Multi-document Summarization with Maximal Marginal Relevance-guided Reinforcement Learning

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1737–1751, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1737 Multi-document Summarization with Maximal Marginal Relevance-guided Reinforcement Learning Yuning Mao1, Yanru Qu1, Yiqing Xie1, Xiang Ren2, Jiawei Han1 1University of Illinois at Urbana-Champaign, IL, USA 2University of Southern California, CA, USA 1{yuningm2, yanruqu2, xyiqing2, hanj}@illinois.edu 2xiangren@usc.edu Abstract While neural sequence learning methods have made significant progress in single-document summarization (SDS), they produce unsatis- factory results on multi-document summariza- tion (MDS). We observe two major challenges when adapting SDS advances to MDS: (1) MDS involves larger search space and yet more limited training data, setting obstacles for neural methods to learn adequate represen- tations; (2) MDS needs to resolve higher in- formation redundancy among the source docu- ments, which SDS methods are less effective to handle. To close the gap, we present RL- MMR, Maximal Margin Relevance-guided Reinforcement Learning for MDS, which uni- fies advanced neural SDS methods and sta- tistical measures used in classical MDS. RL- MMR casts MMR guidance on fewer promis- ing candidates, which restrains the search space and thus leads to better representation learning. Additionally, the explicit redundancy measure in MMR helps the neural represen- tation of the summary to better capture re- dundancy. Extensive experiments demonstrate that RL-MMR achieves state-of-the-art per- formance on benchmark MDS datasets. In par- ticular, we show the benefits of incorporating MMR into end-to-end learning when adapting SDS to MDS in terms of both learning effec- tiveness and efficiency.1 1 Introduction Text summarization aims to produce condensed summaries covering salient and non-redundant in- formation in the source documents. Recent studies on single-document summarization (SDS) bene- fit from the advances in neural sequence learning (Nallapati et al., 2016; See et al., 2017; Chen and Bansal, 2018; Narayan et al., 2018) as well as pre- trained language models (Liu and Lapata, 2019; 1Code can be found at https://github.com/ morningmoni/RL-MMR. Lewis et al., 2019; Zhang et al., 2020) and make great progress. However, in multi-document sum- marization (MDS) tasks, neural models are still facing challenges and often underperform classi- cal statistical methods built upon handcrafted fea- tures (Kulesza and Taskar, 2012). We observe two major challenges when adapt- ing advanced neural SDS methods to MDS: (1) Large search space. MDS aims at producing sum- maries from multiple source documents, which ex- ceeds the capacity of neural SDS models (See et al., 2017) and sets learning obstacles for adequate rep- resentations, especially considering that MDS la- beled data is more limited. For example, there are 287K training samples (687 words on average) on the CNN/Daily Mail SDS dataset (Nallapati et al., 2016) and only 30 on the DUC 2003 MDS dataset (6,831 words). (2) High redundancy. In MDS, the same statement or even sentence can spread across different documents. Although SDS models adopt attention mechanisms as implicit measures to reduce redundancy (Chen and Bansal, 2018), they fail to handle the much higher redundancy of MDS effectively (Sec. 4.2.3). There have been attempts to …

Content truncated to 500 words