Improving Neural Machine Translation through Phrase-based Forced Decoding

11/01/2017 ∙ by Jingyi Zhang, et al. ∙ Carnegie Mellon University National Institute of Information and Communications Technology 0

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.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

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).

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); He et al. (2016), phrase memory Tang et al. (2016); Zhang et al. (2017), and

-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.

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.111In fact, our method can take in the output of any up-stream system, but we experiment exclusively with using it to rerank NMT output. 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 phrase-based 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.

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 -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 -best list.

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.

Figure 1: An example of phrase-based decoding.

2 Attentional NMT

Our baseline NMT model is similar to the attentional model of bahdanau2014neural, which includes an encoder, a decoder and an attention (alignment) model. Given a source sentence

, the encoder learns an annotation for

using a bi-directional recurrent neural network.

The decoder generates the target translation from left to right. The probability of generating next word is,222, and in Equation 1, 2 and 4 are nonlinear, potentially multi-layered, functions.

(1)

where is a decoding state for time step , computed by,

(2)

is a source representation for time , calculated as,

(3)

where scores how well the inputs around position and the output at position match, computed as,

(4)

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 under-translation Cohn et al. (2016); Mi et al. (2016).

3 Phrase-based Forced Decoding for NMT

3.1 Phrase-based SMT

In phrase-based SMT Koehn et al. (2003), a phrase-based translation rule includes a source phrase, a target phrase and a translation score . 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 consists of two rules and . The score of a hypothesis can be calculated as the product of the scores of all applied rules.333In 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. An existing hypothesis can be expanded into a new hypothesis by applying a new rule. As shown in Figure 1, can be expanded into , and . cannot be further expanded, because it covers all source words, while and can (and must) be further expanded. The decoder starts with an initial empty hypothesis and selects the hypothesis with the highest score from all completed hypotheses.

During decoding, hypotheses are stored in stacks. For a source sentence with words, the decoder builds stacks. The hypotheses that cover source words are stored in stack . The decoder expands hypotheses in in turn as shown in Algorithm 1.

Source sentence with length
Translation and decoding path
initialize and
Expand
for  to  do
   for each hypothesis in  do
      Expand    
select best hypothesis in
Algorithm 1 Standard phrase-based decoding.

Here, Expand is expanding to get new hypotheses and putting the new hypotheses into corresponding stacks. For each stack, a beam of the best hypotheses is kept to speed up the decoding process.

3.2 Forced Decoding for NMT

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.

To remedy this problem, inspired by previous work in forced decoding for training phrase-based SMT systems Wuebker et al. (2010, 2012) we propose a soft forced decoding algorithm that can always successfully find a decoding path for a source sentence and an NMT translation .

First, we introduce two new types of rules R and R.

R

A source word can be translated into a special word null. This corresponds to deleting during translation. The score of deleting is calculated as,

(5)

where is how many times is unaligned in the word-aligned training set and is the number of sentence pairs in .

R

A target word can be translated from a special word null, which corresponds to inserting during translation. The score of inserting is calculated as,

(6)

where is how many times is unaligned in .

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.

Words of a practice water
Occur 1.3M 1.0M 2.2K 29K
Unaligned 0.51M 0.41M 0.25K 3.5K
Table 1: The number of times that words occur in the English-to-Chinese training corpus and the number of times that they are unaligned.
Words of a practice water
Occur 1.7M 0.83M 8.8K 7.4K
Unaligned 0.16M 0.12M 0.38K 0.19K
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.

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 the scale of the scores for R and R 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/R.

Source sentence with length and translation with length
Decoding path
initialize and
Expand
expand with rule null
for  to  do
   for each hypothesis in  do
      Expand
      expand with rule null    
for each hypothesis in  do
   update for uncovered source words
select best hypothesis in
Algorithm 2 Forced phrase-based decoding.

