Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Repurposing Decoder-Transformer Language Models for Abstractive Summarization

T0 review · 4 major / 6 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 decoder-only transformer, finetuned with a language-modeling loss alone—no encoder, copy mechanism, coverage, or beam search—reaches ROUGE-1/2 scores competitive with specialized abstractive summarizers on CNN/DailyMail and XSum.

desk verdict A simple, credible demonstration that finetuned decoder-only LMs beat seq2seq baselines on CNN/DM and XSum without beam search or copy mechanisms, but the truncated-input evaluation and missing variance keep it a conditional result. read the letter →

arxiv 1909.00325 v1 pith:L62V7HUY submitted 2019-09-01 cs.CL cs.LG

Luke de Oliveira , Alfredo Láinez Rodrigo This is my paper · ORCID
classification cs.CLcs.LG
keywords abstractivesummarizationdecoder-onlytransformerlanguagemodelingtransferlearningnucleussamplingCNN/DailyMailXSumROUGE
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

This paper sets out to show that abstractive summarization does not require the standard toolbox of neural summarization—encoder–decoder architectures, pointer/copy mechanisms, coverage penalties, reinforcement learning on ROUGE, or beam search. The authors' procedure treats summarization as pure language modeling: each source–summary pair is flattened into a single token sequence with control tokens, and a decoder-only transformer is finetuned to maximize the likelihood of the whole sequence. Starting from a pretrained 117-million-parameter language model, this simple recipe reaches ROUGE-1 40.70 and ROUGE-2 18.03 on CNN/DailyMail with nucleus sampling, which is competitive with published specialized systems; on XSum it exceeds the reported topic-aware convolutional baselines. The authors interpret this as evidence that large-scale unsupervised pretraining already encodes much of what the hand-built components were designed to provide.

What carries the argument

The load-bearing object is the decoder-only transformer used as a pure conditional language model. The paper takes a masked self-attention transformer architecture and replaces the sequence-to-sequence setup with a single concatenated sequence: source tokens, then a summary-start token $\beta$, summary tokens, and an end token $\delta$. Positional encodings restart at zero for the summary, and separate segment embeddings mark source versus summary tokens. The model is trained by maximum likelihood on the whole sequence; no beam search is used at inference, only greedy decoding or nucleus sampling with a length-normalized likelihood selection. The pretrained 117M-parameter checkpoint provides the transfer-learning head start; ablations remove finetuning and segment encodings to show their contributions.

What would settle it

Take the CNN/DailyMail test articles and identify reference summaries whose key facts (named entities, numbers, events) appear only after token 400 of the source. Retrain or evaluate the same model with a longer context, say 1024 tokens matching the model's context size, and compare ROUGE. If the longer-context model materially improves on those articles, the reported ROUGE scores are partly an artifact of the truncation; if it does not, the 400-token cap is not the binding constraint claimed here.

Watch

Extended reading notes

Core claim

The central claim is that a decoder-only transformer, pretrained on open-domain text and finetuned on source–summary pairs with nothing but a next-token language-modeling loss, performs abstractive summarization at a level competitive with systems that use encoders, copy mechanisms, coverage, or reinforcement learning. The model maps each pair to a single sequence $S = [\alpha, x_0, \ldots, x_m, \beta, y_0, \ldots, y_k, \delta]$, where $\alpha$, $\beta$, $\delta$ mark source start, summary start, and end; positional encodings reset at $\beta$, and separate segment embeddings distinguish source tokens ($\sigma$) from summary tokens ($\tau$). Training maximizes the likelihood of the full sequence; inference uses either greedy decoding or nucleus sampling with $p = 0.3$ over five independent decodes, selecting the one with the lowest length-normalized negative log-likelihood. On CNN/DailyMail the best run scores ROUGE-1 40.70, ROUGE-2 18.03, and ROUGE-L 29.62; the ablation without the pretrained initialization drops to ROUGE-1 36.10, and removing the segment embeddings drops to 38.80. The authors acknowledge that ROUGE-L lags well behind the strong baselines and, from a manual review of the lowest-ROUGE-L outputs, attribute the gap to the model attending to details the reference summaries exclude.

