[{"abstract":"Semantic parsing methods are used for capturing and representing semantic meaning of text. Meaning representation capturing all the concepts in the text may not always be available or may not be sufficiently complete. Ontologies provide a structured and reasoning-capable way to model the content of a collection of texts. In this work, we present a novel approach to joint learning of ontology and semantic parser from text. The method is based on semi-automatic induction of a context-free grammar from semantically annotated text. The grammar parses the text into semantic trees. Both, the grammar and the semantic trees are used to learn the ontology on several levels -- classes, instances, taxonomic and non-taxonomic relations. The approach was evaluated on the first sentences of Wikipedia pages describing people.","figures_and_tables":[{"caption":"Figure 1. The proposed approach to ontology and grammar induction gets a collection of text and the seed ontology as input and outputs a new ontology. The collection of text is first annotated at different levels including annotations with the concepts from existing ontology. The annotated text is then used for grammar induction, where the text is represented as semantic trees. These are used together with the grammar to induce a new ontology.","file":"3-Figure1-1.png"},{"caption":"Figure 2. Grammar induction","file":"4-Figure2-1.png"},{"caption":"Table 1. Layered representation of a sentence. Null tokens are expressed as ”-”.","file":"5-Table1-1.png"},{"caption":"Figure 3. An example of a semantic tree, which is the result of parsing the input from Table 1. The tree is fully parsed. Each node has three rows: the class, the rule and the term interpreted in the lexical layer.","file":"7-Figure3-1.png"},{"caption":"Table 2. Two step generalization of term from Table 1 (Phil Madiera is a musician from Nashville). The layer list was constructed in a reverse order from Table 1. In this example, the rule ::= Profession from Location is the only rule that was used in the bottom-up parsing.","file":"9-Table2-1.png"},{"caption":"Figure 4. Given the relation (r a1 a2) and semantic tree on 4a, and relation (r b1 b2) and semantic tree on 4b, the same variable sub-tree is extracted (4c).","file":"11-Figure4-1.png"},{"caption":"Table 3. A sample of induced rules.","file":"14-Table3-1.png"},{"caption":"Figure 5. Statistics of the iterative process","file":"15-Figure5-1.png"},{"caption":"Table 5. Instance relations per category.","file":"15-Table5-1.png"},{"caption":"Table 6. Performance of various relation extraction models.","file":"16-Table6-1.png"},{"caption":"Figure 6. The automaton model for institution relation. On the bottom of each node, the fraction of training variable trees that contain this node, is displayed.","file":"17-Figure6-1.png"}],"full_text":[{"paragraphs":["One of the ultimate goals of Natural Language Processing (NLP) is machine reading BIBREF0 , the automatic, unsupervised understanding of text. One way of pursuing machine reading is by semantic parsing, which transforms text into its meaning representation. However, capturing the meaning is not the final goal, the meaning representation needs to be predefined and structured in a way that supports reasoning. Ontologies provide a common vocabulary for meaning representations and support reasoning, which is vital for understanding the text. To enable flexibility when encountering new concepts and relations in text, in machine reading we want to be able to learn and extend the ontology while reading. Traditional methods for ontology learning BIBREF1 , BIBREF2 are only concerned with discovering the salient concepts from text. Thus, they work in a macro-reading fashion BIBREF3 , where the goal is to extract facts from a large collection of texts, but not necessarily all of them, as opposed to a micro-reading fashion, where the goal is to extract every fact from the input text. Semantic parsers operate in a micro-reading fashion. Consequently, the ontologies with only the salient concepts are not enough for semantic parsing. Furthermore, the traditional methods learn an ontology for a particular domain, where the text is used just as a tool. On the other hand, ontologies are used just as tool to represent meaning in the semantic parsing setting. When developing a semantic parser it is not trivial to get the best meaning representation for the observed text, especially if the content is not known yet. Semantic parsing datasets have been created by either selecting texts that can be expressed with a given meaning representation, like Free917 dataset BIBREF4 , or by manually deriving the meaning representation given the text, like Atis dataset BIBREF5 . In both datasets, each unit of text has its corresponding meaning representation. While Free917 uses Freebase BIBREF6 , which is a very big multi-domain ontology, it is not possible to represent an arbitrary sentence with Freebase or any other existing ontology.","In this paper, we propose a novel approach to joint learning of ontology and semantic parsing, which is designed for homogeneous collections of text, where each fact is usually stated only once, therefore we cannot rely on data redundancy. Our approach is text-driven, semi-automatic and based on grammar induction. It is presented in Figure 1 .The input is a seed ontology together with text annotated with concepts from the seed ontology. The result of the process is an ontology with extended instances, classes, taxonomic and non-taxonomic relations, and a semantic parser, which transform basic units of text, i.e sentences, into semantic trees. Compared to trees that structure sentences based on syntactic information, nodes of semantic trees contain semantic classes, like location, profession, color, etc. Our approach does not rely on any syntactic analysis of text, like part-of-speech tagging or dependency parsing. The grammar induction method works on the premise of curriculum learning BIBREF7 , where the parser first learns to parse simple sentences, then proceeds to learn more complex ones. The induction method is iterative, semi-automatic and based on frequent patterns. A context-free grammar (CFG) is induced from the text, which is represented by several layers of semantic annotations. The motivation to use CFG is that it is very suitable for the proposed alternating usage of top-down and bottom-up parsing, where new rules are induced from previously unparsable parts. Furthermore, it has been shown by BIBREF8 that CFGs are expressive enough to model almost every language phenomena. The induction is based on a greedy iterative procedure that involves minor human involvement, which is needed for seed rule definition and rule categorization. Our experiments show that although the grammar is ambiguous, it is scalable enough to parse a large dataset of sentences.","The grammar and semantic trees serve as an input for the new ontology. Classes, instances and taxonomic relations are constructed from the grammar. We also propose a method for discovering less frequent instances and their classes, and a supervised method to learn relations between instances. Both methods work on semantic trees.","For experimentation, first sentences of Wikipedia pages describing people are taken as a dataset. These sentences are already annotated with links to other pages, which are also instances of DBpedia knowledge base BIBREF9 . Using relations from DBpedia as a training set, several models to predict relations have been trained and evaluated.","The rest of the paper is organized in the following way. The grammar induction approach is presented in Section \"Grammar induction\" . The ontology induction approach follows in Section \"Ontology induction\" . In Section \"Experiments\" we present the conducted experiments with grammar induction, and instance and relation extraction. We examine the related work in Section \"Related Work\" , and conclude with the discussion in Section \"Discussion\" ."],"section_name":"Introduction"},{"paragraphs":["In this section, we propose a semi-automatic bootstrapping procedure for grammar induction, which searches for the most frequent patterns and constructs new production rules from them. One of the main challenges is to make the induction in a way that minimizes human involvement and maximizes the quality of semantic trees.","The input to the process, which is illustrated in Figure 2 , is a set of predefined seed grammar rules (see Section \"Seed rules\" ) and a sample of sentences in a layered representation (see Section \"Experiments\" ) from the dataset. The output of the process is a larger set of rules forming the induced grammar. One rule is added to the grammar on each iteration. At the beginning of each iteration all the sentences are parsed with a top-down parser. The output of parsing a single sentence is a semantic tree – a set of semantic nodes connected into a tree. Here we distinguish two possible outcomes of the parsing: 1) the sentence was completely parsed, which is the final goal and 2) there is at least one part of the sentence that cannot be parsed. From the perspective of a parser the second scenario happens when there is a node that cannot be parsed by any of the rules. We name these nodes – null nodes – and they serve as the input for the next step, the rule induction. In this step several rules are constructed by generalization of null nodes. The generalization (see Section \"Rule induction\" ) is based on utilization of semantic annotations and bottom-up composition of the existing rules. Out of the induced rules, a rule with the highest frequency (the one that was generalized from the highest number of null nodes) is added to the grammar. To improve quality of the grammar, the rules are marked by so called property, which instructs the parser how to use the rule (eg., us it in parsing but not in induction). The property vitally affects result of the parsing in the following iterations potentially causing a huge semantic drift for the rest of process. Consequently, a human user needs to mark the property of each rule. The iterative process runs until a predefined stopping criteria is met. The criteria is either connected to the quality of the grammar or time limitation.","For the sake of transparency of the experiments, the human is involved in the beginning, when the seed rules are created and later when the rule properties are specified. However, in another setting the user could also define new rules in the middle of the bootstrapping procedure.","In the following sections, we describe each component of the process in more details.","Our goal was to develop a semi-automatic method that induces a grammar suitable for our scenario, in which an ontology is extracted, and text is parsed into semantic trees. A survey by BIBREF27 compares several papers on grammar induction. According to their classification, our method falls into unsupervised, text-based (no negative examples of sentences) methods. Many such methods induce context-free grammars. However, their focus is more on learning syntactic structures rather than semantic. This is evident in evaluation strategies, where their parse trees are compared against golden parse trees in treebanks, like Penn treebank BIBREF28 , which are annotated according to syntactic policies. Furthermore, our grammar should not limited to a specific form, like for instance Chomsky normal form or Greibach normal form, instead it may contain arbitrary context-free rules. Several algorithms, like ours, employ the greedy strategy of grammar induction, where the grammar is updated with the best decision at each step. Whereas our method adds a rule after all sentences are parsed, The Incremental Parsing algorithm BIBREF29 updates the grammar after each sentence. This is also done in ADIOS method BIBREF30 , where it has been shown that order of sentences affects the grammar. Our method employs frequency analysis and human supervision to control the grammar construction, while others use Minimum Description Length principle BIBREF31 , clustering of sequences BIBREF32 , or significance of word co-occurrences BIBREF33 ."],"section_name":"Grammar induction"},{"paragraphs":["The input textual data needs to be properly structured in order to work best with the proposed algorithms. Shallow NLP tools, like sentence splitting, word tokenization, named entity recognition, might help obtaining this structure. The basic unit is a sentence, represented by several layers. An example is presented in Table 1 . Each layer consists of several tokens, which span over one or more words. The basic layer is the lexical layer, where each token represents a single word. All other layers are created from the annotations. Some annotations, like named-entities, may span over several words; some of the words may not have an annotation, thus they are given a null token. It is crucial that all algorithms are aware how to deal with a particular layer. For instance, the parser must not break apart a multi-word annotation. Some layers may be derived from others using the seed ontology. For example, instance layer contains annotations to instances of the ontology and the derived class layer represents the classes of these annotations, which are also from the ontology. Annotation layers are valuable if they provide good means for generalization or connection with the ontology. A term is a subpart of the sentence, defined by the starting and ending position in the sentence. It has different interpretation in each layer. If the interpretation breaks any of the tokens, it is not valid. For instance, term representing Madeira is not valid in named-entity layer in Table 1 because it breaks Person."],"section_name":"Textual data representation"},{"paragraphs":["Our context-free grammar $G$ is defined by the 5-tuple: $G = (V, \\sigma , P, S, R)$ , where"," $V$ is a set of non-terminals. Each non-terminal represents a semantic class, e.g. $\\langle \\text{Person} \\rangle $ , $\\langle \\text{Color} \\rangle $ , $\\langle \\text{Organization} \\rangle $ . There is also a universal non-terminal $\\langle * \\rangle $ , which can be replaced by any other non-terminal. The same non-terminal replaces all occurrences in a rule. It is used to represent several rules, with a notation. The grammar is still context-free. See seed rule examples in Section \"Seed rules\" ."," $\\sigma $ is a set of terminals. Terminal is any existing non-null token from any sentence layer. We denote a terminal by value{layer}.","For instance, [location]{named-entity}, Phil_Madeira{instance}. If the terminal is from the lexical layer, the layer is skipped in the denotation."," $P$ is a set of production rules that represents a relation from $V \\rightarrow (V \\cup E)^*$ . For example,"," $S$ is the starting non-terminal symbol. Since non-terminals represent semantic classes, the starting symbol is chosen based on the semantic class of the input examples. If the input examples are sentences, then the appropriate category may be $\\langle \\text{Relation} \\rangle $ . While if the input examples are noun phrases, the starting symbol may be a more specific category, like $\\langle \\text{Job Title} \\rangle $ ."," $R$ is a set of properties: positive, neutral, negative, non-inducible. The property controls the usage of the rule in the parsing and in the rule induction phase. More details are given in the following subsections."],"section_name":"Grammar Definition"},{"paragraphs":["For parsing, a recursive descent parser with backtracking was developed. This is a top-down parser, which first looks at the higher level sentence structure and then proceeds down the parse tree to identify low level details of the sentence. The advantage of top-down parsing is the ability to partially parse sentences and to detect unparsable parts of sentences.","The parser takes a layered sentence as an input and returns a semantic tree as an output (see Figure 3 ). The recursive structure of the program closely follows the structure of the parse tree. The recursive function Parse (see Algorithm \"Parser\" ) takes a term and a non-terminal as input and returns a parse node as an output. The parse node contains the class of node (non-terminal), the rule that parsed the node, the term, and the list of children nodes. In order for the rule to parse the node, the left-hand side must match the input non-terminal and the right-hand side must match the layered input. In the pattern matching function Match (line \"Parser\" ), the right hand side of a rule is treated like a regular expression; non-terminals present the ( $+$ ) wildcard characters, which match at least one word. The terminals are treated as literal characters, which are matched against the layer that defines them. The result of successfully matched pattern is a list of terms, where each term represents a non-terminal of the pattern. Due to ambiguity of pattern matching there might be several matches. For each of the term – non-terminal pair in every list the parse function is recursively called (line \"Parser\" ).","GetEligibleRulesGetEligibleRules CreateNodeNode MergeSelectBestNode MatchMatch ParseParse","Grammargrammar Patternpattern NonTerminalsnon terminals","Rulesrules ChildTreechild node","Unexpandedinduction nodes","AListambiguous lists PListterm list Childchild nodes Finalfinal node Ambiguousnodes Inputinput Outputoutput","FnFunction","Phrase $p$ , Non-terminal $n$ parse node"," $\\leftarrow $ $\\lbrace \\rbrace $ rule $r$ of n = left side of $r$ $\\leftarrow $ right hand side of $r$ $\\leftarrow $ , $p$ ","of $\\leftarrow $ $\\lbrace \\rbrace $ $i\\leftarrow 0$ size of $\\leftarrow $ $_i$ , $._i$ add to add $type$ , $p$ , $r$ , to","is empty $\\leftarrow $ $type$ , $p$ , null, $\\lbrace \\rbrace $ $\\leftarrow $ $\\operatornamewithlimits{arg\\,max}_{n \\in nodes} r(n)$ is not fully parsed add to Pseudocode of the main function parse of the top-down parser.","Since the grammar is ambiguous, a term can be parsed in multiple ways. There are two types of ambiguity. Two or more rules can expand the same term and one rule can expand the term in more than one way. For each ambiguity one node is created, and the best node according to reliability measure is selected to be the result (line \"Parser\" ). The reliability measure $r(n)$ is ","$$r(n)=\n{\\left\\lbrace \\begin{array}{ll}\n1, & \\text{if node is fully parsed} \\\\\n\\beta \\cdot (1 -tp(n)) + (1 - \\beta )\\frac{\\displaystyle \\sum \\limits _{c \\in C(n)} |c|\\cdot r(c)}{\\displaystyle \\sum \\limits _{c \\in C(n)} |c|} ,& \\text{if node is partially parsed} \\\\\n0, & \\text{if node is null} \\\\\n\\end{array}\\right.}$$ (Eq. 14) ","where $tp(n)$ is the trigger probability of the rule that parsed the node $n$ , $\\beta $ is a predefined weight, $C(n)$ is the set of children of $n$ , and $|c|$ is the length of the term of node $c$ . The trigger probability of the rule is the probability that a the right-hand side of the rule pattern matches a random term in the dataset and it is estimated after the rule is induced. The range of the measure is between 0 and 1. The measure was defined in such a way that the more text the node parses, the higher is the reliability (the second summand in the middle row of Eq. 14 ). On the other hand, nodes with rules that are more frequently matched have lower reliability; this penalizes rules that are very loosely defined (the first summand in the middle row of Eq. 14 ). The $\\beta $ parameter was set to 0.05, using grid search, with average F1 score from relation extraction experiment from Section \"Relation extraction\" as a performance measure.","If none of the rules match the term, a null node is created and added to the list of nodes, which will be later used for grammar induction (line \"Parser\" ). Note that even if a null node is discarded, because it is not the most reliable, it will still be used in the grammar induction step. A node is fully parsed if the node itself and all of its descendants are parsed. If a node is parsed and if at least one of its descendants is not parsed, then the node is partially parsed. All nodes that are not fully parsed are added to the list for induction.","Since the ambiguity of the grammar may make parsing computationally infeasible, several optimization techniques are used. Memoization BIBREF10 is used to reduce the complexity from exponential time to $\\mathcal {O}(n^3)$ BIBREF11 , where $n$ is the length of the sentence. The parser does not support $\\epsilon $ productions mainly because the grammar induction will not produce them. The patterns that do not contain terminals are the most ambiguous. At most two non-terminals are allowed, and the maximal length of the term that corresponds to the first non-terminal is three tokens. We argue that this is not a huge limitation, since the way human languages are structured, usually two longer terms are connected with a word, like comma or a verb. Furthermore, the way how our induction works, these connectors do not get generalized and become a terminal in the rule. There was an attempt to introduce rules with negative property. Whenever such rule fully parses a node, that indicates that the current parsing path is incorrect. This allows the parser to backtrack sooner and also prevents adding null sister nodes (null sister nodes are in this case usually wrong) to the rule induction. However, it turned out that negative rules actually slow down the parsing, since the grammar gets bigger. It is better to mark these rules as neutral, therefore they are not added to the grammar."],"section_name":"Parser"},{"paragraphs":["The goal of the rule induction step is to convert the null nodes from the parsing step into rules. Out of these rules, the most frequent one is promoted. The term from the null node is generalized to form the right side of the rule. The class non-terminal of the null node will present the left side of the rule. Recently induced rule will parse all the nodes, from which it was induced, in the following iterations. Additionally, some rules may parse the children of those nodes.","Generalization is done in two steps. First, terms are generalized on the layer level. The output of this process is a sequence of tokens, which might be from different layers. For each position in the term a single layer is selected, according to predefined layer order. In the beginning, term is generalized with the first layer. All the non-null tokens from this layer are taken to be part of the generalized term. All the positions of the term that have not been generalized are attempted to be generalized with the next layer, etc. The last layer is without null-tokens, therefore each position of the term is assigned a layer. Usually, this is the lexical layer. For example, top part of Table 2 shows generalization of term from Table 1 . The layer list is constructed manually. Good layers for generalization are typically those that express semantic classes of individual terms. Preferably, these types are not too general (loss of information) and not too specific (larger grammar).","In the next step of generalization, tokens are further generalized using a greedy bottom-up parser using the rules from the grammar. The right sides of all the rules are matched against the input token term. If there is a match, the matched sub-term is replaced with the left side of the rule. Actually, in each iteration all the disjunct matches are replaced. To get only the disjunct matches, overlapping matches are discarded greedily, where longer matches have the priority. This process is repeated until no more rules match the term. An example is presented in the lower part of Table 2 .","The bottom-up parsing algorithm needs to be fast because the number of unexpanded nodes can be very high due to ambiguities in the top-down parsing. Consequently, the algorithm is greedy, instead of exhaustive, and yields only one result. Aho-Corasick string matching algorithm BIBREF12 is selected for matching for its ability to match all the rules simultaneously. Like the top-down parser, this parser generates partial parses because the bottom-up parser will never fully parse – the output is the same as the non-terminal type in the unexpanded node. This would generate a cyclical rule, i.e. $<$ Class $>$ :== $<$ Class $>$ . However, this never happens because the top-down parser would already expand the null node.","The last step of the iteration is assigning the property to the newly induced rule. Property controls the role of the rule in the parsing and induction. The default property is positive, which defines the default behavior of the rule in all procedures. Rules with neutral property are not used in any procedure. They also cannot be re-induced. Some rules are good for parsing, but may introduce errors in the induction. These rules should be given non-inducible property. For instance, rule $<$ Date $>$ :== $<$ Number $>$ is a candidate for the non-inducible property, since years are represented by a single number. On the contrary, not every number is a date.","In our experiments, the assignment was done manually. The human user sees the induced rule and few examples of the null nodes, from which it was induced. This should provide enough information for the user to decide in a few seconds, which property to assign. After the stopping criteria is met, the iterative procedure can continue automatically by assigning positive property to each rule. Initial experimenting showed that just a single mistake in the assignment can cause a huge drift, making all further rules wrong."],"section_name":"Rule induction"},{"paragraphs":["Before the start, a list of seed rules may be needed in order for grammar induction to be successful. Since this step is done manually, it is reasonable to have a list of seed rules short and efficient. Seed rules can be divided in three groups: domain independent linguistic rules, class rules, top-level domain rules. Domain independent linguistic rules, such as","parse the top and mid-level nodes. They can be applied on many different datasets. Class rules connect class tokens, like named-entity tokens with non-terminals. For example,","They parse the leaf nodes of the trees. On the other hand, top-level domain rules, define the basic structure of the sentence. For example,","As the name suggests, they parse nodes close to the root. Altogether, these rule groups parse on all levels of the tree, and may already be enough to parse the most basic sentences, but more importantly, they provide the basis for learning to parse more complex sentences.","The decision on which and how many seed rules should be defined relies on human judgment whether the current set of seed rules is powerful enough to ignite the bootstrapping procedure. This judgment may be supported by running one iteration and inspecting the top induced rules."],"section_name":"Seed rules"},{"paragraphs":["This section describes how to utilize the grammar and manipulate semantic trees to discover ontology components in the textual data."],"section_name":"Ontology induction"},{"paragraphs":["We propose a procedure for mapping grammar components to ontology components. In particular, classes, instances and taxonomic relations are extracted.","First, we distinguish between instances and classes in the grammar. Classes are represented by all non-terminals and terminals that come from a layer populated with classes, for example, named-entity layer and class layer from Table 1 . Instances might already exist in the instance layer, or they are created from rules, whose right hand side contains only tokens from the lexical layer. These tokens represent the label of the new instance. For instance rule $<$ Profession $>$ ::= software engineer is a candidate for instance extraction.","Furthermore, we distinguish between class and instance rules. Class rules have a single symbol representing a class on the right-hand side. Class rules map to subClassOf relations in the ontology. If the rule is positive, then the class on the right side is the subclass of the class on the left side. For instance, rule $<$ Organization $>$ ::= $<$ Company $>$ yields relation (subClassOf Company Organization).","On the other hand, instance rules have one or more symbols representing an instance on the right side, and define the isa relation. If the rule is positive, then the instance on the right side is a member of a class on the left side. For instance, rule $<$ Profession $>$ ::= software engineer yields relation (isa SoftwareEngineer Profession). If class or instance rule is neutral then the relation can be treated as false. Note that many other relations may be inferred by combing newly induced relations and relations from the seed ontology. For instance, induced relation (subClassOf new-class seed-class) and seed relation (isa seed-class seed-instance) are used to infer a new relation (isa new-class seed-instance).","In this section, we described how to discover relations on the taxonomic level. In the next section, we describe how to discover relations between instances."],"section_name":"Ontology induction from grammar"},{"paragraphs":["We propose a method for learning relations from semantic trees, which tries to solve the same problem as the classical relation extraction methods. Given a dataset of positive relation examples that represent one relation type, e.g. birthPlace, the goal is to discover new unseen relations.","The method is based on the assumption that a relation between entities is expressed in the shortest path between them in the semantic tree BIBREF13 . The input for training are sentences in layered representation, corresponding parse trees, and relation examples. Given a relation from the training set, we first try to identify the sentence containing each entity of the relation. The relation can have one, two, or even more entities. Each entity is matched to the layer that corresponds to the entity type. For example, strings are matched to the lexical layer; ontology entities are matched to the layer containing such entities. The result of a successfully matched entity is a sub-term of the sentence. In the next step, the corresponding semantic tree is searched for a node that contains the sub-term.","At this point, each entity has a corresponding entity node. Otherwise, the relation is discarded from the learning process. Given the entity nodes, a minimum spanning tree containing all off them is extracted. If there is only one entity node, then the resulting subtree is the path between this node and the root node. The extracted sub-tree is converted to a variable tree, so that different semantic trees can have the same variable sub-trees, for example see Figure 4 . The semantic nodes of the sub-tree are converted into variable nodes, by retaining the class and the rule of the node, as well as the places of the children in the original tree. For entity nodes also the position in the relation is memorized. Variable tree extracted from a relation is a positive example in the training process. For negative examples all other sub-trees that do not present any relations are converted to variable trees. Each variable node represents one feature. Therefore, a classification algorithm, such as logistic regression can be used for training.","When predicting, all possible sub-trees of the semantic tree are predicted. If a sub-tree is predicted as positive, then the terms in the leaf nodes represent the arguments of the relation."],"section_name":"Relation extraction from semantic trees"},{"paragraphs":["In this section, we present experiments evaluating the proposed approach. We have conducted experimentation on Wikipedia–DBpedia dataset (Section \"Datasets\" ). First, we have induced a grammar on the Wikipedia dataset (Section \"Grammar Induction Experiments\" ) to present its characteristics, and the scalability of the approach. In the next experiment, we present a method for discovering less prominent instances (Section \"Instance extraction\" ). The last experiment demonstrates one application of semantic parsing – the supervised learning of DBpedia relations(Section \"Relation extraction\" )."],"section_name":"Experiments"},{"paragraphs":["The datasets for experiments were constructed from English Wikipedia and knowledge bases DBpedia BIBREF9 and Freebase BIBREF6 . DBpedia provides structured information about Wikipedia articles that was scraped out of their infoboxes. First sentences of Wikipedia pages describing people were taken as the textual dataset, while DBpedia relations expressing facts about the same people were taken as the dataset for supervised relation learning. Note that each DBpedia instance has a Wikipedia page. A set of person instances was identified by querying DBpedia for instances that have a person class. For the textual dataset, Wikipedia pages representing these entities were parsed by the in-house Wikipedia markup parser to convert the markup into plain text. Furthermore, the links to other Wikipedia pages were retained. Here is an example of a sentence in plain text:","Victor Francis Hess (24 June 1883 – 17 December 1964) was an Austrian-American physicist, and Nobel laureate in physics, who discovered cosmic rays.","Using the Standford OpenNLP BIBREF14 on plain texts we obtained sentence and token splits, and named-entity annotation. Notice, that only the first sentence of each page was retained and converted to the proposed layered representation (see Section \"Experiments\" ). The layered representation contains five layers: lexical (plain text), named-entity (named entity recognizer), wiki-link (Wikipedia page in link – DBpedia instance), dbpedia-class (class of Wikipedia page in Dbpedia) and freebase-class (class of Wikipedia page in Freebase). Freebase also contains its own classes of Wikipedia pages. For the last two layers, there might be several classes per Wikipedia page. Only one was selected using a short priority list of classes. If none of the categories is on the list, then the category is chosen at random. After comparing the dbpedia-class and freebase-class layers, only freebase-class was utilized in the experiments because more wiki-link tokens has a class in freebase-class layer than in dbpedia-class layer.","There are almost 1.1 million sentences in the collection. The average length of a sentence is 18.3 words, while the median length is 13.8 words. There are 2.3 links per sentence.","The dataset for supervised relation learning contains all relations where a person instance appears as a subject in DBpedia relation. For example,","dbpedia:Victor_Francis_Hess dbpedia-owl:birthDate 1883-06-24","There are 119 different relation types (unique predicates), having from just a few relations to a few million relations. Since DBpedia and Freebase are available in RDF format, we used the RDF store for querying and for storage of existing and new relations."],"section_name":"Datasets"},{"paragraphs":["The grammar was induced on 10.000 random sentences taken from the dataset described in Section \"Datasets\" . First, a list of 45 seed nodes was constructed. There were 22 domain independent linguistic rules, 17 category rules and 6 top-level rules. The property assignment was done by the authors. In every iteration, the best rule is shown together with the number of nodes it was induced from, and ten of those nodes together with the sentences they appear in. The goal was set to stop the iterative process after two hours. We believe this is the right amount of time to still expect quality feedback from a human user.","There were 689 new rules created. A sample of them is presented in Table 3 . Table 4 presents the distributions of properties. Around $36 \\%$ of rules were used for parsing (non neutral rules). Together with the seed rules there are 297 rules used for parsing. Different properties are very evenly dispersed across the iterations. Using the procedure for conversion of grammar rules into taxonomy presented in Section \"Ontology induction\" , 33 classes and subClassOf relations, and 95 instances and isa relations were generated.","The grammar was also tested by parsing a sample of 100.000 test sentences. A few statistic are presented in Table 4 . More than a quarter of sentences were fully parsed, meaning that they do not have any null leaf nodes. Coverage represents the fraction of words in a sentence that were parsed (words that are not in null-nodes). The number of operations shows how many times was the Parse function called during the parsing of a sentences. It is highly correlated with the time spend for parsing a sentence, which is on average 0.16ms. This measurement was done on a single CPU core. Consequently, it is feasible to parse a collection of a million sentences, like our dataset. The same statistics were also calculated on the training set, the numbers are very similar to the test set. The fully parsed % and coverage are even slightly lower than on the test set. Some of the statistics were calculated after each iteration, but only when a non neutral rule was created. The graphs in Figure 5 show how have the statistics changed over the course of the grammar induction. Graph 5 shows that coverage and the fraction of fully parsed sentences are correlated and they grow very rapidly at the beginning, then the growth starts to slow down, which indicates that there is a long tail of unparsed nodes/sentences. In the following section, we present a concept learning method, which deals with the long tail. Furthermore, the number of operations per sentence also slows down (see Graph 5 ) with the number of rules, which gives a positive sign of retaining computational feasibility with the growth of the grammar. Graph 5 somewhat elaborates the dynamics of the grammar induction. In the earlier phase of induction many rules that define the upper structure of the tree are induced. These rules can rapidly increase the depth and number of null nodes, like rule 1 and rule 2 . They also explain the spikes on Graph 5 . Their addition to the grammar causes some rules to emerge on the top of the list with a significantly higher frequency. After these rules are induced the frequency gets back to the previous values and slowly decreases over the long run."],"section_name":"Grammar Induction Experiments"},{"paragraphs":["In this section, we present an experiment with a method for discovering new instances, which appear in the long tail of null nodes. Note that the majority of the instances were already placed in the ontology by the method in Section \"Ontology induction from grammar\" . Here, less prominent instances are extracted to increase the coverage of semantic parsing. The term and the class of the null node will form an isa relation. The class of the node represents the class of the relation. The terms are converted to instances. They are first generalized on the layer level (see Section \"Experiments\" ). The goal is to exclude non-atomic terms, which do not represent instances. Therefore, only terms consisting of one wiki-link token or exclusively of lexical tokens are retained. The relations were sorted according to their frequency. We observe that accuracy of the relations drops with the frequency. Therefore, relations that occurred less than three times were excluded. The number and accuracy for six classes is reported in Table 5 . Other classes were less accurate. For each class, the accuracy was manually evaluated on a random sample of 100 instance relations. Taking into account the estimated accuracy, there were more than 13.000 correct isa relations."],"section_name":"Instance extraction"},{"paragraphs":["In this section, we present an experiment of the relation extraction methods presented in Section \"Relation extraction from semantic trees\" . The input for the supervision is the DBpedia relation dataset from Section \"Datasets\" . The subject (first argument) of every relation is a person DBpedia instance – person Wikipedia page. In the beginning, the first sentence of that Wikipedia page has been identified in the textual dataset. If the object (last argument) of this relation matches a sub-term of this sentence, then the relation is eligible for experiments. We distinguish three types of values in objects. DBpedia resources are matched with wiki-link layer. Dates get converted to the format that is used in English Wikipedia. They are matched against the lexical layer, and so are the string objects.","Only relation types that have 200 or more eligible relations have been retained. This is 74 out of 119 relations. The macro average number of eligible relations per relation type is 17.7%. While the micro average is 23.8%, meaning that roughly a quarter of all DBpedia person relations are expressed in the first sentence of their Wikipedia page. For the rest of this section, all stated averages are micro-averages.","The prediction problem is designed in the following way. Given the predicate (relation type) and the first argument of the relation (person), the model predicts the second argument of the relation (object). Because not all relations are functional, like for instance child relation, there can be several values per predicate–person pair; on average there are 1.1. Since only one argument of the relation is predicted, the variable trees presented in Section \"Relation extraction from semantic trees\" , will be paths from the root to a single node. Analysis of variable tree extraction shows that on average 60.8% of eligible relations were successfully converted to variable trees (the object term exactly matches the term in the node). Others were not converted because 8.2% of the terms were split between nodes and 30.9% terms are sub-terms in nodes instead of complete terms. Measuring the diversity of variable trees shows that a distinct variable tree appeared 2.7 times on average.","Several models based on variable trees were trained for solving this classification problem:","Basic (Basic model) – The model contains positive trained variable trees. In the prediction, if the test variable tree matches one of the trees in the model, then the example is predicted positive.","Net (Automaton model) – All positive variable trees are paths with start and end points. In this model they are merged into a net, which acts as a deterministic automaton. If the automaton accepts the test variable tree, than it is predicted positive. An example of automaton model is presented in Figure 6 .","LR (Logistic regression) – A logistic regression model is trained with positive and negative examples, where nodes in variable trees represents features.","LRC (Logistic regression + Context nodes) – All leaf nodes that are siblings of any of the nodes in the variable tree are added to the LR model.","LRCL (Logistic regression + Context nodes + Lexical Tokens) – Tokens from the lexical layer of the entity nodes are added to the LRC as features.","For training all or a maximum of 10.000 eligible relations was taken for each of 74 relation types. A 10-fold cross validation was performed for evaluation. The results are presented in Table 6 . The converted recall and converted F1 score presents recall and F1 on converted examples, which are the one, where relations were successfully converted into variable trees. The performance increases with each model, however the interpretability decreases. We also compared our method to the conditional random fields(CRF). In the CRF method, tokens from all layers with window size 7 were taken as features for sequence prediction. On the converted examples CRF achieved F1 score of 80.8, which is comparable to our best model's (LRCL) F1 score of 80.0."],"section_name":"Relation extraction"},{"paragraphs":["There are many known approaches to ontology learning and semantic parsing, however, to the best of our knowledge, this is the first work to jointly learn an ontology and semantic parser. In the following sections, we make comparisons to other work on semantic parsing, ontology learning, grammar induction and others."],"section_name":"Related Work"},{"paragraphs":["The goal of semantic parsing is to map text to meaning representations. Several approaches have used Combinatory categorial grammar (CCG) and lambda calculus as a meaning representation BIBREF15 , BIBREF16 . CCG grammar closely connects syntax and semantics with a lexicon, where each entry consist of a term, a syntactical category and a lambda statement. Similarly, our context-free grammar contains production rules. Some of these rules do not contain lexical tokens (the grammar is not lexicalized), which gives ability to express some relations with a single rule. For instance, to parse jazz drummer, rule $<$ Musician_Type $>$ ::= $<$ Musical_Genre $>$ $<$ Musician_Type $>$ is used to directly express the relation, which determines the genre of the musician. Lambda calculus may provide a more formal meaning representation than semantic trees, but the lexicon of CCG requires mappings to lambda statements. Other approaches use dependency-based compositional semantics BIBREF17 , ungrounded graphs BIBREF18 , etc. as meaning representations.","Early semantic parsers were trained on datasets, such as Geoquery BIBREF19 and Atis BIBREF5 , that map sentences to domain-specific databases. Later on datasets for question answering based on Freebase were created – Free917 BIBREF4 and WebQuestions BIBREF20 These datasets contain short questions from multiple domains, and since the meaning representations are formed of Freebase concepts, they allow reasoning over Freebase's ontology, which is much richer than databases in GeoQuery and Atis. All those datasets were constructed by either forming sentences given the meaning representation or vice-versa. Consequently, systems that were trained and evaluated on these datasets, might not work on sentences that cannot be represented by the underlying ontology. To overcome this limitation BIBREF16 developed a open vocabulary semantic parser. Their approach uses a CCG parser on questions to from labmda statements, which besides Freebase vocabulary contain underspecified predicates. These lambda statements are together with answers – Freebase entities – used to learn a low-dimensional probabilistic database, which is then used to answer fill-in-the-blank natural language questions. In a very similar fashion, BIBREF21 defines underspecified entities, types and relations, when the corresponding concept does not exist in Freebase. In contrast, the purpose of our method is to identify new concepts and ground them in the ontology."],"section_name":"Semantic parsing"},{"paragraphs":["Many ontology learning approaches address the same ontology components as our approach. However, their goal is to learn only the salient concepts for a particular domain, while our goal is to learn all the concepts (including instances, like particular organizations), so that they can be used in the meaning representation. As survey by BIBREF22 summarizes, the learning mechanisms are based either on statistics, linguistics, or logic. Our approach is unique because part of our ontology is constructed from the grammar. Many approaches use lexico-syntactic patterns for ontology learning. These are often based on dependency parses, like in BIBREF2 , BIBREF23 . Our approach does not rely on linguistic preprocessing, which makes it suitable for non-standard texts and poorly resourced languages. Our approach also build patterns, however in form of grammar rules. Instead of lexico-syntactic patterns, which contain linguistic classes, our approach models semantic patterns, which contain semantic classes, like Person and Color. These patterns are constructed in advance, which is sometimes difficult because the constructor is not always aware of all the phenomena that is expressed in the input text. Our approach allows to create a small number of seed patterns in advance, then explore other patterns through process of grammar learning. A similar bootstrapping semi-automatic approach to ontology learning was developed in BIBREF24 , where the user validates lexicalizations of a particular relation to learn new instances, and in BIBREF25 , where the user validates newly identified terms, while in our approach the user validates grammar rules to learn the composition of whole sentences. A similar approach with combining DBpedia with Wikipedia for superised learning has been taken in BIBREF26 , however their focus is more on lexicalization of relations and classes."],"section_name":"Ontology Learning"},{"paragraphs":["Related work linking short terms to ontology concepts BIBREF34 is designed similarly as our approach in terms of bootstrapping procedure to induce patterns. But instead of inducing context-free grammar production rules, suggestions for rewrite rules that transform text directly to ontology language are provided. Another bootstrapping semi-automatic approach was developed for knowledge base population BIBREF35 . The task of knowledge base population is concerned only with extracting instances and relations given the ontology. In our work we also extract the backbone of the ontology – classes and taxonomic relations. Also, many other approaches focus only on one aspect of knowledge extraction, like taxonomy extraction BIBREF36 , BIBREF37 or relation extraction BIBREF13 , BIBREF38 . Combining these approaches can lead to cumbersome concept matching problems. This problem was also observed by BIBREF39 . Their system OntoUSP tries to overcome this by unsupervised inducing and populating a probabilistic grammar to solve question answering problem. However, the result are logical-form clusters connected in an isa hierarchy, not grounded concepts, which are connected with an existing ontology."],"section_name":"Other Approaches"},{"paragraphs":["We have presented an approach for joint ontology learning and semantic parsing. The approach was evaluated by building an ontology representing biographies of people. The first sentences of person Wikipedia pages and the combination of DBpedia and Freebase were used as a dataset. This dataset was suitable for our approach, because the text is equipped with human tagged annotations, which are already linked to the ontology. In other cases a named entity disambiguation would be needed to obtain the annotations. The next trait of the dataset, that is suitable for our approach, is the homogeneous style of writing. Otherwise, if the style was more heterogeneous, the users would have to participate in more iterations to achieve the same level of coverage. The participation of the users may be seen a cost, but on the other hand it allows them to learn about the dataset without reading it all. The users does not learn so much about specific facts as they learn about the second order information, like what types of relations are expressed and their distribution.","Semantic trees offer a compact tree-structured meaning representation, which could be exploited for scenarios not covered by this paper, like relation type discovery and question answering. Furthermore, they can be used for more interpretable representation of meaning, like the automaton representation in Figure 6 , compared to some other methods, like the one based on neural networks BIBREF40 . Our approach may not be superior on one specific part of the ontology learning, but it rather provides an integrated approach for learning on several levels of the ontology. Also, our approach does not use syntactic analysis, like part of speech tags or dependency parsing, which makes our approach more language independent and useful for non-standard texts, where such analysis is not available. On the other hand, we are looking into integrating syntactic analysis for future work. One scenario is to automatically detect the property of the rule. Another idea for future work is to integrate some ideas from other grammar induction methods to detect meaningful patterns without relying on the annotation of text.","This work was supported by Slovenian Research Agency and the ICT Programme of the EC under XLike (FP7-ICT-288342-STREP) and XLime (FP7-ICT-611346)."],"section_name":"Discussion"}],"paper_id":"1601.00901","title":"Joint learning of ontology and semantic parser from text"},{"abstract":"We describe a question answering model that applies to both images and structured knowledge bases. The model uses natural language strings to automatically assemble neural networks from a collection of composable modules. Parameters for these modules are learned jointly with network-assembly parameters via reinforcement learning, with only (world, question, answer) triples as supervision. Our approach, which we term a dynamic neural model network, achieves state-of-the-art results on benchmark datasets in both visual and structured domains.","figures_and_tables":[{"caption":"Figure 1: A learned syntactic analysis (a) is used to assemble a collection of neural modules (b) into a deep neural network (c), and applied to a world representation (d) to produce an answer.","file":"1-Figure1-1.png"},{"caption":"Figure 2: Simple neural module networks, corresponding to the questions What color is the bird? and Are there any states? (a) A neural find module for computing an attention over pixels. (b) The same operation applied to a knowledge base. (c) Using an attention produced by a lower module to identify the color of the region of the image attended to. (d) Performing quantification by evaluating an attention directly.","file":"2-Figure2-1.png"},{"caption":"Figure 3: Generation of layout candidates. The input sentence (a) is represented as a dependency parse (b). Fragments of this dependency parse are then associated with appropriate modules (c), and these fragments are assembled into full layouts (d).","file":"5-Figure3-1.png"},{"caption":"Table 1: Results on the VQA test server. NMN is the parameter-tying model from Andreas et al. (2015), and D-NMN is the model described in this paper.","file":"7-Table1-1.png"},{"caption":"Figure 4: Sample outputs for the visual question answering task. The second row shows the final attention provided as input to the top-level describe module. For the first two examples, the model produces reasonable parses, attends to the correct region of the images (the ear and the woman’s clothing), and generates the correct answer. In the third image, the verb is discarded and a wrong answer is produced.","file":"7-Figure4-1.png"},{"caption":"Table 2: Results on the GeoQA dataset, and the GeoQA dataset with quantification. Our approach outperforms both a purely logical model (LSP-F) and a model with learned perceptual predicates (LSP-W) on the original dataset, and a fixedstructure NMN under both evaluation conditions.","file":"8-Table2-1.png"},{"caption":"Figure 5: Example layouts and answers selected by the model on the GeoQA dataset. For incorrect predictions, the correct answer is shown in parentheses.","file":"8-Figure5-1.png"}],"full_text":[{"paragraphs":["This paper presents a compositional, attentional model for answering questions about a variety of world representations, including images and structured knowledge bases. The model translates from questions to dynamically assembled neural networks, then applies these networks to world representations (images or knowledge bases) to produce answers. We take advantage of two largely independent lines of work: on one hand, an extensive literature on answering questions by mapping from strings to logical representations of meaning; on the other, a series of recent successes in deep neural models for image recognition and captioning. By constructing neural networks instead of logical forms, our model leverages the best aspects of both linguistic compositionality and continuous representations.","Our model has two components, trained jointly: first, a collection of neural “modules” that can be freely composed (fig:teasera); second, a network layout predictor that assembles modules into complete deep networks tailored to each question (fig:teaserb). Previous work has used manually-specified modular structures for visual learning BIBREF0 . Here we:","Training data consists of (world, question, answer) triples: our approach requires no supervision of network layouts. We achieve state-of-the-art performance on two markedly different question answering tasks: one with questions about natural images, and another