| { |
| "paper_id": "P12-1033", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T09:27:14.604237Z" |
| }, |
| "title": "Smaller Alignment Models for Better Translations: Unsupervised Word Alignment with the \u2113 0 -norm", |
| "authors": [ |
| { |
| "first": "Ashish", |
| "middle": [], |
| "last": "Vaswani", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Southern California Information Sciences Institute", |
| "location": {} |
| }, |
| "email": "avaswani@isi.edu" |
| }, |
| { |
| "first": "Liang", |
| "middle": [], |
| "last": "Huang", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Southern California Information Sciences Institute", |
| "location": {} |
| }, |
| "email": "lhuang@isi.edu" |
| }, |
| { |
| "first": "David", |
| "middle": [], |
| "last": "Chiang", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "University of Southern California Information Sciences Institute", |
| "location": {} |
| }, |
| "email": "chiang@isi.edu" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "Two decades after their invention, the IBM word-based translation models, widely available in the GIZA++ toolkit, remain the dominant approach to word alignment and an integral part of many statistical translation systems. Although many models have surpassed them in accuracy, none have supplanted them in practice. In this paper, we propose a simple extension to the IBM models: an \u2113 0 prior to encourage sparsity in the word-to-word translation model. We explain how to implement this extension efficiently for large-scale data (also released as a modification to GIZA++) and demonstrate, in experiments on Czech, Arabic, Chinese, and Urdu to English translation, significant improvements over IBM Model 4 in both word alignment (up to +6.7 F1) and translation quality (up to +1.4 Bleu).", |
| "pdf_parse": { |
| "paper_id": "P12-1033", |
| "_pdf_hash": "", |
| "abstract": [ |
| { |
| "text": "Two decades after their invention, the IBM word-based translation models, widely available in the GIZA++ toolkit, remain the dominant approach to word alignment and an integral part of many statistical translation systems. Although many models have surpassed them in accuracy, none have supplanted them in practice. In this paper, we propose a simple extension to the IBM models: an \u2113 0 prior to encourage sparsity in the word-to-word translation model. We explain how to implement this extension efficiently for large-scale data (also released as a modification to GIZA++) and demonstrate, in experiments on Czech, Arabic, Chinese, and Urdu to English translation, significant improvements over IBM Model 4 in both word alignment (up to +6.7 F1) and translation quality (up to +1.4 Bleu).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Abstract", |
| "sec_num": null |
| } |
| ], |
| "body_text": [ |
| { |
| "text": "Automatic word alignment is a vital component of nearly all current statistical translation pipelines. Although state-of-the-art translation models use rules that operate on units bigger than words (like phrases or tree fragments), they nearly always use word alignments to drive extraction of those translation rules. The dominant approach to word alignment has been the IBM models (Brown et al., 1993) together with the HMM model (Vogel et al., 1996) . These models are unsupervised, making them applicable to any language pair for which parallel text is available. Moreover, they are widely disseminated in the open-source GIZA++ toolkit (Och and Ney, 2004) . These properties make them the default choice for most statistical MT systems.", |
| "cite_spans": [ |
| { |
| "start": 383, |
| "end": 403, |
| "text": "(Brown et al., 1993)", |
| "ref_id": "BIBREF5" |
| }, |
| { |
| "start": 432, |
| "end": 452, |
| "text": "(Vogel et al., 1996)", |
| "ref_id": "BIBREF25" |
| }, |
| { |
| "start": 641, |
| "end": 660, |
| "text": "(Och and Ney, 2004)", |
| "ref_id": "BIBREF19" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "In the decades since their invention, many models have surpassed them in accuracy, but none has supplanted them in practice. Some of these models are partially supervised, combining unlabeled parallel text with manually-aligned parallel text (Moore, 2005; Taskar et al., 2005; Riesa and Marcu, 2010) . Although manually-aligned data is very valuable, it is only available for a small number of language pairs. Other models are unsupervised like the IBM models (Liang et al., 2006; Gra\u00e7a et al., 2010; Dyer et al., 2011) , but have not been as widely adopted as GIZA++ has.", |
| "cite_spans": [ |
| { |
| "start": 242, |
| "end": 255, |
| "text": "(Moore, 2005;", |
| "ref_id": "BIBREF18" |
| }, |
| { |
| "start": 256, |
| "end": 276, |
| "text": "Taskar et al., 2005;", |
| "ref_id": "BIBREF23" |
| }, |
| { |
| "start": 277, |
| "end": 299, |
| "text": "Riesa and Marcu, 2010)", |
| "ref_id": "BIBREF20" |
| }, |
| { |
| "start": 460, |
| "end": 480, |
| "text": "(Liang et al., 2006;", |
| "ref_id": "BIBREF15" |
| }, |
| { |
| "start": 481, |
| "end": 500, |
| "text": "Gra\u00e7a et al., 2010;", |
| "ref_id": "BIBREF12" |
| }, |
| { |
| "start": 501, |
| "end": 519, |
| "text": "Dyer et al., 2011)", |
| "ref_id": "BIBREF10" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "In this paper, we propose a simple extension to the IBM/HMM models that is unsupervised like the IBM models, is as scalable as GIZA++ because it is implemented on top of GIZA++, and provides significant improvements in both alignment and translation quality. It extends the IBM/HMM models by incorporating an \u2113 0 prior, inspired by the principle of minimum description length (Barron et al., 1998) , to encourage sparsity in the word-to-word translation model (Section 2.2). This extension follows our previous work on unsupervised part-ofspeech tagging (Vaswani et al., 2010) , but enables it to scale to the large datasets typical in word alignment, using an efficient training method based on projected gradient descent (Section 2.3). Experiments on Czech-, Arabic-, Chinese-and Urdu-English translation (Section 3) demonstrate consistent significant improvements over IBM Model 4 in both word alignment (up to +6.7 F1) and translation quality (up to +1.4 Bleu). Our implementation has been released as a simple modification to the GIZA++ toolkit that can be used as a drop-in replacement for GIZA++ in any existing MT pipeline.", |
| "cite_spans": [ |
| { |
| "start": 376, |
| "end": 397, |
| "text": "(Barron et al., 1998)", |
| "ref_id": "BIBREF0" |
| }, |
| { |
| "start": 554, |
| "end": 576, |
| "text": "(Vaswani et al., 2010)", |
| "ref_id": "BIBREF24" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "We start with a brief review of the IBM and HMM word alignment models, then describe how to extend them with a smoothed \u2113 0 prior and how to efficiently train them.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1" |
