Liberal Event Extraction and Event Schema Induction

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 258–268, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Liberal Event Extraction and Event Schema Induction Lifu Huang1, Taylor Cassidy2, Xiaocheng Feng3, Heng Ji1, Clare R. Voss2, Jiawei Han4, Avirup Sil5 1 Rensselaer Polytechnic Institute, 2 US Army Research Lab, 3 Harbin Institute of Technology, 4 Univerisity of Illinois at Urbana-Champaign, 5 IBM T.J. Watson Research Center 1{huangl7,jih}@rpi.edu, 2{taylor.cassidy.civ,clare.r.voss.civ}@rpi.edu, 3xcfeng@ir.hit.edu.cn, 4hanj@illinois.edu, 5avi@us.ibm.com Abstract We propose a brand new “Liberal” Event Extraction paradigm to extract events and discover event schemas from any input corpus simultaneously. We incorporate symbolic (e.g., Abstract Meaning Repre- sentation) and distributional semantics to detect and represent event structures and adopt a joint typing framework to simulta- neously extract event types and argument roles and discover an event schema. Ex- periments on general and specific domains demonstrate that this framework can con- struct high-quality schemas with many event and argument role types, covering a high proportion of event types and argu- ment roles in manually defined schemas. We show that extraction performance us- ing discovered schemas is comparable to supervised models trained from a large amount of data labeled according to pre- defined event types. The extraction quality of new event types is also promising. 1 Introduction Event extraction aims at identifying and typ- ing trigger words and participants (arguments). It remains a challenging and costly task. The first question is what to extract? The TIP- STER (Onyshkevych et al., 1993), MUC (Grish- man and Sundheim, 1996), CoNLL (Tjong et al., 2003; Pradhan et al., 2011), ACE 1 and TAC- KBP (Ji and Grishman, 2011) programs found that it was feasible to manually define an event schema based on the needs of potential users. An ACE event schema example is shown in Figure 1. This process is very expensive because consumers and 1http://www.itl.nist.gov/iad/mig/tests/ace/ expert linguists need to examine a lot of data be- fore specifying the types of events and argument roles and writing detailed annotation guidelines for each type in the schema. Manually-defined event schemas often provide low coverage and fail to generalize to new domains. For example, none of the aforementioned programs include “dona- tion” and “evacuation” in their schema in spite of their potential relevance to users. In this paper we propose Liberal Event Extrac- tion, a new paradigm to take humans out of the loop and enable systems to extract events in a more liberal fashion. It automatically discovers a com- plete event schema, customized for a specific input corpus. Figure 1 compares the ACE event extrac- tion paradigm and our proposed Liberal event ex- traction paradigm. We use the following examples to explain and motivate our approach, where event triggers are in bold and arguments are in italics and underlined: E1. Two Soldiers were killed and one injured in the close-quarters fighting in Kut. E2. Bill Bennet’s glam gambling loss changed my opinion. E3. Gen. Vincent Brooks announced the capture of Barzan Ibrahim Hasan al-Tikriti, telling re- porters …

Content truncated to 500 words