectively, we formally define an undirected graph GP1,P2 following Equation 7.
GP1,P2 = 〈V, E〉 where V = P1 ∪ P2 E = P1 × P2
(7)
Edge Weights. We specify the edge weight between two nodes representing predicates p1 ∈ P1 and p2 ∈ P2 as a weighted linear combination of the similarity measures S described in the next section.
wp1p2 = |S|∑ i λi ∗ simi(p1, p2) (8)
Initially we set all weighting parameters λi to have uniform weights by default. We describe a tuning routine to find an optimized weighting scheme for the individual measures in the experimental evaluation of our approach (Section 4.3).
4.2.2 Similarity Measures. We use a number of similarity measures that make use of complementary information on the predicates, arguments, and discourse context of two predicate–argument structures. Given two lemmatized predicates p1, p2 and their sets of arguments A1 = args(p1), A2 = args(p2), we define seven measures in total. Three of them are specific to the predicates themselves, two take into account information on associated arguments, and two measures capture discourse-level properties. The predicate-specific measures as well as one argument-specific measure correspond to measures previously described in Roth and Frank (2012b). This article extends previous work by considering discourse-level information and an additional argument-specific measure that takes into account argument labels. We demonstrate the benefits of these measures in practice in Section 4.3.
Similarity in WordNet. Given all synsets that contain the two predicates p1, p2, we compute their similarity in WordNet (Fellbaum 1998) as the maximal pairwise score calculated using the information content based measure proposed by Lin (1998). We rely on the WordNet hierarchy to find the least common subsumer (lcs) of two synsets and use the pre-computed Information Content (IC) files from Pedersen (2010) to compute this measure as defined in Equation (9).
simWN(p1, p2) = max 〈s1,s2〉:si∈synsets(pi ) IC(lcs(s1, s2)) IC(s1) ∗ IC(s2)
(9)
Similarity in VerbNet. We additionally make use of VerbNet (Kipper et al. 2008) to compute similarities between verb pairs that cannot be captured by WordNet relations. Verbs in VerbNet are categorized into classes according to their meaning as well as syntactic behavior. A verb class C can recursively embed sub-classes Cs ∈ sub(C) that represent finer semantic and syntactic distinctions. In Equation (10), we define a simple similarity function that assigns fixed scores to pairs of predicates p1, p2 depending on their relatedness within VerbNet.6
simVN(p1, p2) =  1.0 if ∃C : p1, p2 ∈ C 0.8 if ∃C, Cs : Cs ∈ sub(C)
∧ ((p1 ∈ C, p2 ∈ Cs) ∨ (p1 ∈ Cs, p2 ∈ C)) default else
(10)
We empirically set the default value to the average VerbNet similarity (with unrelated pairs counted as 0.0) computed over one million random pairs of predicates in our corpus.
Similarity in a Semantic Space. As predicates can be absent from WordNet and VerbNet, or distributed over separate hierarchies due to different parts-of-speech (verbal vs. nominal predicates), we additionally calculate similarity based on distributional meaning in a semantic space (Landauer and Dumais 1997). This measure is based on the similarity of contexts of two given predicates over all their instances in a corpus. To compute this measure, we first calculate the Pointwise Mutual Information (PMI) for each predicate p ∈ {p1, p2} and the n most frequent context words c ∈ C following Equation (11).
pmi(p, c) = freq(p, c)
freq(p) ∗ freq(c) (11)
As we are dealing with predicates of different parts-of-speech, we calculate joint frequencies in terms of context windows instead of relying on syntactic dependencies as proposed in more recent approaches to distributional semantics (Padó and Lapata 2007; Erk and Padó 2008; Baroni and Lenci 2010). More precisely, we extract context windows of five words to the left and to the right from the Gigaword corpus (Parker et al. 2011), and compute the PMI for the 2,000 most frequent context words c1 . . . c2,000 ∈ C. The same setting has been successfully applied in related tasks, including word sense disambiguation (Guo and Diab 2011) and measuring phrase similarity (Mitchell and Lapata 2010). Vector representations are computed following Equation (12), and similarities are calculated as the cosine function of the angle between two vectors, as defined in Equation (13).
~p = (pmi(p, c1), pmi(p, c2), . . . , pmi(p, c2,000)) (12)
simDist(p1, p2) = ~p1 · ~p2
||~p1|| ∗ ||~p2|| (13)
6 Note that the weight of 0.8 was set in an ad hoc manner (instead of being optimized) in order to avoid overfitting on our small development corpus.
Bag-of-Words Similarity. As a simple argument-specific measure, we compute the overlap of word tokens over all (concatenated) arguments of each predicate–argument structure (PAS). Formally, this similarity measure considers all arguments a1 ∈ A1 and a2 ∈ A2 associated with the predicates p1 and p2, respectively, and calculates overlap as defined in Equation (14). In order to control the impact of frequently occurring words, we weight each word by its Inverse Document Frequency (IDF), which we calculate over all documents d in our corpus D:
simABoW(p1, p2) =
∑ w∈A1∩A2 idf(w)∑
w∈A1 idf(w) + ∑ w∈A2 idf(w) (14)
idf(w) = log |D|
|{d ∈ D|w ∈ words(d)}| (15)
Head of Arguments Similarity. We further define an argument-specific measure that only compares the semantic heads of arguments that have the same argument label. For example, given two PASs that consist of predicates p1, p2 and arguments labeled A0 and A1, we compute the similarity of the two arguments labeled A0 (also denoted as label(a) = ‘A0’) and the similarity of the two arguments labeled A1 (label(a) = ‘A1’). Each similarity between argument heads is computed using the WordNet-based measure described earlier. We extract the semantic head of each argument by considering the dependency tree, as predicted by MATE tools, in which we look for the noun or verb on the highest level within the argument span. Finally, we collapse all pairwise argument similarities into one measure by taking the average following Equation (16).
simAheads(p1, p2) =
∑ {a1,a2|label(a1 )==label(a2 )} simWN(head(a1), head(a2))
|{a1, a2|label(a1) == label(a2)}| (16)
Relative Discourse Position. On the discourse level, we measure the distance of two predicate–argument structures with respect to their relative positions in the text. Given two predicates, we compute this measure based on the absolute difference between their relative positions. The relative position in discourse is computed as the sentence index in which the predicate p1 or p2 occurs, divided by the total number of sentences in the affected document (d1 or d2, respectively). The measure, as defined in Equation (17), ranges from 1.0 (relative positions are exactly the same) to 0.0 (one predicate at the beginning, the other at the end of a text).
simDPos(p1, p2) = 1− (∣∣∣sentence index(p1)length(d1) − sentence index(p2)length(d2) ∣∣∣) (17)
Context Similarity. We further consider occurrences of (shared) predicates in the immediate discourse context of two predicates. Our measure for this type of similarity is
computed as the relative number of overlapping predicate types within the preceding and succeeding n neighboring predicate instances as defined in Equation (18).
simDCon(p1, p2) = context(p1 )∩context(p2 ) context(p1 )∪context(p2 ) , with
context(p) = {p′|index(p′) ∈ [index(p)− n : index(p) + n]} (18)
We compute the index of a predicate as the number of preceding predicates within the same text, that is, starting with zero. Based on preliminary experiments on our development set, we empirically set n to 5.
4.2.3 Alignment via Clustering. The goal of this step of our induction framework is to identify pairs of PASs that describe the same event, state, or entity. As a means to achieve this goal, we represent pairs of comparable texts as graphs and aim to find those edges in a graph that represent connections between predicates that need to be aligned. Although our aim is to find edges between nodes, we note that the majority of predicates (nodes) in our data set are not aligned and hence a crucial prerequisite to generate precise alignments is to filter out those nodes that are unlikely to be good alignment candidates. To achieve the filtering and alignment goals at the same time, we rely on graph clustering techniques that have successfully been applied in the NLP literature (Su and Markert 2009; Cai and Strube 2010; Chen and Ji 2010, inter alia) and that can be used to partition a graph into singleton nodes and smaller subgraphs.
The clustering method applied in our model relies on so-called minimum cuts (henceforth also called mincuts) in order to partition a bipartite graph, representing pairs of texts, into clusters of alignable predicate–argument structures. A mincut operation divides a given graph into two disjoint subgraphs. Each cut is performed between some source node s and some target node t, such that (1) each of the two nodes will be in a different subgraph and (2) the sum of weights of all removed edges will be as small as possible. We implement basic graph operations using the freely available Java library JGraphT7 and determine each mincut using the method described in Goldberg and Tarhan (1986).
Given a constructed input graph G, our algorithm recursively applies mincuts in three steps as follows:
1. Identify edge e with the lowest weight in the current (sub)graph G.
2. Perform a mincut such that the nodes connected by e will be in two different subgraphs G′ and G′′.
3. Recursively apply Steps 1 and 2 to subgraphs G′ and G′′.
As our goal is to induce clusters that correspond to pairs of corresponding structures, we apply Step 3 of the clustering approach outlined above only to subgraphs that contain more than two nodes. Algorithm 1 provides a pseudocode implementation of this procedure. An example of an input graph and the applied minimum cuts is illustrated in Figure 1. As shown in the illustration, we only use edges in our initial graph representation that represent alignment candidates with a similarity above a
7 http://jgrapht.org/.
Algorithm 1. Pseudo code of our clustering algorithm. function CLUSTER(G)
clusters← ∅ E← GETEDGES(G) . Step 1 e← GETEDGEWITHLOWESTWEIGHT(E) s← GETSOURCENODE(e) t← GETTARGETNODE(e) G′ ← MINCUT(G, s, t) . Step 2 C ← GETCONNECTEDCOMPONENTS(G′) for all Gs ∈ C do . Step 3
if SIZE(Gs) <= 2 then clusters← clusters ∪ Gs else clusters← clusters ∪ CLUSTER(Gs)
end if end for return clusters;
end function
threshold determined on the development part of our manually aligned data. Based on the initial representation, the first cut (Cut 1) is performed between the nodes connected by the edge with the lowest weight in the overall graph (13.0). This cut separates the nodes representing the “earnings NNS” predicates from the rest of the graph. Similarly, Cut 2 separates another cluster of two nodes. Finally, Cuts 3 and 4 remove a single node from the only remaining cluster that had more than two nodes.
