emnlp emnlp2010 emnlp2010-5 knowledge-graph by maker-knowledge-mining
Source: pdf
Author: Minh-Thang Luong ; Preslav Nakov ; Min-Yen Kan
Abstract: We propose a language-independent approach for improving statistical machine translation for morphologically rich languages using a hybrid morpheme-word representation where the basic unit of translation is the morpheme, but word boundaries are respected at all stages of the translation process. Our model extends the classic phrase-based model by means of (1) word boundary-aware morpheme-level phrase extraction, (2) minimum error-rate training for a morpheme-level translation model using word-level BLEU, and (3) joint scoring with morpheme- and word-level language models. Further improvements are achieved by combining our model with the classic one. The evaluation on English to Finnish using Europarl (714K sentence pairs; 15.5M English words) shows statistically significant improvements over the classic model based on BLEU and human judgments.
Reference: text
sentIndex sentText sentNum sentScore
1 Our model extends the classic phrase-based model by means of (1) word boundary-aware morpheme-level phrase extraction, (2) minimum error-rate training for a morpheme-level translation model using word-level BLEU, and (3) joint scoring with morpheme- and word-level language models. [sent-5, score-0.462]
2 1 Introduction The fast progress of statistical machine translation (SMT) has boosted translation quality significantly. [sent-9, score-0.404]
3 This is fine for languages with limited morphology like English and French, or no morphology at all like Chinese, but it is inadequate for morphologically rich languages like Arabic, Czech or Finnish (Lee, 2004; Goldwater and McClosky, 2005; Yang and Kirchhoff, 2006). [sent-11, score-0.161]
4 148 There has been a line of recent SMT research that incorporates morphological analysis as part of the translation process, thus providing access to the information within the individual words. [sent-13, score-0.33]
5 Below we propose a language-independent ap- proach to SMT of morphologically rich languages using a hybrid morpheme-word representation where the basic unit of translation is the morpheme, but word boundaries are respected at all stages of the translation process. [sent-15, score-0.608]
6 Sections 3 and 4 present our morphological and phrase merging enhancements. [sent-19, score-0.368]
7 We should also note the importance of the translation direction: it is much harder to translate from a morphologically poor to a morphologically rich language, where morphological distinctions not present in the source need to be generated in the target language. [sent-30, score-0.614]
8 Research in translating into morphologically rich languages, has attracted interest for languages like Arabic (Badr et al. [sent-31, score-0.207]
9 We use a hybrid morpheme-word representation where the basic unit of translation is the morpheme, but word boundaries are respected at all stages of the translation process. [sent-38, score-0.447]
10 This is in contrast with previous work, where morphological enhancements are typically performed as pre-/postprocessing steps only. [sent-39, score-0.258]
11 In addition to changing the basic translation token unit from a word to a morpheme, our model extends the phrase-based SMT model with the following: 1. [sent-40, score-0.202]
12 1 Morphological Representation Our morphological representation is based on the output of an unsupervised morphological analyzer. [sent-52, score-0.256]
13 2 Word Boundary-aware Phrase Extraction The core translation structure of a phrase-based SMT model is the phrase table, which is learned from a bilingual parallel sentence-aligned corpus, typically using the alignment template approach (Och and Ney, 2004). [sent-61, score-0.334]
14 It contains a set of bilingual phrase pairs, each associated with five scores: forward and backward phrase translation probabilities, forward and backward lexicalized translation probabilities, and a constant phrase penalty. [sent-62, score-0.948]
15 The maximum phrase length n is normally limited to seven words; higher values of n increase the table size exponentially without actually yielding performance benefit (Koehn et al. [sent-63, score-0.132]
16 However, things are different when translating with morphemes, for two reasons: (1) morpheme-token phrases of length n can span less than n words; and (2) morphemetoken phrases may only partially span words. [sent-65, score-0.388]
17 The first point means that morpheme-token phrase pairs span fewer word tokens, and thus cover a smaller context, which may result in fewer total extracted pairs compared to a word-level approach. [sent-66, score-0.179]
18 The second point is more interesting: morphemelevel phrases may span words partially, making them potentially usable in translating unknown inflected forms of known source language words, but also creates the danger of generating sequences of morphemes that are not legal target language words. [sent-72, score-0.362]
19 For example, let us consider the phrase in Figure 1: unPRE+ democraticSTM. [sent-73, score-0.132]
20 The original algorithm will extract the spurious phrase ep¨ aPRE+ demokraat S TM+ t SUF+ i SUF+ s SUF+ , beside the correct one that has enSUF appended at the end. [sent-74, score-0.132]
21 Such a spurious phrase does not generally help in translating unknown inflected forms, especially for morphologically-rich languages that feature multiple affixes, but negatively affects the translation model in terms of complexity and quality. [sent-75, score-0.418]
22 3 Morpheme-Token MERT Optimizing Word-Token BLEU Modern phrase-based SMT systems use a log-linear model with the following typical feature functions: language model probabilities, word penalty, distortion cost, and the five parameters from the phrase table. [sent-82, score-0.132]
23 This motivates us to perform MERT at the wordtoken level, although our input consists of morphemes. [sent-89, score-0.15]
24 In particular, for each iteration of MERT, as soon as the decoder generates a morpheme-token translation for a sentence, we convert it into a wordtoken sequence, which is used to calculate BLEU. [sent-90, score-0.352]
25 We thus achieve MERT optimization at the wordtoken level while translating a morpheme-token input and generating a morpheme-token output. [sent-91, score-0.196]
26 This has the advantage of alleviating the problem of data sparseness, especially when translating into a morphologically rich language, since the LM would be able to handle some new unseen inflected forms of known words. [sent-94, score-0.245]
27 On the negative side, a morphemetoken LM spans fewer word-tokens and thus has a more limited word “horizon” compared to one operating at the word level. [sent-95, score-0.15]
28 As with the maximum phrase length, mechanically increasing the order of the morpheme-token LM has a limited impact. [sent-96, score-0.132]
29 The interaction of the twin LMs is illustrated in Figure 2. [sent-104, score-0.15]
30 While the factored translation model (Koehn and Hoang, 2007) in Moses does allow scoring with models of different granularity, e. [sent-109, score-0.258]
31 Note that scoring with twin LMs is conceptually superior to n-best re-scoring with a word-token LM, e. [sent-112, score-0.206]
32 3We use the term “hypothesis” to collectively refer to the following (Koehn, 2003): the source phrase covered, the corresponding target phrase, and most importantly, a reference to the previous hypothesis that it extends. [sent-118, score-0.132]
33 151 However, for phrase-based SMT systems, it is theoretically more appealing to combine their phrase tables since this allows the translation models of both systems to influence the hypothesis search directly. [sent-119, score-0.406]
34 1 Building a Twin Translation Model Figure 3 shows a general scheme of our twin translation model. [sent-123, score-0.352]
35 We then build separate phrase tables (PT) for the two inputs: a word-token PTw and a morphemetoken PTm. [sent-125, score-0.354]
36 Second, we re-tokenize PTw at the morpheme level, thus obtaining a new phrase table PTw→m, which is of the same granularity as PTm. [sent-126, score-0.262]
37 Finally, we merge PTw→m and PTm, and we input the resulting phrase table to the decoder. [sent-127, score-0.132]
38 WordPahlirgnsemEnGxtWIrZPaTAocr+"wdonPMhGroaIsZPpAeThmE+Mmxoterapch"eomnaligment Mseogrmphenoltaog"iocna lP Tw→m PT merging Decoding Figure 3: Building a twin phrase table (PT). [sent-128, score-0.39]
39 Second, the wordtoken PT is retokenized at the morpheme-token level. [sent-130, score-0.15]
40 2 Merging and Normalizing Phrase Tables Below we first describe the two general phrase table combination strategies used in previous work: (1) direct merging using additional feature functions, and (2) phrase table interpolation. [sent-133, score-0.372]
41 The first line of research on phrase table merging is exemplified by (Niehues et al. [sent-136, score-0.24]
42 The idea is to select one of the phrase tables as primary and to add to it all nonduplicating phrase pairs from the second table together with their associated scores. [sent-140, score-0.421]
43 Instead, the conditional probabilities in the two phrase tables are often interpolated directly, e. [sent-149, score-0.204]
44 The above phrase merging approaches have been proposed for phrase tables derived from different sources. [sent-155, score-0.444]
45 This is in contrast with our twin translation scenario, where the morphemetoken phrase tables are built from the same training dataset; the main difference being that word alignments and phrase extraction were performed at the 152 word-token level for PTw→m and at the morphemetoken level for PTm. [sent-156, score-0.988]
46 Thus, we propose different merging approaches for the phrase translation probabilities and for the lexicalized probabilities. [sent-157, score-0.504]
47 In phrase-based SMT, phrase translation probabilities are computed using maximum likelihood (ML) estimation φ(f¯| ¯e) = P#f¯(#f¯,( ¯ef¯), e¯), where #(f¯,¯ e) is the number of times theP Ppair (f¯, e) is extracted from the training dataset (Koehn et al. [sent-158, score-0.417]
48 Instead, we use the raw counts for the two models #m(f¯, e) and #w→m(f¯, e) directly as follows: φ(f¯, e¯) =Pf¯##mm((f ¯ , e¯ e¯ ) + + #Pwf¯→#mw(→f¯m, e¯ ()f¯, e¯ ) For lexicalizPed translation proPbabilities, we would × like to use simple interpolation. [sent-161, score-0.202]
49 However, we notice that when a phrase pair belongs to only one of the phrase tables, the corresponding lexicalized score for the other table would be zero. [sent-162, score-0.326]
50 We thus perform interpolation from PTm and PTw according to the following formula: lex(f¯| e¯) = α lexm(f¯m| e¯m) + (1 α) lexw(f¯w| e¯w) where the concatenation of f¯m and em into wordtoken sequences yields f¯w and ew, respectively. [sent-164, score-0.221]
51 If both (f¯m, em) and (f¯w, ew) are present in PTm and PTw, respectively, we have a simple interpolation of their corresponding lexicalized scores lexm and lexw. [sent-165, score-0.17]
52 We then estimate a lexicalized phrase score using the original formula given in − (Koehn et al. [sent-168, score-0.194]
53 , 2003), where we plug this induced word alignment and word-token lexical translation probabilities estimated from the word-token dataset The case when (f¯w, ew) is present in PTw, but (f¯m, em) is not, is solved similarly. [sent-169, score-0.285]
54 We further split the training dataset into four subsets T1, T2, T3, and T4 of sizes 40K, 80K, 160K, and 320K parallel sentence pairs, which we use for studying the impact of training data size on translation performance. [sent-172, score-0.285]
55 , 2007): w-system: works at the word-token level, extracts phrases of up to seven words, and uses a 4-gram word-token LM (as typical for phrase-based SMT); m-system: works at the morpheme level, tokenized using Morfessor4 and augmented with “+” as described in Section 3. [sent-193, score-0.179]
56 None of the enhancements described previously is applied yet. [sent-197, score-0.13]
57 To evaluate the translation quality, we compute BLEU (Papineni et al. [sent-199, score-0.202]
58 We further introduce a morpheme-token version of BLEU, which we call m-BLEU: it first segments the system output and the reference translation into morpheme-tokens and then calculates a BLEU score as usual. [sent-201, score-0.202]
59 Shown are word BLEU and morpheme mBLEU scores for the w-system and m-system. [sent-225, score-0.13]
60 higher m-BLEU scores, indicating that it may have better morpheme coverage5. [sent-226, score-0.13]
