Fill in the Blanks: Imputing Missing Sentences for Larger-Context Neural Machine Translation

10/30/2019
by   Sébastien Jean, et al.
0

Most neural machine translation systems still translate sentences in isolation. To make further progress, a promising line of research additionally considers the surrounding context in order to provide the model potentially missing source-side information, as well as to maintain a coherent output. One difficulty in training such larger-context (i.e. document-level) machine translation systems is that context may be missing from many parallel examples. To circumvent this issue, two-stage approaches, in which sentence-level translations are post-edited in context, have recently been proposed. In this paper, we instead consider the viability of filling in the missing context. In particular, we consider three distinct approaches to generate the missing context: using random contexts, applying a copy heuristic or generating it with a language model. In particular, the copy heuristic significantly helps with lexical coherence, while using completely random contexts hurts performance on many long-distance linguistic phenomena. We also validate the usefulness of tagged back-translation. In addition to improving BLEU scores as expected, using back-translated data helps larger-context machine translation systems to better capture long-range phenomena.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 1

page 2

page 3

page 4

10/11/2020

Lexically Cohesive Neural Machine Translation with Copy Mechanism

Lexically cohesive translations preserve consistency in word choices in ...
02/20/2021

Understanding and Enhancing the Use of Context for Machine Translation

To understand and infer meaning in language, neural models have to learn...
03/12/2019

Context-Aware Learning for Neural Machine Translation

Interest in larger-context neural machine translation, including documen...
05/25/2018

Context-Aware Neural Machine Translation Learns Anaphora Resolution

Standard machine translation systems process sentences in isolation and ...
10/09/2020

Dynamic Context Selection for Document-level Neural Machine Translation via Reinforcement Learning

Document-level neural machine translation has yielded attractive improve...
12/06/2018

Context is Key: New Approaches to Neural Coherence Modeling

We formulate coherence modeling as a regression task and propose two nov...
08/14/2018

Discrete Structural Planning for Neural Machine Translation

Structural planning is important for producing long sentences, which is ...
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 (Sutskever et al., 2014; Bahdanau et al., 2015) has gained considerable popularity, attracting significant research interest and being widely adopted in industry. While results are impressive, sometimes arguably rivaling human quality (Hassan et al., 2018), there is room for further progress as most neural translation systems are trained to translate sentences in isolation (Läubli et al., 2018). This may lead to undesirable behaviour when translating longer text segments, such as paragraphs or documents. For example, the output may contain incorrect gender markers or generally be inconsistent Popescu-Belis (2019).

Given these limitations, there has been growing interest in document-level neural machine translation, such as in the En-De and En-Cs shared tasks at WMT’19 Barrault et al. (2019). In this work, we consider a scenario where we are provided a combination of document-level and sentence-level only parallel data. Previous work Voita et al. (2019) has tackled this problem with a two-stage approach, initially generating sentences separately and then correcting the output in context. Here, we instead explore the viability of generating the missing context.

We consider multiple techniques to complete the missing data. In all cases, we do not modify the examples where context is already provided. As baselines, we may leave the context-less examples intact, or simply add random sentence pairs to such examples (Junczys-Dowmunt, 2019). We also present a simple partial copy heuristic, where training examples are constructed by combining multiple copies of a source-target pair with random ones. Alternatively, we may sample target-side context with a language model and obtain the corresponding source context through back-translation. Finally, we also consider retrieving similar sentence pairs within the training data to use as context.

By evaluating these context-aware models on multiple challenge sets Voita et al. (2019) targeting specific longe-range phenomena, we observe that the choice of context completion technique clearly impacts performance. In particular, the partial copy heuristic leads to significantly better lexical coherence. In contrast, creating examples where all sentences are unrelated is harmful on all challenge sets, even if this negative impact is not necessarily visible from BLEU scores.

Given the importance of appropriate data to capture long-range phenomena, we also validate the effectiveness of (tagged) back-translation Sennrich et al. (2016); Caswell et al. (2019a); Junczys-Dowmunt (2019) for document-level machine translation. Adding back-translated data is clearly helpful across all specific linguistic phenomena we evaluate, although the initial context completion technique still visibly impacts performance. Back-translation also increases the general translation quality as measured by BLEU, although we see similar gains for sentence-level models.