| }, |
| { |
| "text": "f = f 1 \u2022 \u2022 \u2022 f j \u2022 \u2022 \u2022 f m and an English string e = e 1 \u2022 \u2022 \u2022 e i \u2022 \u2022", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "\u2022 e \u2113 , these models describe the process by which the French string is generated by the English string via the alignment a = a 1 , . . . , a j , . . . , a m . Each a j is a hidden variables, indicating which English word e a j the French word f j is aligned to.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "In IBM Model 1-2 and the HMM model, the joint probability of the French sentence and alignment given the English sentence is", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "P(f, a | e) = m j=1 d(a j | a j\u22121 , j)t( f j | e a j ). (1)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "The parameters of these models are the distortion probabilities d(a j | a j\u22121 , j) and the translation probabilities t( f j | e a j ). The three models differ in their estimation of d, but the differences do not concern us Maximum likelihood training is prone to overfitting, especially in models with many parameters. In word alignment, one well-known manifestation of overfitting is that rare words can act as \"garbage collectors\" (Moore, 2004) , aligning to many unrelated words. This hurts alignment precision and rule-extraction recall. Previous attempted remedies include early stopping, smoothing (Moore, 2004) , and posterior regularization (Gra\u00e7a et al., 2010) .", |
| "cite_spans": [ |
| { |
| "start": 433, |
| "end": 446, |
| "text": "(Moore, 2004)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 604, |
| "end": 617, |
| "text": "(Moore, 2004)", |
| "ref_id": "BIBREF17" |
| }, |
| { |
| "start": 649, |
| "end": 669, |
| "text": "(Gra\u00e7a et al., 2010)", |
| "ref_id": "BIBREF12" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "We have previously proposed another simple remedy to overfitting in the context of unsupervised part-of-speech tagging (Vaswani et al., 2010) , which is to minimize the size of the model using a smoothed \u2113 0 prior. Applying this prior to an HMM improves tagging accuracy for both Italian and English.", |
| "cite_spans": [ |
| { |
| "start": 119, |
| "end": 141, |
| "text": "(Vaswani et al., 2010)", |
| "ref_id": "BIBREF24" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "Here, our goal is to apply a similar prior in a word-alignment model to the word-to-word translation probabilities t( f | e). We leave the distortion models alone, since they are not very large, and there is not much reason to believe that we can profit from compacting them.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "With the addition of the \u2113 0 prior, the MAP (maximum a posteriori) objective function i\u015d", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u03b8 = arg min \u03b8 \u2212 log P(f | e, \u03b8)P(\u03b8)", |
| "eq_num": "(4)" |
| } |
| ], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "where", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "P(\u03b8) \u221d exp \u2212\u03b1 \u03b8 \u03b2 0", |
| "eq_num": "(5)" |
| } |
| ], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "and", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u03b8 \u03b2 0 = e, f 1 \u2212 exp \u2212t( f | e) \u03b2", |
| "eq_num": "(6)" |
| } |
| ], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "is a smoothed approximation of the \u2113 0 -norm. The hyperparameter \u03b2 controls the tightness of the approximation, as illustrated in Figure 1 . Substituting back into (4) and dropping constant terms, we get the following optimization problem: minimize", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 130, |
| "end": 138, |
| "text": "Figure 1", |
| "ref_id": "FIGREF0" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u2212 log P(f | e, \u03b8) \u2212 \u03b1 e, f exp \u2212t( f | e) \u03b2", |
| "eq_num": "(7)" |
| } |
| ], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "subject to the constraints f t( f | e) = 1 for all e.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "We can carry out the optimization in (7) with the MAP-EM algorithm (Bishop, 2006) . EM and MAP-EM share the same E-step; the difference lies in the M-step. For vanilla EM, the M-step is:", |
| "cite_spans": [ |
| { |
| "start": 67, |
| "end": 81, |
| "text": "(Bishop, 2006)", |
| "ref_id": "BIBREF2" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u03b8 = arg min \u03b8 \uf8eb \uf8ec \uf8ec \uf8ec \uf8ec \uf8ec \uf8ec \uf8ec \uf8ed \u2212 e, f E[C(e, f )] log t( f | e) \uf8f6 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f7 \uf8f8", |
| "eq_num": "(9)" |
| } |
| ], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "again subject to the constraints (8). The count C(e, f ) is the number of times that f occurs aligned to e. For MAP-EM, it is:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u03b8 = arg min \u03b8 \u2212 e, f E[C(e, f )] log t( f | e) \u2212 \u03b1 e, f exp \u2212t( f | e) \u03b2", |
| "eq_num": "(10)" |
| } |
| ], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "This optimization problem is non-convex, and we do not know of a closed-form solution. Previously (Vaswani et al., 2010) , we used ALGENCAN, a nonlinear optimization toolkit, but this solution does not scale well to the number of parameters involved in word alignment models. Instead, we use a simpler and more scalable method which we describe in the next section.", |
| "cite_spans": [ |
| { |
| "start": 98, |
| "end": 120, |
| "text": "(Vaswani et al., 2010)", |
| "ref_id": "BIBREF24" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Given a French string", |
| "sec_num": null |
| }, |
| { |
| "text": "Following Schoenemann (2011b), we use projected gradient descent (PGD) to solve the M-step (but with the \u2113 0 -norm instead of the \u2113 1 -norm). Gradient projection methods are attractive solutions to constrained optimization problems, particularly when the constraints on the parameters are simple (Bertsekas, 1999) . Let F(\u03b8) be the objective function in (10); we seek to minimize this function. As in previous work (Vaswani et al., 2010) , we optimize each set of parameters {t(\u2022 | e)} separately for each English word type e. The inputs to the PGD are the expected counts E[C(e, f )] and the current word-toword conditional probabilities \u03b8. We run PGD for K iterations, producing a sequence of intermediate parameter vectors \u03b8 1 , . . . , \u03b8 k , . . . , \u03b8 K . Each iteration has two steps, a projection step and a line search.", |
| "cite_spans": [ |
| { |
| "start": 296, |
| "end": 313, |
| "text": "(Bertsekas, 1999)", |
| "ref_id": "BIBREF1" |
| }, |
| { |
| "start": 415, |
| "end": 437, |
| "text": "(Vaswani et al., 2010)", |
