Modelling selectional preferences in a lexical hierarchy

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 170–179, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Modelling selectional preferences in a lexical hierarchy Diarmuid ´O S´eaghdha Computer Laboratory University of Cambridge Cambridge, UK do242@cam.ac.uk Anna Korhonen Computer Laboratory University of Cambridge Cambridge, UK Anna.Korhonen@cl.cam.ac.uk Abstract This paper describes Bayesian selectional preference models that incorporate knowledge from a lexical hierarchy such as WordNet. In- spired by previous work on modelling with WordNet, these approaches are based either on “cutting” the hierarchy at an appropriate level of generalisation or on a “walking” model that selects a path from the root to a leaf. In an evaluation comparing against human plau- sibility judgements, we show that the mod- els presented here outperform previously pro- posed comparable WordNet-based models, are competitive with state-of-the-art selectional preference models and are particularly well- suited to estimating plausibility for items that were not seen in training. 1 Introduction The concept of selectional preference captures the intuitive fact that predicates in language have a bet- ter semantic “fit” for certain arguments than oth- ers. For example, the direct object argument slot of the verb eat is more plausibly filled by a type of food (I ate a pizza) than by a type of vehicle (I ate a car), while the subject slot of the verb laugh is more plausibly filled by a person than by a veg- etable. Human language users’ knowledge about selectional preferences has been implicated in anal- yses of metaphor processing (Wilks, 1978) and in psycholinguistic studies of comprehension (Rayner et al., 2004). In Natural Language Processing, au- tomatically acquired preference models have been shown to aid a number of tasks, including semantic role labelling (Zapirain et al., 2009), parsing (Zhou et al., 2011) and lexical disambiguation (Thater et al., 2010; ´O S´eaghdha and Korhonen, 2011). It is tempting to assume that with a large enough corpus, preference learning reduces to a simple lan- guage modelling task that can be solved by counting predicate-argument co-occurrences. Indeed, Keller and Lapata (2003) show that relatively good perfor- mance at plausibility estimation can be attained by submitting queries to a Web search engine. How- ever, there are many scenarios where this approach is insufficient: for languages and language domains where Web-scale data is unavailable, for predicate types (e.g., inference rules or semantic roles) that cannot be retrieved by keyword search and for ap- plications where accurate models of rarer words are required. ´O S´eaghdha (2010) shows that the Web- based approach is reliably outperformed by more complex models trained on smaller corpora for less frequent predicate-argument combinations. Models that induce a level of semantic representation, such as probabilistic latent variable models, have a further advantage in that they can provide rich structured in- formation for downstream tasks such as lexical dis- ambiguation ( ´O S´eaghdha and Korhonen, 2011) and semantic relation mining (Yao et al., 2011). Recent research has investigated the potential of Bayesian probabilistic models such as Latent Dirichlet Allocation (LDA) for modelling …

Content truncated to 500 words