The contributions of this paper are three folds:

  • In a scenario where only some of the parallel data has document-level context, we establish the viability of completing the missing data.

  • We present multiple approaches to fill in the missing context and evaluate their impact on multiple linguistic phenomena.

  • We validate the effectiveness of (tagged) back-translation for document-level machine translation.

2 Background

2.1 Neural machine translation

A typical neural machine translation system (Sutskever et al., 2014; Bahdanau et al., 2015)

is a neural network mapping an input sequence

to its translation . Words (or subword tokens) within are mapped to a continuous space with word embeddings. The embeddings are consumed by an encoder network which produces one or multiple source representations . Using these source representations, a decoder network predicts the output, often auto-regressively , although other factorizations may be employed Mansimov et al. (2019). Both the encoder and decoder may be implemented with recurrent, convolutional or self-attentive networks, among others (Bahdanau et al., 2015; Gehring et al., 2017; Vaswani et al., 2017). The encoder and decoder are generally separate networks, although they can alternatively be merged (He et al., 2018).

Given a set of parallel sentence pairs of size , neural machine translation systems are often optimized by minimizing the negative log-likelihood:

To generate translations, given the intractability of , at least under the auto-regressive framework, it is common to use approximate search algorithms, i.e. beam search, which prunes the output of a breadth-first search at every time step.

2.2 Document-level neural machine translation

Sentence-level systems are unable to model some long-distance phenomena correctly. For example, in some languages such as Spanish, pronouns may be dropped. As such, without additional context, an English translation would have to guess the correct gender. Moreover, politeness markers, named entities and other words may also be translated inconsistently across sentences.

To correct such weaknesses, models may use additional context information, for example neighbouring sentences within a document. Document-level machine translation approaches can currently be broadly classified into three categories. First, the additional context may simply be concatenated to the sentence to translate 

(Tiedemann and Scherrer, 2017; Junczys-Dowmunt, 2019), which allows using the same models as for sentence-level translation. The additional target-side context may either be predicted or not by such models. Alternatively, contextual information may be used as a separate input to the model, in which case the model architecture may be adapted in consequence (Jean et al., 2017; Wang et al., 2017; Voita et al., 2018; Zhang et al., 2018; Miculicich et al., 2018; Maruf and Haffari, 2018; Tu et al., 2018). Finally, sentence-by-sentence translations may be post-edited to integrate contextual information (Voita et al., 2019, 2019).

2.3 Evaluating contextual translation systems

The most commonly used metric to evaluate machine translation output is BLEU (Papineni et al., 2002)

. BLEU is a precision-based metric that considers n-gram overlap between a system output and one or more human reference translations. Despite its usefulness, it is not necessarily ideal to evaluate contextual phenomena as only n-grams of size of to 4 are considered.

Various challenge sets have been constructed to specifically evaluate specific phenomena, such as coreference resolution, pronoun choice and coherence (Bawden et al., 2018; Müller et al., 2018). For this paper, we consider four challenge sets targeting lexical cohesion, deixis (politeness marker), verb phrase ellipsis and morphological inflexion for EnglishRussian (EnRu).

These four challenge sets consist of multiple-choice questions with identical source sentences, source contexts and target contexts. To answer an example appropriately, the model must assign a higher probability to the correct answer than to the distractors. However, as these are scoring challenges, choosing the right answer does not necessarily entail that the model would generate that exact sentence.

Lexical cohesion

This challenge set evaluates whether a model can consistently translate named entities across sentences. For each example, a named entity in multiple ways, each acceptable without context, but with only one matching a previous translation. For a model to be successful on this task, use of target-side context is beneficial.

While this challenge set only evaluates EnRu, maintaining lexical cohesion is a desirable property across many languages. It is split into validation and test subsets, of size 500 and 1500 respectively.

Deixis

While deixis covers multiple sub-phenomena, this challenge set specifically targets the T-V distinction (from the latin and ). The T form is informal, while the V form is more polite, used for example when addressing people with higher social standing. The T-V distinction may involve changes to multiple words within the target sentence in order to maintain grammatical correctness.

