emnlp emnlp2012 emnlp2012-42 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Wang Ling ; Joao Graca ; Isabel Trancoso ; Alan Black
Abstract: Phrase-based machine translation models have shown to yield better translations than Word-based models, since phrase pairs encode the contextual information that is needed for a more accurate translation. However, many phrase pairs do not encode any relevant context, which means that the translation event encoded in that phrase pair is led by smaller translation events that are independent from each other, and can be found on smaller phrase pairs, with little or no loss in translation accuracy. In this work, we propose a relative entropy model for translation models, that measures how likely a phrase pair encodes a translation event that is derivable using smaller translation events with similar probabilities. This model is then applied to phrase table pruning. Tests show that considerable amounts of phrase pairs can be excluded, without much impact on the transla- . tion quality. In fact, we show that better translations can be obtained using our pruned models, due to the compression of the search space during decoding.
Reference: text
sentIndex sentText sentNum sentScore
1 t aawcab@, icss Abstract Phrase-based machine translation models have shown to yield better translations than Word-based models, since phrase pairs encode the contextual information that is needed for a more accurate translation. [sent-5, score-0.879]
2 However, many phrase pairs do not encode any relevant context, which means that the translation event encoded in that phrase pair is led by smaller translation events that are independent from each other, and can be found on smaller phrase pairs, with little or no loss in translation accuracy. [sent-6, score-2.356]
3 In this work, we propose a relative entropy model for translation models, that measures how likely a phrase pair encodes a translation event that is derivable using smaller translation events with similar probabilities. [sent-7, score-1.509]
4 Tests show that considerable amounts of phrase pairs can be excluded, without much impact on the transla- . [sent-9, score-0.489]
5 , 2003) model n-to-m translations of n source words to m target words, which are encoded in phrase pairs and stored in the translation model. [sent-13, score-0.904]
6 On the other hand, in Phrase-based Models, we would have a phrase pair p(in the box, dentro da caixa) and p(in china, na china), where the words “in the box” and “in China” can be translated together to “dentro da caixa” and “na China”, which substantially reduces the ambiguity. [sent-21, score-0.848]
7 In this case, both the translation and language models contribute to find the best translation based on the local context, which generally leads to better translations. [sent-22, score-0.504]
8 , 2010) does not discriminate which phrases pairs encode contextual information, and extracts all phrase pairs with consistent alignments. [sent-26, score-0.681]
9 This is undesirable because we are populating translation models with redundant phrase pairs, whose translations can be obtained using com962 PLraoncge uadgineg Lse oafr tnhineg 2,0 p1a2g Jeosin 96t C2–o9n7f1e,re Jnecjue Iosnla Enmd,p Kiroicraela, M 1e2t–h1o4ds Ju ilny N 20a1tu2r. [sent-28, score-0.762]
10 For backoff language models, multiple pruning strategies based on relative entropy have been proposed (Seymore and Rosenfeld, 1996) (Stolcke, 1998), where the objective is to prune n-grams in a way to minimize the relative entropy between the model before and after pruning. [sent-32, score-0.666]
11 While the concept of using relative entropy for pruning is not new and frequently used in backoff language models, there are no such models for machine translation. [sent-33, score-0.494]
12 Thus, the main contribution of our work is to propose a relative entropy pruning model for translation models used in Phrase-based Machine Translation. [sent-34, score-0.691]
13 It is shown that our pruning algorithm can eliminate phrase pairs with little or no impact in the predictions made in our translation model. [sent-35, score-1.068]
14 We perform experiments with our pruning algorithm based on phrase pair independence and analyse the results in section 5. [sent-41, score-0.8]
15 2 Phrase Table Pruning Phrase table pruning algorithms are important in translation, since they efficiently reduce the size of the translation model, without having a large negative impact in the translation quality. [sent-43, score-0.859]
16 This is especially relevant in environments where memory constraints are imposed, such as translation systems for small devices like cellphones, and also when time constraints for the translation are defined, such as online Speech-to-Speech systems. [sent-44, score-0.504]
17 1 Significance Pruning A relevant reference in phrase table pruning is the work of (Johnson and Martin, 2007), where it is shown that a significant portion of the phrase table can be discarded without a considerable negative impact on translation quality, or even positive one. [sent-46, score-1.309]
18 This work computes the probability, named p-value, that the joint occurrence event of the source phrase s and target phrase t occurring in same sentence pair happens by chance, and are actually statistically independent. [sent-47, score-0.906]
19 , 2009), where phrase pairs are treated discriminately based on their complexity. [sent-50, score-0.461]
20 Our work has a similar objective, but instead of trying to predict the independence between the source and target phrases in each phrase pair, we attempt to predict the independence between a phrase pair and other phrase pairs in the model. [sent-52, score-1.398]
21 This algorithm decodes the training corpora and extracts the number of times each phrase pair is used in the 1-best translation hypothesis. [sent-55, score-0.691]
22 Thus, phrase pairs that are rarely used during decoding are excluded first during pruning. [sent-56, score-0.512]
23 This method considers the relationship between phrase pairs in the model, since it tests whether the decoder is more prone to use some phrase pairs than others. [sent-57, score-1.016]
24 This approach will lean towards pruning the phrase pairs in the second hypothesis, since the decoder will use the first hypothesis. [sent-60, score-0.845]
25 This is generally not desired, since the 2 smaller phrase pairs can be used to trans- late the same source sentence with a small probability loss (5%), even if the longer phrase is pruned. [sent-61, score-0.97]
26 On the other hand, if the smaller phrases are pruned, the longer phrase can not be used to translate smaller chunks, such as “the key in Portugal”. [sent-62, score-0.559]
27 This matter is aggravated due to the fact that the training corpora is used to decode, so longer phrase pairs will be used more frequently than when translating unseen sentences, which will make the model more biased into pruning shorter phrase pairs. [sent-63, score-1.252]
28 3 Relative Entropy Model For Phrase-based Translation Models In this section, we shall define our entropy model for phrase pairs. [sent-64, score-0.434]
29 We start by introducing some notation to distinguish different types of phrase pairs and show why some phrase pairs are more redundant than others. [sent-65, score-0.95]
30 Afterwards, we illustrate our notion of relative entropy between phrase pairs. [sent-66, score-0.463]
31 Then, we describe our entropy model, its computation and its application to phrase table pruning. [sent-67, score-0.434]
32 1 Atomic and Composite Phrase Pairs We discriminate between 2 types of phrase pairs: atomic phrase pairs and composite phrase pairs. [sent-69, score-1.334]
33 Atomic phrase pairs define the smallest translation units, such that given an atomic phrase pair that translates from s to t, the same translation cannot be obtained using any combination of other phrase pairs. [sent-70, score-1.869]
34 Removing these phrase pairs reduces the range of translations that our model is capable of translating and also the possible translations. [sent-71, score-0.645]
35 Composite phrase pairs define translations of a given sequence of words that can also be obtained using atomic or other smaller composite phrase pairs. [sent-72, score-1.133]
36 Removing these phrase pairs does not change the amount of sentences that the model can translate, since all translations encoded in these phrases can still be translated using other phrases, but these will lead to different translation probabilities. [sent-74, score-1.079]
37 Considering table 1, we can see that atomic phrases encode one elementary translation event, while composite phrases encode joint events that are encoded in atomic phrase pairs. [sent-75, score-1.146]
38 If we look at the source phrase “in”, there is a multitude of possible translations for this word in most target languages. [sent-76, score-0.486]
39 Thus, we define the joint event of “in” translating to “em” (A1) and “Portugal” to “Portugal” (B1), denoted as A1 ∩ B1, in the phrase pair p(in Portugal, em Portugal). [sent-79, score-0.681]
40 Without this phrase pair it is assumed that these are independent events with probability given by P(A1)P(B1) 1, which would be 10%, leading to a 60% reduction. [sent-80, score-0.553]
41 By definition, if the presence of phrase p(John, John) does not influence the translation of p(in, em) and viceversa, we can say that probability of the joint event P(A1 ∩ C1) is equal to the product of the probabilities of th∩e Cevents P(A1)P(C1). [sent-83, score-0.726]
42 In another words, the model’s predictions even, without this phrase pair will remain the same. [sent-85, score-0.465]
43 The example above shows an extreme case, where the event encoded in the phrase pair p(John in, John em) is decomposed into independent events, and can be removed without changing the model’s prediction. [sent-86, score-0.575]
44 However, finding and pruning phrase pairs that are independent, based on smaller events is impractical, since most translation events are not strictly independent. [sent-87, score-1.233]
45 However, many phrase pairs can be replaced with derivations using smaller phrases with a small loss in the model’s pre1For simplicity, model is used we assume at this stage that no reordering Table 1: Phrase Translation Table with associated events dictions. [sent-88, score-0.771]
46 Hence, we would like to define a metric for phrase pairs that allows us evaluate how discarding each phrase pair will affect the pruned model’s predictions. [sent-89, score-1.021]
47 By removing phrase pairs that can be derived using smaller phrase pairs with similar probability, × × it is possible to discard a significant portion of the translation model, while minimizing the impact on the model’s predictions. [sent-90, score-1.253]
48 2 Relative Entropy Model for Machine Translation For each phrase pair pa, we define the supporting set SP(pa(s, t)) = S1, . [sent-92, score-0.439]
49 As expected, the event encoded in the phrase pair p itself is A1∩ B1∩ C1, which assumes that A1, B1 and C1 are all dependent. [sent-108, score-0.575]
50 We can see that if any of the events S1, S2 or S3 has a “similar probability” as the event coded in the phrase pair, we can remove this phrase pair with a minimal impact in the phrase prediction. [sent-109, score-1.32]
51 Our objective is to minimize D(Pp| |P), which can be done locally by removing phrase pairs p(s, t) −P(s,t)logPPp((tt||ss)). [sent-112, score-0.461]
52 However, minimizing such an objective function would be intractable due to reordering, since the probability assigned to a phrase pair in a sentence pair by each model would depend on the positioning of all other phrase pairs used in the sentence. [sent-114, score-1.031]
53 Thus, we assume that all phrase pairs have the same probability regardless of their context in a sentence. [sent-116, score-0.504]