with more compositional questions about United States geography."],"section_name":"Introduction"},{"paragraphs":["We begin with a high-level discussion of the kinds of composed networks we would like to learn.","Andreas15NMN describe a heuristic approach for decomposing visual question answering tasks into sequence of modular sub-problems. For example, the question What color is the bird? might be answered in two steps: first, “where is the bird?” (fig:examplesa), second, “what color is that part of the image?” (fig:examplesc). This first step, a generic module called find, can be expressed as a fragment of a neural network that maps from image features and a lexical item (here bird) to a distribution over pixels. This operation is commonly referred to as the attention mechanism, and is a standard tool for manipulating images BIBREF1 and text representations BIBREF2 .","The first contribution of this paper is an extension and generalization of this mechanism to enable fully-differentiable reasoning about more structured semantic representations. fig:examplesb shows how the same module can be used to focus on the entity Georgia in a non-visual grounding domain; more generally, by representing every entity in the universe of discourse as a feature vector, we can obtain a distribution over entities that corresponds roughly to a logical set-valued denotation.","Having obtained such a distribution, existing neural approaches use it to immediately compute a weighted average of image features and project back into a labeling decision—a describe module (fig:examplesc). But the logical perspective suggests a number of novel modules that might operate on attentions: e.g. combining them (by analogy to conjunction or disjunction) or inspecting them directly without a return to feature space (by analogy to quantification, fig:examplesd). These modules are discussed in detail in sec:model. Unlike their formal counterparts, they are differentiable end-to-end, facilitating their integration into learned models. Building on previous work, we learn behavior for a collection of heterogeneous modules from (world, question, answer) triples.","The second contribution of this paper is a model for learning to assemble such modules compositionally. Isolated modules are of limited use—to obtain expressive power comparable to either formal approaches or monolithic deep networks, they must be composed into larger structures. fig:examples shows simple examples of composed structures, but for realistic question-answering tasks, even larger networks are required. Thus our goal is to automatically induce variable-free, tree-structured computation descriptors. We can use a familiar functional notation from formal semantics (e.g. Liang et al., 2011) to represent these computations. We write the two examples in fig:examples as"," (describe[color] find[bird])","and"," (exists find[state])","respectively. These are network layouts: they specify a structure for arranging modules (and their lexical parameters) into a complete network. Andreas15NMN use hand-written rules to deterministically transform dependency trees into layouts, and are restricted to producing simple structures like the above for non-synthetic data. For full generality, we will need to solve harder problems, like transforming What cities are in Georgia? (fig:teaser) into"," (and"," find[city]"," (relate[in] lookup[Georgia]))","In this paper, we present a model for learning to select such structures from a set of automatically generated candidates. We call this model a dynamic neural module network."],"section_name":"Deep networks as functional programs"},{"paragraphs":["There is an extensive literature on database question answering, in which strings are mapped to logical forms, then evaluated by a black-box execution model to produce answers. Supervision may be provided either by annotated logical forms BIBREF3 , BIBREF4 , BIBREF5 or from (world, question, answer) triples alone BIBREF6 , BIBREF7 . In general the set of primitive functions from which these logical forms can be assembled is fixed, but one recent line of work focuses on inducing new predicates functions automatically, either from perceptual features BIBREF8 or the underlying schema BIBREF9 . The model we describe in this paper has a unified framework for handling both the perceptual and schema cases, and differs from existing work primarily in learning a differentiable execution model with continuous evaluation results.","Neural models for question answering are also a subject of current interest. These include approaches that model the task directly as a multiclass classification problem BIBREF10 , models that attempt to embed questions and answers in a shared vector space BIBREF11 and attentional models that select words from documents sources BIBREF2 . Such approaches generally require that answers can be retrieved directly based on surface linguistic features, without requiring intermediate computation. A more structured approach described by Yin15NeuralTable learns a query execution model for database tables without any natural language component. Previous efforts toward unifying formal logic and representation learning include those of Grefenstette13Logic, Krishnamurthy13CompVector, Lewis13DistributionalLogical, and Beltagy13Markov.","The visually-grounded component of this work relies on recent advances in convolutional networks for computer vision BIBREF12 , and in particular the fact that late convolutional layers in networks trained for image recognition contain rich features useful for other vision tasks while preserving spatial information. These features have been used for both image captioning BIBREF1 and visual QA BIBREF13 .","Most previous approaches to visual question answering either apply a recurrent model to deep representations of both the image and the question BIBREF14 , BIBREF15 , or use the question to compute an attention over the input image, and then answer based on both the question and the image features attended to BIBREF13 , BIBREF16 . Other approaches include the simple classification model described by Zhou15ClassVQA and the dynamic parameter prediction network described by Noh15DPPVQA. All of these models assume that a fixed computation can be performed on the image and question to compute the answer, rather than adapting the structure of the computation to the question.","As noted, Andreas15NMN previously considered a simple generalization of these attentional approaches in which small variations in the network structure per-question were permitted, with the structure chosen by (deterministic) syntactic processing of questions. Other approaches in this general family include the “universal parser” sketched by Bottou14Reasoning, the graph transformer networks of Bottou97GraphTransformers, the knowledge-based neural networks of Towell94KBNN and the recursive neural networks of Socher13CVG, which use a fixed tree structure to perform further linguistic analysis without any external world representation. We are unaware of previous work that simultaneously learns both parameters for and structures of instance-specific networks."],"section_name":"Related work"},{"paragraphs":["Recall that our goal is to map from questions and world representations to answers. This process involves the following variables:","Our model is built around two distributions: a layout model $p(z|x;\\theta _\\ell )$ which chooses a layout for a sentence, and a execution model $p_z(y|w;\\theta _e)$ which applies the network specified by $z$ to $w$ .","For ease of presentation, we introduce these models in reverse order. We first imagine that $z$ is always observed, and in sec:model:modules describe how to evaluate and learn modules parameterized by $\\theta _e$ within fixed structures. In sec:model:assemblingNetworks, we move to the real scenario, where $z$ is unknown. We describe how to predict layouts from questions and learn $\\theta _e$ and $\\theta _\\ell $ jointly without layout supervision."],"section_name":"Model"},{"paragraphs":["Given a layout $z$ , we assemble the corresponding modules into a full neural network (fig:teaserc), and apply it to the knowledge representation. Intermediate results flow between modules until an answer is produced at the root. We denote the output of the network with layout $z$ on input world $w$ as $\\llbracket z \\rrbracket _w$ ; when explicitly referencing the substructure of $z$ , we can alternatively write $\\llbracket m(h^1, h^2) \\rrbracket $ for a top-level module $m$ with submodule outputs $h^1$ and $h^2$ . We then define the execution model:","$$p_z(y|w) = (\\llbracket z \\rrbracket _w)_y$$ (Eq. 13) "," (This assumes that the root module of $z$ produces a distribution over labels $y$ .) The set of possible layouts $z$ is restricted by module type constraints: some modules (like find above) operate directly on the input representation, while others (like describe above) also depend on input from specific earlier modules. Two base types are considered in this paper are Attention (a distribution over pixels or entities) and Labels (a distribution over answers).","Parameters are tied across multiple instances of the same module, so different instantiated networks may share some parameters but not others. Modules have both parameter arguments (shown in square brackets) and ordinary inputs (shown in parentheses). Parameter arguments, like the running bird example in sec:programs, are provided by the layout, and are used to specialize module behavior for particular lexical items. Ordinary inputs are the result of computation lower in the network. In addition to parameter-specific weights, modules have global weights shared across all instances of the module (but not shared with other modules). We write $A, a, B, b, \\dots $ for global weights and $u^i, v^i$ for weights associated with the parameter argument $i$ . $\\oplus $ and $\\odot $ denote (possibly broadcasted) elementwise addition and multiplication respectively. The complete set of global weights and parameter-specific weights constitutes $\\theta _e$ . Every module has access to the world representation, represented as a collection of vectors $w^1, w^2, \\dots $ (or $W$ expressed as a matrix). The nonlinearity $\\sigma $ denotes a rectified linear unit.","The modules used in this paper are shown below, with names and type constraints in the first row and a description of the module's computation following.","=3mm","|p0.95|","Lookup ( $\\rightarrow $ Attention)","lookup[ $i$ ] produces an attention focused entirely at the index $f(i)$ , where the relationship $f$ between words and positions in the input map is known ahead of time (e.g. string matches on database fields). ","$$\\llbracket {\\texttt {lookup[i]}} \\rrbracket = e_{f(i)}$$ (Eq. 14) ","where $e_i$ is the basis vector that is 1 in the $i$ th position and 0 elsewhere.","Find ( $\\rightarrow $ Attention)","find[ $i$ ] computes a distribution over indices by concatenating the parameter argument with each position of the input feature map, and passing the concatenated vector through a MLP: ","$$\\llbracket {\\texttt {find[i]}} \\rrbracket = \\textrm {softmax}(a \\odot \\sigma (B v^i \\oplus C W \\oplus d))$$ (Eq. 15) ","Relate (Attention $\\rightarrow $ Attention)","relate directs focus from one region of the input to another. It behaves much like the find module, but also conditions its behavior on the current region of attention $h$ . Let $\\bar{w}(h) = \\sum _k h_k w^k$ , where $h_k$ is the $k^{th}$ element of $h$ . Then, ","$$& \\llbracket {\\texttt {relate[i]}}(h) \\rrbracket = \\textrm {softmax}(a \\ \\odot \\nonumber \\\\\n&\\qquad \\sigma (B v^i \\oplus C W \\oplus D\\bar{w}(h) \\oplus e))$$ (Eq. 16) ","And (Attention * $\\rightarrow $ Attention)"," and performs an operation analogous to set intersection for attentions. The analogy to probabilistic logic suggests multiplying probabilities: ","$$\\llbracket {\\texttt {and}}(h^1, h^2, \\ldots ) \\rrbracket = h^1 \\odot h^2 \\odot \\cdots $$ (Eq. 17) ","Describe (Attention $\\rightarrow $ Labels)"," describe[ $i$ ] computes a weighted average of $w$ under the input attention. This average is then used to predict an answer representation. With $\\bar{w}$ as above, ","$$\\llbracket {\\texttt {describe[i]}}(h) \\rrbracket = \\textrm {softmax}(A \\sigma (B \\bar{w}(h) + v^i))$$ (Eq. 18) ","Exists (Attention $\\rightarrow $ Labels)"," exists is the existential quantifier, and inspects the incoming attention directly to produce a label, rather than an intermediate feature vector like describe: ","$$\\llbracket {\\texttt {exists]}}(h) \\rrbracket = \\textrm {softmax}\\Big (\\big (\\max _k h_k\\big )a + b\\Big )$$ (Eq. 19) ","With $z$ observed, the model we have described so far corresponds largely to that of Andreas15NMN, though the module inventory is different—in particular, our new exists and relate modules do not depend on the two-dimensional spatial structure of the input. This enables generalization to non-visual world representations.","Learning in this simplified setting is straightforward. Assuming the top-level module in each layout is a describe or exists module, the fully- instantiated network corresponds to a distribution over labels conditioned on layouts. To train, we maximize"," $\n\\sum _{(w,y,z)} \\log p_z(y|w;\\theta _e)\n$ directly. This can be understood as a parameter-tying scheme, where the decisions about which parameters to tie are governed by the observed layouts $z$ ."],"section_name":"Evaluating modules"},{"paragraphs":["Next we describe the layout model $p(z|x;\\theta _\\ell )$ . We first use a fixed syntactic parse to generate a small set of candidate layouts, analogously to the way a semantic grammar generates candidate semantic parses in previous work BIBREF17 .","A semantic parse differs from a syntactic parse in two primary ways. First, lexical items must be mapped onto a (possibly smaller) set of semantic primitives. Second, these semantic primitives must be combined into a structure that closely, but not exactly, parallels the structure provided by syntax. For example, state and province might need to be identified with the same field in a database schema, while all states have a capital might need to be identified with the correct (in situ) quantifier scope.","While we cannot avoid the structure selection problem, continuous representations simplify the lexical selection problem. For modules that accept a vector parameter, we associate these parameters with words rather than semantic tokens, and thus turn the combinatorial optimization problem associated with lexicon induction into a continuous one. Now, in order to learn that province and state have the same denotation, it is sufficient to learn that their associated parameters are close in some embedding space—a task amenable to gradient descent. (Note that this is easy only in an optimizability sense, and not an information-theoretic one—we must still learn to associate each independent lexical item with the correct vector.) The remaining combinatorial problem is to arrange the provided lexical items into the right computational structure. In this respect, layout prediction is more like syntactic parsing than ordinary semantic parsing, and we can rely on an off-the-shelf syntactic parser to get most of the way there. In this work, syntactic structure is provided by the Stanford dependency parser BIBREF18 .","The construction of layout candidates is depicted in fig:layout, and proceeds as follows:","Represent the input sentence as a dependency tree.","Collect all nouns, verbs, and prepositional phrases that are attached directly to a wh-word or copula.","Associate each of these with a layout fragment: Ordinary nouns and verbs are mapped to a single find module. Proper nouns to a single lookup module. Prepositional phrases are mapped to a depth-2 fragment, with a relate module for the preposition above a find module for the enclosed head noun.","Form subsets of this set of layout fragments. For each subset, construct a layout candidate by joining all fragments with an and module, and inserting either a measure or describe module at the top (each subset thus results in two parse candidates.)","All layouts resulting from this process feature a relatively flat tree structure with at most one conjunction and one quantifier. This is a strong simplifying assumption, but appears sufficient to cover most of the examples that appear in both of our tasks. As our approach includes both categories, relations and simple quantification, the range of phenomena considered is generally broader than previous perceptually-grounded QA work BIBREF8 , BIBREF19 .","Having generated a set of candidate parses, we need to score them. This is a ranking problem; as in the rest of our approach, we solve it using standard neural machinery. In particular, we produce an LSTM representation of the question, a feature-based representation of the query, and pass both representations through a multilayer perceptron (MLP). The query feature vector includes indicators on the number of modules of each type present, as well as their associated parameter arguments. While one can easily imagine a more sophisticated parse-scoring model, this simple approach works well for our tasks.","Formally, for a question $x$ , let $h_q(x)$ be an LSTM encoding of the question (i.e. the last hidden layer of an LSTM applied word-by-word to the input question). Let $\\lbrace z_1, z_2, \\ldots \\rbrace $ be the proposed layouts for $x$ , and let $f(z_i)$ be a feature vector representing the $i$ th layout. Then the score $s(z_i|x)$ for the layout $z_i$ is ","$$s(z_i|x) = a^\\top \\sigma (B h_q(x) + C f(z_i) + d)$$ (Eq. 26) ","i.e. the output of an MLP with inputs $h_q(x)$ and $f(z_i)$ , and parameters $\\theta _\\ell = \\lbrace a, B, C, d\\rbrace $ . Finally, we normalize these scores to obtain a distribution: ","$$p(z_i|x;\\theta _\\ell ) = e^{s(z_i|x)} \\Big / \\sum _{j=1}^n e^{s(z_j|x)}$$ (Eq. 27) ","Having defined a layout selection module $p(z|x;\\theta _\\ell )$ and a network execution model $p_z(y|w;\\theta _e)$ , we are ready to define a model for predicting answers given only (world, question) pairs. The key constraint is that we want to minimize evaluations of $p_z(y|w;\\theta _e)$ (which involves expensive application of a deep network to a large input representation), but can tractably evaluate $p(z|x;\\theta _\\ell )$ for all $z$ (which involves application of a shallow network to a relatively small set of candidates). This is the opposite of the situation usually encountered semantic parsing, where calls to the query execution model are fast but the set of candidate parses is too large to score exhaustively.","In fact, the problem more closely resembles the scenario faced by agents in the reinforcement learning setting (where it is cheap to score actions, but potentially expensive to execute them and obtain rewards). We adopt a common approach from that literature, and express our model as a stochastic policy. Under this policy, we first sample a layout $z$ from a distribution $p(z|x;\\theta _\\ell )$ , and then apply $z$ to the knowledge source and obtain a distribution over answers $p(y|z,w;\\theta _e)$ .","After $z$ is chosen, we can train the execution model directly by maximizing $\\log p(y|z,w;\\theta _e)$ with respect to $\\theta _e$ as before (this is ordinary backpropagation). Because the hard selection of $z$ is non-differentiable, we optimize $p(z|x;\\theta _\\ell )$ using a policy gradient method. The gradient of the reward surface $J$ with respect to the parameters of the policy is ","$$\\nabla J(\\theta _\\ell ) = \\mathbb {E}[ \\nabla \\log p(z|x;\\theta _\\ell ) \\cdot r ]$$ (Eq. 28) ","(this is the reinforce rule BIBREF20 ). Here the expectation is taken with respect to rollouts of the policy, and $r$ is the reward. Because our goal is to select the network that makes the most accurate predictions, we take the reward to be identically the negative log-probability from the execution phase, i.e. ","$$\\mathbb {E}[(\\nabla \\log p(z|x;\\theta _\\ell )) \\cdot \\log p(y|z,w;\\theta _e)]$$ (Eq. 29) ","Thus the update to the layout-scoring model at each timestep is simply the gradient of the log-probability of the chosen layout, scaled by the accuracy of that layout's predictions. At training time, we approximate the expectation with a single rollout, so at each step we update $\\theta _\\ell $ in the direction $\n(\\nabla \\log p(z|x;\\theta _\\ell )) \\cdot \\log p(y|z,w;\\theta _e)\n$ for a single $z \\sim p(z|x;\\theta _\\ell )$ . $\\theta _e$ and $\\theta _\\ell $ are optimized using adadelta BIBREF21 with $\\rho =0.95,$ $\\varepsilon =1\\mathrm {e}{-6}$ and gradient clipping at a norm of 10."],"section_name":"Assembling networks"},{"paragraphs":["The framework described in this paper is general, and we are interested in how well it performs on datasets of varying domain, size and linguistic complexity. To that end, we evaluate our model on tasks at opposite extremes of both these criteria: a large visual question answering dataset, and a small collection of more structured geography questions."],"section_name":"Experiments"},{"paragraphs":["Our first task is the recently-introduced Visual Question Answering challenge (VQA) BIBREF22 . The VQA dataset consists of more than 200,000 images paired with human-annotated questions and answers, as in fig:vqa:qualitative-results.","We use the VQA 1.0 release, employing the development set for model selection and hyperparameter tuning, and reporting final results from the evaluation server on the test-standard set. For the experiments described in this section, the input feature representations $w_i$ are computed by the the fifth convolutional layer of a 16-layer VGGNet after pooling BIBREF12 . Input images are scaled to 448 $\\times $ 448 before computing their representations. We found that performance on this task was best if the candidate layouts were relatively simple: only describe, and and find modules are used, and layouts contain at most two conjuncts.","One weakness of this basic framework is a difficulty modeling prior knowledge about answers (of the form most bears are brown). This kinds of linguistic “prior” is essential for the VQA task, and easily incorporated. We simply introduce an extra hidden layer for recombining the final module network output with the input sentence representation $h_q(x)$ (see eq:layout-score), replacing eq:simple-execution with: ","$$\\log p_z(y|w,x) = (A h_q(x) + B \\llbracket z \\rrbracket _w)_y$$ (Eq. 32) ","(Now modules with output type Labels should be understood as producing an answer embedding rather than a distribution over answers.) This allows the question to influence the answer directly.","Results are shown in tbl:vqa:quantitative-results. The use of dynamic networks provides a small gain, most noticeably on \"other\" questions. We achieve state-of-the-art results on this task, outperforming a highly effective visual bag-of-words model BIBREF23 , a model with dynamic network parameter prediction (but fixed network structure) BIBREF24 , a more conventional attentional model BIBREF13 , and a previous approach using neural module networks with no structure prediction BIBREF0 .","Some examples are shown in fig:vqa:qualitative-results. In general, the model learns to focus on the correct region of the image, and tends to consider a broad window around the region. This facilitates answering questions like Where is the cat?, which requires knowledge of the surroundings as well as the object in question."],"section_name":"Questions about images"},{"paragraphs":["The next set of experiments we consider focuses on GeoQA, a geographical question-answering task first introduced by Krish2013Grounded. This task was originally paired with a visual question answering task much simpler than the one just discussed, and is appealing for a number of reasons. In contrast to the VQA dataset, GeoQA is quite small, containing only 263 examples. Two baselines are available: one using a classical semantic parser backed by a database, and another which induces logical predicates using linear classifiers over both spatial and distributional features. This allows us to evaluate the quality of our model relative to other perceptually grounded logical semantics, as well as strictly logical approaches.","The GeoQA domain consists of a set of entities (e.g. states, cities, parks) which participate in various relations (e.g. north-of, capital-of). Here we take the world representation to consist of two pieces: a set of category features (used by the find module) and a different set of relational features (used by the relate module). For our experiments, we use a subset of the features originally used by Krishnamurthy et al. The original dataset includes no quantifiers, and treats the questions What cities are in Texas? and Are there any cities in Texas? identically. Because we are interested in testing the parser's ability to predict a variety of different structures, we introduce a new version of the dataset, GeoQA+Q, which distinguishes these two cases, and expects a Boolean answer to questions of the second kind.","Results are shown in tbl:geo:quantitative. As in the original work, we report the results of leave-one-environment-out cross-validation on the set of 10 environments. Our dynamic model (D-NMN) outperforms both the logical (LSP-F) and perceptual models (LSP-W) described by BIBREF8 , as well as a fixed-structure neural module net (NMN). This improvement is particularly notable on the dataset with quantifiers, where dynamic structure prediction produces a 20% relative improvement over the fixed baseline. A variety of predicted layouts are shown in fig:geo:qualitative."],"section_name":"Questions about geography"},{"paragraphs":["We have introduced a new model, the dynamic neural module network, for answering queries about both structured and unstructured sources of information. Given only (question, world, answer) triples as training data, the model learns to assemble neural networks on the fly from an inventory of neural models, and simultaneously learns weights for these modules so that they can be composed into novel structures. Our approach achieves state-of-the-art results on two tasks. We believe that the success of this work derives from two factors:","Continuous representations improve the expressiveness and learnability of semantic parsers: by replacing discrete predicates with differentiable neural network fragments, we bypass the challenging combinatorial optimization problem associated with induction of a semantic lexicon. In structured world representations, neural predicate representations allow the model to invent reusable attributes and relations not expressed in the schema. Perhaps more importantly, we can extend compositional question-answering machinery to complex, continuous world representations like images.","Semantic structure prediction improves generalization in deep networks: by replacing a fixed network topology with a dynamic one, we can tailor the computation performed to each problem instance, using deeper networks for more complex questions and representing combinatorially many queries with comparatively few parameters. In practice, this results in considerable gains in speed and sample efficiency, even with very little training data.","These observations are not limited to the question answering domain, and we expect that they can be applied similarly to tasks like instruction following, game playing, and language generation."],"section_name":"Conclusion"},{"paragraphs":["JA is supported by a National Science Foundation Graduate Fellowship. MR is supported by a fellowship within the FIT weltweit-Program of the German Academic Exchange Service (DAAD). This work was additionally supported by DARPA, AFRL, DoD MURI award N000141110688, NSF awards IIS-1427425 and IIS-1212798, and the Berkeley Vision and Learning Center."],"section_name":"Acknowledgments"}],"paper_id":"1601.01705","title":"Learning to Compose Neural Networks for Question Answering"},{"abstract":"Sequence model learning algorithms typically maximize log-likelihood minus the norm of the model (or minimize Hamming loss + norm). In cross-lingual part-of-speech (POS) tagging, our target language training data consists of sequences of sentences with word-by-word labels projected from translations in $k$ languages for which we have labeled data, via word alignments. Our training data is therefore very noisy, and if Rademacher complexity is high, learning algorithms are prone to overfit. Norm-based regularization assumes a constant width and zero mean prior. We instead propose to use the $k$ source language models to estimate the parameters of a Gaussian prior for learning new POS taggers. This leads to significantly better performance in multi-source transfer set-ups. We also present a drop-out version that injects (empirical) Gaussian noise during online learning. Finally, we note that using empirical Gaussian priors leads to much lower Rademacher complexity, and is superior to optimally weighted model interpolation.","figures_and_tables":[],"full_text":[{"paragraphs":["The people of the world speak about 6,900 different languages. Open-source off-the-shelf natural language processing (NLP) toolboxes like OpenNLP and CoreNLP cover only 6–7 languages, and we have sufficient labeled training data for inducing models for about 20–30 languages. In other words, supervised sequence learning algorithms are not sufficient to induce POS models for but a small minority of the world's languages.","What can we do for all the languages for which no training data is available? Unsupervised POS induction algorithms have methodological problems (in-sample evaluation, community-wide hyper-parameter tuning, etc.), and performance is prohibitive of downstream applications. Some work on unsupervised POS tagging has assumed other resources such as tag dictionaries BIBREF0 , but such resources are also only available for a limited number of languages. In our experiments, we assume that no training data or tag dictionaries are available. Our only assumption is a bit of text translated into multiple languages, specifically, fragments of the Bible. We will use Bible data for annotation projection, as well as for learning cross-lingual word embeddings (§3).","Unsupervised learning with typologically informed priors BIBREF1 is an interesting approach to unsupervised POS induction that is more applicable to low-resource languages. Our work is related to this work, but we learn informed priors rather than stipulate them and combine these priors with annotation projection (learning from noisy labels) rather than unsupervised learning.","Annotation projection refers to transferring annotation from one or more source languages to the target language (for which no labeled data is otherwise available), typically through word alignments. In our experiments below, we use an unsupervised word alignment algorithm to align $15\\times 12$ language pairs. For 15 languages, we have predicted POS tags for each word in our multi-parallel corpus. For each word in one of our 12 target language training datasets, we thus have up to 15 votes for each word token, possibly weighted by the confidence of the word alignment algorithm. In this paper, we simply use the majority votes. This is the set-up assumed throughout in this paper (see §3 for more details):"],"section_name":"Cross-lingual transfer learning of sequence models"},{"paragraphs":["We will apply empirical Gaussian priors to linear-chain conditional random fields (CRFs; BIBREF3 ) and averaged structured perceptrons BIBREF4 . Linear-chain CRFs are trained by maximising the conditional log-likelihood of labeled sequences $LL(\\mathbf {w},\\mathcal {D})=\\sum _{\\langle \\mathbf {x},\\mathbf {y}\\rangle \\in \\mathcal {D}}\\log P(\\mathbf {y}|\\mathbf {x})$ with $\\mathbf {w}\\in \\mathbb {R}^m$ and $\\mathcal {D}$ a dataset consisting of sequences of discrete input symbols $\\mathbf {x}=x_1,\\ldots ,x_n$ associated with sequences of discrete labels $\\mathbf {y}=y_1,\\ldots ,y_n$ . L $k$ -regularized CRFs maximize $LL(\\mathbf {w},\\mathcal {D})-|\\mathbf {w}|^k$ with typically $k\\in \\lbrace 0,1,2,\\infty \\rbrace $ , which all introduce costant-width, zero-mean regularizers. We refer to L $k$ -regularized CRFs as L2-CRF. L $k$ regularizers are parametric priors where the only parameter is the width of the bounding shape. The L2-regularizer is a Gaussian prior with zero mean, for example. The regularised log-likelihood with a Gaussian prior is $\\mathbf {w}\\in \\mathbb {R}^m$0 . For practical reasons, hyper-parameters $\\mathbf {w}\\in \\mathbb {R}^m$1 and $\\mathbf {w}\\in \\mathbb {R}^m$2 are typically assumed to be constant for all values of $\\mathbf {w}\\in \\mathbb {R}^m$3 . This also holds for recent work on parametric noise injection, e.g., BIBREF5 . If these parameters are assumed to be constant, the above objective becomes equivalent to L2-regularization. However, you can also try to learn these parameters. In empirical Bayes BIBREF6 , the parameters are learned from $\\mathbf {w}\\in \\mathbb {R}^m$4 itself. BIBREF7 suggest learning the parameters from a validation set. In our set-up, we do not assume that we can learn the priors from training data (which is noisy) or validation data (which is generally not available in cross-lingual learning scenarios). Instead we estimate these parameters directly from source language models.","When we estimate Gaussian priors from source language models, we will learn which features are invariant across languages, and which are not. We thereby introduce an ellipsoid regularizer whose centre is the average source model. In our experiments, we consider both the case where variance is assumed to be constant – which we call L2-regularization with priors (L2-Prior) — and the case where both variances and means are learned – which we call empirical Gaussian priors (EmpGauss). L2-Prior is the L2-CRF objective with $\\sigma ^2_j=C$ with $C$ a regularization parameter, and $\\mu _j=\\hat{\\mu _j}$ the average value of the corresponding parameter in the observed source models. EmpGauss replaces the above objective with $LL(\\lambda )+\\sum _j\\log \\frac{1}{\\sigma \\sqrt{2\\pi }}e^{-\\frac{(\\lambda _j-\\mu _j)^2}{2\\sigma ^2}}$ , which, assuming model parameters are mutually independent, is the same as jointly optimising model probability and likelihood of the data. Note that minimizing the squared weights is equivalent to maximizing the log probability of the weights under a zero-mean Gaussian prior, and in the same way, this is equivalent to minimising the above objective with empirically estimated parameters $\\hat{\\mu _j}$ and $\\sigma {\\mu _j}$ . In other words, empirical Gaussian priors are bounding ellipsoids on the hypothesis space with learned widths and centres. Also, note that in single-source cross-lingual transfer learning, observed variance is zero, and we therefore replace this with a regularization parameter $C$ shared with the baseline. In the single-source set-up, L2-Prior is thus equivalent to EmpGauss. We use L-BFGS to maximize our baseline L2-regularized objectives, as well as our empirical Gaussian prior objectives."],"section_name":"Empirical Gaussian priors"},{"paragraphs":["We also introduce a drop-out variant of empirical Gaussian priors. Our point of departure is average structured perceptron. We implement empirical Gaussian noise injection with Gaussians $\\langle (\\mu _1,\\sigma _1),\\ldots , (\\mu _m,\\sigma _m)\\rangle $ for $m$ features as follows. We initialise our model parameters with the means $\\mu _j$ . For every instance we pass over, we draw a corruption vector $\\mathbf {g}$ of random values $v_i$ from the corresponding Gaussians $(1,\\sigma _i)$ . We inject the noise in $\\mathbf {g}$ by taking pairwise multiplications of $\\mathbf {g}$ and our feature representations of the input sequence with the relevant label sequences. Note that this drop-out algorithm is parameter-free, but of course we could easily throw in a hyper-parameter controlling the degree of regularization. We give the algorithm in Algorithm 1.","[1] $T=\\lbrace \\langle \\mathbf {x}^1,\\mathbf {y}^1\\rangle ,\\ldots ,\\langle \\mathbf {x}_n,\\mathbf {y}_n\\rangle \\rbrace \\text{~w.~}\\mathbf {x}_i=\\langle v_1,\\ldots \\rangle \\text{ and }v_k=\\langle f_1,\\ldots ,f_m\\rangle , \\mathbf {w}^0=\\langle w_1:\\hat{\\mu _1},\\ldots , w_m:\\hat{\\mu _m}\\rangle $ $i\\le I\\times |T|$ $j\\le n$ $\\mathbf {g}\\leftarrow \\mathbf {sample}(\\mathcal {N}(1,\\sigma _1),\\ldots ,\\mathcal {N}(1,\\sigma _m))$ $\\hat{\\mathbf {y}}\\leftarrow \\arg \\max _{\\mathbf {y}}\\mathbf {w}^i \\cdot \\mathbf {g}$ $\\mathbf {w}^{i+1}\\leftarrow \\mathbf {w}^i+\\Phi (\\mathbf {x}_j,\\mathbf {y}_j)\\cdot \\mathbf {g}-\\Phi (\\mathbf {x}_j,\\hat{\\mathbf {y}})\\cdot \\mathbf {g}$ Averaged structured perceptron with empirical Gaussian noise "],"section_name":"Empirical Gaussian noise injection"},{"paragraphs":["We make the following additional observations: (i) Following the procedure in BIBREF11 , we can compute the Rademacher complexity of our models, i.e., their ability to learn noise in the labels (overfit). Sampling POS tags randomly from a uniform distribution, chance complexity is 0.083. With small sample sizes, L2-CRFs actually begin to learn patterns with Rademacher complexity rising to 0.086, whereas both L2-Prior and EmpGauss never learn a better fit than chance. (ii) BIBREF2 present a simple approach to explicitly studying bias-variance trade-offs during learning. They draw subsamples of $l< m$ training data points $\\mathcal {D}_1, \\ldots , \\mathcal {D}_k$ and use a validation dataset of $m^{\\prime }$ data points to define the integrated variance of our methods. Again, we see that using empirical Gaussian priors lead to less integrated variance. (iii) An empirical Gaussian prior effectively limits us to hypotheses in $\\mathcal {H}$ in a ellipsoid around the average source model. When inference is exact, and our loss function is convex, we learn the model with the smallest loss on the training data within this ellipsoid. Model interpolation of (some weighting of) the average source model and the unregularized target model can potentially result in the same model, but since model interpolation is limited to the hyperplane connecting the two models, the probability of this to happen is infinitely small ( $\\frac{1}{\\infty }$ ). Since for any effective regularization parameter value (such that the regularized model is different from the unregularized model), the empirical Gaussian prior can be expected to have the same Rademacher complexity as model interpolation, we conclude that using empirical Gaussian priors is superior to model interpolation (and data concatenation)."],"section_name":"Observations"}],"paper_id":"1601.02166","title":"Empirical Gaussian priors for cross-lingual transfer learning"},{"abstract":"Speech based solutions have taken center stage with growth in the services industry where there is a need to cater to a very large number of people from all strata of the society. While natural language speech interfaces are the talk in the research community, yet in practice, menu based speech solutions thrive. Typically in a menu based speech solution the user is required to respond by speaking from a closed set of words when prompted by the system. A sequence of human speech response to the IVR prompts results in the completion of a transaction. A transaction is deemed successful if the speech solution can correctly recognize all the spoken utterances of the user whenever prompted by the system. The usual mechanism to evaluate the performance of a speech solution is to do an extensive test of the system by putting it to actual people use and then evaluating the performance by analyzing the logs for successful transactions. This kind of evaluation could lead to dissatisfied test users especially if the performance of the system were to result in a poor transaction completion rate. To negate this the Wizard of Oz approach is adopted during evaluation of a speech system. Overall this kind of evaluations is an expensive proposition both in terms of time and cost. In this paper, we propose a method to evaluate the performance of a speech solution without actually putting it to people use. We first describe the methodology and then show experimentally that this can be used to identify the performance bottlenecks of the speech solution even before the system is actually used thus saving evaluation time and expenses.","figures_and_tables":[{"caption":"Fig. 1. Schematic of a typical menu based ASR system (Wn is spoken word).","file":"3-Figure1-1.png"},{"caption":"Fig. 2. A typical speech recognition system. In a menu based system the language model is typically the set of words that need to be recognized at a given node.","file":"4-Figure2-1.png"},{"caption":"Fig. 3. Call flow of Indian Railway Inquiry System (Wn is spoken word)","file":"5-Figure3-1.png"},{"caption":"Table 1. List of Active Words at node 1","file":"6-Table1-1.png"},{"caption":"Table 2. Distance Measurement for Active Words at Node 1 of the Railway Inquiry System","file":"7-Table2-1.png"}],"full_text":[{"paragraphs":["There are several commercial menu based ASR systems available around the world for a significant number of languages and interestingly speech solution based on these ASR are being used with good success in the Western part of the globe BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . Typically, a menu based ASR system restricts user to speak from a pre-defined closed set of words for enabling a transaction. Before commercial deployment of a speech solution it is imperative to have a quantitative measure of the performance of the speech solution which is primarily based on the speech recognition accuracy of the speech engine used. Generally, the recognition performance of any speech recognition based solution is quantitatively evaluated by putting it to actual use by the people who are the intended users and then analyzing the logs to identify successful and unsuccessful transactions. This evaluation is then used to identifying any further improvement in the speech recognition based solution to better the overall transaction completion rates. This process of evaluation is both time consuming and expensive. For evaluation one needs to identify a set of users and also identify the set of actual usage situations and perform the test. It is also important that the set of users are able to use the system with ease meaning that even in the test conditions the performance of the system, should be good, while this can not usually be guaranteed this aspect of keeping the user experience good makes it necessary to employ a wizard of Oz (WoZ) approach. Typically this requires a human agent in the loop during actual speech transaction where the human agent corrects any mis-recognition by actually listening to the conversation between the human user and the machine without the user knowing that there is a human agent in the loop. The use of WoZ is another expense in the testing a speech solution. All this makes testing a speech solution an expensive and time consuming procedure.","In this paper, we describe a method to evaluate the performance of a speech solution without actual people using the system as is usually done. We then show how this method was adopted to evaluate a speech recognition based solution as a case study. This is the main contribution of the paper. The rest of the paper is organized as follows. The method for evaluation without testing is described in Section SECREF2 . In Section SECREF3 we present a case study and conclude in Section SECREF4 ."],"section_name":"Introduction"},{"paragraphs":["Fig. FIGREF1 shows the schematic of a typical menu based speech solution having 3 nodes. At each node there are a set of words that the user is expected to speak and the system is supposed to recognize. In this particular schematic, at the entry node the user can speak any of the INLINEFORM0 words, namely INLINEFORM1 or INLINEFORM2 or INLINEFORM3 or INLINEFORM4 ; INLINEFORM5 is usually called the perplexity of the node in the speech literature. The larger the INLINEFORM6 the more the perplexity and higher the confusion and hence lower the recognition accuracies. In most commercial speech solutions the perplexity is kept very low, typically a couple of words. Once the word at the entry node has been recognized (say word INLINEFORM7 has been recognized), the system moves on to the second node where the active list of words to be recognized could be one of INLINEFORM8 , INLINEFORM9 , INLINEFORM10 , ... INLINEFORM11 if the perplexity at the INLINEFORM12 node is INLINEFORM13 . This is carried on to the third node. A transaction is termed successful if and only if the recognition at each of the three nodes is correct. For example, typically in a banking speech solution the entry node could expect someone to speak among /credit card/, /savings account/, /current account/, /loan product/, /demat/, and /mutual fund transfer/ which has a perplexity of 6. Once a person speaks, say, /savings account/ and is recognized correctly by the system, at the second node it could be /account balance/ or /cheque/ or /last 5 transactions/ (perplexity 3) and at the third node (say, on recognition of /cheque/) it could be /new cheque book request/, /cheque status/, and /stop cheque request/ (perplexity 3). Though we will not dwell on this, it is important to note that an error in recognition at the entry node is more expensive than a recognition error at a lower node.","Based on the call flow, and the domain the system can have several nodes for completion of a transaction. Typical menu based speech solutions strive for a 3 - 5 level nodes to make it usable. In any speech based solution (see Fig. FIGREF3 ) first the spoken utterance is hypothesized into a sequence of phonemes using the acoustic models. Since the phoneme recognition accuracy is low, instead of choosing one phoneme it identifies l-best (typically INLINEFORM0 ) matching phonemes. This phone lattice is then matched with all the expected words (language model) at that node to find the best match. For a node with perplexity INLINEFORM1 the constructed phoneme lattice of the spoken utterance is compared with the phoneme sequence representation of all the INLINEFORM2 words (through the lexicon which is one of he key components of a speech recognition system). The hypothesized phone lattice is declared one of the INLINEFORM3 words depending on the closeness of the phoneme lattice to the phoneme representation of the INLINEFORM4 words.","We hypothesize that we can identify the performance of a menu based speech system by identifying the possible confusion among all the words that are active at a given node. If active words at a given node are phonetically similar it becomes difficult for the speech recognition system to distinguish them which in turn leads to recognition errors. We used Levenshtein distance BIBREF4 , BIBREF5 a well known measure to analyze and identify the confusion among the active words at a given node. This analysis gives a list of all set of words that have a high degree of confusability among them; this understanding can be then used to (a) restructure the set of active words at that node and/or (b) train the words that can be confused by using a larger corpus of speech data. This allows the speech recognition engine to be equipped to be able to distinguish the confusing words better. Actual use of this analysis was carried out for a speech solution developed for Indian Railway Inquiry System to identify bottlenecks in the system before its actual launch."],"section_name":"Evaluation without Testing"},{"paragraphs":["A schematic of a speech based Railway Information system, developed for Hindi language is shown in Fig. FIGREF4 . The system enables user to get information on five different services, namely, (a) Arrival of a given train at a given station, (b) Departure of a given train at a given station, (c) Ticket availability on a given date in a given train between two stations, and class, (d) Fare in a given class in a given train between two stations, and (e) PNR status. At the first recognition node (node-1), there are one or more active words corresponding to each of these services. For example, for selecting the service Fare, the user can speak among /kiraya jankari/, /kiraya/, /fare/. Similarly, for selecting service Ticket availability, user can speak /upalabdhata jankari/ or /ticket availability/ or /upalabdhata/. Generally the perplexity at a node is greater than on equal to the number of words that need to be recognized at that node. In this manner each of the services could have multiple words or phrases that can mean the same thing and the speaker could utter any of these words to refer to that service. The sum of all the possible different ways in which a service can be called ( INLINEFORM0 ) summed over all the 5 services gives the perplexity ( INLINEFORM1 ) at that node, namely, DISPLAYFORM0 ","The speech recognition engine matches the phoneme lattice of the spoken utterance with all the INLINEFORM0 words which are active. The active word (one among the INLINEFORM1 words) with highest likelihood score is the recognized word. In order to avoid low likelihood recognitions a threshold is set so that even the best likelihood wordis returned only if the likelihood score is greater than the predefined threshold. Completion of a service requires recognitions at several nodes with different perplexity at each node. Clearly depending on the type of service that the user is wanting to use; the user has to go through different number of recognition nodes. For example, to complete the Arrival service it is required to pass through 3 recognition nodes namely (a) selection of a service, (b) selection of a train name and (c) selection of the railway station. While the perplexity (the words that are active) at the service selection node is fixed the perplexity at the station selection node could depend on the selection of the train name at an earlier node. For example, if the selected train stops at 23 stations, then the perplexity at the station selection node will be INLINEFORM2 .","For confusability analysis at each of the node, we have used the Levenshtein distance BIBREF5 or the edit distance as is well known in computer science literature. We found that the utterances /Sahi/ and /Galat/ have 100% recognition. These words Sahi is represented by the string of phonemes in the lexicon as S AA HH I and the word Galat is represented as the phoneme sequence G L AX tT in the lexicon. We identified the edit distance between these two words Sahi and Galat and used that distance measure as the threshold that is able to differentiate any two words (say INLINEFORM0 ). So if the distance between any two active words at a given recognition node is lower than the threshold INLINEFORM1 , then there is a greater chance that those two active words could get confused (one word could be recognized as the other which is within a distance of INLINEFORM2 ). There are ways in which this possible misrecognition words could be avoided. The easiest way is to make sure that these two words together are not active at a given recognition node.","Table TABREF6 shows the list of active word at the node 1 when the speech application was initially designed and Table TABREF7 shows the edit distance between all the active words at the node service given in Fig. FIGREF4 . The distance between words Sahi and Galat was found to be INLINEFORM0 which was set at the threshold, namely INLINEFORM1 . This threshold value was used to identify confusing active words. Clearly, as seen in the Table the distance between word pairs fare, pnr and pnr, prasthan is INLINEFORM2 and INLINEFORM3 respectively, which is very close to the threshold value of INLINEFORM4 . This can cause a high possibility that /fare/ may get recognized as pnr and vice-versa.","One can derive from the analysis of the active words that fare and pnr can not coexist as active words at the same node. The result of the analysis was to remove the active words fare and pnr at that node.","When the speech system was actually tested by giving speech samples, 17 out of 20 instances of /pnr/ was was recognized as fare and vice-versa. Similarly 19 out of 20 instances /pnr/ was misrecognized as prasthan and vice versa This confusion is expected as can be seen from the edit distance analysis of the active words in the Table TABREF7 . This modified active word list (removal of fare and pnr) increased the recognition accuracy at the service node (Fig. FIGREF4 ) by as much as 90%.","A similar analysis was carried out at other recognition nodes and the active word list was suitably modified to avoid possible confusion between active word pair. This analysis and modification of the list of active words at a node resulted in a significant improvement in the transaction completion rate. We will present more experimental results in the final paper."],"section_name":"Case Study"},{"paragraphs":["In this paper we proposed a methodology to identify words that could lead to confusion at any given node of a speech recognition based system. We used edit distance as the metric to identifying the possible confusion between the active words. We showed that this metric can be used effectively to enhance the performance of a speech solution without actually putting it to people test. There is a significant saving in terms of being able to identify recognition bottlenecks in a menu based speech solution through this analysis because it does not require actual people testing the system. This methodology was adopted to restructuring the set of active words at each node for better speech recognition in an actual menu based speech recognition system that caters to masses."],"section_name":"Conclusion"}],"paper_id":"1601.02543","title":"Evaluating the Performance of a Speech Recognition based System"},{"abstract":"Events of various kinds are mentioned and discussed in text documents, whether they are books, news articles, blogs or microblog feeds. The paper starts by giving an overview of how events are treated in linguistics and philosophy. We follow this discussion by surveying how events and associated information are handled in computationally. In particular, we look at how textual documents can be mined to extract events and ancillary information. These days, it is mostly through the application of various machine learning techniques. We also discuss applications of event detection and extraction systems, particularly in summarization, in the medical domain and in the context of Twitter posts. We end the paper with a discussion of challenges and future directions.","figures_and_tables":[{"caption":"Table I. Nomenclatures used by linguists to classify events","file":"4-TableI-1.png"},{"caption":"Table II. Features used by linguists to classify events","file":"6-TableII-1.png"},{"caption":"Fig. 2. Sample Comlex Subcategorization Frames","file":"13-Figure2-1.png"},{"caption":"Table III. PropBank entry for attack.01, a sense of the verb attack","file":"19-TableIII-1.png"},{"caption":"Table IV. Roles for the verbs purchase, buy and sell in PropBank","file":"19-TableIV-1.png"},{"caption":"Table V. Simplified VerbNet entry for Hit-18.1 class. Every class is numbered.","file":"21-TableV-1.png"},{"caption":"Table VI. VerbNet statistics","file":"21-TableVI-1.png"},{"caption":"Table VII. Event types and their arguments in the Genia event task. The type of arguments, primary and secondary, to be extracted from the text, are also given.","file":"42-TableVII-1.png"},{"caption":"Table VIII. Structure of an Event in BioNLP 2011 Contest, corresponding to Task 1 (GE)","file":"42-TableVIII-1.png"},{"caption":"Table IX. Structure of an Event in BioNLP 2011 Contest, corresponding to Task 2 (EPI)","file":"42-TableIX-1.png"},{"caption":"Table X. Best results for various sub-tasks in BioNLP ST 2011. Recall/precision/F-score %","file":"43-TableX-1.png"},{"caption":"Table XI. Evaluation results (recall / precision / f-score for Task 1 in Whole data set (W), Abstracts only (A) and Full papers only (F)","file":"44-TableXI-1.png"},{"caption":"Fig. 