A Survey of Automatic Text Summarization Using Graph Neural Networks

Proceedings of the 29th International Conference on Computational Linguistics, pages 6139–6150 October 12–17, 2022. 6139 A Survey of Automatic Text Summarization using Graph Neural Networks Marco Ferdinand Salchner marco.salchner@student .uibk.ac.at Department of Computer Science, University of Innsbruck, Austria Adam Jatowt adam.jatowt@uibk.ac.at Department of Computer Science & Digital Science Center, University of Innsbruck, Austria Abstract Although automatic text summarization (ATS) has been researched for several decades, the application of graph neural networks (GNNs) to this task started relatively recently. In this survey we provide an overview on the rapidly evolving approach of using GNNs for the task of automatic text summarization. In particular we provide detailed information on the func- tionality of GNNs in the context of ATS, and a comprehensive overview of models utilizing this approach. 1 Introduction The advent of the internet has led to an explosion in the amount of textual information available online. The extensive availability of textual information paired with a need to quickly understand it has led to major efforts in the field of automatic text summarization (ATS). For a comprehensive review and survey of ATS as a task we recommend a recent survey by El-Kassas et al. (2021). The goal of ATS is to produce a concise, correct and fluent summary of a given text. Although this definition is intuitively understandable, there is no commonly agreed upon formal definition for these qualities. This is in part due to the difficulty of the task itself as producing a summary with the above properties is challenging, even for humans. Although the field of ATS has made major steps forward, generally, we differentiate between two basic approaches to the task of ATS. The first being extractive ATS and the second being abstractive ATS. Extractive ATS involves extracting text spans from the original document such that a summary is generated. Thus the main challenge consists in identifying useful spans of text from the original document. This approach is popular as it elimi- nates the non-trivial task of generating factually correct and coherent sentences. Abstractive ATS on the other hand involves the challenging task of generating novel sentences for a summary. Despite impressive advancements, generating factually cor- rect and fluent sentences is still a major challenge in ATS (Kry´sci´nski et al., 2020). In addition to the above, one differentiates between the task of single- document and multi-document summarization. 1.1 Why graph neural networks ? Contemporary solutions to the task of ATS suf- fer from a number of issues, chiefly an inconsis- tent evaluation protocol and, somewhat, a lack of progress, as noted by Kry´sci´nski et al. (2019). In recent years GNNs have been successfully applied to a number of downstream NLP tasks such as clas- sification (Liu et al., 2020b) (Zhang et al., 2020b) and translation (Xu et al., 2021) (Yin et al., 2020). Although GNNs may not be able to solve all prob- lems related to the task of ATS, we believe that they can at least give a new perspective to this task. Generally GNNs bring a …

Content truncated to 500 words