61 However, the m-system is outperformed by the w-system on the classic wordtoken BLEU, which means that it either does not perform as well as the w-system or that word-token BLEU is not capable of measuring the morphemelevel improvements. [sent-227, score-0.297]
62 47 points on the full dataset, followed by lm and phr. [sent-235, score-0.144]
63 Table 3 (iii) further shows that using phr and lm together yields absolute improvements of 0. [sent-236, score-0.303]
64 Overall, the morphological enhancements are on par with the w-system baseline, and yield sizable im5Note that these morphemes were generated automatically and thus many of them are erroneous. [sent-240, score-0.365]
65 3510 Table 3: Impact of the morphological enhancements (on test dataset). [sent-263, score-0.258]
66 Shown are BLEU scores (in %) for training on T1 and on the full dataset for (i) baselines, (ii) enhancements individually, and (iii) combined. [sent-264, score-0.213]
67 4 Combining Translation Tables Finally, we investigate the effect of combining phrase tables derived from a word-token and a morpheme-token input, as described in Section 4. [sent-272, score-0.204]
68 T1 (40K) PTm is primary PTw→m is primary Full (714K) 11. [sent-277, score-0.17]
69 19 w→m Table 4: Effect of selection of primary phrase table for add-1 (on dev dataset): PTw→m, derived from a wordtoken input, vs. [sent-281, score-0.367]
70 For add-1 and add-2, we need to decide which (PTw→m or PTm) phrase table should be considw→m 6The feature values are e1, e2/3 or e1/3 (e=2. [sent-284, score-0.132]
71 ); when the phrase pair comes from both tables, from the primary table only, and from the secondary table only, respectively. [sent-288, score-0.217]
72 7The feature values are (e1, e1), (e1, e0) or (e0, e1) when the phrase pair comes from both tables, from the primary table only, and from the secondary table only, respectively. [sent-289, score-0.217]
73 As we can see, using PTw→m as primary performs better on T1 and on the full training dataset; thus, we will use it as primary on the test dataset for add-1 and add-2. [sent-292, score-0.253]
74 Due to time constraints, we use the same value for the phrase translation probabilities and for the lexicalized probabilities, and we perform grid search for α ∈ {0. [sent-294, score-0.396]
75 6e turns out to work best on the development dataset; we will use this value in our experiments on the test dataset both for interpolate and for ourMethod8. [sent-304, score-0.131]
76 We integrate the morphologically enhanced system m+phr+lm and the word-token based w-system using the four merging methods above. [sent-318, score-0.231]
77 The number of phrase pairs in (i) individual PTs and (ii) PT overlap, is shown. [sent-349, score-0.132]
78 5M) of the phrase pairs in PTm+phr are also in PTm, which confirms that boundary-aware phrase extraction selects good phrase pairs from PTm to be retained in PTm+phr. [sent-356, score-0.396]
79 Thus, enriching the translation model with PTw→m helps improve coverage. [sent-364, score-0.202]
80 We performed automatic comparison based on corresponding phrases between the translation output (out) and the reference (ref), using the source (src) test dataset as a pivot. [sent-390, score-0.334]
81 The decoding log gave us the phrases used to translate src to out, and we only needed to find correspondences between src and ref, which we accomplished by appending the test dataset to training and performing IBM Model 4 word alignments. [sent-391, score-0.334]
82 We then looked for phrase triples (src, out, ref), where there was a high character-level similarity between out and ref, measured using longest common subsequence ratio with a threshold of 0. [sent-392, score-0.207]
83 re f: maltillisen kokoomuspuolueen edustajana suhtaudun erittain saastavaisesti veronmaksajien rahoihin . [sent-409, score-0.225]
84 our: konservatiivinen , olen erittain saastavaisesti veronmaksajien rahoja . [sent-410, score-0.393]
85 w : konservatiivinen , olen aarettoman tarkeaa kanssa veronmaksajien rahoja . [sent-411, score-0.243]
86 m : kuten konservatiivinen , olen erittain saastavaisesti veronmaksajien rahoja . [sent-412, score-0.43]
87 re f: olimme erittain rakentavia ja neuvottelimme haagissa viime hetkeen saakka . [sent-427, score-0.545]
88 our: olemme olleet hyvin rakentavia ja olemme neuvotelleet viime hetkeen saakka naiden neuvottelujen haagissa . [sent-428, score-0.769]
89 w : olemme olleet hyvin rakentavia ja olemme neuvotelleet viime tippaan niin naiden neuvottelujen haagissa . [sent-429, score-0.657]
90 m : olimme erittain rakentavan ja neuvottelimme viime hetkeen saakka naiden neuvotteluiden haagissa . [sent-430, score-0.545]
91 re f: olisi eritt ¨ain vaarallinen tilanne ,jos eurooppalaiset tulisivat logistisesti riippuvaisiksi ven a¨j a¨st a¨ . [sent-439, score-0.47]
92 our: olisi eritt ¨ain vaarallinen tilanne ,jos eurooppalaiset tulee logistisesti riippuvaisia ven a¨j a¨n . [sent-440, score-0.526]
93 w : se olisi eritt ¨ain vaarallinen tilanne ,jos eurooppalaisten tulisi logistically riippuvaisia ven a¨j a¨n . [sent-441, score-0.489]
94 m : se olisi hyvin vaarallinen tilanne ,jos eurooppalaiset haluavat tulla logistisesti riippuvaisia ven a¨j a¨n . [sent-442, score-0.526]
95 A comparative study of hypothesis alignment and its improvement for machine translation system combination. [sent-492, score-0.202]
96 Mining a comparable text corpus for a Vietnamese-French statistical machine translation system. [sent-511, score-0.202]
97 Improved statis- tical machine translation for resource-poor languages using related resource-rich languages. [sent-549, score-0.202]
98 The Universit a¨t Karlsruhe translation system for the EACL-WMT 2009. [sent-553, score-0.202]
99 Morphology-aware statistical machine translation based on morphs induced in an unsupervised manner. [sent-590, score-0.202]
100 Phrase-based backoff models for machine translation of highly inflected languages. [sent-598, score-0.24]
wordName wordTfidf (topN-words)
[('ptm', 0.305), ('ptw', 0.3), ('bleu', 0.212), ('translation', 0.202), ('phr', 0.159), ('smt', 0.151), ('morphemetoken', 0.15), ('twin', 0.15), ('wordtoken', 0.15), ('lm', 0.144), ('phrase', 0.132), ('enhancements', 0.13), ('morpheme', 0.13), ('finnish', 0.128), ('suf', 0.128), ('morphological', 0.128), ('morphologically', 0.123), ('oflazer', 0.112), ('merging', 0.108), ('morphemes', 0.107), ('koehn', 0.105), ('src', 0.101), ('ja', 0.096), ('erittain', 0.094), ('lms', 0.093), ('primary', 0.085), ('dataset', 0.083), ('ensuf', 0.075), ('eritt', 0.075), ('haagissa', 0.075), ('hyvin', 0.075), ('logistically', 0.075), ('maahanmuuttopolitiikan', 0.075), ('morphemelevel', 0.075), ('olemme', 0.075), ('olisi', 0.075), ('oursystem', 0.075), ('stm', 0.075), ('tilanne', 0.075), ('vaarallinen', 0.075), ('veronmaksajien', 0.075), ('viime', 0.075), ('triples', 0.075), ('classic', 0.072), ('tables', 0.072), ('interpolation', 0.071), ('mert', 0.068), ('ep', 0.064), ('lexicalized', 0.062), ('ref', 0.059), ('pre', 0.058), ('ven', 0.058), ('scoring', 0.056), ('badr', 0.056), ('demokraattisen', 0.056), ('eurooppalaiset', 0.056), ('hetkeen', 0.056), ('konservatiivinen', 0.056), ('logistisesti', 0.056), ('maahanmuuttopre', 0.056), ('naiden', 0.056), ('olen', 0.056), ('ourmethod', 0.056), ('rahoja', 0.056), ('rakentavia', 0.056), ('riippuvaisia', 0.056), ('saakka', 0.056), ('saastavaisesti', 0.056), ('philipp', 0.056), ('ii', 0.056), ('ain', 0.053), ('arabic', 0.049), ('phrases', 0.049), ('buckwalter', 0.048), ('creutz', 0.048), ('interpolate', 0.048), ('morfessor', 0.048), ('turku', 0.048), ('virpioja', 0.048), ('ew', 0.047), ('span', 0.047), ('translating', 0.046), ('pts', 0.043), ('respected', 0.043), ('backward', 0.043), ('mm', 0.04), ('nakov', 0.04), ('moses', 0.04), ('tune', 0.039), ('iii', 0.038), ('rich', 0.038), ('inflected', 0.038), ('pt', 0.037), ('avramidis', 0.037), ('europeans', 0.037), ('kuten', 0.037), ('lexm', 0.037), ('neuvotelleet', 0.037), ('neuvottelimme', 0.037), ('neuvottelujen', 0.037)]
simIndex simValue paperId paperTitle
same-paper 1 1.0000004 5 emnlp-2010-A Hybrid Morpheme-Word Representation for Machine Translation of Morphologically Rich Languages
Author: Minh-Thang Luong ; Preslav Nakov ; Min-Yen Kan
Abstract: We propose a language-independent approach for improving statistical machine translation for morphologically rich languages using a hybrid morpheme-word representation where the basic unit of translation is the morpheme, but word boundaries are respected at all stages of the translation process. Our model extends the classic phrase-based model by means of (1) word boundary-aware morpheme-level phrase extraction, (2) minimum error-rate training for a morpheme-level translation model using word-level BLEU, and (3) joint scoring with morpheme- and word-level language models. Further improvements are achieved by combining our model with the classic one. The evaluation on English to Finnish using Europarl (714K sentence pairs; 15.5M English words) shows statistically significant improvements over the classic model based on BLEU and human judgments.
2 0.17950772 39 emnlp-2010-EMNLP 044
Author: George Foster
Abstract: We describe a new approach to SMT adaptation that weights out-of-domain phrase pairs according to their relevance to the target domain, determined by both how similar to it they appear to be, and whether they belong to general language or not. This extends previous work on discriminative weighting by using a finer granularity, focusing on the properties of instances rather than corpus components, and using a simpler training procedure. We incorporate instance weighting into a mixture-model framework, and find that it yields consistent improvements over a wide range of baselines.
3 0.1726598 47 emnlp-2010-Example-Based Paraphrasing for Improved Phrase-Based Statistical Machine Translation
Author: Aurelien Max
Abstract: In this article, an original view on how to improve phrase translation estimates is proposed. This proposal is grounded on two main ideas: first, that appropriate examples of a given phrase should participate more in building its translation distribution; second, that paraphrases can be used to better estimate this distribution. Initial experiments provide evidence of the potential of our approach and its implementation for effectively improving translation performance.
4 0.17092367 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice
Author: Samidh Chatterjee ; Nicola Cancedda
Abstract: Minimum Error Rate Training is the algorithm for log-linear model parameter training most used in state-of-the-art Statistical Machine Translation systems. In its original formulation, the algorithm uses N-best lists output by the decoder to grow the Translation Pool that shapes the surface on which the actual optimization is performed. Recent work has been done to extend the algorithm to use the entire translation lattice built by the decoder, instead of N-best lists. We propose here a third, intermediate way, consisting in growing the translation pool using samples randomly drawn from the translation lattice. We empirically measure a systematic im- provement in the BLEU scores compared to training using N-best lists, without suffering the increase in computational complexity associated with operating with the whole lattice.
5 0.1647301 50 emnlp-2010-Facilitating Translation Using Source Language Paraphrase Lattices
Author: Jinhua Du ; Jie Jiang ; Andy Way
Abstract: For resource-limited language pairs, coverage of the test set by the parallel corpus is an important factor that affects translation quality in two respects: 1) out of vocabulary words; 2) the same information in an input sentence can be expressed in different ways, while current phrase-based SMT systems cannot automatically select an alternative way to transfer the same information. Therefore, given limited data, in order to facilitate translation from the input side, this paper proposes a novel method to reduce the translation difficulty using source-side lattice-based paraphrases. We utilise the original phrases from the input sentence and the corresponding paraphrases to build a lattice with estimated weights for each edge to improve translation quality. Compared to the baseline system, our method achieves relative improvements of 7.07%, 6.78% and 3.63% in terms of BLEU score on small, medium and large- scale English-to-Chinese translation tasks respectively. The results show that the proposed method is effective not only for resourcelimited language pairs, but also for resourcesufficient pairs to some extent.