6. (a) Sentence with target event structure, (b) Projection to labeled graph. ***Redraw the image***","file":"44-Figure6-1.png"},{"caption":"Fig. 7. Turku BioNLP Pipeline","file":"47-Figure7-1.png"},{"caption":"Fig. 8. TwiCal Architecture","file":"49-Figure8-1.png"},{"caption":"Table XII. Some of Examples of event types extracted by TwiCal","file":"49-TableXII-1.png"},{"caption":"Table XIII. Examples of events extracted by TwiCal","file":"49-TableXIII-1.png"},{"caption":"Table XIV. Examples of Topics Detected by LDA from Singapore based tweets on June 16, 2010","file":"51-TableXIV-1.png"},{"caption":"Table XV. Examples of Events Detected by EDCoW in June 2010","file":"51-TableXV-1.png"}],"full_text":[{"paragraphs":["Among the several senses that The Oxford English Dictionary, the most venerable dictionary of English, provides for the word event, are the following."," Although an event may refer to anything that happens, we are usually interested in occurrences that are of some importance. We want to extract such events from textual documents. In order to extract important events or events of a specific type, it is likely that we have to identify all events in a document to start with.","Consider the first paragraphs of the article on the Battle of Fredericksburg in the English Wikipedia, accessed on May 5, 2012. We have highlighted the “events\" in the paragraph.","The Battle of Fredericksburg was fought December 11–15, 1862, in and around Fredericksburg, Virginia, between General Robert E. Lee's Confederate Army of Northern Virginia and the Union Army of the Potomac, commanded by Maj. Gen. Ambrose E. Burnside. The Union army's futile frontal assaults on December 13 against entrenched Confederate defenders on the heights behind the city is remembered as one of the most one-sided battles of the American Civil War, with Union casualties more than twice as heavy as those suffered by the Confederates. ","The paragraph contains two fairly long sentences with several “events\", mentioned using the following words: fought, commanded, assaults, entrenched, remembered, casualties and suffered. Some of these “events\" are described in terms of verbs whereas the others are in terms of nouns. Here fought, commanded, assaults, battles definitely seem to be “events\" that have durations or are durative. Entrenched seems to talk about a state, whereas it is possible that suffered talks about something punctual (i.e., takes a moment or point of time) or can be durative (i.e., takes a longer period of time) as well. The act of remembering by an individual is usually considered to happen momentarily, i.e., forgotten things come back to mind at an instant of time. But, in this paragraph it is given in passive voice and hence, it is unclear who the actor is, possibly a lot different people at different points of time.","Thus, depending on who is asked, the “events” picked out may be slightly different, but the essence is that there are several events mentioned in the paragraph and the objective in event extraction is to extract as many of them as possible in an automated fashion. For example, someone may not pick out remembered as an event that took place. Some others may not want to say that entrenched is an event. In addition, if one is asked to pick an important event, responses may vary from person to person. Finally, if one is asked to summarize the paragraph, depending on the person asked the summary may vary. A summary prepared by the author of this article is given below.","The Battle of Fredericksburg, fought December 11-12, 1862, was one of the most one-sided battles of the American Civil War, with heavy Union casualties. ","Obviously, there are many other possibilities for summarization. However, the idea is that identification of events and their participants may play a significant role in summarizing a document.","This paper discusses the extraction of events and their attributes from unstructured English text. It is an survey of research in extracting event descriptions from textual documents. In addition, we discuss how the idea of event extraction can be used in application domains such as summarization of a document. We also discuss application of event extraction in the biomedical domain and in the context of Twitter messages.","The rest of the paper is organized in the following manner. Section \"Events in Linguistics and Philosophy\" provides a description of research in linguistics and philosophy. The author believes that such a background, at least at a basic level, is necessary to understand and develop the approaches and algorithms for automatic computational detection and extraction of events and their participants from textual documents. Section \"Extracting Events from Textual Documents\" discusses approaches used in extracting events from textual documents. Most approaches these days use machine learning techniques."],"section_name":"Introduction"},{"paragraphs":["Real world events are things that take place or happen. In this section, we present an overview of how real events are represented in terms of language. In particular, we discuss classification of events and features necessary for such classification. We follow this by presenting the preferred way among philosophers to represent events in terms of logic. We bring this section to an end by presenting some of the structures ascribed to events by linguists or philosophers working at an abstract level.","The reason for the inclusion of this section in the paper is to set the context for the discussions in the following sections on the practical task of extracting events. Practical systems do not usually follow linguistic niceties although they draw inspiration from linguistics or philosophy."],"section_name":"Events in Linguistics and Philosophy"},{"paragraphs":["There have been many attempts at classifying linguistic events. Below, we briefly discuss a few. The primary focus when linguists discuss events is on the verb present in a sentence. Nouns, adjectives and other elements present in a sentence provide arguments for the verb.","Aristotle (as presented in BIBREF0 ) classified verbs that denote something happening into three classes: actuality, movement and action. An actuality represents the existence of a thing or things; this is called state by others (e.g., BIBREF1 ). An examples of actuality can be seen in the sentence Jon is ill. A movement is an incomplete process or something that takes time but doesn't have an inherent end. An example of movement is seen in the sentence Jon is running. An action is something that takes time and has an inherent end. An example of an action is present in the sentence Jon is building a house. In other words, Aristotle distinguished between states and events and then events."," BIBREF2 lists verbs that belong to the three Aristotelian classes and develops membership criteria for the classes. Kenny renamed the classes as states, activities (actions without inherent end) and performances (actions with inherent ends). Kenny's membership criteria are based on semantic entailments about whether the event can be considered to have taken place when it is still in progress. For example, during any point when we say Jon is running, we can consider that the activity of running has taken place. In other words Jon is running entails Jon has run. Thus, run is an activity. In contrast, when we say Joh is taking the final, we cannot say that Jon has taken the final. In other words, the first does not entail the second. Thus, the main difference between an activity and a performance is what is called delimitation. A delimited event has a natural end."," BIBREF3 developed a 4-way classification scheme for linguistic events and BIBREF4 developed a set of criteria for membership in the classes. The classes enumerated by Dowty are: states, activities, achievements and accomplishments. The definitions are given below.","Activities: Events that take place over a certain period of time, but do not necessarily have a fixed termination point. Examples; Jon walked for an hour, and Jon is driving the car.","Accomplishments: Events that happen over a certain period of time and then end. Examples: Jon built a house in a month, and Jon is taking the final.","Achievements: These are events that occur instantaneously and lack continuous tenses. Examples: Jon finished the final in 45 minutes and The vase broke.","States: These are non-actions that hold for a certain period of time, but lack continuous tenses. Examples: Jon knows the answer and Jon likes Mary."," BIBREF5 adopts the same classification as Vendler and Dowty, but divides achievements into two classes. The first one is still called achievements, but the second one is called semelfactives. In this new scheme, achievements are instantaneous (that is, the beginning of the event is the same as its end) culminating events, but semelfactives are events with no duration that result in no change of state. An example of a semelfactive is: Jon knocked on the door.","Table 1 presents the nomenclatures introduced by various linguists in one place. There are many variations of the schemes given here, although we do not discuss them in this paper.","In the early work on event classification, Aristotle, Vendler and others assume that what needs to be classified is the verb. However, many have concluded that it is impossible to classify a verb into a specific class. It is more appropriate to say that a clause containing an event has a class, and the classification of such a clause depends not only upon the verb, but also on other material present in the clause BIBREF6 , BIBREF4 , BIBREF7 , BIBREF8 . In other words, the classification must be compositional or must depend on various features of the clause, not exclusively verb-based. There is also substantial evidence that sentence material other than the verb can change the overall event type. For example, addition of a direct object can change an activity to an accomplishment BIBREF1 , as in the following examples.","Bill ran for five minutes/*in five minutes: activity","Bill ran the mile *for 5 minutes/in 5 minutes: accomplishment"],"section_name":"Classifying Events"},{"paragraphs":["Many authors in linguistics have delved deeper into the nature of event classes and have tried to come up with features or characteristics that can be used to identify whether something (verb or a clause) belongs to a particular event class or not. These features or characteristics are necessary to describe the structure of events in a theoretical sense. Description of event structure usually refers to the actual words used (lexical features or characteristics) and also the structure of clause or sentence (syntactic features or characteristics). Identification of such features may be described as finding parameters of event types or parameterization of event types.","A lot of the work on parameterization of event types/classes use the classes espoused by Vendler. These include BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 and others. We will only briefly touch upon such work in this paper. Our objective is to impress upon the reader that identification of features of event classes is considered an important task by linguists.","For example, BIBREF9 describes Vendler's classes with two binary features or parameters: continuousness: whether an event has duration, and boundedness: whether an event has a (natural) terminal point or endpoint. Using these two features, the four Vendler classes can be parameterized as follows.",": -bounded, -continuous",": -bounded, +continuous",": +bounded, -continuous",": +bounded, +continuous"," BIBREF12 , BIBREF13 introduce the notion of countability while discussing event classes. This is similar to the mass-count opposition in nouns. Terminating events can be counted, but non-terminating processes cannot. Hoeksema introduces two binary features: count and duration to obtain Vendler's classes as seen below. The feature duration refers to whether the event takes place over time.",": -count, -duration",": -count, +duration",": +count, -duration",": +count, +duration"," BIBREF11 refines Vendler's classes by adding a class much like Smith's semelfactives BIBREF5 . He suggests that, in addition to states, there are four event types: culmination, culminated process, point, and process. He uses two binary features or parameters: consequence identifying termination or culmination, and atomic or non-atomic (which Moens called extended). Atomic is also called momentous or pointed. Moen's classification is given below, along with the features and examples.",": +consequence, +atomic (examples: recognize, win the race)",": +consequence, -atomic (examples: build a house)",": -consequence, +atomic (example: hiccup, tap, wink)",": -consequence, -atomic (example: run, swim, play the piano)",": (examples: understand, love, resemble)","Moens also claims that culminated process is an event class whose members are made up of smaller atomic units. In particular, a culminated process is a process with a consequent state. This insight that events can be decomposed into sub-events was used later by others working on the lexical analysis of events e.g., BIBREF16 , BIBREF17 . Others such as BIBREF18 , BIBREF19 , BIBREF20 have claimed that arguments of verbs are related to sub-events.","We summarize the various features that linguists have used to classify events in Table 2 . Of course, we do not discuss many other proposals for features in this brief discussion.","Classification of events and their parameterization of verbs or predicates (or clauses) are only the first steps in developing a deeper linguistic understanding of events. In particular, in order to understand the linguistic representation of events, linguists need to go beyond classification schemes."],"section_name":"Parameters of Event Classes"},{"paragraphs":["Mathematical logic is used to represent the semantics of language. In particular, we use logic to represent the meaning of single sentences. Early work on events, e.g., Panini (as discussed by BIBREF21 and BIBREF22 ) stated that language encodes two kinds of information–actions and non-actions. Verbs represent actions and nouns represent non-actions or things."," BIBREF23 proposes that one needs an event variable $e$ to represent events in mathematical logic. This variable $e$ is used to represent relations represented by the event denoted by the verb and other constituents in the sentence, such as modifiers. Davidson claims that logically speaking, events are like things in that they can be represented by a variable and this variable can be modified and quantified. A question that arises is: how many arguments should an event predicate (in logic) take BIBREF2 ? Just like nominal modifiers modify nouns, event modifiers can modify event predicates. An event predicate can take any number of modifiers just like noun (nominal) modifiers. Examples of event modifiers are: time, place, manner and instrument. Davidson proposed that an event predicate may take one or more required arguments (is this true?) and any number of adjuncts or optional modifiers. Consider the following examples from BIBREF23 . The English sentence and the corresponding logical representation or logical form is given for each example.","John buttered the toast."," $\\exists e \\; buttered (Jones, the\\_toast, e)$ ","John buttered the toast slowly."," $\\exists e \\; buttered (Jones, the\\_toast, e) \\wedge slowly (e)$ ","John buttered the toast slowly, in the bathroom."," $\\exists e \\; buttered (Jones, the\\_toast, e) \\wedge slowly (e) \\wedge in\\_the\\_bathroom (e)$ ","John buttered the toast slowly, in the bathroom, with a knife."," $\\exists e \\; buttered (Jones, the\\_toast, e) \\wedge slowly (e) \\wedge in\\_the\\_bathroom (e) \\wedge with\\_a\\_knife (e)$ ","John buttered the toast slowly, in the bathroom, with a knife, at midnight."," $\\exists e \\; buttered (Jones, the\\_toast, e) \\wedge slowly (e) \\wedge in\\_the\\_bathroom (e) \\wedge with\\_a\\_knife (e) \\wedge at\\_midnight (e)$ ","Thus we can see that Davidson's approach places the event variable $e$ in the main predicate of a clause and distributes it among the modifiers of the clause in logical representation. In writing the meaning in Davidsonian logic, the author creates predicates such as $the\\_toast$ and $in\\_the\\_bathroom$ , just for illustration, without going into details.","Davidsonian representation allows events to be represented in logic (logical semantics) without requiring verbs to have multiple arities, i.e., without taking different arguments in different situations. Because the event is represented as a variable, the event variable $e$ can be included in the representation of logical meaning of each modifier or adjunct. Another benefit is that using Davidson's representation, one can analyze events represented syntactically as nouns (nominals) or verbs BIBREF21 . For example, one can refer to an event using the verb to burn or the noun a burn. Parsons also observes that using a variable to represent an event allows quantification over events the same way quantification applies to things. The following examples are from BIBREF21 .","In every burning, oxygen is consumed."," $\\forall e \\; burning (e) \\rightarrow \\exists e^{\\prime } (consuming (e^{\\prime }) \\wedge object (e, oxygen) \\wedge in (e,e^{\\prime })$ ","Agatha burned the wood."," $\\exists e \\; burning (e) \\wedge subject (e, Agatha) \\wedge object (e, wood)$ ","Oxygen was consumed."," $\\exists e^{\\prime } \\; consuming (e^{\\prime }) \\wedge object (e^{\\prime }, oxygen)$ ","We do not go into details of containment of events as expressed by $in$ in the first example above, and also the representation of passives as in the third example above. In these three examples, the author uses predicates such as object and subject which represent more fine-grained relationship with the main predicate (corresponding to the verb usually) than the examples earlier. Extending this work, BIBREF21 , BIBREF24 , BIBREF25 have demonstrated that using Davidson's $e$ variable allows one to express tense dependency between perception verbs and their infinitival compliments in a natural way."," BIBREF21 extends Davidson's approach to logical representation by adding an extra term corresponding to the event type of the predicate. He distinguishes between two types of eventualities: eventualities that culminate called Cul containing achievements and accomplishments, and those that do not, called Hold containing states and activities.","John buttered the toast."," $\\exists e \\; buttering (e) \\wedge agent (e, Jones) \\wedge theme (e, toast) \\wedge (\\exists t \\; (t where is a noun phrase headed by the noun in question, e.g., John had a party.","Check if the sentence can be used in present progressive as in John is having a party. If the sentence is felicitous, it adds to the noun being inventive. If it sounds odd, it adds to the evidence that the noun is stative.","Check if the sentence can be used in a pseudo-cleft construction such as What John did was have a party. If it is felicitous, the noun is more likely to be eventive. If not, it is more likely to be stative.","Check if the sentence suggests iterative or habitual action using the simple present such as John has a party every Friday. If so, it adds evidence that the noun is eventive. If the sentence suggests that the situation is taking place at that very moment that it is uttered, it adds evidence that the noun is stative as in John has a cold.","In addition to OntoNotes, there have been other efforts at obtaining large-scale annotated corpora such at the GLARF project BIBREF115 that tries to capture information from various Treebanks and superimpose a predicate argument structure. The Unified Linguistic Annotation (ULA) project BIBREF116 is a collaborative effort that aims to merge PropBank, NomBank, the Penn Discourse Treebank BIBREF117 and TimeBank BIBREF118 with co-reference information."],"section_name":"OntoNotes and Other Large-scale Annotated Corpora"},{"paragraphs":["Different models of events have been used in computational linguistics work geared toward information extraction.","We discuss TimeML events next followed by events. We discuss biomedical event extraction in Section , and extraction of events from Twitter in Section ."],"section_name":"Extracting Events from Textual Documents"},{"paragraphs":["TimeML is a rich specification language for event and temporal expressions in natural language text. In the TimeML BIBREF119 , BIBREF120 annotation scheme, an event is a general term for situations that happen or occur. Events can be punctual or momentary, or last for a period of time. Events in TimeML format may also include predicates describing states or circumstances in which something holds true. Only those states that participate in an opposition structure, as discussed in Subsection UID42 , are annotated. In general, an event can be expressed in terms of verbs, nominalizations, adjectives, predicative clauses, or prepositional phrases. TimeML allows an event, annotated with the EVENT tag, to be one of seven types: occurrence, state, report, i-action, i-state, aspectual and perception. The first five are special cases. The last two, Occurrence and State are used for general cases that do not fall in the special ones.","Reporting: A reporting event describes an action declaring something, narrating an event, informing about a situation, and so on. Some verbs which express this kind of event are say, report, tell, explain, and state. An example sentence with the verb say is Punongbayan said that the 4,795-foot-high volcano was spewing gases up to 1,800 degrees.","I-Action: I stands for intensional. According to the TimeML annotation guidelines, an i-action is a dynamic event that takes an event-denoting argument, which must be explicitly present in the text. Examples of verbs that are used to express i-actions include attempt, try, promise and offer. An example sentence with the verb try is Companies such as Microsoft or a combined worldcom MCI are trying to monopolize Internet access.","I-State: I-State stands for intensional state. Like an I-Action, an I-State event takes an argument that expresses an event. Unlike an I-Action, the I-State class is used for events which are states. An example sentence that uses the verb believe is We believe that his words cannot distract the world from the facts of Iraqi aggression. Other verbs used to express i-states include intend, want, and think.","Aspectual: An aspectual predicate takes an event as an argument, and points to a part of the temporal structure of the event. Such a part may be the beginning, the middle or the end of an event. Verbs such as begin, finish and continue are such aspectual predicates. An example sentence with the verb begin is All non-essential personnel should begin evacuating the sprawling base.","Perception: This class includes events involving the physical perception of another event. Such events are typically expressed by verbs such as see, watch, glimpse, hear, listen, and overhear. An example sentence with the verb see is Witnesses tell Birmingham police they saw a man running.","Occurrence: An occurrence is a general event that occurs or happens in the world. An example of an occurrence is given in the following sentence: The Defense Ministry said 16 planes have landed so far with protective equipment against biological and chemical warfare. The occurrence has been highlighted in bold.","State: A state describes circumstances in which something obtains or holds true. An example sentence that shows two states is It is the US economic and political embargo which has kept Cuba in a box.","TimeML allows one to mark up temporal expressions using the TIMEX3 tag. Temporal expressions are of three types: (a) Fully specified temporal expressions such as June 11, 2013, (b) Underspecified temporal expressions such as Monday, (c) Durations such as three days. TimeML uses the SIGNAL tag to annotate sections of text, usually function words, that indicate how temporal objects are related to each other. The material marked by SIGNAL may contain different types of linguistic elements: indicators of temporal relations such as prepositions such as on and during, other temporal connectives such as when, etc. The TIMEX3 and SIGNAL tags were introduced by BIBREF122 , BIBREF123 .","A major innovation of TimeML is the LINK tags that encode relations between temporal elements of a document and also help establish ordering between the events in a document. There are three types of links: TLINK showing temporal relationships between events, or between an event and a time; SLINK or a subordination link to show context that introduces relations between two events, or an event and a signal; ALINK or an aspectual link to show relationship between an aspectual event and its argument event. TLINK allows for 13 temporal relations introduced by BIBREF124 , BIBREF125 . SLINK is used to express contexts such as use of modal verbs, negatives, positive and negative evidential relations, factives which require the event argument to be true, and counterfactives which require the event argument to be false. ALINK expresses initiation, culmination, termination or continuation relationships between an event and its argument event. Finally, TimeML is able to express three types of causal relations: an event causing an event, an entity causing an event, and the special situation where the use of the discourse marker and as a signal to introduce a TLINK indicating that one event happened before another as in He kicked the ball and it rose into the air.","The creators of TimeML have spent significant efforts to develop a fairly large corpus annotated with TimeML tags. This corpus is called the TIMEBANK corpus BIBREF118 and has 300 annotated articles. This corpus has been used to learn to extract events and temporal relations among events."],"section_name":"TimeML Events"},{"paragraphs":["In the ACE model, only “interesting” events are annotated in corpora and thus extracted by a trained program. ACE annotators specify the event types they want to be extracted. For example, in one information extraction contest, an ACE 2005 event was of 8 types, each with one has one or more sub-types. The types are given below. ***Maybe, give some examples***","Life: Be-born, marry, divorce, injure and die","Movement: Transport","Transaction: Transfer-ownership, Transfer money","Business: Start-organization, Merge-organization, Declare-bankruptcy","Contact: Meet, Phone-write","Conflict: Attack, demonstrate","Personnel: Start position, End position, Nominate, Elect, and","Justice: Arrest-Jail, Release-Parole, Trial-Hearing, Charge-Indict, Sue, Convict, Sentence, Fine, Execute, Extradite, Acquit, Appeal, Pardon.","Each event also has four categorial attributes. The attributes and their values are given below.","Modality: Asserted and Other where Other includes, but is not limited to: Believed events; Hypothetical events; Commanded and requested events; Threatened, Proposed and Discussed events; and Promised events.","Polarity: Positive and Negative.","Genericity: Specific, Generic","Tense: Past, Present, Future and Unspecified.","ACE events have arguments. Each event type has a set of possible argument roles, which may be filled by entities, time expressions or other values. Each event type has a set of possible argument roles. There are a total of 35 role types although no single event can have all 35 roles. A complete description of which roles go with which event type can be found in the annotation guidelines for ACE 2005 events. In an ACE event, time is noted if when explicitly given. Others have defined events or event profiles themselves to suit their purpose. For example, Cybulska and Vossen BIBREF126 , BIBREF127 describe an historical information extraction system where they extract event and participant information from Dutch historical archives. They extract information using what they call profiles. For example, they have developed 402 profiles for event extraction although they use only 22 of them in the reported system. For extraction of participants, they use 314 profiles. They also 43 temporal profiles and 23 location profiles to extract temporal and locational information. Profiles are created using semantic and syntactic information as well as information gleaned from Wordnet BIBREF50 .","The ACE annotation scheme, discussed earlier, was developed by NIST in 1999, and the ERE (Entities, Relations and Events) scheme was defined as a simpler version of ACE BIBREF128 . One of ERE's goals is also to make annotating easier and annotations more consistent across annotators. ERE attempts to achieve these goals by removing the most problematic annotations in ACE and consolidating others. We will discuss the three types annotations now: Entities, Relations and Events.","For example, consider Entities. ACE and ERE both have Person, Organization, Geo-Political Entity and Location as types of entities. ACE has two additional types, Weapon and Vehicle, which ERE does not have. ERE doesn't distinguish between Facility and Location types and merge them into Location. ERE has a type called Title for titles, honorifics, roles and professions. ACE has subtypes for entity mentions, which ERE does not. In addition to subtypes, ACE classifies entity mentions into classes (e.g., Specific, Generic and Underspecified), ERE has only Specific. ACE and ERE also have differences in how extents and heads are marked, and levels of entity mentions.","The purpose of Relation annotation in both ACE and ERE is to extract a representation of the meaning of the text, not necessarily tied to the underlying syntactic or lexical representation. Both schemes include Physical, Part-Whole, Affiliation and Social relations although the details are a bit different. Both tag relations inside a single sentence and tags only explicit mentions. Nesting of tags is not allowed. Each relation can have up to two ordered Argument slots. Neither model tags negative relations. However, ERE annotates only asserted (\"real\") events whereas ACE allows others as well, e.g., Believed Events, Hypothetical Events, Desired Events and Requested Events. There is no explicit trigger word in ACE, which annotates the full clause that serves as the trigger for a relation whereas ERE attempts to minimize the annotated span by allowing for the tagging of an optional trigger word or phrase. ACE justifies tagging of each Relation by assigning Syntactic Clauses to them, such as Possessive, PreMod and Coordination. The three types of Relations inn ERE and ACE have sub-types: Physical, Part-Whole, and Social and Affiliation, but ERE collapses ACE types and sub-types to make them more concise, possibly less specific. BIBREF128 discuss the similarities and differences between ACE and ERE in detail.","Events in both ACE and ERE are defined as `specific occurrences' involving `specific participants'. Like entities and relations, ERE is less specific and simplified compared to ACE. Both annotation schemes annotate the same event types: Life, Movement, Transaction, Business, Conflict, Contact, Personnel, and Justice."," BIBREF129 use another annotation scheme called Richer Event Description (RED), synthesizing co-reference BIBREF130 , BIBREF131 and THYME-TimeML temporal relations BIBREF132 . BIBREF129 discusses challenges in annotating documents with the RED schema, in particular cause-effect relations. The usual way to annotate cause-effect relations is using the counter-factual definition of causation in philosophy BIBREF133 , BIBREF134 :","“X causes Y\" means if X had not occurred, Y would not have happened.","However, BIBREF129 found that this definition leads to many difficult and sometimes erroneous annotations, and that's why while performing RED annotations, they used another definition BIBREF135 , BIBREF136 which treats causation as “a local relation depending on intrinsic properties of the events and what goes on between then, and nothing else\". In particular, the definition is","“X causes Y\" means Y was inevitable given X.","In fact, in the annotations performed by BIBREF129 , they use the new definition to make judgements, but use the old definition as a precondition to the new one.","The Knowledge Base Population Track (TAC-BKP) was started by NIST in 2009 to evaluate knowledge bases (KBs) created from the output of information extraction systems. The primary tasks are a) Entity linking–linking extracted entities to entities in knowledge bases, and b) Slot filling–adding information to entity profiles, information that is missing from the knowledge base BIBREF137 . Wikipedia articles have been used as reference knowledge bases in evaluating TAC-KBP tasks. For example, given an entity, the goal is to identify individual nuggets of information using a fixed list of inventory relations and attributes. For example, given a celebrity name, the task is to identify attributes such as schools attended, occupations, important jobs held, names of immediate family members, etc., and then insert them into the knowledge base. Many people compare slot filling to answering a fixed set of questions, obtaining the answers and filling in the appropriate slots in the knowledge base. Slot filling in TAC-KBP differs from extraction in ACE and ERE notations in several ways such as TAC-KBP seeks out information for named entities only, chiefly PERs and ORGs, TAC-KBP seeks to obtain values for slots and not mentions, and events are handled as uncorrelated slots, and assessment is like in question-answering.","Our focus on this paper has been on extracting events, and we know that to extract events properly, we need to explicitly extract event mentions, and also extract associated attributes such as agents, locations, time of occurrence, duration, etc. Rather than explicitly modeling events, TAC-KBP does so implicitly as it captures various relations associated with for example the agent of the event. For example, given a sentence “Jobs is the founder and CEO of Apple\", TAC-KBP may pick \"Apple\" as the focal entity and identify \"Jobs\" as the filler of its founder slot, and \"Jobs\" as the filler of its CEO slot. However, an ACE or ERE annotation program will ideally pick the event as Founding, with Jobs as an argument (say the first argument or arg1, or the Actor) of the event, and \"Apple\" as another argument, say arg2."],"section_name":"ACE Events"},{"paragraphs":["Many even extraction systems have been built over the years. A big motivator for development of event extraction systems seem to be various contests that are held every few years, although there has been considerable amount of non-contest related research as well. Although we discuss extraction of events represented by various formats, the methods are not really different from each other. That is why we discuss TimeML events in more detail and present the others briefly in this section.","We describe a few of the approaches that have been used for extracting TimeML type events. Quite a few papers that attempt to do so have been published BIBREF121 , BIBREF138 , BIBREF139 , BIBREF140 , BIBREF141 , and we pick just a few representative papers.",": BIBREF121 implemented an event and event feature extraction system called EVITA and showed that a linguistically motivated rule-based system, with some help using statistical disambiguation perfumed well on this task. Evita is claimed to be a unique tool within the TimeML framework in that it is very general, being not based on any pre-established list of event patterns and being domain-independent. Evita can also identify, based on linguistic cues, grammatical information associated with event referring expressions, such as tense, aspect, polarity and modality, as stated in the TimeML specification. Evita does not directly identify event participants, but can work with named entity taggers to link arguments to events.","Evita breaks down the event recognition problem to a number of sub-tasks. Evita preprocesses the input text using the Alembic Workbench POS tagger, lemmatizer to find lexical stems, and chunkier to obtain phrase chunks, verbal, nominal and adjectival, the three that are commonly used as event referring expressions BIBREF142 . For each subtask after pre-processing, it combines linguistic- and statistically-based knowledge. Linguistic knowledge is used in local and limited contexts such as verb phrases and to extract morphological information. Statistical knowledge is used to disambiguate nominal events. The sub-tasks in event recognition in Evita are: determination of event candidates and then the events, identification of grammatical features of events, additional clustering of event chunks for event detection and grammatical feature identification in some situations.","For event identification, Evita looks at the lexical items tagged by the preprocessing step. It uses different strategies for identifying events in the three categories: verbs, nouns and adjectives. For identifying events in a verbal chunk, Evita performs lexical look-up and limited contextual parsing in order to exclude weak stative predicates such as be and generics such as verbs with bare plural subjects. Identifying events expressed by nouns involves a phase of lexical look-up and disambiguation using WordNet, and by mapping events SemCor and TimeBank 1.2 to WordNet synsets. Evita consults 25 subtrees from WordNet where all the synsets denote events. One of these, the largest, is the tree underneath the sunset that contains the word event. If the result of this lexical look-up is not conclusive (i.e., if a nominal occurs as both event and non-event in WordNet), a disambiguation step is applied, based on rules learned by a Bayesian classifier trained on SemCor. To identify events from adjectives, Evita uses a conservative approach, where it tags only those adjectives that were annotated as such in TimeBank 1.2, when such adjectives occur as the head of a predicative complement.","To identify grammatical features (e.g., tense, aspect, modality, polarity and non-finite morphology) of events, Evita uses different procedures based on the part of speech of the event denoting expression. But, in general it involves using morphology, pattern matching, and applying a large number (e.g., 140 such rules for verbal chunks) simple linguistic rules. However, to identify the event class, it performs lexical look-up and word sense disambiguation. Clustering is used to identify chunks from the preprocessing stage, that contribute information about the same event, e.g., when some modal auxiliaries and use of copular verbs. Clustering is activated by specific triggers such as the presence of a chunk headed by an auxiliary verb or a copular verb.","Evaluation of Evita was performed by comparing its performance against TimeBanck 1.2. The reported performance was that Evita had 74.03% precision, 87.31% recall and an F-measure of 80.12% in event detection. Accuracy (precision?) for polarity, aspect and modality was over 97% in each case.",": BIBREF138 use TimeBank-annotated events and identify which words and phrases are events. They consider event identification as a classification task that works on word-chunks. They use the BIO formulation that augments each class label with whether the word is the Beginning, Inside or Outside of a chunk BIBREF143 .","They use a number of features, categorized into various classes, for machine learning. These include affix features (e.g., three or four characters from the beginning and end of each word), morphological features (e.g., base form of the word, and base form of any verb associated with the word if the word is a noun or gerund, for example), word-class features (e.g., POS tags, which noun or verb cluster a word belongs to where the clusters are obtained using co-occurrence statistics in the manner of BIBREF144 ), governing features (e.g., governing light verb, determiner type—cardinal or genitive, for example), and temporal features (e.g., a BIO label indicating whether the word is contained inside a TIMEX2 temporal annotation, a governing temporal preposition like since, till, before, etc.). They also use negation features and Wordnet hypernyms as features. For classification, they use the TinySVM implementation of SVM by BIBREF145 .","They perform experiments with TimeBank documents using a 90% stratified sampling for training and 10% for testing. They obtained 82% precision and 71% recall, with an F-measure of 0.759. They did compare their algorithm with an version of Evita they programmed themselves; this system obtained 0.727 F-measure, and thus Bethard and Martin's approached performed about 4% better. When Bethard and Martin's system was extended to identifying semantic class of an event, it did not perform as well, obtaining precision of 67%, recall of 51%, and F-measure of 0.317. However, the system was much better at identifying the classes of verbs with F-measure of 0.707 compared to finding classes of nouns with an F-measure of 0.337 only.","TIPSem (Temporal Information Processing based on Semantic information) is a system that participated in the TemEval-2 Competition BIBREF146 in 2010, which presented several tasks to participants, although we are primarily interested in the event extraction task. TIPSem achieved the best F1 score in all the tasks in TempEval-2 for Spanish, and for English it obtained the best F1 metric in the task of extracting events, which required the recognition and classification of events as defined by TimeML EVENT tag.","TIPSem learns Conditional Random Field (CRF) models using features for different language analysis levels, although the approach focuses on semantic information, primarily semantic roles and semantic networks. Conditional Random Fields present a popular and efficient machine learning technique for supervised sequence labeling BIBREF147 .","The features used for training the CRF models are similar to one used by others such as Bethard and Martin, although details vary. However, they add semantic role labels to the mix of features. In particular, they identify roles for each governing verb. Semantic role labeling BIBREF148 , BIBREF149 , BIBREF150 identifies for each predicate in a sentence, semantic roles and determine their arguments (agent, patient, etc.) and their adjuncts (locative, temporal, etc.). The previous two features were combined in TIPSem to capture the relation between them. The authors think this combination introduces additional information by distinguishing roles that are dependent on different verbs. The importance of this falls especially on the numbered roles (A0, A1, etc.) meaning different things when depending on different verbs.","The test corpus consists of 17K words for English and 10K words for Spanish, provided by the organizers of TempEval-2. For English, they obtained precision of 0.81, recall of 0.86 and F-measure of 0.83 for recognition with event classification accuracy of 0.79; for Spanish the numbers were 0.90, 0.86, 0.88 for recognition and 0.66 for classification accuracy. We provide these numbers although we know that it is difficult to compare one system with another, for example Bethard and Martin's system with TIPSem since the corpora used are difference.","As in TempEval-2, TempEval-3 BIBREF151 participants took part in a task where they had to determine the extent of the events in a text as defined by the TimeML EVENT tag. In addition, systems may determine the value of the features CLASS, TENSE, ASPECT, POLARITY, MODALITY and also identify if the event is a main event or not. The main attribute to annotate is CLASS.","The TempEval-3 dataset was mostly automatically generated, using a temporal merging system. The half-million token text corpus from English Gigaword2 was automatically annotated using TIPSem, TIPSem-B BIBREF140 and TRIOS BIBREF152 . These systems were re-trained on the TimeBank and AQUAINT corpus, using the TimeML temporal relation set. The outputs of these three state-of-the-art system were merged using a merging algorithm BIBREF151 . The dataset used comprised about 500K tokens of “silver\" standard data and about 100K tokens of “gold\" standard data for training, compared to the corpus of roughly 50K tokens corpus used in TempEval 1 and 2.","There were seven participants and all the participants except one used machine learning approaches. The top performing system was ATT-1 BIBREF153 with precision 81.44, recall 80;67 and F1 of 81.05 for event recognition, and 71.88 for event classification. Close behind was the ATT-2 system BIBREF153 with precision, recall and F-1 of 81.02, 80.81 and 80.92 for event recognition respectively, and 71.10 for event classification. Both systems used MaxEnt classifiers with","Obviously, different sets of features impact on the performance of event recognition and classification BIBREF154 , BIBREF155 , BIBREF156 . In particular, BIBREF157 also examined performance based on different sizes of n-grams in a small scale (n=1,3). Inspired by such work, in building the ATT systems, the creators intended to systematically investigate the performance of various models and for each task, they trained twelve models exploring these two dimensions, three of which we submitted for TempEval-3, and of these three performed among the top ten in TempEval-3 Competition.","The ATT-1 models include lexical, syntactic and semantic features, ATT-2 models include only lexical and syntactic features, and ATT-3 models include only lexical features, i.e., words. They experimented with context windows of 0, 1, 3, and 7 words preceding and following the token to be labeled. For each window size, they trained ATT-1, ATT-2 and ATT-3 models. The ATT-1 models had 18 basic features per token in the context window for up to 15 tokens, so up to 270 basic feaures for each token to be labeled. The ATT-2 models had 16 basic features per token in the context window, so up to 240 basic features for each token to be labeled. The ATT-3 models had just 1 basic feature per token in the context window, so up to 15 basic features for each token to be labeled.","For event extraction and classification, and event feature classification, they used the efficient binary MaxEnt classifier for multi-class classification, available in the machine learning toolkit LLAMA BIBREF158 . They also used LLAMA's pre-processor to build unigram, bigram and trigram extended features from basic features.","For event and time expression extraction, they trained BIO classifiers. It was found that the absence of semantic features causes only small changes in F1. The absence of syntactic features causes F1 to drop slightly (less than 2.5% for all but the smallest window size), with recall decreasing while precision improves somewhat. F1 is also impacted minimally by the absence of semantic features, and about 2-5% by the absence of syntactic features for all but the smallest window size.1","A was surprising that that ATT-3 models that use words only performed well, especially in terms of precision (precision, recall and F2 of 81.95, 75.57 and 78.63 for event recognition, and 69.55 F1 for event classification) . It is also surprising that the words only models with window sizes of 3 and 7 performed as well as the models with a window size of 15. These results are promising for “big dataÓ text analytics, where there may not be time to do heavy preprocessing of input text or to train large models.","We have already discussed several approaches to extraction of events represented by TimeML representation. Extracting events that use other representation is not very different, but different representations have existed and exist, and therefore we briefly present some such attempts. Some of these predate the time TimeML became popular. For example, the various Message Understanding Conferences (MUCs, seven were organize by DARPA from 1987 to 1997), asked participants to extract a small number of relations and events. For instance, MUC-7, the last one, called for the extraction of 3 relations (person-employer, maker-product, and organization-location) and 1 event spacecraft launches.","For example, the MUC-7 and ACE events did not attempt to cover all events, but a limited number of pre-specified event types or classes that participants need to detect during a contest period, based on which the contestants submit papers for publication. The number and the type of arguments covered are also limited and are pre-specified before the competitions start."," BIBREF159 discuss a relation and event extraction system covering areas such as political, financial, business, military, and life-related topics. The system consists of tagging modules, a co-reference resolution module, and a temple generation module. They store the events generated in MUC-7 BIBREF160 format, which is not very unlike the ACE format.","Events are extracted along with their event participants, e.g., who did what to whom when and where? For example, for a BUYING event, REES extracts the buyer, the artifact, the seller, and the time and location of the BUYING event. REES covers 61 types of events. There are 39 types of relations.","The tagging component consists of three modules: NameTagger, NPTagger and EventTagger. Each module relies on the same pattern-based extraction engine, but uses different sets of patterns. The NameTagger recognizes names of people, organizations, places, and artifacts (only vehicles in the implemented system). The NPTagger then takes the output of the NameTagger and first recognizes non-recursive Base Noun Phrase (BNP) BIBREF143 , and then complex NPs for only the four main semantic types of NPs, i.e., Person, Organization, Location, and Artifact (vehicle, drug and weapon). The EventTagger recognizes events applying its lexicon-driven, syntactically-based generic patterns.","REES uses a declarative, lexicon-driven approach. This approach requires a lexicon entry for each event-denoting word, which is generally a verb. The lexicon entry specifies the syntactic and semantic restrictions on the verb's arguments. After the tagging phase, REES sends the output through a rule-based co-reference resolution module that resolves: definite noun phrases of Organization, Person, and Location types, and singular personal pronouns. REES outputs the extracted information in the form of either MUC-style templates or XML.","One of the challenges of event extraction is to be able to recognize and merge those event descriptions which refer to the same event. The Template Generation module uses a set of declarative, customizable rules to merge co- referring events into a single event.","The system's recall, precision, and F-Measure scores for the training set (200 texts) and the blind set (208 texts) from about a dozen news sources. On the so-called training set, the system achieved F-measure of 64.75 for event extraction and 75.35 for relation extraction. The blind set F-Measure for 31 types of relations (73.95","As seen earlier in Subsection \"ACE Events\" , the way ACE events are specified, they have a lot of details that need to be extracted. BIBREF161 follows several steps to extract events and uses machine learning algorithms at every step. The steps are pre-processing of text data, identifying anchors, assigning event types, extracting arguments identifying attributes of events such as modality, polarity, genericity and tense, and finally identifyings event co-referents of the same individuated event. In other words, Ahn attempts to cover all the steps sequentially, making the simplifying assumption that they are unrelated to each other.","A single place in a textual document which may be considered the primary place of reference or discussion about an event is called the event anchor. Ahn treats finding the anchor for an event within a document as a word classification task, using a two-stage classification process. He uses a binary classifier to classify a word as being an event anchor or not. He then classifies those identified as event anchors into one of the event classes. Ahn used one classifier for binary classification and then another classifiers to classify only the positive instances.","Ahn treats identifying event arguments as a pair classification task. Each event mention is paired with each of the entity, time and value mentions occurring in the same sentence to form a single classification instance. There were 35 role types in the ACE 2006 task, but no event type allows arguments of all types. Each event type had its own set of allowable roles. The classification experiment run was a multi-class classification where a separate multi-class classifier was used for each event type. Ahn trains a separate classifier for each attribute. Genericity, modality, and polarity are each binary classification tasks, while tense is a multi-class task. For event coreference, Ahn follows the approach given in BIBREF162 . Each event mention in a document is paired with every other event mention, and a classifier assigns to each pair of mentions the probability that the paired mentions corefer. These probabilities are used in a left-to-right entity linking algorithm in which each mention is compared with all already-established events (i.e., event mention clusters) to determine whether it should be added to an existing event or start a new one.","Ahn experimented with various combinations of a maximum entropy classifier MegaM BIBREF163 and a memory-based nearest neighbor classifier called TIMBL BIBREF164 , for the various tasks.","The ACE specification provided a way to measure the performance of an event extraction system. The evaluation called ACE value is obtained by scoring each of the component tasks individually and then obtaining a normalized summary value. Overall, using the best learned classifiers for the various subtasks, they achieve an ACE value score of 22.3%, where the maximum score is 100%. The value is low, but other systems at the time had comparable performance."," BIBREF165 describe an approach to classify sentences in a document as specifying one or more events from a certain ACE 2006 class. They classify each sentence in a document as containing an instance of a certain type or not. Unlike BIBREF161 , they are not interested in identifying arguments or any additional processing. Also, unlike Ahn who classifies each word as possibly being an event anchor for a specific type of ACE event, Naughton et al. perform a classification of each sentence in a document as being an on-event sentence or an off-event sentence. An on-event sentence is a sentence that contains one or more instances of the target event type. An off-event sentence is a sentence that does not contain any instances of the target event type. They use several approaches to classify a sentence as on-event or off-event. These include the following: SVM-based machine learning BIBREF166 , language modeling approaches using count smoothing, and a manual approach which looks for Wordnet synonyms or hypernyms of certain trigger words in a sentence.","Naughton et al. found that 1) use of a large number of features to start but then reduction of these features using information gain, and 2) use of SVM produces the best results although all versions of SVM (i.e., with all features with no reduction, just the terms without complex features, or a selection of terms and other features) all work very well. A “surprising\" finding was that the “manual\" trigger-based classification approach worked almost as well as the SVM based approaches."],"section_name":"Extracting Events"},{"paragraphs":["When an event is mentioned in several places within a document, finding which references are to the same event is called determining event coreference. These are co-referents to the event. Determining when two event mentions in text talk about the same event or co-refer is a difficult problem. As BIBREF167 point out that the events may be actual occurrences or hypothetical events."," BIBREF162 present a statistical language-independent framework for identifying and tracking named, nominal and pronominal references to entities within unrestricted text documents, and chaining them into groups corresponding to each logical entity present in the text. The model can use arbitrary