Algorithm 2 shows the forced decoding algorithm that integrates the new rules. Because the translation is given for the forced decoding algorithm, the proposed forced decoding algorithm keeps stacks, where is the length of . 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 in Algorithm 2 contains all hypotheses in which the first target words have been generated. We expand hypotheses in in turn. When expanding a hypothesis in , besides expanding it using the original rule table Expand,444The new introduced word inserting/deleting rules are not used when performing Expand. we also expand by inserting the next target word at the end of to get an additional hypothesis and put into . For a final hypothesis in stack , it may not cover all source words. We update its score by translating uncovered words into null.

Because different decoding paths can generate the same final translation, there can be different decoding paths that fit the NMT translation . We use the score of the single decoding path with the highest decoding score as the forced decoding score for .

4 Reranking NMT Outputs with Phrase-based Decoding Score

We rerank the -best NMT outputs using the phrase-based forced decoding score according to Equation 7.

(7)

where is the original NMT translation probability as calculated by Equation 1;

(8)

is the forced decoding score, which is the score of the decoding path with the highest decoding score as described above;

(9)

and are weights that can be tuned on the -best list of the development set.

The easiest way to get an -best list for NMT is by using the

-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.

The standard method for sampling hypotheses in NMT is ancestral sampling, where we randomly select a word from the vocabulary according to 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 -best list may not contain any candidate with better translation quality than the standard beam search output.

Instead, we take an alternative approach that proved empirically better in our experiments: at each time step , we use sampling to randomly select the next word from and according to Equation 10. Here, and are the two target words with the highest probability according to Equation 1.

(10)

The sampling process ends when is selected as the next word.

We repeat the decoding process times to sample 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 .

5 Experiments

5.1 Settings

SOURCE TARGET
en-de TRAIN #Sents 1.90M
#Words 52.2M 49.7M
#Vocab 113K 376K
DEV #Sents 3,003
#Words 67.6K 63.0K
TEST #Sents 2,169
#Words 46.8K 44.0K
en-fr TRAIN #Sents 1.99M
#Words 54.4M 60.4M
#Vocab 114K 137K
DEV #Sents 3,003
#Words 71.1K 81.1K
TEST #Sents 1.5K
#Words 27.1K 29.8K
en-zh TRAIN #Sents 954K
#Words 40.4M 37.2M
#Vocab 504K 288K
DEV #Sents 2K
#Words 77.5K 75.4K
TEST #Sents 2K
#Words 58.1K 55.5K
en-ja TRAIN #Sents 3.14M
#Words 104M 118M
#Vocab 273K 150K
DEV #Sents 2K
#Words 66.5K 74.6K
TEST #Sents 2K
#Words 70.6K 78.5K
Table 3: Data sets.
en-zh en-ja en-de en-fr
dev test dev test dev test dev test
PBMT 30.73 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() 34.38 33.23 38.92 34.18 12.34 13.59 23.13 23.61
NMT+lex+rerank() 36.17 34.09 42.91 40.16 13.08 15.29 24.28 25.71
NMT+lex+rerank(+) 37.94 35.59 45.34 41.75 14.56 16.61 25.96 27.12
NMT+lex+rerank(+WP) 37.44 34.93 45.81 41.90 13.75 15.46 24.47 25.09
NMT+lex+rerank(+WP) 36.44 33.73 43.52 40.49 13.39 15.71 24.74 26.25
NMT+lex+rerank(++WP) 38.69 35.75 46.92 43.17 14.61 16.65 25.98 27.15
Table 4: Translation results (BLEU). NMT+lex: Arthur et al. (2016); NMT+lex+rerank: we rerank the -best outputs of NMT+lex using different features (, and WP).
en-zh en-ja en-de en-fr
METEOR chrF METEOR chrF METEOR chrF METEOR chrF
PBMT 34.70 37.87 35.22 39.45 26.66 50.02 32.33 56.36
NMT 34.51 39.91 35.07 42.02 24.91 44.50 29.58 49.99
NMT+lex 35.56 42.22 36.48 44.34 25.49 45.67 30.10 50.89
NMT+lex+rerank() 34.56 40.80 32.63 38.57 23.57 40.35 29.15 48.64
NMT+lex+rerank() 36.02 42.65 36.87 44.85 26.48 48.73 31.56 54.42
NMT+lex+rerank(+) 36.40 43.73 37.22 45.69 26.26 47.27 31.62 53.99
NMT+lex+rerank(+WP) 36.04 42.86 36.90 44.93 25.03 44.05 30.21 50.78
NMT+lex+rerank(+WP) 36.34 42.78 37.05 45.03 26.16 47.82 31.32 53.75
NMT+lex+rerank(++WP) 36.88 44.09 37.94 46.66 26.20 47.12 31.61 53.98
Table 5: METEOR and chrF scores on the test sets for different system outputs in Table 6.
en-zh en-ja en-de en-fr
dev test dev test dev test dev test
PBMT 1.008 1.018 1.005 0.998 1.077 1.069 0.986 1.004
NMT 0.953 0.954 0.960 0.961 1.059 1.038 0.985 0.977
NMT+lex 0.936 0.966 0.955 0.963 1.054 1.019 1.030 0.977
NMT+lex+rerank() 0.875 0.898 0.814 0.775 0.874 0.854 0.904 0.900
NMT+lex+rerank() 0.973 0.989 0.985 0.981 1.062 1.060 1.030 1.031
NMT+lex+rerank(+) 0.949 0.965 0.945 0.936 1.000 0.992 0.999 0.992
NMT+lex+rerank(+WP) 0.996 1.019 0.999 0.983 1.000 0.975 0.998 1.001
NMT+lex+rerank(+WP) 1.000 1.024 1.001 1.001 1.011 1.007 0.999 0.989
NMT+lex+rerank(++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 6.

