A Multi-Document Coverage Reward for RELAXed Multi-Document Summarization

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5112 - 5128 May 22-27, 2022 c⃝2022 Association for Computational Linguistics A Multi-Document Coverage Reward for RELAXed Multi-Document Summarization Jacob Parnell1,2, Inigo Jauregi Unanue1,2, Massimo Piccardi1 1University of Technology Sydney, NSW, Australia 2RoZetta Technology, NSW, Australia {jacob.parnell,inigo.jauregi}@rozettatechnology.com massimo.piccardi@uts.edu.au Abstract Multi-document summarization (MDS) has made significant progress in recent years, in part facilitated by the availability of new, dedi- cated datasets and capacious language models. However, a standing limitation of these mod- els is that they are trained against limited refer- ences and with plain maximum-likelihood ob- jectives. As for many other generative tasks, reinforcement learning (RL) offers the poten- tial to improve the training of MDS models; yet, it requires a carefully-designed reward that can ensure appropriate leverage of both the reference summaries and the input docu- ments. For this reason, in this paper we pro- pose fine-tuning an MDS baseline with a re- ward that balances a reference-based metric such as ROUGE with coverage of the input documents. To implement the approach, we utilize RELAX (Grathwohl et al., 2018), a con- temporary gradient estimator which is both low-variance and unbiased, and we fine-tune the baseline in a few-shot style for both sta- bility and computational efficiency. Experi- mental results over the Multi-News and WCEP MDS datasets show significant improvements of up to +0.95 pp average ROUGE score and +3.17 pp METEOR score over the baseline, and competitive results with the literature. In addition, they show that the coverage of the in- put documents is increased, and evenly across all documents. 1 Introduction Multi-document summarization (MDS) aims to consolidate salient points of information across a set of documents into a concise summary. The main requirement for the summary is that it ade- quately represent the document set, with low re- dundancy and high coverage across all documents, while at the same time being readable and fluent. Combined with this, is the need to develop tech- niques that can handle the significant memory com- plexity required to tackle MDS. Recently, the re- lease of dedicated datasets (Fabbri et al., 2019; Gholipour Ghalandari et al., 2020), and intelli- gently designed Transformer models (Liu et al., 2018; Liu and Lapata, 2019; Beltagy et al., 2020), have helped drive advancements in multi-document summarization, generally improving the accuracy and fluency of the predicted summaries. However, aspects such as the requirement to cover as much salient information from the input documents as possible, whilst still maintaining low repetition and low redundancy, have certainly been less explored to date (Nayeem et al., 2018; Mao et al., 2020). Within the sphere of contemporary neural MDS models, two main lines of investigation can be iden- tified: graph-based approaches (Li et al., 2020; Pa- sunuru et al., 2021), and concatenation approaches (Liu et al., 2018; Zhang et al., 2020a). The for- mer are approaches that rely on the construction of graphs to capture the inter- and intra-document relations. While powerful, they need to …

Content truncated to 500 words