6 0.15820029 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
7 0.13833266 63 emnlp-2010-Improving Translation via Targeted Paraphrasing
8 0.13611293 33 emnlp-2010-Cross Language Text Classification by Model Translation and Semi-Supervised Learning
9 0.12952548 29 emnlp-2010-Combining Unsupervised and Supervised Alignments for MT: An Empirical Study
10 0.12891336 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
11 0.12571071 57 emnlp-2010-Hierarchical Phrase-Based Translation Grammars Extracted from Alignment Posterior Probabilities
12 0.12051705 22 emnlp-2010-Automatic Evaluation of Translation Quality for Distant Language Pairs
13 0.1124296 99 emnlp-2010-Statistical Machine Translation with a Factorized Grammar
14 0.10064247 36 emnlp-2010-Discriminative Word Alignment with a Function Word Reordering Model
15 0.094271153 86 emnlp-2010-Non-Isomorphic Forest Pair Translation
16 0.092456102 98 emnlp-2010-Soft Syntactic Constraints for Hierarchical Phrase-Based Translation Using Latent Syntactic Distributions
17 0.085466042 42 emnlp-2010-Efficient Incremental Decoding for Tree-to-String Translation
18 0.082514912 92 emnlp-2010-Predicting the Semantic Compositionality of Prefix Verbs
19 0.081934281 9 emnlp-2010-A New Approach to Lexical Disambiguation of Arabic Text
20 0.075560234 76 emnlp-2010-Maximum Entropy Based Phrase Reordering for Hierarchical Phrase-Based Translation
topicId topicWeight
[(0, 0.277), (1, -0.273), (2, -0.045), (3, -0.015), (4, -0.091), (5, -0.023), (6, -0.002), (7, -0.004), (8, -0.045), (9, -0.018), (10, 0.04), (11, 0.057), (12, 0.016), (13, -0.037), (14, -0.021), (15, -0.064), (16, -0.127), (17, 0.105), (18, -0.137), (19, -0.039), (20, 0.057), (21, 0.134), (22, 0.058), (23, 0.184), (24, 0.058), (25, -0.017), (26, 0.118), (27, 0.108), (28, -0.024), (29, -0.101), (30, 0.038), (31, 0.142), (32, -0.029), (33, -0.009), (34, -0.067), (35, -0.042), (36, 0.062), (37, -0.067), (38, 0.01), (39, -0.027), (40, -0.047), (41, -0.012), (42, 0.042), (43, -0.077), (44, 0.021), (45, -0.04), (46, 0.121), (47, -0.023), (48, 0.049), (49, 0.013)]
simIndex simValue paperId paperTitle
same-paper 1 0.92832333 5 emnlp-2010-A Hybrid Morpheme-Word Representation for Machine Translation of Morphologically Rich Languages
Author: Minh-Thang Luong ; Preslav Nakov ; Min-Yen Kan
Abstract: We propose a language-independent approach for improving statistical machine translation for morphologically rich languages using a hybrid morpheme-word representation where the basic unit of translation is the morpheme, but word boundaries are respected at all stages of the translation process. Our model extends the classic phrase-based model by means of (1) word boundary-aware morpheme-level phrase extraction, (2) minimum error-rate training for a morpheme-level translation model using word-level BLEU, and (3) joint scoring with morpheme- and word-level language models. Further improvements are achieved by combining our model with the classic one. The evaluation on English to Finnish using Europarl (714K sentence pairs; 15.5M English words) shows statistically significant improvements over the classic model based on BLEU and human judgments.
2 0.71757174 78 emnlp-2010-Minimum Error Rate Training by Sampling the Translation Lattice
Author: Samidh Chatterjee ; Nicola Cancedda
Abstract: Minimum Error Rate Training is the algorithm for log-linear model parameter training most used in state-of-the-art Statistical Machine Translation systems. In its original formulation, the algorithm uses N-best lists output by the decoder to grow the Translation Pool that shapes the surface on which the actual optimization is performed. Recent work has been done to extend the algorithm to use the entire translation lattice built by the decoder, instead of N-best lists. We propose here a third, intermediate way, consisting in growing the translation pool using samples randomly drawn from the translation lattice. We empirically measure a systematic im- provement in the BLEU scores compared to training using N-best lists, without suffering the increase in computational complexity associated with operating with the whole lattice.
3 0.66798973 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
Author: Guillaume Wisniewski ; Alexandre Allauzen ; Francois Yvon
Abstract: Extant Statistical Machine Translation (SMT) systems are very complex softwares, which embed multiple layers of heuristics and embark very large numbers of numerical parameters. As a result, it is difficult to analyze output translations and there is a real need for tools that could help developers to better understand the various causes of errors. In this study, we make a step in that direction and present an attempt to evaluate the quality of the phrase-based translation model. In order to identify those translation errors that stem from deficiencies in the phrase table (PT), we propose to compute the oracle BLEU-4 score, that is the best score that a system based on this PT can achieve on a reference corpus. By casting the computation of the oracle BLEU-1 as an Integer Linear Programming (ILP) problem, we show that it is possible to efficiently compute accurate lower-bounds of this score, and report measures performed on several standard benchmarks. Various other applications of these oracle decoding techniques are also reported and discussed. 1 Phrase-Based Machine Translation 1.1 Principle A Phrase-Based Translation System (PBTS) consists of a ruleset and a scoring function (Lopez, 2009). The ruleset, represented in the phrase table, is a set of phrase1pairs {(f, e) }, each pair expressing that the source phrase f can ,bee) r}e,w earicthten p (atirra enxslparteedss)i inngto t a target phrase e. Trarsaens flation hypotheses are generated by iteratively rewriting portions of the source sentence as prescribed by the ruleset, until each source word has been consumed by exactly one rule. The order of target words in an hypothesis is uniquely determined by the order in which the rewrite operation are performed. The search space ofthe translation model corresponds to the set of all possible sequences of 1Following the usage in statistical machine translation literature, use “phrase” to denote a subsequence of consecutive words. we 933 rules applications. The scoring function aims to rank all possible translation hypotheses in such a way that the best one has the highest score. A PBTS is learned from a parallel corpus in two independent steps. In a first step, the corpus is aligned at the word level, by using alignment tools such as Gi z a++ (Och and Ney, 2003) and some symmetrisation heuristics; phrases are then extracted by other heuristics (Koehn et al., 2003) and assigned numerical weights. In the second step, the parameters of the scoring function are estimated, typically through Minimum Error Rate training (Och, 2003). Translating a sentence amounts to finding the best scoring translation hypothesis in the search space. Because of the combinatorial nature of this problem, translation has to rely on heuristic search techniques such as greedy hill-climbing (Germann, 2003) or variants of best-first search like multi-stack decoding (Koehn, 2004). Moreover, to reduce the overall complexity of decoding, the search space is typically pruned using simple heuristics. For instance, the state-of-the-art phrase-based decoder Moses (Koehn et al., 2007) considers only a restricted number of translations for each source sequence2 and enforces a distortion limit3 over which phrases can be reordered. As a consequence, the best translation hypothesis returned by the decoder is not always the one with the highest score. 1.2 Typology of PBTS Errors Analyzing the errors of a SMT system is not an easy task, because of the number of models that are combined, the size of these models, and the high complexity of the various decision making processes. For a SMT system, three different kinds of errors can be distinguished (Germann et al., 2004; Auli et al., 2009): search errors, induction errors and model errors. The former corresponds to cases where the hypothesis with the best score is missed by the search procedure, either because of the use of an ap2the 3the option of Moses, defaulting to 20. dl option of Moses, whose default value is 7. tt l ProceMedITin,g Ms oasfs thaceh 2u0se1t0ts C,o UnSfAer,e n9c-e11 on O Ectmobpeir ic 2a0l1 M0.e ?tc ho2d0s10 in A Nsastouciraatlio Lnan fogru Cagoem Ppruotcaetisosninagl, L pinaggeusis 9t3ic3s–943, proximate search method or because of the restrictions of the search space. Induction errors correspond to cases where, given the model, the search space does not contain the reference. Finally, model errors correspond to cases where the hypothesis with the highest score is not the best translation according to the evaluation metric. Model errors encompass several types oferrors that occur during learning (Bottou and Bousquet, 2008)4. Approximation errors are errors caused by the use of a restricted and oversimplistic class of functions (here, finitestate transducers to model the generation of hypotheses and a linear scoring function to discriminate them) to model the translation process. Estimation errors correspond to the use of sub-optimal values for both the phrase pairs weights and the parameters of the scoring function. The reasons behind these errors are twofold: first, training only considers a finite sample of data; second, it relies on error prone alignments. As a result, some “good” phrases are extracted with a small weight, or, in the limit, are not extracted at all; and conversely that some “poor” phrases are inserted into the phrase table, sometimes with a really optimistic score. Sorting out and assessing the impact of these various causes of errors is of primary interest for SMT system developers: for lack of such diagnoses, it is difficult to figure out which components of the system require the most urgent attention. Diagnoses are however, given the tight intertwining among the various component of a system, very difficult to obtain: most evaluations are limited to the computation of global scores and usually do not imply any kind of failure analysis. 1.3 Contribution and organization To systematically assess the impact of the multiple heuristic decisions made during training and decoding, we propose, following (Dreyer et al., 2007; Auli et al., 2009), to work out oracle scores, that is to evaluate the best achievable performances of a PBTS. We aim at both studying the expressive power of PBTS and at providing tools for identifying and quantifying causes of failure. Under standard metrics such as BLEU (Papineni et al., 2002), oracle scores are difficult (if not impossible) to compute, but, by casting the computation of the oracle unigram recall and precision as an Integer Linear Programming (ILP) problem, we show that it is possible to efficiently compute accurate lower-bounds of the oracle BLEU-4 scores and report measurements performed on several standard benchmarks. The main contributions of this paper are twofold. We first introduce an ILP program able to efficiently find the best hypothesis a PBTS can achieve. This program can be easily extended to test various improvements to 4We omit here optimization errors. 934 phrase-base systems or to evaluate the impact of different parameter settings. Second, we present a number of complementary results illustrating the usage of our oracle decoder for identifying and analyzing PBTS errors. Our experimental results confirm the main conclusions of (Turchi et al., 2008), showing that extant PBTs have the potential to generate hypotheses having very high BLEU4 score and that their main bottleneck is their scoring function. The rest of this paper is organized as follows: in Section 2, we introduce and formalize the oracle decoding problem, and present a series of ILP problems of increasing complexity designed so as to deliver accurate lowerbounds of oracle score. This section closes with various extensions allowing to model supplementary constraints, most notably reordering constraints (Section 2.5). Our experiments are reported in Section 3, where we first introduce the training and test corpora, along with a description of our system building pipeline (Section 3. 1). We then discuss the baseline oracle BLEU scores (Section 3.2), analyze the non-reachable parts of the reference translations, and comment several complementary results which allow to identify causes of failures. Section 4 discuss our approach and findings with respect to the existing literature on error analysis and oracle decoding. We conclude and discuss further prospects in Section 5. 