A Robust and Extensible Exemplar-Based Model of Thematic Fit

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 826–834, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics A robust and extensible exemplar-based model of thematic fit Bram Vandekerckhovea, Dominiek Sandraa, Walter Daelemansb aCenter for Psycholinguistics, bCenter for Dutch Language and Speech (CNTS) University of Antwerp Antwerp, Belgium {bram.vandekerckhove,dominiek.sandra,walter.daelemans}@ua.ac.be Abstract This paper presents a new, exemplar-based model of thematic fit. In contrast to pre- vious models, it does not approximate thematic fit as argument plausibility or ‘fit with verb selectional preferences’, but directly as semantic role plausibility for a verb-argument pair, through similarity- based generalization from previously seen verb-argument pairs. This makes the model very robust for data sparsity. We argue that the model is easily extensible to a model of semantic role ambiguity reso- lution during online sentence comprehen- sion. The model is evaluated on human seman- tic role plausibility judgments. Its predic- tions correlate significantly with the hu- man judgments. It rivals two state-of-the- art models of thematic fit and exceeds their performance on previously unseen or low- frequency items. 1 Introduction Thematic fit (or semantic role plausibility) is the plausibility of a noun phrase referent playing a specific semantic role (like agent or patient) in the event denoted by a verbal predicate, e.g. the plausibility that a judge sentences someone (which makes the judge the agent of the sentencing event) or that a judge is sentenced him- or herself (which makes the judge the patient). Thematic fit has been an important concept in psycholinguistics as a pre- dictor variable in models of human sentence com- prehension, either to discriminate between pos- sible structural analyses during initial processing in constraint-based models (see MacDonald and Seidenberg (2006) for a recent overview), or af- ter initial syntactic processing in modular models (e.g. Frazier (1987)). In fact, thematic fit is at the core of the most-studied of all structural ambiguity phenomena, the ambiguity between a main clause or a reduced relative clause interpretation of an NP verb-ed sequence (the MV/RR ambiguity), which is essentially a semantic role ambiguity. If the temporarily ambiguous sentence The judge sen- tenced ... is continued as a main clause (e.g. The judge sentenced him to 10 years in prison), the noun phrase the judge would be the agent of the verb sentenced, while it would be the patient of sentenced in a reduced relative clause continuation (e.g. The judge sentenced to 4 years in prison for indecent exposure could also lose his state pen- sion). Apart from its importance in psycholinguis- tics, the concept of thematic fit is also relevant for computational linguistics in general (see Pad´o et al. (2007) for some examples). A number of models that try to capture hu- man thematic fit preferences have been developed in recent years (Resnik, 1996; Pad´o et al., 2006; Pad´o et al., 2007). These previous approaches rely on the linguistic notion of verb selectional pref- erences. The plausibility that an argument plays a specific semantic role in the …

Content truncated to 500 words