One Model to Learn Both: Zero Pronoun Prediction and Translation

Longyue Wang; Shuming Shi; Xing Wang; Zhaopeng Tu

REVIEW 3 major objections 5 minor 36 references

One Model to Learn Both: Zero Pronoun Prediction and Translation

T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A single neural translation model that predicts zero pronouns as sequence labels and folds in discourse context raises Chinese-English BLEU from 31.8 to 37.1 and zero-pronoun F1 from 0.66 to 0.77, with similar gains on Japanese-English.

desk verdict Solid Chinese-English joint model for zero pronoun prediction and translation, but the Japanese-English universality claim needs more evidence before it can be taken at face value. read the letter →

arxiv 1909.00369 v1 pith:FQLOMSCL submitted 2019-09-01 cs.CL

Longyue Wang , Zhaopeng Tu , Xing Wang , Shuming Shi This is my paper · ORCID

classification cs.CL

keywords zeropronounspro-droplanguagesmachinetranslationjointlearningdiscoursecontexthierarchicalneuralnetworksChinese-EnglishJapanese-English

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Zero pronouns dropped in Chinese and Japanese often must be made explicit in English, and machine translation systems routinely miss or mistranslate them. This paper argues that instead of calling a separate zero-pronoun predictor before translation, one neural model should predict and translate them together, with discourse context from previous sentences folded in. On Chinese-English data the joint model raises BLEU from a 31.80 baseline to 36.04 and, with hierarchical discourse context, to 37.11, while zero-pronoun prediction F1 rises from 0.66 to 0.77; Japanese-English also improves from 19.94 to 22.00. If the approach holds, translation from pro-drop languages becomes simpler and more accurate without an external pronoun-prediction stage.

What carries the argument

The carrying mechanism is the encoder-decoder-reconstructor with a sequence-labeling ZP predictor attached to the reconstructor. The reconstructor reads encoder and decoder states through two interactive attention mechanisms; its hidden states are the inputs to a labeler that predicts, for each source position, no pronoun or a specific dropped pronoun. The training objective sums translation likelihood, source reconstruction likelihood, and ZP-labeling likelihood, so the auxiliary loss pushes the shared representations to retain pronoun information. Discourse context is supplied by a two-layer hierarchical encoder over previous sentences—word-level encoder shared with the NMT encoder, then sentence-level encoder—producing a context vector concatenated to the reconstructor states. This is the load-bearing point: context enters exactly where ZP labels are predicted, not at the decoder.

What would settle it

Take a random sample of source sentences from the Japanese-English test set, have humans mark every dropped pronoun, and compare with the automatic alignment-based labels used to train the joint model. If automatic Japanese annotation accuracy is far below the reported Chinese level, or if retraining on human-verified Japanese labels does not reproduce the +2.06 BLEU gain, the universality claim would be weakened.

Watch

Extended reading notes

Core claim

The central claim is that zero-pronoun prediction and translation are mutually reinforcing and should be learned as one task. The model casts zero-pronoun prediction as sequence labeling over source words, with labels supervised by automatically annotated ZPs, and trains this labeler jointly with an encoder-decoder-reconstructor translation model. Because the labeling loss shapes the hidden representations during training, decoding needs no external ZP predictor and no annotated source input. Adding a hierarchical encoder over the previous three sentences, whose summary is concatenated into the reconstructor state, further improves both translation and prediction; the paper reports 37.11 BLEU and 0.77 F1 on Chinese-English, beating the best external-prediction pipeline, and 22.00 BLEU on Japanese-English.

Load-bearing premise

The training labels for the zero-pronoun prediction component come from an automatic alignment-based annotation method reported to exceed 90% accuracy on Chinese, but the paper does not measure annotation accuracy on its Japanese data, where alignment is harder, so noisy Japanese labels could overstate the universality and auxiliary-loss gains.