We evaluated the proposed approach for English-to-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 en-ja tasks, we used datasets provided for the patent machine translation task at NTCIR-9 Goto et al. (2011).555Note 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. 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 Mecab666http://sourceforge.net/projects/mecab/files/ for Japanese.

We built attentional NMT systems with Lamtram777https://github.com/neubig/lamtram. 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.

Source for hypophysectomized (hypop hy sec to mized) rats , the drinking water additionally contains 5 % glucose .
Reference 对于(for) 去(remove) 垂体(hypophysis) 大(big) 鼠(rat) , 饮用水(drinking water) 中(in) 另外(also) 含有(contain) 5 % 葡萄糖(glucose) 。
PBMT 用于(for) 大(big) 鼠(rat) 垂体(hypophysis) HySecto,(Hy Sec to ,) 饮用水(drinking water) 另外(also) 含有(contain) 5 % 葡萄糖(glucose) 。
NMT 对于(for) 过(pass) 盲肠(cecum) 的(of) 大(big) 鼠(rat) , 饮用水(drinking water) 另外(also) 含有(contain) 5 % 葡萄糖(glucose) 。
NMT+lex 对于(for) 低(low) 酪(cheese) 蛋白(protein) 切除(remove) 的(of) 大(big) 鼠(rat) , 饮用水(drinking water) 另外(also) 含有(contain) 5 % 葡萄糖(glucose) 。
NMT+lex+
NMT+lex++WP
NMT+lex+ 对于(for) 垂体(hypophysis) 在(is) 切除(remove) 大(big) 鼠(rat) 中(in) , 饮用水(drinking water) 另外(also) 含有(contain) 5 % 葡萄糖(glucose) 。
NMT+lex++WP
NMT+lex++ 对于(for) 垂体(hypophysis) 在(is) 切除(remove) 的(of) 大(big) 鼠(rat) 中(in) , 饮用水(drinking water) 另外(also) 含有(contain) 5 % 葡萄糖(glucose) 。
NMT+lex+++WP
Table 7: An example of improving inaccurate rare word translation by using for reranking.

To obtain a phrase-based translation rule table for our forced decoding algorithm, we used GIZA++ Och and Ney (2003) and grow-diag-final-andheuristic to obtain symmetric word alignments for the training set. Then we extracted the rule table using Moses Koehn et al. (2007).