Load-bearing premise

Source articles are cut to 400 tokens and summaries to 100 tokens at training time, so the method's reported success depends on the assumption that everything needed for the reference summaries fits in the first 400 tokens of each article.

Editorial extensions

If this is right

  • A plain language-modeling objective, without any summarization-specific inductive bias, is sufficient to reach competitive ROUGE-1/ROUGE-2 results, so the field's earlier hand-built components are not strictly necessary for news summarization.
  • Transfer learning from a large pretrained decoder-only transformer supplies most of the capability: the finetuning ablation loses 4.6 ROUGE-1 points, showing the pretrained weights are a load-bearing part of the recipe.
  • Beam search is not required: greedy decoding already gives ROUGE-1 39.12 on CNN/DailyMail, and nucleus sampling with length-normalized likelihood selection improves it.
  • The same finetuned setup transfers to a different dataset distribution (XSum) and outperforms the published topic-aware sequence-to-sequence baselines, suggesting the recipe generalizes across summarization styles.
  • The residual weakness is ROUGE-L: because LCS rewards uninterrupted matching, the model's tendency to omit or add details that the reference excludes caps this metric; explicitly learned masking is a suggested direction.

Reading between the lines

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

  • The implicit copy ability visible in the appendix examples suggests decoder-only language models acquire a soft copy mechanism from pretraining alone; a direct probe would be to measure how often generated summaries reproduce source n-grams versus true abstractions, across context lengths.
  • The 400-token truncation of CNN/DailyMail is likely the ceiling on the reported scores; with the 1024-token context of the architecture, the same recipe may do better on long articles, which is testable with the authors' procedure.
  • Because the selection among nucleus samples is by length-normalized likelihood, the method is essentially picking the most probable continuation; using a reference-based scorer or a quality estimator as the selector might push the scores higher without changing the model.
  • The same control-token flattening recipe should apply to other conditional generation tasks such as paraphrase, question generation, or style transfer, where the input–output gap is smaller than in summarization; if true, the paper's claim extends beyond summarization to transduction generally.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes treating abstractive summarization as a language-modeling problem with a decoder-only transformer. A pretrained GPT-2-style 117M parameter model is finetuned on source-summary pairs concatenated into a single sequence with control tokens, reset positional encodings, and segment-specific embeddings, trained with a plain language-modeling objective. At inference, summaries are produced by greedy decoding or by nucleus sampling with five independent decodings, choosing the candidate with the lowest length-normalized negative log-likelihood. The authors report ROUGE-1/ROUGE-2/ROUGE-L scores on CNN/DailyMail and XSum, ablate finetuning and segment encoding, and claim competitive performance without sequence-to-sequence architectures, copy or coverage mechanisms, beam search, or reinforcement learning.

Significance. If the central claim holds, the paper is a useful conceptual demonstration that a generic pretrained language model, with minimal task-specific inductive bias, can reach competitive abstractive summarization performance on standard benchmarks. The paper's strengths include its simple formulation, the ablations showing the contribution of finetuning and segment encoding, and the qualitative examples illustrating that the model can synthesize abstractive summaries rather than merely extract. The empirical support, however, is currently limited by three factors: the reported scores are single point estimates with no variance or significance testing; ROUGE-L on CNN/DailyMail is substantially below strong baselines; and the relationship between the 400-token source truncation and the evaluation protocol is unspecified. These issues do not disprove the central claim but they need to be addressed before the claim is fully supported.

