End-to-End Segmentation-based News Summarization

Findings of the Association for Computational Linguistics: ACL 2022, pages 544 - 554 May 22-27, 2022 c⃝2022 Association for Computational Linguistics End-to-End Segmentation-based News Summarization Yang Liu, Chenguang Zhu and Michael Zeng Microsoft Cognitive Services Research {yaliu10, chezhu, nzeng}@microsoft.com Abstract In this paper, we bring a new way of digest- ing news content by introducing the task of segmenting a news article into multiple sec- tions and generating the corresponding sum- mary to each section. We make two contribu- tions towards this new task. First, we create and make available a dataset, SEGNEWS, con- sisting of 27k news articles with sections and aligned heading-style section summaries. Sec- ond, we propose a novel segmentation-based language generation model adapted from pre- trained language models that can jointly seg- ment a document and produce the summary for each section. Experimental results on SEG- NEWS demonstrate that our model can out- perform several state-of-the-art sequence-to- sequence generation models for this new task. 1 Introduction In recent years, automatic summarization has re- ceived extensive attention in the natural language processing community, due to its potential for pro- cessing redundant information. The evolution of neural network models and availability of large- scale datasets have driven the rapid development of summarization systems. Despite promising results, there are specific char- acteristics of the traditional summarization task that impedes it to provide more beneficial ways of di- gesting long news articles. For instance, current news summarization system only provides one ge- netic summary of the whole article, and when users want to read in more details, the generated sum- mary is not capable of helping navigate the reading. For example, given a news report, current system will output several highlight summaries (Nallapati et al., 2017; Liu and Lapata, 2019; Zhang et al., 2020). Under this circumstance, if a user expect to read more details about one highlight, he will still need to browse the whole article to locate related paragraphs. Meanwhile, when processing a long news article, current systems usually truncate the text and only generate a summary based on the par- tial article (Cheng and Lapata, 2016a; Zhang et al., 2020). Although this is reasonable since most im- portant content usually lies in the initial portion, it also makes it difficult for users to quickly access information beyond the truncated portion. In this paper, we propose a new task of Segmentation-based News Summarization. Given a news article, we aim to identify its potential sec- tions and at the same time, to generate the cor- responding summary for each section. This new task provides a novel alternative to summarizing a news article. We argue that it can lead to a more organized way of understanding long articles and facilitates a more effective style of reading docu- ments. First, segmenting a news article can provide a structural organisation of the content, which is not only helpful to reading but also benefit many im- portant NLP tasks. For example, Brown et al. (1983) states that this kind of multi-paragraph di- vision …

Content truncated to 500 words