Source such changes in reaction conditions include , but are not limited to , an increase in temperature or change in ph .
Reference 所(such) 述(said) 反应(reaction) 条件(condition) 的(of) 改变(change) 包括(include) 但(but) 不(not) 限于(limit) 温度(temperature) 的(of) 增加(increase) 或(or) pH 值(value) 的(of) 改变(change)
PBMT 中(in) 的(of) 这种(such) 变化(change) 的(of) 反应(reaction) 条件(condition) 包括(include) , 但(but) 不(not) 限于(limit) , 增加(increase) 的(of) 温度(temperature) 或(or) pH 变化(change)
NMT 这种(such) 反应(reaction) 条件(condition) 的(of) 变化(change) 包括(include) 但(but) 不(not) 限于(limit) pH 或(or) pH 的(of) 变化(change)
NMT+lex 这种(such) 反应(reaction) 条件(condition) 的(of) 变化(change) 包括(include) , 但(but) 不(not) 限于(limit) , pH 的(of) 升高(increase) 或(or) pH 变化(change)
NMT+lex+
NMT+lex+ 这种(such) 反应(reaction) 条件(condition) 的(of) 变化(change) 包括(include) 但(but) 不(not) 限于(limit) , 温度(temperature) 的(of) 升高(increase) 或(or) 改变(change) pH 值(value)
NMT+lex++ 这种(such) 反应(reaction) 条件(condition) 的(of) 变化(change) 包括(include) , 但(but) 不(not) 限于(limit) , 温度(temperature) 的(of) 升高(increase) 或(or) 改变(change) pH 值(value)
NMT+lex++WP 这种(such) 反应(reaction) 条件(condition) 的(of) 变化(change) 包括(include) , 但(but) 不(not) 限于(limit) , pH 的(of) 升高(increase) 或(or) 改变(change) pH 值(value)
NMT+lex++WP 这种(such) 反应(reaction) 条件(condition) 的(of) 变化(change) 包括(include) , 但(but) 不(not) 限于(limit) , 温度(temperature) 的(of) 升高(increase) 或(or) 改变(change) pH 值(value)
NMT+lex+++WP
Table 8: An example of improving under-translation and over-translation by using for reranking.

5.2 Results and Analysis

Table 6 shows results of the phrase-based SMT system888We used the default Moses settings for phrase-based SMT., 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 , the forced decoding score 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 level using bootstrap resampling Koehn (2004) are shown in bold font.

As we can see, integrating lexical translation probabilities improved the baseline NMT system 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 and for reranking gave better results than only using or alone. We also show METEOR and chrF scores on the test sets in Table 6. Our reranking method improved both METEOR and chrF significantly.

The Word Penalty Feature

The word penalty feature generally improved the reranking results, especially when only the NMT score was used for reranking. As we can see, using only for reranking decreased the translation quality compared 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 reranking results do have higher NMT scores, but have lower BLEU scores according to Table 6. To explain why this happened, we show the ratio of translation length to reference length in Table 6. As we can see, the 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 reranking results considerably, by preferring longer sentences. Because the forced decoding score as shown in Equation 9 does not obviously prefer shorter or longer sentences, when was used for reranking, the word penalty feature became less helpful. When both and were used for reranking, the word penalty feature only achieved further significant improvement on the English-to-Japanese task.

T (NMT+lex):
for 对于(for) -3.04
: hy 低(low) -12.19
: null酪(cheese) -21.99
: null蛋白(protein) -13.83
to mized 切除(remove) -6.22
null的(of) -1.53
rats 大(big) 鼠(rat) -1.52
, the drinking water , 饮用水(drinking water) -1.38
additionally contains 另外(also) 含有(contain) -3.68
5 % 5 % -0.51
glucose . 葡萄糖(glucose) 。 -0.60
: hypopnull -25.33
secnull -20.66
T (NMT+lex++):
for 对于(for) -3.04
hypop hy 垂体(hypophysis) -5.09
the 在(is) -5.32
to mized 切除(remove) -6.22
null的(of) -1.53
rats 大(big) 鼠(rat) -1.52
, 中(in) , -4.11
drinking water 饮用水(drinking water) -1.03
additionally contains 另外(also) 含有(contain) -3.68
5 % 5 % -0.51
glucose . 葡萄糖(glucose) 。 -0.60
secnull -20.66
Table 9: Forced decoding paths for T and T: used rules and log scores. The translation rules with shade are used only for T or T.

