Generating Coherent Event Schemas at Scale

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1721–1731, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Generating Coherent Event Schemas at Scale Niranjan Balasubramanian, Stephen Soderland, Mausam, Oren Etzioni Computer Science & Engineering University of Washington Seattle, WA 98195, USA {niranjan,ssoderlan, mausam, etzioni}@cs.washington.edu Abstract Chambers and Jurafsky (2009) demonstrated that event schemas can be automatically in- duced from text corpora. However, our analy- sis of their schemas identifies several weak- nesses, e.g., some schemas lack a common topic and distinct roles are incorrectly mixed into a single actor. It is due in part to their pair-wise representation that treats subject- verb independently from verb-object. This of- ten leads to subject-verb-object triples that are not meaningful in the real-world. We present a novel approach to inducing open-domain event schemas that overcomes these limitations. Our approach uses co- occurrence statistics of semantically typed re- lational triples, which we call Rel-grams (re- lational n-grams). In a human evaluation, our schemas outperform Chambers’s schemas by wide margins on several evaluation criteria. Both Rel-grams and event schemas are freely available to the research community. 1 Introduction Event schemas (also known as templates or frames) have been widely used in information extraction. An event schema is a set of actors (also known as slots) that play different roles in an event, such as the perpetrator, victim, and instrument in a bomb- ing event. They provide essential guidance in ex- tracting information related to events from free text (Patwardhan and Riloff, 2009), and can also aid in other NLP tasks, such as coreference (Irwin et al., 2011), summarization (Owczarzak and Dang, 2010), and inference about temporal ordering and causality. Actor Rel Actor A1:<person> failed A2:test A1:<person> was suspended for A3:<time period> A1:<person> used A4:<substance, drug> A1:<person> was suspended for A5:<game, activity> A1:<person> was in A6:<location> A1:<person> was suspended by A7:<org, person> Actor Instances: A1: {Murray, Morgan, Governor Bush, Martin, Nelson} A2: {test} A3: {season, year, week, month, night} A4: {cocaine, drug, gasoline, vodka, sedative} A5: {violation, game, abuse, misfeasance, riding} A6: {desert, Simsbury, Albany, Damascus, Akron} A7: {Fitch, NBA, Bud Selig, NFL, Gov Jeb Bush} Table 1: An event schema produced by our system, rep- resented as a set of (Actor, Rel, Actor) triples, and a set of instances for each actor A1, A2, etc. For clarity we show unstemmed verbs. Until recently, all event schemas in use in NLP were hand-engineered, e.g., the MUC templates and ACE event relations (ARPA, 1991; ARPA, 1998; Doddington et al., 2004). This led to technology that could only focus on specific domains of interest and has not been applicable more broadly. The seminal work of Chambers and Jurafsky (2009) showed that event schemas can also be in- duced automatically from text corpora. Instead of labeled roles these schemas have a set of relations and actors that serve as arguments.1 Their system is fully automatic, domain-independent, and scales to large text corpora. However, we identify several limitations in the schemas produced by …

Content truncated to 500 words