Joint Learning-based Heterogeneous Graph Attention Network for Timeline Summarization

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4091 - 4104 July 10-15, 2022 ©2022 Association for Computational Linguistics Joint Learning-based Heterogeneous Graph Attention Network for Timeline Summarization Jingyi You, Dongyuan Li, Hidetaka Kamigaito, Kotaro Funakoshi and Manabu Okumura Tokyo Institute of Technology {youjy,lidy,kamigaito,funakoshi,oku}@lr.pi.titech.ac.jp Abstract Previous studies on the timeline summariza- tion (TLS) task ignored the information in- teraction between sentences and dates, and adopted pre-defined unlearnable representa- tions for them. They also considered date se- lection and event detection as two independent tasks, which makes it impossible to integrate their advantages and obtain a globally optimal summary. In this paper, we present a joint learning-based heterogeneous graph attention network for TLS (HeterTLS), in which date se- lection and event detection are combined into a unified framework to improve the extraction accuracy and remove redundant sentences si- multaneously. Our heterogeneous graph in- volves multiple types of nodes, the representa- tions of which are iteratively learned across the heterogeneous graph attention layer. We evalu- ated our model on four datasets, and found that it significantly outperformed the current state- of-the-art baselines with regard to ROUGE scores and date selection metrics. 1 Introduction Timeline summarization (TLS) is designed to ex- tract sentences that describe an evolutionary story from a massive amount of web articles with respect to a specific topic in chronological order. TLS has drawn much attention in recent years (Chen et al., 2019; Martschat and Markert, 2018; You et al., 2021b; Ghalandari and Ifrim, 2020; Yu et al., 2021) since it releases people from burdensome manual creation of summaries and gives readers a faster but comprehensive access to track events from many aspects, such as start and end, causality, and the main protagonists involved. Most studies on TLS seek ways to combine two individual subtasks: date selection and event detec- tion. Depending on different strategies for them, current methods are generally divided into three categories (Ghalandari and Ifrim, 2020): 1) direct summarization approaches (Chieu and Lee, 2004; Tran et al., 2013; Martschat and Markert, 2018; Duan et al., 2020) directly identify topic-related sentences from a collection of news articles to form a timeline; 2) date-wise summarization methods (Wang et al., 2016; Ghalandari and Ifrim, 2020; Li et al., 2021; Quatra et al., 2021) first select salient dates then construct a timeline for each date indi- vidually with sentences of the highest score; and 3) event detection algorithms (Steen and Markert, 2019; Duan et al., 2020; Yu et al., 2021) detect events by clustering sentences from multi-timeline news articles then identify several of the most im- portant events and summarize them separately. Although great successes have been achieved in conducting TLS, several issues remain unsolved. First, current TLS methods mainly adopt statistical hand-designed features to represent dates, e.g., the number of published articles and topic-related sen- tences in a specific time duration (Yu et al., 2021; Ghalandari and Ifrim, 2020), and employ sentence- BERT (Reimers and Gurevych, 2019) and …

Content truncated to 500 words