Learning to predict script events from domain-specific text

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 205–210, Denver, Colorado, June 4–5, 2015. Learning to predict script events from domain-specific text Rachel Rudinger1, Vera Demberg3, Ashutosh Modi3, Benjamin Van Durme1,2, Manfred Pinkal3 Center for Language and Speech Processing1 Human Language Technology Center of Excellence2 Johns Hopkins University, Baltimore, MD USA MMCI Cluster of Excellence3 Saarland University, Saarbr¨ucken, Germany rudinger@jhu.edu, {vera,pinkal}@coli.uni-saarland.de, amodi@mmci.uni-saarland.de, vandurme@cs.jhu.edu Abstract The automatic induction of scripts (Schank and Abelson, 1977) has been the focus of many recent works. In this paper, we employ a variety of these methods to learn Schank and Abelson’s canonical restaurant script, us- ing a novel dataset of restaurant narratives we have compiled from a website called “Din- ners from Hell.” Our models learn narrative chains, script-like structures that we evaluate with the “narrative cloze” task (Chambers and Jurafsky, 2008). 1 Introduction A well-known theory from the intersection of psy- chology and artificial intelligence posits that humans organize certain kinds of general knowledge in the form of scripts, or common sequences of events (Schank and Abelson, 1977). Though many early AI systems employed hand-encoded scripts, more re- cent work has attempted to induce scripts with auto- matic and scalable techniques. In particular, several related techniques approach the problem of script in- duction as one of learning narrative chains from text corpora (Chambers and Jurafsky, 2008; Chambers and Jurafsky, 2009; Jans et al., 2012; Pichotta and Mooney, 2014). These statistical approaches have focused on open-domain script acquisition, in which a large number of scripts may be learned, but the ac- quisition of any particular set of scripts is not guar- anteed. For many specialized applications, however, knowledge of a few relevant scripts may be more useful than knowledge of many irrelevant scripts. With this scenario in mind, we attempt to learn the famous “restaurant script” (Schank and Abel- son, 1977) by applying the aforementioned narrative chain learning methods to a specialized corpus of dinner narratives we compile from the website “Din- ners from Hell.” Our results suggest that applying these techniques to a domain-specific dataset may be reasonable way to learn domain-specific scripts. 2 Background Previous work in the automatic induction of scripts or script-like structures has taken a number of dif- ferent approaches. Regneri et al. (2010) attempt to learn the structure of specific scripts by eliciting event sequence descriptions (ESDs) from humans to which they apply multiple sequence alignment (MSA) to yield one global structure per script. (Orr et al. (2014) learn similar structures in a probabilis- tic framework with Hidden Markov Models.) Al- though Regneri et al. (2010), like us, are concerned with learning pre-specified scripts, our approach is different in that we apply unsupervised techniques to scenario-specific collections of natural, pre-existing texts. Note that while the applicability of our approach to script learning may appear limited to domains for which a corpus conveniently already exists, pre- vious work demonstrates the feasibility of assem- bling such a corpus by automatically retrieving rel- evant documents from …

Content truncated to 500 words