| "ref_id": "BIBREF24" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "Projection step In this step, we compute:", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u03b8 k = \u03b8 k \u2212 s\u2207F(\u03b8 k ) \u2206", |
| "eq_num": "(11)" |
| } |
| ], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "This moves \u03b8 in the direction of steepest descent (\u2207F) with step size s, and then the function [\u2022] \u2206 projects the resulting point onto the simplex; that is, it finds the nearest point that satisfies the constraints (8).", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "The gradient \u2207F(\u03b8 k ) is", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "\u2202F \u2202t( f | e) = \u2212 E[C( f, e)] t( f | e) + \u03b1 \u03b2 exp \u2212t( f | e) \u03b2", |
| "eq_num": "(12)" |
| } |
| ], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "In contrast to Schoenemann (2011b), we use an O(n log n) algorithm for the projection step due to Duchi et. al. (2008) , shown in Pseudocode 1.", |
| "cite_spans": [ |
| { |
| "start": 98, |
| "end": 118, |
| "text": "Duchi et. al. (2008)", |
| "ref_id": "BIBREF9" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "Pseudocode 1 Project input vector u \u2208 R n onto the probability simplex. v = u sorted in non-decreasing order \u03c1 = 0", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "for i = 1 to n do if v i \u2212 1 i i r=1 v r \u2212 1 > 0 then \u03c1 = i end if end for \u03b7 = 1 \u03c1 \u03c1 r=1 v r \u2212 1 w r = max{v r \u2212 \u03b7, 0} for 1 \u2264 r \u2264 n return w", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "Line search Next, we move to a point between \u03b8 k and \u03b8 k that satisfies the Armijo condition,", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "EQUATION", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [ |
| { |
| "start": 0, |
| "end": 8, |
| "text": "EQUATION", |
| "ref_id": "EQREF", |
| "raw_str": "F(\u03b8 k + \u03b4 m ) \u2264 F(\u03b8 k ) + \u03c3 \u2207F(\u03b8 k ) \u2022 \u03b4 m", |
| "eq_num": "(13)" |
| } |
| ], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "where \u03b4 m = \u03b3 m (\u03b8 k \u2212 \u03b8 k ) and \u03c3 and \u03b3 are both constants in (0, 1). We try values m = 1, 2, . . . until the Armijo condition (13) is satisfied or the limit m = 20", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "Pseudocode 2 Find a point between \u03b8 k and \u03b8 k that satisfies the Armijo condition.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "F min = F(\u03b8 k ) \u03b8 min = \u03b8 k for m = 1 to 20 do \u03b4 m = \u03b3 m \u03b8 k \u2212 \u03b8 k if F(\u03b8 k + \u03b4 m ) < F min then F min = F(\u03b8 k + \u03b4 m ) \u03b8 min = \u03b8 k + \u03b4 m end if if F(\u03b8 k + \u03b4 m ) \u2264 F(\u03b8 k ) + \u03c3 \u2207F(\u03b8 k ) \u2022 \u03b4 m then break end if end for \u03b8 k+1 = \u03b8 min return \u03b8 k+1", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "is reached. (Note that we don't allow m = 0 because this can cause \u03b8 k + \u03b4 m to land on the boundary of the probability simplex, where the objective function is undefined.) Then we set \u03b8 k+1 to the point in", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "{\u03b8 k } \u222a {\u03b8 k + \u03b4 m | 1 \u2264 m \u2264 20} that minimizes F.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "The line search algorithm is summarized in Pseudocode 2.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "In our implementation, we set \u03b3 = 0.5 and \u03c3 = 0.5. We keep s fixed for all PGD iterations; we experimented with s \u2208 {0.1, 0.5} and did not observe significant changes in F-score. We run the projection step and line search alternately for at most K iterations, terminating early if there is no change in \u03b8 k from one iteration to the next. We set K = 35 for the large Arabic-English experiment; for all other conditions, we set K = 50. These choices were made to balance efficiency and accuracy. We found that values of K between 30 and 75 were generally reasonable.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Projected gradient descent", |
| "sec_num": "2.3" |
| }, |
| { |
| "text": "To demonstrate the effect of the \u2113 0 -norm on the IBM models, we performed experiments on four translation tasks: Arabic-English, Chinese-English, and Urdu-English from the NIST Open MT Evaluation, and the Czech-English translation from the Workshop on Machine Translation (WMT) shared task. We measured the accuracy of word alignments generated by GIZA++ with and without the \u2113 0 -norm, and also translation accuracy of systems trained using the word alignments. Across all tests, we found strong improvements from adding the \u2113 0 -norm.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Experiments", |
| "sec_num": "3" |
| }, |
| { |
| "text": "We have implemented our algorithm as an opensource extension to GIZA++. 1 Usage of the extension is identical to standard GIZA++, except that the user can switch the \u2113 0 prior on or off, and adjust the hyperparameters \u03b1 and \u03b2.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "For vanilla EM, we ran five iterations of Model 1, five iterations of HMM, and ten iterations of Model 4. For our approach, we first ran one iteration of Model 1, followed by four iterations of Model 1 with smoothed \u2113 0 , followed by five iterations of HMM with smoothed \u2113 0 . Finally, we ran ten iterations of Model 4. 2 We used the following parallel data:", |
| "cite_spans": [ |
| { |
| "start": 320, |
| "end": 321, |
| "text": "2", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u2022 Chinese-English: selected data from the constrained task of the NIST 2009 Open MT Evaluation. 3", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u2022 Arabic-English: all available data for the constrained track of NIST 2009, excluding United Nations proceedings (LDC2004E13), ISI Automatically Extracted Parallel Text (LDC2007E08), and Ummah newswire text (LDC2004T18), for a total of 5.4+4.3 million words. We also experimented on a larger Arabic-English parallel text of 44+37 million words from the DARPA GALE program.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u2022 Urdu-English: all available data for the constrained track of NIST 2009.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "1 The code can be downloaded from the first author's website at http://www.isi.edu/\u02dcavaswani/giza-pp-l0.html.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "2 GIZA++ allows changing some heuristic parameters for efficient training. Currently, we set two of these to zero: mincountincrease and probcutoff. In the default setting, both are set to 10 \u22127 . We set probcutoff to 0 because we would like the optimization to learn the parameter values. For a fair comparison, we applied the same setting to our vanilla EM training as well. To test, we ran GIZA++ with the default setting on the smaller of our two Arabic-English datasets with the same number of iterations and found no change in F-score.