54 Using this distribution, the model is more biased in pruning phrase pairs with s, t pairs that do not occur frequently. [sent-125, score-0.898]
55 Using the unpruned model, the possible derivations are either using phrase p(s, t) or one element of its support set S1, S2 or S3. [sent-133, score-0.585]
56 First, if the probability ofthe phrase pairp(s, t) is lower than the highest probability element in SP(p(s, t)), then both the models will choose that element, in which case, = 1. [sent-136, score-0.496]
57 Secondly, if the probability of p(s, t) is equal to the most likely element in SP(p(s, t)), regardless of whether the unpruned model choses to use p(s, t) or that element, the probability emissions of the pruned and unpruned model will be identical. [sent-139, score-0.476]
58 1 Translation Model The translation model in Moses is composed by a phrase translation model and a phrase reordering model. [sent-146, score-1.314]
59 The first one models, for each phrase pair p(s, t), the probability of translating the s to t by combining multiple features φi, weighted by φi(p)wTi. [sent-147, score-0.567]
60 wiT, as PT(p) = Qin=1 The reordering model is similar, butQ models the local reordering between p, given the previous and next phrase accord- PinRg( tpo|p tPhe, tpaNrg)e =t siQdeim,= p1Pψia(npd|p pPN,p,P o)rw miRore formally, 4. [sent-148, score-0.501]
61 In practice, we are not searching in theP space Po(ft possible translations, bout ts eianr tchhespace of possible derivations, which are sequences of phrase translations p1(s1, t1) , . [sent-154, score-0.45]
62 Secondly, the score of a given translation hypothesis does not depend on the language model probability P(t), since all derivations in this search space have the same t, thus we discard this probability from the score function. [sent-160, score-0.477]
63 This is possible, since phrase pairs are tˆ tˆ generally smaller than text (less than 8 words), and because we are constraining the search space to t, which is an order of magnitude smaller than the reg967 ular search space with all possible translations. [sent-162, score-0.611]
64 Algorithm 1 Independence Pruning Algorithm 1Independence Pruning Require: pruning threshold δ, unpruned model {p1(s1, t1) , . [sent-171, score-0.432]
65 While this appears relatively simple and similar to a document decoding task, the size of our task is on a different order of magnitude, since we need to decode every phrase pair in the translation model, which might not be tractable for large models with millions of phrase pairs. [sent-178, score-1.131]
66 Another problem with this algorithm is that the decision to prune each phrase pair is made assuming that all other phrase pairs will remain in the model. [sent-181, score-0.986]
67 Thus, there is a chance a phrase pair p1 is pruned because of a derivation using p2 and p3 that leads to the same translation. [sent-182, score-0.604]
68 One possible solution to address this problem is to perform pruning iteratively, from the smallest phrase pairs (number of words) and increase the size at each iteration. [sent-184, score-0.788]
69 However, we find this undesirable, since the model will be biased into removing smaller phrase pairs, which are generally more useful, since they can be used in multiple derivation to replace larger phrase pairs. [sent-185, score-0.797]
70 In the example above, the model would eliminate p3 and keep p1, yet the best decision could be to keep p3 and remove p1, if p3 is also frequently used in derivations of other phrase pairs. [sent-186, score-0.421]
71 In both experiments, the translation model was built using the phrase extraction algorithm (Paul et al. [sent-205, score-0.603]
72 3 Pruning Setup Our pruning algorithm is applied after the translation model weight optimization with MERT. [sent-209, score-0.579]
73 com/p/geppetto/ 968 ate multiple translation models by setting different values for δ, so that translation models of different sizes are generated at intervals of 5%. [sent-212, score-0.538]
74 While the IWSLT translation model has only 88,424 phrase pairs, for the EUROPARL experiment, the translation model was composed by 48,762,372 phrase pairs, which had to be decoded. [sent-214, score-1.239]
75 The average time to decode each phrase pair using the full translation model is 4 seconds per sentence, since the table must be read from disk due to its size. [sent-215, score-0.729]
76 To address this problem, we divide the phrase pairs in the translation model into blocks of K phrase pairs, that are processed separately. [sent-217, score-1.13]
77 For each block, we resort to the approach used in MERT tuning, where the model is filtered to only include the phrase pairs that are used for translating tuning sentences. [sent-218, score-0.571]
78 We filter each block with phrase pairs from K to 2K with the source sentences sK, . [sent-219, score-0.525]
79 We used blocks of 10,000 phrase pairs and each filtered table was reduced to less than 1% of the translation table on average, reducing the average decoding time to 0. [sent-227, score-0.83]
80 4 Results Figure 1 shows the BLEU results for different sizes of the translation model for the IWSLT experiment using the uniform and multinomial distributions for P(s, t) . [sent-231, score-0.446]
81 54, where only atomic phrase pairs remain and both the multinomial and uniform distribution have the same performance, obviously. [sent-237, score-0.702]
82 parison between the uniform and multinomial distribution, we can see that both distributions yield similar results, specially when a low number of phrase pairs is pruned. [sent-244, score-0.621]
83 In theory, the multinomial distribution should yield better results, since the pruning model will prefer to prune phrase pairs that are more likely to be observed. [sent-245, score-0.926]
84 However, longer phrase pairs, which tend compete with other long phrase pairs on which get pruned first. [sent-246, score-0.961]
85 These phrase pairs generally occur only once or twice, so the multinomial model will act similarly to the uniform model regarding longer phrase pairs. [sent-247, score-0.976]
86 The comparison between our pruning model and pruning based on significance is shown in table 2. [sent-249, score-0.701]
87 For instance, the significance metric can either keep or remove all phrase pairs that only appear once, leaving a large gap of phrase table sizes that cannot be attained. [sent-251, score-0.893]
88 The same happens with our metric that cannot distinguish atomic phrase pairs. [sent-253, score-0.43]
89 In the EUROPARL experiment, we cannot generate phrase tables with sizes smaller than 15%. [sent-254, score-0.436]
90 Significant improvements are observed in the 969 Table 2: Comparison between Significance Pruning (Significance Pruning) and Entropy-based pruning using the uniform (Entropy (u) Pruning) and multinomial distributions (Entropy (m) Pruning). [sent-256, score-0.487]
91 In our metric, we do not have any means for detecting spurious phrase pairs, in fact, spurious phrase pairs are probably kept in the phrase table, since each distinct spurious phrase pair is only extracted once, and thus, they have very few derivations in its support set. [sent-262, score-1.825]
92 We believe that in language pairs such as ChineseEnglish with large distance reorderings between phrases are more prone to search errors and benefit more from our pruning algorithm. [sent-265, score-0.541]
93 By pruning the hypothesis “married in” a priori, we contribute in pre- venting such search errors. [sent-278, score-0.396]
94 We observe that some categories of phrase pairs that are systematically pruned, but these cannot be generalized in rules, since there are many exceptions. [sent-279, score-0.461]
95 The most obvious type of phrase pairs are phrases with punctuations, such as “谢谢. [sent-280, score-0.531]
96 Thus, phrase pairs for number sequences can be removed, since those numbers can be translated one by one. [sent-292, score-0.629]
97 However, for sequences such as “一(1) 八(8)”, we need a phrase pair to represent this specifically. [sent-293, score-0.439]
98 6 Conclusions We present a pruning algorithm for Machine Translation based on relative entropy, where we assess whether the translation event encoded in a phrase pair can be decomposed into combinations of events encoded in other phrase pairs. [sent-296, score-1.661]
99 We show that such phrase pairs can be removed from the translation model with little negative impact or even a positive one in the overall translation quality. [sent-297, score-0.993]
100 We also plan to improve the pruning step ofour algorithm to find the optimal set of phrase pairs to prune given the pruning threshold. [sent-300, score-1.175]
wordName wordTfidf (topN-words)
[('portugal', 0.498), ('phrase', 0.351), ('pruning', 0.327), ('translation', 0.252), ('translated', 0.168), ('dentro', 0.165), ('married', 0.158), ('iwslt', 0.14), ('pruned', 0.121), ('pairs', 0.11), ('europarl', 0.107), ('unpruned', 0.105), ('translations', 0.099), ('sp', 0.093), ('composite', 0.092), ('chave', 0.091), ('john', 0.089), ('pair', 0.088), ('translating', 0.085), ('entropy', 0.083), ('event', 0.08), ('atomic', 0.079), ('multinomial', 0.078), ('em', 0.077), ('na', 0.076), ('reordering', 0.075), ('hid', 0.073), ('events', 0.071), ('ppp', 0.071), ('derivations', 0.07), ('pi', 0.068), ('blocks', 0.066), ('china', 0.061), ('prune', 0.06), ('element', 0.059), ('uniform', 0.058), ('decoder', 0.057), ('encoded', 0.056), ('backoff', 0.055), ('escondeu', 0.055), ('logppp', 0.055), ('decoding', 0.051), ('moses', 0.051), ('spurious', 0.051), ('smaller', 0.051), ('significance', 0.047), ('ling', 0.047), ('hypothesis', 0.045), ('argmax', 0.045), ('derivation', 0.044), ('phrases', 0.043), ('probability', 0.043), ('gra', 0.042), ('tn', 0.041), ('encode', 0.04), ('box', 0.039), ('tt', 0.039), ('decode', 0.038), ('prone', 0.037), ('argmpax', 0.037), ('caixa', 0.037), ('fct', 0.037), ('mtsummit', 0.037), ('seymore', 0.037), ('source', 0.036), ('sn', 0.035), ('translate', 0.035), ('translates', 0.035), ('sizes', 0.034), ('independence', 0.034), ('eight', 0.033), ('composed', 0.033), ('undesirable', 0.032), ('tomeh', 0.031), ('eedd', 0.031), ('isabel', 0.031), ('calculating', 0.031), ('koehn', 0.03), ('ss', 0.029), ('pn', 0.029), ('relative', 0.029), ('eleven', 0.028), ('redundant', 0.028), ('jo', 0.028), ('block', 0.028), ('longer', 0.028), ('impact', 0.028), ('contextual', 0.027), ('obvious', 0.027), ('remain', 0.026), ('eck', 0.026), ('matthias', 0.026), ('pa', 0.026), ('si', 0.026), ('ps', 0.026), ('bleu', 0.025), ('tuning', 0.025), ('howard', 0.025), ('search', 0.024), ('distributions', 0.024)]
simIndex simValue paperId paperTitle
same-paper 1 0.9999994 42 emnlp-2012-Entropy-based Pruning for Phrase-based Machine Translation
Author: Wang Ling ; Joao Graca ; Isabel Trancoso ; Alan Black
Abstract: Phrase-based machine translation models have shown to yield better translations than Word-based models, since phrase pairs encode the contextual information that is needed for a more accurate translation. However, many phrase pairs do not encode any relevant context, which means that the translation event encoded in that phrase pair is led by smaller translation events that are independent from each other, and can be found on smaller phrase pairs, with little or no loss in translation accuracy. In this work, we propose a relative entropy model for translation models, that measures how likely a phrase pair encodes a translation event that is derivable using smaller translation events with similar probabilities. This model is then applied to phrase table pruning. Tests show that considerable amounts of phrase pairs can be excluded, without much impact on the transla- . tion quality. In fact, we show that better translations can be obtained using our pruned models, due to the compression of the search space during decoding.