feature types, integrating a variety of lexical, syntactic and semantic features. The mention detection model also uses feature streams derived from different named entity classifiers.","For mention detection, the approach used is based on a log-linear Maximum Entropy classifier BIBREF168 and a linear Robust Risk Minimization classifier BIBREF169 . Then they use a MaxEnt model for predicting whether a mention should or should not be linked to an existing entity, and to build entity chains. Both classifiers can integrate arbitrary types of information and are converted into suitable for sequence classification for both tasks.","For entity tracking, the process works from left to right. It starts with an initial entity consisting of the first mention of a document, and the next mention is processed by either linking it with one of the existing entities, or starting a new entity. Atomic features used by the entity linking algorithm include string match, context, mention count, distance between the two mentions in words and sentences, editing distance, properties of pronouns such gender, number and reflexiveness. The best combination of features was able to obtain slightly more than 73% F-1 value using both RRM and MaxEnt algorithms for mention detection.","Entity tracking was evaluated in terms of what is called the ACE value. A gauge of the performance of an EDT system is the ACE value, a measure developed especially for this purpose. It estimates the normalized weighted cost of detection of specific-only entities in terms of misses, false alarms and substitution errors. Florian et al. achieved an ACE value of 73.4 out of 100 for the MaxEnt classifier and 69.7 for the RRM classifier."," BIBREF161 follows the approach by BIBREF162 for entity coreference determination. He uses a binary classifier to determine if any two event mentions in the document refer to the same event. Thus, he pairs each event with every other event, and the classifier assigns each pair a probability that they are the same. The probability is used with entity linking/matching algorithm to determine event co-reference. Event co-referencing requires event mentions to be clustered to event clusters. Event mentions in a cluster are the same event. The system described here obtained an ACE value of between 88-91%, where the maximum ACE value is 100%."," BIBREF161 uses the following features for event co-reference determination. Let the candidate be the earlier event mention and the anaphor be the later mention.","The anchors for the candidate and the anaphor, the full or original form, and also in lowercase, and POS tag.","Type of the candidate event and the anaphor event.","Depth of candidate anchor word in parse tree.","Distance between the candidate and anchor, measured in sentences.","Number, heads, and roles of shared arguments, etc.","Supervised approaches to solving event coreference use linguistic properties to decide if a pair of event mentions is coreferential BIBREF170 , BIBREF171 , BIBREF161 , BIBREF172 . These models depend on labeled training data, and annotating a large corpus with event coreference information requires substantial manual effort. In addition, since these models make local pairwise decisions, they are unable to capture a global event distribution at topic or document collection level. BIBREF173 present how nonparametric Bayesian models can be applied to an open-domain event coreference task in an unsupervised manner.","The first model extends the hierarchical Dirichlet process BIBREF174 to take into account additional properties associated with event mentions. The second model overcomes some of the limitations of the first model, and uses the infinite factorial hidden Markov model BIBREF175 coupled to the infinite hidden Markov model BIBREF176 in order to consider a potentially infinite number of features associated with observable objects which are event mentions here, perform an automatic selection of the most salient features, and capture the structural dependencies of observable objects or event mentions at the discourse level. Furthermore, both models can work with a potentially infinite number of categorical outcomes or events in this case.","Two event mentions corefer if they have the same event properties and share the same event participants. To find coreferring event mentions, Bejan and Harabagiu describe words that may be possible event mentions with lexical features, class features such as POS and event classes such BIBREF119 as occurrence, state and action, Wordnet features, semantic features obtained by a semantic parse BIBREF177 and the predicate argument structures encoded in PropBank annotations BIBREF89 as well as semantic annotations encoded in the FrameNet corpus BIBREF79 .","The first model represents each event mention by a finite number of feature types, and is also inspired by the Bayesian model proposed by BIBREF178 . In this model, a Dirichlet process (DP) BIBREF179 is associated with each document, and each mixture component (i.e., event) is shared across documents since In the process of generating an event mention, an event index z is first sampled by using a mech- anism that facilitates sampling from a prior for in- finite mixture models called the Chinese restaurant franchise (CRF) representation, as reported in BIBREF174 .","The second model they use is called the iHMM-iFHMM model (infinite hidden Markov model–infinite factorial hidden Markov model). The iFHMM framework uses the Markov Indian buffet process (mIBP) BIBREF175 in order to represent each object as a sparse subset of a potentially unbounded set of latent features BIBREF180 , BIBREF181 , Specifically, the mIBP defines a distribution over an unbounded set of binary Markov chains, where each chain can be associated with a binary latent feature that evolves over time according to Markov dynamics. The iFHMM allows a flexible representation of the latent structure by letting the number of parallel Markov chains be learned from data, it cannot be used where the number of clustering components is infinite. An iHMM represents a nonparametric extension of the hidden Markov model (HMM) BIBREF182 that allows performing inference on an infinite number of states. To further increase the representational power for modeling discrete time series data, they develop a nonparametric extension that combines the best of the two models, and lets the two parameters M and K be learned from data Each step in the new iHMM-iFHMM generative process is performed in two phases: (i) the latent feature variables from the iFHMM framework are sampled using the mIBP mechanism; and (ii) the features sampled so far, which become observable during this second phase, are used in an adapted version of the beam sampling algorithm BIBREF183 to infer the clustering components (i.e., latent events).","They report results in terms of recall (R), precision (P), and F-score (F) by employing the mention-based B3 metric BIBREF184 , the entity-based CEAF metric BIBREF185 , and the pairwise F1 (PW) metric. Their experiments for show that both of these models work well when the feature and cluster numbers are treated as free parameters, and the selection of feature values is performed automatically."," BIBREF167 argue that events represent complex phenomena and can therefore co-refer fully, being identical, like other researchers have discussed, or co-refer partially, being quasi-identical or only partially identical. Two event mentions fully co-refer if their activity, event or state representation is identical in terms of all features used (e.g., agent, location or time). Two event mentions are quasi-identical if they partially co-refer, i.e., most features are the same, but there may be additional details to one or the other.","When two events fully co-refer, Hovy et al. state they may be lexically identical (i.e., the same senses of the same word, e.g., destroy and destruction), synonymous words, one mention is a wider reading of the other (e.g., The attack took place yesterday and The bombing killed four people), one mention is a paraphrase of the other with possibly some syntactic differences (e.g., He went to Boston and He came to Boston), and one mention deictically refers to the other (e.g., the party and that event). Quasi-identity or partial co-reference may arise in two ways: membership identity or subevent identity. Membership identity occurs when one mention, say A, is a set of multiple instances of the same type of event, and the other mention, say B, is one of the individual events in A (e.g., I attended three parties last week. The first one was the best.). Subevent identity is found when one mention, say A, is a stereotypical sequence (or script) of events whereas the other mention, say B, is one of the actions or events within the script (e.g., The family ate at the restaurant. The dad paid the waitress at the end.).","Hovy et al.attempt to build a corpus containing event co-reference links with high quality annotations, i.e., annotations with high inter-annotator agreement, to be useful for machine learning. They have created two corpora to assist with a project on automated deep reading of texts. One corpus is in the domain of violent events (e..g., bombings, killens and wars), and the other one containing texts about the lives of famous people. In both of these corpora, they have annotated a limited number of articles with full and partial co-references."," BIBREF186 claims that performing event co-reference with high accuracy requires deep understanding of the text and statistically-based methods, both supervised and unsupervised, do not perform well. He claims that this is the case because because it is absolutely necessary to identify arguments of an event reliably before event co-references can be found. Arguments are difficult to identify because many are implicit and linguistically unexpressed. Successful even co-reference identification needs determination of spatio-temporal anchoring and locations in time and space are also very often implicit.","The system he builds uses a linguistically based semantic module, which has a number of different submodules which take care of Spatio-Temporal Reasoning, Discourse Level Anaphora Resolution, and determining Topic Hierarchy. The coreference algorithm works as follows: for each possible referent it check all possible coreference links, at first using only the semantic features, which are: wordform and lemma identity; then semantic similarity measured on the basis of a number of similarity criteria which are lexically based. The system searches WordNet synsets and assign a score according to whether the possible referents are directly contained in the same synset or not. A different score is assigned if their relation can be inferred from the hierarchy. Other computational lexical resources they use include FrameNet and Frames hierarchy; SumoMilo and its semantic classification.","After collecting all possible coreferential relations, the system filters out those links that are inconsistent or incompatible. Argument structure and spatiotemporal relations are computed along with dependence relations; temporal logical relations as computed using an adaptation of Allen's algorithm. The system also computes semantic similarity, where high values are preferred. The paper does not give any results to support the initial hypothesis, although the ideas are interesting."," BIBREF187 use granularity in computing event coreference. The intuition is, that an event with a longer duration, that happens on a bigger area and with multiple particpants (for instance a war between Russia and Ukraine) might be related to but will probably not fully corefer with a Òlower levelÓ event of shorter duration and with single participants involved (e.g. A Russian soldier has shot dead a Ukrainian naval officer).","Coreference between mentions of two events is determined by computing compatibility of contents of event attributes. The attributes used are event trigger, time, location, human and non-human participant slots BIBREF188 . Granularity size is mentioned in terms of durations of event actions BIBREF189 and granularity levels of event participants, time and locations. Granularity is given in terms of partonomic relations or through the part-of relation, between entities and events, using the taxonomy of meronymic relations by BIBREF190 . Granularity levels of the human participant slot are contained within WinstonÕs et al. Member-Collection relations. The temporal granularity levels make part of WinstonÕs Portion-Mass relationships and locational levels are in line with Place-Area relations in WinstonÕs taxonomy.","Cybulska and Vossen experimented with a decision-tree supervised pairwise binary classifier to determine coreference of pairs of event mentions. They also ran experiments with a linear SVM and a multinomial Naive Bayes classifier but the decision-tree classifier outperformed both of them.","For the experiments, Cybulska and Vossen use the ECB+ dataset BIBREF191 . The ECB+ corpus contains a new corpus component, consisting of 502 texts, describing different instances of event types. They provide results in terms of several metrics: recall, precision and F-score, MUC BIBREF192 , B3 BIBREF184 , mention-based CEAF BIBREF185 , BLANC BIBREF193 , and CoNLL F1 BIBREF194 , and find that the introduction of the granularity concept into similarity computation improves results for every metric."],"section_name":"Determining Event Coreference"},{"paragraphs":["Researchers are interested in extracting information from the huge amount of biomedical literature published on a regular basis. Of course, one aspect of information extraction is event extraction, the focus of this paper. In the biomedical context, an event extraction system tries to extract details of bimolecular interactions among biomedical entities such as proteins and genes, and the processes they take part in, as described in terms of textual documents. Manually annotated corpora are used to train machine learning techniques and evaluate event extraction techniques.","There have been several workshops on biomedical natural language processing. We focus on the BioNLP Shared Tasks in recent years that had competitions on event extraction. There have been three BioNLP Shared Task competitions so far: 2009, 2011, and 2013. The BioNLP 2009 Shared Task BIBREF195 was based on the GENIA corpus BIBREF196 which contains PubMed abstracts of articles on transcription factors in human blood cells. There was a second BioNLP Shared Task competition organized in 2011 to measure the advances in approaches and associated results BIBREF197 . The third BioNLP ST was held in 2013. We discuss some notable systems from BioNLP ST 2011 and 2013.","Before the BioNLP Shared Tasks, event extraction in the biomedical domain usually classified each pair of named entities (usually protein names) co-occurring in the text as interacting or not. BioNLP Shared Tasks extended such an approach by adding relations such as direction, type and nesting. An event defines the type of interaction, such as phosphorylation, and is usually marked in the text with a trigger word (e.g., phosphorylates) describing the interaction. This word forms the core of the event description. A directed event has roles that have inherent directionality such as cause or theme, the agent or target of the biological process. In addition, events can act as arguments of other events, creating complex nested structures. For example, in the sentence Stat3 phosphorylation is regulated by Vav, a phosphorylation-event is the argument of the regulation-event.","The BioNLP Shared Tasks provide task definitions, benchmark data and evaluations, and participants compete by developing systems to perform the specified tasks. The theme of BioNLP-ST 2011 was a generalization of the 2009 contest, generalized in three ways: text types, event types, and subject domains. The 2011 event-related tasks were arranged in four tracks: GENIA task (GE) BIBREF197 , Epigenetics and Post-translational Modifications (EPI) BIBREF198 , Infectious Diseases (ID) BIBREF199 , and the Bacteria Track BIBREF200 , BIBREF201 .","Of the four event-related shared tasks in BioNLP 2011, the first three were related to event extraction. The Genia task was focused on the domain of transcription factors in human blood cell. Trascription is a complex but just the first step in the process in which the instructions contained in the DNA in the nucleus of a cell are used to produce proteins that control most life processes. Transcription factors are proteins that control the transcription process. The EPI task was focused on events related to epigenetics, dealing with protein and DNA modifications, with 14 new event types, including major protein modification types and their reverse reactions. Epigenesis refers to the development of a plant or animal from a seed, spore or egg, through a sequence of steps in which cells differentiate and organs form. The EPI task was designed toward pathway extraction and curation of domain databases BIBREF202 , BIBREF203 . A biological pathway refers to a sequence of actions among molecules in a cell that leads to a certain product or a change in the cell. The ID task was focused on extraction of events relevant to biomolecular mechanisms of infectious diseases from full length publications. Tasks other than ID focused on abstracts only.","In this paper, we discuss the systems and approaches for only the 2011 GE Task. This is because several of the winning systems for the GE Task did well in the other two relevant tasks as well. The Genia Task is described in Table 7 . The table shows for each event type, the primary and secondary arguments to be extracted. For example, a phosphorylation event is primarily extracted with the protein to be phosphorylated, which is the addition of a phosphate group to a protein or other organic molecule. As secondary information, the specific site to be phosphorylated may be extracted. From a computational viewpoint, the event types represent different levels of complexity. When only primary arguments are considered, the first five event types in Table 7 are classified as simple events, requiring only unary arguments. The binding and regulation types are more complex. Binding requires the detection of an arbitrary number of arguments, and Regulation requires detection of recursive event structure.","Consider the sentence In this study we hypothesized that the phosphorylation of TRAF2 inhibits binding to the CD40 cytoplasmic domain. Here there are two protein (entity) names: TRAF2 and CD40. The word phosphorylation refers to an event; this string is a trigger word. Thus, the goal of the GE task was to identify a structure like the ones in Tables 8 and 9 . In the tables, $T_i$ represents a trigger word, and $E_i$ represents an event associated with the corresponding trigger word. There are three events, $E_1$ is the phosphorylation event, $E_2$ is the binding event and $E_3$ is the negative regulation event. For each trigger word, we see the starting and ending character positions in the entire string. For each event, we see the participants in it. The second task identifies an additional site argument.","Table 10 shows the best results for various tasks in the BioNLP 2011 contests. BIBREF197 note an improvement of 10% over the basic GE task, in 2011 (Task GEa), compared to 2009. The results of the GE tasks show that automatic extraction of simple events–those with unary arguments, e.g., gene expression, localization and phosphorylation—can be achieved at about 70% in F-score, but the extraction of complex events, e.g., binding and regulation is very challenging, with only 40% performance level. The GE and ID results show that generalization to full papers is possible, with just a small loss in performance. The results of phosphorylation events in GE and EP are similar (GEp vs. EPIp), which leads BIBREF197 to conclude that removal of the GE domain specificity does not reduce event extraction performance by much. EPIc results indicate that there are challenges to extracting similar event types that need to be overcome; EPIf results indicate that there are difficult challenges in extracting additional arguments. The complexity of the ID task is similar to that of the GE task; this shows up in the final results, also indicating that it is possible to generalize to new subject domains and new argument (entity) types.","Below, we provide a brief description of some of the approaches to biomedical event extraction from the BioNLP 2011 contests."],"section_name":"Biomedical Event Extraction"},{"paragraphs":["The team that won the GE Task was the FAUST system BIBREF204 , followed by the UMass system BIBREF205 , then the UTurku system BIBREF206 . The performance of these three systems on the various tasks is given in Table 11 . In addition, we have the Stanford system in the table because it performed fairly well on the tasks.","The UMass system BIBREF205 looks at a sentence as having an event structure, and then projects it onto a labeled graph. See Figure 6 for a target event structure and the projected graph for the sentence fragment Phosphorylation of TRAF2 inhibits binding to CD40. The system searches for a structure that connects the event and its participating entities and imposes certain constraints on the structure. Thus, the UMass system treats the search for such a structure as an optimization problem. To formulate this optimization problem, the system represents the structure in terms of a set of binary variables, inspired by the work of BIBREF207 , BIBREF208 . These binary variables are based on the projection of the events to the labeled graph. An example of a binary variable is $a_{i,l.r}$ to indicate that between positions $i$ and $l$ in the sentence, there is an edge labeled $r$ from a set of possible edge labels $R$ . Another such binary variable is $t_{i,p,q}$ that indicates that at position $i$ , there is a binding event with arguments $p$ and $q$ . Given a number of such variables, it is possible to write an objective function to optimize in order to obtain events and entity bindings. The system decomposes the biomedical event extraction task into three sub-tasks: (a) event triggers and outgoing edges on arguments, (b) event triggers and incoming edges on arguments, and (c) and protein-protein bindings. The system obtains an objective function for each of the sub-tasks. It solves the three optimization problems one by one in a loop, till no changes take place, or up to a certain number of iterations. The approach uses optimizing by dual decomposition BIBREF209 , BIBREF210 since the dual of the original optimization problem is solved.","The Stanford system BIBREF211 exploits the observation that event structures bear a close relation to dependency graphs BIBREF212 . They cast bimolecular events in terms of these structures which are pseudo-syntactic in nature. They claim that standard parsing tools such as maximum-spanning tree parsers and parse rerankers can be applied to perform event extraction with minimum domain specific training. They use an off-the-shelf dependency parser, MSTParser BIBREF213 , BIBREF214 , but extend it with event-specific features. Their approach requires conversion to and from dependency trees, at the beginning and and at the end. The features in the MSTParser are quite local (i.e., able to examine a portion of each event at a time); the decoding necessary can be performed globally, allowing the dependency parser some trade-offs. Event parsing is performed using three modules: 1) anchor detection to identify and label event anchors, 2) event parsing to form candidate event structures by linking entries and event anchors, and 3) event reranking to select the best candidate event structure. First, they parse the sentences with a reranking parser BIBREF215 with the biomedical parsing model from BIBREF216 , using the set of Stanford dependencies BIBREF217 . After the parsing, they perform anchor detection using a technique inspired by techniques for named entity recogntion to label each token with an event type or none, using a logistic regression classifier. The classifier uses features inspired by BIBREF208 . They change a parameter to obtain high recall to overgenerate event anchors. Multiword event anchors are reduced to their syntactic head. The event anchors and the included entities become a “reduced\" sentence, input to the event parser. Thus, the event parser gets words that are believed to directly take part in the events. This stage uses the MSTParser with additional event parsing features. The dependency trees are decoded and converted back to event structures. Finally, for event reranking, the system gets $n$ best list of event structures from each decoder in the previous step of event parsing. The reranker uses global features of an event structure to restore and output the highest scoring structure. The reranking approach is based on parse reranking BIBREF218 , but is based on features of event structures instead of syntactic constituency structure. They use the cvlm estimator BIBREF215 when learning weights for the reranking model. Since the reranker can work with outputs of multiple decoders, they use it as an ensemble technique as in BIBREF219 .","The FAUST system BIBREF204 shows that using a straightforward model combination strategy with two competitive systems, the UMass system BIBREF205 and the Stanford system BIBREF211 just described, can produce a new system with substantially high accuracy. The new system uses the framework of stacking BIBREF220 . The new system does it by including the predictions of the Stanford system into the UMass system, simply as a feature. Using this simple model of stacking, the FAUST system was able to obtain first place in three tasks out of four where it participated.","The Turku Event Extraction System BIBREF206 , BIBREF221 can be easily adapted to different event schemes, following the theme of event generalization in BioNLP 2011. The system took part in eight tasks in BioNLP 2011 and demonstrated the best performance in four of them. The Turku system divides event extraction into three main steps: i) Perform named entity recognition in the sentence, ii) Predict argument relations between entities, and iii) Finally, separate entity/argument sets into individual events. The Turku system uses a graph notation with trigger and protein/gene entities as nodes and relations (e.g., theme) as edges. In particular, an event in the graph representation is a trigger node along with its outgoing edges. The steps are shown in Figure 7 .","The Turku system uses Support Vector Machines BIBREF222 , BIBREF223 at various stages to perform each of the sub-tasks. To use an SVM classifier, one needs to convert text into features understood by the classifier. The Turku system performs a number of analyses on the sentences, to obtain features, which are mostly binary. The features are categorized into token features (e.g., Porter-stem BIBREF224 , Penn Treebank part-of-speech tags BIBREF102 , character bi- and tri-grams, presence of punctuation on numeric characters), sentence features (e.g., the number of named entities in the sentence), dependency chains (up to a depth of three, to define the context of the words), dependency with $n$ -grams (joining a token with two flanking dependencies as well as each dependency with two flanking tokens), trigger features (e.g., the trigger word a gene or a protein) and external features (e.g., Wordnet hypernyms, the presence of a word in a list of key terms). Applicable combinations of these features are then used by the three steps in event detection: trigger detection, edge detection and unmerging. Trigger words are detected by classifying each token as negative or as one of the positive trigger classes using SVMs. Sometimes several triggers overlap, in which case a merged class (e.g. phosphorylation–regulation) is used. After trigger prediction, triggers of merged classes are split into their component classes. Edge detection is used to predict event arguments or triggerless events and relations, all of which are defined as edges in the graph representation. The edge detector defines one example per direction for each pair of entities in the sentence, and uses the SVM classifier to classify the examples as negatives or as belonging to one of the positive classes. When edges are predicted between these nodes, the result is a merged graph where overlapping events are merged into a single node and its set of outgoing edges. To produce the final events, these merged nodes need to be Òpulled apartÓ into valid trigger and argument combinations. Unmerging is also performed using the SVM classifier. Speculation and negation are detected independently, with binary classification of trigger nodes using SVMs. The features used are mostly the same as for trigger detection, with the addition of a list of speculation-related words."],"section_name":"Technical Methods Used in BioNLP Shared Tasks 2011"},{"paragraphs":["With the explosive expansion of the Internet during the past twenty years, the volume of socially generated text has skyrocketed. Socially generated text includes blogs and microblogs. For example, Twitter, started in 2006, has become a social phenomenon. It allows individuals with accounts to post short messages that are up to 140 characters long. Currently, more than 340 million tweets are sent out every day. While a majority of posts are conversational or not particularly meaningful, about 3.6% of the posts concern topics of mainstream news. Twitter has been credited with providing the most current news about many important events before traditional media, such as the attacks in Mumbai in November 2008. Twitter also played a prominent role in the unfolding of the troubles in Iran in 2009 subsequent to a disputed election, and the so-called Twitter Revolutions in Tunisia and Egypt in 2010-11.","Most early work on event extraction of information from documents found on the Internet has focussed on news articles BIBREF225 , BIBREF226 , BIBREF227 . However, as noted earlier, social networking sites such as Twitter and Facebook have become important complimentary sources of such information. Individual tweets, like SMS messages, are usually short and self-contained and therefore are not composed of complex discourse structures as is the case with texts containing narratives. However, extracting structured representation of events from short or informal texts is also challenging because most tweets are about mundane things, without any news value and of interest only to the immediate social network. Individual tweets are also very terse, without much context or content. In addition, since Twitter users can talk about any topic, it is not clear a priori what event types may be appropriate for extraction.","The architecture of the system called TwiCal for event extraction BIBREF228 from Twitter messages is given in Figure 8 . Given a stream of raw tweets, TwiCal extract events with associated named entities and times of occurrence. First the tweets are POS tagged using a tagger BIBREF228 , especially trained with Twitter data. Then named entities are recognized BIBREF229 using a recognizer trained with Twitter data as well. After this, phrases that mention events (or, event triggers or event phrases or just events) are extracted using supervised learning. BIBREF228 annotated 1,000 tweets with event phrases, following guidelines for annotation of EVENT tags in Timebank BIBREF119 . The system recognizes event triggers as a sequence labeling task using Conditional Random Fields BIBREF147 . It uses a contextual dictionary, orthographic features, features based on the Twitter-tuned POS tagger, and dictionaries of event terms gathered from WordNet BIBREF121 . Once a large number of events have been extracted by this CRF learner, TwiCal categorizes these events into types using an unsupervised approach based on latent variable models, inspired by work on modeling selectional preferences BIBREF230 , BIBREF231 , BIBREF232 , BIBREF233 and unsupervised information extraction BIBREF234 , BIBREF225 , BIBREF235 . This automatic discovery of event types is similar to topic modeling, where one automatically identifies the extant topics in a corpus of text documents. The automatically discovered types (topics) are quickly inspected by a human effort to filter out incoherent ones, and the rest are annotated with informative labels. Examples of event types discovered along with top event phrases and top entities are given in Table 12 . The resulting set of types are applied to categorize millions of extracted events without the use of any manually annotated examples. For inference, the system uses collapsed Gibbs sampling BIBREF236 and prediction is performed using a streaming approach to inference BIBREF237 . To resolve temporal expressions, TwiCal uses TempEx BIBREF238 , which takes as input a reference date, some text and POS tags, and marks temporal expressions with unambiguous calendar references. Finally, the system measures the strength of association between each named entity and date based on the number of tweets they co-occur in, in order to determine if the event is significant. Examples of events extracted by TwiCal are given in Table 13 . Each event is a 4-tuple including a named entity, event phrase, calendar date and event type.","The TwiCal system describe above used topic modeling using latent variables as one of the several computational components; it is used to capture events captured using supervised learning into types or topics. BIBREF239 point out some drawbacks of using such an approach. The main problem is that frequently the result generated by Latent Dirichlet Analysis (LDA) is difficult to interpret because it simply gives a list of words associate with the topic. For example, when BIBREF239 attempt to find the four most important topics using LDA based on a Twitter collection emanating from Singapore on June 16, 2010, they find the topics listed in Table 14 . Therefore, Weng et al. present another approach to detect events from a corpus of Twitter messages. Their focus is on detection and therefore, not on extraction of components that describe an event. Event detection is based on the assumption that when an event is taking place, some related words show an increase in usage. In this scheme, an event is represented by a number of keywords showing a burst in appearance count BIBREF240 , BIBREF241 . Although it is clear that tweets report events, but such reports are usually overwhelmed by high flood of meaningless “babbles\". In addition, the algorithms for event detection must be scalable to handle the torrent of Twitter posts. The EDCoW (Event Detection with Clustering of Wavelet-based Signals) system builds signals for individual words by applying wavelet analysis on frequency-based raw signals of words occurring in the Twitter posts. These signals capture only the bursts in the words' appearance. The signals are computed efficiently by wavelet analysis BIBREF242 , BIBREF243 . Wavelets are quickly vanishing oscillating functions and unlike sine and cosine functions used in Discrete Fourier Transformation (DFT) which are localized in frequency but extend infinitely in time, wavelets are localized both in time and frequency. Therefore, wavelet transformation is able to provide precise measurements about when and to what extent bursts take place in a signal. BIBREF239 claim that this makes it a better choice for event detection when building signals for individual words. Wavelet transformation converts signals from time domain to time-scale domain where scale can be considered the inverse of frequency. Such signals also take less space for storage. Thus, the first thing EDCoW does is convert frequencies over time to wavelets, using a sliding window interval. It removes trivial words by examining signal auto-correlations. The remaining words are then clustered to form events with a modularity-based graph partitioning technique, which uses a scalable eigenvalue algorithm. It detects events by grouping sets of words with similar patterns of burst. To cluster, similarities between words need to be computed. It does so by using cross correlation, which is a common measure of similarity between two signals BIBREF244 . Cross correlation is a pairwise operation. Cross correlation values among a number of signals can be represented in terms of a correlation matrix $\\mathcal {M}$ , which happens to be a symmetric sparse matrix of adjacent similarities. With this graph setup, event detection can be formulated as a graph partitioning problem, i.e., to cut the graph into subgraphs. Each subgraph corresponds to an event, which contains a set of words with high cross correlation, and also that the cross correlation between words in different subgraphs are low. The quality of such partitioning is measures using a metric called modularity BIBREF245 , BIBREF246 . The modularity of a graph is defined as the sum of weights of all the edges that fall within subgraphs (after partitioning) subtracted by the expected edge weight sum if the edges were placed at random. The main computation task in this component is finding the largest eigenvalue and corresponding eigenvector, of the sparse symmetric modularity matrix. This is solved using power iteration, which is able to scale up with the increase in the number of words in the tweets BIBREF247 . EDCoW requires each individual event to contain at least two words. To differentiate big events from trivial ones, EDCoW quantifies the events' significance, which depends on two factors, the number of words and cross-correlation among the words related to the event. To make EDCoW work with TwiCal to see if it improves performance, the topic detection module will have to be replaced. EDCoW associates fewer words to topics because it filters words away before associating with a topic. Table 15 gives a few event words obtained by EDCoW and the corresponding event description. Please note that the event description was created by the authors and not the system."],"section_name":"Extracting Events from Socially Generated Documents"},{"paragraphs":[" BIBREF248 use event-based features to represent sentences and shows that their approach improves the quality of the final summaries compared to a baseline bag-of-words approach."],"section_name":"Summarization"},{"paragraphs":["Event recognition is a core task in question-answering since the majority of web questions have been found to be relate to events and situations in the world BIBREF121 . For example, to answer the question How many people were killed in Baghdad in March?, or Who was the Prime MInister of India in when China and India fought their only war?, the question-answering system may have to identify events across a bunch of documents before creating an answer."],"section_name":"Question Answering"},{"paragraphs":["It also seems like when doctors take notes on a patient's history or medical record, the information is not written in order of events or in temporal order all the time. It will be good to take notes from here and there and put them in an event ordered fashion or temporally ordered manner. Extracting an event based structure of the medical record would help understand the medical history better.","Most systems process sentences in isolation, like most event extraction systems at the current time. Therefore, events crossing sentence boundaries cannot be detected.","... "],"section_name":"Future Directions of Research"}],"paper_id":"1601.04012","title":"Detecting and Extracting Events from Text Documents"},{"abstract":"One of the limitations of semantic parsing approaches to open-domain question answering is the lexicosyntactic gap between natural language questions and knowledge base entries -- there are many ways to ask a question, all with the same answer. In this paper we propose to bridge this gap by generating paraphrases of the input question with the goal that at least one of them will be correctly mapped to a knowledge-base query. We introduce a novel grammar model for paraphrase generation that does not require any sentence-aligned paraphrase corpus. Our key idea is to leverage the flexibility and scalability of latent-variable probabilistic context-free grammars to sample paraphrases. We do an extrinsic evaluation of our paraphrases by plugging them into a semantic parser for Freebase. Our evaluation experiments on the WebQuestions benchmark dataset show that the performance of the semantic parser significantly improves over strong baselines.","figures_and_tables":[{"caption":"Figure 1: An example word lattice for the question What language do people in Czech Republic speak? using the lexical and phrasal rules from the PPDB.","file":"4-Figure1-1.png"},{"caption":"Figure 2: Trees used for bi-layered L-PCFG training. The questions what day is nochebuena, when is nochebuena and when is nochebuena celebrated are paraphrases from the Paralex corpus. Each nonterminal is decorated with a syntactic label and two identifiers, e.g., for WP-7-254, WP is the syntactic label assigned by the BLLIP parser, 7 is the syntactic latent state, and 254 is the semantic latent state.","file":"5-Figure2-1.png"},{"caption":"Figure 3: Ungrounded graphs for an input question and its paraphrases along with its correct grounded graph. The green squares indicate NL or Freebase entities, the yellow rectangles indicate unary NL predicates or Freebase types, the circles indicate NL or Freebase events, the edge labels indicate binary NL predicates or Freebase relations, and the red diamonds attach to the entity of interest (the answer to the question).","file":"7-Figure3-1.png"},{"caption":"Table 1: Oracle statistics and results on the WebQuestions development set.","file":"9-Table1-1.png"},{"caption":"Table 2: Results on WebQuestions test dataset.","file":"9-Table2-1.png"}],"full_text":[{"paragraphs":["Semantic parsers map sentences onto logical forms that can be used to query databases BIBREF0 , BIBREF1 , instruct robots BIBREF2 , extract information BIBREF3 , or describe visual scenes BIBREF4 . In this paper we consider the problem of semantically parsing questions into Freebase logical forms for the goal of question answering. Current systems accomplish this by learning task-specific grammars BIBREF5 , strongly-typed CCG grammars BIBREF6 , BIBREF7 , or neural networks without requiring any grammar BIBREF8 . These methods are sensitive to the words used in a question and their word order, making them vulnerable to unseen words and phrases. Furthermore, mismatch between natural language and Freebase makes the problem even harder. For example, Freebase expresses the fact that “Czech is the official language of Czech Republic” (encoded as a graph), whereas to answer a question like “What do people in Czech Republic speak?” one should infer people in Czech Republic refers to Czech Republic and What refers to the language and speak refers to the predicate official language.","We address the above problems by using paraphrases of the original question. Paraphrasing has shown to be promising for semantic parsing BIBREF9 , BIBREF10 , BIBREF11 . We propose a novel framework for paraphrasing using latent-variable PCFGs (L-PCFGs). Earlier approaches to paraphrasing used phrase-based machine translation for text-based QA BIBREF12 , BIBREF13 , or hand annotated grammars for KB-based QA BIBREF10 . We find that phrase-based statistical machine translation (MT) approaches mainly produce lexical paraphrases without much syntactic diversity, whereas our grammar-based approach is capable of producing both lexically and syntactically diverse paraphrases. Unlike MT based approaches, our system does not require aligned parallel paraphrase corpora. In addition we do not require hand annotated grammars for paraphrase generation but instead learn the grammar directly from a large scale question corpus.","The main contributions of this paper are two fold. First, we present an algorithm (§ \"Paraphrase Generation Using Grammars\" ) to generate paraphrases using latent-variable PCFGs. We use the spectral method of narayan-15 to estimate L-PCFGs on a large scale question treebank. Our grammar model leads to a robust and an efficient system for paraphrase generation in open-domain question answering. While CFGs have been explored for paraphrasing using bilingual parallel corpus BIBREF14 , ours is the first implementation of CFG that uses only monolingual data. Second, we show that generated paraphrases can be used to improve semantic parsing of questions into Freebase logical forms (§ \"Semantic Parsing using Paraphrasing\" ). We build on a strong baseline of reddylargescale2014 and show that our grammar model competes with MT baseline even without using any parallel paraphrase resources."],"section_name":"Introduction"},{"paragraphs":["Our paraphrase generation algorithm is based on a model in the form of an L-PCFG. L-PCFGs are PCFGs where the nonterminals are refined with latent states that provide some contextual information about each node in a given derivation. L-PCFGs have been used in various ways, most commonly for syntactic parsing BIBREF15 , BIBREF16 , BIBREF17 , BIBREF18 , BIBREF19 , BIBREF20 .","In our estimation of L-PCFGs, we use the spectral method of narayan-15, instead of using EM, as has been used in the past by matsuzaki-2005 and petrov-2006. The spectral method we use enables the choice of a set of feature functions that indicate the latent states, which proves to be useful in our case. It also leads to sparse grammar estimates and compact models.","The spectral method works by identifying feature functions for “inside” and “outside” trees, and then clusters them into latent states. Then it follows with a maximum likelihood estimation step, that assumes the latent states are represented by clusters obtained through the feature function clustering. For more details about these constructions, we refer the reader to cohen-13 and narayan-15.","The rest of this section describes our paraphrase generation algorithm."],"section_name":"Paraphrase Generation Using Grammars"},{"paragraphs":["We define our paraphrase generation task as a sampling problem from an L-PCFG $G_{\\mathrm {syn}}$ , which is estimated from a large corpus of parsed questions. Once this grammar is estimated, our algorithm follows a pipeline with two major steps.","We first build a word lattice $W_q$ for the input question $q$ . We use the lattice to constrain our paraphrases to a specific choice of words and phrases that can be used. Once this lattice is created, a grammar $G_{\\mathrm {syn}}^{\\prime }$ is then extracted from $G_{\\mathrm {syn}}$ . This grammar is constrained to the lattice.","We experiment with three ways of constructing word lattices: naïve word lattices representing the words from the input question only, word lattices constructed with the Paraphrase Database BIBREF14 and word lattices constructed with a bi-layered L-PCFG, described in § \"Bi-Layered L-PCFGs\" . For example, Figure 1 shows an example word lattice for the question What language do people in Czech Republic speak? using the lexical and phrasal rules from the PPDB.","Once $G_{\\mathrm {syn}}^{\\prime }$ is generated, we sample paraphrases of the input question $q$ . These paraphrases are further filtered with a classifier to improve the precision of the generated paraphrases.","We train the L-PCFG $G_{\\mathrm {syn}}$ on the Paralex corpus BIBREF9 . Paralex is a large monolingual parallel corpus, containing 18 million pairs of question paraphrases with 2.4M distinct questions in the corpus. It is suitable for our task of generating paraphrases since its large scale makes our model robust for open-domain questions. We construct a treebank by parsing 2.4M distinct questions from Paralex using the BLLIP parser BIBREF25 .","Given the treebank, we use the spectral algorithm of narayan-15 to learn an L-PCFG for constituency parsing to learn $G_{\\mathrm {syn}}$ . We follow narayan-15 and use the same feature functions for the inside and outside trees as they use, capturing contextual syntactic information about nonterminals. We refer the reader to narayan-15 for more detailed description of these features. In our experiments, we set the number of latent states to 24.","Once we estimate $G_{\\mathrm {syn}}$ from the Paralex corpus, we restrict it for each question to a grammar $G_{\\mathrm {syn}}^{\\prime }$ by keeping only the rules that could lead to a derivation over the lattice. This step is similar to lexical pruning in standard grammar-based generation process to avoid an intermediate derivation which can never lead to a successful derivation BIBREF26 , BIBREF27 .","Sampling a question from the grammar $G_{\\mathrm {syn}}^{\\prime }$ is done by recursively sampling nodes in the derivation tree, together with their latent states, in a top-down breadth-first fashion. Sampling from the pruned grammar $G_{\\mathrm {syn}}^{\\prime }$ raises an issue of oversampling words that are more frequent in the training data. To lessen this problem, we follow a controlled sampling approach where sampling is guided by the word lattice $W_q$ . Once a word $w$ from a path $e$ in $W_q$ is sampled, all other parallel or conflicting paths to $e$ are removed from $W_q$ . For example, generating for the word lattice in Figure 1 , when we sample the word citizens, we drop out the paths “human beings”, “people's”, “the population”, “people” and “members of the public” from $W_q$ and accordingly update the grammar. The controlled sampling ensures that each sampled question uses words from a single start-to-end path in $W_q$ . For example, we could sample a question what is Czech Republic 's language? by sampling words from the path (what, language, do, people 's, in, Czech, Republic, is speaking, ?) in Figure 1 . We repeat this sampling process to generate multiple potential paraphrases.","The resulting generation algorithm has multiple advantages over existing grammar generation methods. First, the sampling from an L-PCFG grammar lessens the lexical ambiguity problem evident in lexicalized grammars such as tree adjoining grammars BIBREF27 and combinatory categorial grammars BIBREF28 . Our grammar is not lexicalized, only unary context-free rules are lexicalized. Second, the top-down sampling restricts the combinatorics inherent to bottom-up search BIBREF29 . Third, we do not restrict the generation by the order information in the input. The lack of order information in the input often raises the high combinatorics in lexicalist approaches BIBREF30 . In our case, however, we use sampling to reduce this problem, and it allows us to produce syntactically diverse questions. And fourth, we impose no constraints on the grammar thereby making it easier to maintain bi-directional (recursive) grammars that can be used both for parsing and for generation BIBREF31 ."],"section_name":"Paraphrases Generation Algorithm"},{"paragraphs":["As mentioned earlier, one of our lattice types is based on bi-layered PCFGs introduced here.","In their traditional use, the latent states in L-PCFGs aim to capture syntactic information. We introduce here the use of an L-PCFG with two layers of latent states: one layer is intended to capture the usual syntactic information, and the other aims to capture semantic and topical information by using a large set of states with specific feature functions.","To create the bi-layered L-PCFG, we again use the spectral algorithm of narayan-15 to estimate a grammar $G_{\\mathrm {par}}$ from the Paralex corpus. We use the word alignment of paraphrase question pairs in Paralex to map inside and outside trees of each nonterminals in the treebank to bag of word features. The number of latent states we use is 1,000.","Once the two feature functions (syntactic in $G_{\\mathrm {syn}}$ and semantic in $G_{\\mathrm {par}}$ ) are created, each nonterminal in the training treebank is assigned two latent states (cluster identifiers). Figure 2 shows an example annotation of trees for three paraphrase questions from the Paralex corpus. We compute the parameters of the bi-layered L-PCFG $G_{\\mathrm {layered}}$ with a simple frequency count maximum likelihood estimate over this annotated treebank. As such, $G_{\\mathrm {layered}}$ is a combination of $G_{\\mathrm {syn}}$ and $G_{\\mathrm {par}}$ , resulting in 24,000 latent states (24 syntactic x 1000 semantic).","Consider an example where we want to generate paraphrases for the question what day is nochebuena. Parsing it with $G_{\\mathrm {layered}}$ will lead to the leftmost hybrid structure as shown in Figure 2 . The assignment of the first latent states for each nonterminals ensures that we retrieve the correct syntactic representation of the sentence. Here, however, we are more interested in the second latent states assigned to each nonterminals which capture the paraphrase information of the sentence at various levels. For example, we have a unary lexical rule (NN-*-142 day) indicating that we observe day with NN of the paraphrase type 142. We could use this information to extract unary rules of the form (NN-*-142 $w$ ) in the treebank that will generate words $w$ which are paraphrases to day. Similarly, any node WHNP-*-291 in the treebank will generate paraphrases for what day, SBARQ-*-403, for what day is nochebuena. This way we will be able to generate paraphrases when is nochebuena and when is nochebuena celebrated as they both have SBARQ-*-403 as their roots.","To generate a word lattice $W_q$ for a given question $q$ , we parse $q$ with the bi-layered grammar $G_{\\mathrm {layered}}$ . For each rule of the form $X$ - $m_1$ - $m_2 \\rightarrow w$ in the bi-layered tree with $X \\in {\\cal P}$ , $m_1 \\in \\lbrace 1, \\ldots , 24 \\rbrace $ , $m_2 \\in \\lbrace 1, \\ldots , 1000 \\rbrace $ and $q$0 a word in $q$1 , we extract rules of the form $q$2 - $q$3 - $q$4 from $q$5 such that $q$6 . For each such $q$7 , we add a path $q$8 parallel to $q$9 in the word lattice."],"section_name":"Bi-Layered L-PCFGs"},{"paragraphs":["Our sampling algorithm overgenerates paraphrases which are incorrect. To improve its precision, we build a binary classifier to filter the generated paraphrases. We randomly select 100 distinct questions from the Paralex corpus and generate paraphrases using our generation algorithm with various lattice settings. We randomly select 1,000 pairs of input-sampled sentences and manually annotate them as “correct” or “incorrect” paraphrases. We train our classifier on this manually created training data. We follow madnani2012, who used MT metrics for paraphrase identification, and experiment with 8 MT metrics as features for our binary classifier. In addition, we experiment with a binary feature which checks if the sampled paraphrase preserves named entities from the input sentence. We use WEKA BIBREF32 to replicate the classifier of madnani2012 with our new feature. We tune the feature set for our classifier on the development data."],"section_name":"Paraphrase Classification"},{"paragraphs":["In this section we describe how the paraphrase algorithm is used for converting natural language to Freebase queries. Following reddylargescale2014, we formalize the semantic parsing problem as a graph matching problem, i.e., finding the Freebase subgraph (grounded graph) that is isomorphic to the input question semantic structure (ungrounded graph).","This formulation has a major limitation that can be alleviated by using our paraphrase generation algorithm. Consider the question What language do people in Czech Republic speak?. The ungrounded graph corresponding to this question is shown in Figure 3 . The Freebase grounded graph which results in correct answer is shown in Figure 3 . Note that these two graphs are non-isomorphic making it impossible to derive the correct grounding from the ungrounded graph. In fact, at least 15% of the examples in our development set fail to satisfy isomorphic assumption. In order to address this problem, we use paraphrases of the input question to generate additional ungrounded graphs, with the aim that one of those paraphrases will have a structure isomorphic to the correct grounding. Figure 3 and Figure 3 are two such paraphrases which can be converted to Figure 3 as described in sec:groundedGraphs.","For a given input question, first we build ungrounded graphs from its paraphrases. We convert these graphs to Freebase graphs. To learn this mapping, we rely on manually assembled question-answer pairs. For each training question, we first find the set of oracle grounded graphs—Freebase subgraphs which when executed yield the correct answer—derivable from the question's ungrounded graphs. These oracle graphs are then used to train a structured perceptron model. These steps are discussed in detail below."],"section_name":"Semantic Parsing using Paraphrasing"},{"paragraphs":["We use GraphParser BIBREF7 to convert paraphrases to ungrounded graphs. This conversion involves three steps: 1) parsing the paraphrase using a CCG parser to extract syntactic derivations BIBREF33 , 2) extracting logical forms from the CCG derivations BIBREF34 , and 3) converting the logical forms to an ungrounded graph. The ungrounded graph for the example question and its paraphrases are shown in Figure 3 , Figure 3 and Figure 3 , respectively."],"section_name":"Ungrounded Graphs from Paraphrases"},{"paragraphs":["The ungrounded graphs are grounded to Freebase subgraphs by mapping entity nodes, entity-entity edges and entity type nodes in the ungrounded graph to Freebase entities, relations and types, respectively. For example, the graph in Figure 3 can be converted to a Freebase graph in Figure 3 by replacing the entity node Czech Republic with the Freebase entity CzechRepublic, the edge (speak.arg $_2$ , speak.in) between $x$ and Czech Republic with the Freebase relation (location.country.official_language.2, location.country.official_language.1), the type node language with the Freebase type language.human_language, and the target node remains intact. The rest of the nodes, edges and types are grounded to null. In a similar fashion, Figure 3 can be grounded to Figure 3 , but not Figure 3 to Figure 3 . If no paraphrase is isomorphic to the target grounded grounded graph, our grounding fails."],"section_name":"Grounded Graphs from Ungrounded Graphs"},{"paragraphs":["We use a linear model to map ungrounded graphs to grounded ones. The parameters of the model are learned from question-answer pairs. For example, the question What language do people in Czech Republic speak? paired with its answer $\\lbrace \\textsc {CzechLanguage}\\rbrace $ . In line with most work on question answering against Freebase, we do not rely on annotated logical forms associated with the question for training and treat the mapping of a question to its grounded graph as latent.","Let $q$ be a question, let $p$ be a paraphrase, let $u$ be an ungrounded graph for $p$ , and let $g$ be a grounded graph formed by grounding the nodes and edges of $u$ to the knowledge base $\\mathcal {K}$ (throughout we use Freebase as the knowledge base). Following reddylargescale2014, we use beam search to find the highest scoring tuple of paraphrase, ungrounded and grounded graphs $(\\hat{p}, \\hat{u}, \\hat{g})$ under the model $\\theta \\in \\mathbb {R}^n$ : $\n({\\hat{p},\\hat{u},\\hat{g}}) = \\operatornamewithlimits{arg\\,max}_{(p,u,g)} \\theta \\cdot \\Phi (p,u,g,q,\\mathcal {K})\\,,\n$ ","where $\\Phi (p, u, g, q, \\mathcal {K}) \\in \\mathbb {R}^n$ denotes the features for the tuple of paraphrase, ungrounded and grounded graphs. The feature function has access to the paraphrase, ungrounded and grounded graphs, the original question, as well as to the content of the knowledge base and the denotation $|g|_\\mathcal {K}$ (the denotation of a grounded graph is defined as the set of entities or attributes reachable at its target node). See sec:details for the features employed. The model parameters are estimated with the averaged structured perceptron BIBREF35 . Given a training question-answer pair $(q,\\mathcal {A})$ , the update is: $\n\\theta ^{t+1} \\leftarrow \\theta ^{t} + \\Phi (p^+, u^+, g^+, q,\n\\mathcal {K}) - \\Phi (\\hat{p}, \\hat{u}, \\hat{g}, q, \\mathcal {K})\\,,\n$ ","where $({p^+,u^+,g^+})$ denotes the tuple of gold paraphrase, gold ungrounded and grounded graphs for $q$ . Since we do not have direct access to the gold paraphrase and graphs, we instead rely on the set of oracle tuples, $\\mathcal {O}_{\\mathcal {K}, \\mathcal {A}}(q)$ , as a proxy: $\n(p^{+},u^{+},{g^{+}}) = \\operatornamewithlimits{arg\\,max}_{(p,u,g) \\in \\mathcal {O}_{\\mathcal {K},\\mathcal {A}}(q)} \\theta \\cdot \\Phi ({p,u,g,q,\\mathcal {K}})\\,,\n$ ","where $\\mathcal {O}_{\\mathcal {K}, \\mathcal {A}}(q)$ is defined as the set of tuples ( $p$ , $u$ , $g$ ) derivable from the question $q$ , whose denotation $|g|_\\mathcal {K}$ has minimal $F_1$ -loss against the gold answer $\\mathcal {A}$ . We find the oracle graphs for each question a priori by performing beam-search with a very large beam."],"section_name":"Learning"},{"paragraphs":["Below, we give details on the evaluation dataset and baselines used for comparison. We also describe the model features and provide implementation details."],"section_name":"Experimental Setup"},{"paragraphs":["We evaluate our approach on the WebQuestions dataset BIBREF5 . WebQuestions consists of 5,810 question-answer pairs where questions represents real Google search queries. We use the standard train/test splits, with 3,778 train and 2,032 test questions. For our development experiments we tune the models on held-out data consisting of 30% training questions, while for final testing we use the complete training data. We use average precision (avg P.), average recall (avg R.) and average F $_1$ (avg F $_1$ ) proposed by berantsemantic2013 as evaluation metrics."],"section_name":"Evaluation Data and Metric"},{"paragraphs":["We use GraphParser without paraphrases as our baseline. This gives an idea about the impact of using paraphrases.","We compare our paraphrasing models with monolingual machine translation based model for paraphrase generation BIBREF24 , BIBREF36 . In particular, we use Moses BIBREF37 to train a monolingual phrase-based MT system on the Paralex corpus. Finally, we use Moses decoder to generate 10-best distinct paraphrases for the test questions."],"section_name":"Baselines"},{"paragraphs":["For WebQuestions, we use 8 handcrafted part-of-speech patterns (e.g., the pattern (DT)?