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "3 LDC catalog numbers LDC2003E07, LDC2003E14, LDC2005E83, LDC2005T06, LDC2006E24, LDC2006E34, LDC2006E85, LDC2006E86, LDC2006E92, and LDC2006E93. p r e s i d e n t o f t h e f o r e i g n a f f a i r s i n s t i t u t e s h u q i n l i u w a s a l s o p r e s e n t a t t h e m e e t i n g . Figure 2 : Smoothed-\u2113 0 alignments (red circles) correct many errors in the baseline GIZA++ alignments (black squares), as shown in four Chinese-English examples (the red circles are almost perfect for these examples, except for minor mistakes such as liu-sh\u016bq\u012bng and meeting-z\u00e0izu\u00f2 in (a) and .-, in (c)). In particular, the baseline system demonstrates typical \"garbage-collection\" phenomena in proper name \"shuqing\" in both languages in (a), number \"4000\" and word \"l\u00e1ib\u012bn\" (lit. \"guest\") in (b), word \"troublesome\" and \"l\u00f9l\u00f9\" (lit. \"land-route\") in (c), and \"blockhouses\" and \"di\u0101ob\u01ceo\" (lit. \"bunker\") in (d). We found this garbage-collection behavior to be especially common with proper names, numbers, and uncommon words in both languages. Most interestingly, in (c), our smoothed-\u2113 0 system correctly aligns \"extremely\" to \"h\u011bn h\u011bn h\u011bn h\u011bn\" (lit. \"very very very very\") which is rare in the bitext. column shows the average fertility of once-seen source words. For Czech-English, the year refers to the WMT shared task; for all other language pairs, the year refers to the NIST Open MT Evaluation. * Half of this test set was also used for tuning feature weights.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 292, |
| "end": 300, |
| "text": "Figure 2", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u00d9 \u00d9 w\u00e0iji\u0101o \u00d9 xu\u00e9hu\u00ec \u00d9 hu\u00eczh\u01ceng \u00d9 li\u00fa \u00d9 \u00d9 sh\u016bq\u012bng \u00d9 hu\u00ecji\u00e0n sh\u00ed \u00d9 \u00d9 \u00d9 \u00d9 z\u00e0izu\u00f2 \u00d9 . o v", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u00d9 zh\u00e0 \u00d9 le \u00d9 s\u00ecge \u00d9 di\u0101ob\u01ceo \u00d9 . (c) (d)", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "\u2022 Czech-English: A corpus of 4 million words of Czech-English data from the News Commentary corpus. 4 We set the hyperparameters \u03b1 and \u03b2 by tuning on gold-standard word alignments (to maximize F1) when possible. For Arabic-English and Chinese-English, we used 346 and 184 hand-aligned sentences from LDC2006E86 and LDC2006E93. Similarly, for Czech-English, 515 hand-aligned sentences were available (Bojar and Prokopov\u00e1, 2006) . But for Urdu-English, since we did not have any gold alignments, we used \u03b1 = 10 and \u03b2 = 0.05. We did not choose a large \u03b1, as the dataset was small, and we chose a conservative value for \u03b2.", |
| "cite_spans": [ |
| { |
| "start": 100, |
| "end": 101, |
| "text": "4", |
| "ref_id": null |
| }, |
| { |
| "start": 399, |
| "end": 426, |
| "text": "(Bojar and Prokopov\u00e1, 2006)", |
| "ref_id": "BIBREF4" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "We ran word alignment in both directions and symmetrized using grow-diag-final (Koehn et al., 2003) . For models with the smoothed \u2113 0 prior, we tuned \u03b1 and \u03b2 separately in each direction.", |
| "cite_spans": [ |
| { |
| "start": 79, |
| "end": 99, |
| "text": "(Koehn et al., 2003)", |
| "ref_id": "BIBREF13" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Training", |
| "sec_num": "3.1" |
| }, |
| { |
| "text": "First, we evaluated alignment accuracy directly by comparing against gold-standard word alignments.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Alignment", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "The results are shown in the alignment F1 column of Table 1 . We used balanced F-measure rather than alignment error rate as our metric (Fraser and Marcu, 2007) .", |
| "cite_spans": [ |
| { |
| "start": 136, |
| "end": 160, |
| "text": "(Fraser and Marcu, 2007)", |
| "ref_id": "BIBREF11" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 52, |
| "end": 59, |
| "text": "Table 1", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Alignment", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "Following Dyer et al. (2011) , we also measured the average fertility,\u03c6 sing. , of once-seen source words in the symmetrized alignments. Our alignments show smaller fertility for once-seen words, suggesting that they suffer from \"garbage collection\" effects less than the baseline alignments do.", |
| "cite_spans": [ |
| { |
| "start": 10, |
| "end": 28, |
| "text": "Dyer et al. (2011)", |
| "ref_id": "BIBREF10" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Alignment", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "The fact that we had to use hand-aligned data to tune the hyperparameters \u03b1 and \u03b2 means that our method is no longer completely unsupervised. However, our observation is that alignment accuracy is actually fairly robust to the choice of these hyperparameters, as shown in Table 2 . As we will see below, we still obtained strong improvements in translation quality when hand-aligned data was unavailable.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 272, |
| "end": 279, |
| "text": "Table 2", |
| "ref_id": "TABREF2" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Alignment", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "We also tried generating 50 word classes using the tool provided in GIZA++. We found that adding word classes improved alignment quality a little, but more so for the baseline system (see Table 3 ). We used the alignments generated by training with word classes for our translation experiments. Table 3 : Adding word classes improves the F-score in both directions for Arabic-English alignment by a little, for the baseline system more so than ours. Figure 2 shows four examples of Chinese-English alignment, comparing the baseline with our smoothed-\u2113 0 method. In all four cases, the baseline produces incorrect extra alignments that prevent good translation rules from being extracted while the smoothed-\u2113 0 results are correct. In particular, the baseline system demonstrates typical \"garbage collection\" behavior (Moore, 2004) in all four examples.", |
| "cite_spans": [ |
| { |
| "start": 817, |
| "end": 830, |
| "text": "(Moore, 2004)", |
| "ref_id": "BIBREF17" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 188, |
| "end": 195, |
| "text": "Table 3", |
| "ref_id": null |
| }, |
| { |
| "start": 295, |
| "end": 302, |
| "text": "Table 3", |
| "ref_id": null |
| }, |
| { |
| "start": 450, |
| "end": 458, |
