Abstractive Multi-Document Summarization via Joint Learning with Single-Document Summarization

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2545–2554 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2545 Abstractive Multi-Document Summarization via Joint Learning with Single-Document Summarization Hanqi Jin, Xiaojun Wan Center for Data Science, Peking University Wangxuan Institute of Computer Technology, Peking University The MOE Key Laboratory of Computational Linguistics, Peking University {jinhanqi,wanxiaojun}@pku.edu.cn Abstract Single-document and multi-document summa- rizations are very closely related in both task definition and solution method. In this work, we propose to improve neural abstrac- tive multi-document summarization by jointly learning an abstractive single-document sum- marizer. We build a unified model for single-document and multi-document summa- rizations by fully sharing the encoder and de- coder and utilizing a decoding controller to ag- gregate the decoder’s outputs for multiple in- put documents. We evaluate our model on two multi-document summarization datasets: Multi-News and DUC-04. Experimental re- sults show the efficacy of our approach, and it can substantially outperform several strong baselines. We also verify the helpfulness of single-document summarization to abstractive multi-document summarization task. 1 Introduction Document summarization aims at producing a flu- ent, condensed summary for the given document or document set. It involves identifying important information and filtering out redundant information from input sources. While single-document sum- marization takes a single source document as input, multi-document summarization requires producing a summary from a cluster of thematically related documents. There are two primary methodologies for document summarization: extractive and ab- stractive. Extractive methods directly select impor- tant sentences from the original documents, which are relatively simple but face the drawbacks of in- formation redundancy and incoherence between sentences. Abstractive methods enable generating new words, phrases, and sentences, which are able to generate better summaries with higher readabil- ity and conciseness. In this paper, we focus on abstractive document summarization. Empowered by large parallel datasets auto- matically harvested from online news websites, sequence-to-sequence learning has shown promis- ing results on abstractive single-document sum- marization (See et al., 2017; Paulus et al., 2018; Tan et al., 2017; C¸ elikyilmaz et al., 2018). Com- pared with single-document summarization, anno- tated multi-document summarization datasets are often scarce. Several works have explored adapt- ing the neural encoder-decoder model trained for single-document summarization to multi-document summarization. Zhang et al. (2018) add a doc- ument set encoder to extend the neural abstrac- tive model trained on large scale single-document summarization corpus to the multi-document sum- marization task. Lebanoff et al. (2018) incorpo- rate the maximal marginal relevance method into a neural encoder-decoder model trained for single- document summarization to address the informa- tion redundancy for multi-document summariza- tion. Single-document and multi-document summa- rizations are very closely related in both task defini- tion and solution method (Wan, 2010). Both tasks need to deal with document-level input, identify the important content of documents, and paraphrase the important information to generate the summary, while the main difference is that multi-document summarization involves summarizing multiple in- put documents. Since the two tasks are closely re- lated, it is promising to learn for …

Content truncated to 500 words