(JJ.? $\\mid $ NN.?){0,2}NN.? matches the noun phrase the big lebowski) to identify candidate named entity mention spans. We use the Stanford CoreNLP caseless tagger for part-of-speech tagging BIBREF38 . For each candidate mention span, we retrieve the top 10 entities according to the Freebase API. We then create a lattice in which the nodes correspond to mention-entity pairs, scored by their Freebase API scores, and the edges encode the fact that no joint assignment of entities to mentions can contain overlapping spans. We take the top 10 paths through the lattice as possible entity disambiguations. For each possibility, we generate $n$ -best paraphrases that contains the entity mention spans. In the end, this process creates a total of $10n$ paraphrases. We generate ungrounded graphs for these paraphrases and treat the final entity disambiguation and paraphrase selection as part of the semantic parsing problem.","We use the features from reddylargescale2014. These include edge alignments and stem overlaps between ungrounded and grounded graphs, and contextual features such as word and grounded relation pairs. In addition to these features, we add two new real-valued features – the paraphrase classifier's score and the entity disambiguation lattice score.","We use beam search to infer the highest scoring graph pair for a question. The search operates over entity-entity edges and entity type nodes of each ungrounded graph. For an entity-entity edge, there are two operations: ground the edge to a Freebase relation, or skip the edge. Similarly, for an entity type node, there are two operations: ground the node to a Freebase type, or skip the node. We use a beam size of 100 in all our experiments."],"section_name":"Implementation Details"},{"paragraphs":["In this section, we present results from five different systems for our question-answering experiments: original, mt, naive, ppdb and bilayered. First two are baseline systems. Other three systems use paraphrases generated from an L-PCFG grammar. naive uses a word lattice with a single start-to-end path representing the input question itself, ppdb uses a word lattice constructed using the PPDB rules, and bilayered uses bi-layered L-PCFG to build word lattices. Note that naive does not require any parallel resource to train, ppdb requires an external paraphrase database, and bilayered, like mt, needs a parallel corpus with paraphrase pairs. We tune our classifier features and GraphParser features on the development data. We use the best setting from tuning for evaluation on the test data."],"section_name":"Results and Discussion"},{"paragraphs":["We described a grammar method to generate paraphrases for questions, and applied it to a question answering system based on semantic parsing. We showed that using paraphrases for a question answering system is a useful way to improve its performance. Our method is rather generic and can be applied to any question answering system."],"section_name":"Conclusion"},{"paragraphs":["The authors would like to thank Nitin Madnani for his help with the implementation of the paraphrase classifier. We would like to thank our anonymous reviewers for their insightful comments. This research was supported by an EPSRC grant (EP/L02411X/1), the H2020 project SUMMA (under grant agreement 688139), and a Google PhD Fellowship for the second author."],"section_name":"Acknowledgements"}],"paper_id":"1601.06068","title":"Paraphrase Generation from Latent-Variable PCFGs for Semantic Parsing"},{"abstract":"Urban legends are a genre of modern folklore, consisting of stories about rare and exceptional events, just plausible enough to be believed, which tend to propagate inexorably across communities. In our view, while urban legends represent a form of\"sticky\"deceptive text, they are marked by a tension between the credible and incredible. They should be credible like a news article and incredible like a fairy tale to go viral. In particular we will focus on the idea that urban legends should mimic the details of news (who, where, when) to be credible, while they should be emotional and readable like a fairy tale to be catchy and memorable. Using NLP tools we will provide a quantitative analysis of these prototypical characteristics. We also lay out some machine learning experiments showing that it is possible to recognize an urban legend using just these simple features.","figures_and_tables":[{"caption":"Table 1: Examples of Urban Legend Claims","file":"3-Table1-1.png"},{"caption":"Table 2: Coverage of Named Entities","file":"12-Table2-1.png"},{"caption":"Table 3: Coverage of Temporal Expressions","file":"12-Table3-1.png"},{"caption":"Table 4: Average Sentiment and Emotions","file":"13-Table4-1.png"},{"caption":"Table 5: Readability difficulty Scores","file":"13-Table5-1.png"},{"caption":"Table 6: Classification Results","file":"15-Table6-1.png"},{"caption":"Table 7: Classification Results on Single Classes","file":"16-Table7-1.png"},{"caption":"Table 9: Overall Feature performances","file":"18-Table9-1.png"},{"caption":"Table 10: Classification Results, WWN vs. NYT","file":"19-Table10-1.png"}],"full_text":[{"paragraphs":["Urban legends are a genre of modern folklore consisting of stories told as true – and plausible enough to be believed – about some rare and exceptional events that supposedly happened to a real person or in a real place.","Whether urban legends are produced by individual authors or emerge spontaneously, they typically spread “virally\" across communities and tend to change over time with repetition and embellishment, like memes BIBREF0 . For example the sewer alligator, that originally “appeared” in New York City BIBREF1 , also appeared in different cities to suit regional variations. Though it is considered synonymous of “false belief,\" the term urban legend refers to a subtler and more complex phenomenon. The crucial factor is that the story is told as true in the absence of verification. Folklorists are generally more interested in the social context and meaning of urban legends than their truth value. From an NLP point of view, instead, it is interesting to computationally explore those linguistic characteristics that make them appealing and bring people to circulate them. With the advent of the Internet, urban legends gained new lifeblood, as they began to be circulated by e-mail.","In BIBREF2 , the authors discuss the idea of “stickiness\" popularized by the book “The Tipping Point” BIBREF3 , seeking to explain what makes an idea or concept memorable or interesting. They also focus on urban legends and claim that, by following the acronym “SUCCES\" (each letter referring to a characteristic that makes an idea “sticky\"), it is possible to describe their prototypical structure:","Such features are allegedly placed at the core of persuasive and viral language; urban legends constitute an ideal framework with which to computationally verify these assertions. Table 1 displays a few examples of urban legends claims.","In particular we will investigate some of the prototypical characteristics that can be found in urban legends as compared to similar literary genres. In our view, urban legends are viral since they are stressed by a tension between credible and incredible: credible like a news and incredible like a fairy tale. We will focus on the idea that urban legends should mimic the details of news (who, where, when) to be credible, and they should be emotional and readable like the story of a fairy tale to be catchy and memorable. We will verify these psychological hypotheses – appeared in the literature – using NLP tools, to drive a quantitative analysis of these qualitative theories. For example, the idea that urban legends derive much of their credibility from details concerning the location where the situation took place, is presented in BIBREF4 . Anecdotically, the television series “1000 Ways to Die\" – that recreates unusual supposed deaths and debunked urban legends in a way similar to the Darwin Awards – introducing each story with the location and date of each supposed incident, to render it more credible.","In the tension between credible and incredible, details should be neither too specific, like in the news, nor too few, as in fairy tales: effective urban legends should be credible but not verifiable. Similarly, emotions should be enough to make it sticky/catchy but not too much to render it not-credible. Finally urban legends should be easy to read, similar to fairy tales, to render them more memorable. As an example consider the following excerpt, taken from the “Kidney Theft” urban legend, as reported by snopes.com:","There is no very strong emotional wording in this example, it is the situation itself that is scary; on the contrary the email contains locations, the signature of a presumed Jerry Mayfield, and – noticeably – credibility is also explicitly addressed in the text with the adjectives “real”, “documented” and “confirmable”.","In the following sections we first review relevant work that addresses the problem of deceptive language and behavior both in online and offline scenarios, followed by an overview of work that addresses the virality of online content. Then we describe the data collected for our experiments and the features extracted to model the aforementioned prototypical characteristics of urban legends. We use these features in both descriptive statistics and generalization tasks and we report the best performing features. Finally we discuss future research on further prototypical characteristics of urban legends."],"section_name":"Introduction"},{"paragraphs":["The topic of deceptive and/or false messages is a burning topic within the NLP community. A seminal work on the linguistic recognition of lies can be found in BIBREF5 . Still, defense from subtle persuasive language in broadcast messages, including social networks, is needed in many applied scenarios. Viral messages have become a very important factor for persuasion and are currently almost entirely out of control. So, protection from fraudulent communication is needed, especially in competitive commercial situations. Two main approaches are currently under investigation in the literature:","1) Recognizing the linguistic characteristics of deceptive content in the social web: for example preventing deceptive consumer reviews BIBREF6 on sites like Trip Advisor is fundamental both for consumers seeking genuine reviews, and for the reputation of the site itself. Deceptive consumer reviews are fictitious opinions that have been deliberately written to sound authentic. Another example concerns online advertising BIBREF7 : detecting fraudulent ads is in the interest of users, of service providers (e.g. Google AdWords system), and other advertisers. An interesting phenomenon at the crossroad of viral phenomena and deceptive customer reviews, where ironic reviews (such as the case of the mountain three wolf moon) create phenomena of social contagion, is discussed in BIBREF8 .","2) Recognizing on-line behavioral patterns of deceptive users: For example recognizing groups of propagandists or fake accounts that are used to push the virality of content BIBREF9 . Four main patterns are recognized: (i) sending high volumes of tweets over short periods of time, (ii) retweeting while publishing little original content, (iii) quickly retweeting, and (iv) colluding with other, seemingly unrelated, users to send duplicate or near-duplicate messages on the same topic simultaneously. Another example is BIBREF10 where the authors hypothesize that there is a set of representative distributions of review rating scores. Deceptive business entities that hire people to write fake reviews can then be recognized since they will necessarily distort distribution of review scores, leaving “distributional footprints\" behind.","We want to consider a third point, which is linked to the previous two but different at the same time: deceptive content that spreads quickly but without an explicit strategy of making them spread, which is the case with urban legends. Finally, the spreading dynamics of an urban legend on one hand closely resembles those of memes that undergo many variations while spreading BIBREF11 ; on the other hand their characteristics resemble those of viral content. Several researchers have studied information flow, community building and similar processes using Social Networking sites as a reference BIBREF12 , BIBREF13 , BIBREF14 . However, the great majority concentrate on network-related features without taking into account the actual content spreading within the network BIBREF15 . A hybrid approach focusing on both product characteristics and network related features is presented in BIBREF16 : in particular, the authors study the effect of passive-broadcast and active-personalized notifications embedded in an application to foster word of mouth.","Recently, the correlation between content characteristics and virality has begun to be investigated, especially with regard to textual content; in BIBREF17 , for example, features derived from sentiment analysis of comments are used to predict stories' popularity. The work in BIBREF18 uses New York Times articles to examine the relationship between emotions evoked by the content and virality, using semi-automated sentiment analysis to quantify the affectivity and emotionality of each article. Results suggest a strong relationship between affect and virality, where virality corresponds to the number of times the article was email forwarded.","The relevant work in BIBREF19 measures a different form of content spreading by analyzing which features of a movie quote make it “memorable\" online. Another approach to content virality, somehow complementary to the previous one, is presented in BIBREF11 , and takes the perspective of understanding which modification dynamics make a meme spread from one person to another (while movie quotes spread remaining exactly the same). More recently, some works tried to investigate how different textual contents give rise to different reactions in the audience: the work presented in BIBREF20 correlates several viral phenomena with the wording of a post, while BIBREF21 shows that specific content features variations (like the readability level of an abstract) differentiate among virality level of downloads, bookmarking, and citations."],"section_name":"Related Work"},{"paragraphs":["To explore the characteristics of urban legends and understand the effectiveness of our ideas we collected a specific dataset. It is composed of roughly 8000 textual examples: 2518 Urban Legends (UL), 1860 Fairy Tales (FT) and 3575 Google News articles (GN). The description of how the datasets have been created follows."],"section_name":"Datasets"},{"paragraphs":["After collecting the datasets we extracted four different groups of features, relevant to the prototypical characteristics we want to analyze.","Named Entities, (NE). To annotate named entities we used the TextPro toolkit BIBREF24 , and in particular its Named Entities recognition module. The output of the tool is in the IOB2 format and includes the tags Person (PER), Organization (ORG), Location (LOC) and Miscellaneous (MISC).","Temporal Expressions, (TIMEX). To annotate temporal expressions we used the toolkit TTK BIBREF25 . The output of the tool is in TimeML annotation language format BIBREF26 . In particular time expressions are flagged with TIMEX3 tags (tern.mitre.org). The tags considered are DATE, DURATION and TIME.","To compute the importance of the aforementioned features, and to explore the characteristics of urban legend texts, we used the method proposed in BIBREF5 . We calculate a score associated with a given set of entities (features), as a measure of saliency for the given word class inside the text, called coverage.","More formally, given a set of feature instances present in a text, C $=$ { $W_1$ , $W_2$ , ..., $W_N$ }, we define the feature coverage in that text (or corpus) $A$ as the percentage of words from $A$ belonging to the feature set $C$ : ","$$Coverage_{A}(C) = \\frac{\\sum _{W_{i} \\in C} Frequency_A(W_i)}{Words_{A}}$$ (Eq. 14) ","where $Frequency_A(W_i)$ represents the total number of feature occurrences $W_i$ inside the text A, and $Words_{A}$ represents the total size (in words) of the text. Note that we computed features' coverage regardless of their actual length: “New York City\" or “Paris” both count as one LOC even if the former is composed of three tokens while the latter only of one. Note also that this approach normalizes according to text length, avoiding biases due to different corpus characteristics.","Sentiment (SENT). Since the three corpora have different characteristics, rather than computing word polarity using specialized bag-of-words approaches, we resort to words' prior polarity - i.e. if a word out of context evokes something positive or something negative. This technique, even if less precise, guarantee that the same score is given to the same word in different contexts, and that none of the corpora is either overestimated or underestimated. To this end, we follow the methodology proposed in BIBREF27 , using SentiWordNet 3.0 BIBREF28 , that assigns prior polarities to words starting from their posterior polarities. In particular we choose the best performing approach. This formula uses a weighted mean, i.e. each sense weight is chosen according to a harmonic series. The rationale behind this choice is based on the assumption that more frequent senses should bear more “affective weight” than very rare senses when computing the prior polarity of a word. In particular, for each word we returned its positive (POS) and negative (NEG) prior polarity score: ","$$\\qquad POS =\\frac{\\sum _ {i=1}^{n}( \\frac{1}{i} \\times posScore_ {i})}{\\sum _ {i=1}^{n}( \\frac{1}{i})}$$ (Eq. 15) ","where $posScore_ {i}$ represents the modulus of the positive polarity of the ith sense of that word. The NEG score is computed following the same procedure.","Emotions (EMO). To sense emotions from text we used the methodology described in BIBREF29 . The idea underlying the method is the distinction between direct and indirect affective words. For direct affective words, we refer to the WordNet Affect BIBREF30 lexicon, an extension of the WordNet database which employs six basic emotion labels (anger, disgust, fear, joy, sadness, surprise) to annotate WordNet synsets. LSA is then used to learn, in an unsupervised setting, a vector space from the British National Corpus. In the LSA space, each emotion label can be represented in various way. In particular, we employ the `LSA Emotion Synset' setting, in which the synsets of direct emotion words are considered. The affective load is computed in terms of its lexical similarity with respect to one of the six emotion labels. The overall affective load of a text is then calculated as the average of its similarity with each emotion label.","Emotions and Sentiment features are grouped under the label Affect (AFF).","Readability (READ). We further analyzed the texts in the three datasets according to readability indices, to understand whether there is a difference in the language difficulty among them. Basically, the task of readability assessment consists of quantifying how difficult a text is for a reader. This kind of assessment has been widely used for several purposes, such as evaluating the reading level of children and impaired persons and improving Web content accessibility, see for example what reported in BIBREF31 . We use three indices to compute the difficulty of a text: the Gunning Fog BIBREF32 , Flesch BIBREF33 and Kincaid BIBREF34 indices. These metrics combine factors such as word and sentence length that are easy to compute and approximate the linguistic elements that have an impact on readability. In the following formulae, $Sent_A$ represents the number of sentences in text $A$ , $Cpx_A$ the number of complex words (those with three or more syllables), and $Syll_A$ the total number of syllables.","The Fog index is a rough measure of how many years of schooling it would take someone to understand the content; higher scores indicate material that is harder to read. Texts requiring near-universal understanding have an index less than 8. Academic papers usually have a score between 15 and 20. The score, for a given text $A$ , is calculated according to the formula: ","$$Fog_{A} = 0.4 \\Big ( \\frac{Words_A}{Sent_A} + 100 \\frac{Cpx_A}{Words_A} \\Big )$$ (Eq. 16) ","The Flesch Index rates texts on a 100-point scale. Higher scores indicate material that is easier to read while lower numbers mark passages that are more difficult to read. Scores can be interpreted as: 90-100 for content easily understood by an average 11-year-old student, while 0-30 for content best understood by university graduates. The score is calculated with the following formula: ","$$Flesch_{A} = 206.835 - 1.015 \\frac{Words_A}{Sent_A} -84.6 \\frac{Syll_A}{Words_A}$$ (Eq. 17) ","The Kincaid Index or “Flesch–Kincaid Grade Level Formula\" translates the 0-100 score of the Flesch Index to a U.S. grade level. It can be interpreted as the number of years of education required to understand this text, similar to the Gunning Fog index. The grade level is calculated with the following formula: ","$$Kincaid_{A} = 0.39 \\frac{Words_A}{Sent_A} + 11.8 \\frac{Syll_A}{Words_A} - 15.59$$ (Eq. 18) "],"section_name":"Feature Extraction"},{"paragraphs":["As can be seen from Tables 2 (Named Entities) and 3 (Temporal Expressions), urban legends place half-way between fairy tales and news, as we expected. While fairy tales represent out-of-time, out-of-place and always-true stories (“a long time ago in a faraway land\"), news represent circumstantial description of events. This is reflected by the overall use of named entities (respectively almost three and four times more in UL and GN) and of temporal expressions (respectively almost two and three times more). Interestingly the use of person names is the only case where FT reduce the lead of UL and GN, and can be explained by the fact that characters in FT are usually addressed with proper names (e.g. “Hansel and Gretel”).","In Table 4 , statistics for sentiment and emotion coverage are reported. As can be seen, in the SENT group of features the differences are less marked and, quite surprisingly, ULs have the lowest scores. As we would expect, FTs have the highest score. Sentiment does not meet our initial expectation and seems in contrast with previous works – see for example what reported in BIBREF35 on UL and evoked emotions; still the results on sentiment as compared to emotions can be explained by the distinction between affective impact and affective language. In fact, affective impact can either derive from the wording of the text itself (usage of strong affect words), or from the depicted situation (i.e. emotions are evoked by describing a vivid situation with a plain language). In our experiment we tested the `wording' using SENT features while the `evoked emotions' with the EMO features. So, UL seem to use a plain and objective language, similar to GN, to gain credibility, but tend to evoke strong emotions (similar to FT) to be catchy. Let us consider the “Kidney Theft” excerpt described in Section 1, as stated, there is no very strong emotional wording in this UL, it is the depicted situation that is scary per se.","In Table 5 , statistics for readability are reported. As can be seen, ULs are readable in a way similar to fairy tales. Still, depending on the readability indices, that grasp different aspects of text difficulty, ULs are either slightly easier than FTs or half-way between FTs and ULs similar to the cases of Tables 2 and 3 .","This behavior can be explained by the fact that ULs have a simpler syntax than FTs but a more complex lexicon. In fact, inspecting the individual elements of the formulae, as reported in the second part of Table 5 , we see that while the percentage of complex words (either $\\frac{Cpx_A}{Words_A}$ or $\\frac{Syll_A}{Words_A}$ ) puts UL halfway between FT and GN, the average length of sentences ( $\\frac{Words_A}{Sent_A}$ ) is surprisingly higher for FT than GN and in turn UL. So, depending on the weight given either to complex words or to sentence length, the results in Table 5 can be interpreted.","All differences in the means reported in the tables are statistically significant (Student's t-test, $p<0.001$ ) apart from TIME, between UL and FT, and DURATION, between UL and GN, (signalled with * in Table 3 ).","Turning to the analysis of variance, we see that FT is – on average – a more cohesive genre, with lower standard deviations, while GN and UL have higher and closer standard deviations. In fact, all differences in the standard deviations reported in the tables are statistically significant (f-test, $p<0.001$ ) apart between UL and GN in Fog, Kincaid and in ALL sentiment (signalled with * in the respective Tables)."],"section_name":"Descriptive Statistics"},{"paragraphs":["The goal of our experiments is to understand to what extent it is possible to assign a text to one of the aforementioned classes using just the prototypical characteristics (features) discussed above, and whether there is a subset of features that stands out among the others in this classification task. For every feature combination we conducted a binary classification experiment with ten-fold cross validation on the dataset. We always randomly downsampled the majority class in order to make the dataset balanced, i.e. 50% of positive examples and 50% of negative examples; this accounts for a random baseline of 0.5. We also normalized all features according to z-score. Experiments were carried out using SVM BIBREF36 , in particular libSVM BIBREF37 under its default settings. Results are reported in Table 6 ; all significance tests discussed below are computed using an approximate randomization test BIBREF38 .","Urban Legends vs News. In the UL vs. GN classification task, while all the features together performed well (F1 = 0.833), improving over all other subgroups of features ( $p<0.001$ ), no single group of features performed so well, apart from READ (F1 = 0.763, $p<0.001$ ). Particularly, the temporal features (TIMEX) performed worse than AFF and NE ( $p<0.001$ ). Still, all features improved over the baseline ( $p<0.001$ ).","Urban Legends vs Fairy Tales. In the UL vs. FT classification task, all the features together performed better than the previous experiment (F1 = 0.897), again improving over all the other subgroups of features alone ( $p<0.001$ ). Interestingly, the best discriminative subgroup of features (still READ, F1 = 0.868) in this case reduces the lead with respect to all the features together (ALL) and improves over the others subgroups ( $p<0.001$ ) apart from the AFF group – from which has no significant difference – that in this case performs better than in the previous experiment. On the contrary, the TIMEX group had similar performances as the previous experiment, while NE improved its performance. Finally, all groups of features had a statistically significant improvement over the baseline ( $p<0.001$ ).","In Table 7 we report the performances of the various classification tasks in term of precision, recall and F1 over the single classes. Interestingly, for almost all feature combinations the classifiers had slightly higher precision than recall for UL, while the contrary holds for FT and GN.","News vs Fairy Tales. Finally, we wanted to check whether UL being “half-way\" between GN and FT can be observed in our classification experiments as well. If this hypothesis is correct, by classifying GN vs. FT we would expect to find higher performance than previous experiments. Results show that this is in fact the case. All features together performed better than all previous experiment and incredibly well (F1= 0.978), again improving over all the other subgroups of features alone ( $p<0.001$ ) apart from READ that performs equally well (F1=0.973, no statistically significant difference). Notably, all other groups of features improves over the UL vs. GN and the UL vs. FT tasks. Finally, all groups of features had a statistically significant improvement over the random baseline ( $p<0.001$ ).","Three Class Classification. Finally we also tested feature predictivity on a three class classification task (UL vs GN vs FT). Since in this case we did not performed downsampling, we use the ZeroR classifier as a baseline. For the sake of interpretability of results, along with precision, recall and F1 we also provide the Matthews Correlation Coefficient (MCC) which is useful for unbalanced datasets, as presented in BIBREF39 for the multiclass case. MCC returns a value between -1 and +1, where +1 represents a perfect prediction, 0 no better than random and -1 indicates total disagreement. Results are consistent with previous experiments. In Table 8 , all feature configurations show an improvement over the baseline ( $p<0.001$ ) but the temporal features (TIMEX) have far lower discriminative power as compared to others groups of features (MCC=0.339)."],"section_name":"Classification Experiments"},{"paragraphs":["While between UL and GN the discrimination is given by a skillful mixture of all the prototypical features together, where none has a clear predominance over the others, between UL and FT, readability (READ) and affect (AFF) play a major role. From the summary in Table 8 we see that while ALL features together have the highest averaged F1, READ is the best performing subset of features in all experiments, followed by AFF, NER and TIMEX that perform reasonably well. n general, these experiments proved the goodness of our features in discriminating UL against FT and GN in a machine learning framework, confirming the results emerged from the quantitative analysis part. In particular, as expected, these features gave the best results in the GN vs FT experiments, showing that these two genres represent the extremes of a continuum where ULs are placed."],"section_name":"Discussion"},{"paragraphs":["As a final validation of our feature importance we also set up experiments where we controlled for the medium where the message is delivered, specifically the online news domain. Since Newspapers exist for all kinds of stories and with all sorts of reputations for reliability we focused on two specific websites. One is the Weekly World News (WWN), a news website with very low reliability where many of the stories have the qualities of urban legends (the WWN was famous for stories about Bigfoot and UFOs, etc.). The other website is The New York Times (NYT), known for its high reliability and fact-checking procedures.","We scraped the WWN for a total of 225 stories, and then randomly selected an equal amount of stories from the NYT. For both datasets we extracted the same set of features discussed in the previous sections. For every feature combination we conducted a binary classification experiment with ten-fold cross validation on the dataset. Since the dataset is balanced, this accounts for a random baseline of 0.5. We also normalized all features according to z-score. Results are reported in Table 9 .","Also in this case our features are able to discriminate between reliable and non-reliable stories (namely those coming from NYT and WWN). In particular, all the features together performed very well (F1 = 0.864), improving over all other subgroups of features (p $<$ 0.001), and NE, TIMEX, READ performed equally well improving over AFF that was the least effective (p $<$ 0.001). Still, AFF improves over the random baseline (p $<$ 0.001).","With this last experiment we were able to show that stories from different newspapers of differing reliability might be classified correctly using the features learned for discriminating regular news from urban legends. So, also in more applicative and ecological scenarios, where stories come from the same medium (online news) these features are useful in discriminating stories on the basis of their UL-ness or GN-ness."],"section_name":"Ecological Experiments in the News Domain"},{"paragraphs":["In this paper we have presented a study on urban legends, a genre of modern folklore consisting of stories about some rare and exceptional events plausible enough to be believed. We argued that urban legends represent a form of “sticky” deceptive text, marked by a tension between the credible and incredible. To be credible they should resemble a news article while being incredible like a fairy tale. In particular we focused on the idea that ULs should mimic the details of news (who, where, when) to be credible, while being emotional and readable like a fairy tale to be catchy and memorable. Using NLP tools we presented a quantitative analysis of these simple yet effective features and provided some machine learning experiments showing that it is possible to recognize an urban legend using just these prototypical characteristics. In the future we want to explore other prototypical aspects of urban legends like, for example, linguistic style BIBREF40 , BIBREF41 . With regard to sentiment, besides the simple word polarities we used, we will explore the emotions expressed in UL, FT and GN, using an approach similar to the one described in BIBREF29 . Exploiting knowledge-based and corpus-based methods, that approach deals with automatic recognition of affect, annotating texts with six basic emotions. We believe that fine-grained emotion annotation of urban legends could shed more light in the understanding the mechanisms behind persuasive language."],"section_name":"Conclusions"}],"paper_id":"1601.06081","title":"Why Do Urban Legends Go Viral?"},{"abstract":"We introduce a model for the linguistic hedges `very' and `quite' within the label semantics framework, and combined with the prototype and conceptual spaces theories of concepts. The proposed model emerges naturally from the representational framework we use and as such, has a clear semantic grounding. We give generalisations of these hedge models and show that they can be composed with themselves and with other functions, going on to examine their behaviour in the limit of composition.","figures_and_tables":[{"caption":"Figure 1: Example of a Bayesian network of thresholds. Dependencies between thresholds εi are represented by arrows.","file":"4-Figure1-1.png"},{"caption":"Figure 2: Illustration of the expansion and contraction modifiers proposed in [2]. The original set F can be described in trapezoidal set notation as (2, 4, 6, 8). The approximate equality function is parametrised by a set Z = (−1,−0.5, 0.5, 1). Dilating the set as described in [2] gives the set EZ(F ) = (1, 3.5, 6.5, 9). Concentrating the set as described results in the set EZ(F ) = (3, 4.5, 5.5, 7)","file":"6-Figure2-1.png"},{"caption":"Figure 4: Plots of µL(x), µL(x) 1/2, and µqL(x) = µL(x) − µL(x) lnµL(x). Values of µqL(x) near to the prototype P = 5 are greater than equivalent values of µL(x) 1/2.","file":"8-Figure4-1.png"},{"caption":"Figure 5: Plots of µL(x), µL(x) 2, and µvL(x) = µL(x)+(1−µL(x)) ln(1−µL(x)). Values of µvL(x) decrease more quickly as we move from the prototype P = 5 than do equivalent values of µL(x) 2.","file":"9-Figure5-1.png"},{"caption":"Figure 6: Plot of µvery tall and quite thin(x1, x2) = µvery tall(x1)µquite thin(x2). Notice how appropriateness drops off quickly as the variable x1, i.e. height, decreases. In contrast, appropriateness decreases slowly as we increase the variable x2, or waist size","file":"10-Figure6-1.png"},{"caption":"Figure 7: Plots of µL(x) and µqL(x) = ∆(d(x, Pq)) − ∆(d(x, Pq)) ln(∆(d(x, Pq))). The prototype of L has been expanded.","file":"11-Figure7-1.png"},{"caption":"Figure 8: Plots of µL(x) and µvL(x) = ∆(d(x, Pv) + (1−∆(d(x, Pv)) ln(1−∆(d(x, Pv))). The prototype of L has been reduced.","file":"12-Figure8-1.png"},{"caption":"Figure 9: Plots of µL(x), µqL(x) and µvL(x). The prototype, a single point, remains constant on application of the hedges.","file":"13-Figure9-1.png"},{"caption":"Figure 10: Plots of µL(x), µqL(x) and µvL(x). In this case, the set {x ∈ Ω : µhL(x) = 1}, where h = v, q or nothing, expands or contracts on application of the hedges.","file":"14-Figure10-1.png"},{"caption":"Figure 12: Plots of µL(x), µL(x) 4, and µvvL(x). Values of µvvL(x) decrease more quickly as we move from the prototype P = 5 than do equivalent values of µL(x) 4.","file":"18-Figure12-1.png"}],"full_text":[{"paragraphs":["The modelling of natural language relies on the idea that languages are compositional, i.e. that the meaning of a sentence is a function of the meanings of the words in the sentence, as proposed by BIBREF0 . Whether or not this principle tells the whole story, it is certainly important as we undoubtedly manage to create and understand novel combinations of words. Fuzzy set theory has long been considered a useful framework for the modelling of natural language expressions, as it provides a functional calculus for concept combination BIBREF1 , BIBREF2 .","A simple example of compositionality is hedged concepts. Hedges are words such as `very', `quite', `more or less', `extremely'. They are usually modelled as transforming the membership function of a base concept to either narrow or broaden the extent of application of that concept. So, given a concept `short', the term `very short' applies to fewer objects than `short', and `quite short' to more. Modelling a hedge as a transformation of a concept allows us to determine membership of an object in the hedged concept as a function of its membership in the base concept, rather than building the hedged concept from scratch BIBREF3 .","Linguistic hedges have been widely applied, including in fuzzy classifiers BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 and database queries BIBREF8 , BIBREF9 . Using linguistic hedges in these applications allows increased accuracy in rules or queries whilst maintaining human interpretability of results BIBREF10 , BIBREF11 . This motivates the need for a semantically grounded account of linguistic hedges: if hedged results are more interpretable then the hedges used must themselves be meaningful.","In the following we provide an account of linguistic hedges that is both functional, and semantically grounded. In its most basic formulation, the operation requires no additional parameters, although we also show that the formulae can be generalised if necessary. Our account of linguistic hedges uses the label semantics framework to model concepts BIBREF12 . This is a random set approach which quantifies an agent's subjective uncertainty about the extent of application of a concept. We refer to this uncertainty as semantic uncertainty BIBREF13 to emphasise that it concerns the definition of concepts and categories, in contrast to stochastic uncertainty which concerns the state of the world. In BIBREF13 the label semantics approach is combined with conceptual spaces BIBREF14 and prototype theory BIBREF15 , to give a formalisation of concepts as based on a prototype and a threshold, located in a conceptual space. This approach is discussed in detail in section \"Conceptual Spaces\" . An outline of the paper is then as follows: section \"Approaches to linguistic hedges\" discusses different approaches to linguistic hedges from the literature, and compares these with our model. Subsequently, in section \"Label semantics approach to linguistic hedges\" , we give formulations of the hedges `very' and `quite'. These are formed by considering the dependence of the threshold of a hedged concept on the threshold of the original concept. We give a basic model and two generalisations, show that the models can be composed and investigate the behaviour in the limit of composition. Section \"Discussion\" compares our results to those in the literature and proposes further lines of research."],"section_name":"Introduction"},{"paragraphs":["Prototype theory views concepts as being defined in terms of prototypes, rather than by a set of necessary and sufficient conditions. Elements from an underlying metric space then have graded membership in a concept depending on their similarity to a prototype for the concept. There is some evidence that humans use natural categories in this way, as shown in experiments reported in BIBREF15 . Fuzzy set theory BIBREF1 was proposed as a calculus for combining and modifying concepts with graded membership, and extended these ideas in BIBREF2 to linguistic variables as variables taking words as values, rather than numbers. For example, `height' can be viewed as a linguistic variable taking values `short,' `tall', `very tall', etc. The variable relates to an underlying universe of discourse $\\Omega $ , which for the concept `tall' could be $\\mathbb {R}^+$ . Then each value $L$ of the variable is associated with a fuzzy subset of $\\Omega $ , and a function $\\mu _L:\\Omega \\rightarrow [0,1]$ associates with each $x \\in \\Omega $ the value of its membership in $L$ . Prototype theory gives a semantic basis to fuzzy sets through the notion of similarity to a prototype, as described in BIBREF16 . In this context, concepts are represented by fuzzy sets and membership of an element in a concept is quantified by its similarity to the prototype. In this situation the fuzziness of the concept is seen as inherent to the concept. An alternative interpretation for fuzzy sets is random set theory, see BIBREF16 for an exposition. Here, the fuzziness of a set comes from uncertainty about a crisp set, i.e. semantic uncertainty, rather than fuzziness inherent in the world. This second approach is the stance taken by BIBREF13 , and which we now adopt in this paper."],"section_name":"Prototype theory and fuzzy set theory"},{"paragraphs":["Conceptual spaces are proposed by Gärdenfors in BIBREF14 as a framework for representing information at the conceptual level. Gärdenfors contrasts his theory with both a symbolic, logical approach to concepts, and an associationist approach where concepts are represented as associations between different kinds of basic information elements. Rather, conceptual spaces are geometrical structures based on quality dimensions such as weight, height, hue, brightness, etc. It is assumed that conceptual spaces are metric spaces, with an associated distance measure. This might be Euclidean distance, or any other appropriate metric. The distance measure can be used to formulate a measure of similarity, as needed for prototype theory - similar objects are close together in the conceptual space, very different objects are far apart.","To develop the conceptual space framework, Gärdenfors also introduces the notion of integral and separable dimensions. Dimensions are integral if assignment of a value in one dimension implies assignment of a value in another, such as depth and breadth. Conversely, separable dimensions are those where there is no such implication, such as height and sweetness. A domain is then defined as a set of quality dimensions that are separable from all other dimensions, and a conceptual space is defined as a collection of one or more domains.","Gärdenfors goes on to define a property as a convex region of a domain in a conceptual space. A concept is defined as a set of such regions that are related via a set of salience weights. This casting of (at least) properties as convex regions of a domain sits very well with prototype theory, as Gärdenfors points out. If properties are convex regions of a space, then it is possible to say that an object is more or less central to that region. Because the region is convex, its centroid will lie within the region, and this centroid can be seen as the prototype of the property."],"section_name":"Conceptual Spaces"},{"paragraphs":["The label semantics framework was proposed by BIBREF12 and related to prototype theory and conceptual spaces in BIBREF13 . In this framework, agents use a set of labels $LA = \\lbrace L_1, L_2, ..., L_n\\rbrace $ to describe an underlying conceptual space $\\Omega $ which has a distance metric $d(x,y)$ between points. In fact, it is sufficient that $d(x,y)$ be a pseudo-distance. When $x$ or $y$ is a set, say $Y$ , we take $d(x,Y) = \\text{min}\\lbrace d(x,y): y \\in Y\\rbrace $ . In this case, the set $Y$ is seen as an ontic set, i.e., a set where all elements are jointly prototypes, as opposed to an epistemic set describing a precise but unknown prototype, as described in BIBREF17 . Each label $L_i$ is associated with firstly a set of prototype values $\\Omega $0 , and secondly a threshold $\\Omega $1 , about which the agents are uncertain. The thresholds $\\Omega $2 are drawn from probability distributions $\\Omega $3 . Labels $\\Omega $4 are associated with neighbourhoods $\\Omega $5 . The neighbourhood can be seen as the extension of the concept $\\Omega $6 . The intuition here is that $\\Omega $7 captures the idea of being sufficiently close to prototypes $\\Omega $8 . In other words, $\\Omega $9 is sufficiently close to $d(x,y)$0 to be appropriately labelled as $d(x,y)$1 providing that $d(x,y)$2 .","Given an element $x \\in \\Omega $ , we can ask how appropriate a given label is to describe it. This is quantified by an appropriateness measure, denoted $\\mu _{L_i}(x)$ . We are intentionally using the same notation as for the membership function of a fuzzy set. This quantity is the probability that the distance from $x$ to $P_i$ , the prototype of $L_i$ , is less than the threshold $\\varepsilon _i$ , as given by: $\n\\mu _{L_i}(x) = P(\\varepsilon _i : x \\in \\mathcal {N}^{\\varepsilon _i}_{L_i}) = P(\\varepsilon _i : d(x, P_i) \\le \\varepsilon _i) = \\int _{d(x, P_i)}^\\infty \\delta _{\\varepsilon _i}(\\varepsilon _i) \\mathrm {d}\\varepsilon _i\n$ ","We also use the notation $\\int _{d}^\\infty \\delta _{\\varepsilon _i} (\\varepsilon _i)\\mathrm {d}\\varepsilon _i = \\Delta _i(d)$ , according to which $\\mu _{L_i}(x) = \\Delta _i(d(x, P_i))$ . The above formulation provides a link to the random set interpretation of fuzzy sets. Random sets are random variables taking sets as values. If we view $\\mathcal {N}^{\\varepsilon _i}_{L_i}$ as a random set from $\\mathbb {R}^+$ into $2^\\Omega $ , then $\\mu _{L_i}(x)$ is the single point coverage function of $\\mathcal {N}^{\\varepsilon _i}_{L_i}$ , as defined in BIBREF18 , and also commonly called a contour function BIBREF19 .","Labels can often be semantically related to each other. For example, the label `pet fish' is semantically related to the labels `pet' and `fish', and the label `very tall' related to the label `tall'. This prompts two questions: firstly, how the prototypes of each concept are related to each other, and secondly, how the thresholds of each concept are related. Two simple models for the relationships between the thresholds are given in BIBREF13 . The consonant model takes all thresholds as being dependent on one common underlying threshold. So, all thresholds have the same distance metric $d$ and are related to a base threshold $\\varepsilon $ by the dependency that $\\varepsilon _i = f_i(\\varepsilon )$ for increasing functions $f_i$ . In contrast, the independence model takes all thresholds as being independent of each other. This might hold when labels are taken from different conceptual spaces.","Between these two extremes, we model dependencies between thresholds as a Bayesian network - i.e., a directed acyclic graph whose edges encode conditional dependence between variables. The key property of this type of network is that the joint distribution of all variables can be broken into factors that depend only on each individual variable and its parents. So, for example, the network in figure 1 can be factorised as $\\delta (\\varepsilon _1,\\varepsilon _2, \\varepsilon _3, \\varepsilon _4, \\varepsilon _5) = \\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _2}(\\varepsilon _2)\\delta _{\\varepsilon _3|\\varepsilon _1, \\varepsilon _2}(\\varepsilon _3|\\varepsilon _1,\\varepsilon _2)\\delta _{\\varepsilon _4|\\varepsilon _2}(\\varepsilon _4|\\varepsilon _2)\\delta _{\\varepsilon _5|\\varepsilon _3}(\\varepsilon _5|\\varepsilon _3)$ .","This enables calculation of the joint distribution and therefore marginal distributions in an efficient manner.","One intuitively easy example is where the dependency of one threshold $\\varepsilon _2$ on another $\\varepsilon _1$ is that $\\varepsilon _2 \\le \\varepsilon _1$ . This could be taken to model the dependency of the threshold of the concept `very tall' on the threshold of `tall'. The label `very tall' should be appropriate to describe fewer people than the label `tall'. Therefore, the threshold for describing someone as `very tall' will be narrower than the threshold for describing someone as `tall', i.e. $\\varepsilon _{\\text{very tall}} \\le \\varepsilon _{\\text{tall}}$ . This simple model will form part of the approach to modelling linguistic hedges, as outlined in the sequel."],"section_name":"Label Semantics"},{"paragraphs":["Linguistic hedges have been given varying treatments in the literature. In this section we summarise these different approaches and state the approach that we wish to take, discussing properties that hedge modifiers may need. We give two specific approaches from the literature with which we will compare our results.","In BIBREF3 the idea of linguistic hedges as operators modifying fuzzy sets was introduced, so that the membership function $\\mu _{hL}(x)$ of a hedged concept, $hL$ , is a function of the membership of the base concept $L$ , i.e. $\\mu _{hL}(x) = f(\\mu _L(x))$ . Furthermore, truth can be considered as a linguistic variable and hence a fuzzy set BIBREF2 , so that the application of a hedge can be seen as modifying the truth value of a sentence using that concept BIBREF20 , BIBREF21 , BIBREF2 . This second view is useful in approximate reasoning, and allows for an algebraic approach to investigating the properties of linguistic hedges, as introduced in BIBREF2 , and expanded upon in BIBREF22 , BIBREF20 , BIBREF21 . The approach we take, however, is to view a hedge as modifying the fuzzy set associated with a concept directly, as taken by BIBREF23 , BIBREF10 , BIBREF24 , BIBREF25 . Rather than examining the algebraic properties of hedges or their role in reasoning, we look at how hedges are semantically grounded and argue that our approach provides a particularly clear semantics.","We will propose a set of operations that may be used for both expansion and refinement of single concepts. This is in contrast to the work presented in BIBREF26 in which information coarsening is effected by taking disjunctions of labels. The idea of a hedged concept has some similarities to that of the bipolar model of concepts described in BIBREF27 , since if it is appropriate to describe someone as `very tall', it must be appropriate to describe them as `tall', and similarly describing someone as `quite tall' implies that it is not entirely inappropriate to describe them as `tall'. However, we see the concepts derived by application of hedges as labels in their own right which can be used to describe data or objects.","Zadeh divides hedges into two types. A type 1 hedge can be seen as an operator acting on a single fuzzy set. Examples are `very', `more or less', `quite', or `extremely' BIBREF3 . Type 2 hedges are more complicated and include modifiers such as `technically' or `practically'. In BIBREF3 concepts are considered as made up of various different components, with the membership function a weighted sum of the memberships of the individual components. Type 1 hedges operate on all components equally, whereas type 2 hedges differentiate between components. For example, the hedge `essentially' might give more weight to the most important components in a concept. Type 2 hedges are further explored in BIBREF28 , BIBREF29 , where components of a concept are categorised as definitional, primary or secondary, and the hedges `technically', `strictly speaking' and `loosely speaking' are analysed in terms of these categories. Although in the following we restrict ourselves to consideration of type 1 hedges only, the treatment of concepts as having different components is mirrored by the conceptual spaces view, where each component might be seen as a dimension in the conceptual space. Further development of the framework may therefore allow a treatment of type 2 hedges.","A further distinction between types of hedge lies in the difference between powering or shifting modifiers. Powering modifiers are of the form $\\mu _{hL}(x) = (\\mu _L(x))^k$ , where $hL$ refers to the hedged concept and $k$ is some real value, and shifting modifiers are of the form $\\mu _{hL}(x) = (\\mu _L(x-a))$ . Zadeh introduces both types of modifier in his discussion of type 1 hedges BIBREF3 , however his powering modifiers are most frequently cited. These are the concentration operator $CON(\\mu _{\\text{tall}}(x)) = (\\mu _{\\text{tall}}(x))^2$ , and the dilation operator $DIL(\\mu _{\\text{tall}}(x)) = (\\mu _{\\text{tall}}(x))^\\frac{1}{2}$ , which are often taken to implement the hedges `very' and `quite', (alternatively `more or less'), respectively. ","The operators $CON$ and $DIL$ leave the core, $\\lbrace x \\in \\Omega : \\mu _L(x) = 1\\rbrace $ , and support $\\lbrace x \\in \\Omega : \\mu _L(x) \\ne 0\\rbrace $ , of the fuzzy sets unchanged, which is often argued to be undesirable BIBREF9 , BIBREF23 , BIBREF25 , BIBREF30 . In particular, BIBREF9 argue that in a fuzzy database, if a concentrating hedge is being used to refine a query that is returning too many objects, the hedge needs to reduce the number of objects returned, and hence narrow down the core. Furthermore, BIBREF7 find that classifiers using the $CON$ and $DIL$ operators (classical hedges) do not perform as well as those with hedges that modify the core and support of the fuzzy sets. In contrast, Zadeh himself argues that the core should not be altered. The application of a modifier `very' to a property given by a crisp set should leave that property unchanged: `very square' is the same as `square'. A fuzzy set is made up of a non-fuzzy part, the core, and a fuzzy part, $\\lbrace x \\in \\Omega : 0 < \\mu _L(x) <1\\rbrace $ . Since the core of a fuzzy set is a crisp set, it should be left unchanged. The use of classical hedges does improve performance over non-hedged fuzzy rules in expert systems BIBREF4 , BIBREF5 , BIBREF6 , so the argument against classical hedges is a matter of degree.","The use of the $CON$ and $DIL$ operators to model the hedges `very' and `quite' is further criticised on the basis that the modifiers are arbitrary and semantically ungrounded. No justification is given for these modifiers other than that they have what seem to be intuitively the right properties BIBREF23 , BIBREF31 , BIBREF25 . Grounding hedges semantically is important for a theoretical account of what happens when we use terms like `very' and also for retaining interpretability in fuzzy systems. BIBREF23 , BIBREF31 both ground modifiers using a resemblance relation which takes into account how objects in the universe are similar to each other. BIBREF25 takes a horizon shifting approach.","In BIBREF25 the class of finite numbers is used as an example of the horizon shifting approach. Some numbers are certainly finite, however as numbers get larger, finiteness becomes impossible to verify. Mapping this idea onto the concept `small', we can say that there is a class of numbers that are definitely small, say $[0, c]$ . As numbers get larger than $c$ we approach the horizon past which the concept `small' no longer applies, expressed as $1 - \\epsilon (x)(x-c)$ . So: $\n\\mu _{\\text{small}}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{if } x \\in [0, c] \\\\\n1 - \\epsilon (x)(x-c) & \\text{if } x \\ge c\n\\end{array}\\right.