Detecting Subevents using Discourse and Narrative Features

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4780–4790 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4780 Detecting Subevents using Discourse and Narrative Features Mohammed Aldawsari & Mark A. Finlayson School of Computing and Information Sciences Florida International University Miami FL, 33199 {malda021,markaf}@fiu.edu Abstract Recognizing the internal structure of events is a challenging language processing task of great importance for text understanding. We present a supervised model for automatically identifying when one event is a subevent of an- other. Building on prior work, we introduce several novel features, in particular discourse and narrative features, that significantly im- prove upon prior state-of-the-art performance. Error analysis further demonstrates the utility of these features. We evaluate our model on the only two annotated corpora with event hi- erarchies: HiEve and the Intelligence Commu- nity corpus. No prior system has been evalu- ated on both corpora. Our model outperforms previous systems on both corpora, achieving 0.74 BLANC F1 on the Intelligence Commu- nity corpus and 0.70 F1 on the HiEve corpus, respectively a 15 and 5 percentage point im- provement over previous models. 1 Introduction An event is something that occurs in a certain place at a certain time (Pustejovsky et al., 2003). Understanding events plays a major role in various natural language processing tasks such as infor- mation extraction (Humphreys et al., 1997), ques- tion answering (Narayanan and Harabagiu, 2004), textual entailment (Haghighi et al., 2005), event coreference (Choubey and Huang, 2018) and con- tradiction detection (De Marneffe et al., 2008). There has been a significant amount of work on automatic processing of events in text including systems for events extraction, event coreference resolution, and temporal relation detection (Araki, 2018; Ning et al., 2017). However, events are not atomic entities: they often have complex inter- nal structure that can be expressed in a variety of ways (Huttunen et al., 2002; Bejan and Harabagiu, 2008; Hovy et al., 2013). One of the unsolved problems related to event understanding is the detection of subevents, also referred to as event hierarchy construction. As de- scribed by Glavaˇs and ˇSnajder (2014a), there have been efforts that have focused on detecting tem- poral and spatial subevent containment individu- ally. However, it is clear that subevent detection requires both simultaneously. The subevent rela- tionship is defined in terms of (e1,e2), where e1 and e2 are events: event e2 is a subevent of event e1 if e2 is spatiotemporally contained by e1. More precisely, we say that an event e1 is a parent event of event e2, and e2 is a child event of e1 if (1) e1 is collector event that contains a complex se- quence of activities; (2) e2 is one of these activ- ities; and (3) e2 is spatially and temporally con- tained within e1 (i.e., e2 occur at the same time and same place as e1) (Hovy et al., 2013; Glavaˇs and ˇSnajder, 2014b). This subevent relationship is independent of other types of relationships, e.g., …

Content truncated to 500 words