2 Oracle Decoder 2.1 The Oracle Decoding Problem Definition To get some insights on the errors of phrasebased systems and better understand their limits, we propose to consider the oracle decoding problem defined as follows: given a source sentence, its reference translation5 and a phrase table, what is the “best” translation hypothesis a system can generate? As usual, the quality of an hypothesis is evaluated by the similarity between the reference and the hypothesis. Note that in the oracle decoding problem, we are only assessing the ability of PBT systems to generate good candidate translations, irrespective of their ability to score them properly. We believe that studying this problem is interesting for various reasons. First, as described in Section 3.4, comparing the best hypothesis a system could have generated and the hypothesis it actually generates allows us to carry on both quantitative and qualitative failure analysis. The oracle decoding problem can also be used to assess the expressive power of phrase-based systems (Auli et al., 2009). Other applications include computing acceptable pseudo-references for discriminative training (Tillmann and Zhang, 2006; Liang et al., 2006; Arun and 5The oracle decoding problem can be extended to the case of multiple references. For the sake of simplicity, we only describe the case of a single reference. Koehn, 2007) or combining machine translation systems in a multi-source setting (Li and Khudanpur, 2009). We have also used oracle decoding to identify erroneous or difficult to translate references (Section 3.3). Evaluation Measure To fully define the oracle decoding problem, a measure of the similarity between a translation hypothesis and its reference translation has to be chosen. The most obvious choice is the BLEU-4 score (Papineni et al., 2002) used in most machine translation evaluations. However, using this metric in the oracle decoding problem raises several issues. First, BLEU-4 is a metric defined at the corpus level and is hard to interpret at the sentence level. More importantly, BLEU-4 is not decomposable6: as it relies on 4-grams statistics, the contribution of each phrase pair to the global score depends on the translation of the previous and following phrases and can not be evaluated in isolation. Because of its nondecomposability, maximizing BLEU-4 is hard; in particular, the phrase-level decomposability of the evaluation × metric is necessary in our approach. To circumvent this difficulty, we propose to evaluate the similarity between a translation hypothesis and a reference by the number of their common words. This amounts to evaluating translation quality in terms of unigram precision and recall, which are highly correlated with human judgements (Lavie et al., ). This measure is closely related to the BLEU-1 evaluation metric and the Meteor (Banerjee and Lavie, 2005) metric (when it is evaluated without considering near-matches and the distortion penalty). We also believe that hypotheses that maximize the unigram precision and recall at the sentence level yield corpus level BLEU-4 scores close the maximal achievable. Indeed, in the setting we will introduce in the next section, BLEU-1 and BLEU-4 are highly correlated: as all correct words of the hypothesis will be compelled to be at their correct position, any hypothesis with a high 1-gram precision is also bound to have a high 2-gram precision, etc. 2.2 Formalizing the Oracle Decoding Problem The oracle decoding problem has already been considered in the case of word-based models, in which all translation units are bound to contain only one word. The problem can then be solved by a bipartite graph matching algorithm (Leusch et al., 2008): given a n m binary matarligxo describing possible t 2r0an08sl)a:ti goinv elinn aks n b×emtw beeinna source words and target words7, this algorithm finds the subset of links maximizing the number of words of the reference that have been translated, while ensuring that each word 6Neither at the sentence (Chiang et al., 2008), nor at the phrase level. 7The (i, j) entry of the matrix is 1if the ith word of the source can be translated by the jth word of the reference, 0 otherwise. 935 is translated only once. Generalizing this approach to phrase-based systems amounts to solving the following problem: given a set of possible translation links between potential phrases of the source and of the target, find the subset of links so that the unigram precision and recall are the highest possible. The corresponding oracle hypothesis can then be easily generated by selecting the target phrases that are aligned with one source phrase, disregarding the others. In addition, to mimic the way OOVs are usually handled, we match identical OOV tokens appearing both in the source and target sentences. In this approach, the unigram precision is always one (every word generated in the oracle hypothesis matches exactly one word in the reference). As a consequence, to find the oracle hypothesis, we just have to maximize the recall, that is the number of words appearing both in the hypothesis and in the reference. Considering phrases instead of isolated words has a major impact on the computational complexity: in this new setting, the optimal segmentations in phrases of both the source and of the target have to be worked out in addition to links selection. Moreover, constraints have to be taken into account so as to enforce a proper segmentation of the source and target sentences. These constraints make it impossible to use the approach of (Leusch et al., 2008) and concur in making the oracle decoding problem for phrase-based models more complex than it is for word-based models: it can be proven, using arguments borrowed from (De Nero and Klein, 2008), that this problem is NP-hard even for the simple unigram precision measure. 2.3 An Integer Program for Oracle Decoding To solve the combinatorial problem introduced in the previous section, we propose to cast it into an Integer Linear Programming (ILP) problem, for which many generic solvers exist. ILP has already been used in SMT to find the optimal translation for word-based (Germann et al., 2001) and to study the complexity of learning phrase alignments (De Nero and Klein, 2008) models. Following the latter reference, we introduce the following variables: fi,j (resp. ek,l) is a binary indicator variable that is true when the phrase contains all spans from betweenword position i to j (resp. k to l) of the source (resp. target) sentence. We also introduce a binary variable, denoted ai,j,k,l, to describe a possible link between source phrase fi,j and target phrase ek,l. These variables are built from the entries of the phrase table according to selection strategies introduced in Section 2.4. In the following, index variables are so that: 0 ≤ i< j ≤ n, in the source sentence and 0 ≤ k < l ≤ m, in the target sentence, where n (resp. m) is the length of the source (resp. target) sentence. Solving the oracle decoding problem then amounts to optimizing the following objective function: mi,j,akx,li,Xj,k,lai,j,k,l· (l − k), (1) under the constraints: X ∀x ∈ J1,mK : ek,l ≤ 1 (2) = (3) 1∀,kn,lK : Xai,j,k,l = fk,l (4) ∀i,j : Xai,j,k,l (5) k,l s.tX. Xk≤x≤l ∀∀xy ∈∈ J11,,mnKK : X i,j s.tX. Xi≤y≤j fi,j 1 Xi,j = ei,j Xk,l The objective function (1) corresponds to the number of target words that are generated. The first set of constraints (2) ensures that each word in the reference e ap- pears in no more than one phrase. Maximizing the objective under these constraints amounts to maximizing the unigram recall. The second set of constraints (3) ensures that each word in the source f is translated exactly once, which guarantees that the search space of the ILP problem is the same as the search space of a phrase-based system. Constraints (4) bind the fk,l and ai,j,k,l variables, ensuring that whenever a link ai,j,k,l is active, the corresponding phrase fk,l is also active. Constraints (5) play a similar role for the reference. The Relaxed Problem Even though it accurately models the search space of a phrase-based decoder, this programs is not really useful as is: due to out-ofvocabulary words or missing entries in the phrase table, the constraint that all source words should be translated yields infeasible problems8. We propose to relax this problem and allow some source words to remain untranslated. This is done by replacing constraints (3) by: ∀y ∈ J1,nK : X i,j s.tX. Xi≤y≤j fi,j ≤ 1 To better ref∀lyec ∈t th J1e, bneKh :avior of phrase-based decoders, which attempt to translate all source words, we also need to modify the objective function as follows: X i,Xj,k,l ai,j,k,l · (l − k) +Xfi,j · (j − i) Xi,j (6) The second term in this new objective ensures that optimal solutions translate as many source words as possible. 8An ILP problem is said to be infeasible when tion violates at least one constraint. every possible solu- 936 The Relaxed-Distortion Problem A last caveat with the Relaxed optimization program is caused by frequently occurring source tokens, such as function words or punctuation signs, which can often align with more than one target word. For lack of taking distortion information into account in our objective function, all these alignments are deemed equivalent, even if some of them are clearly more satisfactory than others. This situation is illustrated on Figure 1. le chat et the cat and le the chien dog Figure 1: Equivalent alignments between “le” and “the”. The dashed lines corresponds to a less interpretable solution. To overcome this difficulty, we propose a last change to the objective function: X i,Xj,k,l ai,j,k,l · (l − k) +Xfi,j · (j − i) X ai,j,k,l|k − i| Xi,j −α (7) i Xk ,l X,j, Compared to the objective function of the relaxed problem (6), we introduce here a supplementary penalty factor which favors monotonous alignments. For each phrase pair, the higher the difference between source and target positions, the higher this penalty. If α is small enough, this extra term allows us to select, among all the optimal alignments of the re l axed problem, the one with the lowest distortion. In our experiments, we set α to min {n, m} to ensure that the penalty factor is always smminall{enr, ,tmha}n tthoe e rneswuarred t fhoart aligning atwltyo single iwso ardlwsa. 2.4 Selecting Indicator Variables In the approach introduced in the previous sections, the oracle decoding problem is solved by selecting, among a set of possible translation links, the ones that yield the solution with the highest unigram recall. We propose two strategies to build this set of possible translation links. In the first one, denoted exact match, an indicator ai,j,k,l is created if there is an entry (f, e) so that f spans from word position ito j in the source and e from word position k to l in the target. In this strategy, the ILP program considers exactly the same ruleset as conventional phrase-based decoders. We also consider an alternative strategy, which could help us to identify errors made during the phrase extraction process. In this strategy, denoted inside match, an indicator ai,j,k,l is created when the following three criteria are met: i) f spans from position ito j of the source; ii) a substring of e, denoted e, spans from position k to l of the reference; iii) (f, e¯) is not an entry of the phrase table. The resulting set of indicator variables thus contains, at least, all the variables used in the exact match strategy. In addition, we license here the use of phrases containing words that do not occur in the reference. In fact, using such solutions can yield higher BLEU scores when the reward for additional correct matches exceeds the cost incurred by wrong predictions. These cases are symptoms of situations where the extraction heuristic failed to extract potentially useful subphrases. 2.5 Oracle Decoding with Reordering Constraints The ILP problem introduced in the previous section can be extended in several ways to describe and test various improvements to phrase-based systems or to evaluate the impact of different parameter settings. This flexibility mainly stems from the possibility offered by our framework to express arbitrary constraints over variables. In this section, we illustrate these possibilities by describing how reordering constraints can easily be considered. As a first example, the Moses decoder uses a distortion limit to constrain the set of possible reorderings. This constraint “enforces (...) that the last word of a phrase chosen for translation cannot be more than d9 words from the leftmost untranslated word in the source” (Lopez, 2009) and is expressed as: ∀aijkl , ai0j0k0l0 s.t. k > k0, aijkl · ai0j0k0l0 · |j − i0 + 1| ≤ d, The maximum distortion limit strategy (Lopez, 2009) is also easily expressed and take the following form (assuming this constraint is parameterized by d): ∀l < m − 1, ai,j,k,l·ai0,j0,l+1,l0 · |i0 − j − 1| 71is%t e6hs.a distortion greater that Moses default distortion limit. alignment decisions enabled by the use of larger training corpora and phrase table. To evaluate the impact ofthe second heuristic, we computed the number of phrases discarded by Moses (be- cause of the default ttl limit) but used in the oracle hypotheses. In the English to French NEWSCO setting, they account for 34.11% of the total number of phrases used in the oracle hypotheses. When the oracle decoder is constrained to use the same phrase table as Moses, its BLEU-4 score drops to 42.78. This shows that filtering the phrase table prior to decoding discards many useful phrase pairs and is seriously limiting the best achievable performance, a conclusion shared with (Auli et al., 2009). Search Errors Search errors can be identified by comparing the score of the best hypothesis found by Moses and the score of the oracle hypothesis. If the score of the oracle hypothesis is higher, then there has been a search error; on the contrary, there has been an estimation error when the score of the oracle hypothesis is lower than the score of the best hypothesis found by Moses. 