| "text": "Figure 2", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Alignment", |
| "sec_num": "3.2" |
| }, |
| { |
| "text": "We then tested the effect of word alignments on translation quality using the hierarchical phrasebased translation system Hiero (Chiang, 2007) . We used a fairly standard set of features: seven inherited from Pharaoh (Koehn et al., 2003) Table 4 : Optimizing hyperparameters on alignment F1 score does not necessarily lead to optimal Bleu. The first two columns indicate whether we used the first-or second-best alignments in each direction (according to F1); the third column shows the F1 of the symmetrized alignments, whose corresponding Bleu scores are shown in the last two columns.", |
| "cite_spans": [ |
| { |
| "start": 128, |
| "end": 142, |
| "text": "(Chiang, 2007)", |
| "ref_id": "BIBREF7" |
| }, |
| { |
| "start": 217, |
| "end": 237, |
| "text": "(Koehn et al., 2003)", |
| "ref_id": "BIBREF13" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 238, |
| "end": 245, |
| "text": "Table 4", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Translation", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "ond language model, and penalties for the glue rule, identity rules, unknown-word rules, and two kinds of number/name rules. The feature weights were discriminatively trained using MIRA (Chiang et al., 2008) . We used two 5-gram language models, one on the combined English sides of the NIST 2009 Arabic-English and Chinese-English constrained tracks (385M words), and another on 2 billion words of English. For each language pair, we extracted grammar rules from the same data that were used for word alignment. The development data that were used for discriminative training were: for Chinese-English and Arabic-English, data from the NIST 2004 and NIST 2006 test sets, plus newsgroup data from the GALE program (LDC2006E92); for Urdu-English, half of the NIST 2008 test set; for Czech-English, a training set of 2051 sentences provided by the WMT10 translation workshop.", |
| "cite_spans": [ |
| { |
| "start": 186, |
| "end": 207, |
| "text": "(Chiang et al., 2008)", |
| "ref_id": "BIBREF6" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Translation", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "The results are shown in the Bleu column of Table 1. We used case-insensitive IBM Bleu (closest reference length) as our metric. Significance testing was carried out using bootstrap resampling with 1000 samples (Koehn, 2004; Zhang et al., 2004) .", |
| "cite_spans": [ |
| { |
| "start": 211, |
| "end": 224, |
| "text": "(Koehn, 2004;", |
| "ref_id": "BIBREF14" |
| }, |
| { |
| "start": 225, |
| "end": 244, |
| "text": "Zhang et al., 2004)", |
| "ref_id": "BIBREF26" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Translation", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "All of the tests showed significant improvements (p < 0.01), ranging from +0.4 Bleu to +1.4 Bleu. For Urdu, even though we didn't have manual alignments to tune hyperparameters, we got significant gains over a good baseline. This is promising for languages that do not have any manually aligned data.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Translation", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "Ideally, one would want to tune \u03b1 and \u03b2 to maximize Bleu. However, this is prohibitively expensive, especially if we must tune them separately in each alignment direction before symmetrization. We ran some contrastive experiments to investigate the impact of hyperparameter tuning on translation quality. For the smaller Arabic-English corpus, we symmetrized all combinations of the two top-scoring alignments (according to F1) in each direction, yielding four sets of alignments. Table 4 shows Bleu scores for translation models learned from these alignments. Unfortunately, we find that optimizing F1 is not optimal for Bleu-using the second-best alignments yields a further improvement of 0.5 Bleu on the NIST 2009 data, which is statistically significant (p < 0.05).", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 481, |
| "end": 488, |
| "text": "Table 4", |
| "ref_id": null |
| } |
| ], |
| "eq_spans": [], |
| "section": "Translation", |
| "sec_num": "3.3" |
| }, |
| { |
| "text": "Schoenemann (2011a), taking inspiration from Bodrumlu et al. (2009) , uses integer linear programming to optimize IBM Model 1-2 and the HMM with the \u2113 0 -norm. This method, however, does not outperform GIZA++. In later work, Schoenemann (2011b) used projected gradient descent for the \u2113 1norm. Here, we have adopted his use of projected gradient descent, but using a smoothed \u2113 0 -norm. Liang et al. (2006) show how to train IBM models in both directions simultaneously by adding a term to the log-likelihood that measures the agreement between the two directions. Gra\u00e7a et al. (2010) explore modifications to the HMM model that encourage bijectivity and symmetry. The modifications take the form of constraints on the posterior distribution over alignments that is computed during the E-step. Mermer and Sara\u00e7lar (2011) explore a Bayesian version of IBM Model 1, applying sparse Dirichlet priors to t. However, because this method requires the use of Monte Carlo methods, it is not clear how well it can scale to larger datasets.", |
| "cite_spans": [ |
| { |
| "start": 45, |
| "end": 67, |
| "text": "Bodrumlu et al. (2009)", |
| "ref_id": "BIBREF3" |
| }, |
| { |
| "start": 387, |
| "end": 406, |
| "text": "Liang et al. (2006)", |
| "ref_id": "BIBREF15" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Related Work", |
| "sec_num": "4" |
| }, |
| { |
| "text": "We have extended the IBM models and HMM model by the addition of an \u2113 0 prior to the word-to-word translation model, which compacts the word-toword translation table, reducing overfitting, and, in particular, the \"garbage collection\" effect. We have shown how to perform MAP-EM with this prior efficiently, even for large datasets. The method is implemented as a modification to the open-source toolkit GIZA++, and we have shown that it significantly improves translation quality across four different language pairs. Even though we have used a small set of gold-standard alignments to tune our hyperparameters, we found that performance was fairly robust to variation in the hyperparameters, and translation performance was good even when goldstandard alignments were unavailable. We hope that our method, due to its simplicity, generality, and effectiveness, will find wide application for training better statistical translation systems.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Conclusion", |
| "sec_num": "5" |
| }, |
| { |
| "text": "This data is available at http://statmt.org/wmt10.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "", |
| "sec_num": null |
| } |
| ], |
| "back_matter": [ |
| { |