The main benefit of our method compared with off-the-shelf clustering techniques is that we can define the termination criterion in line with the goal of our task, namely,
to align pairs of structures across comparable texts, instead of having to optimize additional parameters that require careful fine-tuning (such as the number of clusters or a clustering threshold). In the next section, we provide empirical evidence for the advantage of this approach. This section describes the evaluation of our graph-based clustering model on the task of aligning predicate–argument structures across comparable texts. We define two variants of our graph-based model: Full makes use of all similarity measures described in Section 4.2.2, and EMNLP’12, the model introduced in Roth and Frank (2012b), only uses the predicate-specific measures and the bag-of-words measure for argument similarity. Both models, Full and EMNLP’12, make use of the clustering algorithm introduced in Section 4.2.3.
For evaluation, we compare our model against two baselines and a model from the literature that has recently been proposed for this task (Wolfe et al. 2013). Similar to our approach, Wolfe et al. use various resources to calculate the similarity of two predicate– argument structures. Differences to our model lie in the kind of utilized resources, the use of additional data to learn feature weights, and the fact that each alignment decision is made using a binary classifier. We evaluate the predictions of each model on the manually annotated test set described in Section 4.1.2.8
As evaluation measures, we use precision, recall, and F1-score. Following previous work aligning monolingual texts (Cohn, Callison-Burch, and Lapata 2008), we measure precision as the number of predicted alignments also annotated in the gold standard divided by the total number of predictions, and recall as the number of correctly predicted sure alignments divided by the total number of sure alignments in the gold standard. F1-score is computed as as the harmonic mean between precision and recall.
4.3.1 Baselines. A simple baseline for predicate alignment is to simply cluster all predicates that have identical lemmata (henceforth called LemmaId). To assess the benefits of the clustering step, we propose a second baseline that uses the same similarity measures as our Full model but does not use the mincut clustering described in Section 4.2.3. Instead, it greedily merges as many 1-to-1 alignments as possible, starting with the highest similarity (Greedy). As a more sophisticated baseline, we make use of alignment tools commonly used in statistical machine translation. We train our own word alignment model using the state-of-the-art word alignment tool Berkeley Aligner (Liang, Taskar, and Klein 2006). As word alignment tools require pairs of sentences as input, we first extract paraphrases using a re-implementation of a previously proposed paraphrase detection system based on lemma and n-gram overlap (Wan et al. 2006). In the following section, we abbreviate the alignment based model as WordAlign.
4.3.2 Results. The results for the alignment task are presented in Table 2. From all approaches, Greedy and WordAlign yield the lowest performance. For WordAlign, we observe two main reasons. On the one hand, sentence paraphrase detection does not perform perfectly. Hence, the extracted sentence pairs do not always contain gold
8 We also performed an evaluation on sentence-level predicate alignment, but skipped the discussion here as this task is not relevant for our induction framework. As the additional discourse-level measures of the Full model are not needed for alignment within sentences, we refer the interested reader to the EMNLP’12 model and its evaluation in Roth and Frank (2012b).
alignments. On the other hand, even sentence pairs that contain gold alignments can differ considerably in content and length, making them hard to align for statistical word alignment tools. The difficulty of this task can also be seen in the results for the Greedy model, which only achieves an F1-score of 17.2%. In contrast, we observe that the majority of all sure alignments can be retrieved by applying the LemmaId model (60.3% recall).
The Full model achieves a recall of 48.9% but significantly outperforms all baselines (p < 0.01) in terms of precision (71.8%). This is an important factor for us as we plan to use the alignments in subsequent steps of our framework. With 58.2%, Full also achieves the best overall F1-score. By comparing the results with those of the EMNLP’12 model, we can see that the discourse-level similarity measures provide a significant improvement in terms of precision without a considerable loss in recall. This advantage in precision can also be seen in comparison to Wolfe et al. In contrast, their system outperforms our model with respect to recall. There are two main reasons for this: On the one hand, their model makes use of much larger resources to compute alignments, including a paraphrasing database that contains over 7 million rewriting rules; on the other hand, their model is supervised and makes use of additional data to learn weights for each of their features. In contrast, Full and EMNLP’12 only make use of a small development data set to determine a threshold for graph construction. Though the difference is not significant, it is worth noting that our model outperforms that by Wolfe et al. by 0.6 percentage points in F1-score, despite not making use of any additional training data.
Ablating Similarity Measures. All aforementioned results were conducted in experiments with a uniform weighting scheme of similarity measures as introduced in Section 4.2.2. Table 2 additionally shows the performance impact of individual similarity measures by removing them completely (i.e., setting their weight to 0.0). The numbers indicate that almost all measures contribute positively to the overall performance when using equal weights. Except for the argument head similarity, all ablation tests revealed significant
drops in performance, either with respect to precision or recall. This result highlights the importance of incorporating predicate-specific, argument-specific, and discoursespecific information regarding individual predications in this task.
Tuning Weights for High Precision. Subsequently, we tested various combinations of weights on our development set in order to estimate a better weighting scheme. This tuning procedure is implemented as a grid search technique, in which random weights between 0.0 and 1.0 are assigned to each measure. For graph construction, all weights are normalized again to sum to 1.0. We additionally try different thresholds for adding edges in the graph representation. To achieve high precision, we weight precision three times higher than recall while evaluating different parameters. In total, we tested 2,000 different parameter assignments on our development set. Following this process, we found the best result to be achieved with a threshold of 0.85 and the following weights:r 0.11, 0.14, and 0.21 for simWN, simVN, and simDist, respectively,
(i.e., 46% of the total weight for predicate-specific measures)r 0.21 and 0.05 for simABoW and simAheads, respectively, (i.e., 26% of the total weight for argument-specific measures)r 0.21 and 0.07 for simDPos and simDCon, respectively (i.e., 28% of the total weight for discourse-specific measures)
The weighting scheme shows that information from all categories is considered. When applying the tuned model on our evaluation data set, we note that results in recall drop to 29.1% (−19.8 percentage points). Precision, on the other hand, increases to 86.2% (+14.4 percentage points). In this section, we introduced the task of aligning predicate–argument structures across monolingual comparable texts. We designed annotation guidelines and created a data set of gold-standard alignments. Based on this data set, we developed and evaluated a novel clustering-based alignment model that uses a combination of various similarity measures and a graph-based clustering algorithm that we specifically designed for this task. In an intrinsic evaluation, we showed that our novel model outperforms a range of baselines as well as previous approaches to this particular task. As an additional contribution, we defined a tuning routine that can be utilized to train a high precision model for the discourse-level alignment task. Our results show that, by using this tuning step, corresponding structures in our evaluation set can be identified with a precision of 86.2%. This intermediate result is essential for the success of our overall framework. In the next section, we present Step 2 of our implicit argument induction technique, in which we examine pairs of automatically aligned predicate–argument structures as a means to identify and link implicit arguments. 5 induction framework step 2: extracting implicit arguments :In the second step of our induction framework, we rely on alignments between PAS to detect implicit arguments. That is, we aim to identify argument instances that are present, or explicit, in one PAS but absent, or implicit, from the aligned PAS. Based on the identified instances, our model tries to find antecedents of implicit arguments
within their respective discourse contexts. To perform this task automatically, we rely on several preprocessing tools, which we apply on the full corpus of over 160,000 document pairs introduced in Roth and Frank (2012a).
We describe the applied preparatory steps and the used preprocessing tools in Section 5.1. Using automatically computed annotations as input, we describe a heuristic method to detect implicit arguments and their discourse antecedents in Section 5.2. As a basis for the actual induction, we rely on several preparatory steps that identify information two documents have in common (cf. Figure 2). In particular, we compute and align PAS using the graph-based model described in Section 4, and determine co-referring entities across pairs of texts using coreference resolution techniques on concatenated document pairs (Lee et al. 2012). In theory, arguments implicit in one structure can straightforwardly be induced based on this information by looking for co-referring mentions of the argument explicit in the aligned structure. In practice, we make use of additional checks and filters to ensure that only reliable information is being used. We describe the preprocessing steps in the following paragraphs and provide additional details on our implementation of the induction procedure in Section 5.2.
Single Document Preprocessing. We apply several preprocessing steps to each document in our data set. First, we use the Stanford CoreNLP package (Manning et al. 2014) for tokenization and sentence splitting. We then apply the MATE tools (Björkelund et al. 2010; Bohnet 2010), including the integrated PropBank/NomBank-based semantic parser, to determine local PAS. Finally, we resolve pronouns that occur in a PAS using the coreference resolution system by Martschat et al. (2012), which placed second for English in the CoNLL-2012 Shared Task (Pradhan et al. 2012).
High Precision Alignments. Once all single documents are preprocessed, we align PAS across pairs of comparable texts. We want to induce reliable instances of implicit arguments based on aligned PASs pairs and hence apply our graph-based clustering technique using the high-precision tuning step described in Section 4.3. We run
the high-precision model on all pairs of texts in the corpus. As a result, we extract a total number of 283,588 aligned pairs of PASs. An overview of properties of this data set is given in Table 3. We observe that most alignments involve the same part-of-speech (98.4%) and the same predicate lemma (96.6%). This fact simplifies the task of inducing implicit arguments as it implies that in most cases the PropBank argument labels in a pair of aligned structures correspond to each other and do not have to be mapped via a higher-level set of roles.
