Domain-Independent Novel Event Discovery and Semi-Automatic Event Annotation

Domain-Independent Novel Event Discovery and Semi-Automatic Event Annotation Hao Li a, Xiang Li a, Heng Ji a, Yuval Marton b aComputer Science Department Queens College and Graduate Center City University of New York New York, NY 11367, USA hengji@cs.qc.cuny.edu bCenter of Computational Learning Systems Columbia University, New York, NY 10027, USA ymarton@ccls.columbia.edu Abstract. Information Extraction (IE) is becoming increasingly useful, but it is a costly task to discover and annotate novel events, event arguments, and event types. We exploit both monolingual texts and bilingual sentence-aligned parallel texts to cluster event triggers and discover novel event types. We then generate event argument annotations semi- automatically, framed as a sentence ranking and semantic role labeling task. Experiments on three different corpora -- ACE, OntoNotes and a collection of scientific literature -- have demonstrated that our domain-independent methods can significantly speed up the entire event discovery and annotation process while maintaining high quality. Keywords: Information Extraction, Novel Event Discovery, Domain-Independent, Semantic Role Labeling 1 Introduction Information Extraction (IE) techniques have been effectively applied to different domains (e.g. daily news, Wikipedia, biomedical reports, financial analysis and legal documents). A number of recent IE shared tasks (e.g. NIST Automatic Content Extraction Program (ACE 2005)) identify several common types of events. However, defining and identifying those types heavily rely on expert knowledge, and reaching an agreement among the experts or annotators requires a lot of human labor. Furthermore, annotating a high-quality event extraction corpus proves to be challenging because of highly-ambiguous and complicated event structures. The IE community has been aware of the limitation of this pre-defined event paradigm (e.g. Riloff, 1996; Yangarber et al., 2000; Grishman, 2001). Therefore a central track of IE research is the issue of portability – How can we automatically detect novel event types and rapidly annotate corpora for those types too, in order to alleviate the work load of human experts and annotators? Our hypothesis is that for each domain, there are typical event types that occur frequently, and thus should arouse our interest. For example, Transaction, Start-Organization events are most likely to appear in business domain; Justice events are likely to appear in law/criminal domain; Personnel (Start-Position and End-Position) event may appear frequently in politics and business domains according to the persons involved (President Obama: politics, Bill Gates: business). The general event types defined in the existing shared tasks are far from enough to satisfy the user needs in such domains. The first key problem is to automatically discover novel event types. In this paper we extract candidate event types based on event trigger clustering, and then rank these clusters based on their salience and novelty in the target unlabeled corpus (Section 3). After novel event types PACLIC 24 Proceedings 233 are discovered, it is beneficial to annotate arguments involved in such events. In this paper we demonstrate that such annotation can be realized in a semi-automatic way (Section 4). We take a new view of IE by considering it as a more fine-grained version of semantic role labeling …

Content truncated to 500 words