Editorial extensions

If this is right

A production translation system for Chinese-to-English can drop the separate zero-pronoun prediction step entirely, avoiding error propagation and extra decoding cost.
Discourse context should be routed to the component that predicts missing pronouns, because feeding it directly to the decoder hurt translation in the paper's comparisons.
Zero-pronoun F1 and BLEU improve together, so ZP prediction accuracy can serve as a useful proxy signal for translation quality in pro-drop to non-pro-drop settings.
The same architecture transfers to Japanese-English, indicating the approach is not Chinese-specific, though Japanese gains are smaller.
Errors from subjective ZPs, especially those depending on speaker intention in imperatives, remain the hardest and point to where further context modeling would pay off.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The pattern of results suggests a general design rule for recovering omitted information in MT: attach a prediction head to the representation layer where the missing token would be reconstructed, and feed discourse context into that head rather than into the target-side decoder.
The same sequence-labeling auxiliary loss could be applied to other gaps between source and target, such as dropped articles, tense markers, or other unaligned words, since the reconstructor already aligns source and target representations; the paper lists this direction but does not test it.
Because gains are largest when discourse context is available, jointly trained ZP prediction may serve as a diagnostic for how much discourse a translation model actually uses: a model with high ZP F1 but no BLEU gain would indicate the predicted pronouns are not being realized in the output.
The lower Japanese gains are consistent with noisier automatic labels, so a self-training or unsupervised variant could test whether annotation quality, rather than language structure, is the bottleneck.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this paper. The first is that the Chinese-English results are real and the joint modeling idea is a legitimate step beyond the authors' own prior pipeline work. The second is that the Japanese-English half of the paper is much weaker than the abstract implies: it reports only BLEU, no ZP prediction metrics, and the only supervision for Japanese ZP labels comes from an alignment-based auto-annotation pipeline whose Japanese accuracy is never measured—the paper itself concedes that Japanese SOV structure makes that annotation harder.

What is actually new: earlier joint work (Wang et al. 2018b) still needed an external model to predict ZP positions. Here ZP prediction is a sequence labeling task trained end-to-end with the translation model via the reconstructor, so at decode time you no longer need a separately trained ZP predictor. Adding hierarchical encoding of previous sentences into the reconstructor is a sensible extension. The paper also does some things right on evaluation: it compares against a strong baseline and two prior methods, runs a sign test on BLEU, reports parameter counts to argue the gains are not just capacity, and includes a manual error analysis on 500 sentences showing the improvements come from fixing subjective, objective, and dummy pronoun errors. That last part is credible and useful.

The soft spots are real but concentrated. The Japanese-English experiment is the load-bearing one for the universality claim, and it is under-analyzed. No P/R/F1 for Japanese ZP prediction, no manually annotated Japanese test set, and no validation of auto-annotation quality on Japanese. If those labels are noisy, the auxiliary loss is fitting corrupted targets and the observed +2.06 BLEU could be regularization or added capacity rather than genuine ZP prediction. That gap is not fatal for the Chinese-English contribution, but it means the abstract overstates the evidence. No code or data release is also a problem for reproducibility; the Chinese data link points to prior work but the Japanese data split is described only by counts. F1 differences lack significance tests, though the gap is large. None of this makes the central Chinese-English result doubtful.

Verdict: this is a solid paper for the NMT subfield, worth a serious referee and probably worth citing for the joint-model design. It needs a revision that either adds Japanese ZP prediction evaluation and label-quality checks or softens the universality claim. I'd send it to review.

Referee Report

3 major / 5 minor

Summary. The paper proposes a unified neural model that jointly learns zero pronoun (ZP) prediction and translation. The model casts ZP prediction as sequence labeling, adds an auxiliary reconstruction and labeling loss to the translation objective, and further incorporates a hierarchical discourse encoder over previous source sentences. On Chinese–English subtitle data, the joint model improves BLEU over the baseline and over previous external-ZP-based systems, and improves ZP prediction F1 from 0.66 to 0.70/0.77; discourse-level context adds further gains. On Japanese–English data, only translation BLEU is reported. The paper claims that the approach improves both translation and ZP prediction on both language pairs and reduces reliance on external ZP predictors.

Significance. If the reported results hold, this is a useful advance for translating pro-drop languages: it replaces external ZP prediction modules with an end-to-end joint model, reduces parameter overhead relative to the prior reconstruction-based approach, and provides manual error analysis showing that translation gains come from fixing ZP-related errors. The comparisons to baseline and prior systems, the significance tests for Chinese translation, and the fixed/new error counts are concrete strengths. The main weakness is that the universality claim currently rests on a single language pair with no direct ZP prediction evaluation, and the code is not released, which limits reproducibility.

major comments (3)