In addition to Russian, the T-V distinction is also present in other languages such as French (singular vous) or Spanish ( / usted). As for lexical cohesion, use of target-side context is crucial, and there is both validation and test data, of size 500 and 2500.

Verb phrase ellipsis

This challenge set mostly considers the translation of the verb (including its different conjugations). In Russian and several other languages, it is necessary to specify what is being done. The necessary information may often be found in neighboring sentences, either on the source or target side.

In general, each example within this challenge set contains many more distractors than for lexical cohesion or deixis. In this challenge set only test data is available, so that we can not monitor progress on this phenomenon when training a model.

Ellipsis (inflection)

Russian words are richly inflected. There are two numbers (singular/plural), three genders (masculine/feminine/neuter) and six grammatical cases depending of the usage of a word. In some instances, the grammatical case of a translated word can not be inferred with certainty from the source sentence only, but context may help disambiguate the word. Similarly, in this challenge set only the test data is available.

2.4 Back-translation

Back-translation (Sennrich et al., 2016) is an effective technique to leverage monolingual data, which is generally more plentiful than parallel data, for machine translation. In order to improve a model, an inverse direction model is trained, generally on parallel data. The latter is used to translate many sentences in the target language in order to obtain synthetic source sentences. These examples, with a synthetic source and natural target, are then included in the training data alongside the original parallel sentence pairs.

Back-translation can further be improved by different techniques. The models and training data may be further refined through various rounds of back-translation (Hoang et al., 2018). Adding further noise to the generated source samples may also be beneficial (Edunov et al., 2018). To help the model easily distinguish between the original training data and the back-translated samples, tagged-backtranslation prepends a special token on the source side of back-translated data (Caswell et al., 2019b) which improves the efficacy of the technique further.

3 Context completion

Sentence Ru En
Context-3 Крути педали. Then pedal faster, Third Wheel.
Context-2 Она постоянно выходила из себя. She got mad at everything.
Context-1 Что мне кажется удивительным. Which I find surprising..
Current Она постоянно выходила из себя. She got mad at everything.
Table 1: Example augmented with the partial copy heuristic.

Many sources of parallel data include document-level information, such as recent versions of Europarl111statmt.org/europarl/v9 and News-Commentary222data.statmt.org/news-commentary/v14 (Barrault et al., 2019). Nevertheless, there are many datasets where only out-of-context sentence pairs are available (Voita et al., 2019). In this paper, we consider a scenario where some, but not all, parallel data includes document-level information.

We ask how we may fill in the missing contextual information to improve the model quality. In particular, to match evaluation conditions, we modify the training data so that all examples have three source and target sentence pairs as context. Examples for which there already is context are left untouched. We explore three different techniques.

Random context

A simple technique to fill in missing contextual information is to add random sentence pairs. This approach, although not limited to three contextual sentences exclusively, was previously used to create fake documents (Junczys-Dowmunt, 2019). As the artificial context will most likely be unrelated to the sentences to translate, this approach may have the adverse effect of biasing models towards ignoring context.

Partial copy heuristic

As an alternative to using random context, we design a heuristic designed to encourage the model to sometimes consider the additional context, as well as to copy words or sub-word tokens.

Given a sentence pair, we build the context by adding a copy of the input, as well as two random sentence pairs. To improve robustness to sentence order, we randomly shuffle the four sentence pairs (2 identical, 2 distinct).

We also tried copying the input multiple times, without any unrelated sentences. However, general performance suffered, most likely because the training and evaluation conditions became too different. An augmented example is included in Table 1.

Context generation

With access to distinct target-side monolingual data, we extract consecutive segments of four sentences. Given the last sentence, an encoder-decoder model is trained to predict the previous three, similarly to a skip-thought model (Kiros et al., 2015). Alternatively, a standard language model could be also be used, although the generation process would have to be modified slightly.

This target-to-context model is applied to the original target side of the parallel data to generate plausible contexts. We observed very repetitive outputs with beam search, so we instead use sampling to recover more diverse contexts.

