Event Schema Induction with a Probabilistic Entity-Driven Model

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1797–1807, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Event Schema Induction with a Probabilistic Entity-Driven Model Nathanael Chambers United States Naval Academy Annapolis, MD 21402 nchamber@usna.edu Abstract Event schema induction is the task of learning high-level representations of complex events (e.g., a bombing) and their entity roles (e.g., perpetrator and victim) from unlabeled text. Event schemas have important connections to early NLP research on frames and scripts, as well as modern applications like template extraction. Recent research suggests event schemas can be learned from raw text. In- spired by a pipelined learner based on named entity coreference, this paper presents the first generative model for schema induction that in- tegrates coreference chains into learning. Our generative model is conceptually simpler than the pipelined approach and requires far less training data. It also provides an interesting contrast with a recent HMM-based model. We evaluate on a common dataset for template schema extraction. Our generative model matches the pipeline’s performance, and out- performs the HMM by 7 F1 points (20%). 1 Introduction Early research in language understanding focused on high-level semantic representations to drive their models. Many proposals, such as frames and scripts, used rich event schemas to model the situations de- scribed in text. While the field has since focused on more shallow approaches, recent work on schema induction shows that event schemas might be learn- able from raw text. This paper continues the trend, addressing the question, can event schemas be in- duced from raw text without prior knowledge? We present a new generative model for event schemas, and it produces state-of-the-art induction results, in- cluding a 7 F1 point gain over a different generative proposal developed in parallel with this work. Event schemas are unique from most work in in- formation extraction (IE). Current relation discovery (Banko et al., 2007a; Carlson et al., 2010b) focuses on atomic facts and relations. Event schemas build relations into coherent event structures, often called templates in IE. For instance, an election template jointly connects that obama won a presidential elec- tion with romney was the defeated, the election oc- curred in 2012, and the popular vote was 50-48. The entities in these relations fill specific semantic roles, as in this template schema: Template Schema for Elections (events: nominate, vote, elect, win, declare, concede) Date: Timestamp Winner: Person Loser: Person Position: Occupation Vote: Number Traditionally, template extractors assume fore- knowledge of the event schemas. They know a Win- ner exists, and research focuses on supervised learn- ing to extract winners from text. This paper focuses on the other side of the supervision spectrum. The learner receives no human input, and it first induces a schema before extracting instances of it. Our proposed model contributes to a growing line of research in schema induction. The majority of previous work relies on ad-hoc clustering algo- rithms (Filatova et al., 2006; Sekine, 2006; Cham- bers and Jurafsky, 2011). Chambers and …

Content truncated to 500 words