[§4.3, Table 3] The Japanese–English experiment reports translation BLEU only; it does not report zero-pronoun prediction precision, recall, or F1, nor a manually annotated Japanese ZP test set. Because the abstract and the introduction claim that the approach improves 'both translation performance and ZP prediction accuracy' on 'both Chinese–English and Japanese–English', the Japanese support for the ZP-prediction half of the claim is missing. The observed +2.06 BLEU could in principle come from the added reconstructor and hierarchical encoder as a regularizer or capacity increase rather than from genuine ZP prediction; please add Japanese ZP prediction evaluation or explicitly restrict the universality claim to translation quality.
[§2.2 and §4.3, Eq. (5)] The auxiliary ZP labeling loss in Eq. (5) is supervised by the alignment-based auto-annotation pipeline. The paper cites above-90% accuracy for the Chinese pipeline (Section 2.2) but does not measure annotation quality on the Japanese Opensubtitle data, and Section 4.3 states that Japanese SOV structure 'poses difficulties for ZP annotation via alignment method'. Without a Japanese label-quality estimate or a small gold set, the training signal for the Japanese ZP component is unverified, which weakens the claimed universality of the joint-learning benefit.
[§4.2, Table 2] The paper reports ZP prediction F1 improvements (0.66 to 0.70 and 0.77) as part of the central 'both tasks' claim, but no significance test or confidence interval is provided for these F1 differences, despite significance testing being used for BLEU. Please report a significance test (e.g., paired bootstrap or McNemar on the label sequences) or state explicitly if the test set is too small for such a comparison.

minor comments (5)

[§4.1] The setup states that a sign-test is used for statistical significance, but Table 3, whose caption says the improvements are 'significant', contains no significance markers or p-values. Please mark the significant differences or state that they were not tested.
[§3.1, Eq. (5)] The parameter ψ appears in the objective but is never defined; the text defines only θ and γ. Please clarify whether ψ is intended to denote the labeler parameters or remove it.
[§4.4, Table 6] The model names in the table header contain formatting artifacts ('BASE .', 'EXTE .', 'JOIN .', '+DIS.'); please use the full model names or consistent abbreviations.
[§4.4] The sentence 'Our best model variation outperform that of external ZP prediction by over 2 BLEU points' should be 'outperforms'.
[§1, Table 1] The bracketed repeated pronouns in the input examples, e.g. '(我我我)', are visually confusing; a note that the repeated forms indicate a single omitted pronoun would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the joint objective optimizes translation likelihood, reconstruction, and ZP labeling, while the reported BLEU and F1 are measured on independent test sets.

full rationale

The claimed derivation chain is not circular. The training objective in Eq. (5) is a sum of three terms: the translation log-likelihood, the reconstruction score, and the ZP labeling probability. None of these terms encodes BLEU or F1, so the reported improvements are not forced by construction. The ZP supervision labels come from an automatic alignment-based annotation pipeline (Wang et al., 2016), but the paper evaluates ZP prediction on a manually annotated test set and reports P/R/F1 in Table 2, which is an external, human-defined metric. Translation quality is measured with BLEU against reference translations, also independent of the training objective. The comparisons against Wang et al. (2018a) are self-citations, but they serve as datasets, baselines, and architectural priors rather than as the justification of the central result; the central claim that joint training with a reconstructor and discourse context improves both tasks is empirically supported by independent metrics and manual error analysis in Section 4.4. No equation in the paper reduces a reported prediction to a fitted parameter or to an earlier self-citation, and no uniqueness theorem or ansatz is imported from the authors' prior work to forbid alternatives. The Japanese-English result is weaker because no ZP F1 is reported there, but that is an empirical-evidence limitation, not a circularity: the BLEU gain on Japanese is still computed against held-out references and is not a restatement of the training loss. Therefore the appropriate finding is no significant circularity, score 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim leans on three unverified inputs: auto-annotated ZP labels, manual test annotations, and the discourse window and hidden-size choices. The model itself introduces no new theoretical entities. The main under-supported number is the annotation accuracy of the supervision signal, especially for Japanese.

free parameters (3)

Discourse context window K = 3
The model encodes the previous 3 sentences as context, following Wang et al. (2017) and Tu et al. (2018), with no sweep or sensitivity analysis in this paper. This choice affects how much discourse information the ZP predictor sees.
Reconstructor hidden size = 1000
Set to 1000, while the external baseline uses 2000 and 1000 for two reconstructors. Parameter count comparisons use this value, so it is load-bearing for the claim that improvements are not due to more parameters.
Source and target vocabulary sizes = 30K for Chinese-English, 20K for Japanese-English
Preprocessing choices restrict the training distribution and could affect the comparison, but they are standard and not fitted to the test metric.

