{ "paper_id": "I17-1016", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T07:39:30.945899Z" }, "title": "Improving Neural Machine Translation through Phrase-based Forced Decoding", "authors": [ { "first": "Jingyi", "middle": [], "last": "Zhang", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Institute of Information and Communications Technology", "location": { "country": "Japan" } }, "email": "jingyizhang@nict.go.jp" }, { "first": "Masao", "middle": [], "last": "Utiyama", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Institute of Information and Communications Technology", "location": { "country": "Japan" } }, "email": "mutiyama@nict.go.jp" }, { "first": "Eiichro", "middle": [], "last": "Sumita", "suffix": "", "affiliation": { "laboratory": "", "institution": "National Institute of Information and Communications Technology", "location": { "country": "Japan" } }, "email": "eiichiro.sumita@nict.go.jp" }, { "first": "Graham", "middle": [], "last": "Neubig", "suffix": "", "affiliation": { "laboratory": "", "institution": "Nara Institute of Science and Technology", "location": { "country": "Japan" } }, "email": "gneubig@cs.cmu.edu" }, { "first": "Satoshi", "middle": [], "last": "Nakamura", "suffix": "", "affiliation": { "laboratory": "", "institution": "Nara Institute of Science and Technology", "location": { "country": "Japan" } }, "email": "s-nakamura@is.naist.jp" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "Compared to traditional statistical machine translation (SMT), neural machine translation (NMT) often sacrifices adequacy for the sake of fluency. We propose a method to combine the advantages of traditional SMT and NMT by exploiting an existing phrase-based SMT model to compute the phrase-based decoding cost for an NMT output and then using this cost to rerank the n-best NMT outputs. The main challenge in implementing this approach is that NMT outputs may not be in the search space of the standard phrase-based decoding algorithm, because the search space of phrase-based SMT is limited by the phrase-based translation rule table. We propose a soft forced decoding algorithm, which can always successfully find a decoding path for any NMT output. We show that using the forced decoding cost to rerank the NMT outputs can successfully improve translation quality on four different language pairs.", "pdf_parse": { "paper_id": "I17-1016", "_pdf_hash": "", "abstract": [ { "text": "Compared to traditional statistical machine translation (SMT), neural machine translation (NMT) often sacrifices adequacy for the sake of fluency. We propose a method to combine the advantages of traditional SMT and NMT by exploiting an existing phrase-based SMT model to compute the phrase-based decoding cost for an NMT output and then using this cost to rerank the n-best NMT outputs. The main challenge in implementing this approach is that NMT outputs may not be in the search space of the standard phrase-based decoding algorithm, because the search space of phrase-based SMT is limited by the phrase-based translation rule table. We propose a soft forced decoding algorithm, which can always successfully find a decoding path for any NMT output. We show that using the forced decoding cost to rerank the NMT outputs can successfully improve translation quality on four different language pairs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Neural machine translation (NMT), which uses a single large neural network to model the entire translation process, has recently been shown to outperform traditional statistical machine translation (SMT) such as phrase-based machine translation (PBMT) on several translation tasks (Koehn et al., 2003; Bahdanau et al., 2015; Sennrich et al., 2016a) . Compared to traditional SMT, NMT generally produces more fluent translations, but often sacrifices adequacy, such as translating source words into completely unrelated target words, over-translation or under-translation (Koehn and Knowles, 2017) .", "cite_spans": [ { "start": 281, "end": 301, "text": "(Koehn et al., 2003;", "ref_id": "BIBREF10" }, { "start": 302, "end": 324, "text": "Bahdanau et al., 2015;", "ref_id": "BIBREF2" }, { "start": 325, "end": 348, "text": "Sennrich et al., 2016a)", "ref_id": "BIBREF16" }, { "start": 571, "end": 596, "text": "(Koehn and Knowles, 2017)", "ref_id": "BIBREF9" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "There are a number of methods that combine the two paradigms to address their respective weaknesses. For example, it is possible to incorporate neural features into traditional SMT models to disambiguate hypotheses (Neubig et al., 2015; Stahlberg et al., 2016) . However, the search space of traditional SMT is usually limited by translation rule tables, reducing the ability of these models to generate hypotheses on the same level of fluency as NMT, even after reranking. There are also methods that incorporate knowledge from traditional SMT into NMT, such as lexical translation probabilities (Arthur et al., 2016; , phrase memory (Tang et al., 2016; Zhang et al., 2017) , and n-gram posterior probabilities based on traditional SMT translation lattices (Stahlberg et al., 2017) . These improve the adequacy of NMT outputs, but do not impose hard alignment constraints like traditional SMT systems and therefore cannot effectively solve all over-translation or under-translation problems.", "cite_spans": [ { "start": 215, "end": 236, "text": "(Neubig et al., 2015;", "ref_id": "BIBREF14" }, { "start": 237, "end": 260, "text": "Stahlberg et al., 2016)", "ref_id": "BIBREF20" }, { "start": 597, "end": 618, "text": "(Arthur et al., 2016;", "ref_id": "BIBREF1" }, { "start": 635, "end": 654, "text": "(Tang et al., 2016;", "ref_id": "BIBREF21" }, { "start": 655, "end": 674, "text": "Zhang et al., 2017)", "ref_id": "BIBREF27" }, { "start": 758, "end": 782, "text": "(Stahlberg et al., 2017)", "ref_id": "BIBREF19" } ], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "In this paper, we propose a method that exploits an existing phrase-based translation model to compute the phrase-based decoding cost for a given NMT translation. 1 That is, we force a phrase-based translation system to take in the source sentence and generate an NMT translation. Then we use the cost of this phrase-based forced decoding to rerank the NMT outputs. The phrasebased decoding cost will heavily punish completely unrelated translations, over-translations, and under-translations, as they will not be able to be found in the translation phrase table.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "One challenge in implementing this method is that the NMT output may not be in the search space of the phrase-based translation model, which is limited by the phrase-based translation rule table. To solve this problem, we propose a soft forced decoding algorithm, which is based on the standard phrase-based decoding algorithm and integrates new types of translation rules (deleting a source word or inserting a target word). The proposed forced decoding algorithm can always successfully find a decoding path and compute a phrase-based decoding cost for any NMT output. Another challenge is that we need a diverse NMT n-best list for reranking. Because beam search for NMT often lacks diversity in the beam -candidates only have slight differences, with most of the words overlapping -we use a random sampling method to obtain a more diverse n-best list.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "We test the proposed method on English-to-Chinese, English-to-Japanese, English-to-German and English-to-French translation tasks, obtaining large improvements over a strong NMT baseline that already incorporates discrete lexicon features.