Cross-Document Coreference. For each argument that is explicit in one PAS but implicit in an aligned PAS, we want to determine a suitable antecedent within the discourse context of the implicit instance. We solve this task by viewing the aligned explicit argument as an entity reference and identify co-referring mentions in both texts by applying coreference resolution techniques across pairs of documents. In practice, we follow the methodology of Lee et al. (2012), who propose applying standard coreference methods on pairs of texts by simply concatenating two documents and providing them as a single input document. As merging two different discourses can lead to problems for a coreference system that is based on features and feature weights trained on single documents, we follow Lee et al. and apply a rule-based system. Like them, we use the Stanford Coreference system (Lee et al. 2013), which applies a sequence of coreference “sieves” to the input, ordered according to their precision. To obtain a highly accurate and reliable output, we consider only the most precise resolution sieves:r “String Match”r “Relaxed String Match”r “Precise Constructs”r “Strict Head Match A,” “Strict Head Match B,” “Strict Head Match C”r “Proper Head Noun Match” Note that none of these sieves involve pronoun resolution. Instead, we decided to use the resolved pronouns from the single-document coreference step. This decision
is based on the fact that the system by Martschat et al. (2012) outperforms the Stanford system with all sieves on the CoNLL’11 test set by an average F1-score of 3.0 absolute points. The high-precision sieves are, however, better suited for the cross-document task as we plan to rely on the resulting coreference chains for identifying potential antecedents of implicit arguments. That is, we prefer fewer but more reliable chains in order to minimize the impact of possible preprocessing errors. Given a pair of aligned predicates from two comparable texts, we examine the output of the semantic parser (cf. Section 5.1) to identify arguments in each PAS. We compare the set of labels assigned to the arguments in each structure to determine whether one PAS contains an argument (explicit) that has not been realized in the other PAS (implicit). For each implicit argument, we identify appropriate antecedents by considering the cross-document coreference chain of its explicit counterpart. As our goal is to link implicit arguments within discourse, we require candidate antecedents to be mentions that occur in the same document. We impose a number of restrictions on the resulting pairs of implicit arguments and antecedents to reduce the impact of different types of preprocessing errors:
Mislabeled Arguments. In some cases, the parser annotated the same argument in two texts using different labels. To ensure that mislabeled arguments are not recognized as implicit, we require that pairs of aligned PAS contain a different number of arguments.
Missed Arguments. Depending on sentence structure, the semantic parser is sometimes unable to determine all local arguments. This often leads to the identification of erroneous implicit arguments. To intercept some of these cases, we require that all antecedents from the cross-document coreference chain must be outside of the sentence that contains the affected PAS. We apply the outlined identification and linking approach to all text pairs in our corpus of comparable texts. As a result, we induce a total of 698 implicit argument and antecedent pairs. A summary of properties of the obtained pairs can be found in Table 4. The full data set involves predicates of 294 different lemmata. Each pair was found in a separate document. Note that 698 implicit arguments from 283,588 pairs of PAS seem to represent a fairly low recall. The reason for this is that most PAS pairs in the highprecision data set consist of identically labeled argument sets (84.5%) and in most of the remaining cases an antecedent in discourse cannot be found using the high-precision coreference sieves. This does not mean that implicit arguments are rare in general. As discussed in Section 4.1, 38.9% of all manually aligned PAS pairs involve a different number of arguments.
We manually evaluated a subset of 90 induced implicit arguments and found 80 discourse antecedents to be correct (89%). Examples are displayed in Table 5. A closer analysis revealed that some incorrectly linked instances still result from preprocessing
errors. In particular, combinations of errors can lead to incorrectly identified instances as showcased in Example 19:
Example 19 “The Guatemalan Congress on Thursday ratified 126-12 [a Central America-US]A0 [free trade]A1 agreement, lawmakers said.” Induced missing argument and discourse antecedent: [goods]A2/co-agent
Instead of recognizing Central America and US as two separate arguments (agent and co-agent) of the predicate AGREE, the semantic parser labels both entities as one argument (A0, agent); our system hence tries to determine a discourse antecedent for an argument that is predicted to be missing despite being actually realized (A2, co-agent). In the aligned PAS, the co-agent is realized as a prepositional phrase: “[with the United States]A2”. The cross-document coreference tool incorrectly predicts the United States in this phrase to be coreferent with the phrase U.S. goods and services; our system hence detects goods as the antecedent for the erroneously predicted implicit argument.
Further error sources are incorrectly extracted document pairs and alignments between PAS that do not correspond to each other. Example 20 shows two PAS from a pair of texts that describe different sets of changes in economic activity.
Example 20 “[Production]A1 rose [3.9 percent from October in the capital good sector]A2” “[Industrial production excluding energy, food and construction . . . ]A1 rose [2.6 percent]A2 [in November]TMP from the previous month”
Example 21 shows a pair of aligned structures in a pair of comparable texts that both report on two planned trips by President Obama. Here, the alignment model erroneously aligned a reference to one trip with a reference to both trips.
Example 21 “The postponement ... marks the second time [the president’s]A0 trip [to Indonesia]A1 has been postponed” “Obama was to depart on a [weeklong]TMP trip [to both countries . . . ]A1 on June 13” In this section, we introduced a computational implementation of Step 2 of our rulebased induction method for identifying implicit arguments and their discourse antecedents. Our approach depends on automatic annotations from semantic role labeling, PAS alignment, and coreference resolution. We implement two particular types of measures to minimize the impact of preprocessing errors: (1) we avoid imprecise input by applying high-precision tools instead of methods that are tuned for balanced precision and recall; and (2) we circumvent some common preprocessing errors by formulating
constraints on resulting instances of implicit arguments and discourse antecedents. An intrinsic analysis reveals that we cannot eliminate all error sources this way, although we found the induced data set to be of high precision.
The next sections present extrinsic evaluations in which we demonstrate the utility of the induced instances empirically. As discussed in Section 1, we assume that a proper treatment of implicit arguments can improve the coverage of semantic role labeling systems and entity-based models of coherence. We show how our data set can be utilized in both of these areas: In Section 6, we use our data set of implicit arguments to enhance the training process for models of implicit argument linking; in Sections 7, we demonstrate that implicit arguments can affect the perceived coherence of a text and that instances of aligned (explicit and implicit) arguments in our data set can be used to successfully predict this impact adequately. 6 task setting 1: linking implicit arguments in discourse :Our first extrinsic evaluation assesses the utility of automatically induced pairs of implicit arguments and antecedents for the task of implicit argument linking. For this scenario, we use the data sets from the SemEval 2010 task on “Linking Events and their Participants in Discourse” (Ruppenhofer et al. 2010, henceforth just SemEval task). For direct comparison with previous results and heuristic acquisition techniques, we apply the implicit argument identification and linking model by Silberer and Frank (2012, henceforth S&F) for training and testing. We briefly describe the SemEval task data and the model by S & F in the next sections. Both the training and test sets of the SemEval task are text corpora extracted from Sherlock Holmes novels, with manual frame semantic annotations including implicit arguments. In the actual linking task (“NI-only”), gold labels are provided for local arguments and participating systems have to perform the following three sub-tasks: firstly, non-instantiated roles have to be identified; secondly, they have to be classified as being “accessible to the speaker and hearer” (definite null instantiation, DNI) or as being “only existentially bound within discourse” (indefinite null instantiations, INI); and finally, all resolvable null instantiations have to be linked to discourse antecedents.
The task organizers provide two versions of their data set: one based on FrameNet annotations and one based on PropBank/NomBank annotations. We found, however, that the latter only contains a subset of the implicit argument annotations from the FrameNet-based version. As all previous results in this task have been reported on the FrameNet data set, we adopt the same setting. Note that our automatically induced data set, which we want to apply as additional training data, is labeled with a PropBank/NomBank-style parser. Consequently, we need to map our annotations to FrameNet in order to make use of them in this task. The organizers of the SemEval task provide a manual mapping dictionary for predicates in the annotated data set. We use this manual mapping and additionally use SemLink (Palmer 2009) for mapping predicates and arguments not covered by the dictionary. We make use of the system by S&F to train a new model for the NI-only task. As mentioned in the previous subsection, this task consists of three steps: In Step (1), implicit
arguments are identified as unfilled and non-redundant FrameNet core roles; in Step (2), an SVM classifier is used to predict whether implicit arguments are definite based on a small number of features—semantic type of the affected Frame Element, the relative frequency of its realization type in the SemEval training corpus, and a Boolean feature that indicates whether the affected sentence is in passive voice and does not contain a (deep) subject. In Step (3), we apply the same features and classifier as S&F to find appropriate antecedents for (predicted) definite arguments. S&F report that their best results were obtained considering the following target antecedents: all entity mentions that are syntactic constituents from the present and the past two sentences and all entity mentions that occurred at least five times in the previous discourse (“Chains+Win” setting). In their evaluation, the latter of these two restrictions crucially depended on gold coreference chains. As the automatic coreference chains in our data are rather sparse (and noisy), we only consider syntactic constituents from the present and the past two sentences as antecedents (“SentWin” setting).
Before training and testing a new model with our own data, we perform feature selection using 10-fold cross validation. To find the best set of features, we run the feature selection on a combination of the SemEval training data and our full additional data set.9 The only features that were selected in this process concern the “prominence” of the candidate antecedent, its semantic agreement with the selectional preferences of the predicate, the part-of-speech-tags used in each reference to the candidate entity, and the semantic types of all roles that the entity fills according to local role annotations. These features are a subset of the best features described in Silberer and Frank (2012). Evaluation measures. For direct comparison in the full task, both with S&F’s model and other models, we adopt the precision, recall, and F1 measures as defined in Ruppenhofer et al. (2010).
Baselines. We compare our results with those previously reported on the SemEval task (see Table 6 for a summary): The best performing system in the actual task in 2010 was developed by Chen et al. (2010) and is an adaptation of the semantic role labeling system SEMAFOR (Das et al. 2010). In 2011, Tonelli and Delmonte (2011) presented a revised version of their SemEval system (Tonelli and Delmonte 2010), which outperforms SEMAFOR in terms of recall (6%) and F1-score (8%). The best results in terms of recall and F1-score to date have been reported by Laparra and Rigau (2012), with 25% and 19%, respectively. Our model outperforms their state-of-the-art system in terms of precision (21%) but achieves a lower recall (8%). Two influencing factors for their high recall are probably (1) their improved method for identifying (resolvable) implicit arguments, and (2) their addition of lexicalized and ontological features.