assumptions (3)

domain assumption Auto-annotated ZP labels are accurate enough to supervise joint learning.
Section 2.2 relies on Wang et al. (2016) auto-annotation accuracy above 90%; the paper does not re-measure it, especially for Japanese. If labels are noisy, the auxiliary loss is mis-specified.
domain assumption The manually annotated test set and the 30-pronoun vocabulary fully cover the zero pronouns that matter for translation.
Evaluation F1 and the manual error analysis depend on this coverage. The paper does not report inter-annotator agreement.
standard math The BLEU sign test with p<0.01 is a valid significance test for translation improvements.
Used in Section 4.1 to claim significance; if the test statistic assumptions are violated, the significance claim is weakened.

how reviews work

0 comments

Cite this review

Pith. "Pith review of One Model to Learn Both: Zero Pronoun Prediction and Translation." pith.science (2026). https://pith.science/paper/FQLOMSCL

@misc{pith2026190900369,
  author       = {Pith},
  title        = {Pith review of: One Model to Learn Both: Zero Pronoun Prediction and Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FQLOMSCL}},
  note         = {Machine review of arXiv:1909.00369}
}

read the original abstract

Zero pronouns (ZPs) are frequently omitted in pro-drop languages, but should be recalled in non-pro-drop languages. This discourse phenomenon poses a significant challenge for machine translation (MT) when translating texts from pro-drop to non-pro-drop languages. In this paper, we propose a unified and discourse-aware ZP translation approach for neural MT models. Specifically, we jointly learn to predict and translate ZPs in an end-to-end manner, allowing both components to interact with each other. In addition, we employ hierarchical neural networks to exploit discourse-level context, which is beneficial for ZP prediction and thus translation. Experimental results on both Chinese-English and Japanese-English data show that our approach significantly and accumulatively improves both translation performance and ZP prediction accuracy over not only baseline but also previous works using external ZP prediction models. Extensive analyses confirm that the performance improvement comes from the alleviation of different kinds of errors especially caused by subjective ZPs.

Figures

Figures reproduced from arXiv: 1909.00369 by the authors.

**Figure 2.** Architecture of hierarchical neural encoder. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages

[1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

Rachel Bawden, Rico Sennrich, Alexandra Birch, and Barry Haddow. 2018. Evaluating discourse phenomena in neural machine translation. In NAACL

work page 2018
[4]

Chen Chen and Vincent Ng. 2013. Chinese zero pronoun resolution: Some recent advances. In EMNLP

work page 2013
[5]

Kyunghyun Cho, Bart Van Merri \"e nboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP

work page 2014
[6]

Tagyoung Chung and Daniel Gildea. 2010. Effects of empty categories on machine translation. In EMNLP

work page 2010
[7]

Michael Collins, Philipp Koehn, and Ivona Kucerova. 2005. Clause restructuring for statistical machine translation. In ACL

work page 2005
[8]

Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. 2017. Convolutional sequence to sequence learning. In ICML

work page 2017

Show all 36 references

[9]

Geoffrey E Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan R Salakhutdinov. 2012. Improving neural networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580

2012 arXiv
[10]

Sebastien Jean, Stanislas Lauly, Orhan Firat, and Kyunghyun Cho. 2017 a . Does neural machine translation benefit from larger context? arXiv preprint arXiv:1704.05135

2017 arXiv
[11]

Sebastien Jean, Stanislas Lauly, Orhan Firat, and Kyunghyun Cho. 2017 b . Neural machine translation for cross-lingual pronoun prediction. In Workshop on Discourse in Machine Translation

2017
[12]

Fang Kong and Guodong Zhou. 2010. A tree kernel-based unified framework for chinese zero anaphora resolution. In EMNLP

2010
[13]

Ronan Le Nagard and Philipp Koehn. 2010. Aiding pronoun translation with co-reference resolution. In WMT-MetricsMATR

2010
[14]

Charles N Li and SA Thomson. 1979. Third-person pronouns and zero-anaphora in chinese discourse in discourse and syntax. Syntax and Semantics, 12

1979
[15]

Yang Liu and Qun Liu. 2010. Joint parsing and translation. In COLING

2010
[16]

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. BLEU : A method for automatic evaluation of machine translation. In ACL

2002
[17]

Arum Park, Seunghee Lim, and Munpyo Hong. 2015. Zero object resolution in korean. In PACLIC