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Introduction", "sec_num": "1" }, { "text": "Our baseline NMT model is similar to the attentional model of Bahdanau et al. (2015) , which includes an encoder, a decoder and an attention (alignment) model. Given a source sentence F = {f 1 , ..., f J }, the encoder learns an annotation h j = h j ; \u2190 h j for f j using a bi-directional recurrent neural network.", "cite_spans": [ { "start": 62, "end": 84, "text": "Bahdanau et al. (2015)", "ref_id": "BIBREF2" } ], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "The decoder generates the target translation from left to right. The probability of generating next word e i is, 2", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "P N M T e i |e i\u22121 1 , F = sof tmax (g (e i\u22121 , t i , s i ))", "eq_num": "(1)" } ], "section": "Attentional NMT", "sec_num": "2" }, { "text": "where t i is a decoding state for time step i, computed by,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "t i = f (t i\u22121 , e i\u22121 , s i )", "eq_num": "(2)" } ], "section": "Attentional NMT", "sec_num": "2" }, { "text": "s i is a source representation for time i, calculated as,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "s i = J j=1 \u03b1 i,j \u2022 h j", "eq_num": "(3)" } ], "section": "Attentional NMT", "sec_num": "2" }, { "text": "where \u03b1 i,j scores how well the inputs around position j and the output at position i match, computed as,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "\u03b1 i,j = exp (a (t i\u22121 , h j )) J k=1 exp (a (t i\u22121 , h k ))", "eq_num": "(4)" } ], "section": "Attentional NMT", "sec_num": "2" }, { "text": "As we can see, NMT only learns an attention (alignment) distribution for each target word over all source words and does not provides exact mutually-exclusive word or phrase level alignments. As a result, it is known that attentional NMT systems make mistakes in over-or undertranslation (Cohn et al., 2016; Mi et al., 2016) .", "cite_spans": [ { "start": 288, "end": 307, "text": "(Cohn et al., 2016;", "ref_id": "BIBREF4" }, { "start": 308, "end": 324, "text": "Mi et al., 2016)", "ref_id": "BIBREF13" } ], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "3 Phrase-based Forced Decoding for NMT", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Attentional NMT", "sec_num": "2" }, { "text": "In phrase-based SMT (Koehn et al., 2003) , a phrase-based translation rule r includes a source phrase, a target phrase and a translation score S (r). Phrase-based translation rules can be extracted from the word-aligned training set and then used to translate new sentences. Word alignments for the training set can be obtained by IBM models (Brown et al., 1993) . Phrase-based decoding uses a list of translation rules to translate source phrases in the input sentence and generate target phrases from left to right. A basic concept in phrase-based decoding is hypotheses. As shown in Figure 1 , the hypothesis H 1 consists of two rules r 1 and r 2 . The score of a hypothesis S (H) can be calculated as the product of the scores of all applied rules. 3 An existing hypothesis can be expanded into a new hypothesis by applying a new rule. As shown in Figure 1 , H 1 can be expanded into H 2 , H 3 and H 4 . H 2 cannot be further expanded, because it covers all source words, while H 3 and H 4 can (and must) be further expanded. The decoder starts with an initial empty hypothesis H 0 and selects the hypothesis with the highest score from all completed hypotheses.", "cite_spans": [ { "start": 20, "end": 40, "text": "(Koehn et al., 2003)", "ref_id": "BIBREF10" }, { "start": 342, "end": 362, "text": "(Brown et al., 1993)", "ref_id": "BIBREF3" }, { "start": 753, "end": 754, "text": "3", "ref_id": null } ], "ref_spans": [ { "start": 586, "end": 594, "text": "Figure 1", "ref_id": null }, { "start": 852, "end": 860, "text": "Figure 1", "ref_id": null } ], "eq_spans": [], "section": "Phrase-based SMT", "sec_num": "3.1" }, { "text": "During decoding, hypotheses are stored in stacks. For a source sentence with J words, the decoder builds J stacks. The hypotheses that cover j source words are stored in stack s j . The decoder expands hypotheses in s 1 , s 2 , ..., s J in turn as shown in Algorithm 1. Here, EXPAND(H) is expanding H to get new hypotheses and putting the new hypotheses into corresponding stacks. For each stack, a beam of the best n hypotheses is kept to speed up the decoding process. ", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Phrase-based SMT", "sec_num": "3.1" }, { "text": "S(H1)=S(r1)*S(r2) S(H4)=S(r1)*S(r2)*S(r5) S(H3)=S(r1)*S(r2)*S(r4) S(H2)=S(r1)*S(r2)*S(r3)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Phrase-based SMT", "sec_num": "3.1" }, { "text": "Phrase Table Figure 1: An example of phrase-based decoding. ", "cite_spans": [], "ref_spans": [ { "start": 7, "end": 21, "text": "Table Figure", "ref_id": null } ], "eq_spans": [], "section": "Phrase-based SMT", "sec_num": "3.1" }, { "text": "As stated in the introduction, our goal is not to generate new hypotheses with phrase-based SMT, but instead use the phrase-based model to calculate scores for NMT output. In order to do so, we can perform forced decoding, which is very similar to the algorithm in the previous section but discards all partial hypotheses that do not match the NMT output. However, the NMT output is not limited by the phrase-based rule table, so there may be no decoding path that completely matches the NMT output when using only the phrase-based rules.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "To remedy this problem, inspired by previous work in forced decoding for training phrase-based SMT systems (Wuebker et al., 2010 (Wuebker et al., , 2012 we propose a soft forced decoding algorithm that can always successfully find a decoding path for a source sentence F and an NMT translation E.", "cite_spans": [ { "start": 107, "end": 128, "text": "(Wuebker et al., 2010", "ref_id": "BIBREF24" }, { "start": 129, "end": 152, "text": "(Wuebker et al., , 2012", "ref_id": "BIBREF23" } ], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "First, we introduce two new types of rules R 1 and R 2 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "R 1 A source word f can be translated into a special word null. This corresponds to deleting f during translation. The score of deleting f is cal-culated as,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "s (f \u2192 null) = unalign (f ) |T | (5)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "where unalign (f ) is how many times f is unaligned in the word-aligned training set T and |T | is the number of sentence pairs in T .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "R 2 A target word e can be translated from a special word null, which corresponds to inserting e during translation. The score of inserting e is calculated as,", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "s (null \u2192 e) = unalign (e) |T |", "eq_num": "(6)" } ], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "where unalign (e) is how many times e is unaligned in T . One motivation for Equations 5 and 6 is that function words usually have high frequencies, but do not have as clear a correspondence with a word in the other language as content words. As a result, in the training set function words are more often unaligned than content words. As an example, Table 1 and Table 2 show how many times different words occur and how many times they are unaligned in the word-aligned training set of English-to-Chinese and English-to-French tasks in our experiments. As we can see, generally there are less unaligned words in the English-to-French task, however, function words are more likely to be unaligned in both tasks. Based on Equation 5 and Equation 6, the scores of deleting or inserting \"of\" and \"a\" will be higher.", "cite_spans": [], "ref_spans": [ { "start": 351, "end": 370, "text": "Table 1 and Table 2", "ref_id": "TABREF2" } ], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "In our forced decoding, we choose to model the score of each translation rule that exists in the phrase table as the product of direct and inverse phrase translation probabilities. To make sure that Words of a practice water Occur 1.3M 1.0M 2.2K 29K Unaligned 0.51M 0.41M 0.25K 3.5K Table 2 : The number of times that words occur in the English-to-French training corpus and the number of times that they are unaligned.", "cite_spans": [], "ref_spans": [ { "start": 283, "end": 290, "text": "Table 2", "ref_id": null } ], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "the scale of the scores for R 1 and R 2 match the other phrase (which are the product of two probabilities), we use the square of the score in Equation 5/6 as the rule score for R 1 /R 2 .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "Algorithm 2 shows the forced decoding algorithm that integrates the new rules. Because the translation E is given for the forced decoding algorithm, the proposed forced decoding algorithm keeps I stacks, where I is the length of E. In other words, the stack size is corresponding to the target word size during forced decoding while the stack size is corresponding to the source word size during standard phrase-based decoding. The stack s i in Algorithm 2 contains all hypotheses in which the first i target words have been generated. We expand hypotheses in s 1 , s 2 , ..., s I in turn. When expanding a hypothesis H old in s i , besides expanding it using the original rule table EXPAND(H old ), 4 we also expand H old by inserting the next target word e i+1 at the end of H old to get an additional hypothesis H new and put H new into s i+1 . For a final hypothesis in stack s I , it may not cover all source words. We update its score by translating uncovered words into null.