major comments (4)
  1. [§3, Datasets] The paper states that source articles are limited to 400 tokens and summaries to 100 tokens 'at training time,' but it never states whether the same truncation is applied at evaluation time. Since the model has a 1024-token context and the 400-token source cutoff is therefore not forced by the architecture, the reported ROUGE scores could depend on the truncation assumption. Please specify exactly how source documents and references are handled at test time. If they are truncated, report coverage statistics such as the fraction of reference-summary n-grams that fall within the first 400 source tokens, and if possible show ROUGE as a function of source length. Without this, the comparison to baselines is not well defined.
  2. [§4, Tables 1–3] All ROUGE numbers are reported as single point estimates with no confidence intervals, variance across runs, or significance tests. Some differences from baselines are small (e.g., CNN-DM ROUGE-1 40.70 vs. 41.22 for Bottom-Up, ROUGE-2 18.03 vs. 18.68), and the advantage on XSum is not tested for significance. Additionally, on CNN-DM the ROUGE-L score (29.62 with nucleus sampling) is roughly seven to nine points below the strong baselines (36.38 for Pointer-Generator + Coverage, 38.34 for Bottom-Up). The paper acknowledges the ROUGE-L weakness, but the 'competitive performance' claim should be qualified to ROUGE-1/ROUGE-2, or supported by a statistical comparison that includes ROUGE-L.
  3. [§2, Eq. (7) and Abstract] The abstract and §5 state that the approach relies on 'no beam-search or other decoding-time optimization,' but Eq. (7) describes a decoding-time selection step that chooses among five nucleus-sampled summaries using the model's own length-normalized negative log-likelihood. This is not beam search, but it is a form of decoding-time optimization over generated candidates, and it could bias selection toward summaries that are likely under the model rather than maximally informative. Please adjust the claim or explicitly clarify that the selection procedure is a simple heuristic distinct from beam search, and discuss its potential selection bias.
  4. [§3, Models & Inference] The experimental description is not sufficient for reproduction. The paper does not report the number of transformer layers, dropout values, random seeds, early-stopping details, BPE vocabulary sizes, or the exact preprocessing applied to CNN-DM and XSum, and no code is released. Because the central claim is empirical, these details are load-bearing for verifying the results. Please provide a complete specification of the training and evaluation pipeline, ideally with code or a detailed hyperparameter table.
minor comments (6)
  1. [§4, Manual analysis] The manual analysis of fifty summaries in the bottom 5% of ROUGE-L is a useful qualitative check, but the authors themselves note it is 'not systematic and inconclusive statistically.' It would help to report inter-annotator agreement on the three traits and to explain how the fifty examples were sampled.
  2. [§1 and References] The reference to 'Radford et al., and Radford et al. (2019)' is incomplete in §1; the first entry appears to lack a publication venue or year. Please fix the citation.
  3. [Footnote 1] Footnote 1 contains a typo: 'top 30% of of the probability distribution' should read 'top 30% of the probability distribution.'
  4. [§2, Eqs. (6)–(7)] The indexing in Eqs. (6) and (7) is inconsistent with the definition of S in Eq. (4). If x = [x0,...,xm] and y = [y0,...,yk], the concatenated sequence S has more tokens than the product/sum ranges indicate. Please clarify the indexing or correct the equations so the loss and the decoding-score normalization are unambiguous.
  5. [§3, Datasets] The sentence describing XSum says the best model configuration from the ablation studies is used, but it is not stated whether this configuration is the finetuned model with nucleus sampling, the CNN-DM setup, or a separately tuned model. Please specify the exact configuration used for Table 3.
  6. [§4, Table 5] In Table 5, the example attributed to 'Croatia's milos raonic' appears to be a factual error, since Milos Raonic is Canadian. The example or its caption should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported ROUGE scores are empirical benchmark measurements, not quantities derived from fitted inputs, self-citations, or definitional equivalences.

full rationale

The central claim is an empirical one: fine-tuning a decoder-only transformer with a language-modeling loss yields competitive ROUGE-1/2 scores on CNN/DailyMail and XSum. Those scores are evaluated against fixed reference summaries (Lin, 2004) and compared with published baselines, so they are not derived by construction from any parameter fitted in the paper. The pretrained GPT-2 checkpoint (Radford et al., 2019) is an external transfer-learning source with no author overlap, and the paper's ablation of finetuning (Table 2) shows that transfer learning contributes empirically to the result rather than being assumed. The decoding-time selection in Eq. 7 ranks five nucleus-sampled summaries by the model's own length-normalized negative log-likelihood; this is a heuristic for choosing among candidate generations, not a fitted parameter and not a prediction derived from the target ROUGE score. The paper contains no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in via citation. The 400-token training-time truncation of source articles mentioned in Section 3 could raise a question about whether the evaluation condition matches the general summarization task, but that is a benchmark-validity concern, not circularity in the derivation chain. Under the required standard of exhibiting a specific reduction, no circular step can be identified.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's central result is empirical and relies on standard domain assumptions. No invented entities are introduced. The free parameters listed are decoding and training hyperparameters chosen by hand, not fitted to match the reported ROUGE scores.