To obtain the source-side context, we back-translate the sampled context with a reverse-direction model. We also include the original target sentence within the input, but remove the last generated sentence from the output and replace it with the original source. For back-translation, we use a reverse-direction context-aware model trained with the partial copy heuristic.

Especially if the translation system is trained to predict the target context, the potential mismatch between real target sentences and generated samples can potentially affect translation quality.

4 Experiments

4.1 Parallel data

Both parallel and monolingual data are sourced from OpenSubtitles2018 Lison et al. (2018), which is a collection of movie and TV show subtitles. We use the same 6 million EnRu training examples as in Voita et al. (2019), out of which 1.5 million have context. For these examples, the previous three source and target sentences are provided. Given the fixed-sized context, many of these training instances have overlapping sentences. We also use the same validation and test sets, for both general performance and BLEU.

We train a SentencePiece model (Kudo and Richardson, 2018) separately on the English and Russian side of the training data, with vocabularies of size 32,768. Given the different tokenization, BLEU scores are not directly comparable to Voita et al. (2019, 2019), so we retrain some baseline models. Scores on the lexical cohesion, deixis, ellipsis (VP) and ellipsis (infl.) challenge sets remain comparable.

4.2 Monolingual data

We gather document-level Russian monolingual data from OpenSubtitles2018. As the raw data is time-stamped, we concatenate multiple consecutive sentences within the same movie or TV show as long as the time difference between two sentences is at most 2 seconds.

The retrieved strings, which may contain arbitrarily many sentences, are filtered to avoid overlap with all the validation and test sets. More specifically, the last sentence of any validation or test example is prohibited to be within the monolingual training data, but contextual sentences are allowed.333Whitespace within sentences is removed for filtering to deal with potential tokenization differences. Strings with less than 4 sentences are removed, while longer segments are split into overlapping 4-sentence examples. The final data contains approximately 26 million examples.

This filtering criterion is only applied to the monolingual data we extract, but not to the parallel data (Voita et al., 2019). In that case, the training and test instances came from different movies, which could still have overlapping sentences.

4.3 Models

We use Transformer models Vaswani et al. (2017)

for all experiments. In particular, with bilingual data only, we use Transformer base models, with 6 encoder and decoder layers, 8-head attention, model and feed-forward projection dimensions of 512 and 2048 respectively. To reduce overfitting, the dropout rate is set to 20%.

With additional target-side monolingual data and back-translated source inputs, we additionally train Transformer big models. Even with augmented training data, we observed significant overfitting, and as such set the dropout hyper-parameter to 40%.

Models are trained on v3 TPUs in a 4x4 topology (32 cores). Baseline models are trained with inputs of size 64 by 128 for each core. Multiple sentences, of maximum length 98, are packed within each row to improve computational efficiency. To allow for longer inputs for context-aware models, batches have size 16 x 512, with maximum length 512. Sentences within each example are concatenated to each other, with a reserved token separating each sentence. All our experiments were performed using the Tensorflow Lingvo

(Shen et al., 2019) framework.

Model outputs are generated with beam search, with a beam width of 8. BLEU scores are evaluated with an in-house re-implementation of mteval-v13a.444https://github.com/moses-smt/mosesdecoder/blob/master/scripts/generic/mteval-v13a.pl

Auxiliary systems

The reverse-direction system used to back-translate additional data is trained on the parallel corpus only, with the missing context replaced with the partial copy heuristic. All the extracted monolingual 4-sentence examples are back-translated, unless their length exceeds 512 subword tokens. Following Caswell et al. (2019a), a tag is prepended to the predicted source samples. Note that for the baseline system trained with additional monolingual data, we still use the context-aware RuEn model for back-translation, but only add the last sentence pair (out of 4) to the training data.

To train the context generation model, we reformat the monolingual data so that the last sentence within a block of four is used as input to the encoder, while the three previous ones are predicted in a left-to-right manner. To reduce the distance between the input and the first predicted words, we also considered models that generated either sentences or tokens from right-to-left, but settled on the first approach. We use transformer base models to generate the missing context.