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "Because different decoding paths can generate the same final translation, there can be different decoding paths that fit the NMT translation E. We use the score of the single decoding path with the highest decoding score as the forced decoding score for E. We rerank the n-best NMT outputs using the phrase-based forced decoding score according to Equation 7.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "log P (E|F ) = w1 \u2022 log Pn (E|F ) + w2 \u2022 log S d (E|F ) (7)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "where P n (E|F ) is the original NMT translation probability as calculated by Equation 1;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "Pn (E|F ) = I i=1 PNMT ei|e i\u22121 1 , F", "eq_num": "(8)" } ], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "S d (E|F ) is the forced decoding score, which is the score of the decoding pathD with the highest decoding score as described above;", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "S d (E|F ) = r\u2208D S (r)", "eq_num": "(9)" } ], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "w 1 and w 2 are weights that can be tuned on the n-best list of the development set. The easiest way to get an n-best list for NMT is by using the n-best translations from beam search, which is the standard decoding algorithm for NMT. While beam search is likely to find the highest-scoring hypothesis, it often lacks diversity in the beam: candidates only have slight differences, with most of the words overlapping. In order to obtain a more diverse list of hypotheses for reranking, we additionally augment the 1-best hypothesis discovered by beam search with translations sampled from the NMT conditional probability distribution.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "The standard method for sampling hypotheses in NMT is ancestral sampling, where we randomly select a word from the vocabulary according to P N M T e i |e i\u22121 1 , F (Shen et al., 2016) . This will make a diverse list of hypotheses, but may reduce the probability of selecting a highly scoring hypothesis, and the whole n-best list may not contain any candidate with better translation quality than the standard beam search output.", "cite_spans": [ { "start": 164, "end": 183, "text": "(Shen et al., 2016)", "ref_id": "BIBREF18" } ], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "Instead, we take an alternative approach that proved empirically better in our experiments: at each time step i, we use sampling to randomly select the next word from e and e according to Equation 10. Here, e and e are the two target words with the highest probability according to Equation 1.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "EQUATION", "cite_spans": [], "ref_spans": [], "eq_spans": [ { "start": 0, "end": 8, "text": "EQUATION", "ref_id": "EQREF", "raw_str": "P rdm (e ) = P N M T (e |e i\u22121 1 ,F ) P N M T (e |e i\u22121 1 ,F )+PNMT (e |e i\u22121 1 ,F ) P rdm (e ) = P N M T (e |e i\u22121 1 ,F ) P N M T (e |e i\u22121 1 ,F )+PNMT (e |e i\u22121 1 ,F )", "eq_num": "(10)" } ], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "The sampling process ends when /s is selected as the next word. We repeat the decoding process 1, 000 times to sample 1, 000 outputs for each source sentence. We additionally add the 1-best output of standard beam search, making the size of the list used for reranking to be 1, 001.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Forced Decoding for NMT", "sec_num": "3.2" }, { "text": "We evaluated the proposed approach for Englishto-Chinese (en-zh), English-to-Japanese (en-ja), English-to-German (en-de) and English-to-French (en-fr) translation tasks. For the en-zh and enja tasks, we used datasets provided for the patent machine translation task at NTCIR-9 (Goto et al., 2011) . 5 For the en-de and en-fr tasks, we used version 7 of the Europarl corpus as training data, WMT 2014 test sets as our development sets and WMT 2015 test sets as our test sets. The detailed statistics for training, development and test sets are given in Table 3 . The word segmentation was done by BaseSeg (Zhao et al., 2006) for Chinese and Mecab 6 for Japanese.", "cite_spans": [ { "start": 277, "end": 296, "text": "(Goto et al., 2011)", "ref_id": "BIBREF5" }, { "start": 299, "end": 300, "text": "5", "ref_id": null }, { "start": 604, "end": 623, "text": "(Zhao et al., 2006)", "ref_id": "BIBREF28" } ], "ref_spans": [ { "start": 552, "end": 559, "text": "Table 3", "ref_id": "TABREF5" } ], "eq_spans": [], "section": "Settings", "sec_num": "5.1" }, { "text": "We built attentional NMT systems with Lamtram 7 . Word embedding size and hidden layer size are both 512. We used Byte-pair encoding (BPE) (Sennrich et al., 2016b) and set the vocabulary size to be 50K. We used the Adam algorithm for optimization.", "cite_spans": [ { "start": 139, "end": 163, "text": "(Sennrich et al., 2016b)", "ref_id": "BIBREF17" } ], "ref_spans": [], "eq_spans": [], "section": "Settings", "sec_num": "5.1" }, { "text": "To obtain a phrase-based translation rule table for our forced decoding algorithm, we used GIZA++ (Och and Ney, 2003) and grow-diagfinal-and heuristic to obtain symmetric word alignments for the training set. Then we extracted the rule table using Moses (Koehn et al., 2007) . Table 4 shows results of the phrase-based SMT system 8 , the baseline NMT system, the lexicon integration method (Arthur et al., 2016) and the proposed reranking method. We tested three features for reranking: the NMT score P n , the forced decoding score S d and a word penalty (WP) feature, which is the length of the translation. The best NMT system and the systems that have no significant difference from the best NMT system at the p < 0.05 level using bootstrap resampling (Koehn, 2004) are shown in bold font.", "cite_spans": [ { "start": 98, "end": 117, "text": "(Och and Ney, 2003)", "ref_id": "BIBREF15" }, { "start": 254, "end": 274, "text": "(Koehn et al., 2007)", "ref_id": "BIBREF8" }, { "start": 390, "end": 411, "text": "(Arthur et al., 2016)", "ref_id": "BIBREF1" }, { "start": 756, "end": 769, "text": "(Koehn, 2004)", "ref_id": "BIBREF7" } ], "ref_spans": [ { "start": 277, "end": 284, "text": "Table 4", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Settings", "sec_num": "5.1" }, { "text": "As we can see, integrating lexical translation probabilities improved the baseline NMT system 1.000 1.024 1.001 1.001 1.011 1.007 0.999 0.989 NMT+lex+rerank(Pn+S d +WP) 0.990 1.014 1.000 0.986 1.000 0.989 1.000 0.992 Table 6 : Ratio of translation length to reference length for different system outputs in Table 4. and reranking with the three features all together achieved further improvements for all four language pairs. Even on English-to-Chinese and English-to-Japanese tasks, where the NMT system outperformed the phrase-based SMT system by 7-8 BLEU scores, using the forced decoding score for reranking NMT outputs can still achieve significant improvements. With or without the word penalty feature, using both P n and S d for reranking gave better results than only using P n or S d alone. We also show METEOR and chrF scores on the test sets in Table 5 . Our reranking method improved both METEOR and chrF significantly.", "cite_spans": [], "ref_spans": [ { "start": 217, "end": 224, "text": "Table 6", "ref_id": null }, { "start": 307, "end": 315, "text": "Table 4.", "ref_id": "TABREF7" }, { "start": 857, "end": 864, "text": "Table 5", "ref_id": "TABREF8" } ], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5.2" }, { "text": "The Word Penalty Feature The word penalty feature generally improved the reranking results, especially when only the NMT score P n was used for reranking. As we can see, using only P n for reranking decreased the translation quality com-pared to the standard beam search result of NMT. Because the search spaces of beam search and random sampling are quite different, the best beam search output does not necessarily have the highest NMT score compared