Flexible, Corpus-Based Modelling of Human Plausibility Judgements

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 400–409, Prague, June 2007. c⃝2007 Association for Computational Linguistics Flexible, Corpus-Based Modelling of Human Plausibility Judgements Sebastian Padó and Ulrike Padó Computational Linguistics Saarland University Saarbrücken, Germany {pado,ulrike}@coli.uni-sb.de Katrin Erk Dept. of Linguistics University of Texas at Austin Austin, Texas katrin.erk@mail.utexas.edu Abstract In this paper, we consider the computational modelling of human plausibility judgements for verb-relation-argument triples, a task equivalent to the computation of selectional preferences. Such models have applications both in psycholinguistics and in computa- tional linguistics. By extending a recent model, we obtain a completely corpus-driven model for this task which achieves significant correlations with human judgements. It rivals or exceeds deeper, resource-driven models while exhibit- ing higher coverage. Moreover, we show that our model can be combined with deeper mod- els to obtain better predictions than from ei- ther model alone. 1 Introduction One fundamental and intuitive finding in experimen- tal psycholinguistics is that humans judge the plau- sibility of a verb-argument pair vastly differently de- pending on the semantic relation in the pair. Table 1 lists example human judgements which McRae et al. (1998) elicited by asking about the plausibility of, e.g., a hunter shooting (relation agent) or being shot (relation patient). McRae et al. found that “hunter” is judged to be a very plausible agent of “shoot” and an implausible patient, while the reverse is true for “deer”. In linguistics, this phenomenon is explained by selectional preferences on verbs’ argument po- sitions; we use plausibility and fit with selectional preferences interchangeably. Verb Relation Noun Plausibility shoot agent hunter 6.9 shoot patient hunter 2.8 shoot agent deer 1.0 shoot patient deer 6.4 Table 1: Verb-relation-noun triples with plausibility judgements on a 7-point scale (McRae et al., 1998) In this paper, we consider computational mod- els that predict human plausibility ratings, or the fit of selectional preferences and argument, for such (verb, relation, argument), in short, (v, r, a), triples. Being able to model this type of data is rel- evant in a number of ways. From the point of view of psycholinguistics, selectional preferences have an important effect in human sentence processing (e.g., McRae et al. (1998), Trueswell et al. (1994)), and models of selectional preferences are therefore nec- essary to inform models of this process (Padó et al., 2006). In computational linguistics, a multitude of tasks is sensitive to selectional preferences, such as the resolution of ambiguous attachments (Hindle and Rooth, 1993), word sense disambiguation (McCarthy and Carroll, 2003), semantic role labelling (Gildea and Jurafsky, 2002), or testing the applicability of inference rules (Pantel et al., 2007). A number of approaches has been proposed to model selectional preference data (Padó et al., 2006; Resnik, 1996; Clark and Weir, 2002; Abe and Li, 1996). These models generally operate by general- ising from seen (v, r, a) triples to unseen ones. By relying on resources like corpora with semantic role annotation or the WordNet ontology, these models 400 …

Content truncated to 500 words