Model BLEU
Baseline 32.3 (31.8)
Concat 31.0 (30.8)
Random 31.8 (31.5)
Partial copy 31.6 (31.3)
Context generation 31.5 (31.4)

Table 2: EnDe BLEU scores, with parallel data only. Validation results in parentheses.
Model Deixis Lex. cohesion Ellipsis (infl.) Ellipsis (VP)
Baseline 50.0 (50.0) 45.9 (46.2) 52.0 25.0
Concat 83.4 (87.2) 48.9 (48.2) 76.0 73.8
Random 69.9 (68.8) 45.9 (46.4) 63.6 62.3
Partial copy 86.6 (85.1) 74.9 (75.4) 75.5 77.9
Context generation 85.6 (86.0) 60.0 (59.6) 74.8 74.2
Concat (Voita et al., 2019) 83.5 47.5 76.2 76.6
CADec (Voita et al., 2019) 81.6 58.1 72.2 80.0

Table 3: EnDe challenge set accuracy, with parallel data only. Validation results in parentheses.
Model Run BLEU Challenge sets
Random 1 31.9 61.5
2 32.0 60.2
3 31.5 60.0
Avg. 31.8 0.2 60.6 0.7
Partial copy 1 31.7 78.4
2 31.4 78.8
3 31.6 79.0
Avg. 31.6 0.1 78.7 0.2

Table 4: Robustness experiments across 3 runs with different data and model random seeds. Each challenge set is weighted equally.

5 Results

With parallel data only, EnRu BLEU scores are shown in Table 2. The concat model, where context is available for only a quarter of the data, trails the baseline by more than 1 BLEU. This model is also prone to severe overfitting as the training and test conditions differ significantly.

By completing the missing context in a variety of ways, the BLEU gap is mostly closed, although the baseline still performs better for EnRu. Using the same hyper-parameters, there is still some overfitting, but it is less severe than for the concat model. We note the using random contexts to fill in data is as effective as other methods, at least in terms of BLEU.

In the opposite direction (RuEn), the concat model still trails the baseline and overfits quickly, even with added regularization. With the partial copy heuristic, overall performance rises (39.4 BLEU), becoming very similar to the baseline model (39.2 BLEU).555Given the lack of RuEn challenge sets, and as we mainly use RuEn for back-translation, we did not train models with all the context completion techniques.

On the various challenge sets (Table 3), the concat models already shows gains over the context-agnostic baseline, in agreement with previous results reported by  Voita et al. (2018). However, the gains for lexical cohesion are minimal.

When missing context is replaced with random sentence pairs, performance drops across all tasks compared to the concat model, but remains above the baseline. As such, the model still benefits from data where real context is available (1.5M examples), but the its ability to capture cross-sentential phenomena is impeded by the additional examples.

With the partial copy heuristic, the model obtains an accuracy of 74.9% on the lexical cohesion test set, improving by 26% over the concat model with the original data. Compared to CADec Voita et al. (2019), a two-stage approach in which a baseline model is trained over the entire parallel corpus, and a post-editing model is built from the subset that has context, there is a gain of 15%. Over the other challenge sets, performance is comparable or slightly higher (-0.5% to 4.1% accuracy difference) than with unmodified training data.

Generating the missing context with separate models, accuracy on the lexical cohesion test set reaches 60.0%, an improvement over the original data, but still less than with the copying heuristic. Results on the other challenge sets are similar to those obtained with unaugmented parallel data.

Robustness

To characterize the models’ robustness to random fluctuations, we trained three distinct copies for both the random and partial copy context completion heuristics. Both the data generation and model parameter seeds were modified for each run.

Results are reported in Table 4. For both data completion techniques, BLEU scores differ by 0.5 or less across runs. Averaging results over the four challenge sets, performance variations between runs with similarly generated contexts remain within 2%, while there is almost a 20% difference between the two context imputation approaches.

Model BLEU Deixis Lex. cohesion Ellipsis (infl.) Ellipsis (VP)
Random 31.8 69.9 45.9 63.6 62.3
Partial copy - 2 31.6 86.6 74.9 75.5 77.9
Partial copy - 3 31.5 86.2 76.7 76.2 78.2
Full copy 30.93 85.4 81.3 75.6 76.8

Table 5: Progression from random to copied contexts.