to random sampling outputs. Therefore, even the P n reranking results do have higher NMT scores, but have lower BLEU scores according to Table 4 . To explain why this happened, we show the ratio of translation length to reference length in Table 6 . As we can see, the P n reranking outputs are much shorter. This is because NMT generally prefers shorter translations, since Equation 8 multiplies all target word probabilities together. So the word penalty feature can improve the P n reranking results considerably, by preferring longer sentences. Because the forced decoding score S d as shown in Equation 9 does not obviously prefer shorter or longer sentences, when S d was used for reranking, the word penalty Source for hypophysectomized (hypop hy sec to mized) rats , the drinking water additionally contains 5 % glucose .", "cite_spans": [], "ref_spans": [ { "start": 590, "end": 597, "text": "Table 4", "ref_id": "TABREF7" }, { "start": 693, "end": 700, "text": "Table 6", "ref_id": null } ], "eq_spans": [], "section": "Results and Analysis", "sec_num": "5.2" }, { "text": "\u5bf9\u4e8e(for) \u53bb(remove) \u5782\u4f53(hypophysis) \u5927(big) \u9f20(rat) \uff0c \u996e\u7528\u6c34(drinking water) \u4e2d(in) \u53e6 \u5916(also) \u542b\u6709(contain) 5 \uff05 \u8461\u8404\u7cd6(glucose) \u3002 PBMT \u7528\u4e8e(for) \u5927(big) \u9f20(rat) \u5782\u4f53(hypophysis) HySecto\uff0c(Hy Sec to \uff0c) \u996e\u7528\u6c34(drinking water) \u53e6\u5916(also) \u542b\u6709(contain) 5 \uff05 \u8461\u8404\u7cd6(glucose) \u3002 NMT \u5bf9\u4e8e(for) \u8fc7(pass) \u76f2\u80a0(cecum) \u7684(of) \u5927(big) \u9f20(rat) \uff0c \u996e\u7528\u6c34(drinking water) \u53e6\u5916(also) \u542b\u6709(contain) 5 \uff05 \u8461\u8404\u7cd6(glucose) \u3002 NMT+lex \u5bf9\u4e8e(for) \u4f4e(low) \u916a(cheese) \u86cb\u767d(protein) \u5207\u9664(remove) \u7684(of) \u5927(big) \u9f20(rat) \uff0c \u996e\u7528 \u6c34(drinking water) \u53e6\u5916(also) \u542b\u6709(contain) 5 \uff05 \u8461\u8404\u7cd6(glucose) \u3002 NMT+lex+Pn NMT+lex+Pn+WP NMT+lex+S d \u5bf9\u4e8e(for) \u5782\u4f53(hypophysis) \u5728(is) \u5207\u9664(remove) \u5927(big) \u9f20(rat) \u4e2d(in) \uff0c \u996e\u7528\u6c34(drinking water) \u53e6\u5916(also) \u542b\u6709(contain) 5 \uff05 \u8461\u8404\u7cd6(glucose) \u3002", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "NMT+lex+S d +WP NMT+lex+Pn+S d", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "\u5bf9\u4e8e(for) \u5782\u4f53(hypophysis) \u5728(is) \u5207\u9664(remove) \u7684(of) \u5927(big) \u9f20(rat) \u4e2d(in) \uff0c \u996e\u7528 \u6c34(drinking water) \u53e6\u5916(also) \u542b\u6709(contain) 5 \uff05 \u8461\u8404\u7cd6(glucose) \u3002 NMT+lex+Pn+S d +WP Table 7 : An example of improving inaccurate rare word translation by using S d for reranking. feature became less helpful. When both P n and S d were used for reranking, the word penalty feature only achieved further significant improvement on the English-to-Japanese task. Table 9 : Forced decoding paths for T 1 and T 2 : used rules and log scores. The translation rules with shade are used only for T 1 or T 2 . Table 7 gives translation examples of our reranking method from the English-to-Chinese task. The source English word \"hypophysectomized\" is an unknown word which does not occur in the training set. By employing BPE, this word is split into \"hypop\", \"hy\", \"sec\", \"to\" and \"mized\". The correct translation for \"hypophysectomized\" is \"\u53bb(remove) \u5782 \u4f53(hypophysis)\" as shown in the reference sentence. The original attentional NMT translated it into incorrect translation \"\u8fc7(pass) \u76f2 \u80a0(cecum)\". After integrating lexicons, the NMT system translated it into \"\u4f4e(low) \u916a(cheese) \u86cb\u767d(protein) \u5207 \u9664(remove)\". The last word \"\u5207 \u9664(remove)\" is correct, but the rest of the translation is still wrong. Only by using the forced decoding score S d for reranking, we get the more accurate translation \"\u5782\u4f53(hypophysis) \u5728(is) \u5207\u9664(remove)\".", "cite_spans": [], "ref_spans": [ { "start": 148, "end": 155, "text": "Table 7", "ref_id": null }, { "start": 423, "end": 430, "text": "Table 9", "ref_id": null }, { "start": 564, "end": 571, "text": "Table 7", "ref_id": null } ], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "To further demonstrate how the reranking method works, Table 9 shows translation rules and their log-scores contained in the forced decoding paths found for T 1 , the NMT translation without reranking and T 2 , the NMT translation using both P n and S d for reranking. As we can see, the four rules r a , r b , r c and r d used for T 1 have low scores. r a is an unlikely translation. In r b , r c and r d , \"\u916a(cheese)\", \"\u86cb\u767d(protein)\" and \"hypop\" are content words, which are unlikely to be deleted or inserted during translation. Table 9 also shows that the translation of function words is very flexible. The score of inserting a function word \"\u7684(of)\" is very high. The translation rule \"the \u2192\u5728(is)\" used for T 2 is incorrect, but its score is relatively high, because function words are often Source such changes in reaction conditions include , but are not limited to , an increase in temperature or change in ph . Reference \u6240(such) incorrectly aligned in the training set. The reason why function words are more likely to be incorrectly aligned to each other is that they usually have high frequencies and do not have clear correspondences between different languages.", "cite_spans": [], "ref_spans": [ { "start": 55, "end": 62, "text": "Table 9", "ref_id": null }, { "start": 531, "end": 538, "text": "Table 9", "ref_id": null } ], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "\u8ff0(said) \u53cd \u5e94(reaction) \u6761 \u4ef6(condition) \u7684(of) \u6539 \u53d8(change) \u5305 \u62ec(include) \u4f46(but) \u4e0d(not) \u9650 \u4e8e(limit) \u6e29\u5ea6(temperature) \u7684(of) \u589e\u52a0(increase) \u6216(or) pH \u503c(value) \u7684(of) \u6539\u53d8(change) \u3002 PBMT \u4e2d(in) \u7684(of) \u8fd9 \u79cd(such) \u53d8 \u5316(change) \u7684(of) \u53cd \u5e94(reaction) \u6761 \u4ef6(condition) \u5305 \u62ec(include) \uff0c \u4f46(but) \u4e0d(not) \u9650 \u4e8e(limit) \uff0c \u589e\u52a0(", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "In T 1 , \"hypophysectomized (hypop hy sec to mized)\" is incorrectly translated into \"\u4f4e(low) \u916a(cheese) \u86cb\u767d(protein) \u5207\u9664(remove)\". However, from Table 9, we can see that the forced decoding algorithm learns it as unlikely translation (hy\u2192\u4f4e(low)), over-translation (null\u2192\u916a(cheese), null\u2192\u86cb \u767d(protein)) and under-translation (hypop\u2192null, sec\u2192null), because there is no translation rule between \"hypop\" \"sec\" and \"\u916a(cheese)\" \"\u86cb\u767d(protein)\". Because content words are unlikely to be deleted or inserted during translation, they have low forced decoding scores. So using the forced decoding score for reranking NMT outputs can naturally improve over-translation or under-translation as shown in Table 8 . As we can see, without using S d for reranking, NMT under-translated \"temperature\" and over-translated \"ph\" twice, which will be assigned low scores by forced decoding. By using S d for reranking, the correct translation was selected.", "cite_spans": [], "ref_spans": [ { "start": 684, "end": 691, "text": "Table 8", "ref_id": "TABREF10" } ], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "We did human evaluation on 100 sentences randomly selected from the English-to-Chinese test set to test the effectiveness of our forced decoding method. We compared the outputs of two systems:", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "\u2022 NMT+lex+rerank(P n +WP) \u2022 NMT+lex+rerank(P n +S d +WP)", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "For each source sentence, we compared the two system outputs. Table 10 shows the numbers of sentences that our forced decoding feature helped to reduce completely unrelated translation, over-translation and under-translation. The last line of Table 10 means that for 73 source sentences, our forced decoding feature neither reduced nor caused more unrelated/over/under translation. That is our forced decoding feature never caused more unrelated/over/under translation for the sampled 100 sentences, which shows that our method is very robust for improving unrelated/over/under translation. Reranking PBMT Outputs with NMT We also did experiments that use the NMT score as an additional feature to rerank PBMT outputs (unique 1, 000-best list). The results are shown in Table 11 . We also copy results of baseline PBMT and NMT from Table 4 for direct comparison. As we can see, using NMT to rerank PBMT outputs achieved improvements over the baseline PBMT system. However, when the baseline NMT system is significantly better than the baseline PBMT system (en-zh, en-ja), even using NMT to rerank PBMT outputs still achieved lower translation quality compared to the baseline NMT system. 