Statistical Script Learning with Multi-Argument Events

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 220–229, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics Statistical Script Learning with Multi-Argument Events Karl Pichotta Department of Computer Science The University of Texas at Austin pichotta@cs.utexas.edu Raymond J. Mooney Department of Computer Science The University of Texas at Austin mooney@cs.utexas.edu Abstract Scripts represent knowledge of stereotyp- ical event sequences that can aid text un- derstanding. Initial statistical methods have been developed to learn probabilis- tic scripts from raw text corpora; how- ever, they utilize a very impoverished rep- resentation of events, consisting of a verb and one dependent argument. We present a script learning approach that employs events with multiple arguments. Unlike previous work, we model the interactions between multiple entities in a script. Ex- periments on a large corpus using the task of inferring held-out events (the “narrative cloze evaluation”) demonstrate that mod- eling multi-argument events improves pre- dictive accuracy. 1 Introduction Scripts encode knowledge of stereotypical events, including information about their typical ordered sequences of sub-events and corresponding argu- ments (Schank and Abelson, 1977). The clas- sic example is the “restaurant script,” which en- codes knowledge about what normally happens when dining out. Such knowledge can be used to improve text understanding by supporting in- ference of missing actions and events, as well as resolution of lexical and syntactic ambiguities and anaphora (Rahman and Ng, 2012). For example, given the text “John went to Olive Garden and or- dered lasagna. He left a big tip and left,” an infer- ence that scripts would ideally allow us to make is “John ate lasagna.” There is a small body of recent research on auto- matically learning probabilistic models of scripts from large corpora of raw text (Manshadi et al., 2008; Chambers and Jurafsky, 2008; Chambers and Jurafsky, 2009; Jans et al., 2012). However, this work uses a very impoverished representation of events that only includes a verb and a single de- pendent entity. We propose a more complex multi- argument event representation for use in statistical script models, capable of directly capturing inter- actions between multiple entities. We present a method for learning such a model, and provide ex- perimental evidence that modeling entity interac- tions allows for better prediction of events in docu- ments, compared to previous single-entity “chain” models. We also compare to a competitive base- line not used in previous work, and introduce a novel evaluation metric. 2 Background The idea of representing stereotypical event se- quences for textual inference originates in the seminal work of Schank and Abelson (1977). Early scripts were manually engineered for spe- cific domains; however, Mooney and DeJong (1985) present an early knowledge-based method for learning scripts from a single document. These early scripts (and methods for learning them) were non-statistical and fairly brittle. Chambers and Jurafsky (2008) introduced a method for learning statistical scripts that, using a much simpler event representation that allows for efficient learning and inference. Jans et al. …

Content truncated to 500 words