}\n$ ","Now, to implement the hedge `very', the horizon $c$ is shifted by a factor $\\sigma $ and the membership function altered thus: $\n\\mu _{\\text{very small}}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{if } x \\in [0, \\sigma c] \\\\\n1 - \\epsilon (x, \\sigma )(x- \\sigma c) & \\text{if } x \\ge \\sigma c\n\\end{array}\\right.}\n$ ","In BIBREF25 , examples of different kinds of membership functions that might be used to implement this idea are given. A linear membership function gives $\\epsilon (x) = \\frac{1}{a-c}$ where $a$ is the upper limit of the membership function. To implement the hedge, the function $\\epsilon (x, \\sigma ) = \\frac{1}{\\sigma (a-c)}$ is introduced, giving $\n\\mu _{\\text{small}}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{if } x \\in [0, c] \\\\\n1 - \\frac{x-c}{a-c} & \\text{if } x \\in [c, a]\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","and $\n\\mu _{\\text{very small}}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{if } x \\in [0, \\sigma c] \\\\\n1 - \\frac{x- \\sigma c}{\\sigma (a-c)} & \\text{if } x \\in [\\sigma c, \\sigma a]\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ "," BIBREF23 , BIBREF31 both ground their approaches in the idea of looking at the elements near a fuzzy set in order to contract or dilate the set. The two approaches are similar, so we restrict ourselves to that of BIBREF23 . This approach introduces a fuzzy resemblance relation on the universe of discourse, and either a $T$ -norm in the case of dilation, or a fuzzy implicator for concentration. The modifier is then implemented as follows. Consider a fuzzy set $F$ and a proximity relation $E^Z$ which is approximate equality, parametrised by a fuzzy set $Z$ . As described in BIBREF23 , $E$ is modelled by $(u,v) \\rightarrow E(u,v) = Z(u - v)$ , where $Z$ is a fuzzy interval centred on 0 with finite support. In terms of a trapezoidal membership function, $Z$ can be expressed as $(-z - a, -z, z, z+a)$ . Therefore, if $|u -v| \\le z$ , $F$0 and $F$1 are judged to be approximately equal, i.e. $F$2 . The set $F$3 is dilated by $F$4 , where $F$5 is any $F$6 -norm, $F$7 being the standard.","To understand the effect that this has on a fuzzy set $F$ , suppose that $F$ has a trapezoidal membership function $(A, B, C, D)$ where $[B, C]$ is the core of $F$ and $[A,B]$ , $[C,D]$ the support, and that $Z$ similarly is $(-z - a, -z, z, z+a)$ , with the T-norm $min$ used. Then $F$0 .","Concentration is effected in a similar way: $E_Z(F)(s) = \\text{inf}_{r \\in \\Omega } I(F(r), E^Z(s,r))$ , where I is a fuzzy implication. If $F$ and $Z$ are as above with the condition that $C - B \\ge 2z$ , and $I$ is the Gödel implication, then $E_Z(F)$ = $(A + z + a, B + z, C - z, D - z - a)$ .","For example, suppose we start with a set $F$ described in trapezoidal notation as $F = (A, B, C, D) = (2,4,6,8)$ , and an approximate equality function parametrised by $Z =(-z-a, -z, z, z + a) = (-1, -0.5, 0.5, 1)$ . The dilation of the set $F$ using T-norm $min$ is then: $\nE^Z(F) = (A - z - a, B - z, C +z, D +z +a) = (1, 3.5, 6.5, 9)\n$ ","The concentration of the set $F$ using the Gödel implication is: $\nE_Z(F) = (A + z + a, B + z, C - z, D - z - a) = (3, 4.5, 5.5, 7)\n$ ","These effects are illustrated in figure 2 .","The intuitive idea behind this approach is that if an object $x_1$ resembles another object $x_2$ that is $L$ , then $x_1$ can be said to be `quite $L$ '. Conversely, object $x_2$ that is $L$ can be said to be `very $L$ ' only if all the objects $x$ that resemble it can be said to be $L$ . This formulation alters both the core and support of the fuzzy set $x_2$0 , which has been argued to be a desirable effect.","Following BIBREF23 , BIBREF31 , BIBREF25 , we will propose linguistic modifiers that are semantically grounded rather than attempting to show their utility in classifiers, reasoning or to examine the algebra of modifiers. Our approach to linguistic modifiers arises very naturally from the label semantics framework, and the primary result does not require any parameters additional to the original membership function of the concept. We also show similarities between our model and the two detailed above."],"section_name":"Approaches to linguistic hedges"},{"paragraphs":["We present three formulations of linguistic hedges with increasing levels of generality. The first assumes that prototypes are equal. Secondly, we show that an analogue holds where prototypes are not equal, and thirdly that these hold in the case where the second threshold is a function of the first. We go on to show similarities between our model and those of BIBREF23 , BIBREF31 , BIBREF25 . Furthermore, we show that hedges are compositional, and look at their behaviour in the limit of composition.","As described in section \"Approaches to linguistic hedges\" , $LA$ denotes a finite set of labels $\\lbrace L_i\\rbrace $ that agents use to describe basic categories. $\\Omega $ is the underlying domain of discourse, with prototypes $P_i \\in \\Omega $ and thresholds $\\varepsilon _i$ , drawn from a distribution $\\delta _{\\varepsilon _i}$ . As before, the appropriateness $\\mu _{L_i}(x) = \\Delta _i(d(x, P_i)) = \\int _{d(x, P_i)}^\\infty \\delta _{\\varepsilon _i}(\\varepsilon _i)\\mathrm {d}\\varepsilon _i$ . We use the notation $L_i = $ .","A concept $L_1$ can be narrowed or broadened to a second concept $L_2$ using the linguistic hedges `very' and `quite' respectively, i.e. $L_2$ is defined as `quite $L_1$ '. The directed acyclic graph illustrating this dependency is given in figure 3 . In this case, the threshold $\\varepsilon _2$ associated with $L_2$ is dependent on $\\varepsilon _1$ in that $\\varepsilon _2 \\ge \\varepsilon _1$ . In the case of `very', we have that $\\varepsilon _2 \\le \\varepsilon _1$ . Essentially, for `quite', we are saying that however wide a margin of certainty we apply the label `tall' with, the margin for `quite tall' will be wider, and conversely for `very'."],"section_name":"Label semantics approach to linguistic hedges"},{"paragraphs":["Definition 1 (Dilation and Concentration) A label $L_2 = $ is a dilation of a label $L_1 = $ when $\\varepsilon _2$ is dependent on $\\varepsilon _1$ such that $\\varepsilon _2 \\ge \\varepsilon _1$ . $L_2$ is a concentration of $L_1$ when $\\varepsilon _2$ is dependent on $\\varepsilon _1$ such that $\\varepsilon _2 \\le \\varepsilon _1$ .","Theorem 2 ( $L_2 = $ quite $L_1$ ) Suppose $L_2 = $ is a dilation of $L_1 = $ , so that $\\varepsilon _2 \\ge \\varepsilon _1$ . Suppose also that $P_1 = P_2 = P$ , and that the marginal (unconditional) distribution of $\\varepsilon _2$ , before conditioning on the knowledge that $\\varepsilon _2 \\ge \\varepsilon _1$ , is identical to $\\delta _{\\varepsilon _1}$ , since $L_2$ is a dilation of $L_1$0 . Then $L_1$1 , $L_1$2 . $L_1$3 ","and hence, $\n\\delta (\\varepsilon _1, \\varepsilon _2) &= \\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _2|\\varepsilon _1}(\\varepsilon _2|\\varepsilon _1)\n=\n{\\left\\lbrace \\begin{array}{ll}\n\\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _1(\\varepsilon _1)} & \\text{if } \\varepsilon _2 \\ge \\varepsilon _1\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","Then since $\\varepsilon _2 \\ge \\varepsilon _1$ we have that $\n\\mu _{L_2}(x) &= \\int _0^\\infty \\int _{max(\\varepsilon _1, d(x,P))}^\\infty \\delta (\\varepsilon _1, \\varepsilon _2) d\\varepsilon _2 d\\varepsilon _1 = \\int _0^\\infty \\int _{max(\\varepsilon _1, d(x,P))}^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _1(\\varepsilon _1)} d\\varepsilon _2 d\\varepsilon _1\\\\\n&= \\int _0^{d(x,P)}\\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)}\\int _{d(x,P)}^\\infty \\delta _{\\varepsilon _1}(\\varepsilon _2) d\\varepsilon _2 d\\varepsilon _1 + \\int _{d(x,P)}^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)}\\int _{\\varepsilon _1}^\\infty \\delta _{\\varepsilon _1}(\\varepsilon _2) d\\varepsilon _2 d\\varepsilon _1\\\\\n&= \\mu _{L_1}(x)\\int _0^{d(x,P)}\\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta (\\varepsilon _1)} d\\varepsilon _1 + \\int _{d(x,P)}^\\infty \\delta _{\\varepsilon _1}(\\varepsilon _1) d\\varepsilon _1 = \\mu _{L_1}(x) - \\mu _{L_1}(x)\\ln (\\mu _{L_1}(x))\\nonumber $ "," The following example gives an illustration of the effect of applying this hedge, in comparison with the standard dilation hedge $(\\mu _L(x))^{1/2}$ .","Example 3 Suppose our conceptual space $\\Omega = \\mathbb {R}$ with Euclidean distance and that a label $L$ has prototype $P = 5$ , and threshold $\\varepsilon \\sim $ Uniform $[0,3]$ . Then $\n\\mu _L(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 - \\frac{|x - 5|}{3} & \\text{ if } x \\in [2, 8]\\\\\n0 & otherwise\n\\end{array}\\right.}\n$ ","We can then form a new label $qL$ with prototype $P_q = P = 5$ and threshold $\\varepsilon _q \\ge \\varepsilon $ . Then, according to theorem \"Theorem 2 (L 2 =L_2 = quite L 1 L_1)\" , $\\mu _{qL}(x) = \\mu _L(x) - \\mu _L(x)\\ln \\mu _L(x)$ . The effect of applying a dilation hedge to $L$ can be seen in figure 4 . The dilation hedge given above is contrasted with Zadeh's dilation hedge $(\\mu _L(x))^{1/2}$ .","Theorem 4 ( $L_2 = $ very $L_1$ ) Suppose $L_2 = $ is a concentration of $L_1 = $ , so that $\\varepsilon _2 \\le \\varepsilon _1$ . Suppose also that $P_1 = P_2 = P$ , and that the marginal (unconditional) distribution of $\\varepsilon _2$ , before conditioning on the knowledge that $\\varepsilon _2 \\le \\varepsilon _1$ , is identical to $\\delta _{\\varepsilon _1}$ , since $L_2$ is a concentration of $L_1$0 . Then $L_1$1 , $L_1$2 . $L_1$3 ","and hence, $\n\\delta (\\varepsilon _1, \\varepsilon _2) &= \\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _2|\\varepsilon _1}(\\varepsilon _2|\\varepsilon _1)\n=\n{\\left\\lbrace \\begin{array}{ll}\n\\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _1}(\\varepsilon _2)}{1-\\Delta _1(\\varepsilon _1)} & \\text{if } \\varepsilon _2 \\le \\varepsilon _1\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","So since $\\varepsilon _2 \\le \\varepsilon _1$ we have that: $\n\\mu _{L_2}(x) &= \\int _0^\\infty \\int _{min(\\varepsilon _1, d(x,P))}^{\\varepsilon _1} \\delta (\\varepsilon _1, \\varepsilon _2) d\\varepsilon _2 d\\varepsilon _1 = \\int _0^\\infty \\int _{min(\\varepsilon _1, d(x,P))}^{\\varepsilon _1} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)\\delta _{\\varepsilon _1}(\\varepsilon _2)}{1-\\Delta _1(\\varepsilon _1)} d\\varepsilon _2 d\\varepsilon _1\\\\\n&= \\int _{d(x,P)}^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{1 - \\Delta _1(\\varepsilon _1)} \\int _{d(x,P)}^{\\varepsilon _1} \\delta _{\\varepsilon _1}(\\varepsilon _2) d\\varepsilon _2 d\\varepsilon _1\\\\\n&= \\int _{d(x,P)}^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{1 - \\Delta _1(\\varepsilon _1)} \\left( \\int _0^{\\varepsilon _1} \\delta _{\\varepsilon _1}(\\varepsilon _2) d\\varepsilon _2 - \\int _0^{d(x,P)} \\delta _{\\varepsilon _1}(\\varepsilon _2) d\\varepsilon _2 \\right) d\\varepsilon _1\\\\\n&= \\mu _{L_1}(x) - (1-\\mu _{L_1}(x))\\int _{d(x,P)}^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{1 - \\Delta _1(\\varepsilon _1)} d\\varepsilon _1 = \\mu _{L_1}(x) + (1-\\mu _{L_1}(x))\\ln (1 - \\mu _{L_1}(x))\n$ ","The effect of these hedges are illustrated in the following example.","Example 5 Suppose the label $L$ is as described in example \"Example 3\" . We can form a new label $vL$ with prototype $P_v = P = 5$ and threshold $\\varepsilon _v \\le \\varepsilon $ . Then, according to theorem \"Theorem 4 (L 2 =L_2 = very L 1 L_1)\" , $\\mu _{vL}(x) = \\mu _L(x) + (1 - \\mu _L(x))\\ln (1- \\mu _L(x))$ The effect of applying this contraction hedge is seen in figure 5 , and again, this concentration hedge is contrasted with Zadeh's concentration hedge $(\\mu _L(x))^2$ .","These hedges can also be applied across multiple dimensions, demonstrated in the example below.","Example 6 Suppose we have two labels `tall' and `thin'. `Tall' has prototype $P_{\\text{tall}} = 6.5$ ft and `thin' has prototype $P_{\\text{thin}} = 24$ in. The appropriateness of each label is defined by: $\n\\mu _{\\text{tall}}(x_1) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & if x_1 > 6.5\\\\\n1 - (P_{\\text{tall}} - x_1) & if x_1 \\in [5.5, 6.5]\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","where the variable $x_1$ measures height $\n\\mu _{\\text{thin}}(x_2) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & if x_2 < 24\\\\\n1 - \\frac{(x_2 - P_{\\text{thin}})}{4} & if x_2 \\in [24, 28]\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","where $x_2$ measures waist size.","Suppose further that being tall and being thin are independent of each other. The appropriateness of the label `tall and thin' could then be defined by: $\n\\mu _{\\text{tall and thin}}(x_1, x_2) = \\mu _{\\text{tall}}(x_1)\\mu _{\\text{thin}}(x_2)\n$ ","If `tall' and `thin' are independent, we can treat their hedges separately, so the appropriateness of a label `very tall and quite thin' is: $\n\\mu _{\\text{very tall and quite thin}}(x_1, x_2) &= \\mu _{\\text{very tall}}(x_1)\\mu _{\\text{quite thin}}(x_2)\\\\\n& = (\\mu _{\\text{tall}}(x_1)+ (1 - \\mu _{\\text{tall}}(x_1)\\ln (1 - \\mu _{\\text{tall}}(x_1))) (\\mu _{\\text{thin}}(x_2) - \\mu _{\\text{thin}}(x_2)\\ln (\\mu _{\\text{thin}}(x_2)))\n$ ","This is illustrated in figure 6 "],"section_name":"Hedges with unmodified prototypes"},{"paragraphs":["As they stand, the hedges proposed leave the core and support of the fuzzy sets unchanged, which is often argued to be undesirable BIBREF23 , BIBREF9 , BIBREF25 , BIBREF30 . A slight modification yields models of hedges in which the core, or prototype, of the concept has been changed.","Theorem 7 (Dilation) Suppose that $L_2 = $ quite $L_1$ , as in theorem \"Theorem 2 (L 2 =L_2 = quite L 1 L_1)\" , but that $P_2 \\ne P_1$ . Then $\\mu _{L_2}(x) = \\Delta _1(d(x,P_2)) - \\Delta _1(d(x, P_2))\\ln (\\Delta _1(d(x, P_2)))$ .","Substitute $\\Delta _1(d(x, P_2))$ for $\\mu _{L_1}(x)$ throughout proof of theorem \"Theorem 2 (L 2 =L_2 = quite L 1 L_1)\" ","Theorem 8 (Concentration) Suppose that $L_2 = $ very $L_1$ , as in theorem \"Theorem 4 (L 2 =L_2 = very L 1 L_1)\" , but that $P_2 \\ne P_1$ . Then $\\mu _{L_2}(x) = \\Delta _1(d(x,P_2)) + (1 - \\Delta _1(d(x, P_2)))\\ln (1 - \\Delta _1(d(x, P_2)))$ .","As above.","Corollary 9 If $\\varepsilon _2 \\ge \\varepsilon _1$ and $P_2 \\supseteq P_1$ then $\\mu _{L_2}(x) \\ge \\mu _{L_1}(x) - \\mu _{L_1}(x)\\ln (\\mu _{L_1}(x))$ , and if $\\varepsilon _2 \\le \\varepsilon _1$ and $P_2 \\subseteq P_1$ , then $\\mu _{L_2}(x) \\le \\mu _{L_1}(x) + (1-\\mu _{L_1}(x))\\ln (1-\\mu _{L_1}(x))$ ."," $\\mu _{L_2}(x) = \\Delta _1(d(x,P_2)) - \\Delta _1(d(x, P_2))\\ln (\\Delta _1(d(x, P_2)))$ , but since $P_2 \\supseteq P_1$ , $d(x, P_2) \\le d(x, P_1)$ $\\forall x \\in \\Omega $ , and so $\\Delta _1(d(x, P_2)) \\ge \\Delta _1(d(x, P_1)) = \\mu _{L_1}(x)$ $\\forall x \\in \\Omega $ . Hence, $\\mu _{L_2}(x) \\ge \\mu _{L_1}(x) - \\mu _{L_1}(x)\\ln (\\mu _{L_1}(x))$ . A similar argument shows that $\\mu _{L_2}(x) \\le \\mu _{L_1}(x) + (1-\\mu _{L_1}(x))\\ln (1-\\mu _{L_1}(x))$ .","Example 10 Suppose our conceptual space $\\Omega = \\mathbb {R}$ with Euclidean distance and that a label $L$ has prototype $P = [4.5,5.5]$ , and threshold $\\varepsilon \\sim $ Uniform $[0,3]$ . Then $\n\\mu _L(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } x \\in [4.5,5.5]\\\\\n1 - \\frac{|x - 5|}{3} & \\text{ if } x \\in [1.5, 4.5] \\text{ or } x \\in [5.5, 8.5] \\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","We form the concept $qL$ by setting the prototype to be $P_q = [4,6]$ , and $\\varepsilon _q \\ge \\varepsilon $ . The effect of applying our dilation hedge is illustrated in figure 7 .","Conversely, suppose that a label $L$ has prototype $P = [4, 6]$ and threshold $\\varepsilon \\sim $ Uniform $[0,3]$ . Then $\n\\mu _L(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } x \\in [4,6]\\\\\n1 - \\frac{|x - 5|}{3} & \\text{ if } x \\in [1, 4] \\text{ or } x \\in [6, 9]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","We now form the concept $vL$ by contracting the prototype to $P_v = [4.5,5.5]$ and setting $\\varepsilon _v \\le \\varepsilon $ . The effect of applying the contraction hedge is illustrated in figure 8 ."],"section_name":"Hedges with differing prototypes"},{"paragraphs":["It may be the case that the threshold of a given concept is greater than or less than a function of the original threshold. This could hold when a hedged concept is a very expanded or restricted version of the original concept, such as when the hedge `loosely' or `extremely' is used. Our formulae can also take account of this.","Theorem 11 Suppose $L_2 = $ is a dilation of $L_1 = $ with $P_2 \\ne P_1$ and $\\varepsilon _2 \\ge f(\\varepsilon _1)$ , where $f: \\mathbb {R} \\rightarrow \\mathbb {R}$ is strictly increasing or decreasing. Then $\n\\mu _{L_2}(x) = \\Delta _1(f^{-1}(d(x,P_2))) - \\Delta _1(f^{-1}(d(x,P_2)))\\ln (\\Delta _1(f^{-1}(d(x,P_2))))\n$ ","Rewrite $\\varepsilon _2 \\ge f(\\varepsilon _1)$ as $\\varepsilon _2 \\ge \\varepsilon = f(\\varepsilon _1)$ , where $\\varepsilon \\sim \\delta $ and is associated with a label $L$ with prototype $P$ . Then: $\n\\mu _{L_2}(x) = \\Delta (d(x,P_2)) - \\Delta (d(x, P_2))\\ln (\\Delta (d(x, P_2)))\n$ ","as above","Since $f: \\mathbb {R} \\rightarrow \\mathbb {R}$ is strictly monotone, $f^{-1}$ exists, and $\\Delta (d(x, P)) = P(d(x,P) \\le \\varepsilon ) = P(f^{-1}(d(x,P)) \\le \\varepsilon _1) = \\Delta _1(f^{-1}(d(x,P)))$ .","So $\n\\mu _{L_2}(x) = \\Delta _1(f^{-1}(d(x,P_2))) - \\Delta _1(f^{-1}(d(x,P_2)))\\ln (\\Delta _1(f^{-1}(d(x,P_2))))\n$ ","as required.","Theorem 12 Suppose $L_2 = $ is a concentration of $L_1 = $ with $P_2 \\ne P_1$ and $\\varepsilon _2 \\le f(\\varepsilon _1)$ , where $f: \\mathbb {R} \\rightarrow \\mathbb {R}$ is strictly increasing or decreasing. Then $\n\\mu _{L_2}(x) = \\Delta _1(f^{-1}(d(x,P_2))) +(1- \\Delta _1(f^{-1}(d(x,P_2))))\\ln (1 -\\Delta _1(f^{-1}(d(x,P_2))))\n$ ","The proof is entirely similar to that of theorem \"Theorem 11\" "],"section_name":"Functions of thresholds"},{"paragraphs":["It is possible to specify the dependence of the threshold of the hedged concept on the threshold of the unhedged concept purely deterministically, i.e. by $\\varepsilon _2 = f(\\varepsilon _1)$ , rather than $\\varepsilon _2 \\le f(\\varepsilon _1)$ . In this case, we can show links to other models of hedges from the literature.","A simple example of a deterministic dependency is given below.","Example 13 Suppose $\\Omega = \\mathbb {R}$ , $d$ is Euclidean distance and that $L_1$ has prototype $P = 5$ and $\\varepsilon \\sim $ Uniform $[0,3]$ . Then as before, $\n\\mu _L(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 - \\frac{|x - 5|}{3} & \\text{ if } x \\in [2, 8]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","To implement a dilation hedge, we would form a new label $qL$ with $P_q = P = 5$ and $\\varepsilon _q = k_q\\varepsilon $ with $k_q > 1$ . For a contraction hedge, we would form the label $vL$ by setting $P_v = P = 5$ and $\\varepsilon _v = k_v\\varepsilon $ with $k_v <1$ . Then, $\n\\mu _{hL}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 - \\frac{|x - 5|}{3k} & \\text{ if } x \\in [5 - 3k, 5+ 3k]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","where $h = q$ or $v$ , $k = k_q$ or $k_v$ respectively.","The effect of implementing these hedges is illustrated in figure 9 .","Using this approach, we can also create an effect similar to that of changing the prototype. Suppose that a label $L$ in a conceptual space $\\Omega $ has a single point $P$ as a prototype, but that the minimum value of the threshold $\\varepsilon $ is greater than 0, for example, $\\varepsilon \\sim $ Uniform $[c,a]$ . Then $\n\\mu _{L}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } d(x,P) < c\\\\\n\\frac{a}{a - c} - \\frac{|x - P|}{a - c} & \\text{ if } d(x, P) \\in [a,c]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","Suppose that a hedged concept $hL$ is formed from $L$ by the dependency $\\varepsilon _h = k\\varepsilon $ where $k$ is a constant. Then $\n\\mu _{L}(x) = \\Delta (\\frac{|x-P|}{k}) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } \\frac{|x-P|}{k} < c\\\\\n\\frac{a}{(a - c)} - \\frac{|x - P|}{k(a - c)} & \\text{ if } \\frac{|x-P|}{k} \\in [a,c]\\\\\n0 & otherwise\n\\end{array}\\right.}• =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } |x - P| < kc\\\\\n\\frac{ka - |x - P|}{k(a - c)} & \\text{ if } |x-P| \\in [ka,kc]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","This effect is illustrated in the example below.","Example 14 Suppose that the conceptual space $\\Omega = \\mathbb {R}$ and that a label $L$ has prototype $P = 5$ and threshold $\\varepsilon \\sim $ Uniform $[1, 2]$ . Then $\n\\mu _{L}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } |x - 5| < 1\\\\\n2 - |x - 5| & \\text{ if } |x - 5| \\in [1,2]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","Forming a new label $qL$ by applying the hedge $\\varepsilon _q = 2\\varepsilon $ gives appropriateness measure $\n\\mu _{qL}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } |x - 5| < 2\\\\\n\\frac{4 - |x - 5|}{2} & \\text{ if } |x - 5| \\in [2,4]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","Forming a new label $vL$ by applying the hedge $\\varepsilon _v = 0.5\\varepsilon $ gives appropriateness measure $\n\\mu _{vL}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } |x - 5| < 0.5\\\\\n\\frac{1 - |x - 5|}{0.5} & \\text{ if } |x - 5| \\in [0.5,1]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","These are illustrated in figure 10 .","Notice that if we set $\\Omega = \\mathbb {R}^+$ and label $L$ specified by $P = 0$ , $\\varepsilon \\sim $ Uniform $[c, a]$ , this is identical to the linear membership model given in BIBREF25 . Specifically, we have $\n\\mu _{L}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } x < c\\\\\n\\frac{a - x}{a-c} & \\text{ if } x \\in [c,a]\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","Forming a hedged concept $hL$ by setting $P_{hL}=P=0$ and $\\varepsilon _{hL} = k\\varepsilon $ gives $\n\\mu _{hL}(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 & \\text{ if } x < kc\\\\\n\\frac{ka - x}{ka-kc} & \\text{ if } x \\in [kc,ka]\\\\\n0 & \\text{otherwise}\n\\end{array}\\right.}\n$ ","Comparing this with the model given in section \"Approaches to linguistic hedges\" , we see that this is precisely the model proposed by BIBREF25 in the linear case.","Similarity between the hedging effects illustrated in figure 9 and the effects implemented in the model proposed in BIBREF23 , illustrated in figure 2 , can clearly be seen. To derive the model given in BIBREF23 , we describe the fuzzy sets associated with labels $L$ and $hL$ in trapezoidal notation. Notice that $L= (P-a, P - c, P+c, P+a)$ and $hL = (P-ka, P- kc, P+kc, P+ka)$ . We can render this transformation in the terms employed by BIBREF23 . Consider labels $L$ and $hL$ as fuzzy sets characterised by the appropriateness measure $\\mu _{L}(x)$ and $\\mu _{hL}(x)$ . Then, in the case of dilation, we have: $\nqL = E^Z(L)(s) = \\text{sup}_{r \\in \\Omega }T(\\mu _{L}(s), E^Z(s,r))\n$ ","and for contraction, $\nvL = E_Z(L)(s) = \\text{inf}_{r \\in \\Omega }I(\\mu _{L}(s), E^Z(s,r))\n$ ","When $T$ is the $T$ -norm $min$ , I is the Gödel implication and $Z = (-z-\\alpha , -z, z, z+\\alpha )$ (with the restriction that $c < z$ to ensure a well-defined set), the approach in BIBREF23 gives $qL = (P-a - z - \\alpha , P - c - z, P+c + z, P+a + z + \\alpha )$ . If we set $z = (k - 1)c$ and $\\alpha = (k - 1)(a - c)$ , this is equal to $qL = (P-ka, P- kc, P+kc, P+ka)$ .","However, we also require that $vL = (P-ka, P- kc, P+kc, P+ka)$ . The approach in BIBREF23 gives $vL = (P-a + z + \\alpha , P - c + z, P+c - z, P+a - z - \\alpha )$ , and we therefore need to set $z = (1-k)c$ and $\\alpha = (1-k)(a-c)$ ","This formulation is not as general as given in BIBREF23 , however, note that it only uses one additional parameter and no additional operators, rather than the two parameters and either a $T$ -norm or implication used by BIBREF23 .","Two more key models from the literature are the powering and shifting modifiers proposed in BIBREF3 . Recall that powering modifiers are of the form $\\mu _{hL}(x) = (\\mu _L(x))^k$ and shifting modifiers are of the form $\\mu _{hL}(x) = (\\mu _L(x-a))$ . Shifting modifiers are easy to implement within our model, simply by shifting the prototype by the quantity $a$ .","Powering modifiers can be expressed as a function of the threshold $\\varepsilon $ given a particular distribution of the threshold $\\delta $ . Suppose $\\Omega = \\mathbb {R}$ , $\\varepsilon \\sim U[0,c]$ , giving $\n\\mu _L(x) =\n{\\left\\lbrace \\begin{array}{ll}\n1 - \\frac{d(x, P)}{b} & \\text{ if } x \\in [P-b, P+b]\\\\\n0 & otherwise\n\\end{array}\\right.}•\n$ ","and suppose a new label $hL$ is formed with prototype $P$ and threshold $\\varepsilon _h = f(\\varepsilon )$ such that $\\mu _{hL}(x) = \\mu _L(x)^k$ . Then $\\mu _{hL}(x) = \\Delta (f^{-1}(d(x,P))) = (\\Delta (d(x,P)))^k$ , so $\nf^{-1}(d(x,P)) = \\Delta ^{-1}((\\Delta (d(x,P)))^k) = b - b(\\frac{(b - d(x, P))^k}{b^k}) = b - \\frac{(b - d(x,P))^k}{b^{k-1}}\n$ ","and hence $\n\\varepsilon _{hL} = f(\\varepsilon ) = b - (b^{k-1}(b - \\varepsilon ))^{1/k}\n$ ","This expression seems surprisingly complicated, and there may be better ways of deriving the powering hedges that are not as a function of the threshold $\\varepsilon $ .","In this section we have shown that our general model can capture some of the many approaches found in the literature as special cases. We now go on to look at the property of compositionality that is exhibited by a number of models."],"section_name":"Links to other models of hedges"},{"paragraphs":["One of the features of hedges seen in BIBREF23 , BIBREF31 , BIBREF25 , BIBREF3 is that they can be applied multiple times. Within the label semantics framework, this consists in expanding or reducing the threshold of a concept a number of times. The directed acyclic graph corresponding to this is shown in figure 11 . We show below that expressions for `very' and `quite' as given in theorems \"Theorem 2 (L 2 =L_2 = quite L 1 L_1)\" and \"Theorem 4 (L 2 =L_2 = very L 1 L_1)\" are compositional, and that the appropriateness of a concept after $n$ applications of a hedge can be expressed purely in terms of the appropriateness after $n -1$ applications. We also derive expressions for the composition of deterministic hedges as described in section \"Links to other models of hedges\" .","Theorem 15 Suppose that labels $L_1, L_2, ... , L_n$ are defined by prototypes $P_1 = P_2 = ... = P_n = P$ , thresholds $\\varepsilon _1 \\ge \\varepsilon _2 \\ge ... \\ge \\varepsilon _n$ and with a distance metric $d$ common to all labels. Then $\\mu _{L_n}(x) = \\mu _{L_{n-1}}(x) + (1-\\mu _{L_{n-1}}(x))\\ln (1 - \\mu _{L_{n-1}}(x))$ ","We proceed by induction on $n$ . Theorem \"Theorem 2 (L 2 =L_2 = quite L 1 L_1)\" proves this for $n=2$ . Assuming true for $n=k$ , we have $\n\\mu _{L_{k+1}}(x) &= \\int _0^\\infty \\int _0^\\infty ... \\int _{max(d(x,P), \\varepsilon _k)}^\\infty \\delta (\\varepsilon _1, \\varepsilon _2,...,\\varepsilon _{k+1}) \\mathrm {d}\\varepsilon _{k+1}...\\mathrm {d}\\varepsilon _1 \\\\\n&= \\int _0^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)} \\int _0^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} ... \\int _0^\\infty \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} \\int _{max(d(x,P), \\varepsilon _{k})}^\\infty \\delta _{\\varepsilon _k}(\\varepsilon _{k+1}) \\mathrm {d}\\varepsilon _{k+1}...\\mathrm {d}\\varepsilon _1 \\\\\n&=\\int _0^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)} \\int _0^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} ... \\int _{max(d(x, P), \\varepsilon _{k-1})}^\\infty \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} \\overbrace{\\int _{\\varepsilon _k}^\\infty \\delta _{\\varepsilon _k}(\\varepsilon _{k+1}) \\mathrm {d}\\varepsilon _{k+1}}^{= \\Delta _k(\\varepsilon _k)} \\mathrm {d}\\varepsilon _k...\\mathrm {d}\\varepsilon _1 \\\\\n& \\quad +\\int _0^{d(x,P)} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)} \\int _{\\varepsilon _1}^{d(x,P)} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} ... \\int _{\\varepsilon _{k-1}}^{d(x,P)} \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} \\overbrace{\\int _{d(x,P)}^\\infty \\delta _{\\varepsilon _k}(\\varepsilon _{k+1}) \\mathrm {d}\\varepsilon _{k+1}}^{= \\mu _{L_k}(x)} \\mathrm {d}\\varepsilon _k...\\mathrm {d}\\varepsilon _1 \\\\\n&= \\overbrace{\\int _0^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)} \\int _0^\\infty \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} ... \\int _{max(d(x, P), \\varepsilon _k)}^\\infty \\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)\\mathrm {d}\\varepsilon _k...\\mathrm {d}\\varepsilon _1 }^{= \\mu _{L_k}(x) \\text{ by ind. hyp.}}\\\\\n& \\quad + \\mu _{L_k}(x) \\int _0^{d(x,P)} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)} \\int _{\\varepsilon _1}^{d(x,P)} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} ... \\int _{\\varepsilon _{k-1}}^{d(x,P)} \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} \\mathrm {d}\\varepsilon _k...\\mathrm {d}\\varepsilon _1\\\\\n&= \\mu _{L_k}(x) + \\mu _{L_k}(x) \\int _0^{d(x,P)} \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} ... \\int _0^{\\varepsilon _3} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} \\int _0^{\\varepsilon _2} \\frac{\\delta _{\\varepsilon _1}(\\varepsilon _1)}{\\Delta _1(\\varepsilon _1)} \\mathrm {d}\\varepsilon _1...\\mathrm {d}\\varepsilon _k\\\\\n&= \\mu _{L_k}(x) + \\mu _{L_k}(x) \\overbrace{\\int _0^{d(x,P)} \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} ... \\int _0^{\\varepsilon _3} \\frac{-\\delta _{\\varepsilon _1}(\\varepsilon _2)\\ln (\\Delta _1(\\varepsilon _2))}{\\Delta _2(\\varepsilon _2)} \\mathrm {d}\\varepsilon _2...\\mathrm {d}\\varepsilon _k}^{=A}\n$ ","By the inductive hypothesis, $\\forall i = 0...k$ $\n\\delta _{\\varepsilon _i}(\\varepsilon _i) &= -\\frac{\\mathrm {d}}{\\mathrm {d\\varepsilon _i}}\\Delta _i(\\varepsilon _i)\\\\ \\nonumber &= -\\frac{\\mathrm {d}}{\\mathrm {d\\varepsilon _i}}(\\Delta _{i-1}(\\varepsilon _i) - \\Delta _{i-1}(\\varepsilon _i)\\ln (\\Delta _{i-1}(\\varepsilon _i)))\\\\\n&= -\\delta _{\\varepsilon _{i-1}}(\\varepsilon _i)\\ln (\\Delta _{i-1}(\\varepsilon _i))\n$ ","Recursively substituting in $A$ , we obtain $\n\\mu _{L_{k+1}}(x) &= \\mu _{L_k}(x) + \\mu _{L_k}(x) \\int _0^{d(x,P)} \\frac{\\delta _{\\varepsilon _{k-1}}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} ... \\int _0^{\\varepsilon _3} \\frac{\\delta _{\\varepsilon _2}(\\varepsilon _2)}{\\Delta _2(\\varepsilon _2)} \\mathrm {d}\\varepsilon _2...\\mathrm {d}\\varepsilon _k\\\\\n&= \\mu _{L_k}(x) +\\mu _{L_k}(x) \\int _0^{d(x, P)}\\frac{\\delta _{\\varepsilon _k}(\\varepsilon _k)}{\\Delta _k(\\varepsilon _k)} \\mathrm {d}\\varepsilon _k\\\\\n&= \\mu _{L_k}(x) -\\mu _{L_k}(x)\\ln (\\mu _{L_k}(x))\n$ ","Theorem 16 Suppose labels $L_1, L_2, ..., L_n$ are defined by prototypes $P_1 = P_2 = ... = P_n = P$ , thresholds $\\varepsilon _1 \\le \\varepsilon _2 \\le ... \\le \\varepsilon _n$ , and that distance metric $d$ is common to all. Then $\\mu _{L_n}(x) = \\mu _{L_{n-1}}(x) - \\mu _{L_{n-1}}(x)\\ln (\\mu _{L_{n-1}}(x))$ ","Similar to proof of theorem \"Theorem 15\" .","We can also derive expressions for the composition of deterministic hedges.","Theorem 17 Suppose labels $L_1, L_2, ..., L_n$ are defined by prototypes $P_1 = P_2 = ... = P_n = P$ , thresholds $\\varepsilon _n = f(\\varepsilon _{n-1}), \\varepsilon _{n-1} = f(\\varepsilon _{n-2}), ..., \\varepsilon _2 = f(\\varepsilon _1)$ , where $f$ is monotone increasing or decreasing, and that distance metric $d$ is common to all. Then $\\mu _{L_n}(x) = \\Delta _1(f^{-(n-1)}(d(x,P))$ , where $f^{-k}$ signifies $f^{-1}$ composed $k$ times."," $\\mu _{L_2}(x) = \\Delta _1(f^{-1}(d(x,P))$ . Suppose that $\\mu _{L_k}(x) = \\Delta _k(d(x,P)) = \\Delta _1(f^{-(k-1)}(d(x,P))$ . Since $\\varepsilon _{k+1} = f(\\varepsilon _k)$ , we have $\\mu _{L_{k+1}}(x) = \\Delta _k(f^{-1}(d(x,P)) = \\Delta _1(f^{-k}(d(x,P)))$ .","Therefore $\\mu _{L_n}(x) = \\Delta _1(f^{-(n-1)}(d(x,P))$ by induction.","Since labels can be composed in this way, we can model different degrees of emphasis corresponding to the composition of multiple hedges. So, for example, we could model `extremely L' as `very, very L'. This is illustrated in example \"Example 18\" .","Example 18 Suppose the label $L$ is as described in example \"Example 3\" , i.e. $L$ has prototype $P = 5$ , and threshold $\\varepsilon \\sim $ Uniform $[0,3]$ . We can form a new label $vL$ with prototype $P_v = P = 5$ and threshold $\\varepsilon _v \\le \\varepsilon $ , which has appropriateness $\\mu _{vL}(x) = \\mu _L(x) + (1 - \\mu _L(x))\\ln (1- \\mu _L(x))$ as shown in theorem \"Theorem 4 (L 2 =L_2 = very L 1 L_1)\" . We may then form another new label $vvL$ with prototype $L$0 and threshold $L$1 with appropriateness $L$2 as described in theorem \"Theorem 15\" . The effect of applying this contraction hedge is seen in figure 12 . We have contrasted the effect of the composed hedges with $L$3 .","Since these hedges can be composed only an integral number of times, we cannot obtain the differences in grade that could be achieved with using various powers in a powering modifier, e.g. $\\mu _L(x)^{1.73}$ . However, in section \"Functions of thresholds\" we discuss how to tune the intensity of hedges by using dependencies on functions of thresholds. We have further shown in section \"Links to other models of hedges\" how to derive powering and shifting modifiers within our framework. It would be interesting to explore how other examples of hedges can be expressed in this framework.","We have shown that when multiple hedges of the forms seen in theorems \"Theorem 2 (L 2 =L_2 = quite L 1 L_1)\" and \"Theorem 4 (L 2 =L_2 = very L 1 L_1)\" are used, $\\mu _{L_n}(x)$ can be expressed purely in terms of the appropriateness of the label directly preceding it. We have not been able to find a closed form solution for this recurrence, however, we can investigate the fixed points of the recurrence and examine what happens to the values of $\\mu _{L_n}(x)$ as $n \\rightarrow \\infty $ . We have also shown that deterministic hedges can be composed, and we go on to look at their behaviour in the limit of composition."],"section_name":"Compositionality"},{"paragraphs":["The following results examine the behaviour of $\\mu _{L_n}(x)$ as $n \\rightarrow \\infty $ ","Theorem 19 Suppose $L_1, ..., L_n$ are labels obtained by repeated application of the dilation operator. Then $\\mu _{L_n}$ has a limit $M^+$ and $M^+ = 1$ $\\forall x \\in \\Omega $ such that $\\mu _{L_1}(x) \\ne 0$ , and $M^+ = 0$ otherwise."," $\\mu _{L_{i+1}}(x) = \\mu _{L_i}(x) - \\mu _{L_i}(x)\\ln (\\mu _{L_i}(x))$ , $i = 1,.., n-1$ . If $\\mu _{L_1}(x) = 1$ then $\\mu _{L_i}(x) = 1$ $\\forall i = 1,..., n$ . Also, if $\\mu _{L_1}(x) = 0$ then $\\mu _{L_i}(x) = 0$ $\\forall i = 1,..., n$ . Suppose $\\mu _{L_i}(x) \\in (0,1)$ . Then $\\mu _{L_{i+1}}(x) > \\mu _{L_i}(x)$ , and so for $i = 1,.., n-1$0 , $i = 1,.., n-1$1 is a strictly increasing sequence.","If a limit $M^+$ exists, then we will have $M^+ = M^+ - M^+\\ln (M^+)$ , so either $M^+ = 0$ or $\\ln (M^+) = 0$ . We can't have $M^+ = 0$ , since we assume that $\\mu _{L_1}(x) \\in (0,1)$ and the sequence is strictly increasing. Therefore, we must have $\\ln (M^+) = 0$ and therefore $M^+ = 1$ . So $\n\\mu _{L_\\infty }(x) = \\left\\lbrace \n\\begin{array}{l l}\n1 & \\quad \\mu _{L_1}(x) \\in (0,1]\\\\\n0 & \\quad \\mu _{L_1}(x) = 0\\\\\n\\end{array} \\right.\n$ ","Theorem 20 Suppose $L_1,..., L_n$ are labels obtained by repeated application of the contraction operator. Then $\\mu _{L_n}$ has a limit $M^-$ and $M^- = 0$ $\\forall x \\in \\Omega $ such that $\\mu _{L_1}(x) \\ne 1$ , and $M^- = 1$ otherwise."," $\\mu _{L_{i+1}}(x) = \\mu _{L_i}(x) + (1- \\mu _{L_i}(x))\\ln (1 - \\mu _{L_i}(x))$ , $i = 1,..., n-1$ . Again, if $\\mu _{L_1}(x) = 1$ then $\\mu _{L_i}(x) = 1$ $\\forall i = 1,..., n$ . Also, if $\\mu _{L_1}(x) = 0$ then $\\mu _{L_i}(x) = 0$ $\\forall i = 1,..., n$ , and for $\\mu _{L_1}(x) \\in (0,1)$ , $\\mu _{L_1}(x) >...> \\mu _{L_n}(x)$ is a strictly decreasing sequence.","If a limit $M^-$ exists, then $\nM^- &= M^- + (1 - M^-)\\ln (1- M^-)\\\\\n\\ln (1 - M^-) &= M^-\\ln (1- M^-)\n$ ","So either $M^- = 0$ or $\\ln (1- M^-) = 0$ . If $\\ln (1- M^-) = 0$ then $M^- = 1$ , which is impossible since $\\mu _{L_1}(x) \\in (0,1)$ and the sequence of $\\mu _{L_i}(x)$ is strictly decreasing. Therefore $\n\\mu _{L_\\infty }(x) = \\left\\lbrace \n\\begin{array}{l l}\n0 & \\quad \\mu _{L_1}(x) \\in [0,1)\\\\\n1 & \\quad \\mu _{L_1}(x) = 1\\\\\n\\end{array} \\right.\n$ ","We have shown here that in the limit, the result of applying dilation or contraction modifiers multiple times is to create a crisp set. In the case of dilation, the crisp set includes the whole support of the fuzzy set associated with the original label, whereas in the case of contraction, the concept reduces to include only its prototype.","When deterministic hedges are used, i.e. $\\varepsilon _2 = f(\\varepsilon _1)$ , the behaviour of the limit depends on the behaviour of the function $f$ and its properties in the limit as $n \\rightarrow \\infty $ of $f^{-n}$ .","Example 21 Suppose $f(\\varepsilon ) = 0.5\\varepsilon $ . Applying this hedge multiple times will result in $\\mu _{L_n}(x) = \\Delta _1(2^n d(x,P))$ . As $n\\rightarrow \\infty $ , $2^n d(x,P) \\rightarrow \\infty $ , except where $d(x,P)=0$ . Therefore, $\n\\mu _{L_\\infty }(x) = \\left\\lbrace \n\\begin{array}{l l}\n0 & \\quad d(x,P) > 0\\\\\n1 & \\quad d(x,P) = 0\\\\\n\\end{array} \\right.\n$ ","On the other hand, if $f(\\varepsilon ) = 2\\varepsilon $ , $\\mu _{L_n}(x) = \\Delta _1(2^{-n}d(x,P))$ . As $n\\rightarrow \\infty $ , $2^{-n} d(x,P) \\rightarrow 0$ , and hence $\\mu _{L_\\infty }(x) = 1$ $\\forall x \\in \\Omega $ .","The behaviour of the hedges given in example \"Example 21\" is therefore different from those in theorems \"Theorem 19\" and \"Theorem 20\" , since the concept either shrinks to a single point, in the case of contraction, or, in the case of dilation, expands to fill the entire space $\\Omega $ ."],"section_name":"Limits of Compositions."},{"paragraphs":["We have presented formulae for linguistic hedges which are both functional and semantically grounded. The modifiers presented arise naturally from the label semantics framework, in which concepts are represented by a prototype and threshold. Our hedges have an intuitive meaning: if I think that the threshold for a concept `small' is of a certain width, then the threshold for the concept `very small' will be narrower. On the other hand, the threshold for the concept `quite small' will be broader. The hedges proposed are examples of `type 1' hedges, i.e. they operate equally across all dimensions of the fuzzy set associated with a concept. The first result presented is somewhat similar to a powering modifier since the core and support of the set remain the same. In BIBREF9 , BIBREF7 , it is argued that this property is undesirable for hedges used in fuzzy expert systems, since if a query is returning too large a set of answers, this type of contraction hedge does not reduce this overabundance. However, although the hedges we propose do not at their simplest address the overabundance issue, we argue that they address another problem associated with powering hedges, in that they have a clear semantic grounding that the powering modifiers lack."," BIBREF23 , BIBREF31 also propose modifiers that are semantically grounded, using the idea of resemblance to nearby objects. Their formulations have the properties that the core and support of the fuzzy set are both changed, thereby addressing the issue of overabundant answers BIBREF9 , BIBREF7 . In our most specific case, since the prototype is not altered, the core and support of the fuzzy set representing the concept remain the same. However, our initial proposal can be generalised, as in section \"Hedges with differing prototypes\" , to apply to the case where $P_1 \\ne P_2$ . Specifying a semantically meaningful way of altering the boundaries of the prototype would answer the objection that the core and support of a set should change under a linguistic hedge.","The most general result (section \"Functions of thresholds\" ) shows that the formula still applies when $\\varepsilon _2 \\le f(\\varepsilon _1)$ , or $\\varepsilon _2 \\ge f(\\varepsilon _1)$ . Combined with a distribution $\\delta $ such that the lower bound of the distribution is not zero, the core and support of the fuzzy set are modified. With the condition $\\varepsilon _2 = f(\\varepsilon _1)$ , we are able to recreate the result given in BIBREF25 for linear membership functions, and show how the model proposed by BIBREF23 has strong similarities to our own. In this case we have introduced additional parameters, so the simplicity of the original result is lost. However, the further parameters introduced are no more than those introduced by BIBREF25 , and arguably fewer than those introduced by BIBREF23 , who require that a resemblance relation be specified, using two additional parameters, and also, that a $T$ -norm or fuzzy implication need to be specified. There are various choices of operator that could be used for either of these, and it is not obvious that any one is better than the others.","We have also shown that the basic case operators `very' and `quite' can be composed, which is not immediately obvious from the formulae (section \"Compositionality\" ). Further, we show that in the limit of composition the membership of any object in the fuzzy part of $L$ , i.e. $\\lbrace x: 0 < \\mu _L(x) < 1\\rbrace $ , increases to 1 in the case of `quite' or decreases to 0 in the case of `very' (section \"Limits of Compositions.\" ). This is similar to the limit of applying the powering modifiers, but differs from what would happen with the modifiers proposed by BIBREF23 . In that case, the limit of `very' would shrink to a single point and the limit of `quite' would expand to encompass the whole universe of discourse. This can be modelled using the deterministic hedges described in section \"Links to other models of hedges\" . Although behaviour differs slightly, in fact human discourse does not apply modifiers infinitely, so the difference in behaviour is arguably not important.","Our formulation has the benefit that it can be applied in more situations than simply linguistic hedges. For example, the concept `apple green' has a prototype different to that of just green, and the threshold for `apple green' is likely to be smaller than the threshold for simply `green'. Our model can take account of this."],"section_name":"Discussion"},{"paragraphs":["We have presented formulae for two simple linguistic hedges, `very' and `quite'. These formulae are functional, hence easy to compute, but also semantically grounded, in that they arise naturally from the conceptual framework of label semantics combined with prototype theory and conceptual spaces theory, and in the most specific case require no additional parameters. We have also shown that two other formulations BIBREF23 , BIBREF31 , BIBREF25 , can be derived from this framework with equal or fewer parameters. We have shown that the hedges can be composed and have described their behaviour in the limit of composition.","Further work could look at testing the utility of these hedges in particular classifiers to compare their performance with the classical hedges and with the hedges used by e.g. BIBREF9 , BIBREF7 , and also to examine a trade-off between accuracy and the number of parameters used. Alternatively, investigating semantically grounded ways of expanding or reducing prototypes could have a similar impact.","The model could also be extended to the more complicated type 2 hedges such as `essentially', or `technically', by treating dimensions of the conceptual space heterogeneously. This requires using some type of weighting or necessity measure on the dimensions, work which is currently ongoing."],"section_name":"Conclusions and further work"},{"paragraphs":["Martha Lewis gratefully acknowledges support from EPSRC Grant No. EP/E501214/1"],"section_name":"Acknowledgements"}],"paper_id":"1601.06738","title":"A Label Semantics Approach to Linguistic Hedges"},{"abstract":"As the name suggests, type-logical grammars are a grammar formalism based on logic and type theory. From the prespective of grammar design, type-logical grammars develop the syntactic and semantic aspects of linguistic phenomena hand-in-hand, letting the desired semantics of an expression inform the syntactic type and vice versa. Prototypical examples of the successful application of type-logical grammars to the syntax-semantics interface include coordination, quantifier scope and extraction.This chapter describes the Grail theorem prover, a series of tools for designing and testing grammars in various modern type-logical grammars which functions as a tool . All tools described in this chapter are freely available.","figures_and_tables":[{"caption":"Fig. 9.1 The standard architecture of type-logical grammars","file":"3-Figure9.1-1.png"},{"caption":"Table 9.2 Lambek calculus lexicon","file":"5-Table9.2-1.png"},{"caption":"Fig. 9.2 “Every student aced some exam” with the subject wide scope reading.","file":"6-Figure9.2-1.png"},{"caption":"Fig. 9.3 Deep structure of the derivation of Figure 9.2.","file":"7-Figure9.3-1.png"},{"caption":"Fig. 9.4 Intuitionistic proof and lambda term corresponding to the deep structure of Figure 9.3.","file":"7-Figure9.4-1.png"},{"caption":"Fig. 9.5 “De re” reading for the sentence “Alyssa believes someone committed perjury”.","file":"14-Figure9.5-1.png"},{"caption":"Table 9.4 Natural deduction rules for MILL1","file":"15-Table9.4-1.png"},{"caption":"Table 9.5 Links for multimodal proof nets","file":"19-Table9.5-1.png"},{"caption":"Fig. 9.6 Lexical unfolding of a/ab,b/ac ` a/ac","file":"20-Figure9.6-1.png"},{"caption":"Fig. 9.7 The proof structure of Figure 9.6 after identification of the a and b atoms (left) and after identification of all atoms","file":"20-Figure9.7-1.png"},{"caption":"Fig. 9.8 The proof structure of Figure 9.7 (left) and its abstract proof structure (right)","file":"21-Figure9.8-1.png"},{"caption":"Table 9.6 Contractions — multimodal binary connectives","file":"22-Table9.6-1.png"},{"caption":"Fig. 9.9 Structural rewrites for associativity of mode a.","file":"23-Figure9.9-1.png"},{"caption":"Fig. 9.10 Structural rewrite and contraction for the abstract proof structure of Figure 9.8, showing this is a proof net for a/ab◦a b/ac ` a/ac","file":"24-Figure9.10-1.png"},{"caption":"Fig. 9.11 Interactive Grail output","file":"24-Figure9.11-1.png"},{"caption":"Table 9.7 Logical links for MILL1 proof structures","file":"26-Table9.7-1.png"},{"caption":"Fig. 9.12 Proof structure for a( ∃x.b(x) ` ∃y.[a( b(y)].","file":"27-Figure9.12-1.png"},{"caption":"Fig. 9.13 Correction graph for the proof structure of Figure 9.12 with the cycle indicated, showing a( ∃x.b(x) ` ∃y.[a( b(y)] is underivable","file":"28-Figure9.13-1.png"},{"caption":"Table 9.8 Contractions for first-order linear logic. Conditions: vi 6= v j and, for the u contraction, all free occurrences of x are at v j .","file":"28-Table9.8-1.png"},{"caption":"Fig. 9.14 Contractions for the underivable a( ∃x.b(x) ` ∃y.