Event Detection and Co-reference with Minimal Supervision

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 392–402, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Event Detection and Co-reference with Minimal Supervision Haoruo Peng1 and Yangqiu Song2 and Dan Roth1 1University of Illinois, Urbana-Champaign 2Department of Computer Science and Engineering, Hong Kong University of Science and Technology 1{hpeng7,danr}@illinois.edu, 2yqsong@cse.ust.hk Abstract An important aspect of natural language un- derstanding involves recognizing and catego- rizing events and the relations among them. However, these tasks are quite subtle and an- notating training data for machine learning based approaches is an expensive task, re- sulting in supervised systems that attempt to learn complex models from small amounts of data, which they over-fit. This paper addresses this challenge by developing an event detec- tion and co-reference system with minimal su- pervision, in the form of a few event exam- ples. We view these tasks as semantic similar- ity problems between event mentions or event mentions and an ontology of types, thus fa- cilitating the use of large amounts of out of domain text data. Notably, our semantic re- latedness function exploits the structure of the text by making use of a semantic-role-labeling based representation of an event. We show that our approach to event detection is competitive with the top supervised meth- ods. More significantly, we outperform state- of-the-art supervised methods for event co- reference on benchmark data sets, and support significantly better transfer across domains. 1 Introduction Natural language understanding involves, as a key component, the need to understand events men- tioned in texts. This entails recognizing elements such as agents, patients, actions, location and time, among others. Understanding events also necessi- tates understanding relations among them and, as a minimum, determining whether two snippets of text represent the same event or not – the event co- reference problem. Events have been studied for years, but they still remain a key challenge. One reason is that the frame-based structure of events ne- cessitates addressing multiple coupled problems that are not easy to study in isolation. Perhaps an even more fundamental difficulty is that it is not clear whether our current set of events’ definitions is ade- quate (Hovy et al., 2013). Thus, given the complex- ity and fundamental difficulties, the current evalua- tion methodology in this area focuses on a limited domain of events, e.g. 33 types in ACE 2005 (NIST, 2005) and 38 types in TAC KBP (Mitamura et al., 2015). Consequently, this allows researchers to train supervised systems that are tailored to these sets of events and that overfit the small domain covered in the annotated data, rather than address the realistic problem of understanding events in text. In this paper, we pursue an approach to under- standing events that we believe to be more feasi- ble and scalable. Fundamentally, event detection is about identifying whether an event in context is semantically related to a set of events of a specific type; and, event co-reference is about whether two event mentions are semantically …

Content truncated to 500 words