| "text": "We are indebted to Thomas Schoenemann for initial discussions and pilot experiments that led to this work, and to the anonymous reviewers for their valuable comments. We thank Jason Riesa for providing the Arabic-English and Chinese-English hand-aligned data and the alignment visualization tool, and Chris Dyer for the Czech-English handaligned data. This research was supported in part by DARPA under contract DOI-NBC D11AP00244 and a Google Faculty Research Award to L. H.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Acknowledgments", |
| "sec_num": null |
| } |
| ], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "The minimum description length principle in coding and modeling", |
| "authors": [ |
| { |
| "first": "Andrew", |
| "middle": [], |
| "last": "Barron", |
| "suffix": "" |
| }, |
| { |
| "first": "Jorma", |
| "middle": [], |
| "last": "Rissanen", |
| "suffix": "" |
| }, |
| { |
| "first": "Bin", |
| "middle": [], |
| "last": "Yu", |
| "suffix": "" |
| } |
| ], |
| "year": 1998, |
| "venue": "IEEE Transactions on Information Theory", |
| "volume": "44", |
| "issue": "6", |
| "pages": "2743--2760", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Andrew Barron, Jorma Rissanen, and Bin Yu. 1998. The minimum description length principle in coding and modeling. IEEE Transactions on Information Theory, 44(6):2743-2760.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "Nonlinear Programming", |
| "authors": [ |
| { |
| "first": "Dimitri", |
| "middle": [ |
| "P" |
| ], |
| "last": "Bertsekas", |
| "suffix": "" |
| } |
| ], |
| "year": 1999, |
| "venue": "Athena Scientific", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Dimitri P. Bertsekas. 1999. Nonlinear Programming. Athena Scientific.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Pattern Recognition and Machine Learning", |
| "authors": [ |
| { |
| "first": "Christopher", |
| "middle": [ |
| "M" |
| ], |
| "last": "Bishop", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Christopher M. Bishop. 2006. Pattern Recognition and Machine Learning. Springer.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "A new objective function for word alignment", |
| "authors": [ |
| { |
| "first": "Tugba", |
| "middle": [], |
| "last": "Bodrumlu", |
| "suffix": "" |
| }, |
| { |
| "first": "Kevin", |
| "middle": [], |
| "last": "Knight", |
| "suffix": "" |
| }, |
| { |
| "first": "Sujith", |
| "middle": [], |
| "last": "Ravi", |
| "suffix": "" |
| } |
| ], |
| "year": 2009, |
| "venue": "Proceedings of the NAACL HLT Workshop on Integer Linear Programming for Natural Language Processing", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Tugba Bodrumlu, Kevin Knight, and Sujith Ravi. 2009. A new objective function for word alignment. In Pro- ceedings of the NAACL HLT Workshop on Integer Lin- ear Programming for Natural Language Processing.", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "Czech-English word alignment", |
| "authors": [ |
| { |
| "first": "Ond\u0159ej", |
| "middle": [], |
| "last": "Bojar", |
| "suffix": "" |
| }, |
| { |
| "first": "Magdalena", |
| "middle": [], |
| "last": "Prokopov\u00e1", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "Proceedings of LREC", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Ond\u0159ej Bojar and Magdalena Prokopov\u00e1. 2006. Czech- English word alignment. In Proceedings of LREC.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "The mathematics of statistical machine translation: Parameter estimation", |
| "authors": [ |
| { |
| "first": "F", |
| "middle": [], |
| "last": "Peter", |
| "suffix": "" |
| }, |
| { |
| "first": "Stephen", |
| "middle": [ |
| "A Della" |
| ], |
| "last": "Brown", |
| "suffix": "" |
| }, |
| { |
| "first": "Vincent", |
| "middle": [ |
| "J" |
| ], |
| "last": "Pietra", |
| "suffix": "" |
| }, |
| { |
| "first": "Robert", |
| "middle": [ |
| "L" |
| ], |
| "last": "Della Pietra", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Mercer", |
| "suffix": "" |
| } |
| ], |
| "year": 1993, |
| "venue": "Computational Linguistics", |
| "volume": "19", |
| "issue": "", |
| "pages": "263--311", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The mathemat- ics of statistical machine translation: Parameter esti- mation. Computational Linguistics, 19:263-311.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Online large-margin training of syntactic and structural translation features", |
| "authors": [ |
| { |
| "first": "David", |
| "middle": [], |
| "last": "Chiang", |
| "suffix": "" |
| }, |
| { |
| "first": "Yuval", |
| "middle": [], |
| "last": "Marton", |
| "suffix": "" |
| }, |
| { |
| "first": "Philip", |
| "middle": [], |
| "last": "Resnik", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "Proceedings of EMNLP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "David Chiang, Yuval Marton, and Philip Resnik. 2008. Online large-margin training of syntactic and struc- tural translation features. In Proceedings of EMNLP.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "Hierarchical phrase-based translation", |
| "authors": [ |
| { |
| "first": "David", |
| "middle": [], |
| "last": "Chiang", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Computational Linguistics", |
| "volume": "33", |
| "issue": "2", |
| "pages": "201--208", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "David Chiang. 2007. Hierarchical phrase-based transla- tion. Computational Linguistics, 33(2):201-208.", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Maximum likelihood from incomplete data via the EM algorithm", |
| "authors": [ |
| { |
| "first": "A", |
| "middle": [ |
| "P" |
| ], |
| "last": "Dempster", |
| "suffix": "" |
| }, |
| { |
| "first": "N", |
| "middle": [ |
| "M" |
| ], |
| "last": "Laird", |
| "suffix": "" |
| }, |
| { |
| "first": "D", |
| "middle": [ |
| "B" |
| ], |
| "last": "Rubin", |
| "suffix": "" |
| } |
| ], |
| "year": 1977, |
| "venue": "Computational Linguistics", |
| "volume": "39", |
| "issue": "4", |
| "pages": "1--38", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "A. P. Dempster, N. M. Laird, and D. B. Rubin. 1977. Maximum likelihood from incomplete data via the EM algorithm. Computational Linguistics, 39(4):1-38.", |
| "links": null |
| }, |
| "BIBREF9": { |
| "ref_id": "b9", |
| "title": "Efficient projections onto the \u2113 1 -ball for learning in high dimensions", |
| "authors": [ |
| { |
| "first": "John", |
| "middle": [], |
| "last": "Duchi", |
| "suffix": "" |
| }, |
| { |
| "first": "Shai", |
| "middle": [], |
| "last": "Shalev-Shwartz", |
| "suffix": "" |
| }, |
| { |
| "first": "Yoram", |
| "middle": [], |
| "last": "Singer", |
| "suffix": "" |
| }, |
| { |
| "first": "Tushar", |
| "middle": [], |
| "last": "Chandra", |
| "suffix": "" |
| } |
| ], |
| "year": 2008, |
