Elamur Research⌕
☰
home / papers / 2018 / Multi-Reward Reinforced Summarization with Saliency and Entailment

Multi-Reward Reinforced Summarization with Saliency and Entailment

Summary

This short NAACL 2018 paper argues that a good abstractive summary needs more than n-gram overlap with a reference. It should capture the important content of the source and should not state anything the source does not support. The authors start from a pointer-copy LSTM encoder-decoder with attention and coverage (following See et al., 2017). They fine-tune it with REINFORCE, using a self-critical (SCST) baseline and a mixed cross-entropy/RL objective. They introduce two new rewards:

  • ROUGESal reweights ROUGE matches using token-level saliency probabilities from a tagger trained on SQuAD answer spans.
  • Entail averages the entailment probability of each generated sentence given the reference summary, using a classifier trained on SNLI and Multi-NLI. It is scaled by relative summary length so that very short outputs are not rewarded.

Rather than tuning a weighted sum of rewards, the authors optimize different rewards on alternating mini-batches, an approach borrowed from multi-task learning. The ROUGESal+Entail combination achieved new state-of-the-art scores on CNN/Daily Mail. It also transferred well to DUC-2002 without any retraining. This suggests that explicit saliency and entailment signals make summarizers more robust.

Key points

  • Best results on CNN/Daily Mail: the ROUGESal+Entail model reached ROUGE-1 40.43 and ROUGE-2 18.00. The cross-entropy baseline scored 39.41 / 17.33, and See et al. (2017) reported 39.53 / 17.28.
  • Test-only DUC-2002 transfer: the combined model scored R-1 38.95, R-2 17.05, R-L 35.52 and METEOR 16.47. The baseline scored 35.50 / 14.57 / 32.19 / 14.36. All four differences are significant at p < 0.001.
  • Choice of entailment premise: the reference summary worked better as the premise than the full document. Because the reference is itself entailed by the source, the authors argue this choice is justified by transitivity.
  • Saliency and entailment analysis:
    • Under the SQuAD-based scorer, ROUGESal outputs scored 30.86% salient tokens versus about 28% for the other systems.
    • On the CNN/Daily Mail Cloze setup, ROUGESal outputs scored 64.66%.
    • Entail-reward outputs had the highest entailment score, at 28.98%.
  • Abstractiveness: all RL-trained models produced more novel n-grams than the cross-entropy baseline. ROUGESal copied somewhat more than the other RL models, likely because it favors salient items such as names.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Cited by

90 papers of our library cite it (of the 140884 whose bibliographies we read); 10 of them have a page here, 80 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n18-2102

@inproceedings{pasunuru2018multi,
  title = {{Multi-Reward Reinforced Summarization with Saliency and Entailment}},
  author = {Pasunuru, Ramakanth and Bansal, Mohit},
  booktitle = {Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers)},
  year = {2018},
  doi = {10.18653/v1/n18-2102},
  url = {https://doi.org/10.18653/v1/n18-2102}
}