940 Based on the comparison of the score of Moses hypotheses and of oracle hypotheses for the English to French NEWSCO setting, our preliminary conclusion is that the number of search errors is quite limited: only about 5% of the hypotheses of our oracle decoder are actually getting a better score than Moses solutions. Again, this shows that the scoring function (model error) is one of the main bottleneck of current PBTS. Comparing these hypotheses is nonetheless quite revealing: while Moses mostly selects phrase pairs with high translation scores and generates monotonous alignments, our ILP decoder uses larger reorderings and less probable phrases to achieve better solutions: on average, the reordering score of oracle solutions is −5.74, compared to −76.78 fscoro rMeo osfe osr outputs. iGonivsen is −the5 weight assigned through MERT training to the distortion score, no wonder that these hypotheses are severely penalized. The Impact of Phrase Length The observed outputs do not only depend on decisions made during the search, but also on decisions made during training. One such decision is the specification of maximal length for the source and target phrases. In our framework, evaluating the impact of this decision is simple: it suffices to change the definition of indicator variables so as to consider only alignments between phrases of a given length. In the English-French NEWSCO setting, the most restrictive choice, when only alignments between single words are authorized, yields an oracle BLEU-4 of 48.68; however, authorizing phrases up to length 2 allows to achieve an oracle value of 66.57, very close to the score achieved when considering all extracted phrases (67.77). This is corroborated with a further analysis of our oracle alignments, which use phrases whose average source length is 1.21 words (respectively 1.31 for target words). If many studies have already acknowledged the predomi- nance of “small” phrases in actual translations, our oracle scores suggest that, for this language pair, increasing the phrase length limit beyond 2 or 3 might be a waste of computational resources. 4 Related Work To the best of our knowledge, there are only a few works that try to study the expressive power ofphrase-based machine translation systems or to provide tools for analyzing potential causes of failure. The approach described in (Auli et al., 2009) is very similar to ours: in this study, the authors propose to find and analyze the limits of machine translation systems by studying the reference reachability. A reference is reachable for a given system if it can be exactly generated by this system. Reference reachability is assessed using Moses in forced decoding mode: during search, all hypotheses that deviate from the reference are simply discarded. Even though the main goal of this study was to compare the search space of phrase-based and hierarchical systems, it also provides some insights on the impact of various search parameters in Moses, delivering conclusions that are consistent with our main results. As described in Section 1.2, these authors also propose a typology of the errors of a statistical translation systems, but do not attempt to provide methods for identifying them. The authors of (Turchi et al., 2008) study the learn- ing capabilities of Moses by extensively analyzing learning curves representing the translation performances as a function of the number of examples, and by corrupting the model parameters. Even though their focus is more on assessing the scoring function, they reach conclusions similar to ours: the current bottleneck of translation performances is not the representation power of the PBTS but rather in their scoring functions. Oracle decoding is useful to compute reachable pseudo-references in the context of discriminative training. This is the main motivation of (Tillmann and Zhang, 2006), where the authors compute high BLEU hypotheses by running a conventional decoder so as to maximize a per-sentence approximation of BLEU-4, under a simple (local) reordering model. Oracle decoding has also been used to assess the limitations induced by various reordering constraints in (Dreyer et al., 2007). To this end, the authors propose to use a beam-search based oracle decoder, which computes lower bounds of the best achievable BLEU-4 using dynamic programming techniques over finite-state (for so-called local and IBM constraints) or hierarchically structured (for ITG constraints) sets of hypotheses. Even 941 though the numbers reported in this study are not directly comparable with ours17, it seems that our decoder is not only conceptually much simpler, but also achieves much more optimistic lower-bounds of the oracle BLEU score. The approach described in (Li and Khudanpur, 2009) employs a similar technique, which is to guide a heuristic search in an hypergraph representing possible translation hypotheses with n-gram counts matches, which amounts to decoding with a n-gram model trained on the sole reference translation. Additional tricks are presented in this article to speed-up decoding. Computing oracle BLEU scores is also the subject of (Zens and Ney, 2005; Leusch et al., 2008), yet with a different emphasis. These studies are concerned with finding the best hypotheses in a word graph or in a consensus network, a problem that has various implications for multi-pass decoding and/or system combination techniques. The former reference describes an exponential approximate algorithm, while the latter proves the NPcompleteness of this problem and discuss various heuristic approaches. Our problem is somewhat more complex and using their techniques would require us to built word graphs containing all the translations induced by arbitrary segmentations and permutations of the source sentence. 5 Conclusions In this paper, we have presented a methodology for analyzing the errors of PBTS, based on the computation of an approximation of the BLEU-4 oracle score. We have shown that this approximation could be computed fairly accurately and efficiently using Integer Linear Programming techniques. Our main result is a confirmation of the fact that extant PBTS systems are expressive enough to achieve very high translation performance with respect to conventional quality measurements. The main efforts should therefore strive to improve on the way phrases and hypotheses are scored during training. This gives further support to attempts aimed at designing context-dependent scoring functions as in (Stroppa et al., 2007; Gimpel and Smith, 2008), or at attempts to perform discriminative training of feature-rich models. (Bangalore et al., 2007). We have shown that the examination of difficult-totranslate sentences was an effective way to detect errors or inconsistencies in the reference translations, making our approach a potential aid for controlling the quality or assessing the difficulty of test data. Our experiments have also highlighted the impact of various parameters. Various extensions of the baseline ILP program have been suggested and/or evaluated. In particular, the ILP formalism lends itself well to expressing various constraints that are typically used in conventional PBTS. In 17The best BLEU-4 oracle they achieve on Europarl German to English is approximately 48; but they considered a smaller version of the training corpus and the WMT’06 test set. our future work, we aim at using this ILP framework to systematically assess various search configurations. We plan to explore how replacing non-reachable references with high-score pseudo-references can improve discrim- inative training of PBTS. We are also concerned by determining how tight is our approximation of the BLEU4 score is: to this end, we intend to compute the best BLEU-4 score within the n-best solutions of the oracle decoding problem. Acknowledgments Warm thanks to Houda Bouamor for helping us with the annotation tool. This work has been partly financed by OSEO, the French State Agency for Innovation, under the Quaero program. References Tobias Achterberg. 2007. Constraint Integer Programming. Ph.D. thesis, Technische Universit a¨t Berlin. http : / / opus .kobv .de /tuberl in/vol ltexte / 2 0 0 7 / 16 11/ . Abhishek Arun and Philipp Koehn. 2007. Online learning methods for discriminative training of phrase based statistical machine translation. In Proc. of MT Summit XI, Copenhagen, Denmark. Michael Auli, Adam Lopez, Hieu Hoang, and Philipp Koehn. 2009. A systematic analysis of translation model search spaces. In Proc. of WMT, pages 224–232, Athens, Greece. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proc. of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan. Srinivas Bangalore, Patrick Haffner, and Stephan Kanthak. 2007. Statistical machine translation through global lexical selection and sentence reconstruction. In Proc. of ACL, pages 152–159, Prague, Czech Republic. L e´on Bottou and Olivier Bousquet. 2008. The tradeoffs oflarge scale learning. In Proc. of NIPS, pages 161–168, Vancouver, B.C., Canada. Chris Callison-Burch, Philipp Koehn, Christof Monz, and Josh Schroeder. 2009. Findings of the 2009 Workshop on Statistical Machine Translation. In Proc. of WMT, pages 1–28, Athens, Greece. David Chiang, Steve DeNeefe, Yee Seng Chan, and Hwee Tou Ng. 2008. Decomposability of translation metrics for improved evaluation and efficient algorithms. In Proc. of ECML, pages 610–619, Honolulu, Hawaii. John De Nero and Dan Klein. 2008. The complexity of phrase alignment problems. In Proc. of ACL: HLT, Short Papers, pages 25–28, Columbus, Ohio. Markus Dreyer, Keith B. Hall, and Sanjeev P. Khudanpur. 2007. Comparing reordering constraints for smt using efficient bleu oracle computation. In NAACL-HLT/AMTA Workshop on Syntax and Structure in Statistical Translation, pages 103– 110, Rochester, New York. 942 Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, and Kenji Yamada. 2001 . Fast decoding and optimal decoding for machine translation. In Proc. of ACL, pages 228–235, Toulouse, France. 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 Proc. of NAACL, pages 1–8, Edmonton, Canada. Kevin Gimpel and Noah A. Smith. 2008. Rich source-side context for statistical machine translation. In Proc. of WMT, pages 9–17, Columbus, Ohio. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proc. of NAACL, pages 48–54, Edmonton, Canada. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris CallisonBurch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proc. of ACL, demonstration session. Philipp Koehn. 2004. Pharaoh: A beam search decoder for phrase-based statistical machine translation models. In Proc. of AMTA, pages 115–124, Washington DC. Shankar Kumar and William Byrne. 2005. Local phrase reordering models for statistical machine translation. In Proc. of HLT, pages 161–168, Vancouver, Canada. Alon Lavie, Kenji Sagae, and Shyamsundar Jayaraman. The significance of recall in automatic metrics for MT evaluation. In In Proc. of AMTA, pages 134–143, Washington DC. Gregor Leusch, Evgeny Matusov, and Hermann Ney. 2008. Complexity of finding the BLEU-optimal hypothesis in a confusion network. In Proc. of EMNLP, pages 839–847, Honolulu, Hawaii. Zhifei Li and Sanjeev Khudanpur. 2009. Efficient extraction of oracle-best translations from hypergraphs. In Proc. of NAACL, pages 9–12, Boulder, Colorado. Percy Liang, Alexandre Bouchard-C oˆt´ e, Dan Klein, and Ben Taskar. 2006. An end-to-end discriminative approach to machine translation. In Proc. of ACL, pages 761–768, Sydney, Australia. Adam Lopez. 2009. Translation as weighted deduction. In Proc. of EACL, pages 532–540, Athens, Greece. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Comput. Linguist. , 29(1): 19–5 1. Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proc. of ACL, pages 160–167, Sapporo, Japan. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-jing Zhu. 2002. Bleu: A method for automatic evaluation of machine translation. Technical report, Philadelphia, Pennsylvania. D. Roth and W. Yih. 2005. Integer linear programming inference for conditional random fields. In Proc. of ICML, pages 737–744, Bonn, Germany. Nicolas Stroppa, Antal van den Bosch, and Andy Way. 2007. Exploiting source similarity for smt using context-informed features. In Andy Way and Barbara Proc. of TMI, pages Christoph Tillmann 231–240, Sk¨ ovde, and Tong Zhang. Gawronska, editors, Sweden. 2006. A discriminative global training algorithm for statistical mt. In Proc. of ACL, 721–728, Sydney, Australia. Turchi, Tijl De Bie, and Nello pages Marco Cristianini. 2008. Learn- ing performance of a machine translation system: a statistical and computational analysis. In Proc. of WMT, pages Columbus, Ohio. 35–43, Richard Zens and Hermann Ney. 2005. Word graphs for statistical machine translation. In Proc. of the ACL Workshop on Building and Using Parallel Texts, pages 191–198, Ann Arbor, Michigan. 943
4 0.65587497 39 emnlp-2010-EMNLP 044
Author: George Foster
Abstract: We describe a new approach to SMT adaptation that weights out-of-domain phrase pairs according to their relevance to the target domain, determined by both how similar to it they appear to be, and whether they belong to general language or not. This extends previous work on discriminative weighting by using a finer granularity, focusing on the properties of instances rather than corpus components, and using a simpler training procedure. We incorporate instance weighting into a mixture-model framework, and find that it yields consistent improvements over a wide range of baselines.
5 0.62317783 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
Author: Sankaranarayanan Ananthakrishnan ; Rohit Prasad ; David Stallard ; Prem Natarajan
Abstract: Production of parallel training corpora for the development of statistical machine translation (SMT) systems for resource-poor languages usually requires extensive manual effort. Active sample selection aims to reduce the labor, time, and expense incurred in producing such resources, attaining a given performance benchmark with the smallest possible training corpus by choosing informative, nonredundant source sentences from an available candidate pool for manual translation. We present a novel, discriminative sample selection strategy that preferentially selects batches of candidate sentences with constructs that lead to erroneous translations on a held-out development set. The proposed strategy supports a built-in diversity mechanism that reduces redundancy in the selected batches. Simulation experiments on English-to-Pashto and Spanish-to-English translation tasks demon- strate the superiority of the proposed approach to a number of competing techniques, such as random selection, dissimilarity-based selection, as well as a recently proposed semisupervised active learning strategy.