2 0.48879677 11 emnlp-2012-A Systematic Comparison of Phrase Table Pruning Techniques
Author: Richard Zens ; Daisy Stanton ; Peng Xu
Abstract: When trained on very large parallel corpora, the phrase table component of a machine translation system grows to consume vast computational resources. In this paper, we introduce a novel pruning criterion that places phrase table pruning on a sound theoretical foundation. Systematic experiments on four language pairs under various data conditions show that our principled approach is superior to existing ad hoc pruning methods.
3 0.14120513 67 emnlp-2012-Inducing a Discriminative Parser to Optimize Machine Translation Reordering
Author: Graham Neubig ; Taro Watanabe ; Shinsuke Mori
Abstract: This paper proposes a method for learning a discriminative parser for machine translation reordering using only aligned parallel text. This is done by treating the parser’s derivation tree as a latent variable in a model that is trained to maximize reordering accuracy. We demonstrate that efficient large-margin training is possible by showing that two measures of reordering accuracy can be factored over the parse tree. Using this model in the pre-ordering framework results in significant gains in translation accuracy over standard phrasebased SMT and previously proposed unsupervised syntax induction methods.
4 0.13682316 35 emnlp-2012-Document-Wide Decoding for Phrase-Based Statistical Machine Translation
Author: Christian Hardmeier ; Joakim Nivre ; Jorg Tiedemann
Abstract: Independence between sentences is an assumption deeply entrenched in the models and algorithms used for statistical machine translation (SMT), particularly in the popular dynamic programming beam search decoding algorithm. This restriction is an obstacle to research on more sophisticated discourse-level models for SMT. We propose a stochastic local search decoding method for phrase-based SMT, which permits free document-wide dependencies in the models. We explore the stability and the search parameters ofthis method and demonstrate that it can be successfully used to optimise a document-level semantic language model. 1 Motivation In the field oftranslation studies, it is undisputed that discourse-wide context must be considered care- fully for good translation results (Hatim and Mason, 1990). By contrast, the state of the art in statistical machine translation (SMT), despite significant advances in the last twenty years, still assumes that texts can be translated sentence by sentence under strict independence assumptions, even though it is well known that certain linguistic phenomena such as pronominal anaphora cannot be translated correctly without referring to extra-sentential context. This is true both for the phrase-based and the syntaxbased approach to SMT. In the rest of this paper, we shall concentrate on phrase-based SMT. One reason why it is difficult to experiment with document-wide models for phrase-based SMT is that the dynamic programming (DP) algorithm 1179 which has been used almost exclusively for decoding SMT models in the recent literature has very strong assumptions of locality built into it. DP beam search for phrase-based SMT was described by Koehn et al. (2003), extending earlier work on word-based SMT (Tillmann et al., 1997; Och et al., 2001 ; Tillmann and Ney, 2003). This algorithm con- structs output sentences by starting with an empty hypothesis and adding output words at the end until translations for all source words have been generated. The core models of phrase-based SMT, in particular the n-gram language model (LM), only depend on a constant number of output words to the left of the word being generated. This fact is exploited by the search algorithm with a DP technique called hypothesis recombination (Och et al., 2001), which permits the elimination of hypotheses from the search space if they coincide in a certain number of final words with a better hypothesis and no future expansion can possibly invert the relative ranking of the two hypotheses under the given models. Hypothesis recombination achieves a substantial reduction of the search space without affecting search optimality and makes it possible to use aggressive pruning techniques for fast search while still obtaining good results. The downside of this otherwise excellent approach is that it only works well with models that have a local dependency structure similar to that of an n-gram language model, so they only depend on a small context window for each target word. Sentence-local models with longer dependencies can be added, but doing so greatly increases the risk for search errors by inhibiting hypothesis recombination. Cross-sentence dependencies cannot be directly integrated into DP SMT decoding in LParnogcue agdein Lgesa ornf tihneg, 2 p0a1g2e Jso 1in17t C9–o1n1f9e0re,n Jce ju on Is Elanmdp,ir Kicoarlea M,e 1t2h–o1d4s J iunly N 2a0tu1r2a.l ? Lc a2n0g1u2ag Aes Psorcoicaetsiosin fgo arn Cdo Cmopmutpauti oantiaoln Lailn Ngautiustriacls any obvious way, especially if joint optimisation of a number of interdependent decisions over an entire document is required. Research into models with a more varied, non-local dependency structure is to some extent stifled by the difficulty of decoding such models effectively, as can be seen by the problems some researchers encountered when they attempted to solve discourse-level problems. Consider, for instance, the work on cache-based language models by Tiedemann (2010) and Gong et al. (201 1), where error propagation was a serious issue, or the works on pronominal anaphora by Le Nagard and Koehn (2010), who implemented cross-sentence dependencies with an ad-hoc two-pass decoding strategy, and Hardmeier and Federico (2010) with the use of an external decoder driver to manage backward-only dependencies between sentences. In this paper, we present a method for decoding complete documents in phrase-based SMT. Our decoder uses a local search approach whose state consists of a complete translation of an entire document at any time. The initial state is improved by the application of a series of operations using a hill climbing strategy to find a (local) maximum of the score function. This setup gives us complete freedom to define scoring functions over the entire document. Moreover, by optionally initialising the state with the output of a traditional DP decoder, we can ensure that the final hypothesis is no worse than what would have been found by DP search alone. We start by describing the decoding algorithm and the state operations used by our decoder, then we present empirical results demonstrating the effectiveness of our approach and its usability with a document-level semantic language model, and finally we discuss some related work. 2 SMT Decoding by Hill Climbing In this section, we formally describe the phrasebased SMT model implemented by our decoder as well as the decoding algorithm we use. 2.1 SMT Model Our decoder is based on local search, so its state at any time is a representation of a complete translation of the entire document. Even though the decoder operates at the document level, it is important to keep 1180 track of sentence boundaries, and the individual operations that are applied to the state are still confined to sentence scope, so it is useful to decompose the state of a document into the state of its sentences, and we define the overall state S as a sequence of sentence states: S = S1S2 . . .SN, (1) where N is the number of sentences. This implies that we constrain the decoder to emit exactly one output sentence per input sentence. Let ibe the number of a sentence and mi the number of input tokens of this sentence, p and q (with 1 ≤ p ≤ q ≤ mi) be positions in the input sentence a1n ≤d [p; q] qde ≤no mte the set ofpositions from p up to and including q. We say that [p; q] precedes [p0; q0], or [p; q] ≺ [p0; q0], if q < p0. Let Φi([p; q]) be the set of t[pra;nqs]l ≺atio [pns for the source phrase covering positions [p; q] in the input sentence ias given by the phrase table. We call A = h[p; q] ,φi an anchored phrase pair w.it Wh coverage C(A) = [p; q] nif a φ ∈ Φi([p; q]) sise a target phrase translating =th [ep source w∈o Φrds at positions [p; q] . Then a sequence of ni anchored phrase pairs Si = A1A2 . . .Ani (2) is a valid sentence state for sentence iif the following two conditions hold: 1. The coverage sets C(Aj) for j in 1, . . . , ni are mutually disjoint, and 2. the anchored phrase pairs jointly cover the complete input sentence, or [niC(Aj) = [1;mi]. (3) [j=1 Let f(S) be a scoring function mapping a state S to a real number. As usual in SMT, it is assumed that the scoring function can be decomposed into a linear combination of K feature functions hk(S), each with a constant weight λk, so f(S) =k∑K=1λkhk(S). (4) The problem addressed by the decoder is the search for the state with maximal score, such that Sˆ Sˆ = argSmaxf(S). (5) The feature functions implemented in our baseline system are identical to the ones found in the popular Moses SMT system (Koehn et al., 2007). In particular, our decoder has the following feature functions: 1. phrase translation scores provided by the phrase table including forward and backward conditional probabilities, lexical weights and a phrase penalty (Koehn et al., 2003), 2. n-gram language model scores implemented with the KenLM toolkit (Heafield, 2011), 3. a word penalty score, 4. a distortion model with geometric (Koehn et al., 2003), and decay 5. a feature indicating the number of times a given distortion limit is exceeded in the current state. In our experiments, the last feature is used with a fixed weight of negative infinity in order to limit the gaps between the coverage sets of adjacent anchored phrase pairs to a maximum value. In DP search, the distortion limit is usually enforced directly by the search algorithm and is not added as a feature. In our decoder, however, this restriction is not required to limit complexity, so we decided to add it among the scoring models. 