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 “去(remove) 垂体(hypophysis)” as shown in the reference sentence. The original attentional NMT translated it into incorrect translation “过(pass) 盲肠(cecum)”. After integrating lexicons, the NMT system translated it into “低(low) 酪(cheese) 蛋白(protein) 切除(remove)”. The last word “切除(remove)” is correct, but the rest of the translation is still wrong. Only by using the forced decoding score for reranking, we get the more accurate translation “垂体(hypophysis) 在(is) 切除(remove)”.

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, the NMT translation without reranking and T, the NMT translation using both and for reranking. As we can see, the four rules , , and used for T have low scores. is an unlikely translation. In , and , “酪(cheese)”, “蛋白(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 “的(of)” is very high. The translation rule “the 在(is)” used for T is incorrect, but its score is relatively high, because function words are often 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.

In T, “hypophysectomized (hypop hy sec to mized)” is incorrectly translated into “低(low) 酪(cheese) 蛋白(protein) 切除(remove)”. However, from Table 9, we can see that the forced decoding algorithm learns it as unlikely translation (hy低(low)), over-translation (null酪(cheese), null蛋白(protein)) and under-translation (hypopnull, secnull), because there is no translation rule between “hypop” “sec” and “酪(cheese)” “蛋白(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 for reranking, NMT under-translated “temperature” and over-translated “ph” twice, which will be assigned low scores by forced decoding. By using for reranking, the correct translation was selected.

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:

  • NMT+lex+rerank(+WP)

  • NMT+lex+rerank(++WP)

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.

Reduce both under- and over- translation 2
under-translation 11
over-translation 10
unrelated translation 4
No difference 73
Table 10: Human evaluation results.

Reranking PBMT Outputs with NMT

We also did experiments that use the NMT score as an additional feature to rerank PBMT outputs (unique -best list). The results are shown in Table 11. We also copy results of baseline PBMT and NMT from Table 6 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.

en-zh en-ja en-de en-fr
PBMT+rerank 32.77 37.68 14.23 28.86
PBMT dev 30.73 35.67 12.37 25.96
NMT 34.60 41.67 12.52 23.63
PBMT+rerank 30.04 35.14 15.89 29.77
PBMT test 27.72 33.46 13.95 27.50
NMT 32.71 39.00 14.05 23.99
Table 11: Results of using NMT for reranking PBMT outputs.

6 Related Work

wuebker-mauser-ney:2010:ACL,wuebker-hwang-quirk:2012:WMT applied forced decoding on the training set to improve the training process of phrase-based SMT and prune the phrase-based 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); Meng et al. (2016); Alkhouli et al. (2016) while we rerank NMT outputs by exploiting knowledge from traditional SMT.

There are also existing methods that rerank NMT outputs by using target-bidirectional NMT models Liu et al. (2016); 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.

7 Conclusion

In this paper, we propose to exploit an existing phrase-based SMT model to compute the phrase-based decoding cost for NMT outputs and then use the phrase-based decoding cost to rerank the -best 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.

References

  • Alkhouli et al. (2016) Tamer Alkhouli, Gabriel Bretschner, Jan-Thorsten Peter, Mohammed Hethnawi, Andreas Guta, and Hermann Ney. 2016. Alignment-based neural machine translation. In Proceedings of the First Conference on Machine Translation, pages 54–65.
  • Arthur et al. (2016) 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 Natural Language Processing

    , pages 1557–1567.
  • Bahdanau et al. (2015) Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations.
  • Brown et al. (1993) Peter F Brown, Vincent J Della Pietra, Stephen A Della Pietra, and Robert L Mercer. 1993.

    The mathematics of statistical machine translation: Parameter estimation.

