Semantic Relations between Events and their Time, Locations and Participants for Event Coreference Resolution

Proceedings of Recent Advances in Natural Language Processing, pages 156–163, Hissar, Bulgaria, 7-13 September 2013. Semantic Relations between Events and their Time, Locations and Participants for Event Coreference Resolution Agata Cybulska VU University Amsterdam De Boelelaan 1105 1081HV Amsterdam a.k.cybulska@vu.nl Piek Vossen VU University Amsterdam De Boelelaan 1105 1081HV Amsterdam piek.vossen@vu.nl Abstract In this study, we measure the contribution of different event components and particular se- mantic relations to the task of event coreference resolution. First we calculate what event times, locations and participants add to event coreference resolution. Secondly, we an- alyze the contribution by hyponymy and gran- ularity within the participant component. Coreference of events is then calculated from the coreference match scores of each event component. Coreferent action candidates are accordingly filtered based on compatibility of their time, locations, or participants. We report the success rates of our experiments on a cor- pus annotated with coreferent events. 1 Introduction In this paper, we present an approach to event coreference resolution that employs the im- portance of full and partial linguistic coreference between events and their participants, times and locations. The goal of this work is to measure the contribution of different components of event descriptions to the task of event coreference resolution. Another goal is to calculate what se- mantic relations add to event coreference. Con- sidering the goals, we deliberately do not use machine learning as we want to have a clear pic- ture of what the contributions are by different factors. Having an idea of how various event components influence event coreference, will guide the feature choice for machine learning. Descriptions of one and the same event can differ in specificity and granularity (compare: two students taken hostage in Beslanian school vs. two people taken hostage in a classroom in Beslan Russia). High level events, as war, are more general and abstract with longer time span and group participants; low level events, e.g. a shooting event, are rather specific with shorter duration, and individual participants (Cybulska, Vossen, 2010). To capture differences between event representations and to identify relations between events, we applied an event model that consists of 4 components: a location, time, par- ticipant and an action slot (see Van Hage et al., 2011 for the formal SEM model along the same lines). In our previous work we extracted con- flict-related actions (e.g. war, genocide, shooting or fighting) and their participants, locations and times from text. Next, we determine relations between event mentions, starting with getting some insights into event coreference. 2 Related Work One of the recent approaches to event coreference resolution was proposed by Bejan and Harabagiu (2010), who experimented with nonparametric Bayesian models. Another one, by Chen et al. (2011) employs support vector ma- chines with tree kernels and spectral graph parti- tioning. These approaches do not explicitly ac- count for partial coreference of events, where some of the event components are related through hyponymy or part-of relationship, which is the focus of our work. Bejan and Harabagiu noted in their …

Content truncated to 500 words