6 0.54593378 29 emnlp-2010-Combining Unsupervised and Supervised Alignments for MT: An Empirical Study
7 0.53702289 22 emnlp-2010-Automatic Evaluation of Translation Quality for Distant Language Pairs
8 0.52949744 47 emnlp-2010-Example-Based Paraphrasing for Improved Phrase-Based Statistical Machine Translation
9 0.5251618 50 emnlp-2010-Facilitating Translation Using Source Language Paraphrase Lattices
10 0.49611545 63 emnlp-2010-Improving Translation via Targeted Paraphrasing
11 0.47122616 99 emnlp-2010-Statistical Machine Translation with a Factorized Grammar
12 0.41162214 92 emnlp-2010-Predicting the Semantic Compositionality of Prefix Verbs
13 0.39020193 57 emnlp-2010-Hierarchical Phrase-Based Translation Grammars Extracted from Alignment Posterior Probabilities
14 0.38222113 42 emnlp-2010-Efficient Incremental Decoding for Tree-to-String Translation
15 0.38040045 33 emnlp-2010-Cross Language Text Classification by Model Translation and Semi-Supervised Learning
16 0.37983808 10 emnlp-2010-A Probabilistic Morphological Analyzer for Syriac
17 0.36190352 9 emnlp-2010-A New Approach to Lexical Disambiguation of Arabic Text
18 0.34667367 1 emnlp-2010-"Poetic" Statistical Machine Translation: Rhyme and Meter
19 0.33691996 76 emnlp-2010-Maximum Entropy Based Phrase Reordering for Hierarchical Phrase-Based Translation
20 0.32437754 36 emnlp-2010-Discriminative Word Alignment with a Function Word Reordering Model
topicId topicWeight
[(10, 0.466), (12, 0.023), (29, 0.081), (30, 0.02), (52, 0.072), (56, 0.057), (62, 0.011), (66, 0.102), (72, 0.033), (76, 0.018), (92, 0.017)]
simIndex simValue paperId paperTitle
1 0.98194289 59 emnlp-2010-Identifying Functional Relations in Web Text
Author: Thomas Lin ; Mausam ; Oren Etzioni
Abstract: Determining whether a textual phrase denotes a functional relation (i.e., a relation that maps each domain element to a unique range element) is useful for numerous NLP tasks such as synonym resolution and contradiction detection. Previous work on this problem has relied on either counting methods or lexico-syntactic patterns. However, determining whether a relation is functional, by analyzing mentions of the relation in a corpus, is challenging due to ambiguity, synonymy, anaphora, and other linguistic phenomena. We present the LEIBNIZ system that overcomes these challenges by exploiting the synergy between the Web corpus and freelyavailable knowledge resources such as Freebase. It first computes multiple typedfunctionality scores, representing functionality of the relation phrase when its arguments are constrained to specific types. It then aggregates these scores to predict the global functionality for the phrase. LEIBNIZ outperforms previous work, increasing area under the precisionrecall curve from 0.61 to 0.88. We utilize LEIBNIZ to generate the first public repository of automatically-identified functional relations.
2 0.86469585 74 emnlp-2010-Learning the Relative Usefulness of Questions in Community QA
Author: Razvan Bunescu ; Yunfeng Huang
Abstract: We present a machine learning approach for the task of ranking previously answered questions in a question repository with respect to their relevance to a new, unanswered reference question. The ranking model is trained on a collection of question groups manually annotated with a partial order relation reflecting the relative utility of questions inside each group. Based on a set of meaning and structure aware features, the new ranking model is able to substantially outperform more straightforward, unsupervised similarity measures.
same-paper 3 0.75601768 5 emnlp-2010-A Hybrid Morpheme-Word Representation for Machine Translation of Morphologically Rich Languages
Author: Minh-Thang Luong ; Preslav Nakov ; Min-Yen Kan
Abstract: We propose a language-independent approach for improving statistical machine translation for morphologically rich languages using a hybrid morpheme-word representation where the basic unit of translation is the morpheme, but word boundaries are respected at all stages of the translation process. Our model extends the classic phrase-based model by means of (1) word boundary-aware morpheme-level phrase extraction, (2) minimum error-rate training for a morpheme-level translation model using word-level BLEU, and (3) joint scoring with morpheme- and word-level language models. Further improvements are achieved by combining our model with the classic one. The evaluation on English to Finnish using Europarl (714K sentence pairs; 15.5M English words) shows statistically significant improvements over the classic model based on BLEU and human judgments.
4 0.49840045 51 emnlp-2010-Function-Based Question Classification for General QA
Author: Fan Bu ; Xingwei Zhu ; Yu Hao ; Xiaoyan Zhu
Abstract: In contrast with the booming increase of internet data, state-of-art QA (question answering) systems, otherwise, concerned data from specific domains or resources such as search engine snippets, online forums and Wikipedia in a somewhat isolated way. Users may welcome a more general QA system for its capability to answer questions of various sources, integrated from existed specialized sub-QA engines. In this framework, question classification is the primary task. However, the current paradigms of question classification were focused on some specified type of questions, i.e. factoid questions, which are inappropriate for the general QA. In this paper, we propose a new question classification paradigm, which includes a question taxonomy suitable to the general QA and a question classifier based on MLN (Markov logic network), where rule-based methods and statistical methods are unified into a single framework in a fuzzy discriminative learning approach. Experiments show that our method outperforms traditional question classification approaches.
5 0.39474207 18 emnlp-2010-Assessing Phrase-Based Translation Models with Oracle Decoding
Author: Guillaume Wisniewski ; Alexandre Allauzen ; Francois Yvon
Abstract: Extant Statistical Machine Translation (SMT) systems are very complex softwares, which embed multiple layers of heuristics and embark very large numbers of numerical parameters. As a result, it is difficult to analyze output translations and there is a real need for tools that could help developers to better understand the various causes of errors. In this study, we make a step in that direction and present an attempt to evaluate the quality of the phrase-based translation model. In order to identify those translation errors that stem from deficiencies in the phrase table (PT), we propose to compute the oracle BLEU-4 score, that is the best score that a system based on this PT can achieve on a reference corpus. By casting the computation of the oracle BLEU-1 as an Integer Linear Programming (ILP) problem, we show that it is possible to efficiently compute accurate lower-bounds of this score, and report measures performed on several standard benchmarks. Various other applications of these oracle decoding techniques are also reported and discussed. 1 Phrase-Based Machine Translation 1.1 Principle A Phrase-Based Translation System (PBTS) consists of a ruleset and a scoring function (Lopez, 2009). The ruleset, represented in the phrase table, is a set of phrase1pairs {(f, e) }, each pair expressing that the source phrase f can ,bee) r}e,w earicthten p (atirra enxslparteedss)i inngto t a target phrase e. Trarsaens flation hypotheses are generated by iteratively rewriting portions of the source sentence as prescribed by the ruleset, until each source word has been consumed by exactly one rule. The order of target words in an hypothesis is uniquely determined by the order in which the rewrite operation are performed. The search space ofthe translation model corresponds to the set of all possible sequences of 1Following the usage in statistical machine translation literature, use “phrase” to denote a subsequence of consecutive words. we 933 rules applications. The scoring function aims to rank all possible translation hypotheses in such a way that the best one has the highest score. A PBTS is learned from a parallel corpus in two independent steps. In a first step, the corpus is aligned at the word level, by using alignment tools such as Gi z a++ (Och and Ney, 2003) and some symmetrisation heuristics; phrases are then extracted by other heuristics (Koehn et al., 2003) and assigned numerical weights. In the second step, the parameters of the scoring function are estimated, typically through Minimum Error Rate training (Och, 2003). Translating a sentence amounts to finding the best scoring translation hypothesis in the search space. Because of the combinatorial nature of this problem, translation has to rely on heuristic search techniques such as greedy hill-climbing (Germann, 2003) or variants of best-first search like multi-stack decoding (Koehn, 2004). Moreover, to reduce the overall complexity of decoding, the search space is typically pruned using simple heuristics. For instance, the state-of-the-art phrase-based decoder Moses (Koehn et al., 2007) considers only a restricted number of translations for each source sequence2 and enforces a distortion limit3 over which phrases can be reordered. As a consequence, the best translation hypothesis returned by the decoder is not always the one with the highest score. 1.2 Typology of PBTS Errors Analyzing the errors of a SMT system is not an easy task, because of the number of models that are combined, the size of these models, and the high complexity of the various decision making processes. For a SMT system, three different kinds of errors can be distinguished (Germann et al., 2004; Auli et al., 2009): search errors, induction errors and model errors. The former corresponds to cases where the hypothesis with the best score is missed by the search procedure, either because of the use of an ap2the 3the option of Moses, defaulting to 20. dl option of Moses, whose default value is 7. tt l ProceMedITin,g Ms oasfs thaceh 2u0se1t0ts C,o UnSfAer,e n9c-e11 on O Ectmobpeir ic 2a0l1 M0.e ?tc ho2d0s10 in A Nsastouciraatlio Lnan fogru Cagoem Ppruotcaetisosninagl, L pinaggeusis 9t3ic3s–943, proximate search method or because of the restrictions of the search space. Induction errors correspond to cases where, given the model, the search space does not contain the reference. Finally, model errors correspond to cases where the hypothesis with the highest score is not the best translation according to the evaluation metric. Model errors encompass several types oferrors that occur during learning (Bottou and Bousquet, 2008)4. Approximation errors are errors caused by the use of a restricted and oversimplistic class of functions (here, finitestate transducers to model the generation of hypotheses and a linear scoring function to discriminate them) to model the translation process. Estimation errors correspond to the use of sub-optimal values for both the phrase pairs weights and the parameters of the scoring function. The reasons behind these errors are twofold: first, training only considers a finite sample of data; second, it relies on error prone alignments. As a result, some “good” phrases are extracted with a small weight, or, in the limit, are not extracted at all; and conversely that some “poor” phrases are inserted into the phrase table, sometimes with a really optimistic score. Sorting out and assessing the impact of these various causes of errors is of primary interest for SMT system developers: for lack of such diagnoses, it is difficult to figure out which components of the system require the most urgent attention. Diagnoses are however, given the tight intertwining among the various component of a system, very difficult to obtain: most evaluations are limited to the computation of global scores and usually do not imply any kind of failure analysis. 1.3 Contribution and organization To systematically assess the impact of the multiple heuristic decisions made during training and decoding, we propose, following (Dreyer et al., 2007; Auli et al., 2009), to work out oracle scores, that is to evaluate the best achievable performances of a PBTS. We aim at both studying the expressive power of PBTS and at providing tools for identifying and quantifying causes of failure. Under standard metrics such as BLEU (Papineni et al., 2002), oracle scores are difficult (if not impossible) to compute, but, by casting the computation of the oracle unigram recall and precision as an Integer Linear Programming (ILP) problem, we show that it is possible to efficiently compute accurate lower-bounds of the oracle BLEU-4 scores and report measurements performed on several standard benchmarks. The main contributions of this paper are twofold. We first introduce an ILP program able to efficiently find the best hypothesis a PBTS can achieve. This program can be easily extended to test various improvements to 4We omit here optimization errors. 934 phrase-base systems or to evaluate the impact of different parameter settings. Second, we present a number of complementary results illustrating the usage of our oracle decoder for identifying and analyzing PBTS errors. Our experimental results confirm the main conclusions of (Turchi et al., 2008), showing that extant PBTs have the potential to generate hypotheses having very high BLEU4 score and that their main bottleneck is their scoring function. The rest of this paper is organized as follows: in Section 2, we introduce and formalize the oracle decoding problem, and present a series of ILP problems of increasing complexity designed so as to deliver accurate lowerbounds of oracle score. This section closes with various extensions allowing to model supplementary constraints, most notably reordering constraints (Section 2.5). Our experiments are reported in Section 3, where we first introduce the training and test corpora, along with a description of our system building pipeline (Section 3. 1). We then discuss the baseline oracle BLEU scores (Section 3.2), analyze the non-reachable parts of the reference translations, and comment several complementary results which allow to identify causes of failures. Section 4 discuss our approach and findings with respect to the existing literature on error analysis and oracle decoding. We conclude and discuss further prospects in Section 5. 