2.2 Decoding Algorithm The decoding algorithm we use (algorithm 1) is very simple. It starts with a given initial document state. In the main loop, which extends from line 3 to line 12, it generates a successor state S0 for the current state S by calling the function Neighbour, which non-deterministically applies one of the operations described in section 3 of this paper to S. The score of the new state is compared to that of the previous one. If it meets a given acceptance criterion, S0 becomes the current state, else search continues from the previous state S. For the experiments in this paper, we use the hill climbing acceptance criterion, which simply accepts a new state if its score is higher than that of the current state. Other acceptance criteria are possible and could be used to endow the search algorithm with stochastic behaviour. 1181 The main loop is repeated until a maximum number of steps (step limit) is reached or until a maximum number of moves are rejected in a row (rejection limit). Algorithm 1 Decoding algorithm Input: an initial document state S; search parameters maxsteps and maxrejected Output: a modified document state 1: nsteps ← 0 2: nrejected ← 0 3: nwrhejileec nsteps < maxsteps and nrejected < maxrejected do 4: S0 ← Neighbour (S) 5: if Accept (f(S0) , f(S)) then 6: S ← S0 7: nrejected ← 0 8: elsner 9: nrejected ← nrejected + 1 10: enndr eifj 11: nsteps ← nsteps + 1 12: ennds wtephsile ← 13: return S A notable difference between this algorithm and other hill climbing algorithms that have been used for SMT decoding (Germann et al., 2004; Langlais et al., 2007) is its non-determinism. Previous work for sentence-level decoding employed a steepest ascent strategy which amounts to enumerating the complete neighbourhood of the current state as defined by the state operations and selecting the next state to be the best state found in the neighbourhood of the current one. Enumerating all neighbours of a given state, costly as it is, has the advantage that it makes it easy to prove local optimality of a state by recognising that all possible successor states have lower scores. It can be rather inefficient, since at every step only one modification will be adopted; many of the modifications that are discarded will very likely be generated anew in the next iteration. As we extend the decoder to the document level, the size of the neighbourhood that would have to be explored in this way increases considerably. Moreover, the inefficiency of the steepest ascent approach potentially increases as well. Very likely, a promising move in one sentence will remain promising after a modification has been applied to another sentence, even though this is not guaranteed to be true in the presence of cross-sentence models. We therefore adopt a first-choice hill climbing strategy that non-deterministically generates successor states and accepts the first one that meets the acceptance criterion. This frees us from the necessity of generating the full set of successors for each state. On the downside, if the full successor set is not known, it is no longer possible to prove local optimality of a state, so we are forced to use a different condition for halting the search. We use a combination of two limits: The step limit is a hard limit on the resources the user is willing to expend on the search problem. The value of the rejection limit determines how much of the neighbourhood is searched for better successors before a state is accepted as a solution; it is related to the probability that a state returned as a solution is in fact locally optimal. To simplify notations in the description of the individual state operations, we write Si −→ Si0 (6) to signify that a state operation, when presented with a document state as in equation 1 and acting on sentence i, returns a new document state of S0 = S1 . . .Si−1 Si0 Si+1 . . .SN. (7) Similarly, Si : Aj . . .Aj+h−1 −→ A01 . . .A0h0 (8) is equivalent to Si −→ A1 . . .Aj−1 A01 . . .A0h0 Aj+h . . .Ani (9) and indicates that the operation returns a state in which a sequence of h consecutive anchored phrase pairs has been replaced by another sequence of h0 anchored phrase pairs. 2.3 Efficiency Considerations When implementing the feature functions for the decoder, we have to exercise some care to avoid recomputing scores for the whole document at every iteration. To achieve this, the scores are computed completely only once, at the beginning of the decoding run. In subsequent iterations, scoring functions are presented with the scores of the previous 1182 iteration and a list of modifications produced by the state operation, a set of tuples hi, r, s,A01 . . .A0h0i, each indicating tthioant ,t ahe s edto ocfu tmupelnets s hhio,ru,sld, Abe modifii,e eda as described by Si :Ar . . .As −→ A01 . . .A0h0 . (10) If a feature function is decomposable in some way, as all the standard features developed under the constraints of DP search are, it can then update the state simply by subtracting and adding score components pertaining to the modified parts of the document. Feature functions have the possibility to store their own state information along with the document state to make sure the required information is available. Thus, the framework makes it possible to exploit decomposability for efficient scoring without impos- ing any particular decomposition on the features as beam search does. To make scoring even more efficient, scores are computed in two passes: First, every feature function is asked to provide an upper bound on the score that will be obtained for the new state. In some cases, it is possible to calculate reasonable upper bounds much more efficiently than computing the exact feature value. If the upper bound fails to meet the acceptance criterion, the new state is discarded right away; if not, the full score is computed and the acceptance criterion is tested again. Among the basic SMT models, this two-pass strategy is only used for the n-gram LM, which requires fairly expensive parameter lookups for scoring. The scores of all the other baseline models are fully computed during the first scoring pass. The n-gram model is more complex. In its state information, it keeps track of the LM score and LM library state for each word. The first scoring pass then identifies the words whose LM scores are affected by the current search step. This includes the words changed by the search operation as well as the words whose LM history is modified. The range of the history de- pendencies can be determined precisely by considering the “valid state length” information provided by the KenLM library. In the first pass, the LM scores of the affected words are subtracted from the total score. The model only looks up the new LM scores for the affected words and updates the total score if the new search state passes the first acceptance check. This two-pass scoring approach allows us to avoid LM lookups altogether for states that will be rejected anyhow because of low scores from the other models, e. g. because the distortion limit is violated. Model score updates become more complex and slower as the number of dependencies of a model increases. While our decoding algorithm does not impose any formal restrictions on the number or type of dependencies that can be handled, there will be practical limits beyond which decoding becomes unacceptably slow or the scoring code becomes very difficult to maintain. These limits are however fairly independent of the types of dependencies handled by a model, which permits the exploration of more varied model types than those handled by DP search. 2.4 State Initialisation Before the hill climbing decoding algorithm can be run, an initial state must be generated. The closer the initial state is to an optimum, the less work remains to be done for the algorithm. If the algorithm is to be self-contained, initialisation must be relatively uninformed and can only rely on some general prior assumptions about what might be a good initial guess. On the other hand, if optimal results are sought after, it pays off to invest some effort into a good starting point. One way to do this is to run DP search first. For uninformed initialisation, we chose to implement a very simple procedure based only on the observation that, at least for language pairs involving the major European languages, it is usually a good guess to keep the word order of the output very similar to that of the input. We therefore create the initial state by selecting, for each sentence in the document, a sequence of anchored phrase pairs covering the input sentence in monotonic order, that is, such that for all pairs of adjacent anchored phrase pairs Aj and Aj+1, we have that C(Aj) ≺ C(Aj+1 ). For initialisation with DP search, we first run the Moses decoder (Koehn et al., 2007) with default search parameters and the same models as those used by our decoder. Then we extract the best output hypothesis from the search graph of the decoder and map it into a sequence of anchored phrase pairs in the obvious way. When the document-level decoder is used with models that are incompatible with beam search, Moses can be run with a subset of the models in order to find an approximation of the solution 1183 which is then refined with the complete feature set. 3 State Operations Given a document state S, the decoder uses a neighbourhood function Neighbour to simulate a move in the state space. The neighbourhood function nondeterministically selects a type of state operation and a location in the document to apply it to and returns the resulting new state. We use a set of three operations that has the property that every possible document state can be reached from every other state in a sequence of moves. Designing operations for state transitions in local search for phrase-based SMT is a problem that has been addressed in the literature (Langlais et al., 2007; Arun et al., 2010). Our decoder’s first- choice hill climbing strategy never enumerates the full neighbourhood of a state. We therefore place less emphasis than previous work on defining a compact neighbourhood, but allow the decoder to make quite extensive changes to a state in a single step with a certain probability. Otherwise our operations are similar to those used by Arun et al. (2010). All of the operations described in this paper make changes to a single sentence only. Each time it is called, the Neighbour function selects a sentence in the document with a probability proportional to the number of input tokens in each sentence to ensure a fair distribution ofthe decoder’s attention over the words in the document regardless of varying sentence lengths. 