Modelling Semantic Role Pausibility in Human Sentence Processing

Modelling Semantic Role Plausibility in Human Sentence Processing Ulrike Padó and Matthew Crocker Computational Linguistics Saarland University 66041 Saarbrücken Germany {ulrike,crocker}@coli.uni-sb.de Frank Keller School of Informatics University of Edinburgh 2 Buccleuch Place Edinburgh EH8 9LW, UK keller@inf.ed.ac.uk Abstract We present the psycholinguistically moti- vated task of predicting human plausibility judgements for verb-role-argument triples and introduce a probabilistic model that solves it. We also evaluate our model on the related role-labelling task, and com- pare it with a standard role labeller. For both tasks, our model benefits from class- based smoothing, which allows it to make correct argument-specific predictions de- spite a severe sparse data problem. The standard labeller suffers from sparse data and a strong reliance on syntactic cues, es- pecially in the prediction task. 1 Introduction Computational psycholinguistics is concerned with modelling human language processing. Much work has gone into the exploration of sen- tence comprehension. Syntactic preferences that unfold during the course of the sentence have been successfully modelled using incremental proba- bilistic context-free parsing models (e.g., Jurafsky, 1996; Crocker and Brants, 2000). These models assume that humans prefer the most likely struc- tural alternative at each point in the sentence. If the preferred structure changes during processing, such models correctly predict processing difficulty for a range of experimentally investigated con- structions. They do not, however, incorporate an explicit notion of semantic processing, while there are many phenomena in human sentence process- ing that demonstrate a non-trivial interaction of syntactic preferences and semantic plausibility. Consider, for example, the well-studied case of reduced relative clause constructions. When incre- mentally processing the sentence The deer shot by the hunter was used as a trophy, there is a local ambiguity at shot between continuation as a main clause (as in The deer shot the hunter) or as a re- duced relative clause modifying deer (equivalent to The deer which was shot .. .). The main clause continuation is syntactically more likely. However, there is a second, semantic clue pro- vided by the high plausibility of deer being shot and the low plausibility of them shooting. This influences readers to choose the syntactically dis- preferred reduced relative reading which interprets the deer as an object of shot (McRae et al., 1998). Plausibility has overridden the syntactic default. On the other hand, for a sentence like The hunter shot by the teenager was only 30 years old, se- mantic plausibility initially reinforces the syntac- tic main clause preference and readers show diffi- culty accommodating the subsequent disambigua- tion towards the reduced relative. In order to model effects like these, we need to extend existing models of sentence process- ing by introducing a semantic dimension. Pos- sible ways of integrating different sources of in- formation have been presented e.g. by McRae et al. (1998) and Narayanan and Jurafsky (2002). Our aim is to formulate a model that reliably pre- dicts human plausibility judgements from corpus resources, in parallel to the standard practice of basing the syntax component of psycholinguistic models on corpus probabilities or …

Content truncated to 500 words