[a( b(y)].","file":"29-Figure9.14-1.png"},{"caption":"Table 9.9 The different theorem provers","file":"30-Table9.9-1.png"}],"full_text":[{"paragraphs":["This chapter describes a series of tools for developing and testing type-logical grammars. The Grail family of theorem provers have been designed to work with a variety of modern type-logical frameworks, including multimodal type-logical grammars BIBREF0 , NL $_{cl}$ BIBREF1 , the Displacement calculus BIBREF2 and hybrid type-logical grammars BIBREF3 .","The tools give a transparent way of implementing grammars and testing their consequences, providing a natural deduction proof in the specific type-logical grammar for each of the readings of a sentence. None of this replaces careful reflection by the grammar writer, of course, but in many cases, computational testing of hand-written grammars will reveal surprises, showing unintended consequences of our grammar and such unintended proofs (or unintended absences of proofs) help us improve the grammar. Computational tools also help us speed up grammar development, for example by allowing us to compare several alternative solutions to a problem and investigate where they make different predictions.","This chapter describes the underlying formalism of the theorem provers, as it is visible during an interactive proof trace, and present the general strategy followed by the theorem provers. The presentation in this chapter is somewhat informal, referring the reader elsewhere for full proofs.","The rest of this chapter is structured as follows. Section \"Type-logical grammars\" presents a general introduction to type-logical grammars and illustrates its basic concepts using the Lambek calculus, ending the section with some problems at the syntax-semantics interface for the Lambek calculus. Section \"Modern type-logical grammars\" looks at recent developments in type-logical grammars and how they solve some of the problems at the syntax-semantics interface. Section \"Theorem proving\" looks at two general frameworks for automated theorem proving for type-logical grammars, describing the internal representation of partial proofs and giving a high-level overview of the proof search mechanism."],"section_name":"Introduction"},{"paragraphs":["Type-logical grammars are a family of grammar formalisms built on a foundation of logic and type theory. Type-logical grammars originated when BIBREF4 introduced his Syntactic Calculus (called the Lambek calculus, L, by later authors). Though Lambek built on the work of BIBREF5 , BIBREF6 and others, Lambek's main innovation was to cast the calculus as a logic, giving a sequent calculus and showing decidability by means of cut elimination. This combination of linguistic and computational applications has proved very influential.","In its general form, a type-logical grammar consists of following components:","A sentence $w_1, \\ldots , w_n$ is grammatical iff the statement $A_1,\\ldots , A_n \\vdash C$ is provable in our logic, for some $A_i\n\\in \\textit {lex}(w_i)$ and for some goal formula $C$ . In other words, we use the lexicon to map words to formulas and then ask the logic whether the resulting sequence of formulas is a theorem. Parsing in a type-logical grammar is quite literally a form of theorem proving, a very pure realisation of the slogan “parsing as deduction”.","One of the attractive aspects of type-logical grammars is their simple and transparent syntax-semantics interface. Though there is a variety of logics used for the syntax of type-logical grammars (I will discuss the Lambek calculus in Section \"The Lambek calculus\" and two generalisations of it in Sections \"Multimodal grammars\" and \"First-order linear logic\" ), there is a large consensus over the syntax-semantics interface. Figure 1 gives a picture of the standard architecture of type-logical grammars.","The “bridge” between syntax and semantics in the figure is the Curry-Howard isomorphism between linear lambda terms and proofs in multiplicative intuitionistic linear logic.","Theorem proving occurs in two places of the picture: first when parsing a sentence in a given type-logical grammar and also at the end when we use the resulting semantics for inferences. I will have little to say about this second type of theorem proving BIBREF9 , BIBREF10 ; theorem proving for parsing will be discussed in Section \"Theorem proving\" .","The lexicon plays the role of translating words to syntactic formulas but also specifies the semantic term which is used to compute the semantics later. The lexicon of a categorial grammar is “semantically informed”. The desired semantics of a sentence allows us to reverse-engineer the formula and lexical lambda-term which produce it.","Many current semantic theories do not provide a semantic formula directly, but first provide a proto-semantics on which further computations are performed to produce the final semantics (eg. for anaphora resolution, presuppositions projection etc.). In the current context this means at least some inference is necessary to determine semantic and pragmatic wellformedness."],"section_name":"Type-logical grammars"},{"paragraphs":["To make things more concrete, I will start by presenting the Lambek calculus BIBREF4 . Lambek introduced his calculus as a way to “obtain an effective rule (or algorithm) for distinguishing sentences from nonsentences”, which would be applicable both to formal and to (at least fragments of) natural languages BIBREF4 . The simplest formulas used in the Lambek calculus are atomic formulas, which normally include $s$ for sentence, $n$ for common noun, $np$ for noun phrase. We then inductively define the set of formulas of the Lambek calculus by saying that, they include the atomic formulas, and that, if $A$ and $B$ are formulas (atomic or not), then $A/B$ , $A\\bullet B$ and $B\\backslash A$ are also formulas.","The intended meaning of a formula $A/B$ — called $A$ over $B$ — is that it is looking for an expression of syntactic type $B$ to its right to produce an expression of syntactic type $A$ . An example would be a word like “the” which is assigned the formula $np/n$ in the lexicon, indicating that it is looking for a common noun (like “student”) to its right to form a noun phrase, meaning “the student” would be assigned syntactic type $np$ . Similarly, the intended meaning of a formula $B\\backslash A$ — called $B$ under $A$ — is that it is looking for an expression of syntactic type $A$0 to its left to produce an expression of type $A$1 . This means an intransitive verb like “slept”, when assigned the formula $A$2 in the lexicon, combines with a noun phrase to its left to form a sentence $A$3 . We therefore predict that “the student slept” is a sentence, given the earlier assignment of $A$4 to “the student”. Finally, a formula $A$5 denotes the concatenation of an expression of type $A$6 to an expression of type $A$7 .","Basic statements of the Lambek calculus are of the form $A_1,\\ldots ,A_n \\vdash C$ (with $n\n\\ge 1$ ), indicating a claim that the sequence of formulas $A_1,\\ldots , A_n$ is of type $C$ ; the sequent comma `,' is implicitly associative and non-commutative. Table 1 shows the natural deduction rules for the Lambek calculus. $\\Gamma $ , $\\Delta $ , etc. denote non-empty sequences of formulas.","A simple Lambek calculus lexicon is shown in Table 2 . I have adopted the standard convention in type-logical grammars of not using set notation for the lexicon, but instead listing multiple lexical entries for a word separately. This corresponds to treating $\\textit {lex}$ as a non-deterministic function rather than as a set-valued function.","Proper names, such as “Alyssa” and “Emory” are assigned the category $np$ . Common nouns, such as “student” and “exam” are assigned the category $n$ . Adjectives, such as “difficult” or “erratic” are not assigned a basic syntactic category but rather the category $n/n$ , indicating they are looking for a common noun to their right to form a new common noun, so we predict that both “difficult exam” and “exam” can be assigned category $n$ . For more complex entries, “someone” is looking to its right for a verb phrase to produce a sentence, where $np\\backslash s$ is the Lambek calculus equivalent of verb phrase, whereas “whom” is first looking to its right for a sentence which is itself missing a noun phrase to its right and then to its left for a noun.","Given the lexicon of Table 2 , we can already derive some fairly complex sentences, such as the following, and, as we will see in the next section, obtain the correct semantics.",". Every student aced some exam.",". The student who slept during the exam loves Alyssa.","One of the two derivations of Sentence \"The Lambek calculus\" is shown in Figure 2 . To improve readability, the figure uses a “sugared” notation: instead of writing the lexical hypothesis corresponding to “exam” as $n \\vdash n$ , we have written it as $\\textit {exam} \\vdash n$ . The withdrawn $np$ 's corresponding to the object and the subject are given a labels $p_0$ and $q_0$ respectively; the introduction rules are coindexed with the withdrawn hypotheses, even though this information can be inferred from the rule instantiation.","We can always uniquely reconstruct the antecedent from the labels. For example, the sugared statement “ $p_0\\ \\textrm {aced}\\ q_0 \\vdash s$ ” in the proof corresponds to $np, (np\\backslash s)/np, np \\vdash s$ .","Although it is easy to verify that the proof of Figure 2 has correctly applied the rules of the Lambek calculus, finding such a proof from scratch may look a bit complicated (the key steps at the beginning of the proof involve introducing two $np$ hypotheses and then deriving $s/np$ to allow the object quantifier to take narrow scope). We will defer the question “given a statement $\\Gamma \\vdash C$ , how do we decide whether or not it is derivable?” to Section \"Theorem proving\" but will first discuss how this proof corresponds to the following logical formula. $\n\\forall x. [\\mathit {student}(x) \\Rightarrow \\exists y. [\\mathit {exam}(y) \\wedge \\mathit {ace}(x,y) ] ]\n$ "],"section_name":"The Lambek calculus"},{"paragraphs":["For the Lambek calculus, specifying the homomorphism to multiplicative intuitionistic linear logic is easy: we replace the two implications ` $\\backslash $ ' and ` $/$ ' by the linear implication ` $\\multimap $ ' and the product ` $\\bullet $ ' by the tensor ` $\\otimes $ '. In a statement $\\Gamma \\vdash C$ , $\\Gamma $ is now a multiset of formulas instead of a sequence. In other words, the sequent comma `,' is now associative, commutative instead of associative, non-commutative. For the proof of Figure 2 of the previous section, this mapping gives the proof shown in Figure 3 .","We have kept the order of the premisses of the rules as they were in Figure 2 to allow for an easier comparison. This deep structure still uses the same atomic formulas as the Lambek calculus, it just forgets about the order of the formulas and therefore can no longer distinguish between the leftward looking implication ` $\\backslash $ ' and the rightward looking implication ` $/$ '.","To obtain a semantics in the tradition of BIBREF11 , we use the following mapping from syntactic types to semantic types, using Montague's atomic types $e$ (for entity) and $t$ (for truth value). $\nnp^* & = e\\\\\nn^* & = e\\rightarrow t\\\\\ns^* & = t\\\\\n(A \\multimap B)^* & = A^* \\rightarrow B^*\n$ ","Applying this mapping to the deep structure proof of Figure 3 produces the intuitionistic proof and the corresponding (linear) lambda term as shown in Figure 4 ","The computed term corresponds to the derivational semantics of the proof. To obtain the complete meaning, we need to substitute, for each of $z_0, \\ldots , z_4$ , the meaning assigned in the lexicon.","For example, “every” has syntactic type $(s/(np\\backslash s))/n$ and its semantic type is $(e\\rightarrow t)\\rightarrow (e\\rightarrow t)\\rightarrow t$ . The corresponding lexical lambda term of this type is $\\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. (\\forall (\\lambda x^e. ((\\Rightarrow (P\\, x)) (Q\\, x))))$ , with ` $\\forall $ ' a constant of type $(e\\rightarrow t)\\rightarrow t$ and ` $\\Rightarrow $ ' a constant of type $t\\rightarrow (t\\rightarrow t)$ . In the more familiar Montague formulation, this lexical term corresponds to $\\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. \\forall x. [ (P\\, x)\n\\Rightarrow (Q\\,x)]$ , where we can see the formula in higher-order logic we are constructing more clearly. Although the derivational semantics is a linear lambda term, the lexical term assigned to “every” is not, since the variable $x$ has two bound occurrences.","The formula assigned to “some” has the same semantic type but a different term $\\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. (\\exists (\\lambda x^e. ((\\wedge (P\\, x)) (Q\\, x))))$ .","The other words are simple, “exam” is assigned $\\mathit {exam}^{e\\rightarrow t}$ , “student” is assigned $\\mathit {student}^{e\\rightarrow t}$ , and “aced” is assigned $\\mathit {ace}^{e\\rightarrow (e\\rightarrow t)}$ .","So to compute the meaning, we start with the derivational semantics, repeated below. $\n((z_0\\,z_1) \\,(\\lambda x. ((z_3\\,z_4)\\,\\lambda y. ((z_2\\,y)\\,x))))\n$ ","Then we substitute the lexical meanings, for $z_0,\\ldots ,z_4$ . $\nz_0& := \\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. (\\forall (\\lambda x^e. ((\\Rightarrow (P\\, x)) (Q\\, x))))\\\\\nz_1&:= \\mathit {student}^{e\\rightarrow t}\\\\\nz_2& := \\mathit {ace}^{e\\rightarrow (e\\rightarrow t)}\\\\\nz_3& := \\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. (\\exists (\\lambda x^e. ((\\wedge (P\\, x)) (Q\\, x))))\\\\\nz_4& := \\mathit {exam}^{e\\rightarrow t}\\\\\n$ ","This produces the following lambda term. $\n((\\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. & (\\forall (\\lambda x^e. ((\\Rightarrow (P\\, x)) (Q\\,\nx))))\\,\\mathit {student}^{e\\rightarrow t}) \\\\\n\\,(\\lambda x. ((\\lambda P^{e\\rightarrow t}.\\lambda Q^{e\\rightarrow t}. & (\\exists (\\lambda x^e. ((\\wedge (P\\,\nx)) (Q\\,\nx))))\\,\\mathit {exam}^{e\\rightarrow t})\\\\\n&\\lambda y. ((\\mathit {ace}^{e\\rightarrow (e\\rightarrow t)}\\,y)\\,x))))\n$ ","Finally, when we normalise this lambda term, we obtain the following semantics for this sentence. $\n(\\forall (\\lambda x^e. ((\\Rightarrow (\\mathit {student}^{e\\rightarrow t})\\, x)) (\\exists (\\lambda y^e. ((\\wedge (\\mathit {exam}^{e\\rightarrow t}\\,\ny)) (((\\mathit {ace}^{e\\rightarrow (e\\rightarrow t)}\\,y)\\,x)))))\n$ ","This lambda term represents the more readable higher-order logic formula. $\n\\forall x. [\\mathit {student}(x) \\Rightarrow \\exists y. [\\mathit {exam}(y) \\wedge \\mathit {ace}(x,y) ] ]\n$ ","Proofs in the Lambek calculus, and in type-logical grammars are subsets of the proofs in intuitionistic (linear) logic and these proofs are compatible with formal semantics in the tradition initiated by BIBREF11 .","For the example in this section, we have calculated the semantics of a simple example in “slow motion”: many authors assign a lambda term directly to a proof in their type-logical grammar, leaving the translation to intuitionistic linear logic implicit.","Given a semantic analysis without a corresponding syntactic proof, we can try to reverse engineer the syntactic proof. For example, suppose we want to assign the reflexive “himself” the lambda term $\\lambda R^{(e\\rightarrow e\\rightarrow t)}\\lambda x^e. ((R\\,x)\\, x)$ , that is, a term of type $(e\\rightarrow e\\rightarrow t)\\rightarrow e\\rightarrow t$ . Then, using some syntactic reasoning to eliminate implausible candidates like $(np\\multimap n)\\multimap n$ , the only reasonable deep structure formula is $(np\\multimap np\\multimap s)\\multimap (np\\multimap s)$ and, reasoning a bit further about which of the implications is left and right, we quickly end up with the quite reasonable (though far from perfect) Lambek calculus formula $((np\\backslash s)/np)\\backslash (np\\backslash s)$ ."],"section_name":"The syntax-semantics interface"},{"paragraphs":["Though the Lambek calculus is a beautiful and simple logic and though it gives a reasonable account of many interesting phenomena on the syntax-semantics interface, the Lambek calculus has a number of problems, which I will discuss briefly below. The driving force of research in type-logical grammars since the eighties has been to find solutions to these problems and some of these solutions will be the main theme of the next section.","The Lambek calculus generates only context-free languages BIBREF12 . There is a rather large consensus that natural languages are best described by a class of languages at least slightly larger than the context-free languages. Classical examples of phenomena better analysed using so-called mildly context-sensitive language include verb clusters in Dutch and in Swiss German BIBREF13 , BIBREF14 .","Though our example grammar correctly predicted two readings for Sentence \"The Lambek calculus\" above, our treatment of quantifiers doesn't scale well. For example, if we want to predict two readings for the following sentence (which is just Sentence \"The Lambek calculus\" where “some” and “every” have exchanged position)",". Some student aced every exam.","then we need to add an additional lexical entry both for “some” and for “every”; this is easily done, but we end up with two lexical formulas for both words. However, this would still not be enough. For example, the following sentence is also grammatical.",". Alyssa gave every student a difficult exam.",". Alyssa believes a student committed perjury.","In Sentence UID18 , “every student” does not occur in a peripheral position, and though it is possible to add a more complex formula with the correct behaviour, we would need yet another formula for Sentence UID18 . Sentence UID18 is generally considered to have two readings: a de dicto reading, where Alyssa doesn't have a specific student in mind (she could conclude this, for example, when two students make contradictory statements under oath, this reading can be felicitously followed by “but she doesn't know which”), and a de re reading where Alyssa believes a specific student perjured. The Lambek calculus cannot generate this second reading without adding yet another formula for “a”.","It seems we are on the wrong track when we need to add a new lexical entry for each different context in which a quantifier phrase occurs. Ideally, we would like a single formula for “every”, “some” and “a” which applied in all these different cases.","Another way to see this is that we want to keep the deep structure formula $n\\multimap ((np\\multimap s) \\multimap s)$ and that we need to replace the Lambek calculus by another logic such that the correct deep structures for the desired readings of sentences like UID18 and UID18 are produced.","The grammar above also overgenerates in several ways.","“ace” implies a (very positive) form of evaluation with respect to the object. “aced the exam” is good, whereas “aced Emory”, outside of the context of a tennis match is bad. “aced logic” can only mean something like “aced the exam for the logic course”.","“during” and similar temporal adverbs imply its argument is a temporal interval: “during the exam” is good, but “during the student” is bad, and “during logic” can only mean something like “during the contextually understood logic lecture”","In the literature on semantics, there has been an influential movement towards a richer ontology of types (compared to the “flat” Montagovian picture presented above) but also towards a richer set of operations for combining terms of specific types, notably allowing type coercions BIBREF15 , BIBREF16 . So an “exam” can be “difficult” (it subject matter, or informational content) but also “take a long time” (the event of taking the exam). The theory of semantics outlined in the previous section needs to be extended if we want to take these and other observations into account."],"section_name":"Going further"},{"paragraphs":["We ended the last section with some problems with using the Lambek calculus as a theory of the syntax-semantics interface. The problems are of two different kinds."],"section_name":"Modern type-logical grammars"},{"paragraphs":["Multimodal type-logical grammars BIBREF0 take the non-associative Lambek calculus as its base, but allow multiple families of connectives.","For the basic statements $\\Gamma \\vdash C$ of the Lambek calculus, we ask the question whether we can derive formula $C$ , the succedent, from a sequence of formulas $\\Gamma $ , the antecedent. In the multimodal Lambek calculus, the basic objects are labeled binary trees. The labels come from a separate set of indices or modes $I$ . Multimodal formulas are then of the form $A/_i\nB$ , $A\\bullet _i B$ and $A\\backslash _i B$ , and antecedent terms are of the form $\\Gamma \\circ _{i} \\Delta $ , with $C$0 an index from $C$1 (we have omitted the outer brackets for the rules, but the operator $C$2 is non-associative). Sequents are still written as $C$3 , but $C$4 is now a binary branching, labeled tree with formulas as its leaves.","Given a set of words $w_1,\\ldots ,w_n$ and a goal formula $C$ , the question is now: is there a labeled tree $\\Gamma $ with formulas $A_1,\\ldots ,A_n$ as its yield, such that $\\Gamma \\vdash C$ is derivable and $A_i \\in \\textit {lex}(w_i)$ for all $i$ (the implementation of Section \"Multimodal proof nets\" will automatically compute such a $\\Gamma $ ).","The rules of multimodal type-logical grammars are shown in Table 3 . In the rules, $\\Gamma [\\Delta ]$ denotes an antecedent tree $\\Gamma $ with distinguished subtree $\\Delta $ — the subtree notation is a non-associative version of the Lambek calculus antecedent $\\Gamma ,\\Delta ,\\Gamma ^{\\prime }$ , where $\\Delta $ is a subsequence instead of a subtree as it is in $\\Gamma [\\Delta ]$ .","Each logical connective with mode $i$ uses a structural connective $\\circ _i$ in its rule. For the $/ E$ , $\\bullet I$ and $\\backslash E$ rules, reading from premisses to conclusions, we build structure. For the $/I$ , $\\bullet E$ and $\\backslash I$ rules we remove a structural connective with the same mode as the logical connective. The natural deduction rules use explicit antecedents, although, for convenience, we will again use coindexation between the introduction rules for the implications ` $/$ ' and ` $\\backslash $ ' and its withdrawn premiss (and similarly for the $\\circ _i$0 rule and its two premisses).","The main advantage of adding modes to the logic is that modes allow us to control the application of structural rules lexically. This gives us fine-grained control over the structural rules in our logic.","For example, the base logic is non-associative. Without structural rules, the sequent $a/b, b/c \\vdash a/c$ , which is derivable in the Lambek calculus is not derivable in its multimodal incarnation $a/_a b, b/_a c \\vdash a/_a c$ . The proof attempt below, with the failed rule application marked by the `' symbol, shows us that the elimination rules and the introduction rule for this sequent do not match up correctly. $\n[[/ I]]{a/_ab \\circ _{a} b/_ac\\vdash a/_a c\n}{[\\text{}]{(a/_ab \\circ _{a} b/_ac) \\circ _{a} c \\vdash a}{[[/ E]]{a/_a b \\circ _{a} (b/_a c \\circ _{a} c)\\vdash a}{a/_a b\\vdash a/_a b & [[/ E]]{b/_a c \\circ _{a} c \\vdash b}{b/_a\nc \\vdash b/_a c & c\\vdash c}}}}\n$ ","This is where the structural rules, shown at the bottom of Table 3 come in. The general form, read from top to bottom, states that we take a structure $\\Gamma $ containing a distinguished subtree $\\Xi $ which itself has $n$ subtrees $\\Delta _1,\\ldots ,\\Delta _n$ , and we replace this subtree $\\Xi $ with a subtree $\\Xi ^{\\prime }$ which has the same number of subtrees, though not necessarily in the same order ( $\\pi $ is a permutation on the leaves). In brief, we replace a subtree $\\Xi $ by another subtree $\\Xi ^{\\prime }$ and possibly rearrange the leaves (subtrees) of $\\Xi $ , without deleting or copying any subtrees. Examples of structural rules are the following.","The first structural rule is one of the structural rules for associativity. It is the simplest rule which will make the proof attempt above valid (with $\\Gamma []$ the empty context, $\\Delta _1 = a/_a\nb$ , $\\Delta _2 = b/_a c$ and $\\Delta _3 = c$ ). This structural rule keeps the order of the $\\Delta _i$ the same.","The rule above on the right is slightly more complicated. There, the positions of $\\Delta _2$ and $\\Delta _3$ are swapped as are the relative positions of modes 0 and 1. Rules like this are called “mixed commutativity”, they permit controlled access to permutation. One way to see this rule, seen from top to bottom, is that is “moves out” a $\\Delta _3$ constituent which is on the right branch of mode 1. Rules of this kind are part of the solution to phenomena like Dutch verb clusters BIBREF17 .","Many modern type-logical grammars, such as the Displacement calculus and NL $_{cl}$ can be seen as multimodal grammars BIBREF18 , BIBREF1 ."],"section_name":"Multimodal grammars"},{"paragraphs":["We have seen that multimodal type-logical grammars generalise the Lambek calculus by offering the possibility of fine-tuned controlled over the application of structural rules. In this section, I will introduce a second way of extending the Lambek calculus.","Many parsing algorithms use pairs of integers to represent the start and end position of substrings of the input string. For example, we can represent the sentence",". Alyssa believes someone committed perjury.","as follows (this is a slightly simplified version of Sentence UID18 from Section \"Going further\" ); we have treated “committed perjury” as a single word.","[node distance=5em] 0) 0; 1) [right of=0]1; 2) [right of=1]2; 3) [right of=2]3; 4) [node distance=10em, right of=3]4; (0) edge node[above] [label] Alyssa (1); (1) edge node[above] [label] believes $_{\\rule {0pt}{1ex}}$ (2); (2) edge node[above] [label] someone $_{\\rule {0pt}{1ex}}$ (3); (3) edge node[above] [label] committed perjury $_{\\rule {0pt}{1ex}}$ (4);","The basic idea of first-order linear logic as a type-logical grammar is that we can code strings as pairs (or, more generally, tuples) of integers representing string positions. So for deciding the grammaticality of a sequence of words $w_1,\\ldots , w_n \\vdash C$ , with a goal formula $C$ , we now give a parametric translation from $\\Vert \nA_i \\Vert ^{i-1,i}$ for each lexical entry $w_i$ and $\\Vert C\\Vert ^{0,n}$ for the conclusion formula.","Given these string positions, we can assign the noun phrase “Alyssa” the formula $np(0,1)$ , that is a noun phrase from position 0 to position 1. The verb “believes”, which occurs above between position 1 and 2, can then be assigned the complex formula $\\forall x_2. [\ns(2,x_2) \\multimap \\forall x_1. [ np(x_1,1) \\multimap s(x_1,x_2)] ]$ , meaning that it first selects a sentence to its right (that is, starting at its right edge, position 2 and ending anywhere) and then a noun phrase to its left (that is, starting anywhere and ending at its left edge, position 1) to produce a sentence from the left position of the noun phrase argument to the right position of the sentence argument.","We can systematise this translation, following BIBREF19 , and obtain the following translation from Lambek calculus formulas to first-order linear logic formulas. $\n\\Vert p \\Vert ^{x,y} & = p(x,y) \\\\\n\\Vert A / B \\Vert ^{x,y} &= \\forall z. \\Vert B \\Vert ^{y,z} \\multimap \\Vert A \\Vert ^{x,z}\n\\\\\n\\Vert B\\backslash A \\Vert ^{y,z} &= \\forall x. \\Vert B \\Vert ^{x,y} \\multimap \\Vert A \\Vert ^{x,z}\n\\\\\n\\Vert A \\bullet B \\Vert ^{x,z} &= \\exists y. \\Vert A \\Vert ^{x,y} \\otimes \\Vert B\n\\Vert ^{y,z}\n$ ","Given this translation, the lexical entry for “believes” discussed above is simply the translation of the Lambek calculus formula $(np\\backslash s)/s$ , with position pair $1,2$ , to first-order linear logic. Doing the same for “committed perjury” with formula $np\\backslash s$ and positions $3,4$ gives $\\forall z. [np(z,3) \\multimap s(z,4)]$ . For “someone” we would simply translate the Lambek calculus formula $s/(np\\backslash s)$ , but we can do better than that: when we translate “someone” as $\\forall y_1. \\forall y_2. [ (np(2,3) \\multimap s(y_1,y_2)) \\multimap s(y_1,y_2) ]$ , we improve upon the Lambek calculus analysis.","As we noted in Section \"Going further\" , the Lambek calculus cannot generate the “de re” reading, where the existential quantifier has wide scope. Figure 5 shows how the simple first-order linear logic analysis does derive this reading.","Besides the Lambek calculus, first-order linear logic has many other modern type-logical grammars as fragments. Examples include lambda grammars BIBREF20 , the Displacement calculus BIBREF2 and hybrid type-logical grammars BIBREF3 . We can see first-order linear logic as a sort of “machine language” underlying these different formalisms, with each formalism introducing its own set of abbreviations convenient for the grammar writer. Seeing first-order linear logic as an underlying language allows us to compare the analyses proposed for different formalisms and find, in spite of different starting points, a lot of convergence. In addition, as discussed in Section \"First-order proof nets\" , we can use first-order linear logic as a uniform proof strategy for these formalisms.","As usual, we obtain the deep structure of a syntactic derivation by defining a homomorphism from the syntactic proof to a proof in multiplicative intuitionistic linear logic. For first-order linear logic, the natural mapping simply forgets all first-order quantifiers and replaces all atomic predicates $p(x_1,\\ldots ,x_n)$ by propositions $p$ . Since the first-order variables have, so far, only been used to encode string positions, such a forgetful mapping makes sense.","However, other solutions are possible. When we add semantically meaningful terms to first-order linear logic, the Curry-Howard isomorphism for the first-order quantifiers will give us dependent types and this provides a natural connection to the work using dependent types for formal semantics BIBREF21 , BIBREF22 , BIBREF23 , BIBREF24 ."],"section_name":"First-order linear logic"},{"paragraphs":["In the previous sections, we have discussed two general solutions to the problems of the syntax-semantics interface of the Lambek calculus. Both solutions proposed a more flexible syntactic logic. In this section, we will discuss a different type of added flexibility, namely in the syntax-semantics interface itself.","The basic motivating examples for a more flexible composition have been amply debated in the literature BIBREF15 , BIBREF16 . Our solution is essentially the one proposed by BIBREF25 , called the Montagovian Generative Lexicon. I will only give a brief presentation of this framework. More details can be found in Chapter 6.","Like many other solutions, the first step consists of splitting Montague's type $e$ for entities into several types: physical objects, locations, informational objects, eventualities, etc. Although there are different opinions with respect to the correct granularity of types BIBREF15 , BIBREF16 , BIBREF26 , nothing much hinges on this for the present discussion.","The second key element is the move to the second-order lambda calculus, system F BIBREF27 , which allows abstraction over types as well as over terms. In our Lambek calculus, the determiner “the” was assigned the formula $np/n$ and the type of its lexical semantics was therefore $(e\\rightarrow t) \\rightarrow e$ , which we implement using the $\\iota $ operators of type $(e\\rightarrow t) \\rightarrow e$ , which, roughly speaking, selects a contextually salient entity from (a characteristic function of) a set. When we replace the single type $e$ by several different types, we want to avoid listing several separate syntactically identical by semantically different entries for “the” in the lexicon, and therefore assign it a polymorphic term $\\Lambda \\alpha . \\iota ^{(\\alpha \\rightarrow t)\\rightarrow \\alpha }$ of type $\\Pi \\alpha . ((\\alpha \\rightarrow t)\\rightarrow \\alpha )$ , quantifying over all types $\\alpha $ . Though this looks problematic, the problem is resolved once we realise that only certain function words (quantifiers, conjunctions like “and”) are assigned polymorphic terms and that we simply use universal instantiation to obtain the value of the quantifier variable. So if “student” is a noun of type human, that is of type $h\\rightarrow t$ , then “the student” will be of type $h$ , instantiating $(e\\rightarrow t) \\rightarrow e$0 to $(e\\rightarrow t) \\rightarrow e$1 . Formally, we use $(e\\rightarrow t) \\rightarrow e$2 reduction as follows (this is substitution of types instead of terms, substituting type $(e\\rightarrow t) \\rightarrow e$3 for $(e\\rightarrow t) \\rightarrow e$4 ). $(e\\rightarrow t) \\rightarrow e$5 ","The final component of the Montagovian Generative Lexicon is a set of lexically specified, optional transformations. In case of a type mismatch, an optional transformation can “repair” the term.","As an example from BIBREF28 and BIBREF29 , one of the classic puzzles in semantics are plurals and collective and distributive readings. For example, verbs like “meet” have collective readings, they apply to groups of individuals collectively, so we have the following contrast, where collectives like committees and plurals like students can meet, but not singular or distributively quantified noun phrases. The contrast with verbs like “sneeze”, which force a distributive reading is clear.",". The committee met.",". All/the students met",". *A/each/the student met.",". All/the students sneezed.",". A/each/the student sneezed.","In the Montagovian Generative lexicon, we can models these fact as follows. First, we assign the plural morphology “-s” the semantics $\\Lambda \\alpha \\lambda P^{\\alpha \\rightarrow t} \\lambda Q^{\\alpha \\rightarrow t}. | Q | > 1 \\wedge \\forall x^{\\alpha }. Q(x) \\Rightarrow P(x)$ , then “students” is assigned the following term $\\lambda Q^{h\\rightarrow t}. | Q | > 1 \\wedge \\forall x^h. Q(x) \\Rightarrow \\textit {student}(x)$ , that is the sets of cardinality greater than one such that all its members are students. Unlike “student” which was assigned a term of type $h\\rightarrow t$ , roughly a property of humans, the plural “students” is assigned a term of type $(h\\rightarrow t)\\rightarrow t$ , roughly a property of sets of humans. Consequently, the contrast between “the student” and “the students” is that the first is of type $h$ (a human) and the second of type $h\\rightarrow t$ (a set of humans) as indicated below.","Therefore, the meaning of “the students” is the contextually determined set of humans, from the sets of more than one human such that all of them are students.","Then we distinguish the verbs “meet” and “sneeze” as follows, with the simpler verb “sneeze” simply selecting for a human subject and the collective verb “meet” selecting for a set of humans (of cardinality greater than one) as its subject.","Given these basic lexical entries, we already correctly predict that “the student met” is ill-formed semantically (there is an unresolvable type mismatch) but “the students met” and “the student sneezed” are given the correct semantics.","The interesting case is “the students sneezed” which has as its only reading that each student sneezed individually. Given that “the students” is of type $h\\rightarrow t$ and that “sneezed” requires an argument of type $h$ , there is a type mismatch when we apply the two terms. However, “sneeze” has the optional distributivity operator `*', which when we apply it to the lexical semantics for “sneeze” produces the term $\\lambda P^{h\\rightarrow t}. \\forall x^h. P(x) \\Rightarrow \\textit {sneeze}(x)$ , which combines with “the students” to produce the reading. $\n\\forall x^h. (\\iota (\\lambda Q^{h\\rightarrow t}. | Q | > 1 \\wedge \\forall y^h Q(y) \\Rightarrow \\textit {student}(y))\\, x) \\Rightarrow \\textit {sneeze}(x)\n$ ","In other words, all of the members of the contextually determined set of more than human which are all students, sneeze.","The basic idea for the Montagovian Generative Lexicon is that lexical entries specify optional transformations which can repair certain sorts of type mismatches in the syntax-semantics interface. This adaptability allows the framework to solve many semantic puzzles.","Though a proof-of-concept application of these ideas exists, more robust and scalable applications, as well as efforts incorporate these ideas into wide-coverage semantics, are ongoing research."],"section_name":"The Montagovian Generative Lexicon"},{"paragraphs":["When looking at the rules and examples for the different logics, the reader may have wondered: how do we actually find proofs for type-logical grammars? This question becomes especially urgent once our grammars become more complex and the consequences of our lexical entries, given our logic, become hard to oversee. Though pen and paper generally suffice to show that a given sentence is derivable for the desired reading, it is generally much more laborious to show that a given sentence is underivable or that it has only the desired readings. This is where automated theorem provers are useful: they allow more extensive and intensive testing of your grammars, producing results more quickly and with less errors (though we should be careful about too naively assuming the implementation we are using is correct: when a proof is found it is generally easy to verify its correctness by hand, but when a proof isn't found because of a programming error this can be hard to detect).","Though the natural deduction calculi we have seen so far can be used for automated theorem proving BIBREF30 , BIBREF31 , and though BIBREF4 already gave a sequent calculus decision procedure, both logics have important drawbacks for proof search.","Natural deduction proofs have a 1-1 correspondence between proofs and readings, though this is somewhat complicated to enforce for a logic with the $\\bullet \\textit {E}$ rule (and the related $\\Diamond \\textit {E}$ rule). For the sequent calculus, the product rule is just like the other rules, but sequent calculus suffers from the so-called “spurious ambiguity” problem, which means that it generates many more proofs than readings.","Fortunately, there are proof systems which combine the good aspects of natural deduction and sequent calculus, and which eliminate their respective drawbacks. Proof nets are a graphical representation of proofs first introduced for linear logic BIBREF32 . Proof nets suffer neither from spurious ambiguity nor from complications for the product rules.","Proof nets are usually defined as a subset of a larger class, called proof structures. Proof structures are “candidate proofs”: part of the search space of a naive proof search procedure which need not correspond to actual proofs. Proof nets are those proof structures which correspond to sequent proofs. Perhaps surprisingly, we can distinguish proof nets from other proof structures by looking only at graph-theoretical properties of these structures.","Proof search for type-logical grammars using proof nets uses the following general procedure.","In Sections \"Multimodal proof nets\" and \"First-order proof nets\" we will instantiate this general procedure for multimodal type-logical grammar and for first-order linear logic respectively."],"section_name":"Theorem proving"},{"paragraphs":["Table 5 presents the links for multimodal proof nets. The top row list the links corresponding to the elimination rules of natural deduction, the bottom row those corresponding to the introduction rules. There are two types of links: tensor links, with an open center, and par links, with a filled center. Par links have a single arrow pointing to the main formula of the link (the complex formula containing the principal connective). The top and bottom row are up-down symmetric with tensor and par reversed. The tensor links correspond to the logical rules which build structure when we read them from top to bottom, the par links to those rules which remove structure.","The formulas written above the central node of a link are its premisses, whereas the formulas written below it are its conclusions. Left-to-right order of the premisses as well as the conclusions is important.","A proof structure is a set of formula occurrences and a set of links such that:","each formula is at most once the premiss of a link,","each formula is at most once the conclusion of a link.","A formula which is not the premiss of any link is a conclusion of the proof structure. A formula which is not the conclusion of any link is a hypothesis of the proof structure. We say a proof structure with hypotheses $\\Gamma $ and conclusions $\\Delta $ is a proof structure of $\\Gamma \\vdash \\Delta $ (we are overloading of the ` $\\vdash $ ' symbol here, though this use should always be clear from the context; note that $\\Delta $ can contain multiple formulas).","After the first step of lexical lookup we have a sequent $\\Gamma \\vdash C$ , and we can enumerate its proof structures as follows: unfold the formulas in $\\Gamma , C$ , unfolding them so that the formulas in $\\Gamma $ are hypotheses and the formula $C$ is a conclusion of the resulting structure, until we reach the atomic subformulas (this is step 2 of the general procedure), then identify atomic subformulas (step 3 of the general procedure, we turn to the last step, checking correctness, below). This identification step can, by the conditions on proof structures only identify hypotheses with conclusions and must leave all formulas of $\\Gamma $ , including atomic formulas, as hypotheses and $C$ as a conclusion.","Figure 6 shows the lexical unfolding of the sequent $a/_a b, b/_a c \\vdash a/_a c$ . It is already a proof structure, though a proof structure of $a, a/_a b, b, b/_a c, c \\vdash a, a/_a c, b, c$ (to the reader familiar with the proof nets of linear logic: some other presentations of proof nets use more restricted definitions of proof structures where a “partial proof structure” such as shown in the figure is called a module).","To turn this proof structure into a proof structure of $a/_a b, b/_a c\n\\vdash a/_a c$ , we identify the atomic formulas. In this case, there is only a single way to do this, since $a$ , $b$ and $c$ all occur once as a hypothesis and once as a conclusion, though in general there may be many possible matchings. Figure 7 shows, on the left, the proof structure after identifying the $a$ and $b$ formulas. Since left and right (linear order), up and down (premiss, conclusion) have meaning in the graph, connecting the $c$ formulas is less obvious: $c$ is a conclusion of the $/I$ link and must therefore be below it, but a premiss of the $/E$ link and must therefore be above it. This is hard to achieve in the figure shown on the left. Though a possible solution would be to draw the figure on a cylinder, where “going up” from the topmost $a$0 we arrive at the bottom one, for ease of type-setting and reading the figure, I have chosen the representation shown in Figure 7 on the right. The curved line goes up from the $a$1 premiss of the $a$2 link and arrives from below at the $a$3 link, as desired. One way so see this strange curved connection is as a graphical representation of the coindexation of a premiss with a rule in the natural deduction rule for the implication.","Figure 7 therefore shows, on the right, a proof structure for $a/_a b, b/_a c\n\\vdash a/_a c$ . However, is it also a proof net, that is, does it correspond to a proof? In a multimodal logic, the answer depends on the available structural rules. For example, if no structural rules are applicable to mode $a$ then $a/_a b, b/_a c\n\\vdash a/_a c$ is underivable, but if mode $a$ is associative, then it is derivable.","We decide whether a proof structure is a proof net based only on properties of the graph. As a first step, we erase all formula information from the internal nodes of the graph; for administrative reasons, we still need to be able to identify which of the hypotheses and conclusion of the structure correspond to which formula occurrence. All relevant information for correctness is present in this graph, which we call an abstract proof structure.","We talked about how the curved line in proof structures (and abstract proof structure) corresponds to the coindexation of discharged hypotheses with rule names for the implication introduction rules. However, the introduction rules for multimodal type-logical grammars actually do more than just discharge a hypothesis, they also check whether the discharged hypothesis is the immediate left (for $\\backslash I$ ) or right (for $/ I$ ) daughter of the root node, that is, that the withdrawn hypothesis $A$ occurs as $A\\circ _i \\Gamma $ (for $\\backslash I$ and mode $i$ ) or $\\Gamma \\circ _i A$ (for $/I$ and mode $i$ ). The par links in the (abstract) proof structure represent a sort of “promise” that will produce the required structure. We check whether it is satisfied by means of contractions on the abstract proof structure.","The multimodal contractions are shown in Table 6 . All portrayed configurations contract to a single vertex: we erase the two internal vertices and the paired links and we identify the two external vertices, keeping all connections of the external vertices to the rest of the abstract proof structure as they were: the vertex which is the result of the contraction will be a conclusion of the same link as the top external vertex (or a hypothesis of the abstract proof structure in case it wasn't) and it will be a premiss of the same link as the bottom external vertex (or a conclusion of the abstract proof structure in case it wasn't).","The contraction for $/I$ checks if the withdrawn hypothesis is the right daughter of a tensor link with the same mode information $i$ , and symmetrically for the $\\backslash I$ contraction. The $\\bullet E$ contraction contracts two hypotheses occurring as sister nodes.","All contractions are instantiations of the same pattern: a tensor link and a par link are connected, respecting left-right and up-down the two vertices of the par link without the arrow.","To get a better feel for the contractions, we will start with its simplest instances. When we do pattern matching on the contraction for $/ I$ , we see that it corresponds to the following patterns, depending on our choice for the tensor link (the par link is always $/ I$ ). $\nC/_i B &\\vdash C/_i B \\\\\nA & \\vdash (A\\bullet _i B)/_i B \\\\\nA & \\vdash C/_i (A\\backslash _i C)\n$ ","A proof structure is a proof net iff it contracts to a tree containing only tensor links using the contractions of Table 6 and any structural rewrites, discussed below — BIBREF33 present full proofs. In other words, we need to contract all par links in the proof structure according to their contraction, each contraction ensuring the correct application of the rule after which it is named. The abstract proof structure on the right of Figure 8 does not contract, since there is no substructure corresponding to the $/I$ contraction: for a valid contraction, a par link is connected to both “tentacles” of a single tensor link, and in the figure the two tentacles without arrow are connected to different tensor links. This is correct, since $a/_a b, b/_a c\\vdash a/_a c$ is underivable in a logic without structural rules for $a$ .","However, we have seen that this statement becomes derivable once we add associativity of $a$ and it is easily verified to be a theorem of the Lambek calculus. How can we add a modally controlled version of associativity to the proof net calculus? We can add such a rule by adding a rewrite from a tensor tree to another tensor tree with the same set of leaves. The rewrite for associativity is shown in Figure 9 . To apply a structural rewrite, we replace the tree on the left hand side of the arrow by the one on the right hand side, reattaching the leaves and the root to the rest of the proof net.","Just like the structural rules, a structural rewrite always has the same leaves on both sides of the arrow — neither copying nor deletion is allowed, though we can reorder the leaves in any way (the associativity rule doesn't reorder the leaves).","Figure 10 shows how the contractions and the structural rewrites work together to derive $a/_a b, b/_a c \\vdash a/_a c$ .","We start with a structural rewrite, which rebrackets the pair of tensor links. The two hypotheses are now the premisses of the same link, and this also produces a contractible structure for the $/I$ link. Hence, we have shown the proof structure to be a proof net.","In the Grail theorem prover, the representation of abstract proof structures looks as shown in Figure 11 (this is an automatically produced subgraph close to the graph on the left of Figure 10 , though with a non-associative mode $n$ and therefore not derivable). This graph is used during user interaction. The graphs are drawn using GraphViz, an external graph drawing program which does not guarantee respecting our desires for left, right and top/bottom, so tentacles are labeled 1, 2 and 3 (for left, right and top/bottom respectively) to allow us to make these distinctions regardless of the visual representation. Vertices are given unique identifiers for user interaction, for example to allow specifying which pair of atoms should be identified or which par link should be contracted.","Although the structural rules give the grammar writer a great deal of flexibility, such flexibility complicates proof search. As discussed at the beginning of Section \"Theorem proving\" , theorem proving using proof nets is a four step process, which in the current situation looks as follows: 1) lexical lookup, 2) unfolding, 3) identification of atoms, 4) graph rewriting. In the current case, both the graph rewriting and the identification of atoms are complicated and since we can interleave the atom connections and the graph rewriting it is not a priori clear which strategy is optimal for which set of structural rules. The current implementation does graph rewriting only once all atoms have been connected.","The Grail theorem prover implements some strategies for early failure. Since all proofs in multimodal type-logical grammars are a subset of the proofs in multiplicative linear logic, we can reject (partial) proof structures which are invalid in multiplicative linear logic, a condition which is both powerful and easy to check.","As a compromise between efficiency and flexibility, Grail allows the grammar writer to specify a first-order approximation of her structural rules. Unlike the test for validity in multiplicative linear logic which is valid for any set of structural rules, specifying such a first-order approximation is valid only when there is a guarantee that all derivable sequents in the multimodal grammar are a subset of their approximations derivable in first-order linear logic. Errors made here can be rather subtle and hard to detect. It is recommended to use such methods to improve parsing speed only when a grammar has been sufficiently tested and where it is possible to verify whether no valid readings are excluded, or, ideally, to prove that the subset relation holds between the multimodal logic and its first-order approximation.","The next section will discuss first-order proof nets in their own right. Though these proof nets have been used as an underlying mechanism in Grail for a long time, we have seen in Section \"First-order linear logic\" that many modern type-logical grammars are formulated in a way which permits a direct implementation without an explicit set of structural rules.","As to the proof search strategy used by Grail, it is an instance of the “dancing links” algorithm BIBREF35 : when connecting atomic formulas, we always link a formula which has the least possibilities and we rewrite the abstract proof structures only once a fully linked proof structure has been produced. Though the parser is not extremely fast, evaluation both on randomly generated statements and on multimodal statements extracted from corpora show that the resulting algorithm performs more than well enough BIBREF36 ."],"section_name":"Multimodal proof nets"},{"paragraphs":["Proof nets for first-order linear logic BIBREF37 are a simple extension of the proof nets for standard, multiplicative linear logic BIBREF38 . Compared to the multimodal proof nets of the previous section, all logical links have the main formula of the link as their conclusion but there is now a notion of polarity, corresponding to whether or not the formula occurs on the left hand side of the turnstile (negative polarity) or on the right hand side (positive polarity).","We unfold a sequent $A_1,\\ldots ,A_n \\vdash C$ by using the negative unfolding for each of the $A_i$ and the positive unfolding for $C$ . The links for first-order proof nets are shown in Table 7 .","Contrary to multimodal proof nets, where a tensor link was drawn with an open central node and a par link with a filled central node, here par links are drawn as a connected pair of dotted lines and tensor links as a pair of solid lines.","As before, premisses are drawn above the link and conclusions are drawn below it. With the exception of the cut and axiom links, the order of the premisses and the conclusions is important. We assume without loss of generality that every quantifier link uses a distinct eigenvariable.","A set of formula occurrences connected by links is a proof structure if every formula is at most once the premiss of a link and if every formula is exactly once the conclusion of a link. Those formulas which are not the premiss of any link are the conclusions of the proof structure — note the difference with multimodal proof nets: a proof structure has conclusions but no hypotheses and, as a consequence, each formula in the proof net must be the conclusion of exactly one (instead of at most one) link.","For polarised proof nets, unfolding the formulas according to the links of Table 7 no longer produces a proof structure, since the atomic formulas after unfolding are not the conclusions of any link. Such “partial proof structures” are called a modules. To turn a module into a proof structure, we connect atomic formulas of opposite polarity by axiom links until we obtain a complete matching of the atomic formulas, that is until every atomic formula is the conclusion of an axiom link.","The negative $\\forall $ and the positive $\\exists $ link, are defined using substitution of an arbitrary term $t$ for the eigenvariable of the link. In actual proof search, we use unification of these variables when the axiom links are performed.","As usual, not all proof structures are proof nets. However, since the logical rules for the quantifiers make essential use of the notion of “free occurrence of a variable”, this should be reflected in out correctness condition. BIBREF37 uses a notion of switching for proof structures which extends the switchings of BIBREF38 .","A switching is, for each of the binary par links a choice of its left or right premiss and for each of the unary par links with eigenvariable $x$ a choice of one of the formulas in the structure with a free occurrence of $x$ or of the premiss of the rule.","Given a switching, a correction graph replaces a binary par link by a connection from the conclusion of the link to the premiss chosen by the switching, and it replace a unary par link by a link from the conclusion to the formula chosen by the switching.","Finally, a proof structure is a proof net when all its correction graphs are both acyclic and connected BIBREF37 .","As an example, look at the proof structure of $a\\multimap \\exists x.b(x) \\vdash \\exists y. [a\\multimap b(y)]$ shown in Figure 12 . This statement is not derivable in first-order linear logic (nor in intuitionistic logic). Consider therefore the switching connecting the binary par link to its left premiss $a$ and the link for $x$ to the formula $a\\multimap b(x)$ (it has a free occurrence of $x$ , so this like is a valid switching).","This switching produces the correction graph shown in Figure 13 . It contains a cycle, drawn with bold edges, and is therefore not a proof structure (in addition, the $b$ axiom is disconnected from the rest of the structure, giving a second reason for rejecting the proof structure).","Though switching conditions for proof nets are simple and elegant, they don't lend themselves to naive application: already for the example proof structure of Figure 12 there are six possible switchings to consider and, as the reader can verify, only the switching shown in Figure 13 is cyclic (and disconnected). In general, it is often the case that all switchings but one are acyclic and connected, as it is here.","Though there are efficient ways of testing acyclicity and connectedness for multiplicative proof nets BIBREF39 , BIBREF40 and it seems these can be adapted to the first-order case (though some care needs to be taken when we allow complex terms), the theorem prover for first-order linear logic uses a extension of the contraction criterion of BIBREF41 .","Given a proof structure we erase all formulas from the vertices and keep only a set of the free variables at this vertex. We then use the contractions of Table 8 to contract the edges of the graph. The resulting vertex of each contraction has the union of the free variables of the two vertices of the redex (we remove the eigenvariable $x$ of a $\\forall $ contraction, “ $\\Rightarrow _u$ ”). A proof structure is a proof net iff it contracts to a single vertex using the contractions of Table 8 .","To give an example of the contractions, Figure 14 shows the contractions for the underivable proof structure of Figure 12 . The initial structure, which simply takes the proof structure of Figure 12 and replaces the formulas by the corresponding set of free variables, is shown on the left. Contracting the five solid edges using the $c$ contraction produces the structure shown in the figure on the right.","No further contractions apply: the two connected dotted links from the binary par link do not end in the same vertex, so the par contraction $p$ cannot apply. In addition, the universal contraction $u$ cannot apply either, since it requires all vertices with its eigenvariable $x$ to occur at the node from which the arrow is leaving and there is another occurrence of $x$ at the bottom node of the structure. We have therefore shown that this is not a proof net.","Since there are no structural rewrites, the contractions for first-order linear logic are easier to apply than those for multimodal type-logical grammars: it is rather easy to show confluence for the contractions (the presence of structural rules, but also the unary versions of the multimodal contractions, means confluence is not guaranteed for multimodal proof nets). We already implicitly used confluence when we argued