3.1 Changing Phrase Translations The change-phrase-translation operation replaces the translation of a single phrase with a random translation with the same coverage taken from the phrase table. Formally, the operation selects an anchored phrase pair Aj by drawing uniformly from the elements of Si and then draws a new translation φ0 uniformly from the set Φi(C(Aj)). The new state is given by Si : Aj −→ hC(Aj), φ0i. (11) 3.2 Changing Word Order The swap-phrases operation affects the output word order without changing the phrase translations. It exchanges two anchored phrase pairs Aj and Aj+h, resulting in an output state of Si : Aj . . .Aj+h −→ Aj+h Aj+1 . . .Aj+h−1 Aj. (12) The start location j is drawn uniformly from the eligible sentence positions; the swap range h comes from a geometric distribution with configurable decay. Other word-order changes such as a one-way move operation that does not require another movement in exchange or more advanced permutations can easily be defined. 3.3 Resegmentation The most complex operation is resegment, which allows the decoder to modify the segmentation ofthe source phrase. It takes a number of anchored phrase pairs that form a contiguous block both in the input and in the output and replaces them with a new set of phrase pairs covering the same span of the input sentence. Formally, Si : Aj . . .Aj+h−1 −→ A01 . . .A0h0 (13) such that j+[h−1 [h0 [ C(Aj0) = [ C(A0j0) = [p;q] j[0=j (14) j[0=1 for some p and q, where, for j0 = 1, . . . ,h0, we have that A0j0 = h[pj0; qj0] , φj0i, all [pj0; qj0] are mutually disjoint =an hd[ peach φj0 isi randomly drawn from Φi([pj0;qj0]). Regardless of the ordering of Aj . . .Aj+h−1 , the resegment operation always generates a sequence of anchored phrase pairs in linear order, such that C(A0j0) ≺ C(A0j0+1 ) for j0 = 1, . . . ,h0 −1 . As )f o≺r Cth(eA other operations, j is− generated uniformly and h is drawn from a geometric distribution with a decay parameter. The new segmentation is generated by extending the sequence of anchored phrase pairs with random elements starting at the next free position, proceeding from left to right until the whole range [p; q] is covered. 4 Experimental Results In this section, we present the results of a series of experiments with our document decoder. The 1184 goal of our experiments is to demonstrate the behaviour of the decoder and characterise its response to changes in the fundamental search parameters. The SMT models for our experiments were created with a subset of the training data for the English-French shared task at the WMT 2011workshop (Callison-Burch et al., 2011). The phrase table was trained on Europarl, news-commentary and UN data. To reduce the training data to a manageable size, singleton phrase pairs were removed before the phrase scoring step. Significance-based filtering (Johnson et al., 2007) was applied to the resulting phrase table. The language model was a 5gram model with Kneser-Ney smoothing trained on the monolingual News corpus with IRSTLM (Federico et al., 2008). Feature weights were trained with Minimum Error-Rate Training (MERT) (Och, 2003) on the news-test2008 development set using the DP beam search decoder and the MERT implementation of the Moses toolkit (Koehn et al., 2007). Experimental results are reported for the newstest2009 test set, a corpus of 111 newswire documents totalling 2,525 sentences or 65,595 English input tokens. 4.1 Stability An important difference between our decoder and the classical DP decoder as well as previous work in SMT decoding with local search is that our decoder is inherently non-deterministic. This implies that repeated runs of the decoder with the same search parameters, input and models will not, in general, find the same local maximum of the score space. The first empirical question we ask is therefore how different the results are under repeated runs. The results in this and the next section were obtained with random state initialisation, i. e. without running the DP beam search decoder. Figure 1 shows the results of 7 decoder runs with the models described above, translating the newstest2009 test set, with a step limit of 227 and a rejection limit of 100,000. The x-axis of both plots shows the number of decoding steps on a logarithmic scale, so the number of steps is doubled between two adjacent points on the same curve. In the left plot, the y-axis indicates the model score optimised by the decoder summed over all 2525 sentences of the document. In the right plot, the case-sensitive BLEU score (Papineni et al., 2002) of the current decoder Figure 1: Score stability in repeated decoder runs state against a reference translation is displayed. We note, as expected, that the decoder achieves a considerable improvement of the initial state with diminishing returns as decoding continues. Between 28 and 214 steps, the score increases at a roughly logarithmic pace, then the curve flattens out, which is partly due to the fact that decoding for some documents effectively stopped when the maximum number of rejections was reached. The BLEU score curve shows a similar increase, from an initial score below 5 % to a maximum of around 21.5 %. This is below the score of 22.45 % achieved by the beam search decoder with the same models, which is not surprising considering that our decoder approximates a more difficult search problem, from which a number of strong independence assumptions have been lifted, without, at the moment, having any stronger models at its disposal to exploit this additional freedom for better translation. In terms of stability, there are no dramatic differences between the decoder runs. Indeed, the small differences that exist are hardly discernible in the plots. The model scores at the end of the decoding run range between −158767.9 and −158716.9, a g re rlautniv rea ndgieffe breetnwceee nof − only a6b7.o9ut a n0d.0 −3 %15.8 F1i6n.a9l, BLEU scores range from 21.41 % to 21.63 %, an interval that is not negligible, but comparable to the variance observed when, e. g., feature weights from repeated MERT runs are used with one and the same SMT system. Note that these results were obtained with random state initialisation. With DP initialisation, score differences between repeated runs rarely 1185 exceed 0.02 absolute BLEU percentage points. Overall, we conclude that the decoding results of our algorithm are reasonably stable despite the nondeterminism inherent in the procedure. In our subsequent experiments, the evaluation scores reported are calculated as the mean of three runs for each experiment. 4.2 Search Algorithm Parameters The hill climbing algorithm we use has two parameters which govern the trade-off between decoding time and the accuracy with which a local maximum is identified: The step limit stops the search process after a certain number of steps regardless of the search progress made or lack thereof. The rejection limit stops the search after a certain number of unsuccessful attempts to make a step, when continued search does not seem to be promising. In most of our experiments, we used a step limit of 227 ≈ 1.3 · 108 and a rejection limit of 105. In practice, decoding terminates by reaching the rejection limit for the vast majority of documents. We therefore examined the effect of different rejection limits on the learning curves. The results are shown in figure 2. The results show that continued search does pay off to a certain extent. Indeed, the curve for rejection limit 107 seems to indicate that the model score increases roughly logarithmically, albeit to a higher base, even after the curve has started to flatten out at 214 steps. At a certain point, however, the probability of finding a good successor state drops rather sharply by about two orders of magnitude, as Figure 2: Search performance at different rejection limits evidenced by the fact that a rejection limit of 106 does not give a large improvement over one of 105, while one of 107 does. The continued model score improvement also results in an increase in BLEU scores, and with a BLEU score of 22. 1% the system with rejection limit 107 is fairly close to the score of 22.45 % obtained by DP beam search. Obviously, more exact search comes at a cost, and in this case, it comes at a considerable cost, which is an explosion of the time required to decode the test set from 4 minutes at rejection limit 103 to 224 minutes at rejection limit 105 and 38 hours 45 minutes at limit 107. The DP decoder takes 3 1 minutes for the same task. We conclude that the rejection limit of 105 selected for our experiments, while technically suboptimal, realises a good trade-off between decoding time and accuracy. 4.3 A Semantic Document Language Model In this section, we present the results of the application of our decoder to an actual SMT model with cross-sentence features. Our model addresses the problem of lexical cohesion. In particular, it rewards the use of semantically related words in the translation output by the decoder, where semantic distance is measured with a word space model based on Latent Semantic Analysis (LSA). LSA has been applied to semantic language modelling in previous research with some success (Coccaro and Jurafsky, 1998; Bellegarda, 2000; Wandmacher and Antoine, 2007). In SMT, it has mostly been used for domain adaptation (Kim and Khudanpur, 2004; Tam et al., 1186 2007), or to measure sentence similarities (Banchs and Costa-juss a`, 2011). The model we use is inspired by Bellegarda (2000). It is a Markov model, similar to a standard n-gram model, and assigns to each content word a score given a history of n preceding content words, where n = 30 below. Scoring relies on a 30dimensional LSA word vector space trained with the S-Space software (Jurgens and Stevens, 2010). The score is defined based on the cosine similarity between the word vector of the predicted word and the mean word vector of the words in the history, which is converted to a probability by histogram lookup as suggested by Bellegarda (2000). The model is structurally different from a regular n-gram model in that word vector n-grams are defined over content words occurring in the word vector model only and can cross sentence boundaries. Stop words, identified by an extensive stop word list and amounting to around 60 % of the tokens, are scored by a different mechanism based on their relative frequency (undiscounted unigram probability) in the training corpus. In