5.1 From random to fully copied contexts

Using either random or partially copied contexts lead to similar BLEU scores, but very different results on the challenge sets. As such, we are interested in the progression from random contexts to exclusively copying the input when training models. Training examples with existing context (1.5M out of 6M), as well as all test data, remain unchanged.

As the input sentence pair is copied more extensively, BLEU starts to decrease, down to 30.9 when missing context is exclusively replaced by copies of the input sentence pair. This behaviour can likely be attributed to the increasing mismatch between training and inference conditions. Moreover, the model may be over-incentivized to copy the context at the expense of translating the source sentence.

As the input is copied more frequently, performance on the lexical cohesion challenge set improves, reaching 81.3% accuracy in the most extreme scenario. Nevertheless, as Russian proper nouns are declined, a copying mechanism may not capture all relevant phenomena, especially for words that are not split into multiple tokens.

On each of the other challenge sets, as long as missing context is copied at least once, performance remains within a tight 2% window. However, as previously remarked, using random contexts leads to much worse results on these test sets.

5.2 With monolingual data

With additional monolingual data, which is back-translated in context, and also using larger models, BLEU scores unsurprisingly improve (Table 7), reaching scores between 33.0 and 33.4. The context-aware models no longer trail behind the baseline, and the concat model now performs comparably to others. Note that, with bilingual and back-translated data mixed in a ratio close to 1, only approximately of the data has either missing or automatically generated target context, as opposed to with the original bilingual data.

Performance on all challenge sets increases significantly for all models, except for the context-less baseline (Table 6). With the original bilingual data and additional back-translated data (concat), accuracy reaches 86.1% or better on all four test sets. If missing contexts are instead replaced by random ones, performance drops on average by 3.65%. A model trained with the partial copy heuristic has the best performance on the lexical cohesion test set, at 90.1%, and other comparable performance to the concat model.

Figure 1: Lexical Cohesion validation accuracy over the training process. Orange line corresponds to the Concat model, the blue line corresponds to random context augmentation and the green line represents the partial copy heuristic.
Model Deixis Lex. cohesion Ellipsis (infl.) Ellipsis (VP)
Baseline 50.0 (50.0) 45.9 (46.2) 54.0 28.8
Concat 87.8 (88.8) 86.1 (85.6) 87.6 88.8
Random 85.4 (87.6) 82.5 (80.0) 82.8 85.2
Partial copy 89.6 (89.8) 90.1 (89.8) 86.6 88.6
Context generation 90.6 (90.8) 85.3 (83.2) 85.6 87.2
DocRepair (Voita et al., 2019) 91.8 80.6 86.4 75.2

Table 6: EnDe challenge set accuracy, with additional back-translated data. Validation results in parentheses.

Figure 1 displays the progress of lexical cohesion validation accuracy for a few models. The model trained with the copy heuristic, which achieves the overall best performance, also learns the fastest. If we instead use random contexts, final performance remains within 10%, while learning is much slower.

We also present results for the DocRepair model Voita et al. (2019), where the outputs of a sentence-to-sentence baseline are refined in context. The post-editing model is trained from round-trip translated monolingual data (without context), so that the output is contextually coherent, but not necessarily the input. Note that while we use monolingual data from the same corpus, and a similar amount of examples, our filtering procedures differ, so the training data is not exactly the same.

DocRepair obtains the best performance on the deixis test set, with a margin of 1.2% over our best model on this task. Results on the ellipsis (infl.) challenge sets are similar to those obtained with single-pass context-aware translation systems. However, these single-pass models perform up to 9.5% and 13.6% better on the lexical cohesion and VP ellipsis challenge sets respectively. In particular, while VP ellipsis may be a hard phenomena to capture with monolingual data only Voita et al. (2019), a context-aware translation system trained on sufficiently many examples may perform well on this task.

