Improving event prediction by representing script participants

Proceedings of NAACL-HLT 2016, pages 546–551, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Improving event prediction by representing script participants Simon Ahrendt and Vera Demberg Saarland University 66123 Saarbr¨ucken Germany {simona,vera}@coli.uni-saarland.de Abstract Automatically learning script knowledge has proved difficult, with previous work not or just barely beating a most-frequent baseline. Script knowledge is a type of world knowl- edge which can however be useful for various task in NLP and psycholinguistic modelling. We here propose a model that includes partici- pant information (i.e., knowledge about which participants are relevant for a script) and show, on the Dinners from Hell corpus as well as the InScript corpus, that this knowledge helps us to significantly improve prediction perfor- mance on the narrative cloze task. 1 Introduction Scripts represent knowledge about typical event se- quences (Schank and Abelson, 1977), for exam- ple the sequence of events happening when eating at a restaurant. Script knowledge thereby includes events like order, bring and eat as well as par- ticipants of those events, e.g., menu, waiter, food, guest. Script knowledge is a form of structured world knowledge that is useful in NLP applications for natural language understanding tasks (e.g., ambi- guity resolution Rahman and Ng, 2012), as well as for psycholinguistic models of human language pro- cessing, which need to represent event knowledge to model human expectations (Zwaan et al., 1995; Sch¨utz-Bosbach and Prinz, 2007) of upcoming ref- erents and utterances. One recent line of research has tried to learn scripts in an unsupervised way from large text col- lections. The core idea in Chambers and Jurafsky (2008, 2009); Jans et al. (2012) is to use corefer- ence chains to identify events involving the same en- tity, with the intuition that these events would, if ob- served in many texts, be likely to represent a proto- typical event sequence. Rudinger et al. (2015) show that this method is also applicable for learning spe- cific targeted scripts from a domain-specific corpus, shown at the example of “Dinners From Hell” sto- ries and the restaurant script. Pichotta and Mooney (2014) (P&M) have demon- strated that using richer event representations con- taining multiple arguments improves prediciton ac- curacy on the narrative cloze task over the simpler models by Chambers and Jurafsky (2008). While they represent a script event as a pair of a verb and a dependency (an example of an event chain would be <call,obj>; <bring,subj>; <take,subj>), which is problematic for weak verbs and verb ambigu- ity, P&M represent events using a multi-argument event representation, e.g., call(guest,waiter,*); bring(waiter,menu,*); take(waiter,order,*). This richer event representation however still has some shortcomings. As the representation is based on coreference chains, the model runs into diffi- culties for entities that are in a chain of length one. Entities in a chain are internally mapped onto variables, but all single entities are mapped onto a common category Other. This means that all information about such referents is lost, e.g. enjoy(customer, fish, ∗) can not be distin- guished from enjoy(customer, silence, …

Content truncated to 500 words