free parameters (5)
  • nucleus sampling threshold p = 0.3
    Manual choice from Holtzman et al. (2019), used for decoding; no ablation reported. Affects all nucleus-sampling ROUGE scores in Tables 1 to 3.
  • number of independent decodings for nucleus sampling = 5
    Manual choice; best-of-N selection by lowest normalized negative log-likelihood (Eq. 7) improves ROUGE. No sensitivity analysis.
  • length normalization exponent = 0.6
    Borrowed from Wu et al. (2016), used in Eq. 7 scoring to avoid short summary bias; chosen by prior art, not tuned here.
  • training hyperparameters = learning rate 5e-5, batch size 10, context 1024
    Minimal tuning per Section 3; early stopping on validation loss. These affect convergence but not the central derivation.
  • truncation lengths = source 400 tokens, summary 100 tokens
    Section 3: source articles and summaries are truncated at training time; this is a preprocessing choice that determines what information the model can use.
assumptions (5)
  • domain assumption ROUGE F1 scores are a valid proxy for summarization quality in this comparison.
    Section 3 Evaluation uses ROUGE-1/2/L; Section 4 acknowledges known deficiencies but still relies on ROUGE for the central claim.
  • domain assumption Finetuning the pretrained 117M GPT-2 model from Radford et al. (2019) is a beneficial initialization for summarization.
    Section 3 finetunes this model; ablation in Table 2 shows finetuning helps, but the transfer assumption is imported from prior work.
  • domain assumption Concatenating source and summary into one sequence with control tokens, segment embeddings, and reset positional encodings is a valid way to perform transduction with a language model.
    Section 2, Eqs. 4 and 5; relies on Liu et al. (2018) and Devlin et al. (2018) design choices.
  • domain assumption Truncating source articles to 400 tokens and summaries to 100 tokens at training time preserves enough information for meaningful summaries.
    Section 3 Datasets; if salient information appears beyond 400 tokens, the model cannot include it, potentially lowering ROUGE.
  • domain assumption The language modeling loss over the full concatenated sequence (Eq. 6) is an appropriate training signal for summary generation.
    The model is trained to predict source and summary tokens alike; there is no task-specific loss that distinguishes summary quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Repurposing Decoder-Transformer Language Models for Abstractive Summarization." pith.science (2026). https://pith.science/paper/L62V7HUY

@misc{pith2026190900325,
  author       = {Pith},
  title        = {Pith review of: Repurposing Decoder-Transformer Language Models for Abstractive Summarization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L62V7HUY}},
  note         = {Machine review of arXiv:1909.00325}
}
read the original abstract

Neural network models have shown excellent fluency and performance when applied to abstractive summarization. Many approaches to neural abstractive summarization involve the introduction of significant inductive bias, exemplified through the use of components such as pointer-generator architectures, coverage, and partially extractive procedures, designed to mimic the process by which humans summarize documents. We show that it is possible to attain competitive performance by instead directly viewing summarization as a language modeling problem and effectively leveraging transfer learning. We introduce a simple procedure built upon decoder-transformers to obtain highly competitive ROUGE scores for summarization performance using a language modeling loss alone, with no beam-search or other decoding-time optimization, and instead relying on efficient nucleus sampling and greedy decoding.

Figures

Figures reproduced from arXiv: 1909.00325 by the authors.

Figure 1
Figure 1. Schematic of our Decoder-only Transformer [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 6 canonical work pages

  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]

    Asli Celikyilmaz, Antoine Bosselut, Xiaodong He, and Yejin Choi. 2018. Deep communicating agents for abstractive summarization. arXiv preprint arXiv:1803.10357

    arXiv 2018

  4. [4]

    Sumit Chopra, Michael Auli, Alexander M Rush, and SEAS Harvard. 2016. Abstractive sentence summarization with attentive recurrent neural networks. Proceedings of NAACL-HLT16, pages 93--98

    work page 2016

  5. [5]

    Andrew M Dai and Quoc V Le. 2015. Semi-supervised sequence learning. In Advances in neural information processing systems, pages 3079--3087