Model BLEU
Baseline 33.0 (33.1)
Concat 33.2 (33.1)
Random 33.4 (33.3)
Partial copy 33.4 (33.1)
Context generation 33.4 (33.2)

Table 7: EnDe BLEU scores, with additional back-translated data. Validation results in parentheses.

6 Conclusion

When document-level context is only partially available, we evaluate the effectiveness of various data completion techniques, using both BLEU and challenge sets targeting specific linguistic phenomena. In particular, a simple copy heuristic helps models achieve much better lexical cohesion, even for a highly inflected language such as Russian. Only adding random context sentence pairs however reduces a model’s ability to capture cross-sentence interactions, yet these effects are not visible from BLEU scores. Additionally, we confirm the effectiveness of back-translation on overall translation quality, while also demonstrating its usefulness on the four challenge sets.

As even simple context completion techniques have a clear impact on model performance, it may be worthwhile to explore additional approaches. In particular, to obtain more natural contexts, while limiting model generation errors, we can envision embedding sentences such that neighbours in vector space are probable contexts of each other. Moreover, it might be useful to understand how different data augmentation schemes interact with more sophisticated model architectures.

Acknowledgments

We would like to thank Kyunghyun Cho for insightful comments and discussions, the Google Translate and Google Brain teams for their useful inputs.

References

  • D. Bahdanau, K. Cho, and Y. Bengio (2015) Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations (ICLR), Cited by: §1, §2.1.
  • L. Barrault, O. Bojar, M. R. Costa-jussà, C. Federmann, M. Fishel, Y. Graham, B. Haddow, M. Huck, P. Koehn, S. Malmasi, C. Monz, M. Müller, S. Pal, M. Post, and M. Zampieri (2019) Findings of the 2019 conference on machine translation (WMT19). In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), Florence, Italy, pp. 1–61. External Links: Link Cited by: §1, §3.
  • R. Bawden, R. Sennrich, A. Birch, and B. Haddow (2018) Evaluating discourse phenomena in neural machine translation. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), Vol. 1, pp. 1304–1313. Cited by: §2.3.
  • I. Caswell, C. Chelba, and D. Grangier (2019a) Tagged back-translation. In Proceedings of the Fourth Conference on Machine Translation, Cited by: §1, §4.3.
  • I. Caswell, C. Chelba, and D. Grangier (2019b) Tagged back-translation. arXiv preprint arXiv:1906.06442. Cited by: §2.4.
  • S. Edunov, M. Ott, M. Auli, and D. Grangier (2018) Understanding back-translation at scale. In

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing

    ,
    pp. 489–500. Cited by: §2.4.
  • J. Gehring, M. Auli, D. Grangier, D. Yarats, and Y. N. Dauphin (2017) Convolutional sequence to sequence learning. In