Our Results. Comparison of our results with those reported by S&F, whose system we use, shows that our additional data improves precision (from 6% to 21%) and F1-score (from 7% to 12%). The loss of one percentage point in recall is marginal given the size of the test set (only 259 implicit arguments have an annotated antecedent). Our
9 Note that this feature selection procedure is the same as applied by Silberer and Frank. Hence, the evaluated models are directly comparable and all differences in results can directly be traced back to the use of additional data.
result in precision is the second highest score reported on this task. Interestingly, the improvements are higher than those achieved in the original study by Silberer and Frank (2012), even though their best additional training set is three times bigger than ours and contains manual semantic annotations.
We conjecture that their low gain in precision could be a side effect triggered by two different factors. Firstly, the heuristically created training instances, induced by treating anaphoric pronouns as implicit argument instances, might not reflect the same properties as actual implicit arguments. For example, pronouns can occur in syntactic constructions in which an actual argument cannot be omitted in practice, leading an incorrect overgeneralization. An additional negative factor might be that their model relies on coreference chains, which are automatically generated for the test set and hence rather noisy. In contrast, our automatically induced data does not contain manual annotations of semantic roles and coreference chains, hence we do not rely on gold information during training and testing. The results show that, despite this limitation, our new model outperforms previous models trained using the same system, indicating the utility and high reliability of our automatically induced data.
Impact of training data size. To assess the impact of training data size, we perform an additional experiment with subsets of automatically induced implicit arguments. Specifically, we train different classifiers using the full SemEval training data and varying amounts of our automatically induced training data (random samples of 1%, 10%, and 25%). The model uses the best feature set determined on the combination of SemEval and our full additional data set in all settings. For each setting, we report average results obtained over four runs. The outcomes of this experiment are summarized in Table 7. The numbers reveal that using only 1% of the additional data for training already leads to a classification performance of 0.13 in F1-score. The large improvement over the S&F model without additional data, which achieves an F1-score of 0.07, can be explained by the fact that the features selected by our model generalize better than those selected on the SemEval training data only. The improvements are highest when using between 10% and 25% of the additional data, indicating that the use of additional induced instances indeed increases performance but that utilizing more out-of-domain than in-domain data for training seems to be harmful.
10 Results as reported by Tonelli and Delmonte (2011). 11 Results computed as an average over the scores given for both test files; rounded towards the number
given for the test file that contained more instances. In this section, we presented a NLP application of the automatically induced data set of implicit arguments that we introduced in Section 5. We found that automatically induced implicit arguments can successfully be used as training data to improve a system for linking implicit arguments in discourse. Although the presented model cannot compete with state-of-the-art systems, the addition of our data led to an enhanced performance compared with the same system with different and without additional training data. Compared with the model without additional training data, our induced data set increased results in terms of precision and F1-score by 15 and 5 percentage points, respectively. 7 task setting 2: modeling local coherence :In our second experiment, we examine whether argument realization decisions affect the perceived coherence of a text and investigate how and which factors related to their impact can be used to model realization decisions computationally. We approach this question in the following way: We exploit PAS alignments across comparable documents to identify contexts with implicit and explicit arguments; we then make use of these automatically induced contexts in order to train a discourse coherence model that is able to predict whether—in a given context—an argument should be realized or remain implicit.
Induction of such a model and its evaluation will be approached as follows: First, we assemble a data set of document pairs that differ only with respect to a single realization decision; given each pair in this data set, we ask human annotators to indicate their preference for the implicit or explicit argument instance in the prespecified context (Section 7.1); secondly, we attempt to emulate the decision process computationally using a discriminative model based on discourse and entity-specific features (Section 7.2). To assess the performance of the new model, we train it on automatically induced training data and evaluate it, in comparison with previous models of local coherence, against human annotations (Section 7.3). We use the data set of automatically induced implicit arguments (henceforth source data), described in Section 5, as a starting point for composing a set of document pairs that involve implicit and explicit arguments. To make sure that each document pair in this data set only differs with respect to a single realization decision, we first create
two copies of each document from the source data: One copy remains in its original form, and the other copy will be modified with respect to a single argument realization. Example (22) illustrates an original and modified sentence.
Example 22 (a) [The Dalai Lama’s]A0 visit [to France]A1 ends on Tuesday.
(a‘) [The Dalai Lama’s]A0 visit ends on Tuesday.
Note that adding or removing arguments at random can lead to structures that are semantically implausible. Hence, we consider two restrictions. First, we ensure that the remaining context is still understandable after an argument is removed by only considering texts in which follow-up references can still be resolved based on earlier antecedents. Second, we only modify arguments of PAS that actually occur and are aligned across two texts. Given a pair of PAS that differ with respect to an argument realization, we create modifications by replacing the specific implicit or explicit argument in one text with the corresponding argument in the paired text. Examples (22) and (23) show two such comparable sentences. The original PAS in Example (22a) contains an explicit argument that is implicit in the aligned PAS and hence removed in the modified version. Similarly, the original text in (23a) involves an implicit argument that is made explicit in the modified version (23a‘).
Example 23 (a) [The Dalai Lama’s]A0 visit coincides with the Beijing Olympics.
(a‘) [The Dalai Lama’s]A0 visit [to France]A1 coincides with the Beijing Olympics.
We ensure that the modified structure fits into the given context grammatically by only considering pairs of PASs with identical predicate form and constituent order. We found that this restriction constrains affected arguments to be modifiers, prepositional phrases, and direct objects. We argue that this is actually a desirable property because more complicated alternations could affect coherence by themselves. In other words, resulting interplays would make it difficult to distinguish between the isolated effect of argument realization itself and other effects, triggered for example by sentence order (Gordon, Grosz, and Gilliom 1993).
Annotation. We set up a Web experiment using the evaluation toolkit by Belz and Kow (2011) to collect ratings of local coherence for implicit and explicit arguments. For this experiment, we compiled a data set of 150 document pairs. Each text in such a pair consists of the same content, with the only difference being one argument realization.
We presented all 150 pairs to two annotators12 and asked them to indicate their preference for one alternative over the other using a continuous slider scale. The annotators got to see the full texts, with the alternatives presented next to each other. To make texts easier to read and differences easier to spot, we collapsed all identical sentences into one column and highlighted the aligned predicate (in both texts) and the affected argument (in the explicit case). An example is shown in Figure 3. To avoid any bias in the annotation process, we shuffled the sequence of text pairs and randomly assigned the side of display (left/right) of each realization type (explicit/implicit). Instead of
12 Both annotators are native speakers of English and undergraduate students in literature and linguistics.
providing a definition of local coherence ourselves, we asked annotators to rate how “natural” a realization reads given the discourse context. This procedure is in line with previous work by Pitler and Nenkova (2008), who “view text readability and text coherence as equivalent properties,” given that the annotators are “competent language users.”
We found that annotators made use of the full rating scale, with the extremes indicating either a strong preference for the text on the left-hand side or the righthand side, respectively. However, most ratings were concentrated more towards the center of the scale (i.e., around zero). This seems to imply that the use of implicit or explicit arguments did not make a considerable difference most of the time. The annotators affirmed that some cases do not read naturally when a specific argument is omitted or redundantly realized at a given position in discourse. For example, the text fragment in Example (24) shows two sentences in which an argument has been realized twice, leading to a perceived redundancy in the second sentence (A4, destination); conversely, Example (25) showcases an excerpt in which a non-redundant argument (A2, co-signer) has been omitted.
Example 24 The remaining contraband was picked up at Le Havre. The containers had arrived [in Le Havre] from China.
Example 25 Lt.-Gen. Mohamed Lamari (. . . ) denied his country wanted South African weapons to fight Muslim rebels fighting the government. “We are not going to fight a flea with a hammer,” Lamari told reporters after signing the agreement of intent [∅].
We computed correlation between ratings of both annotators using Spearman’s ρ (Spearman 1904) and found a low but significant correlation (ρ = 0.22, p < 0.01). To
construct a test set with gold annotations, we mapped continuous ratings to discrete labels (implicit, explicit, neutral) and selected a subset of 70 instances for which clear13 preferences were observed. Table 8 provides some statistics on the collected data. Even though correlation with respect to continuous ratings is relatively low, we find that both annotators have the same general preference in most cases, with a raw agreement of 75% overall. Our test set contains 47 (31%) cases of explicit arguments and 23 (15%) cases of implicit arguments. Those cases in which annotators disagreed (25%) or had no preference (28%) were discarded. We model the decision process that underlies the (non-)realization of arguments using an SVM-based model and a range of discourse features. The features are based on the following three factors: the affected predicate–argument structure (Parg), the (automatic) coreference chain of the affected entity (Coref), and the discourse context (Disc).
Parg. The first group of features is concerned with the characteristics of the affected PAS: This includes the absolute and relative number of explicitly realized arguments in the structure, the number of modifiers in it, and the total length of the structure as well as of the complete sentence (in words).
Coref. The coreference-specific features include transition patterns as inspired by the entity grid model (cf. Section 3), the absolute number of previous and follow-up mentions of the (non-)realized argument, their POS tags, and the distance between the current PAS to the closest previous and follow-up mention (in number of words and sentences). In contrast to previous work on the entity grid model, we do not type transition features with respect to the grammatical function of explicit realizations. The reason for skipping this information lies in the insignificant amount of relevant samples in our training data (see the following).
Disc. On the discourse level, we define a small set of additional features that include the total number of coreference chains in the text, the occurrence of pronouns in the current sentence, lexical repetitions in the previous and follow-up sentence, the current position in discourse (begin, middle, end), and a feature indicating whether the affected argument occurred in the first sentence.
13 Absolute continuous rating of >1 standard deviation from zero.
Most of these features overlap with those successfully applied in previous work on modeling coherence. For example, the transition patterns are inspired by the entity grid model. In addition to entity-grid like features, Pitler and Nenkova (2008) also use text length, word overlap, and pronoun occurrences as features for predicting readability. Our own contribution lies in the definition of PAS-specific features and the adaptation of all features to the task of predicting (non-)realization of arguments in a PAS. In the evaluation (cf. Section 7.3), we report results for two models: A simplified model that only makes use of entity-grid like features and a full model that uses all features described here. To learn feature weights, we make use of the training data described in the following section. The goal of our model is to correctly predict the realization type (implicit or explicit) of an argument that maximizes the perceived coherence of the document. As a proxy for coherence, we use the naturalness ratings given by our annotators. We evaluate classification performance on the 70 data points in our annotated test set for which clear preferences have been established. We report results in terms of precision, recall, and F1-score per class as well as micro- and macro-averaged F1-score across classes. We compute precision as the fraction of correct classifier decisions divided by the total number of classifications made for a specific class label; recall as the fraction of correct classifier decisions divided by the total number of test items with the specific label; and F1 as the harmonic mean between precision and recall.