6 Related Work Wuebker et al. (2010 Wuebker et al. ( , 2012 applied forced decoding on the training set to improve the training process of phrase-based SMT and prune the phrasebased rule table. They also used word insertions and deletions for forced decoding, but they used a high penalty for all insertions and deletions. In contrast, our soft forced decoding algorithm for NMT outputs uses a small penalty for function words and a high penalty for content words, because function words are usually translated very flexibly and more likely to be inserted or deleted compared to content words. For example, the under-translation of a content word can hurt the adequacy of the translation heavily. But function words may naturally disappear during translation (e.g. the English word \"the\" disappears in Chinese). By assigning a high penalty to words that should not be deleted or inserted during translation, our soft forced decoding method aims to improve the adequacy of NMT, which is very different from previous forced decoding methods that are used to improve general SMT training (Yu et al., 2013; Xiao et al., 2016) . A major difference of traditional SMT and NMT is that the alignment model in traditional SMT provides exact word or phrase level alignments between the source and target sentences while the attention model in NMT only computes an alignment probability distribution for each target word over all source words, which is the main reason why NMT is more likely to produce completely unrelated translations, over-translation or under-translation compared to traditional SMT. To relieve NMT of these problems, there are methods that modify the NMT neural network structure (Tu et al., 2016; Alkhouli et al., 2016) while we rerank NMT outputs by exploiting knowledge from traditional SMT.", "cite_spans": [ { "start": 1203, "end": 1223, "text": "Wuebker et al. (2010", "ref_id": "BIBREF24" }, { "start": 1224, "end": 1247, "text": "Wuebker et al. ( , 2012", "ref_id": "BIBREF23" }, { "start": 2273, "end": 2290, "text": "(Yu et al., 2013;", "ref_id": "BIBREF26" }, { "start": 2291, "end": 2309, "text": "Xiao et al., 2016)", "ref_id": "BIBREF25" }, { "start": 2879, "end": 2896, "text": "(Tu et al., 2016;", "ref_id": "BIBREF22" }, { "start": 2897, "end": 2919, "text": "Alkhouli et al., 2016)", "ref_id": "BIBREF0" } ], "ref_spans": [ { "start": 62, "end": 70, "text": "Table 10", "ref_id": "TABREF2" }, { "start": 243, "end": 251, "text": "Table 10", "ref_id": "TABREF2" }, { "start": 770, "end": 778, "text": "Table 11", "ref_id": "TABREF2" }, { "start": 832, "end": 839, "text": "Table 4", "ref_id": "TABREF7" } ], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "There are also existing methods that rerank NMT outputs by using target-bidirectional NMT models Sennrich et al., 2016a) . Their reranking method aims to overcome the issue of unbalanced accuracy in NMT outputs while our reranking method aims to solve the inadequacy problem of NMT.", "cite_spans": [ { "start": 97, "end": 120, "text": "Sennrich et al., 2016a)", "ref_id": "BIBREF16" } ], "ref_spans": [], "eq_spans": [], "section": "Reference", "sec_num": null }, { "text": "In this paper, we propose to exploit an existing phrase-based SMT model to compute the phrasebased decoding cost for NMT outputs and then use the phrase-based decoding cost to rerank the nbest NMT outputs, so we can combine the advantages of both PBMT and NMT. Because an NMT output may not be in the search space of standard phrase-based SMT, we propose a forced decoding algorithm, which can always successfully find a decoding path for any NMT output by deleting source words and inserting target words. Results show that using the forced decoding cost to rerank NMT outputs improved translation accuracy on four different language pairs.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Conclusion", "sec_num": "7" }, { "text": "In fact, our method can take in the output of any upstream system, but we experiment exclusively with using it to rerank NMT output.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "g, f and a in Equation 1, 2 and 4 are nonlinear, potentially multi-layered, functions.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "In actual phrase-based decoding it is common to integrate reordering probabilities in the forced decoding score defined in Equation 9. However, because NMT generally produces more properly ordered sentences than traditional SMT, in this work we do not consider reordering probabilities in our forced decoding algorithm.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "The new introduced word inserting/deleting rules are not used when performing EXPAND(H old ).", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "Note that NTCIR-9 only contained a Chinese-to-English translation task, we used English as the source language in our experiments. In NTCIR-9, the development and test sets were both provided for the zh-en task while only the test set was provided for the en-ja task. We used the sentences from the NTCIR-8 en-ja and ja-en test sets as the development set in our experiments.6 http://sourceforge.net/projects/mecab/files/ 7 https://github.com/neubig/lamtram", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null }, { "text": "We used the default Moses settings for phrase-based SMT.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "", "sec_num": null } ], "back_matter": [], "bib_entries": { "BIBREF0": { "ref_id": "b0", "title": "Alignment-based neural machine translation", "authors": [ { "first": "Tamer", "middle": [], "last": "Alkhouli", "suffix": "" }, { "first": "Gabriel", "middle": [], "last": "Bretschner", "suffix": "" }, { "first": "Jan-Thorsten", "middle": [], "last": "Peter", "suffix": "" }, { "first": "Mohammed", "middle": [], "last": "Hethnawi", "suffix": "" }, { "first": "Andreas", "middle": [], "last": "Guta", "suffix": "" }, { "first": "Hermann", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the First Conference on Machine Translation", "volume": "", "issue": "", "pages": "54--65", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tamer Alkhouli, Gabriel Bretschner, Jan-Thorsten Pe- ter, Mohammed Hethnawi, Andreas Guta, and Her- mann Ney. 2016. Alignment-based neural machine translation. In Proceedings of the First Conference on Machine Translation, pages 54-65.", "links": null }, "BIBREF1": { "ref_id": "b1", "title": "Incorporating discrete translation lexicons into neural machine translation", "authors": [ { "first": "Philip", "middle": [], "last": "Arthur", "suffix": "" }, { "first": "Graham", "middle": [], "last": "Neubig", "suffix": "" }, { "first": "Satoshi", "middle": [], "last": "Nakamura", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1557--1567", "other_ids": {}, "num": null, "urls": [], "raw_text": "Philip Arthur, Graham Neubig, and Satoshi Nakamura. 2016. Incorporating discrete translation lexicons into neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Nat- ural Language Processing, pages 1557-1567.", "links": null }, "BIBREF2": { "ref_id": "b2", "title": "Neural machine translation by jointly learning to align and translate", "authors": [ { "first": "Dzmitry", "middle": [], "last": "Bahdanau", "suffix": "" }, { "first": "Kyunghyun", "middle": [], "last": "Cho", "suffix": "" }, { "first": "Yoshua", "middle": [], "last": "Bengio", "suffix": "" } ], "year": 2015, "venue": "International Conference on Learning Representations", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. 2015. Neural machine translation by jointly learning to align and translate. In International Con- ference on Learning Representations.", "links": null }, "BIBREF3": { "ref_id": "b3", "title": "The mathematics of statistical machine translation: Parameter estimation", "authors": [ { "first": "Vincent J Della", "middle": [], "last": "Peter F Brown", "suffix": "" }, { "first": "Stephen A Della", "middle": [], "last": "Pietra", "suffix": "" }, { "first": "Robert L", "middle": [], "last": "Pietra", "suffix": "" }, { "first": "", "middle": [], "last": "Mercer", "suffix": "" } ], "year": 1993, "venue": "Computational Linguistics", "volume": "19", "issue": "2", "pages": "263--311", "other_ids": {}, "num": null, "urls": [], "raw_text": "Peter F Brown, Vincent J Della Pietra, Stephen A Della Pietra, and Robert L Mercer. 