    Computational Linguistics, 19(2):263–311.
  • Cohn et al. (2016) Trevor Cohn, Cong Duy Vu Hoang, Ekaterina Vymolova, 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 Computational Linguistics: Human Language Technologies, pages 876–885.
  • Goto et al. (2011) Isao Goto, Bin Lu, Ka Po Chow, Eiichiro Sumita, and Benjamin K Tsou. 2011. Overview of the patent machine translation task at the NTCIR-9 workshop. In Proc. NTCIR-9, pages 559–578.
  • He et al. (2016) Wei He, Zhongjun He, Hua Wu, and Haifeng Wang. 2016. Improved neural machine translation with SMT features. In

    Thirtieth AAAI conference on artificial intelligence

    .
  • Koehn (2004) Philipp Koehn. 2004. Statistical significance tests for machine translation evaluation. In Proceedings of the 2004 Conference on Empirical Methods in Natural Language Processing, pages 388–395.
  • Koehn et al. (2007) 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 Association for Computational Linguistics Companion Volume Proceedings of the Demo and Poster Sessions, pages 177–180.
  • Koehn and Knowles (2017) Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28–39.
  • Koehn et al. (2003) Philipp Koehn, Franz Josef Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology-Volume 1, pages 48–54.
  • Liu et al. (2016) Lemao Liu, Masao Utiyama, Andrew Finch, and Eiichiro Sumita. 2016. Agreement on target-bidirectional neural machine translation. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 411–416.
  • Meng et al. (2016) Fandong Meng, Zhengdong Lu, Hang Li, and Qun Liu. 2016. Interactive attention for neural machine translation. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 2174–2185.
  • Mi et al. (2016) 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.
  • Neubig et al. (2015) Graham Neubig, Makoto Morishita, and Satoshi Nakamura. 2015. Neural reranking improves subjective quality of machine translation: NAIST at WAT2015. In Proceedings of the 2nd Workshop on Asian Translation (WAT2015).
  • Och and Ney (2003) Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51.
  • Sennrich et al. (2016a) Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016a. Edinburgh neural machine translation systems for WMT 16. In Proceedings of the First Conference on Machine Translation, pages 371–376.
  • Sennrich et al. (2016b) Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016b. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715–1725.
  • Shen et al. (2016) Shiqi Shen, Yong Cheng, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. 2016. Minimum risk training for neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1683–1692.
  • Stahlberg et al. (2017) Felix Stahlberg, Adrià 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 Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 362–368.
  • Stahlberg et al. (2016) Felix Stahlberg, Eva Hasler, Aurelien Waite, and Bill Byrne. 2016. Syntactically guided neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 299–305.
  • Tang et al. (2016) Yaohua Tang, Fandong Meng, Zhengdong Lu, Hang Li, and Philip LH Yu. 2016. Neural machine translation with external phrase memory. arXiv preprint arXiv:1606.01792.
  • Tu et al. (2016) Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu, and Hang Li. 2016. Modeling coverage for neural machine translation. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 76–85.
  • Wuebker et al. (2012) Joern Wuebker, Mei-Yuh Hwang, and Chris Quirk. 2012. Leave-one-out phrase model training for large-scale deployment. In Proceedings of the Seventh Workshop on Statistical Machine Translation, pages 460–467.
  • Wuebker et al. (2010) Joern Wuebker, Arne Mauser, and Hermann Ney. 2010. Training phrase translation models with leaving-one-out. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 475–484.
  • Xiao et al. (2016) Tong Xiao, Derek F Wong, and Jingbo Zhu. 2016. A loss-augmented approach to training syntactic machine translation systems. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 24(11):2069–2083.
  • Yu et al. (2013) Heng Yu, Liang Huang, Haitao Mi, and Kai Zhao. 2013.

    Max-violation perceptron and forced decoding for scalable MT training.

    In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1112–1123.
  • Zhang et al. (2017) Jiacheng Zhang, Yang Liu, Huanbo Luan, Jingfang Xu, and Maosong Sun. 2017. Prior knowledge integration for neural machine translation using posterior regularization. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1514–1523.
  • Zhao et al. (2006) 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 Processing, pages 162–165.