A Hierarchical Bayesian Model for Unsupervised Induction of Script Knowledge

Proceedings of the 14th Conference of the European Chapter of the Association for Computational Linguistics, pages 49–57, Gothenburg, Sweden, April 26-30 2014. c⃝2014 Association for Computational Linguistics A Hierarchical Bayesian Model for Unsupervised Induction of Script Knowledge Lea Frermann1 l.frermann@ed.ac.uk Ivan Titov2 titov@uva.nl 1 ILCC, School of Informatics, University of Edinburgh, United Kingdom 2 ILLC, University of Amsterdam, Netherlands 3 Department of Computational Linguistics, Saarland University, Germany Manfred Pinkal3 pinkal@coli.uni-sb.de Abstract Scripts representing common sense knowledge about stereotyped sequences of events have been shown to be a valu- able resource for NLP applications. We present a hierarchical Bayesian model for unsupervised learning of script knowledge from crowdsourced descriptions of human activities. Events and constraints on event ordering are induced jointly in one unified framework. We use a statistical model over permutations which captures event ordering constraints in a more flexible way than previous approaches. In order to alleviate the sparsity problem caused by using relatively small datasets, we incorporate in our hierarchical model an informed prior on word distributions. The resulting model substantially outperforms a state-of-the-art method on the event ordering task. 1 Introduction A script is a “predetermined, stereotyped se- quence of actions that define a well-known sit- uation” (Schank and Abelson, 1975). While humans acquire such common-sense knowledge over their lifetime, it constitutes a bottleneck for many NLP systems. Effective question answer- ing and summarization are impossible without a form of story understanding, which in turn has been shown to benefit from access to databases of script knowledge (Mueller, 2004; Miikkulainen, 1995). Knowledge about the typical ordering of events can further help assessing document co- herence and generating coherent text. Here, we present a general method for acquiring data bases of script knowledge. Our work may be regarded as complementary to existing work on learning script knowledge from natural text (cf. (Chambers and Jurafsky, 2008)), as not all types of scripts are elaborated in natural text – being left implicit because of assumed read- ers’ world knowledge. Our model, operating on data obtained in a cheap way by crowdsourcing, is applicable to any kind of script and can fill this gap. We follow work in inducing script knowl- edge from explicit instantiations of scripts, so- called event sequence descriptions (ESDs) (Reg- neri et al., 2010). Our data consists of sets of ESDs, each set describing a well-known situation we will call scenario (e.g., “washing laundry”). An ESD consists of a sequence of events, each describing an action defining part of the scenario (e.g., “place the laundry in the washing machine”). We refer to descriptions of the same event across ESDs as event types. We refer to entities involved in a scenario as participants (e.g., a “washing ma- chine” or a “detergent”), and to sets of participant descriptions describing the same entity as partici- pant types. For each type of scenario, our model clusters descriptions which refer to the same type of event, and infers constraints on the temporal order in which the events types occur in a particular …

Content truncated to 500 words