7.3.1 Baselines. For comparison, we apply a couple of coherence models proposed in previous work: the original entity grid model by Barzilay and Lapata (2005), a modified version that uses topic models (Elsner and Charniak 2011a), and an extended version that includes entity-specific features (Elsner and Charniak 2011b); we further apply the discourse-new model by Elsner and Charniak (2008), and the pronoun-based model by Charniak and Elsner (2009). For all of the aforementioned models, we use their respective implementation provided with the Brown Coherence Toolkit.14 Note that the toolkit only returns one coherence score for each document. To use the model scores for predicting argument realization, we use two documents per data point—one that contains the affected argument explicitly and one that does not (implicit argument)— and treat the higher scoring variant as classification output. If both documents achieve the same score, we count the test item neither as correctly nor as incorrectly classified.
7.3.2 Our Models. Like the applied baseline models, our models do not make use of any manually labeled data for training. Instead, we utilize the automatically identified instances of explicit and implicit arguments from pairs of comparable texts, which we described in Section 5. To train our own model, we prepare this data set as follows: Firstly, we remove all data points that were selected for the test set; secondly, we split all pairs of texts into two groups—texts that contain a PAS in which an implicit argument has been identified (IA), and their comparable counterparts, which contain the aligned PAS with an explicit argument (EA). All texts are labeled according to their group. For all texts in group EA, we remove the explicit argument from the aligned PAS. This way, the feature extractor always gets to see the text and automatic annotations as if
14 http://www.ling.ohio-state.edu/%7Emelsner/.
the realization decision had not been performed and can thus extract unbiased feature values for the affected entity and argument position. Given each feature representation, we train a classifier using the default parameters of the LIBSVM package (Chang and Lin 2011).15
We apply our own model on each data point in the small annotated test set, where we always treat the affected argument, regardless of its actual annotation, as implicit to extract unbiased feature values for classification. Based on the features described in Section 7.2 and trained on the automatically constructed PAS alignments, our model predicts the realization type of each argument in the given context. We note that our model has an advantage here because it is specifically designed for this task and trained on corresponding data. All models compute local coherence ratings based on entity occurrences, however, and should thus be able to predict which realization type coheres best with the given discourse context. That is, because the input document pairs are identical except for the affected argument position, the coherence scores assigned by each model to pairs of text only differ with respect to the affected entity realization.
7.3.3 Results. The results are summarized in Table 9. We begin our analysis with a discussion on results in terms of micro-averaged F1-scores. As observable from the last column in Table 9, each of the baseline models achieves an F1-score between 31% and 41%, whereas our (full) model achieves 74%. As our data set is imbalanced and biased towards explicit realizations, we also provide macro-averaged F1-scores that show average performance over both classes without taking into account their respective sizes. Here, we find that baseline results lie between 27% and 39%, whereas our (full) model achieves 69%. Before discussing the results of our model in more detail, we investigate baselines performances.
We observe that the original entity grid model exhibits a preference for the implicit realization type: It predicts this class in 61 (87%) cases, resulting in only 9% of all explicit arguments being correctly classified. Overall, the entity grid achieves a
15 The default settings in our LIBSVM version are: equal costs of both classes and use of a sigmoid kernel.
(micro-averaged) F1-score of 31%. Taking a closer look at the features of the model reveals that this is an expected outcome: In its original setting, the entity grid learns realization patterns in the form of sentence-to-sentence transitions. As discussed in the paper that introduced the entity grid model (Barzilay and Lapata 2008), most entities are, however, only mentioned a few times in a text. Consequently, non-realizations constitute the “most frequent” class, independently of whether they are relevant in a given context or not. The models by Charniak and Elsner (2009) and Elsner and Charniak (2011a), which are not based on an entity grid, suffer less from such an effect and achieve better results, with F1-scores up to 35% and 41%, respectively. The extended entity grid model also alleviates the bias towards non-realizations, resulting in a slightly improved F1-score of 33%. To abstract away from this issue, we train a simplified version of our own model that only uses features that involve entity transition patterns. The main difference between this simplified model and the original entity grid model lies in the different use of training data: whereas entity grid models treat all non-realized items equally, our model gets to “see” actual examples of entities that are implicit. In other words, our simplified model takes into account implicit mentions of entities, not only explicit ones. The results confirm that this extra information has a significant impact (p < 0.01, using a randomization test; Yeh 2000) on test set performance, and raises the ratio of correctly classified explicit arguments to 73%. Yet, the simplified model only provides a correct label for 30% of instances in the test data that are marked as implicit arguments, with a class-specific F1-score of only 42%. As demonstrated by the performance of our full model, a combination of all features is needed to achieve the best overall results of 69% and 74% in macro and micro-averaged F1-scores, respectively. Applied to the two classes separately, our model achieves an F1-score of 53% on arguments that are annotated as implicit and 82% on explicit arguments. Both of these scores are the best across all tested models.
To determine the impact of the three different feature groups, we derive the weight of each feature from the model learned by LIBSVM. Table 10 gives an overview over the ten highest weights for implicit and explicit realization classification. We use the following terminology in the feature description: “the entity” refers to the entity that is referred to by the to-be-classified argument, “next/previous mention” denotes a co-referring mention to the same entity, “the PAS” refers to the predicate–argument structure that contains the affected argument (implicitly), and “the sentence” refers to the sentence in which this PAS is realized. All “distances” refer to the number of tokens that appear between the predicate that heads the PAS and the previous or next mention of the entity. As can be seen at the top and bottom ends of Table 10, the strongest feature for classifying an argument as implicit is whether the entity is also realized in the preceding or following two sentences. The strongest feature for classifying an argument as explicit is whether the next mention is a pronoun. The trained weights indicate that the model is learning some interesting patterns that reflect rules such as “avoid close repetitions,” “keep sentences short,” and “pronouns and proper names can more often be dropped than definite noun phrases.” In this chapter, we presented a computational linguistic application of the automatically induced data set of implicit arguments that we introduced in Section 5. This data set has been induced from pairs of comparable text and is a unique resource in that it contains automatic annotations of implicit arguments, aligned explicit arguments, and discourse antecedents. In our experiments on perceived coherence, we found that the
use of implicit vs. explicit arguments, although often being a subtle difference, can have a clear impact on readability ratings by human annotators. We showed that our novel coherence model, which is solely trained on automatically induced data, is able to predict this difference in newswire articles with an F1-score of up to 74%. 8 discussion and conclusions :In this article, we introduced a framework for inducing instances of implicit arguments and their discourse antecedents from pairs of comparable texts, and showcased applications of these instances in natural language processing. As described in Section 2, our framework consists of two steps: aligning predicate–argument structures across comparable texts and identifying implicit arguments and antecedents. In the following paragraphs, we summarize our framework and highlight specific contributions.
Predicate–Argument Structure Alignments. With the goal of inducing instances of implicit arguments, we proposed a novel task that aims to align pairs of PASs across pairs of comparable texts. In Section 4, we introduced a manually annotated data set for the development and evaluation of models for this particular task. We found that pairs of PASs can be aligned across documents with good inter-annotator agreement given appropriate annotation guidelines. Based on the development part of our corpus, we designed and fine-tuned a novel graph-based clustering model. To apply this model, we represent PAS in pairs of documents as bipartite graphs and recursively divide this graph into subgraphs. All clustering decisions by the model are based on pairwise similarities between PASs, combining information on predicates, associated arguments,
and their respective discourse contexts. We empirically evaluated our model against various baselines and a competitive model that has recently been proposed in the literature. The results of our evaluation show that our model outperforms all other current models on the PAS alignment task by a margin of at least 0.6 percentage points in F1-score, despite only a single threshold parameter being adjusted on our development set. As an additional contribution, we defined a tuning procedure, in which we adjust our method for high precision. Following this tuning routine, our model is capable of aligning PAS pairs with a precision of 86.2%.
Heuristic Induction Method. Based on aligned pairs of PASs, the second step in our induction framework is to identify instances of implicit arguments. In Section 5, we described a computational implementation of this step in which aligned argument structures are automatically compared and discourse antecedents for implicit arguments are found by means of entity coreference chains across documents. To reduce the effect of preprocessing errors, our implementation makes use of precise preprocessing methods and a small set of restrictions that exclude instances whose automatic annotations are likely to be erroneous. We found that by combining information from different preprocessing modules, we can induce instances of implicit arguments and discourse antecedents with high precision.
Coherence Modeling and Implicit Argument Linking. To examine the utility and reliability of our data set, we additionally performed extrinsic evaluations in task-based settings. We described two particular applications of our data: linking implicit arguments to discourse antecedents and predicting coherent argument realizations. In the first application, described in Section 6, we used our data set as additional training data to enhance a pre-existing system for identifying and linking implicit arguments in discourse. Experimental results showed that the addition of our training data can improve performance in terms of precision (+15 percentage points) and F1-score (+5 percentage points).
For the second application, we developed a novel model of local coherence, described in Section 7, which predicts whether a specific argument should be explicitly realized at a given point in discourse or whether it can already be inferred from context. Our experiments revealed that this model, when trained on automatically induced data, can predict human judgments on argument realization in newswire text with F1-scores between 53% and 82%. In comparison, we found that entity-based coherence models from previous work only achieve results below 50%, reflecting the fact that they do not capture this phenomenon appropriately.
In conclusion, a considerable amount of work still needs to be done to enhance models for handling implicit arguments in discourse. In the long run, however, this research direction will be beneficial for many applications that involve the understanding or generation of text beyond the sentence level. In this article, we provided several research contributions that form a reliable basis for future work. In particular, we developed a framework for automatically inducing instances of implicit arguments, and we designed a novel coherence model that predicts the effect of argument realizations on perceived textual coherence. From a theoretical perspective, we validated that both explicit and implicit arguments can affect coherence and that automatically induced training data can be utilized to model this phenomenon appropriately. We further showed that our induced data set, which contains instances of implicit arguments and discourse antecedents, can be applied to enhance current models for implicit argument
linking. Future work will be able to build on these insights, further enhance existing models, and apply them to improve current state-of-the-art NLP systems.
