PISA: A measure of Preference In Selection of Arguments to model verb argument recoverability

Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 131–136 Barcelona, Spain (Online), December 12–13, 2020 131 PISA: A Measure of Preference in Selection of Arguments to Model Verb Argument Recoverability Giulia Cappelli Scuola Normale Superiore giulia.cappelli@sns.it Alessandro Lenci CoLing Lab University of Pisa alessandro.lenci@unipi.it Abstract Our paper offers a computational model of the semantic recoverability of verb arguments, tested in particular on direct objects and In- struments. Our fully distributional model is intended to improve on older taxonomy-based models, which require a lexicon in addition to the training corpus. We computed the selec- tional preferences of 99 transitive verbs and 173 Instrument verbs as the mean value of the pairwise cosine similarity between their argu- ments (a weighted mean between all the argu- ments, or an unweighted mean with the top- most k arguments). Results show that our model can predict the recoverability of objects and Instruments, providing a similar result to that of taxonomy-based models but at a much cheaper computational cost. 1 Introduction Verb meaning, together with pragmatic and discourse-related factors, has long been identified as playing a major role in determining argument optionality (Levin, 1993). Here, we specifically fo- cus on the semantic recoverability of arguments, which has often been claimed to be the key determi- nant of object omission (Jespersen, 1927; Hopper and Thompson, 1980; Levin, 1993; Resnik, 1993, 1996; Conklin et al., 2004; Medina, 2007; Glass, 2020). Resnik (1993, 1996) link argument recov- erability to selectional preferences by means of an experiment we detail in Section 2, and also show that it is correlated with plausibility and typicality judgments provided by human subjects. The relation linking a verb to its optional argument is a grammatical function in Resnik (1993, 1996), such as “subject” or “direct object”, but it may also be a semantic role, such as “Instrument” or “Pa- tient”. The choice between the two depends on This work is licensed under a Creative Commons Attri- bution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/. computational requirements rather than on theoret- ical constraints. Let us consider some examples. Recoverability affects the grammaticality of the sentences in (1), since the Patient of an eating event (as in (1a)) is easily recoverable as a member of the class of edi- bles, but there are no such strict constraints on what one can make (as in (1b)). (1) a. John ate ∅object. b. *John made ∅object. The same applies to prepositional phrases filling the Instrument role, albeit in a more nuanced fash- ion. Koenig et al. (2002, 2003, 2007) have shown that verbs describing actions may be divided into two classes based on whether they semantically require an Instrument (as in (2a)) or they merely allow it (as in (2b)). Require-Instrument verbs se- lect for a smaller range of Instruments than Allow- Instrument verbs: For instance, a beheading event is very likely to involve a heavy bladed tool like a sword, while a killing event may happen with a much larger set of tools …

Content truncated to 500 words