2 Oracle Decoder 2.1 The Oracle Decoding Problem Definition To get some insights on the errors of phrasebased systems and better understand their limits, we propose to consider the oracle decoding problem defined as follows: given a source sentence, its reference translation5 and a phrase table, what is the “best” translation hypothesis a system can generate? As usual, the quality of an hypothesis is evaluated by the similarity between the reference and the hypothesis. Note that in the oracle decoding problem, we are only assessing the ability of PBT systems to generate good candidate translations, irrespective of their ability to score them properly. We believe that studying this problem is interesting for various reasons. First, as described in Section 3.4, comparing the best hypothesis a system could have generated and the hypothesis it actually generates allows us to carry on both quantitative and qualitative failure analysis. The oracle decoding problem can also be used to assess the expressive power of phrase-based systems (Auli et al., 2009). Other applications include computing acceptable pseudo-references for discriminative training (Tillmann and Zhang, 2006; Liang et al., 2006; Arun and 5The oracle decoding problem can be extended to the case of multiple references. For the sake of simplicity, we only describe the case of a single reference. Koehn, 2007) or combining machine translation systems in a multi-source setting (Li and Khudanpur, 2009). We have also used oracle decoding to identify erroneous or difficult to translate references (Section 3.3). Evaluation Measure To fully define the oracle decoding problem, a measure of the similarity between a translation hypothesis and its reference translation has to be chosen. The most obvious choice is the BLEU-4 score (Papineni et al., 2002) used in most machine translation evaluations. However, using this metric in the oracle decoding problem raises several issues. First, BLEU-4 is a metric defined at the corpus level and is hard to interpret at the sentence level. More importantly, BLEU-4 is not decomposable6: as it relies on 4-grams statistics, the contribution of each phrase pair to the global score depends on the translation of the previous and following phrases and can not be evaluated in isolation. Because of its nondecomposability, maximizing BLEU-4 is hard; in particular, the phrase-level decomposability of the evaluation × metric is necessary in our approach. To circumvent this difficulty, we propose to evaluate the similarity between a translation hypothesis and a reference by the number of their common words. This amounts to evaluating translation quality in terms of unigram precision and recall, which are highly correlated with human judgements (Lavie et al., ). This measure is closely related to the BLEU-1 evaluation metric and the Meteor (Banerjee and Lavie, 2005) metric (when it is evaluated without considering near-matches and the distortion penalty). We also believe that hypotheses that maximize the unigram precision and recall at the sentence level yield corpus level BLEU-4 scores close the maximal achievable. Indeed, in the setting we will introduce in the next section, BLEU-1 and BLEU-4 are highly correlated: as all correct words of the hypothesis will be compelled to be at their correct position, any hypothesis with a high 1-gram precision is also bound to have a high 2-gram precision, etc. 2.2 Formalizing the Oracle Decoding Problem The oracle decoding problem has already been considered in the case of word-based models, in which all translation units are bound to contain only one word. The problem can then be solved by a bipartite graph matching algorithm (Leusch et al., 2008): given a n m binary matarligxo describing possible t 2r0an08sl)a:ti goinv elinn aks n b×emtw beeinna source words and target words7, this algorithm finds the subset of links maximizing the number of words of the reference that have been translated, while ensuring that each word 6Neither at the sentence (Chiang et al., 2008), nor at the phrase level. 7The (i, j) entry of the matrix is 1if the ith word of the source can be translated by the jth word of the reference, 0 otherwise. 935 is translated only once. Generalizing this approach to phrase-based systems amounts to solving the following problem: given a set of possible translation links between potential phrases of the source and of the target, find the subset of links so that the unigram precision and recall are the highest possible. The corresponding oracle hypothesis can then be easily generated by selecting the target phrases that are aligned with one source phrase, disregarding the others. In addition, to mimic the way OOVs are usually handled, we match identical OOV tokens appearing both in the source and target sentences. In this approach, the unigram precision is always one (every word generated in the oracle hypothesis matches exactly one word in the reference). As a consequence, to find the oracle hypothesis, we just have to maximize the recall, that is the number of words appearing both in the hypothesis and in the reference. Considering phrases instead of isolated words has a major impact on the computational complexity: in this new setting, the optimal segmentations in phrases of both the source and of the target have to be worked out in addition to links selection. Moreover, constraints have to be taken into account so as to enforce a proper segmentation of the source and target sentences. These constraints make it impossible to use the approach of (Leusch et al., 2008) and concur in making the oracle decoding problem for phrase-based models more complex than it is for word-based models: it can be proven, using arguments borrowed from (De Nero and Klein, 2008), that this problem is NP-hard even for the simple unigram precision measure. 2.3 An Integer Program for Oracle Decoding To solve the combinatorial problem introduced in the previous section, we propose to cast it into an Integer Linear Programming (ILP) problem, for which many generic solvers exist. ILP has already been used in SMT to find the optimal translation for word-based (Germann et al., 2001) and to study the complexity of learning phrase alignments (De Nero and Klein, 2008) models. Following the latter reference, we introduce the following variables: fi,j (resp. ek,l) is a binary indicator variable that is true when the phrase contains all spans from betweenword position i to j (resp. k to l) of the source (resp. target) sentence. We also introduce a binary variable, denoted ai,j,k,l, to describe a possible link between source phrase fi,j and target phrase ek,l. These variables are built from the entries of the phrase table according to selection strategies introduced in Section 2.4. In the following, index variables are so that: 0 ≤ i< j ≤ n, in the source sentence and 0 ≤ k < l ≤ m, in the target sentence, where n (resp. m) is the length of the source (resp. target) sentence. Solving the oracle decoding problem then amounts to optimizing the following objective function: mi,j,akx,li,Xj,k,lai,j,k,l· (l − k), (1) under the constraints: X ∀x ∈ J1,mK : ek,l ≤ 1 (2) = (3) 1∀,kn,lK : Xai,j,k,l = fk,l (4) ∀i,j : Xai,j,k,l (5) k,l s.tX. Xk≤x≤l ∀∀xy ∈∈ J11,,mnKK : X i,j s.tX. Xi≤y≤j fi,j 1 Xi,j = ei,j Xk,l The objective function (1) corresponds to the number of target words that are generated. The first set of constraints (2) ensures that each word in the reference e ap- pears in no more than one phrase. Maximizing the objective under these constraints amounts to maximizing the unigram recall. The second set of constraints (3) ensures that each word in the source f is translated exactly once, which guarantees that the search space of the ILP problem is the same as the search space of a phrase-based system. Constraints (4) bind the fk,l and ai,j,k,l variables, ensuring that whenever a link ai,j,k,l is active, the corresponding phrase fk,l is also active. Constraints (5) play a similar role for the reference. The Relaxed Problem Even though it accurately models the search space of a phrase-based decoder, this programs is not really useful as is: due to out-ofvocabulary words or missing entries in the phrase table, the constraint that all source words should be translated yields infeasible problems8. We propose to relax this problem and allow some source words to remain untranslated. This is done by replacing constraints (3) by: ∀y ∈ J1,nK : X i,j s.tX. Xi≤y≤j fi,j ≤ 1 To better ref∀lyec ∈t th J1e, bneKh :avior of phrase-based decoders, which attempt to translate all source words, we also need to modify the objective function as follows: X i,Xj,k,l ai,j,k,l · (l − k) +Xfi,j · (j − i) Xi,j (6) The second term in this new objective ensures that optimal solutions translate as many source words as possible. 8An ILP problem is said to be infeasible when tion violates at least one constraint. every possible solu- 936 The Relaxed-Distortion Problem A last caveat with the Relaxed optimization program is caused by frequently occurring source tokens, such as function words or punctuation signs, which can often align with more than one target word. For lack of taking distortion information into account in our objective function, all these alignments are deemed equivalent, even if some of them are clearly more satisfactory than others. This situation is illustrated on Figure 1. le chat et the cat and le the chien dog Figure 1: Equivalent alignments between “le” and “the”. The dashed lines corresponds to a less interpretable solution. To overcome this difficulty, we propose a last change to the objective function: X i,Xj,k,l ai,j,k,l · (l − k) +Xfi,j · (j − i) X ai,j,k,l|k − i| Xi,j −α (7) i Xk ,l X,j, Compared to the objective function of the relaxed problem (6), we introduce here a supplementary penalty factor which favors monotonous alignments. For each phrase pair, the higher the difference between source and target positions, the higher this penalty. If α is small enough, this extra term allows us to select, among all the optimal alignments of the re l axed problem, the one with the lowest distortion. In our experiments, we set α to min {n, m} to ensure that the penalty factor is always smminall{enr, ,tmha}n tthoe e rneswuarred t fhoart aligning atwltyo single iwso ardlwsa. 2.4 Selecting Indicator Variables In the approach introduced in the previous sections, the oracle decoding problem is solved by selecting, among a set of possible translation links, the ones that yield the solution with the highest unigram recall. We propose two strategies to build this set of possible translation links. In the first one, denoted exact match, an indicator ai,j,k,l is created if there is an entry (f, e) so that f spans from word position ito j in the source and e from word position k to l in the target. In this strategy, the ILP program considers exactly the same ruleset as conventional phrase-based decoders. We also consider an alternative strategy, which could help us to identify errors made during the phrase extraction process. In this strategy, denoted inside match, an indicator ai,j,k,l is created when the following three criteria are met: i) f spans from position ito j of the source; ii) a substring of e, denoted e, spans from position k to l of the reference; iii) (f, e¯) is not an entry of the phrase table. The resulting set of indicator variables thus contains, at least, all the variables used in the exact match strategy. In addition, we license here the use of phrases containing words that do not occur in the reference. In fact, using such solutions can yield higher BLEU scores when the reward for additional correct matches exceeds the cost incurred by wrong predictions. These cases are symptoms of situations where the extraction heuristic failed to extract potentially useful subphrases. 2.5 Oracle Decoding with Reordering Constraints The ILP problem introduced in the previous section can be extended in several ways to describe and test various improvements to phrase-based systems or to evaluate the impact of different parameter settings. This flexibility mainly stems from the possibility offered by our framework to express arbitrary constraints over variables. In this section, we illustrate these possibilities by describing how reordering constraints can easily be considered. As a first example, the Moses decoder uses a distortion limit to constrain the set of possible reorderings. This constraint “enforces (...) that the last word of a phrase chosen for translation cannot be more than d9 words from the leftmost untranslated word in the source” (Lopez, 2009) and is expressed as: ∀aijkl , ai0j0k0l0 s.t. k > k0, aijkl · ai0j0k0l0 · |j − i0 + 1| ≤ d, The maximum distortion limit strategy (Lopez, 2009) is also easily expressed and take the following form (assuming this constraint is parameterized by d): ∀l < m − 1, ai,j,k,l·ai0,j0,l+1,l0 · |i0 − j − 1| 71is%t e6hs.a distortion greater that Moses default distortion limit. alignment decisions enabled by the use of larger training corpora and phrase table. To evaluate the impact ofthe second heuristic, we computed the number of phrases discarded by Moses (be- cause of the default ttl limit) but used in the oracle hypotheses. In the English to French NEWSCO setting, they account for 34.11% of the total number of phrases used in the oracle hypotheses. When the oracle decoder is constrained to use the same phrase table as Moses, its BLEU-4 score drops to 42.78. This shows that filtering the phrase table prior to decoding discards many useful phrase pairs and is seriously limiting the best achievable performance, a conclusion shared with (Auli et al., 2009). Search Errors Search errors can be identified by comparing the score of the best hypothesis found by Moses and the score of the oracle hypothesis. If the score of the oracle hypothesis is higher, then there has been a search error; on the contrary, there has been an estimation error when the score of the oracle hypothesis is lower than the score of the best hypothesis found by Moses. 