sum, the score produced by the semantic document LM has the following form: wh(er|h)α=is tεpαheuncipgors(wp)o|hrtinof w fci os nakutneskotn wpon w ,onerldse,in(ls1teh5) training corpus and ε is a small fixed probability. It is integrated into the decoder as an extra feature function. Since we lack an automatic method for training the feature weights of document-wide features, its weight was selected by grid search over a number of values, comparing translation performance for the newstest2009 test set. In these experiments, we used DP beam search to initialise the state of our local search decoder. Three results are presented (table 1): The first table row shows the baseline performance using DP beam search with standard sentence-local features only. The scores in the second row were obtained by running the hill climbing decoder with DP initialisation, but without adding any models. A marginal increase in scores for all three test sets demonstrates that the hill climbing decoder manages to fix some of the search errors made by the DP search. The last row contains the scores obtained by adding in the semantic language model. Scores are presented for three publicly available test sets from recent WMT Machine Translation shared tasks, of which one (newstest2009) was used to monitor progress during development and select the final model. Adding the semantic language model results in a small increase in NIST scores (Doddington, 2002) for all three test sets as well as a small BLEU score gain (Papineni et al., 2002) for two out of three corpora. We note that the NIST score turned out to react more sensitively to improvements due to the semantic LM in all our experiments, which is reasonable because the model specifically targets content words, which benefit from the information weighting done by the NIST score. While the results we present do not constitute compelling evidence in favour of our semantic LM in its current form, they do suggest that this model could be improved to realise higher gains from cross-sentence semantic information. They support our claim that cross- sentence models should be examined more closely and that existing methods should be adapted to deal with them, a problem addressed by our main contribution, the local search document decoder. 5 Related Work Even though DP beam search (Koehn et al., 2003) has been the dominant approach to SMT decoding in recent years, methods based on local search have been explored at various times. For word-based SMT, greedy hill-climbing techniques were advo1187 cated as a faster replacement for beam search (Germann et al., 2001 ; Germann, 2003; Germann et al., 2004), and a problem formulation specifically targeting word reordering with an efficient word reordering algorithm has been proposed (Eisner and Tromble, 2006). A local search decoder has been advanced as a faster alternative to beam search also for phrasebased SMT (Langlais et al., 2007; Langlais et al., 2008). That work anticipates many of the features found in our decoder, including the use of local search to refine an initial hypothesis produced by DP beam search. The possibility of using models that do not fit well into the beam search paradigm is mentioned and illustrated with the example of a reversed n-gram language model, which the authors claim would be difficult to implement in a beam search decoder. Similarly to the work by Germann et al. (2001), their decoder is deterministic and explores the entire neighbourhood of a state in order to identify the most promising step. Our main contribution with respect to the work by Langlais et al. (2007) is the introduction of the possibility of handling document-level models by lifting the assumption of sentence independence. As a consequence, enumerating the entire neighbourhood becomes too expensive, which is why we resort to a “first-choice” strategy that non-deterministically generates states and accepts the first one encountered that meets the acceptance criterion. More recently, Gibbs sampling was proposed as a way to generate samples from the posterior distribution of a phrase-based SMT decoder (Arun et al., 2009; Arun et al., 2010), a process that resembles local search in its use of a set of state-modifying operators to generate a sequence of decoder states. Where local search seeks for the best state attainable from a given initial state, Gibbs sampling produces a representative sample from the posterior. Like all work on SMT decoding that we know of, the Gibbs sampler presented by Arun et al. (2010) assumes independence of sentences and considers the complete neighbourhood of each state before taking a sample. 6 Conclusion In the last twenty years of SMT research, there has been a strong assumption that sentences in a text newstest2009 newstest2010 newstest201 1 BLEU NIST BLEU NIST BLEU NIST 22.56 6.513 27.27 7.034 24.94 7.170 + hill climbing 22.60 6.518 27.33 7.046 24.97 7.169 with semantic LM 22.71 6.549 27.53 7.087 24.90 7.199 DP search only DP Table 1: Experimental results with a cross-sentence semantic language model are independent of one another, and discourse context has been largely neglected. Several factors have contributed to this. Developing good discourse-level models is difficult, and considering the modest translation quality that has long been achieved by SMT, there have been more pressing problems to solve and lower hanging fruit to pick. However, we argue that the popular DP beam search algorithm, which delivers excellent decoding performance, but imposes a particular kind of local dependency structure on the feature models, has also had its share in driving researchers away from discourse-level problems. In this paper, we have presented a decoding procedure for phrase-based SMT that makes it possible to define feature models with cross-sentence dependencies. Our algorithm can be combined with DP beam search to leverage the quality of the traditional approach with increased flexibility for models at the discourse level. We have presented preliminary results on a cross-sentence semantic language model addressing the problem of lexical cohesion to demonstrate that this kind of models is worth exploring further. Besides lexical cohesion, cross-sentence models are relevant for other linguistic phenomena such as pronominal anaphora or verb tense selection. We believe that SMT research has reached a point of maturity where discourse phenomena should not be ignored any longer, and we consider our decoder to be a step towards this goal. References Abhishek Arun, Chris Dyer, Barry Haddow, Phil Blunsom, Adam Lopez, and Philipp Koehn. 2009. Monte carlo inference and maximization for phrase-based translation. In Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL-2009), pages 102–1 10, Boulder, Colorado, June. Association for Computational Linguistics. Abhishek Arun, Barry Haddow, Philipp Koehn, Adam Lopez, Chris Dyer, and Phil Blunsom. 2010. Monte 1188 Ma- Carlo techniques for phrase-based translation. chine translation, 24(2): 103–121 . Rafael E. Banchs and Marta R. Costa-juss a`. 2011. A semantic feature for Statistical Machine Translation. In Proceedings of Fifth Workshop on Syntax, Semantics and Structure in Statistical Translation, pages 126– 134, Portland, Oregon, USA, June. Association for Computational Linguistics. Jerome R. Bellegarda. 2000. Exploiting latent semantic information in statistical language modeling. Proceedings of the IEEE, 88(8): 1279–1296. Chris Callison-Burch, Philipp Koehn, Christof Monz, and Omar Zaidan. 2011. Findings of the 2011 Workshop on Statistical Machine Translation. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 22–64, Edinburgh, Scotland, July. Association for Computational Linguistics. Noah Coccaro and Daniel Jurafsky. 1998. Towards better integration of semantic predictors in statistical language modeling. In Proceedings of the 5th International Conference on Spoken Language Processing, Sydney. George Doddington. 2002. Automatic evaluation of machine translation quality using n-gram co-occurrence statistics. In Proceedings of the second International conference on Human Language Technology Research, pages 138–145, San Diego. Jason Eisner and Roy W. Tromble. 2006. Local search with very large-scale neighborhoods for optimal permutations in machine translation. In Proceedings of the HLT-NAACL Workshop on Computationally Hard Problems and Joint Inference in Speech and Language Processing, pages 57–75. Marcello Federico, Nicola Bertoldi, and Mauro Cettolo. 2008. IRSTLM: an open source toolkit for handling large scale language models. In Interspeech 2008, pages 1618–1621 . ISCA. Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, and Kenji Yamada. 2001 . Fast decoding and optimal decoding for machine translation. In Proceedings of 39th Annual Meeting of the Association for Computational Linguistics, pages 228–235, Toulouse, France, July. Association for Computational Linguis- tics. Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, and Kenji Yamada. 2004. Fast and optimal decoding for machine translation. Artificial Intelligence, 154(1–2): 127–143. Ulrich Germann. 2003. Greedy decoding for Statistical Machine Translation in almost linear time. In Proceedings of the 2003 Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics. Zhengxian Gong, Min Zhang, and Guodong Zhou. 2011. Cache-based document-level Statistical Machine Translation. In Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 909–919, Edinburgh, Scotland, UK., July. Association for Computational Linguistics. Christian Hardmeier and Marcello Federico. 2010. Modelling Pronominal Anaphora in Statistical Machine Translation. In Proceedings of the seventh International Workshop on Spoken Language Translation (IWSLT), pages 283–289. Basil Hatim and Ian Mason. 1990. Discourse and the Translator. Language in Social Life Series. Longman, London. Kenneth Heafield. 2011. KenLM: faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July. Association for Computational Linguistics. Howard Johnson, Joel Martin, George Foster, and Roland Kuhn. 2007. Improving translation quality by discarding most of the phrasetable. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL), pages 967– 975, Prague, Czech Republic, June. Association for Computational Linguistics. David Jurgens and Keith Stevens. 