Open-Domain Event Detection using Distant Supervision

Proceedings of the 27th International Conference on Computational Linguistics, pages 878–891 Santa Fe, New Mexico, USA, August 20-26, 2018. 878 Open-Domain Event Detection using Distant Supervision Jun Araki and Teruko Mitamura Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA junaraki@cs.cmu.edu teruko@andrew.cmu.edu Abstract This paper introduces open-domain event detection, a new event detection paradigm to address issues of prior work on restricted domains and event annotation. The goal is to detect all kinds of events regardless of domains. Given the absence of training data, we propose a distant su- pervision method that is able to generate high-quality training data. Using a manually annotated event corpus as gold standard, our experiments show that despite no direct supervision, the model outperforms supervised models. This result indicates that the distant supervision enables robust event detection in various domains, while obviating the need for human annotation of events. 1 Introduction Events are a key semantic component integral to natural language understanding. They are a ubiquitous linguistic phenomenon appearing in various domains. For clarification, we use the term ‘domain’ to refer to a specific genre of text, such as biology, finance, and so forth. Prior studies on automatic event detection traditionally focus on limited types of events, mainly defined by several research initiatives and shared tasks in a few domains: • Newswire: TIPSTER (Onyshkevych et al., 1993) and MUC (Grishman and Sundheim, 1996) • Multi-domain: ACE (Doddington et al., 2004) and TAC KBP (Mitamura et al., 2016). • Biology: PASBio (Wattarujeekrit et al., 2004), GENIA (Kim et al., 2008), BioNLP (Kim et al., 2009) and ProcessBank (Berant et al., 2014). Although closed-domain event detection might be of practical use in some domain-specific scenarios, it only contributes to partial understanding of events because it only addresses a subset of events by definition. On the other hand, there is an established consensus that in order to advance natural language applications such as open-domain question answering, we need automatic event identification techniques with larger, wider, and more consistent coverage (Saur´ı et al., 2005; Pradhan et al., 2007). There are several pieces of prior work on open-domain events, but they have some limitations with respect to coverage of events. Lexical databases such as WordNet (Miller et al., 1990), FrameNet (Baker et al., 1998) and PropBank (Palmer et al., 2005) can be viewed as a superset of events, and their subtax- onomies seem to provide an extensional definition of events. However, these databases have a narrow coverage of events because they are generally expected to cover basic terminology due to their dictionary nature. For instance, none of WordNet, FrameNet and PropBank covers current terminology and proper nouns such as ‘Hurricane Katrina’. OntoNotes (Weischedel et al., 2011) is aimed at covering an unre- stricted set of events and entities, but its event annotation is limited to a small number of eventive nouns. TimeML (Pustejovsky et al., 2003) annotates events and times in a domain-agnostic manner, focusing on temporal aspects of events, but it does not deal with multi-word and generic …

Content truncated to 500 words