    International Conference on Machine Learning

    ,
    pp. 1243–1252. Cited by: §2.1.
  • H. Hassan, A. Aue, C. Chen, V. Chowdhary, J. Clark, C. Federmann, X. Huang, M. Junczys-Dowmunt, W. Lewis, M. Li, et al. (2018) Achieving human parity on automatic chinese to english news translation. arXiv preprint arXiv:1803.05567. Cited by: §1.
  • T. He, X. Tan, Y. Xia, D. He, T. Qin, Z. Chen, and T. Liu (2018) Layer-wise coordination between encoder and decoder for neural machine translation. In Advances in Neural Information Processing Systems 31, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (Eds.), pp. 7944–7954. External Links: Link Cited by: §2.1.
  • V. C. D. Hoang, P. Koehn, G. Haffari, and T. Cohn (2018) Iterative back-translation for neural machine translation. In Proceedings of the 2nd Workshop on Neural Machine Translation and Generation, pp. 18–24. Cited by: §2.4.
  • S. Jean, S. Lauly, O. Firat, and K. Cho (2017) Does neural machine translation benefit from larger context?. arXiv preprint arXiv:1704.05135. Cited by: §2.2.
  • M. Junczys-Dowmunt (2019) Microsoft translator at wmt 2019: towards large-scale document-level neural machine translation. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), Florence, Italy, pp. 225–233. External Links: Link Cited by: §1, §1, §2.2, §3.
  • R. Kiros, Y. Zhu, R. R. Salakhutdinov, R. Zemel, R. Urtasun, A. Torralba, and S. Fidler (2015) Skip-thought vectors. In Advances in neural information processing systems, pp. 3294–3302. Cited by: §3.
  • T. Kudo and J. Richardson (2018) SentencePiece: a simple and language independent subword tokenizer and detokenizer for neural text processing. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pp. 66–71. Cited by: §4.1.
  • S. Läubli, R. Sennrich, and M. Volk (2018) Has machine translation achieved human parity? a case for document-level evaluation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 4791–4796. Cited by: §1.
  • P. Lison, J. Tiedemann, and M. Kouylekov (2018) Opensubtitles2018: statistical rescoring of sentence alignments in large, noisy parallel corpora. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC-2018), Cited by: §4.1.
  • E. Mansimov, A. Wang, and K. Cho (2019) A generalized framework of sequence generation with application to undirected sequence models. arXiv preprint arXiv:1905.12790. Cited by: §2.1.
  • S. Maruf and G. Haffari (2018) Document context neural machine translation with memory networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vol. 1, pp. 1275–1284. Cited by: §2.2.
  • L. Miculicich, D. Ram, N. Pappas, and J. Henderson (2018) Document-level neural machine translation with hierarchical attention networks. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 2947–2954. Cited by: §2.2.
  • M. Müller, A. Rios, E. Voita, and R. Sennrich (2018) A large-scale test set for the evaluation of context-aware pronoun translation in neural machine translation. arXiv preprint arXiv:1810.02268. Cited by: §2.3.
  • K. Papineni, S. Roukos, T. Ward, and W. Zhu (2002) BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pp. 311–318. Cited by: §2.3.
  • A. Popescu-Belis (2019) Context in neural machine translation: a review of models and evaluations. arXiv preprint arXiv:1901.09115. Cited by: §1.
  • R. Sennrich, B. Haddow, and A. Birch (2016) Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 86–96. Cited by: §1, §2.4.
  • J. Shen, P. Nguyen, Y. Wu, Z. Chen, et al. (2019) Lingvo: a modular and scalable framework for sequence-to-sequence modeling. External Links: 1902.08295 Cited by: §4.3.
  • I. Sutskever, O. Vinyals, and Q. V. Le (2014) Sequence to sequence learning with neural networks. NIPS. Cited by: §1, §2.1.
  • J. Tiedemann and Y. Scherrer (2017) Neural machine translation with extended context. In Proceedings of the Third Workshop on Discourse in Machine Translation, pp. 82–92. Cited by: §2.2.
  • Z. Tu, Y. Liu, S. Shi, and T. Zhang (2018) Learning to remember translation history with a continuous cache. Transactions of the Association of Computational Linguistics 6, pp. 407–420. Cited by: §2.2.
  • A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin (2017) Attention is all you need. In Advances in Neural Information Processing Systems, pp. 5998–6008. Cited by: §2.1, §4.3.
  • E. Voita, R. Sennrich, and I. Titov (2019) Context-aware monolingual repair for neural machine translation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, Hong Kong, China. Cited by: §2.2, §4.1, §5.2, §5.2, Table 6.
  • E. Voita, R. Sennrich, and I. Titov (2019) When a Good Translation is Wrong in Context: Context-Aware Machine Translation Improves on Deixis, Ellipsis, and Lexical Cohesion. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy. External Links: Link Cited by: §1, §1, §2.2, §3, §4.1, §4.1, §4.2, Table 3, §5.
  • E. Voita, P. Serdyukov, R. Sennrich, and I. Titov (2018) Context-aware neural machine translation learns anaphora resolution. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vol. 1, pp. 1264–1274. Cited by: §2.2, §5.
  • L. Wang, Z. Tu, A. Way, and Q. Liu (2017) Exploiting cross-sentence context for neural machine translation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 2826–2831. Cited by: §2.2.
  • J. Zhang, H. Luan, M. Sun, F. Zhai, J. Xu, M. Zhang, and Y. Liu (2018) Improving the transformer translation model with document-level context. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 533–542. Cited by: §2.2.