Enhancing Multi-Document Summarization with Cross-Document Graph-based Information Extraction

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1696–1707 May 2-6, 2023 ©2023 Association for Computational Linguistics Enhancing Multi-Document Summarization with Cross-Document Graph-based Information Extraction Zixuan Zhang1, Heba Elfardy2, Markus Dreyer2, Kevin Small2, Heng Ji2, Mohit Bansal2 1University of Illinois at Urbana-Champaign 2Amazon Alexa zixuan11@illinois.edu {helfardy, mddreyer, smakevin, jihj, mobansal}@amazon.com Abstract Information extraction (IE) and summarization are closely related, both tasked with present- ing a subset of the information contained in a natural language text. However, while IE ex- tracts structural representations, summarization aims to abstract the most salient information into a generated text summary – thus poten- tially encountering the technical limitations of current text generation methods (e.g., halluci- nation). To mitigate this risk, this work uses structured IE graphs to enhance the abstrac- tive summarization task. Specifically, we focus on improving Multi-Document Summarization (MDS) performance by using cross-document IE output, incorporating two novel components: (1) the use of auxiliary entity and event recog- nition systems to focus the summary genera- tion model and; (2) incorporating an alignment loss between IE nodes and their text spans to reduce inconsistencies between the IE graphs and text representations. Operationally, both the IE nodes and corresponding text spans are projected into the same embedding space and pairwise distance is minimized. Experimental results on multiple MDS benchmarks show that summaries generated by our model are more factually consistent with the source documents than baseline models while maintaining the same level of abstractiveness.1,2 1 Introduction Information extraction (IE) (Lin et al., 2020; Li et al., 2021) and summarization (Xiao et al., 2022; Pasunuru et al., 2021) are inherently similar tasks, sharing the objective of identifying and presenting a targeted subset of the information present in a nat- ural language text. However, there are also concep- tual and methodological distinctions. First of all, IE aims to extract specific structured information 1All our code will be publicly available at https:// github.com/amazon-science/IESum. 2The work was done during the first author’s internship at Amazon Alexa. from natural language text while abstractive text summarization targets abstracting the most salient information of a given text into a natural language summary. Secondly, IE methods frequently have access to world knowledge via external schema and knowledge resources (e.g., Wikidata) whereas summarization methods often rely on the informa- tion encoded in large-scale pretrained embeddings to produce coherent summaries. The complemen- tary aspects of these tasks imply an opportunity to transfer knowledge from one task to another. Hence, in this paper, our primary motivation is to take advantage of the complementary nature of IE and summarization tasks, using the structured output of entity and event extraction systems to improve abstractive text summarization by focus- ing text generation toward explicitly observable grounded concepts. There are a few previous re- search studies exploring the mutual enhancement between IE and summarization. For example, Lu et al. (2022) use text summarization to improve relation extraction and Pasunuru et al. (2021) adopt open-domain IE to provide additional structural in- puts for …

Content truncated to 500 words