References

32 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Anand Gupta, Manpreet Kaur, Adarsh Singh, Aseem Goel, and Shachar Mirkin. 2014. Text summarization through entailment-based minimum vertex cover. Lexical and Computational Semantics (* SEM 2014) page 75. find
  2. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. In ICLR. find our card
  3. Samuel R Bowman, Gabor Angeli, Christopher Potts, and Christopher D Manning. 2015. A large annotated corpus for learning natural language inference. In EMNLP. find
  4. Sanda Harabagiu and Andrew Hickl. 2006. Methods for using textual entailment in open-domain question answering. In ACL. pages 905–912. find
  5. Karl Moritz Hermann, Tomas Kocisky, Edward Grefenstette, Lasse Espeholt, Will Kay, Mustafa Suleyman, and Phil Blunsom. 2015. Teaching machines to read and comprehend. In NIPS. pages 1693–1701. find our card
  6. Qian Chen, Xiaodan Zhu, Zhenhua Ling, Si Wei, and Hui Jiang. 2016. Distraction-based neural networks for modeling documents. In IJCAI. find
  7. Jackie Chi Kit Cheung and Gerald Penn. 2014. Unsupervised sentence enhancement for automatic summarization. In EMNLP. pages 775–786. find
  8. Masaru Isonuma, Toru Fujino, Junichiro Mori, Yutaka Matsuo, and Ichiro Sakata. 2017. Extractive summarization using multi-task learning with document classification. In EMNLP. pages 2091–2100. find
  9. Sumit Chopra, Michael Auli, and Alexander M Rush. 2016. Abstractive sentence summarization with attentive recurrent neural networks. In HLT-NAACL. find our card
  10. Sergio Jimenez, George Duenas, Julia Baquero, Alexander Gelbukh, Av Juan Dios Bátiz, and Av Mendizábal. 2014. UNAL-NLP: Combining soft cardinality features for semantic textual similarity, relatedness and entailment. In In SemEval. pages 732–742. find
  11. James Clarke and Mirella Lapata. 2008. Global inference for sentence compression: An integer linear programming approach. Journal of Artificial Intelligence Research 31:399–429. find
  12. Hongyan Jing. 2000. Sentence reduction for automatic text summarization. In ANLP. find
  13. Ido Dagan, Oren Glickman, and Bernardo Magnini. 2006. The pascal recognising textual entailment challenge. In Machine learning challenges. evaluating predictive uncertainty, visual object classification, and recognising tectual entailment, Springer, pages 177–190. find
  14. Diederik Kingma and Jimmy Ba. 2015. Adam: A method for stochastic optimization. In ICLR. find
  15. Kevin Knight and Daniel Marcu. 2002. Summarization beyond sentence extraction: A probabilistic approach to sentence compression. Artificial Intelligence 139(1):91–107. find
  16. Michael Denkowski and Alon Lavie. 2014. Meteor universal: Language specific translation evaluation for any target language. In EACL. find our card
  17. Alice Lai and Julia Hockenmaier. 2014. Illinois-LH: A denotational and distributional approach to semantics. Proc. SemEval 2:5. find
  18. Shibhansh Dohare and Harish Karnick. 2017. Text summarization using abstract meaning representation. arXiv preprint arXiv:1706.01678 . find
  19. Bradley Efron and Robert J Tibshirani. 1994. An introduction to the bootstrap. CRC press. find
  20. Chin-Yew Lin. 2004. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out: Proceedings of the ACL-04 workshop. volume 8. find
  21. Katja Filippova, Enrique Alfonseca, Carlos A Colmenares, Lukasz Kaiser, and Oriol Vinyals. 2015. Sentence compression by deletion with lstms. In EMNLP. pages 360–368. find our card
  22. Fei Liu, Jeffrey Flanigan, Sam Thomson, Norman Sadeh, and Noah A Smith. 2015. Toward abstractive summarization using semantic representations. In NAACL: HLT. pages 1077–1086. find our card
  23. Yashar Mehdad, Giuseppe Carenini, Frank W Tompa, and Raymond T Ng. 2013. Abstractive meeting summarization with entailment and fusion. In Proc. of the 14th European Workshop on Natural Language Generation. pages 136–146. find
  24. Kavita Ganesan, ChengXiang Zhai, and Jiawei Han. 2010. Opinosis: a graph-based approach to abstractive summarization of highly redundant opinions. In Proceedings of the 23rd international conference on computational linguistics. ACL, pages 340–348. find our card
  25. Ramesh Nallapati, Bowen Zhou, Caglar Gulcehre, Bing Xiang, et al. 2016. Abstractive text summarization using sequence-to-sequence rnns and beyond. In CoNLL. find
  26. Shima Gerani, Yashar Mehdad, Giuseppe Carenini, Raymond T Ng, and Bita Nejat. 2014. Abstractive summarization of product reviews using discourse structure. In EMNLP. volume 14, pages 1602–1613. find
  27. Ankur P Parikh, Oscar Täckström, Dipanjan Das, and Jakob Uszkoreit. 2016. A decomposable attention model for natural language inference. In EMNLP. find our card
  28. George Giannakopoulos. 2009. Automatic summarization from multiple documents. Ph. D. dissertation . find
  29. Yongzheng Zhang, Nur Zincir-Heywood, and Evangelos Milios. 2004. World wide web site summarization. Web Intelligence and Agent Systems: An International Journal 2(1):39–53. find
  30. Ramakanth Pasunuru and Mohit Bansal. 2017. Reinforced video captioning with entailment rewards. In EMNLP. find
  31. Romain Paulus, Caiming Xiong, and Richard Socher. 2017. A deep reinforced model for abstractive summarization. arXiv preprint arXiv:1705.04304 . find our card
  32. Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. In EMNLP. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/pasunuru-2018-multi-reward-reinforced-summarization-saliency/