Probabilistic Frame Induction

Proceedings of NAACL-HLT 2013, pages 837–846, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Probabilistic Frame Induction Jackie Chi Kit Cheung∗ Department of Computer Science University of Toronto Toronto, ON, M5S 3G4, Canada jcheung@cs.toronto.edu Hoifung Poon One Microsoft Way Microsoft Research Redmond, WA 98052, USA hoifung@microsoft.com Lucy Vanderwende One Microsoft Way Microsoft Research Redmond, WA 98052, USA lucyv@microsoft.com Abstract In natural-language discourse, related events tend to appear near each other to describe a larger scenario. Such structures can be formal- ized by the notion of a frame (a.k.a. template), which comprises a set of related events and prototypical participants and event transitions. Identifying frames is a prerequisite for infor- mation extraction and natural language gen- eration, and is usually done manually. Meth- ods for inducing frames have been proposed recently, but they typically use ad hoc proce- dures and are difficult to diagnose or extend. In this paper, we propose the first probabilistic approach to frame induction, which incorpo- rates frames, events, and participants as latent topics and learns those frame and event transi- tions that best explain the text. The number of frame components is inferred by a novel application of a split-merge method from syn- tactic parsing. In end-to-end evaluations from text to induced frames and extracted facts, our method produces state-of-the-art results while substantially reducing engineering effort. 1 Introduction Events with causal or temporal relations tend to oc- cur near each other in text. For example, a BOMB- ING scenario in an article on terrorism might be- gin with a DETONATION event, in which terrorists set off a bomb. Then, a DAMAGE event might en- sue to describe the resulting destruction and any casualties, followed by an INVESTIGATION event ∗This research was undertaken during the author’s internship at Microsoft Research. covering subsequent police investigations. After- wards, the BOMBING scenario may transition into a CRIMINAL-PROCESSING scenario, which begins with police catching the terrorists, and proceeds to a trial, sentencing, etc. A common set of partici- pants serves as the event arguments; e.g., the agent (or subject) of DETONATION is often the same as the theme (or object) of INVESTIGATION and corre- sponds to a PERPETRATOR. Such structures can be formally captured by the notion of a frame (a.k.a. template, scenario), which consists of a set of events with prototypical transi- tions, as well as a set of slots representing the com- mon participants. Identifying frames is an explicit or implicit prerequisite for many NLP tasks. Infor- mation extraction, for example, stipulates the types of events and slots that are extracted for a frame or template. Online applications such as dialogue sys- tems and personal-assistant applications also model users’ goals and subgoals using frame-like represen- tations. In natural-language generation, frames are often used to represent contents to be expressed as well as to support surface realization. Until recently, frames and related representations have been manually constructed, which has limited their applicability to a relatively small number of do- mains and a few slots within a domain. Furthermore, …

Content truncated to 500 words