that the proof structure in Figure 14 was not a proof net. The theorem prover uses a maximally contracted representation of the proof structure to represent the current state of proof search and this means less overhead and more opportunities for early failure during proof search.","Like before, the theorem proving uses four steps, which look as follows in the first-order case: 1) lexical lookup, 2) unfolding, 3) axiom links with unification, 4) graph contraction. Unlike the multimodal proof nets of the previous section, the graph contractions are now confluent and can be performed efficiently (the linear time solutions for the multiplicative case may be adaptable, but a naive implementation already has an $O(n^2)$ worst-case performance). After lexical lookup, theorem proving for first-order linear logic unfolds the formulas as before, but uses a greedy contraction strategy. This maximally contracted partial proof net constrains further axiom links: for example, a vertex containing a free variable $x$ cannot be linked to the conclusion of the edge of its eigenvariable (the vertex to which the arrow of the edge with variable $x$ points) or to one of its descendants, since such a structure would fail to satisfy the condition that the two vertices of a $\\forall $ link for the $u$ contraction of Figure 8 are distinct. Another easily verified constraint is that two atomic formulas can only be connected by an axiom link if these formulas unify. Like for multimodal proof nets, the first-order linear logic theorem prover chooses an axiom link for one of the atoms with the fewest possibilities."],"section_name":"First-order proof nets"},{"paragraphs":["Table 9 lists the different theorem provers which are available. Grail 0 BIBREF42 and Grail 3 BIBREF43 use the multimodal proof net calculus of Section \"Multimodal proof nets\" , whereas LinearOne BIBREF44 uses the first-order proof nets of Section \"First-order proof nets\" . GrailLight BIBREF45 is a special-purpose chart parser, intended for use with an automatically extracted French grammar for wide-coverage parsing and semantics BIBREF34 , BIBREF46 . All provers are provided under the GNU Lesser General Public License — this means, notably, there is no warranty, though I am committed to making all software as useful as possible; so contact me for any comments, feature requests or bug reports. All theorem provers can be downloaded from the author's GitHub site.","https://github.com/RichardMoot/","The columns of table Table 9 indicate whether the theorem provers provide natural deduction output, graph output (of the partial proof nets), whether there is an interactive mode for proof search, whether the implementation is complete and whether the grammar can specify its own set of structural rules; “NA” means the question doesn't apply to the given system (GrailLight doesn't use a graphs to represent proofs and first-order linear logic does not have a grammar-specific set of structural rules). The table should help you select the most adequate tool for your purposes.","LinearOne provides natural deduction output not only for first-order linear logic, but also for the Displacement calculus, hybrid type-logical grammars and lambda grammars. That is, the grammar writer can write a grammar in any of these formalisms, LinearOne will do proof search of the translation of this grammar in first-order linear logic and then translate any resulting proofs back to the source language.","The syntactic example proofs in this chapter have been automatically generated using these tools and the corresponding grammars files, as well as many other example grammars, are included in the repository."],"section_name":"Tools"}],"paper_id":"1602.00812","title":"The Grail theorem prover: Type theory for syntax and semantics"},{"abstract":"In this paper, we propose a novel unsupervised learning method for the lexical acquisition of words related to places visited by robots, from human continuous speech signals. We address the problem of learning novel words by a robot that has no prior knowledge of these words except for a primitive acoustic model. Further, we propose a method that allows a robot to effectively use the learned words and their meanings for self-localization tasks. The proposed method is nonparametric Bayesian spatial concept acquisition method (SpCoA) that integrates the generative model for self-localization and the unsupervised word segmentation in uttered sentences via latent variables related to the spatial concept. We implemented the proposed method SpCoA on SIGVerse, which is a simulation environment, and TurtleBot2, which is a mobile robot in a real environment. Further, we conducted experiments for evaluating the performance of SpCoA. The experimental results showed that SpCoA enabled the robot to acquire the names of places from speech sentences. They also revealed that the robot could effectively utilize the acquired spatial concepts and reduce the uncertainty in self-localization.","figures_and_tables":[{"caption":"Fig. 1. Schematic representation of the target task: (a) Learning targets are three places near the objects. (b) When an utterer and a robot came in front of the TV, the utterer spoke “Here is the front of TV.” The same holds true for the other places. (c) The robot performs word discovery from the uttered sentences and learns the related spatial concepts. Words are related to each place. (d) The robot is in front of TV actually. However, the hypothesis of the self-position of the robot is uncertain. Then, the robot asks a neighbor what the current place is. (e) By utilizing spatial concepts and an uttered sentence, the robot can narrow down the hypothesis of self-position.","file":"1-Figure1-1.png"},{"caption":"Fig. 2. Graphical model of the proposed method SpCoA","file":"4-Figure2-1.png"},{"caption":"TABLE I EACH ELEMENT OF THE GRAPHICAL MODEL","file":"4-TableI-1.png"},{"caption":"Fig. 3. Environment to be used for learning and localization on SIGVerse: This is a pseudo-room in the simulated real world. There is a robot in the center of the room. The size of the room is 500 cm × 1,000 cm, and the size of the robot is 50 cm × 50 cm.","file":"7-Figure3-1.png"},{"caption":"Fig. 4. Learning result of the position distribution: A point group of each color to represent each position distribution is drawn on an map of the considered environment. The colors of the point groups are determined randomly. Each balloon shows the index number for each position distribution.","file":"7-Figure4-1.png"},{"caption":"TABLE II VARIOUS PHRASES OF EACH JAPANESE SENTENCE: “**” IS USED AS A PLACEHOLDER FOR THE NAME OF EACH PLACE. EXAMPLES OF THESE PHRASES ARE “** is here.”, “This place is **.”, “This place’s name is **.”, AND “Came to **.” IN ENGLISH.","file":"7-TableII-1.png"},{"caption":"Fig. 5. Learning result of the multinomial distributions of the names of places W (top); multinomial distributions of the index of the position distribution φl (bottom): All the words obtained during the experiment are shown.","file":"7-Figure5-1.png"},{"caption":"TABLE III COMPARISON OF THE PHONEME ACCURACY RATES OF UTTERED SENTENCES FOR DIFFERENT WORD SEGMENTATION METHODS","file":"8-TableIII-1.png"},{"caption":"TABLE IV EXAMPLES OF WORD SEGMENTATION RESULTS OF UTTERED SENTENCES. “|” DENOTES A WORD SEGMENT POINT.","file":"9-TableIV-1.png"},{"caption":"Fig. 6. Comparison of the accuracy rates of the estimation results of spatial concepts","file":"9-Figure6-1.png"},{"caption":"Fig. 7. PAR scores for the word considered the name of a place","file":"9-Figure7-1.png"},{"caption":"Fig. 10. Teaching places and the names of places shown on the generated map. The teaching places included places having two names each and multiple places having the same names.","file":"10-Figure10-1.png"},{"caption":"Fig. 8. Results of estimation errors and EARs of self-localization","file":"10-Figure8-1.png"},{"caption":"Fig. 9. Autonomous mobile robot TurtleBot 2: The robot is based on Yujin Robot Kobuki and Microsoft Kinect for its use as a range sensor.","file":"10-Figure9-1.png"},{"caption":"Fig. 11. Learning result of each position distribution: A point group of each color denoting each position distribution was drawn on the map. The colors of the point groups were determined randomly. Further, each index number is denoted as it = k.","file":"11-Figure11-1.png"},{"caption":"TABLE V LEARNING RESULT OF HIGH-PROBABILITY WORDS AND INDICES OF THE POSITION DISTRIBUTION FOR EACH SPATIAL CONCEPT","file":"11-TableV-1.png"},{"caption":"Fig. 12. States of particles: before the teaching utterance (top); after the teaching utterance (bottom). The uttered sentence is “kokowa ** dayo,” (which means “Here is **.”) “**” is the name of the place.","file":"13-Figure12-1.png"}],"full_text":[{"paragraphs":["Autonomous robots, such as service robots, operating in the human living environment with humans have to be able to perform various tasks and language communication. To this end, robots are required to acquire novel concepts and vocabulary on the basis of the information obtained from their sensors, e.g., laser sensors, microphones, and cameras, and recognize a variety of objects, places, and situations in an ambient environment. Above all, we consider it important for the robot to learn the names that humans associate with places in the environment and the spatial areas corresponding to these names; i.e., the robot has to be able to understand words related to places. Therefore, it is important to deal with considerable uncertainty, such as the robot's movement errors, sensor noise, and speech recognition errors.","Several studies on language acquisition by robots have assumed that robots have no prior lexical knowledge. These studies differ from speech recognition studies based on a large vocabulary and natural language processing studies based on lexical, syntactic, and semantic knowledge BIBREF0 , BIBREF1 . Studies on language acquisition by robots also constitute a constructive approach to the human developmental process and the emergence of symbols.","The objectives of this study were to build a robot that learns words related to places and efficiently utilizes this learned vocabulary in self-localization. Lexical acquisition related to places is expected to enable a robot to improve its spatial cognition. A schematic representation depicting the target task of this study is shown in Fig. FIGREF3 . This study assumes that a robot does not have any vocabularies in advance but can recognize syllables or phonemes. The robot then performs self-localization while moving around in the environment, as shown in Fig. FIGREF3 (a). An utterer speaks a sentence including the name of the place to the robot, as shown in Fig. FIGREF3 (b). For the purposes of this study, we need to consider the problems of self-localization and lexical acquisition simultaneously.","When a robot learns novel words from utterances, it is difficult to determine segmentation boundaries and the identity of different phoneme sequences from the speech recognition results, which can lead to errors. First, let us consider the case of the lexical acquisition of an isolated word. For example, if a robot obtains the speech recognition results “aporu”, “epou”, and “aqpuru” (incorrect phoneme recognition of apple), it is difficult for the robot to determine whether they denote the same referent without prior knowledge. Second, let us consider a case of the lexical acquisition of the utterance of a sentence. For example, a robot obtains a speech recognition result, such as “thisizanaporu.” The robot has to necessarily segment a sentence into individual words, e.g., “this”, “iz”, “an”, and “aporu”. In addition, it is necessary for the robot to recognize words referring to the same referent, e.g., the fruit apple, from among the many segmented results that contain errors. In case of Fig. FIGREF3 (c), there is some possibility of learning names including phoneme errors, e.g., “afroqtabutibe,” because the robot does not have any lexical knowledge. On the other hand, when a robot performs online probabilistic self-localization, we assume that the robot uses sensor data and control data, e.g., values obtained using a range sensor and odometry. If the position of the robot on the global map is unclear, the difficulties associated with the identification of the self-position by only using local sensor information become problematic. In the case of global localization using local information, e.g., a range sensor, the problem that the hypothesis of self-position is present in multiple remote locations, frequently occurs, as shown in Fig. FIGREF3 (d).","In order to solve the abovementioned problems, in this study, we adopted the following approach. An utterance is recognized as not a single phoneme sequence but a set of candidates of multiple phonemes. We attempt to suppress the variability in the speech recognition results by performing word discovery taking into account the multiple candidates of speech recognition. In addition, the names of places are learned by associating with words and positions. The lexical acquisition is complemented by using certain particular spatial information; i.e., this information is obtained by hearing utterances including the same word in the same place many times. Furthermore, in this study, we attempt to address the problem of the uncertainty of self-localization by improving the self-position errors by using a recognized utterance including the name of the current place and the acquired spatial concepts, as shown in Fig. FIGREF3 (e).","In this paper, we propose nonparametric Bayesian spatial concept acquisition method (SpCoA) on basis of unsupervised word segmentation and a nonparametric Bayesian generative model that integrates self-localization and a clustering in both words and places. The main contributions of this paper are as follows:","The remainder of this paper is organized as follows: In Section SECREF2 , previous studies on language acquisition and lexical acquisition relevant to our study are described. In Section SECREF3 , the proposed method SpCoA is presented. In Sections SECREF4 and SECREF5 , we discuss the effectiveness of SpCoA in the simulation and in the real environment. Section SECREF6 concludes this paper."],"section_name":"Introduction"},{"paragraphs":["Most studies on lexical acquisition typically focus on lexicons about objects BIBREF0 , BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 . Many of these studies have not be able to address the lexical acquisition of words other than those related to objects, e.g., words about places.","Roy et al. proposed a computational model that enables a robot to learn the names of objects from an object image and spontaneous infant-directed speech BIBREF0 . Their results showed that the model performed speech segmentation, word discovery, and visual categorization. Iwahashi et al. reported that a robot properly understands the situation and acquires the relationship of object behaviors and sentences BIBREF2 , BIBREF3 , BIBREF4 . Qu & Chai focused on the conjunction between speech and eye gaze and the use of domain knowledge in lexical acquisition BIBREF6 , BIBREF7 . They proposed an unsupervised learning method that automatically acquires novel words for an interactive system. Qu & Chai's method based on the IBM translation model BIBREF11 estimates the word-entity association probability.","Nakamura et al. proposed a method to learn object concepts and word meanings from multimodal information and verbal information BIBREF9 . The method proposed in BIBREF9 is a categorization method based on multimodal latent Dirichlet allocation (MLDA) that enables the acquisition of object concepts from multimodal information, such as visual, auditory, and haptic information BIBREF12 . Araki et al. addressed the development of a method combining unsupervised word segmentation from uttered sentences by a nested Pitman-Yor language model (NPYLM) BIBREF13 and the learning of object concepts by MLDA BIBREF10 . However, the disadvantage of using NPYLM was that phoneme sequences with errors did not result in appropriate word segmentation.","These studies did not address the lexical acquisition of the space and place that can also tolerate the uncertainty of phoneme recognition. However, for the introduction of robots into the human living environment, robots need to acquire a lexicon related to not only objects but also places. Our study focuses on the lexical acquisition related to places. Robots can adaptively learn the names of places in various human living environments by using SpCoA. We consider that the acquired names of places can be useful for various tasks, e.g., tasks with a movement of robots by the speech instruction."],"section_name":"Lexical acquisition"},{"paragraphs":["The following studies have addressed lexical acquisition related to places. However, these studies could not utilize the learned language knowledge in other estimations such as the self-localization of a robot.","Taguchi et al. proposed a method for the unsupervised learning of phoneme sequences and relationships between words and objects from various user utterances without any prior linguistic knowledge other than an acoustic model of phonemes BIBREF1 , BIBREF14 . Further, they proposed a method for the simultaneous categorization of self-position coordinates and lexical learning BIBREF15 . These experimental results showed that it was possible to learn the name of a place from utterances in some cases and to output words corresponding to places in a location that was not used for learning.","Milford et al. proposed RatSLAM inspired by the biological knowledge of a pose cell of the hippocampus of rodents BIBREF16 . Milford et al. proposed a method that enables a robot to acquire spatial concepts by using RatSLAM BIBREF17 . Further, Lingodroids, mobile robots that learn a language through robot-to-robot communication, have been studied BIBREF18 , BIBREF19 , BIBREF20 . Here, a robot communicated the name of a place to other robots at various locations. Experimental results showed that two robots acquired the lexicon of places that they had in common. In BIBREF20 , the researchers showed that it was possible to learn temporal concepts in a manner analogous to the acquisition of spatial concepts. These studies reported that the robots created their own vocabulary. However, these studies did not consider the acquisition of a lexicon by human-to-robot speech interactions.","Welke et al. proposed a method that acquires spatial representation by the integration of the representation of the continuous state space on the sensorimotor level and the discrete symbolic entities used in high-level reasoning BIBREF21 . This method estimates the probable spatial domain and word from the given objects by using the spatial lexical knowledge extracted from Google Corpus and the position information of the object. Their study is different from ours because their study did not consider lexicon learning from human speech.","In the case of global localization, the hypothesis of self-position often remains in multiple remote places. In this case, there is some possibility of performing an incorrect estimation and increasing the estimation error. This problem exists during teaching tasks and self-localization after the lexical acquisition. The abovementioned studies could not deal with this problem. In this paper, we have proposed a method that enables a robot to perform more accurate self-localization by reducing the estimation error of the teaching time by using a smoothing method in the teaching task and by utilizing words acquired through the lexical acquisition. The strengths of this study are that learning of spatial concept and self-localization represented as one generative model and robots are able to utilize acquired lexicon to self-localization autonomously."],"section_name":"Simultaneous learning of places and vocabulary"},{"paragraphs":["We propose nonparametric Bayesian spatial concept acquisition method (SpCoA) that integrates a nonparametric morphological analyzer for the lattice BIBREF22 , i.e., latticelm, a spatial clustering method, and Monte Carlo localization (MCL) BIBREF23 ."],"section_name":"Spatial Concept Acquisition"},{"paragraphs":["In our study, we define a position as a specific coordinate or a local point in the environment, and the position distribution as the spatial area of the environment. Further, we define a spatial concept as the names of places and the position distributions corresponding to these names.","The model that was developed for spatial concept acquisition is a probabilistic generative model that integrates a self-localization with the simultaneous clustering of places and words. Fig. FIGREF13 shows the graphical model for spatial concept acquisition. Table TABREF14 shows each variable of the graphical model. The number of words in a sentence at time INLINEFORM0 is denoted as INLINEFORM1 . The generative model of the proposed method is defined as equation ( EQREF11 -). DISPLAYFORM0 ","Then, the probability distribution for equation () can be defined as follows: DISPLAYFORM0 ","The prior distribution configured by using the stick breaking process (SBP) BIBREF24 is denoted as INLINEFORM0 , the multinomial distribution as INLINEFORM1 , the Dirichlet distribution as INLINEFORM2 , the inverse–Wishart distribution as INLINEFORM3 , and the multivariate Gaussian (normal) distribution as INLINEFORM4 . The motion model and the sensor model of self-localization are denoted as INLINEFORM5 and INLINEFORM6 in equations () and (), respectively.","This model can learn an appropriate number of spatial concepts, depending on the data, by using a nonparametric Bayesian approach. We use the SBP, which is one of the methods based on the Dirichlet process. In particular, this model can consider a theoretically infinite number of spatial concepts INLINEFORM0 and position distributions INLINEFORM1 . SBP computations are difficult because they generate an infinite number of parameters. In this study, we approximate a number of parameters by setting sufficiently large values, i.e., a weak-limit approximation BIBREF25 .","It is possible to correlate a name with multiple places, e.g., “staircase” is in two different places, and a place with multiple names, e.g., “toilet” and “restroom” refer to the same place. Spatial concepts are represented by a word distribution of the names of the place INLINEFORM0 and several position distributions ( INLINEFORM1 , INLINEFORM2 ) indicated by a multinomial distribution INLINEFORM3 . In other words, this model is capable of relating the mixture of Gaussian distributions to a multinomial distribution of the names of places. It should be noted that the arrows connecting INLINEFORM4 to the surrounding nodes of the proposed graphical model differ from those of ordinal Gaussian mixture model (GMM). We assume that words obtained by the robot do not change its position, but that the position of the robot affects the distribution of words. Therefore, the proposed generative process assumes that the index of position distribution INLINEFORM5 , i.e., the category of the place, is generated from the position of the robot INLINEFORM6 . This change can be naturally introduced without any troubles by introducing equation ( EQREF12 )."],"section_name":"Generative model"},{"paragraphs":["We assume that a robot performs self-localization by using control data and sensor data at all times. The procedure for the learning of spatial concepts is as follows:","An utterer teaches a robot the names of places, as shown in Fig. FIGREF3 (b). Every time the robot arrives at a place that was a designated learning target, the utterer says a sentence, including the name of the current place.","The robot performs speech recognition from the uttered speech signal data. Thus, the speech recognition system includes a word dictionary of only Japanese syllables. The speech recognition results are obtained in a lattice format.","Word segmentation is performed by using the lattices of the speech recognition results.","The robot learns spatial concepts from words obtained by word segmentation and robot positions obtained by self-localization for all teaching times. The details of the learning are given in SECREF23 .","The procedure for self-localization utilizing spatial concepts is as follows:","The words of the learned spatial concepts are registered to the word dictionary of the speech recognition system.","When a robot obtains a speech signal, speech recognition is performed. Then, a word sequence as the 1-best speech recognition result is obtained.","The robot modifies the self-localization from words obtained by speech recognition and the position likelihood obtained by spatial concepts. The details of self-localization are provided in SECREF35 .","The proposed method can learn words related to places from the utterances of sentences. We use an unsupervised word segmentation method latticelm that can directly segment words from the lattices of the speech recognition results of the uttered sentences BIBREF22 . The lattice can represent to a compact the set of more promising hypotheses of a speech recognition result, such as N-best, in a directed graph format. Unsupervised word segmentation using the lattices of syllable recognition is expected to be able to reduce the variability and errors in phonemes as compared to NPYLM BIBREF13 , i.e., word segmentation using the 1-best speech recognition results.","The self-localization method adopts MCL BIBREF23 , a method that is generally used as the localization of mobile robots for simultaneous localization and mapping (SLAM) BIBREF26 . We assume that a robot generates an environment map by using MCL-based SLAM such as FastSLAM BIBREF27 , BIBREF28 in advance, and then, performs localization by using the generated map. Then, the environment map of both an occupancy grid map and a landmark map is acceptable."],"section_name":"Overview of the proposed method SpCoA"},{"paragraphs":["Spatial concepts are learned from multiple teaching data, control data, and sensor data. The teaching data are a set of uttered sentences for all teaching times. Segmented words of an uttered sentence are converted into a bag-of-words (BoW) representation as a vector of the occurrence counts of words INLINEFORM0 . The set of the teaching times is denoted as INLINEFORM1 , and the number of teaching data items is denoted as INLINEFORM2 . The model parameters are denoted as INLINEFORM3 . The initial values of the model parameters can be set arbitrarily in accordance with a condition. Further, the sampling values of the model parameters from the following joint posterior distribution are obtained by performing Gibbs sampling. DISPLAYFORM0 ","where the hyperparameters of the model are denoted as INLINEFORM0 . The algorithm of the learning of spatial concepts is shown in Algorithm SECREF23 .","The conditional posterior distribution of each element used for performing Gibbs sampling can be expressed as follows: An index INLINEFORM0 of the position distribution is sampled for each data INLINEFORM1 from a posterior distribution as follows: DISPLAYFORM0 ","An index INLINEFORM0 of the spatial concepts is sampled for each data item INLINEFORM1 from a posterior distribution as follows: DISPLAYFORM0 ","where INLINEFORM0 denotes a vector of the occurrence counts of words in the sentence at time INLINEFORM1 . A posterior distribution representing word probabilities of the name of place INLINEFORM2 is calculated as follows: DISPLAYFORM0 ","where variables with the subscript INLINEFORM0 denote the set of all teaching times. A word probability of the name of place INLINEFORM1 is sampled for each INLINEFORM2 as follows: DISPLAYFORM0 ","where INLINEFORM0 represents the posterior parameter and INLINEFORM1 denotes the BoW representation of all sentences of INLINEFORM2 in INLINEFORM3 . A posterior distribution representing the position distribution INLINEFORM4 is calculated as follows: DISPLAYFORM0 ","A position distribution INLINEFORM0 , INLINEFORM1 is sampled for each INLINEFORM2 as follows: DISPLAYFORM0 ","where INLINEFORM0 denotes the Gaussian–inverse–Wishart distribution; INLINEFORM1 , and INLINEFORM2 represent the posterior parameters; and INLINEFORM3 indicates the set of the teaching positions of INLINEFORM4 in INLINEFORM5 . A topic probability distribution INLINEFORM6 of spatial concepts is sampled as follows: DISPLAYFORM0 ","A posterior distribution representing the mixed weights INLINEFORM0 of the position distributions is calculated as follows: DISPLAYFORM0 ","A mixed weight INLINEFORM0 of the position distributions is sampled for each INLINEFORM1 as follows: DISPLAYFORM0 ","where INLINEFORM0 denotes a vector counting all the indices of the Gaussian distribution of INLINEFORM1 in INLINEFORM2 .","Self-positions INLINEFORM0 are sampled by using a Monte Carlo fixed-lag smoother BIBREF29 in the learning phase. The smoother can estimate self-position INLINEFORM1 and not INLINEFORM2 , i.e., a sequential estimation from the given data INLINEFORM3 until time INLINEFORM4 , but it can estimate INLINEFORM5 , i.e., an estimation from the given data INLINEFORM6 until time INLINEFORM7 later than INLINEFORM8 INLINEFORM9 . In general, the smoothing method can provide a more accurate estimation than the MCL of online estimation. In contrast, if the self-position of a robot INLINEFORM10 is sampled like direct assignment sampling for each time INLINEFORM11 , the sampling of INLINEFORM12 is divided in the case with the teaching time INLINEFORM13 and another time INLINEFORM14 as follows: DISPLAYFORM0 ","[tb] Learning of spatial concepts [1] INLINEFORM0 , INLINEFORM1 Localization and speech recognition INLINEFORM2 to INLINEFORM3 INLINEFORM4 BIBREF29 the speech signal is observed INLINEFORM5 add INLINEFORM6 to INLINEFORM7 Registering the lattice add INLINEFORM8 to INLINEFORM9 Registering the teaching time Word segmentation using lattices INLINEFORM10 BIBREF22 Gibbs sampling Initialize parameters INLINEFORM11 , INLINEFORM12 , INLINEFORM13 INLINEFORM14 to INLINEFORM15 INLINEFORM16 ( EQREF25 ) INLINEFORM17 ( EQREF26 ) INLINEFORM18 ( EQREF28 ) INLINEFORM19 ( EQREF30 ) INLINEFORM20 ( EQREF31 ) INLINEFORM21 ( EQREF33 ) INLINEFORM22 to INLINEFORM23 INLINEFORM24 ( EQREF34 ) INLINEFORM25 "],"section_name":"Learning of spatial concept"},{"paragraphs":["A robot that acquires spatial concepts can leverage spatial concepts to self-localization. The estimated model parameters INLINEFORM0 and a speech recognition sentence INLINEFORM1 at time INLINEFORM2 are given to the condition part of the probability formula of MCL as follows: DISPLAYFORM0 ","When the robot hears the name of a place spoken by the utterer, in addition to the likelihood of the sensor model of MCL, the likelihood of INLINEFORM0 with respect to a speech recognition sentence is calculated as follows: DISPLAYFORM0 ","The algorithm of self-localization utilizing spatial concepts is shown in Algorithm SECREF35 . The set of particles is denoted as INLINEFORM0 , the temporary set that stores the pairs of the particle INLINEFORM1 and the weight INLINEFORM2 , i.e., INLINEFORM3 , is denoted as INLINEFORM4 . The number of particles is INLINEFORM5 . The function INLINEFORM6 is a function that moves each particle from its previous state INLINEFORM7 to its current state INLINEFORM8 by using control data. The function INLINEFORM9 calculates the likelihood of each particle INLINEFORM10 using sensor data INLINEFORM11 . These functions are normally used in MCL. For further details, please refer to BIBREF26 . In this case, a speech recognition sentence INLINEFORM12 is obtained by the speech recognition system using a word dictionary containing all the learned words. [tb] Self-localization utilizing spatial concepts [1] INLINEFORM13 INLINEFORM14 INLINEFORM15 INLINEFORM16 to INLINEFORM17 INLINEFORM18 () INLINEFORM19 () the speech signal is observed INLINEFORM20 add INLINEFORM21 to INLINEFORM22 INLINEFORM23 to INLINEFORM24 draw INLINEFORM25 with probability INLINEFORM26 add INLINEFORM27 to INLINEFORM28 INLINEFORM29 "],"section_name":"Self-localization of after learning spatial concepts"},{"paragraphs":["In this experiment, we validate the evidence of the proposed method (SpCoA) in an environment simulated on the simulator platform SIGVerse BIBREF30 , which enables the simulation of social interactions. The speech recognition is performed using the Japanese continuous speech recognition system Julius BIBREF31 , BIBREF32 . The set of 43 Japanese phonemes defined by Acoustical Society of Japan (ASJ)'s speech database committee is adopted by Julius BIBREF31 . The representation of these phonemes is also adopted in this study. The Julius system uses a word dictionary containing 115 Japanese syllables. The microphone attached on the robot is SHURE's PG27-USB. Further, an unsupervised morphological analyzer, a latticelm 0.4, is implemented BIBREF22 .","In the experiment, we compare the following three types of word segmentation methods. A set of syllable sequences is given to the graphical model of SpCoA by each method. This set is used for the learning of spatial concepts as recognized uttered sentences INLINEFORM0 .","The remainder of this section is organized as follows: In Section SECREF43 , the conditions and results of learning spatial concepts are described. The experiments performed using the learned spatial concepts are described in Section SECREF49 to SECREF64 . In Section SECREF49 , we evaluate the accuracy of the phoneme recognition and word segmentation for uttered sentences. In Section SECREF56 , we evaluate the clustering accuracy of the estimation results of index INLINEFORM0 of spatial concepts for each teaching utterance. In Section SECREF60 , we evaluate the accuracy of the acquisition of names of places. In Section SECREF64 , we show that spatial concepts can be utilized for effective self-localization."],"section_name":"Experiment I"},{"paragraphs":["We conduct this experiment of spatial concept acquisition in the environment prepared on SIGVerse. The experimental environment is shown in Fig. FIGREF45 . A mobile robot can move by performing forward, backward, right rotation, or left rotation movements on a two-dimensional plane. In this experiment, the robot can use an approximately correct map of the considered environment. The robot has a range sensor in front and performs self-localization on the basis of an occupancy grid map. The initial particles are defined by the true initial position of the robot. The number of particles is INLINEFORM0 .","The lag value of the Monte Carlo fixed-lag smoothing is fixed at 100. The other parameters of this experiment are as follows: INLINEFORM0 , INLINEFORM1 , INLINEFORM2 , INLINEFORM3 , INLINEFORM4 , INLINEFORM5 , INLINEFORM6 , INLINEFORM7 , and INLINEFORM8 . The number of iterations used for Gibbs sampling is 100. This experiment does not include the direct assignment sampling of INLINEFORM9 in equation ( EQREF34 ), i.e., lines 22–24 of Algorithm SECREF23 are omitted, because we consider that the self-position can be obtained with sufficiently good accuracy by using the Monte Carlo smoothing. Eight places are selected as the learning targets, and eight types of place names are considered. Each uttered place name is shown in Fig. FIGREF45 . These utterances include the same name in different places, i.e., “teeburunoatari” (which means near the table in English), and different names in the same place, i.e., “kiqchiN” and “daidokoro” (which mean a kitchen in English). The other teaching names are “geNkaN” (which means an entrance or a doorway in English); “terebimae” (which means the front of the TV in English); “gomibako” (which means a trash box in English); “hoNdana” (which means a bookshelf in English); and “sofaamae” (which means the front of the sofa in English). The teaching utterances, including the 10 types of phrases, are spoken for a total of 90 times. The phrases in each uttered sentence are listed in Table TABREF46 .","The learning results of spatial concepts obtained by using the proposed method are presented here. Fig. FIGREF47 shows the position distributions learned in the experimental environment. Fig. FIGREF47 (top) shows the word distributions of the names of places for each spatial concept, and Fig. FIGREF47 (bottom) shows the multinomial distributions of the indices of the position distributions. Consequently, the proposed method can learn the names of places corresponding to each place of the learning target. In the spatial concept of index INLINEFORM0 , the highest probability of words was “sofamae”, and the highest probability of the indices of the position distribution was INLINEFORM1 ; therefore, the name of a place “sofamae” was learned to correspond to the position distribution of INLINEFORM2 . In the spatial concept of index INLINEFORM3 , “kiqchi” and “daidokoro” were learned to correspond to the position distribution of INLINEFORM4 . Therefore, this result shows that multiple names can be learned for the same place. In the spatial concept of index INLINEFORM5 , “te” and “durunoatari” (one word in a normal situation) were learned to correspond to the position distributions of INLINEFORM6 and INLINEFORM7 . Therefore, this result shows that the same name can be learned for multiple places."],"section_name":"Learning of spatial concepts"},{"paragraphs":["We compared the performance of three types of word segmentation methods for all the considered uttered sentences. It was difficult to weigh the ambiguous syllable recognition and the unsupervised word segmentation separately. Therefore, this experiment considered the positions of a delimiter as a single letter. We calculated the matching rate of a phoneme string of a recognition result of each uttered sentence and the correct phoneme string of the teaching data that was suitably segmented into Japanese morphemes using MeCab, which is an off-the-shelf Japanese morphological analyzer that is widely used for natural language processing. The matching rate of the phoneme string was calculated by using the phoneme accuracy rate (PAR) as follows: DISPLAYFORM0 ","The numerator of equation ( EQREF52 ) is calculated by using the Levenshtein distance between the correct phoneme string and the recognition phoneme string. INLINEFORM0 denotes the number of substitutions; INLINEFORM1 , the number of deletions; and INLINEFORM2 , the number of insertions. INLINEFORM3 represents the number of phonemes of the correct phoneme string.","Table TABREF54 shows the results of PAR. Table TABREF55 presents examples of the word segmentation results of the three considered methods. We found that the unsupervised morphological analyzer capable of using lattices improved the accuracy of phoneme recognition and word segmentation. Consequently, this result suggests that this word segmentation method considers the multiple hypothesis of speech recognition as a whole and reduces uncertainty such as variability in recognition by using the syllable recognition results in the lattice format."],"section_name":"Phoneme recognition accuracy of uttered sentences"},{"paragraphs":["We compared the matching rate with the estimation results of index INLINEFORM0 of the spatial concepts of each teaching utterance and the classification results of the correct answer given by humans. The evaluation of this experiment used the adjusted Rand index (ARI) BIBREF33 . ARI is a measure of the degree of similarity between two clustering results.","Further, we compared the proposed method with a method of word clustering without location information for the investigation of the effect of lexical acquisition using location information. In particular, a method of word clustering without location information used the Dirichlet process mixture (DPM) of the unigram model of an SBP representation. The parameters corresponding to those of the proposed method were the same as the parameters of the proposed method and were estimated using Gibbs sampling.","Fig. FIGREF59 shows the results of the average of the ARI values of 10 trials of learning by Gibbs sampling. Here, we found that the proposed method showed the best score. These results and the results reported in Section SECREF49 suggest that learning by uttered sentences obtained by better phoneme recognition and better word segmentation produces a good result for the acquisition of spatial concepts. Furthermore, in a comparison of two clustering methods, we found that SpCoA was considerably better than DPM, a word clustering method without location information, irrespective of the word segmentation method used. The experimental results showed that it is possible to improve the estimation accuracy of spatial concepts and vocabulary by performing word clustering that considered location information."],"section_name":"Estimation accuracy of spatial concepts"},{"paragraphs":["We evaluated whether the names of places were properly learned for the considered teaching places. This experiment assumes a request for the best phoneme sequence INLINEFORM0 representing the self-position INLINEFORM1 for a robot. The robot moves close to each teaching place. The probability of a word INLINEFORM2 when the self-position INLINEFORM3 of the robot is given, INLINEFORM4 , can be obtained by using equation ( EQREF37 ). The word having the best probability was selected. We compared the PAR with the correct phoneme sequence and a selected name of the place. Because “kiqchiN” and “daidokoro” were taught for the same place, the word whose PAR was the higher score was adopted.","Fig. FIGREF63 shows the results of PAR for the word considered the name of a place. SpCoA (latticelm), the proposed method using the results of unsupervised word segmentation on the basis of the speech recognition results in the lattice format, showed the best PAR score. In the 1-best and BoS methods, a part syllable sequence of the name of a place was more minutely segmented as shown in Table TABREF55 . Therefore, the robot could not learn the name of the teaching place as a coherent phoneme sequence. In contrast, the robot could learn the names of teaching places more accurately by using the proposed method."],"section_name":"Accuracy of acquired phoneme sequences representing the names of places"},{"paragraphs":["In this experiment, we validate that the robot can make efficient use of the acquired spatial concepts. We compare the estimation accuracy of localization for the proposed method (SpCoA MCL) and the conventional MCL. When a robot comes to the learning target, the utterer speaks out the sentence containing the name of the place once again for the robot. The moving trajectory of the robot and the uttered positions are the same in all the trials. In particular, the uttered sentence is “kokowa ** dayo”. When learning a task, this phrase is not used. The number of particles is INLINEFORM0 , and the initial particles are uniformly distributed in the considered environment. The robot performs a control operation for each time step.","The estimation error in the localization is evaluated as follows: While running localization, we record the estimation error (equation ( EQREF66 )) on the INLINEFORM0 plane of the floor for each time step. DISPLAYFORM0 ","where INLINEFORM0 denote the true position coordinates of the robot as obtained from the simulator, and INLINEFORM1 , INLINEFORM2 represent the weighted mean values of localization coordinates. The normalized weight INLINEFORM3 is obtained from the sensor model in MCL as a likelihood. In the utterance time, this likelihood is multiplied by the value calculated using equation ( EQREF37 ). INLINEFORM4 , INLINEFORM5 denote the INLINEFORM6 -coordinate and the INLINEFORM7 -coordinate of index INLINEFORM8 of each particle at time INLINEFORM9 . After running the localization, we calculated the average of INLINEFORM10 .","Further, we compared the estimation accuracy rate (EAR) of the global localization. In each trial, we calculated the proportion of time step in which the estimation error was less than 50 cm.","Fig. FIGREF68 shows the results of the estimation error and the EAR for 10 trials of each method. All trials of SpCoA MCL (latticelm) and almost all trials of the method using 1-best NPYLM and BoS showed relatively small estimation errors. Results of the second trial of 1-best NPYLM and the fifth trial of BoS showed higher estimation errors. In these trials, many particles converged to other places instead of the place where the robot was, based on utterance information. Nevertheless, compared with those of the conventional MCL, the results obtained using spatial concepts showed an obvious improvement in the estimation accuracy. Consequently, spatial concepts acquired by using the proposed method proved to be very helpful in improving the localization accuracy."],"section_name":"Self-localization that utilizes acquired spatial concepts"},{"paragraphs":["In this experiment, the effectiveness of the proposed method was tested by using an autonomous mobile robot TurtleBot 2 in a real environment. Fig. FIGREF70 shows TurtleBot 2 used in the experiments.","Mapping and self-localization are performed by the robot operating system (ROS). The speech recognition system, the microphone, and the unsupervised morphological analyzer were the same as those described in Section SECREF4 ."],"section_name":"Experiment II"},{"paragraphs":["We conducted an experiment of the spatial concept acquisition in a real environment of an entire floor of a building. In this experiment, self-localization was performed using a map generated by SLAM. The initial particles are defined by the true initial position of the robot. The generated map in the real environment and the names of teaching places are shown in Fig. FIGREF73 . The number of teaching places was 19, and the number of teaching names was 16. The teaching utterances were performed for a total of 100 times.","Fig. FIGREF75 shows the position distributions learned on the map. Table TABREF76 shows the five best elements of the multinomial distributions of the name of place INLINEFORM0 and the multinomial distributions of the indices of the position distribution INLINEFORM1 for each index of spatial concept INLINEFORM2 .","Thus, we found that the proposed method can learn the names of places corresponding to the considered teaching places in the real environment. For example, in the spatial concept of index INLINEFORM0 , “torire” was learned to correspond to a position distribution of INLINEFORM1 . Similarly, “kidanokeN” corresponded to INLINEFORM2 in INLINEFORM3 , and “kaigihitsu” was corresponded to INLINEFORM4 in INLINEFORM5 . In the spatial concept of index INLINEFORM6 , a part of the syllable sequences was minutely segmented as “sohatsuke”, “N”, and “tani”, “guchi”. In this case, the robot was taught two types of names. These words were learned to correspond to the same position distribution of INLINEFORM7 . In INLINEFORM8 , “gomibako” showed a high probability, and it corresponded to three distributions of the position of INLINEFORM9 . The position distribution of INLINEFORM10 had the fourth highest probability in the spatial concept INLINEFORM11 . Therefore, “raqkukeN,” which had the fifth highest probability in the spatial concept INLINEFORM12 (and was expected to relate to the spatial concept INLINEFORM13 ), can be estimated as the word drawn from spatial concept INLINEFORM14 . However, in practice, this situation did not cause any severe problems because the spatial concept of the index INLINEFORM15 had the highest probabilities for the word “rapukeN” and the position distribution INLINEFORM16 than INLINEFORM17 . In the probabilistic model, the relative probability and the integrative information are important. When the robot listened to an utterance related to “raqkukeN,” it could make use of the spatial concept of index INLINEFORM18 for self-localization with a high probability, and appropriately updated its estimated self-location. We expected that the spatial concept of index INLINEFORM19 was learned as two separate spatial concepts. However, “watarirooka” and “kaidaNmae” were learned as the same spatial concept. Therefore, the multinomial distribution INLINEFORM20 showed a higher probability for the indices of the position distribution corresponding to the teaching places of both “watarirooka” and “kaidaNmae”.","The proposed method adopts a nonparametric Bayesian method in which it is possible to form spatial concepts that allow many-to-many correspondences between names and places. In contrast, this can create ambiguity that classifies originally different spatial concepts into one spatial concept as a side effect. There is a possibility that the ambiguity of concepts such as INLINEFORM0 will have a negative effect on self-localization, even though the self-localization performance was (overall) clearly increased by employing the proposed method. The solution of this problem will be considered in future work.","In terms of the PAR of uttered sentences, the evaluation value from the evaluation method used in Section SECREF49 is 0.83; this value is comparable to the result in Section SECREF49 . However, in terms of the PAR of the name of the place, the evaluation value from the evaluation method used in Section SECREF60 is 0.35, which is lower than that in Section SECREF60 . We consider that the increase in uncertainty in the real environment and the increase in the number of teaching words reduced the performance. We expect that this problem could be improved using further experience related to places, e.g., if the number of utterances per place is increased, and additional sensory information is provided."],"section_name":"Learning of spatial concepts in the real environment"},{"paragraphs":["In this experiment, we verified the modification results of self-localization by using spatial concepts in global self-localization. This experiment used the learning results of spatial concepts presented in Section SECREF71 . The experimental procedures are shown below. The initial particles were uniformly distributed on the entire floor. The robot begins to move from a little distance away to the target place. When the robot reached the target place, the utterer spoke the sentence containing the name of the place for the robot. Upon obtaining the speech information, the robot modifies the self-localization on the basis of the acquired spatial concepts. The number of particles was the same as that mentioned in Section SECREF71 .","Fig. FIGREF80 shows the results of the self-localization before (the top part of the figure) and after (the bottom part of the figure) the utterance for three places. The particle states are denoted by red arrows. The moving trajectory of the robot is indicated by a green dotted arrow. Figs. FIGREF80 (a), (b), and (c) show the results for the names of places “toire”, “souhatsukeN”, and “gomibako”. Further, three spatial concepts, i.e., those at INLINEFORM0 , were learned as “gomibako”. In this experiment, the utterer uttered to the robot when the robot came close to the place of INLINEFORM1 . In all the examples shown in the top part of the figure, the particles were dispersed in several places. In contrast, the number of particles near the true position of the robot showed an almost accurate increase in all the examples shown in the bottom part of the figure. Thus, we can conclude that the proposed method can modify self-localization by using spatial concepts."],"section_name":"Modification of localization by the acquired spatial concepts"},{"paragraphs":["In this paper, we discussed the spatial concept acquisition, lexical acquisition related to places, and self-localization using acquired spatial concepts. We proposed nonparametric Bayesian spatial concept acquisition method SpCoA that integrates latticelm BIBREF22 , a spatial clustering method, and MCL. We conducted experiments for evaluating the performance of SpCoA in a simulation and a real environment. SpCoA showed good results in all the experiments. In experiments of the learning of spatial concepts, the robot could form spatial concepts for the places of the learning targets from human continuous speech signals in both the room of the simulation environment and the entire floor of the real environment. Further, the unsupervised word segmentation method latticelm could reduce the variability and errors in the recognition of phonemes in all the utterances. SpCoA achieved more accurate lexical acquisition by performing word segmentation using the lattices of the speech recognition results. In the self-localization experiments, the robot could effectively utilize the acquired spatial concepts for recognizing self-position and reducing the estimation errors in self-localization. As a method that further improves the performance of the lexical acquisition, a mutual learning method was proposed by Nakamura et al. on the basis of the integration of the learning of object concepts with a language model BIBREF34 , BIBREF35 . Following a similar approach, Heymann et al. proposed a method that alternately and repeatedly updates phoneme recognition results and the language model by using unsupervised word segmentation BIBREF36 . As a result, they achieved robust lexical acquisition. In our study, we can expect to improve the accuracy of lexical acquisition for spatial concepts by estimating both the spatial concepts and the language model.","Furthermore, as a future work, we consider it necessary for robots to learn spatial concepts online and to recognize whether the uttered word indicates the current place or destination. Furthermore, developing a method that simultaneously acquires spatial concepts and builds a map is one of our future objectives. We believe that the spatial concepts will have a positive effect on the mapping. We also intend to examine a method that associates the image and the landscape with spatial concepts and a method that estimates both spatial concepts and object concepts.","[] Akira Taniguchi received his BE degree from Ritsumeikan University in 2013 and his ME degree from the Graduate School of Information Science and Engineering, Ritsumeikan University, in 2015. He is currently working toward his PhD degree at the Emergent System Lab, Ritsumeikan University, Japan. His research interests include language acquisition, concept acquisition, and symbol emergence in robotics.","[] Tadahiro Taniguchi received the ME and PhD degrees from Kyoto University in 2003 and 2006, respectively. From April 2005 to March 2006, he was a Japan Society for the Promotion of Science (JSPS) research fellow (DC2) in the Department of Mechanical Engineering and Science, Graduate School of Engineering, Kyoto University. From April 2006 to March 2007, he was a JSPS research fellow (PD) in the same department. From April 2007 to March 2008, he was a JSPS research fellow in the Department of Systems Science, Graduate School of Informatics, Kyoto University. From April 2008 to March 2010, he was an assistant professor at the Department of Human and Computer Intelligence, Ritsumeikan University. Since April 2010, he has been an associate professor in the same department. He is currently engaged in research on machine learning, emergent systems, and semiotics.","[] Tetsunari Inamura received the BE, MS and PhD degrees from the University of Tokyo, in 1995, 1997 and 2000, respectively. He was a Researcher of the CREST program, Japanese Science and Technology Cooperation, from 2000 to 2003, and then joined the Department of Mechano-Informatics, School of Information Science and Technology, University of Tokyo as a Lecturer, from 2003 to 2006. He is now an Associate Professor in the Principles of Informatics Research Division, National Institute of Informatics, and an Associate Professor in the Department of Informatics, School of Multidisciplinary Sciences, Graduate University for Advanced Studies (SOKENDAI). His research interests include imitation learning and symbol emergence on humanoid robots, development of interactive robots through virtual reality and so on."],"section_name":"Conclusion and Future Work"}],"paper_id":"1602.01208","title":"Spatial Concept Acquisition for a Mobile Robot that Integrates Self-Localization and Unsupervised Word Discovery from Spoken Sentences"}]