940 Based on the comparison of the score of Moses hypotheses and of oracle hypotheses for the English to French NEWSCO setting, our preliminary conclusion is that the number of search errors is quite limited: only about 5% of the hypotheses of our oracle decoder are actually getting a better score than Moses solutions. Again, this shows that the scoring function (model error) is one of the main bottleneck of current PBTS. Comparing these hypotheses is nonetheless quite revealing: while Moses mostly selects phrase pairs with high translation scores and generates monotonous alignments, our ILP decoder uses larger reorderings and less probable phrases to achieve better solutions: on average, the reordering score of oracle solutions is −5.74, compared to −76.78 fscoro rMeo osfe osr outputs. iGonivsen is −the5 weight assigned through MERT training to the distortion score, no wonder that these hypotheses are severely penalized. The Impact of Phrase Length The observed outputs do not only depend on decisions made during the search, but also on decisions made during training. One such decision is the specification of maximal length for the source and target phrases. In our framework, evaluating the impact of this decision is simple: it suffices to change the definition of indicator variables so as to consider only alignments between phrases of a given length. In the English-French NEWSCO setting, the most restrictive choice, when only alignments between single words are authorized, yields an oracle BLEU-4 of 48.68; however, authorizing phrases up to length 2 allows to achieve an oracle value of 66.57, very close to the score achieved when considering all extracted phrases (67.77). This is corroborated with a further analysis of our oracle alignments, which use phrases whose average source length is 1.21 words (respectively 1.31 for target words). If many studies have already acknowledged the predomi- nance of “small” phrases in actual translations, our oracle scores suggest that, for this language pair, increasing the phrase length limit beyond 2 or 3 might be a waste of computational resources. 4 Related Work To the best of our knowledge, there are only a few works that try to study the expressive power ofphrase-based machine translation systems or to provide tools for analyzing potential causes of failure. The approach described in (Auli et al., 2009) is very similar to ours: in this study, the authors propose to find and analyze the limits of machine translation systems by studying the reference reachability. A reference is reachable for a given system if it can be exactly generated by this system. Reference reachability is assessed using Moses in forced decoding mode: during search, all hypotheses that deviate from the reference are simply discarded. Even though the main goal of this study was to compare the search space of phrase-based and hierarchical systems, it also provides some insights on the impact of various search parameters in Moses, delivering conclusions that are consistent with our main results. As described in Section 1.2, these authors also propose a typology of the errors of a statistical translation systems, but do not attempt to provide methods for identifying them. The authors of (Turchi et al., 2008) study the learn- ing capabilities of Moses by extensively analyzing learning curves representing the translation performances as a function of the number of examples, and by corrupting the model parameters. Even though their focus is more on assessing the scoring function, they reach conclusions similar to ours: the current bottleneck of translation performances is not the representation power of the PBTS but rather in their scoring functions. Oracle decoding is useful to compute reachable pseudo-references in the context of discriminative training. This is the main motivation of (Tillmann and Zhang, 2006), where the authors compute high BLEU hypotheses by running a conventional decoder so as to maximize a per-sentence approximation of BLEU-4, under a simple (local) reordering model. Oracle decoding has also been used to assess the limitations induced by various reordering constraints in (Dreyer et al., 2007). To this end, the authors propose to use a beam-search based oracle decoder, which computes lower bounds of the best achievable BLEU-4 using dynamic programming techniques over finite-state (for so-called local and IBM constraints) or hierarchically structured (for ITG constraints) sets of hypotheses. Even 941 though the numbers reported in this study are not directly comparable with ours17, it seems that our decoder is not only conceptually much simpler, but also achieves much more optimistic lower-bounds of the oracle BLEU score. The approach described in (Li and Khudanpur, 2009) employs a similar technique, which is to guide a heuristic search in an hypergraph representing possible translation hypotheses with n-gram counts matches, which amounts to decoding with a n-gram model trained on the sole reference translation. Additional tricks are presented in this article to speed-up decoding. Computing oracle BLEU scores is also the subject of (Zens and Ney, 2005; Leusch et al., 2008), yet with a different emphasis. These studies are concerned with finding the best hypotheses in a word graph or in a consensus network, a problem that has various implications for multi-pass decoding and/or system combination techniques. The former reference describes an exponential approximate algorithm, while the latter proves the NPcompleteness of this problem and discuss various heuristic approaches. Our problem is somewhat more complex and using their techniques would require us to built word graphs containing all the translations induced by arbitrary segmentations and permutations of the source sentence. 5 Conclusions In this paper, we have presented a methodology for analyzing the errors of PBTS, based on the computation of an approximation of the BLEU-4 oracle score. We have shown that this approximation could be computed fairly accurately and efficiently using Integer Linear Programming techniques. Our main result is a confirmation of the fact that extant PBTS systems are expressive enough to achieve very high translation performance with respect to conventional quality measurements. The main efforts should therefore strive to improve on the way phrases and hypotheses are scored during training. This gives further support to attempts aimed at designing context-dependent scoring functions as in (Stroppa et al., 2007; Gimpel and Smith, 2008), or at attempts to perform discriminative training of feature-rich models. (Bangalore et al., 2007). We have shown that the examination of difficult-totranslate sentences was an effective way to detect errors or inconsistencies in the reference translations, making our approach a potential aid for controlling the quality or assessing the difficulty of test data. Our experiments have also highlighted the impact of various parameters. Various extensions of the baseline ILP program have been suggested and/or evaluated. In particular, the ILP formalism lends itself well to expressing various constraints that are typically used in conventional PBTS. In 17The best BLEU-4 oracle they achieve on Europarl German to English is approximately 48; but they considered a smaller version of the training corpus and the WMT’06 test set. our future work, we aim at using this ILP framework to systematically assess various search configurations. We plan to explore how replacing non-reachable references with high-score pseudo-references can improve discrim- inative training of PBTS. We are also concerned by determining how tight is our approximation of the BLEU4 score is: to this end, we intend to compute the best BLEU-4 score within the n-best solutions of the oracle decoding problem. Acknowledgments Warm thanks to Houda Bouamor for helping us with the annotation tool. This work has been partly financed by OSEO, the French State Agency for Innovation, under the Quaero program. References Tobias Achterberg. 2007. Constraint Integer Programming. Ph.D. thesis, Technische Universit a¨t Berlin. http : / / opus .kobv .de /tuberl in/vol ltexte / 2 0 0 7 / 16 11/ . Abhishek Arun and Philipp Koehn. 2007. Online learning methods for discriminative training of phrase based statistical machine translation. In Proc. of MT Summit XI, Copenhagen, Denmark. Michael Auli, Adam Lopez, Hieu Hoang, and Philipp Koehn. 2009. A systematic analysis of translation model search spaces. In Proc. of WMT, pages 224–232, Athens, Greece. Satanjeev Banerjee and Alon Lavie. 2005. METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. In Proc. of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, pages 65–72, Ann Arbor, Michigan. Srinivas Bangalore, Patrick Haffner, and Stephan Kanthak. 2007. Statistical machine translation through global lexical selection and sentence reconstruction. In Proc. of ACL, pages 152–159, Prague, Czech Republic. L e´on Bottou and Olivier Bousquet. 2008. The tradeoffs oflarge scale learning. In Proc. of NIPS, pages 161–168, Vancouver, B.C., Canada. Chris Callison-Burch, Philipp Koehn, Christof Monz, and Josh Schroeder. 2009. Findings of the 2009 Workshop on Statistical Machine Translation. In Proc. of WMT, pages 1–28, Athens, Greece. David Chiang, Steve DeNeefe, Yee Seng Chan, and Hwee Tou Ng. 2008. Decomposability of translation metrics for improved evaluation and efficient algorithms. In Proc. of ECML, pages 610–619, Honolulu, Hawaii. John De Nero and Dan Klein. 2008. The complexity of phrase alignment problems. In Proc. of ACL: HLT, Short Papers, pages 25–28, Columbus, Ohio. Markus Dreyer, Keith B. Hall, and Sanjeev P. Khudanpur. 2007. Comparing reordering constraints for smt using efficient bleu oracle computation. In NAACL-HLT/AMTA Workshop on Syntax and Structure in Statistical Translation, pages 103– 110, Rochester, New York. 942 Ulrich Germann, Michael Jahr, Kevin Knight, Daniel Marcu, and Kenji Yamada. 2001 . Fast decoding and optimal decoding for machine translation. In Proc. of ACL, pages 228–235, Toulouse, France. 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 Proc. of NAACL, pages 1–8, Edmonton, Canada. Kevin Gimpel and Noah A. Smith. 2008. Rich source-side context for statistical machine translation. In Proc. of WMT, pages 9–17, Columbus, Ohio. Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proc. of NAACL, pages 48–54, Edmonton, Canada. Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris CallisonBurch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proc. of ACL, demonstration session. Philipp Koehn. 2004. Pharaoh: A beam search decoder for phrase-based statistical machine translation models. In Proc. of AMTA, pages 115–124, Washington DC. Shankar Kumar and William Byrne. 2005. Local phrase reordering models for statistical machine translation. In Proc. of HLT, pages 161–168, Vancouver, Canada. Alon Lavie, Kenji Sagae, and Shyamsundar Jayaraman. The significance of recall in automatic metrics for MT evaluation. In In Proc. of AMTA, pages 134–143, Washington DC. Gregor Leusch, Evgeny Matusov, and Hermann Ney. 2008. Complexity of finding the BLEU-optimal hypothesis in a confusion network. In Proc. of EMNLP, pages 839–847, Honolulu, Hawaii. Zhifei Li and Sanjeev Khudanpur. 2009. Efficient extraction of oracle-best translations from hypergraphs. In Proc. of NAACL, pages 9–12, Boulder, Colorado. Percy Liang, Alexandre Bouchard-C oˆt´ e, Dan Klein, and Ben Taskar. 2006. An end-to-end discriminative approach to machine translation. In Proc. of ACL, pages 761–768, Sydney, Australia. Adam Lopez. 2009. Translation as weighted deduction. In Proc. of EACL, pages 532–540, Athens, Greece. Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Comput. Linguist. , 29(1): 19–5 1. Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proc. of ACL, pages 160–167, Sapporo, Japan. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-jing Zhu. 2002. Bleu: A method for automatic evaluation of machine translation. Technical report, Philadelphia, Pennsylvania. D. Roth and W. Yih. 2005. Integer linear programming inference for conditional random fields. In Proc. of ICML, pages 737–744, Bonn, Germany. Nicolas Stroppa, Antal van den Bosch, and Andy Way. 2007. Exploiting source similarity for smt using context-informed features. In Andy Way and Barbara Proc. of TMI, pages Christoph Tillmann 231–240, Sk¨ ovde, and Tong Zhang. Gawronska, editors, Sweden. 2006. A discriminative global training algorithm for statistical mt. In Proc. of ACL, 721–728, Sydney, Australia. Turchi, Tijl De Bie, and Nello pages Marco Cristianini. 2008. Learn- ing performance of a machine translation system: a statistical and computational analysis. In Proc. of WMT, pages Columbus, Ohio. 35–43, Richard Zens and Hermann Ney. 2005. Word graphs for statistical machine translation. In Proc. of the ACL Workshop on Building and Using Parallel Texts, pages 191–198, Ann Arbor, Michigan. 943
6 0.38536471 31 emnlp-2010-Constraints Based Taxonomic Relation Classification
7 0.38021117 37 emnlp-2010-Domain Adaptation of Rule-Based Annotators for Named-Entity Recognition Tasks
8 0.37800458 115 emnlp-2010-Uptraining for Accurate Deterministic Question Parsing
9 0.36816546 72 emnlp-2010-Learning First-Order Horn Clauses from Web Text
10 0.36672539 123 emnlp-2010-Word-Based Dialect Identification with Georeferenced Rules
11 0.36641532 53 emnlp-2010-Fusing Eye Gaze with Speech Recognition Hypotheses to Resolve Exophoric References in Situated Dialogue
12 0.36314633 55 emnlp-2010-Handling Noisy Queries in Cross Language FAQ Retrieval
13 0.35994416 12 emnlp-2010-A Semi-Supervised Method to Learn and Construct Taxonomies Using the Web
14 0.35879487 67 emnlp-2010-It Depends on the Translation: Unsupervised Dependency Parsing via Word Alignment
15 0.34867746 57 emnlp-2010-Hierarchical Phrase-Based Translation Grammars Extracted from Alignment Posterior Probabilities
16 0.34767684 65 emnlp-2010-Inducing Probabilistic CCG Grammars from Logical Form with Higher-Order Unification
17 0.34752101 105 emnlp-2010-Title Generation with Quasi-Synchronous Grammar
18 0.34672898 35 emnlp-2010-Discriminative Sample Selection for Statistical Machine Translation
19 0.34438032 29 emnlp-2010-Combining Unsupervised and Supervised Alignments for MT: An Empirical Study
20 0.34384409 63 emnlp-2010-Improving Translation via Targeted Paraphrasing