2015
[18]

Sudha Rao, Allyson Ettinger, Hal Daum \'e III, and Philip Resnik. 2015. Dialogue focus tracking for zero pronoun resolution. In NAACL

2015
[19]

Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau

Iulian V. Serban, Alessandro Sordoni, Yoshua Bengio, Aaron Courville, and Joelle Pineau. 2016. Building end-to-end dialogue systems using generative hierarchical neural network models. In AAAI

2016
[20]

Alessandro Sordoni, Yoshua Bengio, Hossein Vahabi, Christina Lioma, Jakob Grue Simonsen, and Jian - Yun Nie. 2015. A hierarchical recurrent encoder-decoder for generative context-aware query suggestion. In CIKM

2015
[21]

Hirotoshi Taira, Katsuhito Sudoh, and Masaaki Nagata. 2012. Zero pronoun resolution can improve the quality of J-E translation. In Workshop on Syntax, Semantics and Structure in Statistical Translation

2012
[22]

Shunsuke Takeno, Masaaki Nagata, and Kazuhide Yamamoto. 2017. Controlling target features in neural machine translation via prefix constraints. In The 4th Workshop on Asian Translation

2017
[23]

J \"o rg Tiedemann. 2012. Parallel data, tools and interfaces in opus. In LREC

2012
[24]

Zhaopeng Tu , Yang Liu , Lifeng Shang , Xiaohua Liu , and Hang Li . 2017. Neural machine translation with reconstruction. In AAAI

2017
[25]

Zhaopeng Tu, Yang Liu, Shuming Shi, and Tong Zhang. 2018. Learning to remember translation history with a continuous cache. TACL

2018
[26]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In NIPS

2017
[27]

Elena Voita, Pavel Serdyukov, Rico Sennrich, and Ivan Titov. 2018. Context-aware neural machine translation learns anaphora resolution. In ACL

2018
[28]

Longyue Wang, Zhaopeng Tu, Shuming Shi, Tong Zhang, Yvette Graham, and Qun Liu. 2018 a . Translating pro-drop languages with reconstruction models . In AAAI

2018
[29]

Longyue Wang, Zhaopeng Tu, Andy Way, and Qun Liu. 2017. Exploiting cross-sentence context for neural machine translation. In EMNLP

2017
[30]

Longyue Wang, Zhaopeng Tu, Andy Way, and Qun Liu. 2018 b . Learning to jointly translate and predict dropped pronouns with a shared reconstruction mechanism. In EMNLP

2018
[31]

Longyue Wang, Zhaopeng Tu, Xiaojun Zhang, Hang Li, Andy Way, and Qun Liu. 2016. A novel approach for dropped pronoun translation. In NAACL

2016
[32]

Bing Xiang, Xiaoqiang Luo, and Bowen Zhou. 2013. Enlisting the ghost: Modeling empty categories for machine translation. In ACL

2013
[33]

Xinyan Xiao, Yang Liu, Young-Sook Hwang, Qun Liu, and Shouxun Lin. 2010. Joint tokenization and translation. In COLING

2010
[34]

Qingyu Yin, Yu Zhang, Weinan Zhang, and Ting Liu. 2017. Chinese zero pronoun resolution with deep memory network. In EMNLP

2017
[35]

Matthew D Zeiler. 2012. ADADELTA : an adaptive learning rate method. arXiv preprint arXiv:1212.5701

2012 arXiv
[36]

Shanheng Zhao and Hwee Tou Ng. 2007. Identification and resolution of chinese zero pronouns: A machine learning approach. In EMNLP

2007

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

Rachel Bawden, Rico Sennrich, Alexandra Birch, and Barry Haddow. 2018. Evaluating discourse phenomena in neural machine translation. In NAACL

work page 2018

[4] [4]

Chen Chen and Vincent Ng. 2013. Chinese zero pronoun resolution: Some recent advances. In EMNLP

work page 2013

[5] [5]

Kyunghyun Cho, Bart Van Merri \"e nboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In EMNLP

work page 2014

[6] [6]

Tagyoung Chung and Daniel Gildea. 2010. Effects of empty categories on machine translation. In EMNLP

work page 2010

[7] [7]

Michael Collins, Philipp Koehn, and Ivona Kucerova. 2005. Clause restructuring for statistical machine translation. In ACL

work page 2005

[8] [8]

Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. 2017. Convolutional sequence to sequence learning. In ICML

work page 2017