MemSum: Extractive Summarization of Long Documents Using Multi-Step Episodic Markov Decision Processes

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6507 - 6522 May 22-27, 2022 c⃝2022 Association for Computational Linguistics MemSum: Extractive Summarization of Long Documents Using Multi-Step Episodic Markov Decision Processes Nianlong Gu Institute of Neuroinformatics, University of Zurich and ETH Zurich nianlong@ini.ethz.ch Elliott Ash Department of Humanities, Social and Political Sciences, ETH Zurich ashe@ethz.ch Richard H.R. Hahnloser Institute of Neuroinformatics, University of Zurich and ETH Zurich rich@ini.ethz.ch Abstract We introduce MemSum (Multi-step Episodic Markov decision process extractive SUMma- rizer), a reinforcement-learning-based extrac- tive summarizer enriched at each step with information on the current extraction history. When MemSum iteratively selects sentences into the summary, it considers a broad infor- mation set that would intuitively also be used by humans in this task: 1) the text content of the sentence, 2) the global text context of the rest of the document, and 3) the extraction his- tory consisting of the set of sentences that have already been extracted. With a lightweight ar- chitecture, MemSum obtains state-of-the-art test-set performance (ROUGE) in summariz- ing long documents taken from PubMed, arXiv, and GovReport. Ablation studies demonstrate the importance of local, global, and history in- formation. A human evaluation confirms the high quality and low redundancy of the gen- erated summaries, stemming from MemSum’s awareness of extraction history. 1 Introduction Automatic text summarization is the task of au- tomatically summarizing a long document into a relatively short text while preserving most of the information (Tas and Kiyani, 2007). Text summa- rization methods can be categorized into abstractive and extractive summarization (Gambhir and Gupta, 2017; Nenkova and McKeown, 2012). Given a document d consisting of an ordered list of N sen- tences, extractive summarization aims to pick up M (M≪N) sentences as the summary of the doc- ument. The extracted summaries tend to be both grammatically and semantically more reliable than abstractive summaries (Liu* et al., 2018; Liu and Lapata, 2019a; Luo et al., 2019; Liao et al., 2020), as they are directly selected from the source text. Extractive summarization is usually modeled as two sequential phases (Zhou et al., 2018): 1) sen- tence scoring and 2) sentence selection. In the Step Remaining sentences Extracted sentences Select Stop extraction Select Agent Agent Agent Extracted summary Score & Select Action Figure 1: We modeled extractive summarization as a multi-step iterative process of scoring and selecting sen- tences. si represents the ith sentence in the document D. sentence scoring phase, an affinity score is com- puted for each sentence by neural networks such as bidirectional RNNs (Dong et al., 2018; Narayan et al., 2018; Luo et al., 2019; Xiao and Carenini, 2019) or BERT (Zhang et al., 2019; Liu and Lapata, 2019b). In the sentence selection phase, sentences are selected by either i) predicting a label (1 or 0) for each sentence based on its score, and selecting sentences with label 1 (Zhang et al., 2019; Liu and Lapata, 2019b; Xiao and Carenini, 2019), or ii) ranking sentences based …

Content truncated to 500 words