2010. The S-Space package: An open source package for word space models. In Proceedings of the ACL 2010 System Demonstrations, pages 30–35, Uppsala, Sweden, July. Association for Computational Linguistics. Woosung Kim and Sanjeev Khudanpur. 2004. Crosslingual latent semantic analysis for language modeling. In IEEE international conference on acoustics, speech, and signal processing (ICASSP), volume 1, pages 257–260, Montr ´eal. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 conference of the North American chapter of the Association for Computational Linguistics on Human Language Technology, pages 48– 54, Edmonton. 1189 Philipp Koehn, Hieu Hoang, Alexandra Birch, et al. 2007. Moses: open source toolkit for Statistical Machine Translation. In Annual meeting of the Association for Computational Linguistics: Demonstration session, pages 177–180, Prague. Philippe Langlais, Alexandre Patry, and Fabrizio Gotti. 2007. A greedy decoder for phrase-based statistical machine translation. In TMI-2007: Proceedings of the 11th International Conference on Theoretical and Methodological Issues in Machine Translation, pages 104–1 13, Sk¨ ovde. Philippe Langlais, Alexandre Patry, and Fabrizio Gotti. 2008. Recherche locale pour la traduction statistique par segments. In TALN 2008, pages 119–128, Avignon, France, June. ATALA. Ronan Le Nagard and Philipp Koehn. 2010. Aiding pronoun translation with co-reference resolution. In Proceedings of the Joint Fifth Workshop on Statistical Machine Translation and MetricsMATR, pages 252–261, Uppsala, Sweden, July. Association for Computational Linguistics. Franz Josef Och, Nicola Ueffing, and Hermann Ney. 2001. An efficient A* search algorithm for Statistical Machine Translation. In Proceedings of the DataDriven Machine Translation Workshop, 39th Annual Meeting of the Association for Computational Linguistics (ACL), pages 55–62, Toulouse. Franz Josef Och. 2003. Minimum error rate training in Statistical Machine Translation. In Proceedings of the 41st annual meeting of the Association for Computational Linguistics, pages 160–167, Sapporo (Japan). Kishore Papineni, Salim Roukos, Todd Ward, and WeiJing Zhu. 2002. BLEU: a method for automatic evaluation of Machine Translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 3 11–3 18, Philadelphia. ACL. Yik-Cheung Tam, Ian Lane, and Tanja Schultz. 2007. Bilingual LSA-based adaptation for Statistical Machine Translation. Machine Translation, 21(4): 187– 207. J o¨rg Tiedemann. 2010. To cache or not to cache? Experiments with adaptive models in Statistical Machine Translation. In Proceedings of the ACL 2010 Joint Fifth Workshop on Statistical Machine Translation and Metrics MATR, pages 189–194, Uppsala, Sweden. Association for Computational Linguistics. Christoph Tillmann and Hermann Ney. 2003. Word reordering and a Dynamic Programming beam search algorithm for Statistical Machine Translation. Computational linguistics, 29(1):97–133. Christoph Tillmann, Stephan Vogel, Hermann Ney, and Alex Zubiaga. 1997. A DP-based search using monotone alignments in Statistical Translation. In Proceedings of the 35th Annual Meeting of the Association for Computational Spain, tics. Linguistics, July. Association for 289–296, Madrid, Computational Linguis- pages Tonio Wandmacher and Jean-Yves Antoine. 2007. Methods to integrate a language model with semantic information for a word prediction component. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLPCoNLL), pages 506–5 13, Prague, Czech Republic, June. Association for Computational Linguistics. 1190
5 0.13539718 74 emnlp-2012-Language Model Rest Costs and Space-Efficient Storage
Author: Kenneth Heafield ; Philipp Koehn ; Alon Lavie
Abstract: Approximate search algorithms, such as cube pruning in syntactic machine translation, rely on the language model to estimate probabilities of sentence fragments. We contribute two changes that trade between accuracy of these estimates and memory, holding sentence-level scores constant. Common practice uses lowerorder entries in an N-gram model to score the first few words of a fragment; this violates assumptions made by common smoothing strategies, including Kneser-Ney. Instead, we use a unigram model to score the first word, a bigram for the second, etc. This improves search at the expense of memory. Conversely, we show how to save memory by collapsing probability and backoff into a single value without changing sentence-level scores, at the expense of less accurate estimates for sentence fragments. These changes can be stacked, achieving better estimates with unchanged memory usage. In order to interpret changes in search accuracy, we adjust the pop limit so that accuracy is unchanged and report the change in CPU time. In a GermanEnglish Moses system with target-side syntax, improved estimates yielded a 63% reduction in CPU time; for a Hiero-style version, the reduction is 21%. The compressed language model uses 26% less RAM while equivalent search quality takes 27% more CPU. Source code is released as part of KenLM.
6 0.13383321 86 emnlp-2012-Locally Training the Log-Linear Model for SMT
7 0.12312829 54 emnlp-2012-Forced Derivation Tree based Model Training to Statistical Machine Translation
8 0.11904379 31 emnlp-2012-Cross-Lingual Language Modeling with Syntactic Reordering for Low-Resource Speech Recognition
9 0.11327192 128 emnlp-2012-Translation Model Based Cross-Lingual Language Model Adaptation: from Word Models to Phrase Models
10 0.11034419 109 emnlp-2012-Re-training Monolingual Parser Bilingually for Syntactic SMT
11 0.10565677 82 emnlp-2012-Left-to-Right Tree-to-String Decoding with Prediction
12 0.10081916 1 emnlp-2012-A Bayesian Model for Learning SCFGs with Discontiguous Rules
13 0.088789053 57 emnlp-2012-Generalized Higher-Order Dependency Parsing with Cube Pruning
14 0.088333517 25 emnlp-2012-Bilingual Lexicon Extraction from Comparable Corpora Using Label Propagation
15 0.076391451 127 emnlp-2012-Transforming Trees to Improve Syntactic Convergence
16 0.075461775 72 emnlp-2012-Joint Inference for Event Timeline Construction
17 0.073488697 97 emnlp-2012-Natural Language Questions for the Web of Data
18 0.072367087 58 emnlp-2012-Generalizing Sub-sentential Paraphrase Acquisition across Original Signal Type of Text Pairs
19 0.069209434 39 emnlp-2012-Enlarging Paraphrase Collections through Generalization and Instantiation
20 0.068217807 18 emnlp-2012-An Empirical Investigation of Statistical Significance in NLP
topicId topicWeight
[(0, 0.251), (1, -0.207), (2, -0.33), (3, -0.028), (4, -0.166), (5, -0.154), (6, -0.188), (7, 0.164), (8, 0.054), (9, -0.08), (10, -0.046), (11, -0.112), (12, 0.163), (13, -0.175), (14, 0.053), (15, -0.245), (16, -0.025), (17, -0.069), (18, -0.212), (19, -0.235), (20, -0.013), (21, 0.029), (22, 0.08), (23, -0.009), (24, 0.055), (25, -0.018), (26, -0.004), (27, -0.07), (28, 0.054), (29, -0.036), (30, -0.02), (31, 0.047), (32, -0.089), (33, -0.074), (34, -0.035), (35, -0.072), (36, -0.039), (37, -0.011), (38, -0.092), (39, 0.017), (40, 0.017), (41, 0.005), (42, -0.016), (43, -0.017), (44, 0.055), (45, -0.055), (46, 0.02), (47, -0.049), (48, -0.045), (49, 0.011)]
simIndex simValue paperId paperTitle
same-paper 1 0.9848581 42 emnlp-2012-Entropy-based Pruning for Phrase-based Machine Translation
Author: Wang Ling ; Joao Graca ; Isabel Trancoso ; Alan Black
Abstract: Phrase-based machine translation models have shown to yield better translations than Word-based models, since phrase pairs encode the contextual information that is needed for a more accurate translation. However, many phrase pairs do not encode any relevant context, which means that the translation event encoded in that phrase pair is led by smaller translation events that are independent from each other, and can be found on smaller phrase pairs, with little or no loss in translation accuracy. In this work, we propose a relative entropy model for translation models, that measures how likely a phrase pair encodes a translation event that is derivable using smaller translation events with similar probabilities. This model is then applied to phrase table pruning. Tests show that considerable amounts of phrase pairs can be excluded, without much impact on the transla- . tion quality. In fact, we show that better translations can be obtained using our pruned models, due to the compression of the search space during decoding.
2 0.98178893 11 emnlp-2012-A Systematic Comparison of Phrase Table Pruning Techniques
Author: Richard Zens ; Daisy Stanton ; Peng Xu
Abstract: When trained on very large parallel corpora, the phrase table component of a machine translation system grows to consume vast computational resources. In this paper, we introduce a novel pruning criterion that places phrase table pruning on a sound theoretical foundation. Systematic experiments on four language pairs under various data conditions show that our principled approach is superior to existing ad hoc pruning methods.
3 0.6319719 74 emnlp-2012-Language Model Rest Costs and Space-Efficient Storage
Author: Kenneth Heafield ; Philipp Koehn ; Alon Lavie
Abstract: Approximate search algorithms, such as cube pruning in syntactic machine translation, rely on the language model to estimate probabilities of sentence fragments. We contribute two changes that trade between accuracy of these estimates and memory, holding sentence-level scores constant. Common practice uses lowerorder entries in an N-gram model to score the first few words of a fragment; this violates assumptions made by common smoothing strategies, including Kneser-Ney. Instead, we use a unigram model to score the first word, a bigram for the second, etc. This improves search at the expense of memory. Conversely, we show how to save memory by collapsing probability and backoff into a single value without changing sentence-level scores, at the expense of less accurate estimates for sentence fragments. These changes can be stacked, achieving better estimates with unchanged memory usage. In order to interpret changes in search accuracy, we adjust the pop limit so that accuracy is unchanged and report the change in CPU time. In a GermanEnglish Moses system with target-side syntax, improved estimates yielded a 63% reduction in CPU time; for a Hiero-style version, the reduction is 21%. The compressed language model uses 26% less RAM while equivalent search quality takes 27% more CPU. Source code is released as part of KenLM.