The resources described in this article are available for download at http:// projects.cl.uni-heidelberg.de/india/. In this article, we investigate aspects of sentential meaning that are not expressed in local predicate–argument structures. In particular, we examine instances of semantic arguments that are only inferable from discourse context. The goal of this work is to automatically acquire and process such instances, which we also refer to as implicit arguments, to improve computational models of language. As contributions towards this goal, we establish an effective framework for the difficult task of inducing implicit arguments and their antecedents in discourse and empirically demonstrate the importance of modeling this phenomenon in discourse-level tasks. Our framework builds upon a novel projection approach that allows for the accurate detection of implicit arguments by aligning and comparing predicate–argument structures across pairs of comparable texts. As part of this framework, we develop a graph-based model for predicate alignment that significantly outperforms previous approaches. Based on such alignments, we show that implicit argument instances can be automatically induced and applied to improve a current model of linking implicit arguments in discourse. We further validate that decisions on argument realization, although being a subtle phenomenon most of the time, can considerably affect the perceived coherence of a text. Our experiments reveal that previous models of coherence are not able to predict this impact. Consequently, we develop a novel coherence model, which learns to accurately predict argument realization based on automatically aligned pairs of implicit and explicit arguments. [{""affiliations"": [], ""name"": ""Michael Roth""}, {""affiliations"": [], ""name"": ""Anette Frank""}] SP:e9e9e7f46c40792a25f778c0c791118d43613b2e [{""authors"": [""Arnold"", ""Jennifer E.""], ""title"": ""Reference Form and Discourse Patterns"", ""venue"": ""Ph.D. thesis, Stanford University."", ""year"": 1998}, {""authors"": [""Baroni"", ""Marco"", ""Alessandro Lenci.""], ""title"": ""Distributional memory: A general framework for corpus-based semantics"", ""venue"": ""Computational Linguistics, 36(4):673\u2013721."", ""year"": 2010}, {""authors"": [""Barzilay"", ""Regina"", ""Mirella Lapata.""], ""title"": ""Modeling local coherence: An entity-based approach"", ""venue"": ""Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics, pages 141\u2013148. Ann Arbor, MI."", ""year"": 2005}, {""authors"": [""Barzilay"", ""Regina"", ""Mirella Lapata.""], ""title"": ""Modeling local coherence: An entity-based approach"", ""venue"": ""Computational Linguistics, 34(1):1\u201334."", ""year"": 2008}, {""authors"": [""Belz"", ""Anja"", ""Eric Kow.""], ""title"": ""Discrete vs"", ""venue"": ""continuous rating scales for language evaluation in NLP. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language"", ""year"": 2011}, {""authors"": [""Bj\u00f6rkelund"", ""Anders"", ""Bernd Bohnet"", ""Love Hafdell"", ""Pierre Nugues.""], ""title"": ""A high-performance syntactic and semantic dependency parser"", ""venue"": ""COLING 2010: Demonstration Volume, pages 33\u201336, Beijing."", ""year"": 2010}, {""authors"": [""Bohnet"", ""Bernd.""], ""title"": ""Top accuracy and fast dependency parsing is not a contradiction"", ""venue"": ""Proceedings of the 23rd International Conference on Computational Linguistics, pages 89\u201397, Beijing."", ""year"": 2010}, {""authors"": [""Brockett"", ""Chris.""], ""title"": ""Aligning the RTE 2006 corpus"", ""venue"": ""Technical Report MSR-TR-2007-77, Microsoft Research."", ""year"": 2007}, {""authors"": [""Brown"", ""Cheryl.""], ""title"": ""Topic continuity in written English narrative"", ""venue"": ""Talmy Givon, editor, Topic Continuity in Discourse: A"", ""year"": 1983}, {""authors"": [""Cai"", ""Jie"", ""Michael Strube.""], ""title"": ""End-to-end coreference resolution via hypergraph partitioning"", ""venue"": ""Proceedings of the 23rd International Conference on Computational Linguistics (COLING 2010), pages 143\u2013151,"", ""year"": 2010}, {""authors"": [""Chang"", ""Chih-Chung"", ""Chih-Jen Lin.""], ""title"": ""LIBSVM: A library for support vector machines"", ""venue"": ""ACM Transactions on Intelligent Systems and Technology, 2(27):1\u201327."", ""year"": 2011}, {""authors"": [""Charniak"", ""Eugene"", ""Micha Elsner.""], ""title"": ""EM works for pronoun anaphora resolution"", ""venue"": ""Proceedings of the 12th Conference of the European Chapter of the ACL (EACL 2009), pages 148\u2013156, Athens."", ""year"": 2009}, {""authors"": [""Chen"", ""Desai"", ""Nathan Schneider"", ""Dipanjan Das"", ""Noah A. Smith.""], ""title"": ""SEMAFOR: Frame argument resolution with log-linear models"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic"", ""year"": 2010}, {""authors"": [""Chen"", ""Zheng"", ""Heng Ji.""], ""title"": ""Graph-based clustering for computational linguistics: A survey"", ""venue"": ""Proceedings of TextGraphs-5 - 2010 Workshop on Graph-based Methods for Natural Language Processing, pages 1\u20139,"", ""year"": 2010}, {""authors"": [""Cohn"", ""Trevor"", ""Chris Callison-Burch"", ""Mirella Lapata.""], ""title"": ""Constructing corpora for development and evaluation of paraphrase systems"", ""venue"": ""Computational Linguistics, 34(4):597\u2013614."", ""year"": 2008}, {""authors"": [""Das"", ""Dipanjan"", ""Nathan Schneider"", ""Desai Chen"", ""Noah A. Smith.""], ""title"": ""Probabilistic frame-semantic parsing"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American"", ""year"": 2010}, {""authors"": [""DeNero"", ""John"", ""Alexandre Bouchard-C\u00f4t\u00e9"", ""Dan Klein.""], ""title"": ""Sampling alignment structure under a Bayesian translation model"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language"", ""year"": 2008}, {""authors"": [""Di Eugenio"", ""Barbara.""], ""title"": ""Centering theory and the Italian pronominal system"", ""venue"": ""Proceedings of the 13th Conference on Computational Linguistics, pages 270\u2013275, Helsinki."", ""year"": 1990}, {""authors"": [""Elsner"", ""Micha"", ""Eugene Charniak.""], ""title"": ""Coreference-inspired coherence modeling"", ""venue"": ""Proceedings of ACL-08: HLT, Short Papers, pages 41\u201344, Columbus, OH."", ""year"": 2008}, {""authors"": [""Elsner"", ""Micha"", ""Eugene Charniak.""], ""title"": ""Disentangling chat with local coherence models"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies,"", ""year"": 2011}, {""authors"": [""Elsner"", ""Micha"", ""Eugene Charniak.""], ""title"": ""Extending the entity grid with entity-specific features"", ""venue"": ""Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human"", ""year"": 2011}, {""authors"": [""Erk"", ""Katrin"", ""Sebastian Pad\u00f3.""], ""title"": ""A structured vector space model for word meaning in context"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, Waikiki, HI."", ""year"": 2008}, {""authors"": [""Fellbaum"", ""Christiane""], ""title"": ""WordNet: An Electronic Lexical Database"", ""year"": 1998}, {""authors"": [""Feng"", ""Vanessa Wei"", ""Graeme Hirst.""], ""title"": ""Patterns of local discourse coherence as a feature for authorship attribution"", ""venue"": ""Literary and Linguistic Computing, 29:191\u2013198."", ""year"": 2014}, {""authors"": [""Filippova"", ""Katja"", ""Michael Strube.""], ""title"": ""Extending the entity-grid coherence model to semantically related entities"", ""venue"": ""Proceedings of the 11th European Workshop on Natural Language Generation,"", ""year"": 2007}, {""authors"": [""Fillmore"", ""Charles J.""], ""title"": ""Pragmatically controlled zero anaphora"", ""venue"": ""Proceedings of the Twelfth Annual Meeting of the Berkeley Linguistics Society, pages 95\u2013107, Berkeley, CA."", ""year"": 1986}, {""authors"": [""Fleiss"", ""Joseph L"", ""Bruce Levin"", ""Myunghee Cho Paik.""], ""title"": ""The measurement of interrater agreement"", ""venue"": ""Statistical Methods for Rates and Proportions, 2:212\u2013236."", ""year"": 1981}, {""authors"": [""F\u00fcrstenau"", ""Hagen"", ""Mirella Lapata.""], ""title"": ""Semi-supervised semantic role labeling via structural alignment"", ""venue"": ""Computational Linguistics, 38(1):135\u2013171."", ""year"": 2012}, {""authors"": [""Gerber"", ""Matthew"", ""Joyce Chai.""], ""title"": ""Semantic role labeling of implicit arguments for nominal predicates"", ""venue"": ""Computational Linguistics, 38(4):755\u2013798."", ""year"": 2012}, {""authors"": [""Goldberg"", ""Andrew V."", ""Robert E. Tarjan.""], ""title"": ""A new approach to the maximum flow problem"", ""venue"": ""Proceedings of the Eighteenth Annual ACM Symposium on Theory of Computing, pages 136\u2013146,"", ""year"": 1986}, {""authors"": [""Gordon"", ""Peter C."", ""Barbara J. Grosz"", ""Laura A. Gilliom.""], ""title"": ""Pronouns, names, and the centering of attention in discourse"", ""venue"": ""Cognitive Science, 17:311\u2013347."", ""year"": 1993}, {""authors"": [""Gorinski"", ""Philip"", ""Josef Ruppenhofer"", ""Caroline Sporleder.""], ""title"": ""Towards weakly supervised resolution of null instantiations"", ""venue"": ""Proceedings of the 10th International Conference on Computational"", ""year"": 2013}, {""authors"": [""Grosz"", ""Barbara J."", ""Aravind K. Joshi"", ""Scott Weinstein.""], ""title"": ""Centering: A framework for modeling the local coherence of discourse"", ""venue"": ""Computational Linguistics, 21(2):203\u2013225."", ""year"": 1995}, {""authors"": [""Guo"", ""Weiwei"", ""Mona Diab.""], ""title"": ""Semantic topic models: Combining word distributional statistics and dictionary definitions"", ""venue"": ""Proceedings of the 2011 Conference on Empirical Methods in Natural"", ""year"": 2011}, {""authors"": [""Hou"", ""Yufang"", ""Katja Markert"", ""Michael Strube.""], ""title"": ""Global inference for bridging anaphora