1993. The mathemat- ics of statistical machine translation: Parameter esti- mation. Computational Linguistics, 19(2):263-311.", "links": null }, "BIBREF4": { "ref_id": "b4", "title": "Incorporating structural alignment biases into an attentional neural translation model", "authors": [ { "first": "Trevor", "middle": [], "last": "Cohn", "suffix": "" }, { "first": "Cong Duy Vu", "middle": [], "last": "Hoang", "suffix": "" }, { "first": "Ekaterina", "middle": [], "last": "Vymolova", "suffix": "" }, { "first": "Kaisheng", "middle": [], "last": "Yao", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Dyer", "suffix": "" }, { "first": "Gholamreza", "middle": [], "last": "Haffari", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "876--885", "other_ids": {}, "num": null, "urls": [], "raw_text": "Trevor Cohn, Cong Duy Vu Hoang, Ekaterina Vy- molova, Kaisheng Yao, Chris Dyer, and Gholamreza Haffari. 2016. Incorporating structural alignment biases into an attentional neural translation model. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computa- tional Linguistics: Human Language Technologies, pages 876-885.", "links": null }, "BIBREF5": { "ref_id": "b5", "title": "Overview of the patent machine translation task at the NTCIR-9 workshop", "authors": [ { "first": "Isao", "middle": [], "last": "Goto", "suffix": "" }, { "first": "Bin", "middle": [], "last": "Lu", "suffix": "" }, { "first": "Ka", "middle": [ "Po" ], "last": "Chow", "suffix": "" }, { "first": "Eiichiro", "middle": [], "last": "Sumita", "suffix": "" }, { "first": "Benjamin", "middle": [ "K" ], "last": "Tsou", "suffix": "" } ], "year": 2011, "venue": "Proc. NTCIR-9", "volume": "", "issue": "", "pages": "559--578", "other_ids": {}, "num": null, "urls": [], "raw_text": "Isao Goto, Bin Lu, Ka Po Chow, Eiichiro Sumita, and Benjamin K Tsou. 2011. Overview of the patent ma- chine translation task at the NTCIR-9 workshop. In Proc. NTCIR-9, pages 559-578.", "links": null }, "BIBREF6": { "ref_id": "b6", "title": "Improved neural machine translation with SMT features", "authors": [ { "first": "Wei", "middle": [], "last": "He", "suffix": "" }, { "first": "Zhongjun", "middle": [], "last": "He", "suffix": "" }, { "first": "Hua", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Haifeng", "middle": [], "last": "Wang", "suffix": "" } ], "year": 2016, "venue": "Thirtieth AAAI conference on artificial intelligence", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Wei He, Zhongjun He, Hua Wu, and Haifeng Wang. 2016. Improved neural machine translation with SMT features. In Thirtieth AAAI conference on ar- tificial intelligence.", "links": null }, "BIBREF7": { "ref_id": "b7", "title": "Statistical significance tests for machine translation evaluation", "authors": [ { "first": "Philipp", "middle": [], "last": "Koehn", "suffix": "" } ], "year": 2004, "venue": "Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "388--395", "other_ids": {}, "num": null, "urls": [], "raw_text": "Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Nat- ural Language Processing, pages 388-395.", "links": null }, "BIBREF8": { "ref_id": "b8", "title": "Moses: Open source toolkit for statistical machine translation", "authors": [ { "first": "Philipp", "middle": [], "last": "Koehn", "suffix": "" }, { "first": "Hieu", "middle": [], "last": "Hoang", "suffix": "" }, { "first": "Alexandra", "middle": [], "last": "Birch", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Callison-Burch", "suffix": "" }, { "first": "Marcello", "middle": [], "last": "Federico", "suffix": "" }, { "first": "Nicola", "middle": [], "last": "Bertoldi", "suffix": "" }, { "first": "Brooke", "middle": [], "last": "Cowan", "suffix": "" }, { "first": "Wade", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Christine", "middle": [], "last": "Moran", "suffix": "" }, { "first": "Richard", "middle": [], "last": "Zens", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Dyer", "suffix": "" }, { "first": "Ondrej", "middle": [], "last": "Bojar", "suffix": "" }, { "first": "Alexandra", "middle": [], "last": "Constantin", "suffix": "" }, { "first": "Evan", "middle": [], "last": "Herbst", "suffix": "" } ], "year": 2007, "venue": "Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions", "volume": "", "issue": "", "pages": "177--180", "other_ids": {}, "num": null, "urls": [], "raw_text": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondrej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open source toolkit for statistical machine translation. In Proceedings of the 45th Annual Meeting of the As- sociation for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Ses- sions, pages 177-180.", "links": null }, "BIBREF9": { "ref_id": "b9", "title": "Six challenges for neural machine translation", "authors": [ { "first": "Philipp", "middle": [], "last": "Koehn", "suffix": "" }, { "first": "Rebecca", "middle": [], "last": "Knowles", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the First Workshop on Neural Machine Translation", "volume": "", "issue": "", "pages": "28--39", "other_ids": {}, "num": null, "urls": [], "raw_text": "Philipp Koehn and Rebecca Knowles. 2017. Six chal- lenges for neural machine translation. In Pro- ceedings of the First Workshop on Neural Machine Translation, pages 28-39.", "links": null }, "BIBREF10": { "ref_id": "b10", "title": "Statistical phrase-based translation", "authors": [ { "first": "Philipp", "middle": [], "last": "Koehn", "suffix": "" }, { "first": "Franz", "middle": [ "Josef" ], "last": "Och", "suffix": "" }, { "first": "Daniel", "middle": [], "last": "Marcu", "suffix": "" } ], "year": 2003, "venue": "Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology", "volume": "1", "issue": "", "pages": "48--54", "other_ids": {}, "num": null, "urls": [], "raw_text": "Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computa- tional Linguistics on Human Language Technology- Volume 1, pages 48-54.", "links": null }, "BIBREF11": { "ref_id": "b11", "title": "Agreement on targetbidirectional neural machine translation", "authors": [ { "first": "Lemao", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Masao", "middle": [], "last": "Utiyama", "suffix": "" }, { "first": "Andrew", "middle": [], "last": "Finch", "suffix": "" }, { "first": "Eiichiro", "middle": [], "last": "Sumita", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", "volume": "", "issue": "", "pages": "411--416", "other_ids": {}, "num": null, "urls": [], "raw_text": "Lemao Liu, Masao Utiyama, Andrew Finch, and Eiichiro Sumita. 2016. Agreement on target- bidirectional neural machine translation. In Pro- ceedings of the 2016 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 411-416.", "links": null }, "BIBREF12": { "ref_id": "b12", "title": "Interactive attention for neural machine translation", "authors": [ { "first": "Fandong", "middle": [], "last": "Meng", "suffix": "" }, { "first": "Zhengdong", "middle": [], "last": "Lu", "suffix": "" }, { "first": "Hang", "middle": [], "last": "Li", "suffix": "" }, { "first": "Qun", "middle": [], "last": "Liu", "suffix": "" } ], "year": 2016, "venue": "Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers", "volume": "", "issue": "", "pages": "2174--2185", "other_ids": {}, "num": null, "urls": [], "raw_text": "Fandong Meng, Zhengdong Lu, Hang Li, and Qun Liu. 2016. Interactive attention for neural ma- chine translation. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2174-2185.", "links": null }, "BIBREF13": { "ref_id": "b13", "title": "Coverage embedding models for neural machine translation", "authors": [ { "first": "Haitao", "middle": [], "last": "Mi", "suffix": "" }, { "first": "Zhiguo", "middle": [], "last": "Baskaran Sankaran", "suffix": "" }, { "first": "Abe", "middle": [], "last": "Wang", "suffix": "" }, { "first": "", "middle": [], "last": "Ittycheriah", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "955--960", "other_ids": {}, "num": null, "urls": [], "raw_text": "Haitao Mi, Baskaran Sankaran, Zhiguo Wang, and Abe Ittycheriah. 