| "venue": "Proceedings of ICML", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "John Duchi, Shai Shalev-Shwartz, Yoram Singer, and Tushar Chandra. 2008. Efficient projections onto the \u2113 1 -ball for learning in high dimensions. In Proceed- ings of ICML.", |
| "links": null |
| }, |
| "BIBREF10": { |
| "ref_id": "b10", |
| "title": "Unsupervised word alignment with arbitrary features", |
| "authors": [ |
| { |
| "first": "Chris", |
| "middle": [], |
| "last": "Dyer", |
| "suffix": "" |
| }, |
| { |
| "first": "Jonathan", |
| "middle": [ |
| "H" |
| ], |
| "last": "Clark", |
| "suffix": "" |
| }, |
| { |
| "first": "Alon", |
| "middle": [], |
| "last": "Lavie", |
| "suffix": "" |
| }, |
| { |
| "first": "Noah", |
| "middle": [ |
| "A" |
| ], |
| "last": "Smith", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Chris Dyer, Jonathan H. Clark, Alon Lavie, and Noah A. Smith. 2011. Unsupervised word alignment with ar- bitrary features. In Proceedings of ACL.", |
| "links": null |
| }, |
| "BIBREF11": { |
| "ref_id": "b11", |
| "title": "Measuring word alignment quality for statistical machine translation", |
| "authors": [ |
| { |
| "first": "Alexander", |
| "middle": [], |
| "last": "Fraser", |
| "suffix": "" |
| }, |
| { |
| "first": "Daniel", |
| "middle": [], |
| "last": "Marcu", |
| "suffix": "" |
| } |
| ], |
| "year": 2007, |
| "venue": "Computational Linguistics", |
| "volume": "33", |
| "issue": "3", |
| "pages": "293--303", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Alexander Fraser and Daniel Marcu. 2007. Measuring word alignment quality for statistical machine transla- tion. Computational Linguistics, 33(3):293-303.", |
| "links": null |
| }, |
| "BIBREF12": { |
| "ref_id": "b12", |
| "title": "Learning tractable word alignment models with complex constraints", |
| "authors": [ |
| { |
| "first": "V", |
| "middle": [], |
| "last": "Jo\u00e3o", |
| "suffix": "" |
| }, |
| { |
| "first": "Kuzman", |
| "middle": [], |
| "last": "Gra\u00e7a", |
| "suffix": "" |
| }, |
| { |
| "first": "Ben", |
| "middle": [], |
| "last": "Ganchev", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Taskar", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Computational Linguistics", |
| "volume": "36", |
| "issue": "3", |
| "pages": "481--504", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Jo\u00e3o V. Gra\u00e7a, Kuzman Ganchev, and Ben Taskar. 2010. Learning tractable word alignment models with complex constraints. Computational Linguistics, 36(3):481-504.", |
| "links": null |
| }, |
| "BIBREF13": { |
| "ref_id": "b13", |
| "title": "Statistical phrase-based translation", |
| "authors": [ |
| { |
| "first": "Philipp", |
| "middle": [], |
| "last": "Koehn", |
| "suffix": "" |
| }, |
| { |
| "first": "Franz", |
| "middle": [ |
| "Joseph" |
| ], |
| "last": "Och", |
| "suffix": "" |
| }, |
| { |
| "first": "Daniel", |
| "middle": [], |
| "last": "Marcu", |
| "suffix": "" |
| } |
| ], |
| "year": 2003, |
| "venue": "Proceedings of NAACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Philipp Koehn, Franz Joseph Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceed- ings of NAACL.", |
| "links": null |
| }, |
| "BIBREF14": { |
| "ref_id": "b14", |
| "title": "Statistical significance tests for machine translation evaluation", |
| "authors": [ |
| { |
| "first": "Philipp", |
| "middle": [], |
| "last": "Koehn", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of EMNLP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of EMNLP.", |
| "links": null |
| }, |
| "BIBREF15": { |
| "ref_id": "b15", |
| "title": "Alignment by agreement", |
| "authors": [ |
| { |
| "first": "Percy", |
| "middle": [], |
| "last": "Liang", |
| "suffix": "" |
| }, |
| { |
| "first": "Ben", |
| "middle": [], |
| "last": "Taskar", |
| "suffix": "" |
| }, |
| { |
| "first": "Dan", |
| "middle": [], |
| "last": "Klein", |
| "suffix": "" |
| } |
| ], |
| "year": 2006, |
| "venue": "Proceedings of HLT-NAACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Percy Liang, Ben Taskar, and Dan Klein. 2006. Align- ment by agreement. In Proceedings of HLT-NAACL.", |
| "links": null |
| }, |
| "BIBREF16": { |
| "ref_id": "b16", |
| "title": "Bayesian word alignment for statistical machine translation", |
| "authors": [ |
| { |
| "first": "Co\u015fkun", |
| "middle": [], |
| "last": "Mermer", |
| "suffix": "" |
| }, |
| { |
| "first": "Murat", |
| "middle": [], |
| "last": "Sara\u00e7lar", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of ACL HLT", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Co\u015fkun Mermer and Murat Sara\u00e7lar. 2011. Bayesian word alignment for statistical machine translation. In Proceedings of ACL HLT.", |
| "links": null |
| }, |
| "BIBREF17": { |
| "ref_id": "b17", |
| "title": "Improving IBM wordalignment Model 1", |
| "authors": [ |
| { |
| "first": "Robert", |
| "middle": [ |
| "C" |
| ], |
| "last": "Moore", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Robert C. Moore. 2004. Improving IBM word- alignment Model 1. In Proceedings of ACL.", |
| "links": null |
| }, |
| "BIBREF18": { |
| "ref_id": "b18", |
| "title": "A discriminative framework for bilingual word alignment", |
| "authors": [ |
| { |
| "first": "Robert", |
| "middle": [], |
| "last": "Moore", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of HLT-EMNLP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Robert Moore. 2005. A discriminative framework for bilingual word alignment. In Proceedings of HLT- EMNLP.", |
| "links": null |
| }, |
| "BIBREF19": { |
| "ref_id": "b19", |
| "title": "The alignment template approach to statistical machine translation", |
| "authors": [ |
| { |
| "first": "Joseph", |
| "middle": [], |
| "last": "Franz", |
| "suffix": "" |
| }, |
| { |
| "first": "Hermann", |
| "middle": [], |
| "last": "Och", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Ney", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Computational Linguistics", |
| "volume": "30", |
| "issue": "", |
| "pages": "417--449", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Franz Joseph Och and Hermann Ney. 2004. The align- ment template approach to statistical machine transla- tion. Computational Linguistics, 30:417-449.", |
| "links": null |
| }, |
| "BIBREF20": { |
| "ref_id": "b20", |
| "title": "Hierarchical search for word alignment", |
| "authors": [ |
| { |
| "first": "Jason", |
| "middle": [], |
| "last": "Riesa", |
| "suffix": "" |
| }, |
| { |
| "first": "Daniel", |
| "middle": [], |
| "last": "Marcu", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Proceedings of ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Jason Riesa and Daniel Marcu. 2010. Hierarchical search for word alignment. In Proceedings of ACL.", |
| "links": null |
| }, |
| "BIBREF21": { |
| "ref_id": "b21", |
| "title": "Probabilistic word alignment under the L 0 -norm", |
| "authors": [ |
| { |
| "first": "Thomas", |
| "middle": [], |