4 0.46342006 128 emnlp-2012-Translation Model Based Cross-Lingual Language Model Adaptation: from Word Models to Phrase Models
Author: Shixiang Lu ; Wei Wei ; Xiaoyin Fu ; Bo Xu
Abstract: In this paper, we propose a novel translation model (TM) based cross-lingual data selection model for language model (LM) adaptation in statistical machine translation (SMT), from word models to phrase models. Given a source sentence in the translation task, this model directly estimates the probability that a sentence in the target LM training corpus is similar. Compared with the traditional approaches which utilize the first pass translation hypotheses, cross-lingual data selection model avoids the problem of noisy proliferation. Furthermore, phrase TM based cross-lingual data selection model is more effective than the traditional approaches based on bag-ofwords models and word-based TM, because it captures contextual information in modeling the selection of phrase as a whole. Experiments conducted on large-scale data sets demonstrate that our approach significantly outperforms the state-of-the-art approaches on both LM perplexity and SMT performance.
5 0.45893949 54 emnlp-2012-Forced Derivation Tree based Model Training to Statistical Machine Translation
Author: Nan Duan ; Mu Li ; Ming Zhou
Abstract: A forced derivation tree (FDT) of a sentence pair {f, e} denotes a derivation tree that can tpraainrsl {afte, f} idnetono itetss a acc duerraivtea target etrea tnhsaltat cioann e. In this paper, we present an approach that leverages structured knowledge contained in FDTs to train component models for statistical machine translation (SMT) systems. We first describe how to generate different FDTs for each sentence pair in training corpus, and then present how to infer the optimal FDTs based on their derivation and alignment qualities. As the first step in this line of research, we verify the effectiveness of our approach in a BTGbased phrasal system, and propose four FDTbased component models. Experiments are carried out on large scale English-to-Japanese and Chinese-to-English translation tasks, and significant improvements are reported on both translation quality and alignment quality.
6 0.42912427 31 emnlp-2012-Cross-Lingual Language Modeling with Syntactic Reordering for Low-Resource Speech Recognition
7 0.41690278 86 emnlp-2012-Locally Training the Log-Linear Model for SMT
8 0.39101201 35 emnlp-2012-Document-Wide Decoding for Phrase-Based Statistical Machine Translation
9 0.3665776 1 emnlp-2012-A Bayesian Model for Learning SCFGs with Discontiguous Rules
10 0.35630104 67 emnlp-2012-Inducing a Discriminative Parser to Optimize Machine Translation Reordering
11 0.3548907 25 emnlp-2012-Bilingual Lexicon Extraction from Comparable Corpora Using Label Propagation
12 0.32205641 109 emnlp-2012-Re-training Monolingual Parser Bilingually for Syntactic SMT
13 0.30822843 45 emnlp-2012-Exploiting Chunk-level Features to Improve Phrase Chunking
14 0.29174444 118 emnlp-2012-Source Language Adaptation for Resource-Poor Machine Translation
15 0.29126704 82 emnlp-2012-Left-to-Right Tree-to-String Decoding with Prediction
16 0.2850641 97 emnlp-2012-Natural Language Questions for the Web of Data
17 0.25753555 41 emnlp-2012-Entity based QA Retrieval
18 0.25618085 39 emnlp-2012-Enlarging Paraphrase Collections through Generalization and Instantiation
19 0.25406262 114 emnlp-2012-Revisiting the Predictability of Language: Response Completion in Social Media
20 0.25339124 18 emnlp-2012-An Empirical Investigation of Statistical Significance in NLP
topicId topicWeight
[(2, 0.022), (14, 0.032), (16, 0.026), (25, 0.015), (34, 0.136), (39, 0.012), (41, 0.07), (45, 0.016), (60, 0.108), (63, 0.064), (64, 0.015), (65, 0.021), (70, 0.029), (74, 0.064), (76, 0.047), (80, 0.026), (85, 0.162), (86, 0.022), (94, 0.012), (95, 0.012)]
simIndex simValue paperId paperTitle
same-paper 1 0.83540773 42 emnlp-2012-Entropy-based Pruning for Phrase-based Machine Translation
Author: Wang Ling ; Joao Graca ; Isabel Trancoso ; Alan Black
Abstract: Phrase-based machine translation models have shown to yield better translations than Word-based models, since phrase pairs encode the contextual information that is needed for a more accurate translation. However, many phrase pairs do not encode any relevant context, which means that the translation event encoded in that phrase pair is led by smaller translation events that are independent from each other, and can be found on smaller phrase pairs, with little or no loss in translation accuracy. In this work, we propose a relative entropy model for translation models, that measures how likely a phrase pair encodes a translation event that is derivable using smaller translation events with similar probabilities. This model is then applied to phrase table pruning. Tests show that considerable amounts of phrase pairs can be excluded, without much impact on the transla- . tion quality. In fact, we show that better translations can be obtained using our pruned models, due to the compression of the search space during decoding.
2 0.78746116 11 emnlp-2012-A Systematic Comparison of Phrase Table Pruning Techniques
Author: Richard Zens ; Daisy Stanton ; Peng Xu
Abstract: When trained on very large parallel corpora, the phrase table component of a machine translation system grows to consume vast computational resources. In this paper, we introduce a novel pruning criterion that places phrase table pruning on a sound theoretical foundation. Systematic experiments on four language pairs under various data conditions show that our principled approach is superior to existing ad hoc pruning methods.
3 0.75911516 131 emnlp-2012-Unified Dependency Parsing of Chinese Morphological and Syntactic Structures
Author: Zhongguo Li ; Guodong Zhou
Abstract: Most previous approaches to syntactic parsing of Chinese rely on a preprocessing step of word segmentation, thereby assuming there was a clearly defined boundary between morphology and syntax in Chinese. We show how this assumption can fail badly, leading to many out-of-vocabulary words and incompatible annotations. Hence in practice the strict separation of morphology and syntax in the Chinese language proves to be untenable. We present a unified dependency parsing approach for Chinese which takes unsegmented sentences as input and outputs both morphological and syntactic structures with a single model and algorithm. By removing the intermediate word segmentation, the unified parser no longer needs separate notions for words and phrases. Evaluation proves the effectiveness of the unified model and algorithm in parsing structures of words, phrases and sen- tences simultaneously. 1
4 0.72202331 89 emnlp-2012-Mixed Membership Markov Models for Unsupervised Conversation Modeling
Author: Michael J. Paul
Abstract: Recent work has explored the use of hidden Markov models for unsupervised discourse and conversation modeling, where each segment or block of text such as a message in a conversation is associated with a hidden state in a sequence. We extend this approach to allow each block of text to be a mixture of multiple classes. Under our model, the probability of a class in a text block is a log-linear function of the classes in the previous block. We show that this model performs well at predictive tasks on two conversation data sets, improving thread reconstruction accuracy by up to 15 percentage points over a standard HMM. Additionally, we show quantitatively that the induced word clusters correspond to speech acts more closely than baseline models.
5 0.720101 109 emnlp-2012-Re-training Monolingual Parser Bilingually for Syntactic SMT
Author: Shujie Liu ; Chi-Ho Li ; Mu Li ; Ming Zhou
Abstract: The training of most syntactic SMT approaches involves two essential components, word alignment and monolingual parser. In the current state of the art these two components are mutually independent, thus causing problems like lack of rule generalization, and violation of syntactic correspondence in translation rules. In this paper, we propose two ways of re-training monolingual parser with the target of maximizing the consistency between parse trees and alignment matrices. One is targeted self-training with a simple evaluation function; the other is based on training data selection from forced alignment of bilingual data. We also propose an auxiliary method for boosting alignment quality, by symmetrizing alignment matrices with respect to parse trees. The best combination of these novel methods achieves 3 Bleu point gain in an IWSLT task and more than 1 Bleu point gain in NIST tasks. 1
6 0.71833032 18 emnlp-2012-An Empirical Investigation of Statistical Significance in NLP
7 0.71407509 54 emnlp-2012-Forced Derivation Tree based Model Training to Statistical Machine Translation
8 0.70611137 123 emnlp-2012-Syntactic Transfer Using a Bilingual Lexicon
9 0.70402586 70 emnlp-2012-Joint Chinese Word Segmentation, POS Tagging and Parsing
10 0.70321965 24 emnlp-2012-Biased Representation Learning for Domain Adaptation
11 0.70301026 45 emnlp-2012-Exploiting Chunk-level Features to Improve Phrase Chunking
12 0.70210379 67 emnlp-2012-Inducing a Discriminative Parser to Optimize Machine Translation Reordering
13 0.69846976 136 emnlp-2012-Weakly Supervised Training of Semantic Parsers
14 0.69772851 81 emnlp-2012-Learning to Map into a Universal POS Tagset
15 0.69738185 108 emnlp-2012-Probabilistic Finite State Machines for Regression-based MT Evaluation
16 0.69576842 95 emnlp-2012-N-gram-based Tense Models for Statistical Machine Translation
17 0.69565338 64 emnlp-2012-Improved Parsing and POS Tagging Using Inter-Sentence Consistency Constraints
18 0.69545585 111 emnlp-2012-Regularized Interlingual Projections: Evaluation on Multilingual Transliteration
19 0.69475865 74 emnlp-2012-Language Model Rest Costs and Space-Efficient Storage
20 0.69465011 22 emnlp-2012-Automatically Constructing a Normalisation Dictionary for Microblogs