    2015

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. CoRR, abs/1810.04805

    arXiv 2018

  7. [7]

    Sebastian Gehrmann, Yuntian Deng, and Alexander Rush. 2018. Bottom-up abstractive summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4098--4109

    2018

  8. [8]

    Jiatao Gu, Zhengdong Lu, Hang Li, and Victor OK Li. 2016. Incorporating copying mechanism in sequence-to-sequence learning. arXiv preprint arXiv:1603.06393

    arXiv 2016

Show all 33 references
  1. [9]

    Ari Holtzman , Jan Buys , Maxwell Forbes , and Yejin Choi . 2019. http://arxiv.org/abs/1904.09751 The Curious Case of Neural Text Degeneration . arXiv e-prints, page arXiv:1904.09751

    2019 arXiv

  2. [10]

    Jeremy Howard and Sebastian Ruder. 2018. Universal language model fine-tuning for text classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 328--339

    2018

  3. [11]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  4. [12]

    Taku Kudo and John Richardson. 2018. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226

    2018 arXiv

  5. [13]

    Jimmy Lei Ba , Jamie Ryan Kiros , and Geoffrey E. Hinton . 2016. http://arxiv.org/abs/1607.06450 Layer Normalization . arXiv e-prints, page arXiv:1607.06450

    2016 arXiv

  6. [14]

    Piji Li, Lidong Bing, and Wai Lam. 2018. Actor-critic based training framework for abstractive summarization. arXiv preprint arXiv:1803.11070

    2018 arXiv

  7. [15]

    Chin-Yew Lin. 2004. https://www.aclweb.org/anthology/W04-1013 ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out: Proceedings of the ACL -04 Workshop , pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  8. [16]

    Liu, Mohammad Ahmad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer

    Peter J. Liu, Mohammad Ahmad Saleh, Etienne Pot, Ben Goodrich, Ryan Sepassi, Lukasz Kaiser, and Noam Shazeer. 2018. https://openreview.net/pdf?id=Hyg0vbWC- Generating wikipedia by summarizing long sequences

    2018

  9. [17]

    Elena Lloret, Laura Plaza, and Ahmet Aker. 2018. https://doi.org/10.1007/s10579-017-9399-2 The challenging task of summary evaluation: an overview . Language Resources and Evaluation, 52(1):101--148

    2018 doi

  10. [18]

    Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. arXiv preprint arXiv:1602.06023

    2016 arXiv

  11. [19]

    Cohen, and Mirella Lapata

    Shashi Narayan, Shay B. Cohen, and Mirella Lapata. 2018. Don't give me the details, just the summary! T opic-aware convolutional neural networks for extreme summarization. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium

    2018

  12. [20]

    Romain Paulus, Caiming Xiong, and Richard Socher. 2017. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304

    2017 arXiv

  13. [21]

    Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. 2018. Deep contextualized word representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: H...

    2018

  14. [22]

    Improving language understanding by generative pre-training

    Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training

  15. [23]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners

    2019

  16. [24]

    Alexander M Rush, Sumit Chopra, and Jason Weston. 2015. A neural attention model for abstractive sentence summarization. arXiv preprint arXiv:1509.00685

    2015 arXiv

  17. [25]

    Natalie Schluter. 2017. The limits of automatic summarisation according to rouge. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 41--45

    2017

  18. [26]

    Abigail See, Peter J Liu, and Christopher D Manning. 2017. Get to the point: Summarization with pointer-generator networks. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1073--1083

    2017

  19. [27]

    Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909

    2015 arXiv

  20. [28]

    Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104--3112

    2014

  21. [29]

    Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu, and Hang Li. 2016. Modeling coverage for neural machine translation. arXiv preprint arXiv:1601.04811

    2016 arXiv

  22. [30]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

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

    2017

  23. [31]

    Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly. 2015. Pointer networks. In Advances in Neural Information Processing Systems, pages 2692--2700

    2015

  24. [32]

    Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google's neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144

    2016 arXiv

  25. [33]

    Wenyuan Zeng, Wenjie Luo, Sanja Fidler, and Raquel Urtasun. 2016. Efficient summarization with read-again and copy mechanism. arXiv preprint arXiv:1611.03382

    2016 arXiv

Pith tools

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