| "last": "Schoenemann", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of CoNLL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Thomas Schoenemann. 2011a. Probabilistic word align- ment under the L 0 -norm. In Proceedings of CoNLL.", |
| "links": null |
| }, |
| "BIBREF22": { |
| "ref_id": "b22", |
| "title": "Regularizing mono-and bi-word models for word alignment", |
| "authors": [ |
| { |
| "first": "Thomas", |
| "middle": [], |
| "last": "Schoenemann", |
| "suffix": "" |
| } |
| ], |
| "year": 2011, |
| "venue": "Proceedings of IJCNLP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Thomas Schoenemann. 2011b. Regularizing mono-and bi-word models for word alignment. In Proceedings of IJCNLP.", |
| "links": null |
| }, |
| "BIBREF23": { |
| "ref_id": "b23", |
| "title": "A discriminative matching approach to word alignment", |
| "authors": [ |
| { |
| "first": "Ben", |
| "middle": [], |
| "last": "Taskar", |
| "suffix": "" |
| }, |
| { |
| "first": "Lacoste-Julien", |
| "middle": [], |
| "last": "Simon", |
| "suffix": "" |
| }, |
| { |
| "first": "Klein", |
| "middle": [], |
| "last": "Dan", |
| "suffix": "" |
| } |
| ], |
| "year": 2005, |
| "venue": "Proceedings of HLT-EMNLP", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Ben Taskar, Lacoste-Julien Simon, and Klein Dan. 2005. A discriminative matching approach to word align- ment. In Proceedings of HLT-EMNLP.", |
| "links": null |
| }, |
| "BIBREF24": { |
| "ref_id": "b24", |
| "title": "Efficient optimization of an MDL-inspired objective function for unsupervised part-of-speech tagging", |
| "authors": [ |
| { |
| "first": "Ashish", |
| "middle": [], |
| "last": "Vaswani", |
| "suffix": "" |
| }, |
| { |
| "first": "Adam", |
| "middle": [], |
| "last": "Pauls", |
| "suffix": "" |
| }, |
| { |
| "first": "David", |
| "middle": [], |
| "last": "Chiang", |
| "suffix": "" |
| } |
| ], |
| "year": 2010, |
| "venue": "Proceedings of ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Ashish Vaswani, Adam Pauls, and David Chiang. 2010. Efficient optimization of an MDL-inspired objective function for unsupervised part-of-speech tagging. In Proceedings of ACL.", |
| "links": null |
| }, |
| "BIBREF25": { |
| "ref_id": "b25", |
| "title": "HMM-based word alignment in statistical translation", |
| "authors": [ |
| { |
| "first": "Stephan", |
| "middle": [], |
| "last": "Vogel", |
| "suffix": "" |
| }, |
| { |
| "first": "Hermann", |
| "middle": [], |
| "last": "Ney", |
| "suffix": "" |
| }, |
| { |
| "first": "Christoph", |
| "middle": [], |
| "last": "Tillmann", |
| "suffix": "" |
| } |
| ], |
| "year": 1996, |
| "venue": "Proceedings of COLING", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Stephan Vogel, Hermann Ney, and Christoph Tillmann. 1996. HMM-based word alignment in statistical trans- lation. In Proceedings of COLING.", |
| "links": null |
| }, |
| "BIBREF26": { |
| "ref_id": "b26", |
| "title": "Interpreting BLEU/NIST scores: How much improvement do we need to have a better system", |
| "authors": [ |
| { |
| "first": "Ying", |
| "middle": [], |
| "last": "Zhang", |
| "suffix": "" |
| }, |
| { |
| "first": "Stephan", |
| "middle": [], |
| "last": "Vogel", |
| "suffix": "" |
| }, |
| { |
| "first": "Alex", |
| "middle": [], |
| "last": "Waibel", |
| "suffix": "" |
| } |
| ], |
| "year": 2004, |
| "venue": "Proceedings of LREC", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Ying Zhang, Stephan Vogel, and Alex Waibel. 2004. Interpreting BLEU/NIST scores: How much improve- ment do we need to have a better system? In Proceed- ings of LREC.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "text": "The \u2113 0 -norm (top curve) and smoothed approximations (below) for \u03b2 = 0.05, 0.1, 0.2.", |
| "num": null, |
| "type_str": "figure", |
| "uris": null |
| }, |
| "TABREF0": { |
| "text": "Table 1: Adding the \u2113 0 -norm to the IBM models improves both alignment and translation accuracy across four different language pairs. The word trans column also shows that the number of distinct word translations (i.e., the size of the lexical weighting table) is reduced. The\u03c6 sing.", |
| "content": "<table><tr><td>task</td><td>data (M)</td><td>system</td><td colspan=\"3\">align F1 (%) word trans (M)\u03c6 sing.</td><td>Bleu (%)</td></tr><tr><td/><td/><td/><td/><td/><td/><td>2008 2009 2010</td></tr><tr><td/><td/><td>baseline</td><td>73.2</td><td>3.5</td><td>6.2</td><td>28.7</td></tr><tr><td>Chi-Eng</td><td>9.6+12</td><td>\u2113 0 -norm</td><td>76.5</td><td>2.0</td><td>3.3</td><td>29.5</td></tr><tr><td/><td/><td>difference</td><td>+3.3</td><td>\u221243%</td><td colspan=\"2\">\u221247% +0.8</td></tr><tr><td/><td/><td>baseline</td><td>65.0</td><td>3.1</td><td>4.5</td><td>39.8 42.5</td></tr><tr><td colspan=\"2\">Ara-Eng 5.4+4.3</td><td>\u2113 0 -norm</td><td>70.8</td><td>1.8</td><td>1.8</td><td>41.1 43.7</td></tr><tr><td/><td/><td>difference</td><td>+5.9</td><td>\u221239%</td><td colspan=\"2\">\u221260% +1.3 +1.2</td></tr><tr><td/><td/><td>baseline</td><td>66.2</td><td>15</td><td>5.0</td><td>41.6 44.9</td></tr><tr><td>Ara-Eng</td><td>44+37</td><td>\u2113 0 -norm</td><td>71.8</td><td>7.9</td><td>1.8</td><td>42.5 45.3</td></tr><tr><td/><td/><td>difference</td><td>+5.6</td><td>\u221247%</td><td colspan=\"2\">\u221264% +0.9 +0.4</td></tr><tr><td/><td/><td>baseline</td><td/><td>1.7</td><td>4.5</td><td>25.3 * 29.8</td></tr><tr><td colspan=\"2\">Urd-Eng 1.7+1.5</td><td>\u2113 0 -norm</td><td/><td>1.2</td><td>2.2</td><td>25.9 * 31.2</td></tr><tr><td/><td/><td>difference</td><td/><td>\u221229%</td><td colspan=\"2\">\u221251% +0.6 * +1.4</td></tr><tr><td/><td/><td>baseline</td><td>65.6</td><td>1.5</td><td>3.0</td><td>17.3 18.0</td></tr><tr><td colspan=\"2\">Cze-Eng 2.1+2.3</td><td>\u2113 0 -norm</td><td>72.3</td><td>1.0</td><td>1.4</td><td>17.9 18.4</td></tr><tr><td/><td/><td>difference</td><td>+6.7</td><td>\u221233%</td><td>\u221253%</td><td>+0.6 +0.4</td></tr></table>", |
| "num": null, |
| "html": null, |
| "type_str": "table" |
| }, |
| "TABREF2": { |
| "text": "", |
| "content": "<table><tr><td colspan=\"4\">: Almost all hyperparameter settings achieve higher F-scores than the baseline IBM Model 4 and HMM model</td></tr><tr><td colspan=\"3\">for Arabic-English alignment (\u03b1 = 0).</td><td/></tr><tr><td/><td/><td colspan=\"2\">word classes?</td></tr><tr><td>direction</td><td>system</td><td>no</td><td>yes</td></tr><tr><td/><td>baseline</td><td>49.0</td><td>52.1</td></tr><tr><td>P( f | e)</td><td>\u2113 0 -norm</td><td>63.9</td><td>65.9</td></tr><tr><td/><td colspan=\"3\">difference +14.9 +13.8</td></tr><tr><td/><td>baseline</td><td>64.3</td><td>65.2</td></tr><tr><td>P(e | f )</td><td>\u2113 0 -norm</td><td>69.2</td><td>70.3</td></tr><tr><td/><td colspan=\"2\">difference +4.9</td><td>+5.1</td></tr></table>", |
| "num": null, |
| "html": null, |
| "type_str": "table" |
| } |
| } |
| } |
| } |