2016. Coverage embedding models for neural machine translation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 955-960.", "links": null }, "BIBREF14": { "ref_id": "b14", "title": "Neural reranking improves subjective quality of machine translation: NAIST at WAT2015", "authors": [ { "first": "Graham", "middle": [], "last": "Neubig", "suffix": "" }, { "first": "Makoto", "middle": [], "last": "Morishita", "suffix": "" }, { "first": "Satoshi", "middle": [], "last": "Nakamura", "suffix": "" } ], "year": 2015, "venue": "Proceedings of the 2nd Workshop on Asian Translation (WAT2015)", "volume": "", "issue": "", "pages": "", "other_ids": {}, "num": null, "urls": [], "raw_text": "Graham Neubig, Makoto Morishita, and Satoshi Naka- mura. 2015. Neural reranking improves subjective quality of machine translation: NAIST at WAT2015. In Proceedings of the 2nd Workshop on Asian Trans- lation (WAT2015).", "links": null }, "BIBREF15": { "ref_id": "b15", "title": "A systematic comparison of various statistical alignment models", "authors": [ { "first": "Josef", "middle": [], "last": "Franz", "suffix": "" }, { "first": "Hermann", "middle": [], "last": "Och", "suffix": "" }, { "first": "", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2003, "venue": "Computational Linguistics", "volume": "29", "issue": "1", "pages": "19--51", "other_ids": {}, "num": null, "urls": [], "raw_text": "Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19-51.", "links": null }, "BIBREF16": { "ref_id": "b16", "title": "Edinburgh neural machine translation systems for WMT 16", "authors": [ { "first": "Rico", "middle": [], "last": "Sennrich", "suffix": "" }, { "first": "Barry", "middle": [], "last": "Haddow", "suffix": "" }, { "first": "Alexandra", "middle": [], "last": "Birch", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the First Conference on Machine Translation", "volume": "", "issue": "", "pages": "371--376", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016a. Edinburgh neural machine translation sys- tems for WMT 16. In Proceedings of the First Con- ference on Machine Translation, pages 371-376.", "links": null }, "BIBREF17": { "ref_id": "b17", "title": "Neural machine translation of rare words with subword units", "authors": [ { "first": "Rico", "middle": [], "last": "Sennrich", "suffix": "" }, { "first": "Barry", "middle": [], "last": "Haddow", "suffix": "" }, { "first": "Alexandra", "middle": [], "last": "Birch", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1715--1725", "other_ids": {}, "num": null, "urls": [], "raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016b. Neural machine translation of rare words with subword units. In Proceedings of the 54th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715- 1725.", "links": null }, "BIBREF18": { "ref_id": "b18", "title": "Minimum risk training for neural machine translation", "authors": [ { "first": "Shiqi", "middle": [], "last": "Shen", "suffix": "" }, { "first": "Yong", "middle": [], "last": "Cheng", "suffix": "" }, { "first": "Zhongjun", "middle": [], "last": "He", "suffix": "" }, { "first": "Wei", "middle": [], "last": "He", "suffix": "" }, { "first": "Hua", "middle": [], "last": "Wu", "suffix": "" }, { "first": "Maosong", "middle": [], "last": "Sun", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Liu", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1683--1692", "other_ids": {}, "num": null, "urls": [], "raw_text": "Shiqi Shen, Yong Cheng, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. 2016. Minimum risk training for neural machine translation. In Pro- ceedings of the 54th Annual Meeting of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 1683-1692.", "links": null }, "BIBREF19": { "ref_id": "b19", "title": "Neural machine translation by minimising the Bayes-risk with respect to syntactic translation lattices", "authors": [ { "first": "Felix", "middle": [], "last": "Stahlberg", "suffix": "" }, { "first": "Adri\u00e0", "middle": [], "last": "De Gispert", "suffix": "" }, { "first": "Eva", "middle": [], "last": "Hasler", "suffix": "" }, { "first": "Bill", "middle": [], "last": "Byrne", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 15th Conference of the European Chapter", "volume": "2", "issue": "", "pages": "362--368", "other_ids": {}, "num": null, "urls": [], "raw_text": "Felix Stahlberg, Adri\u00e0 de Gispert, Eva Hasler, and Bill Byrne. 2017. Neural machine translation by minimising the Bayes-risk with respect to syntactic translation lattices. In Proceedings of the 15th Con- ference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Pa- pers, pages 362-368.", "links": null }, "BIBREF20": { "ref_id": "b20", "title": "Syntactically guided neural machine translation", "authors": [ { "first": "Felix", "middle": [], "last": "Stahlberg", "suffix": "" }, { "first": "Eva", "middle": [], "last": "Hasler", "suffix": "" }, { "first": "Aurelien", "middle": [], "last": "Waite", "suffix": "" }, { "first": "Bill", "middle": [], "last": "Byrne", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics", "volume": "2", "issue": "", "pages": "299--305", "other_ids": {}, "num": null, "urls": [], "raw_text": "Felix Stahlberg, Eva Hasler, Aurelien Waite, and Bill Byrne. 2016. Syntactically guided neural machine translation. In Proceedings of the 54th Annual Meet- ing of the Association for Computational Linguistics (Volume 2: Short Papers), pages 299-305.", "links": null }, "BIBREF21": { "ref_id": "b21", "title": "Neural machine translation with external phrase memory", "authors": [ { "first": "Yaohua", "middle": [], "last": "Tang", "suffix": "" }, { "first": "Fandong", "middle": [], "last": "Meng", "suffix": "" }, { "first": "Zhengdong", "middle": [], "last": "Lu", "suffix": "" }, { "first": "Hang", "middle": [], "last": "Li", "suffix": "" }, { "first": "Philip Lh", "middle": [], "last": "Yu", "suffix": "" } ], "year": 2016, "venue": "", "volume": "", "issue": "", "pages": "", "other_ids": { "arXiv": [ "arXiv:1606.01792" ] }, "num": null, "urls": [], "raw_text": "Yaohua Tang, Fandong Meng, Zhengdong Lu, Hang Li, and Philip LH Yu. 2016. Neural machine transla- tion with external phrase memory. arXiv preprint arXiv:1606.01792.", "links": null }, "BIBREF22": { "ref_id": "b22", "title": "Modeling coverage for neural machine translation", "authors": [ { "first": "Zhaopeng", "middle": [], "last": "Tu", "suffix": "" }, { "first": "Zhengdong", "middle": [], "last": "Lu", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Xiaohua", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Hang", "middle": [], "last": "Li", "suffix": "" } ], "year": 2016, "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "76--85", "other_ids": {}, "num": null, "urls": [], "raw_text": "Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu, and Hang Li. 2016. Modeling coverage for neural machine translation. In Proceedings of the 54th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 76-85.", "links": null }, "BIBREF23": { "ref_id": "b23", "title": "Leave-one-out phrase model training for large-scale deployment", "authors": [ { "first": "Joern", "middle": [], "last": "Wuebker", "suffix": "" }, { "first": "Mei-Yuh", "middle": [], "last": "Hwang", "suffix": "" }, { "first": "Chris", "middle": [], "last": "Quirk", "suffix": "" } ], "year": 2012, "venue": "Proceedings of the Seventh Workshop on Statistical Machine Translation", "volume": "", "issue": "", "pages": "460--467", "other_ids": {}, "num": null, "urls": [], "raw_text": "Joern Wuebker, Mei-Yuh Hwang, and Chris Quirk. 2012. Leave-one-out phrase model training for large-scale deployment. In Proceedings of the Sev- enth Workshop on Statistical Machine Translation, pages 460-467.", "links": null }, "BIBREF24": { "ref_id": "b24", "title": "Training phrase translation models with leavingone-out", "authors": [ { "first": "Joern", "middle": [], "last": "Wuebker", "suffix": "" }, { "first": "Arne", "middle": [], "last": "Mauser", "suffix": "" }, { "first": "Hermann", "middle": [], "last": "Ney", "suffix": "" } ], "year": 2010, "venue": "Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics", "volume": "", "issue": "", "pages": "475--484", "other_ids": {}, "num": null, "urls": [], "raw_text": "Joern Wuebker, Arne Mauser, and Hermann Ney. 2010. Training phrase translation models with leaving- one-out. In Proceedings of the 48th Annual Meet- ing of the Association for Computational Linguis- tics, pages 475-484.", "links": null }, "BIBREF25": { "ref_id": "b25", "title": "A loss-augmented approach to training syntactic machine translation systems", "authors": [ { "first": "Tong", "middle": [], "last": "Xiao", "suffix": "" }, { "first": "F", "middle": [], "last": "Derek", "suffix": "" }, { "first": "Jingbo", "middle": [], "last": "Wong", "suffix": "" }, { "first": "", "middle": [], "last": "Zhu", "suffix": "" } ], "year": 2016, "venue": "IEEE/ACM Transactions on Audio, Speech, and Language Processing", "volume": "24", "issue": "11", "pages": "2069--2083", "other_ids": {}, "num": null, "urls": [], "raw_text": "Tong Xiao, Derek F Wong, and Jingbo Zhu. 2016. A loss-augmented approach to training syntactic ma- chine translation systems. IEEE/ACM Transac- tions on Audio, Speech, and Language Processing, 24(11):2069-2083.", "links": null }, "BIBREF26": { "ref_id": "b26", "title": "Max-violation perceptron and forced decoding for scalable MT training", "authors": [ { "first": "Heng", "middle": [], "last": "Yu", "suffix": "" }, { "first": "Liang", "middle": [], "last": "Huang", "suffix": "" }, { "first": "Haitao", "middle": [], "last": "Mi", "suffix": "" }, { "first": "Kai", "middle": [], "last": "Zhao", "suffix": "" } ], "year": 2013, "venue": "Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing", "volume": "", "issue": "", "pages": "1112--1123", "other_ids": {}, "num": null, "urls": [], "raw_text": "Heng Yu, Liang Huang, Haitao Mi, and Kai Zhao. 2013. Max-violation perceptron and forced decod- ing for scalable MT training. