EDU-level Extractive Summarization with Varying Summary Lengths

Findings of the Association for Computational Linguistics: EACL 2023, pages 1655–1667 May 2-6, 2023 ©2023 Association for Computational Linguistics EDU-level Extractive Summarization with Varying Summary Lengths Yuping Wu, Ching-Hsun Tseng, Jiayu Shang, Shengzhong Mao, Goran Nenadic, Xiao-Jun Zeng∗ Department of Computer Science, University of Manchester {yuping.wu-2, ching-hsun.tseng, jiayu.shang, shengzhong.mao}@postgrad.manchester.ac.uk gnenadic, x.zeng@manchester.ac.uk Abstract Extractive models usually formulate text sum- marization as extracting fixed top-k salient sen- tences from the document as a summary. Few works exploited extracting finer-grained Ele- mentary Discourse Unit (EDU) with little anal- ysis and justification for the extractive unit se- lection. Further, the selection strategy of the fixed top-k salient sentences fits the summariza- tion need poorly, as the number of salient sen- tences in different documents varies and there- fore a common or best k does not exist in reality. To fill these gaps, this paper first conducts the comparison analysis of oracle summaries based on EDUs and sentences, which provides evi- dence from both theoretical and experimental perspectives to justify and quantify that EDUs make summaries with higher automatic evalua- tion scores than sentences. Then, considering this merit of EDUs, this paper further proposes an EDU-level extractive model with Varying summary Lengths (EDU-VL1) and develops the corresponding learning algorithm. EDU- VL learns to encode and predict probabilities of EDUs in the document, generate multiple can- didate summaries with varying lengths based on various k values, and encode and score can- didate summaries, in an end-to-end training manner. Finally, EDU-VL is experimented on single and multi-document benchmark datasets and shows improved performances on ROUGE scores in comparison with state-of-the-art ex- tractive models, and further human evalua- tion suggests that EDU-constituent summaries maintain good grammaticality and readability. 1 Introduction Automatic text summarization aims at aggregat- ing information in long document(s) into a shorter piece of text while keeping important information. Extractive summarization and abstractive summa- rization are two categories of it. This paper focuses ∗Corresponding author. 1https://github.com/yuping-wu/EDU-VL Document: (...) [The second audio,] [taken from dash cam video from inside a patrol car,] [captures a phone call between Slager and someone] [CNN believes] [is his wife.] (...) Reference Summary: The second audio cap- tures a phone call between Slager and some- one CNN believes is his wife. Table 1: Example to demonstrate redundant information in sentence. Content within [] indicates an EDU. only on the extractive task which formulates sum- marization as identifying salient textual segments in document (Lunh, 1958). Under the supervised learning framework, this task is further formulated as a label classification task, i.e., encoding textual segments and predicting labels on the encoded vec- tors. Recent state-of-the-art models (Liu and Lap- ata, 2019; Zhong et al., 2020; Liu et al., 2021; Ruan et al., 2022) on this task tend to be Transformer- based since BERT (Devlin et al., 2019) shows sig- nificantly better performance than RNN on most natural language understanding tasks. Most existing works extract sentences from the document and some works further (Xu and Durrett, 2019) propose post-processing steps to prune the generated summary. …

Content truncated to 500 words