resolution"", ""venue"": ""Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational"", ""year"": 2013}, {""authors"": [""Hwa"", ""Rebecca"", ""Philip Resnik"", ""Amy Weinberg"", ""Clara Cabezas"", ""Okan Kolak.""], ""title"": ""Bootstrapping parsers via syntactic projection across parallel texts"", ""venue"": ""Natural Language Engineering, 11(3):311\u2013325."", ""year"": 2005}, {""authors"": [""Joshi"", ""Aravind K."", ""Steve Kuhn.""], ""title"": ""Centered logic: The role of entity centered sentence representation in natural language inferencing"", ""venue"": ""Proceedings of the 6th International Joint Conference on Artificial"", ""year"": 1979}, {""authors"": [""Kameyama"", ""Megumi.""], ""title"": ""Zero Anaphora: The Case of Japanese"", ""venue"": ""Ph.D. thesis, Stanford University."", ""year"": 1985}, {""authors"": [""Karamanis"", ""Nikiforos"", ""Chris Mellish"", ""Massimo Poesio"", ""Jon Oberlander.""], ""title"": ""Evaluating centering for information ordering using corpora"", ""venue"": ""Computational Linguistics, 35(1):29\u201346."", ""year"": 2009}, {""authors"": [""Kay"", ""Martin"", ""Martin R\u00f6scheisen.""], ""title"": ""Text-translation alignment"", ""venue"": ""Computational Linguistics, 19(1):121\u2013142."", ""year"": 1993}, {""authors"": [""Kipper"", ""Karin"", ""Anna Korhonen"", ""Neville Ryant"", ""Martha Palmer.""], ""title"": ""A large-scale classification of English verbs"", ""venue"": ""Language Resources and Evaluation Journal, 42(1):21\u201340."", ""year"": 2008}, {""authors"": [""Kozhevnikov"", ""Mikhail"", ""Ivan Titov""], ""title"": ""Crosslingual transfer of semantic role"", ""year"": 2013}, {""authors"": [""T.K. Landauer"", ""S.T. Dumais.""], ""title"": ""A solution to Plato\u2019s problem: The latent semantic analysis theory of the acquisition, induction, and representation of knowledge"", ""venue"": ""Psychological Review,"", ""year"": 1997}, {""authors"": [""Laparra"", ""Egoitz"", ""German Rigau.""], ""title"": ""Exploiting explicit annotations and semantic types for implicit argument resolution"", ""venue"": ""Proceedings of the Sixth IEEE International Conference on Semantic"", ""year"": 2012}, {""authors"": [""Laparra"", ""Egoitz"", ""German Rigau.""], ""title"": ""Sources of evidence for implicit argument resolution"", ""venue"": ""Proceedings of the 10th International Conference on Computational Semantics (IWCS 2013) \u2013 Long Papers,"", ""year"": 2013}, {""authors"": [""Lee"", ""Heeyoung"", ""Angel Chang"", ""Yves Peirsman"", ""Nathanael Chambers"", ""Mihai Surdeanu"", ""Dan Jurafsky""], ""title"": ""Deterministic coreference resolution based on entity-centric, precision-ranked rules"", ""year"": 2013}, {""authors"": [""Lee"", ""Heeyoung"", ""Marta Recasens"", ""Angel Chang"", ""Mihai Surdeanu"", ""Dan Jurafsky.""], ""title"": ""Joint entity and event coreference resolution across documents"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical"", ""year"": 2012}, {""authors"": [""Liang"", ""Percy"", ""Benjamin Taskar"", ""Dan Klein.""], ""title"": ""Alignment by agreement"", ""venue"": ""North American Association for Computational Linguistics (NAACL), pages 104\u2013111, New York, NY."", ""year"": 2006}, {""authors"": [""Lin"", ""Dekang.""], ""title"": ""An information-theoretic definition of similarity"", ""venue"": ""Proceedings of the 15th International Conference on Machine Learning, pages 296\u2013304, Madison, WI."", ""year"": 1998}, {""authors"": [""Louis"", ""Annie"", ""Ani Nenkova.""], ""title"": ""Creating local coherence: An empirical assessment"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association"", ""year"": 2010}, {""authors"": [""Manning"", ""Christopher"", ""Mihai Surdeanu"", ""John Bauer"", ""Jenny Finkel"", ""Steven Bethard"", ""David McClosky.""], ""title"": ""The Stanford corenlp natural language processing toolkit"", ""venue"": ""Proceedings of the 52nd Annual"", ""year"": 2014}, {""authors"": [""Martschat"", ""Sebastian"", ""Jie Cai"", ""Samuel Broscheit"", ""\u00c9va M\u00fajdricza-Maydt"", ""Michael Strube.""], ""title"": ""A multigraph model for coreference resolution"", ""venue"": ""Joint Conference on EMNLP and CoNLL - Shared"", ""year"": 2012}, {""authors"": [""McIntyre"", ""Neil"", ""Mirella Lapata.""], ""title"": ""Plot induction and evolutionary search for story generation"", ""venue"": ""Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1562\u20131572,"", ""year"": 2010}, {""authors"": [""Meyers"", ""Adam.""], ""title"": ""Annotation guidelines for NomBank\u2013noun argument structure for PropBank"", ""venue"": ""Technical report, New York University. Available from http://nlp. cs.nyu.edu/meyers/nombank/nombank-"", ""year"": 2007}, {""authors"": [""Mitchell"", ""Jeff"", ""Mirella Lapata.""], ""title"": ""Composition in distributional models of semantics"", ""venue"": ""Cognitive Science, 34(8):1388\u20131429."", ""year"": 2010}, {""authors"": [""Moor"", ""Tatjana"", ""Michael Roth"", ""Anette Frank.""], ""title"": ""Predicate-specific annotations for implicit role binding: Corpus annotation, data analysis and evaluation experiments"", ""venue"": ""Proceedings of the 10th"", ""year"": 2013}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Mirella Lapata.""], ""title"": ""Dependency-based construction of semantic space models"", ""venue"": ""Computational Linguistics, 33(2):161\u2013199."", ""year"": 2007}, {""authors"": [""Pad\u00f3"", ""Sebastian"", ""Mirella Lapata.""], ""title"": ""Cross-lingual annotation projection of semantic roles"", ""venue"": ""Journal of Artificial Intelligence Research, 36:307\u2013340."", ""year"": 2009}, {""authors"": [""Palmer"", ""Martha.""], ""title"": ""Semlink: Linking PropBank, VerbNet and FrameNet"", ""venue"": ""Proceedings of the Generative Lexicon Conference, pages 9\u201315, Pisa."", ""year"": 2009}, {""authors"": [""Palmer"", ""Martha"", ""Daniel Gildea"", ""Paul Kingsbury.""], ""title"": ""The Proposition Bank: An annotated corpus of semantic roles"", ""venue"": ""Computational Linguistics, 31(1):71\u2013106."", ""year"": 2005}, {""authors"": [""Palmer"", ""Martha S."", ""Deborah A. Dahl"", ""Rebecca J. Schiffman"", ""Lynette Hirschman"", ""Marcia Linebarger"", ""John Dowding.""], ""title"": ""Recovering implicit information"", ""venue"": ""Proceedings of the 24th Annual Meeting"", ""year"": 1986}, {""authors"": [""Parker"", ""Robert"", ""David Graff"", ""Jumbo Kong"", ""Ke Chen"", ""Kazuaki Maeda.""], ""title"": ""English Gigaword Fifth Edition"", ""venue"": ""Linguistic Data Consortium, Philadelphia, PA."", ""year"": 2011}, {""authors"": [""Pedersen"", ""Ted.""], ""title"": ""Information content measures of semantic similarity perform better without sense-tagged text"", ""venue"": ""Human Language Technologies: The 2010 Annual Conference of the North American Chapter of"", ""year"": 2010}, {""authors"": [""Pitler"", ""Emily"", ""Ani Nenkova.""], ""title"": ""Revisiting readability: A unified framework for predicting text quality"", ""venue"": ""Proceedings of the 2008 Conference on Empirical Methods in Natural Language"", ""year"": 2008}, {""authors"": [""Poesio"", ""Massimo"", ""Rosemary Stevenson"", ""Barbara Di Eugenio"", ""Janet Hitzeman.""], ""title"": ""Centering: A parametric theory and its instantiations"", ""venue"": ""Computational Linguistics, 30(3):309\u2013363."", ""year"": 2004}, {""authors"": [""Postolache"", ""Oana"", ""Dan Cristea"", ""Constantin Orasan.""], ""title"": ""Transferring coreference chains through word alignment"", ""venue"": ""Proceedings of the Fifth International Conference on Language"", ""year"": 2006}, {""authors"": [""Pradhan"", ""Sameer"", ""Alessandro Moschitti"", ""Nianwen Xue"", ""Olga Uryupina"", ""Yuchen Zhang.""], ""title"": ""CoNLL-2012 shared task: Modeling multilingual unrestricted coreference in OntoNotes"", ""venue"": ""Joint"", ""year"": 2012}, {""authors"": [""Pradhan"", ""Sameer S."", ""Lance Ramshaw"", ""Ralph Weischedel"", ""Jessica MacBride"", ""Linnea Micciulla.""], ""title"": ""Unrestricted coreference: Identifying entities and events in OntoNotes"", ""venue"": ""Proceedings of the First IEEE"", ""year"": 2007}, {""authors"": [""Recasens"", ""Marta"", ""Marta Vila.""], ""title"": ""Squibs: On paraphrase and coreference"", ""venue"": ""Computational Linguistics, 36(4):639\u2013647."", ""year"": 2010}, {""authors"": [""Roth"", ""Michael"", ""Anette Frank.""], ""title"": ""Aligning predicate argument structures in monolingual comparable texts: A new corpus for a new task"", ""venue"": ""Proceedings of the First Joint Conference on Lexical and"", ""year"": 2012}, {""authors"": [""Roth"", ""Michael"", ""Anette Frank.""], ""title"": ""Aligning predicates across monolingual comparable texts using graph-based clustering"", ""venue"": ""Proceedings of the 2012 Joint Conference on Empirical Methods in Natural"", ""year"": 2012}, {""authors"": [""Ruppenhofer"", ""Josef"", ""Russell Lee-Goldman"", ""Caroline Sporleder"", ""Roser Morante.""], ""title"": ""Beyond sentence-level semantic role labeling: linking argument structures in discourse"", ""venue"": ""Language Resources and"", ""year"": 2012}, {""authors"": [""Ruppenhofer"", ""Josef"", ""Caroline Sporleder"", ""Roser Morante"", ""Collin Baker"", ""Martha Palmer.""], ""title"": ""SemEval-2010 Task 10: Linking events and their participants in discourse"", ""venue"": ""Proceedings of the 5th"", ""year"": 2010}, {""authors"": [""Sidner"", ""Candace L.""], ""title"": ""Towards a computational theory of definite anaphora comprehension in English"", ""venue"": ""Technical Report AI-Memo 537, Massachusetts Institute of Technology, AI Lab,"", ""year"": 1979}, {""authors"": [""Silberer"", ""Carina"", ""Anette Frank.""], ""title"": ""Casting implicit role linking as an anaphora resolution task"", ""venue"": ""Proceedings of the First Joint Conference on Lexical and Computational Semantics (*SEM 2012),"", ""year"": 2012}, {""authors"": [""Spearman"", ""Charles.""], ""title"": ""The proof