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1112-1123.", "links": null }, "BIBREF27": { "ref_id": "b27", "title": "Prior knowledge integration for neural machine translation using posterior regularization", "authors": [ { "first": "Jiacheng", "middle": [], "last": "Zhang", "suffix": "" }, { "first": "Yang", "middle": [], "last": "Liu", "suffix": "" }, { "first": "Huanbo", "middle": [], "last": "Luan", "suffix": "" }, { "first": "Jingfang", "middle": [], "last": "Xu", "suffix": "" }, { "first": "Maosong", "middle": [], "last": "Sun", "suffix": "" } ], "year": 2017, "venue": "Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics", "volume": "1", "issue": "", "pages": "1514--1523", "other_ids": {}, "num": null, "urls": [], "raw_text": "Jiacheng Zhang, Yang Liu, Huanbo Luan, Jingfang Xu, and Maosong Sun. 2017. Prior knowledge inte- gration for neural machine translation using poste- rior regularization. In Proceedings of the 55th An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1514- 1523.", "links": null }, "BIBREF28": { "ref_id": "b28", "title": "An improved chinese word segmentation system with conditional random field", "authors": [ { "first": "Hai", "middle": [], "last": "Zhao", "suffix": "" }, { "first": "Chang-Ning", "middle": [], "last": "Huang", "suffix": "" }, { "first": "Mu", "middle": [], "last": "Li", "suffix": "" } ], "year": 2006, "venue": "Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing", "volume": "", "issue": "", "pages": "162--165", "other_ids": {}, "num": null, "urls": [], "raw_text": "Hai Zhao, Chang-Ning Huang, Mu Li, et al. 2006. An improved chinese word segmentation system with conditional random field. In Proceedings of the Fifth SIGHAN Workshop on Chinese Language Process- ing, pages 162-165.", "links": null } }, "ref_entries": { "FIGREF0": { "text": "increase) \u7684(of) \u6e29\u5ea6(temperature) \u6216(or) pH \u53d8\u5316(change) \u3002 NMT \u8fd9\u79cd(such) \u53cd\u5e94(reaction) \u6761\u4ef6(condition) \u7684(of) \u53d8\u5316(change) \u5305\u62ec(include) \u4f46(but) \u4e0d(not)\u9650\u4e8e(limit) pH \u6216(or) pH \u7684(of) \u53d8\u5316(change) \u3002 NMT+lex \u8fd9\u79cd(such) \u53cd\u5e94(reaction) \u6761\u4ef6(condition) \u7684(of) \u53d8\u5316(change) \u5305\u62ec(include) \uff0c \u4f46(but) \u4e0d(not) \u9650\u4e8e(limit) \uff0c pH \u7684(of) \u5347\u9ad8(increase) \u6216(or) pH \u53d8\u5316(change) \u3002 NMT+lex+Pn NMT+lex+S d \u8fd9\u79cd(such) \u53cd\u5e94(reaction) \u6761\u4ef6(condition) \u7684(of) \u53d8\u5316(change) \u5305\u62ec(include) \u4f46(but) \u4e0d(not) \u9650\u4e8e(limit) \uff0c \u6e29\u5ea6(temperature) \u7684(of) \u5347\u9ad8(increase) \u6216(or) \u6539\u53d8(change) pH \u503c(value) \u3002 NMT+lex+Pn+S d \u8fd9\u79cd(such) \u53cd\u5e94(reaction) \u6761\u4ef6(condition) \u7684(of) \u53d8\u5316(change) \u5305\u62ec(include) \uff0c \u4f46(but) \u4e0d(not) \u9650\u4e8e(limit) \uff0c \u6e29\u5ea6(temperature) \u7684(of) \u5347\u9ad8(increase) \u6216(or) \u6539\u53d8(change) pH \u503c(value) \u3002 NMT+lex+Pn+WP \u8fd9\u79cd(such) \u53cd\u5e94(reaction) \u6761\u4ef6(condition) \u7684(of) \u53d8\u5316(change) \u5305\u62ec(include) \uff0c \u4f46(but) \u4e0d(not) \u9650\u4e8e(limit) \uff0c pH \u7684(of) \u5347\u9ad8(increase) \u6216(or) \u6539\u53d8(change) pH \u503c(value) \u3002 NMT+lex+S d +WP \u8fd9\u79cd(such) \u53cd\u5e94(reaction) \u6761\u4ef6(condition) \u7684(of) \u53d8\u5316(change) \u5305\u62ec(include) \uff0c \u4f46(but) \u4e0d(not) \u9650\u4e8e(limit) \uff0c \u6e29\u5ea6(temperature) \u7684(of) \u5347\u9ad8(increase) \u6216(or) \u6539\u53d8(change) pH \u503c(value) \u3002 NMT+lex+Pn+S d +WP", "num": null, "type_str": "figure", "uris": null }, "TABREF2": { "html": null, "type_str": "table", "text": "The number of times that words occur in the English-to-Chinese training corpus and the number of times that they are unaligned.", "content": "
Wordsofa practice water
Occur1.7M 0.83M8.8K7.4K
Unaligned 0.16M 0.12M0.38K 0.19K
", "num": null }, "TABREF3": { "html": null, "type_str": "table", "text": "Algorithm 2 Forced phrase-based decoding. Require: Source sentence F with length J and translation E with length I Ensure: Decoding path D initialize H 0 and s 1 , s 2 , ..., s I EXPAND(H 0 ) expand H 0 with rule null\u2192 e 1 for i = 1 to I \u2212 1 do for each hypothesis H ik in s i do EXPAND(H ik ) expand H ik with rule null\u2192 e i+1 for each hypothesis H Ik in s I do update S (H Ik ) for uncovered source words select best hypothesis in s I", "content": "
4 Reranking NMT Outputs with
Phrase-based Decoding Score
", "num": null }, "TABREF5": { "html": null, "type_str": "table", "text": "", "content": "", "num": null }, "TABREF6": { "html": null, "type_str": "table", "text": "27.72 35.67 33.46 12.37 13.95 25.96 27.50 NMT 34.60 32.71 41.67 39.00 12.52 14.05 23.63 23.99 NMT+lex 36.06 34.80 44.47 41.09 13.36 15.60 24.00 24.91 NMT+lex+rerank(Pn) 34.38 33.23 38.92 34.18 12.34 13.59 23.13 23.61 NMT+lex+rerank(S d ) 36.17 34.09 42.91 40.16 13.08 15.29 24.28 25.71 NMT+lex+rerank(Pn+S d ) 37.94 35.59 45.34 41.75 14.56 16.61 25.96 27.12 +WP) 38.69 35.75 46.92 43.17 14.61 16.65 25.98 27.15", "content": "
en-zhen-jaen-deen-fr
devtestdevtestdevtestdevtest
PBMT 30.73 NMT+lex+rerank(Pn+WP) 37.44 34.93 45.81 41.90 13.75 15.46 24.47 25.09
NMT+lex+rerank(S d +WP)36.44 33.73 43.52 40.49 13.39 15.71 24.74 26.25
NMT+lex+rerank(Pn+S d
", "num": null }, "TABREF7": { "html": null, "type_str": "table", "text": "Translation results (BLEU). NMT+lex:(Arthur et al., 2016); NMT+lex+rerank: we rerank the n-best outputs of NMT+lex using different features (P n , S d and WP).", "content": "
en-zhen-jaen-deen-fr
METEOR chrFMETEOR chrFMETEOR chrFMETEOR chrF
PBMT34.7037.87 35.2239.45 26.6650.02 32.3356.36
NMT34.5139.91 35.0742.02 24.9144.50 29.5849.99
NMT+lex35.5642.22 36.4844.34 25.4945.67 30.1050.89
NMT+lex+rerank(Pn)34.5640.80 32.6338.57 23.5740.35 29.1548.64
NMT+lex+rerank(S d )36.0242.65 36.8744.85 26.4848.73 31.5654.42
NMT+lex+rerank(Pn+S d )36.4043.73 37.2245.69 26.2647.27 31.6253.99
NMT+lex+rerank(Pn+WP)36.0442.86 36.9044.93 25.0344.05 30.2150.78
NMT+lex+rerank(S d +WP)36.3442.78 37.0545.03 26.1647.82 31.3253.75
NMT+lex+rerank(Pn+S d +WP) 36.8844.09 37.9446.66 26.2047.12 31.6153.98
", "num": null }, "TABREF8": { "html": null, "type_str": "table", "text": "METEOR and chrF scores on the test sets for different system outputs inTable 4.", "content": "
en-zhen-jaen-deen-fr
devtestdevtestdevtestdevtest
PBMT1.008 1.018 1.005 0.998 1.077 1.069 0.986 1.004
NMT0.953 0.954 0.960 0.961 1.059 1.038 0.985 0.977
NMT+lex0.936 0.966 0.955 0.963 1.054 1.019 1.030 0.977
NMT+lex+rerank(Pn)0.875 0.898 0.814 0.775 0.874 0.854 0.904 0.900
NMT+lex+rerank(S d )0.973 0.989 0.985 0.981 1.062 1.060 1.030 1.031
NMT+lex+rerank(Pn+S d )0.949 0.965 0.945 0.936 1.000 0.992 0.999 0.992
NMT+lex+rerank(Pn+WP)0.996 1.019 0.999 0.983 1.000 0.975 0.998 1.001
NMT+lex+rerank(S
", "num": null }, "TABREF10": { "html": null, "type_str": "table", "text": "An example of improving under-translation and over-translation by using S d for reranking.", "content": "", "num": null }, "TABREF11": { "html": null, "type_str": "table", "text": "Human evaluation results.", "content": "
", "num": null }, "TABREF13": { "html": null, "type_str": "table", "text": "Results of using NMT for reranking PBMT outputs.", "content": "
", "num": null } } } }