and measurement of association between two things"", ""venue"": ""The American Journal of Psychology, 15(1):72\u2013101."", ""year"": 1904}, {""authors"": [""Sporleder"", ""Caroline"", ""Alex Lascarides.""], ""title"": ""Using automatically labelled examples to classify rhetorical relations: A critical assessment"", ""venue"": ""Natural Language Engineering, 14(3):369\u2013416."", ""year"": 2008}, {""authors"": [""Spreyer"", ""Kathrin"", ""Anette Frank.""], ""title"": ""Projection-based acquisition of a temporal labeller"", ""venue"": ""Proceedings of the Third International Joint Conference on Natural Language Processing, pages 489\u2013496,"", ""year"": 2008}, {""authors"": [""Su"", ""Fangzhong"", ""Katja Markert.""], ""title"": ""Subjectivity recognition on word senses via semi-supervised mincuts"", ""venue"": ""Proceedings of Human Language Technologies: The 2009 Annual Conference of the North"", ""year"": 2009}, {""authors"": [""Tonelli"", ""Sara"", ""Rodolfo Delmonte.""], ""title"": ""VENSES++: Adapting a deep semantic processing system to the identification of null instantiations"", ""venue"": ""Proceedings of the 5th International Workshop on Semantic"", ""year"": 2010}, {""authors"": [""Tonelli"", ""Sara"", ""Rodolfo Delmonte""], ""title"": ""Desperately seeking implicit arguments"", ""year"": 2011}, {""authors"": [""Turan"", ""\u00dcmit Deniz.""], ""title"": ""Null vs"", ""venue"": ""Overt Subjects in Turkish: A Centering Approach. Ph.D. thesis, University of Pennsylvania."", ""year"": 1995}, {""authors"": [""van der Plas"", ""Lonneke"", ""Paola Merlo"", ""James Henderson""], ""title"": ""Scaling up automatic cross-lingual semantic role annotation"", ""venue"": ""In Proceedings of the 49th Annual Meeting of the Association"", ""year"": 2011}, {""authors"": [""Walker"", ""Christopher"", ""Stephanie Strassel"", ""Julie Medero"", ""Kazuaki Maeda.""], ""title"": ""ACE 2005 Multilingual Training Corpus"", ""venue"": ""Linguistic Data Consortium, Philadelphia, PA."", ""year"": 2006}, {""authors"": [""Wan"", ""Stephen"", ""Mark Dras"", ""Robert Dale"", ""Cecile Paris.""], ""title"": ""Using dependency-based features to take the \u201cPara-farce\u201d out of paraphrase"", ""venue"": ""Proceedings of the Australasian Language"", ""year"": 2006}, {""authors"": [""Xuchen Yao""], ""title"": ""PARMA: A predicate"", ""year"": 2013}] acknowledgments  :We thank our annotators in Heidelberg and Edinburgh. We are grateful to the anonymous reviewers for helpful feedback and suggestions. The research leading to these results has received funding by the Landes-graduiertenförderung BadenWürttemberg within the research initiative “Coherence in language processing” at Heidelberg University. Work by M. R. at the University of Edinburgh was funded by a DFG Research Fellowship (RO 4848/1-1). a framework and its applications :Michael Roth∗ University of Edinburgh
Anette Frank∗∗ Heidelberg University
In this article, we investigate aspects of sentential meaning that are not expressed in local predicate–argument structures. In particular, we examine instances of semantic arguments that are only inferable from discourse context. The goal of this work is to automatically acquire and process such instances, which we also refer to as implicit arguments, to improve computational models of language. As contributions towards this goal, we establish an effective framework for the difficult task of inducing implicit arguments and their antecedents in discourse and empirically demonstrate the importance of modeling this phenomenon in discourse-level tasks.
Our framework builds upon a novel projection approach that allows for the accurate detection of implicit arguments by aligning and comparing predicate–argument structures across pairs of comparable texts. As part of this framework, we develop a graph-based model for predicate alignment that significantly outperforms previous approaches. Based on such alignments, we show that implicit argument instances can be automatically induced and applied to improve a current model of linking implicit arguments in discourse. We further validate that decisions on argument realization, although being a subtle phenomenon most of the time, can considerably affect the perceived coherence of a text. Our experiments reveal that previous models of coherence are not able to predict this impact. Consequently, we develop a novel coherence model, which learns to accurately predict argument realization based on automatically aligned pairs of implicit and explicit arguments.","8 discussion and conclusions :In this article, we introduced a framework for inducing instances of implicit arguments and their discourse antecedents from pairs of comparable texts, and showcased applications of these instances in natural language processing. As described in Section 2, our framework consists of two steps: aligning predicate–argument structures across comparable texts and identifying implicit arguments and antecedents. In the following paragraphs, we summarize our framework and highlight specific contributions.
Predicate–Argument Structure Alignments. With the goal of inducing instances of implicit arguments, we proposed a novel task that aims to align pairs of PASs across pairs of comparable texts. In Section 4, we introduced a manually annotated data set for the development and evaluation of models for this particular task. We found that pairs of PASs can be aligned across documents with good inter-annotator agreement given appropriate annotation guidelines. Based on the development part of our corpus, we designed and fine-tuned a novel graph-based clustering model. To apply this model, we represent PAS in pairs of documents as bipartite graphs and recursively divide this graph into subgraphs. All clustering decisions by the model are based on pairwise similarities between PASs, combining information on predicates, associated arguments,
and their respective discourse contexts. We empirically evaluated our model against various baselines and a competitive model that has recently been proposed in the literature. The results of our evaluation show that our model outperforms all other current models on the PAS alignment task by a margin of at least 0.6 percentage points in F1-score, despite only a single threshold parameter being adjusted on our development set. As an additional contribution, we defined a tuning procedure, in which we adjust our method for high precision. Following this tuning routine, our model is capable of aligning PAS pairs with a precision of 86.2%.
Heuristic Induction Method. Based on aligned pairs of PASs, the second step in our induction framework is to identify instances of implicit arguments. In Section 5, we described a computational implementation of this step in which aligned argument structures are automatically compared and discourse antecedents for implicit arguments are found by means of entity coreference chains across documents. To reduce the effect of preprocessing errors, our implementation makes use of precise preprocessing methods and a small set of restrictions that exclude instances whose automatic annotations are likely to be erroneous. We found that by combining information from different preprocessing modules, we can induce instances of implicit arguments and discourse antecedents with high precision.
Coherence Modeling and Implicit Argument Linking. To examine the utility and reliability of our data set, we additionally performed extrinsic evaluations in task-based settings. We described two particular applications of our data: linking implicit arguments to discourse antecedents and predicting coherent argument realizations. In the first application, described in Section 6, we used our data set as additional training data to enhance a pre-existing system for identifying and linking implicit arguments in discourse. Experimental results showed that the addition of our training data can improve performance in terms of precision (+15 percentage points) and F1-score (+5 percentage points).
For the second application, we developed a novel model of local coherence, described in Section 7, which predicts whether a specific argument should be explicitly realized at a given point in discourse or whether it can already be inferred from context. Our experiments revealed that this model, when trained on automatically induced data, can predict human judgments on argument realization in newswire text with F1-scores between 53% and 82%. In comparison, we found that entity-based coherence models from previous work only achieve results below 50%, reflecting the fact that they do not capture this phenomenon appropriately.
In conclusion, a considerable amount of work still needs to be done to enhance models for handling implicit arguments in discourse. In the long run, however, this research direction will be beneficial for many applications that involve the understanding or generation of text beyond the sentence level. In this article, we provided several research contributions that form a reliable basis for future work. In particular, we developed a framework for automatically inducing instances of implicit arguments, and we designed a novel coherence model that predicts the effect of argument realizations on perceived textual coherence. From a theoretical perspective, we validated that both explicit and implicit arguments can affect coherence and that automatically induced training data can be utilized to model this phenomenon appropriately. We further showed that our induced data set, which contains instances of implicit arguments and discourse antecedents, can be applied to enhance current models for implicit argument
linking. Future work will be able to build on these insights, further enhance existing models, and apply them to improve current state-of-the-art NLP systems.
The resources described in this article are available for download at http:// projects.cl.uni-heidelberg.de/india/."