Logical Metonymy in a Distributional Model of Sentence Comprehension

In theoretical linguistics, logical metonymy is defined as the combination of an event-subcategorizing verb with an entity-denoting direct object (e.g., The author began the book), so that the interpretation of the VP requires the retrieval of a covert event (e.g., writing). Psycholinguistic studies have revealed extra processing costs for logical metonymy, a phenomenon generally explained with the introduction of new semantic structure. In this paper, we present a general distributional model for sentence comprehension inspired by the Memory, Unification and Control model by Hagoort (2013, 2016). We show that our distributional framework can account for the extra processing costs of logical metonymy and can identify the covert event in a classification task. 1 Logical Metonymy: Psycholinguistic Evidence and Computational Modeling The interpretation of so-called logical metonymy (e.g, The student begins the book) has received an extensive attention in both psycholinguistic and linguistic research. The phenomenon is extremely problematic for traditional theories of compositionality (Asher, 2015) and is generally explained as a type clash between an eventselecting metonymic verb (e.g., begin) and an entity-denoting nominal object (e.g., the book), which triggers the recovery of a hidden event (e.g., reading). Past research work brought extensive evidence that such metonymic constructions also determine extra processing costs during online sentence comprehension (McElree et al., 2001; Traxler et al., 2002), although such evidence is not uncontroversial (Falkum, 2011). According to Frisson and McElree (2008), event recovery is triggered by the type clash, and the extra processing load is due to ”the deployment of operations to construct a semantic representation of the event”. Thus, logical metonymy raises two major questions: i.) How is the hidden event recovered? ii.) What is the relationship between such mechanism and the increase in processing difficulty? One of the first accounts of the phenomenon dates back to the works of Pustejovsky (1995) and Jackendoff (1997), which assume that the covert event is retrieved from complex lexical entries consisting of rich knowledge structures (Pustejovsky’s qualia roles). For example, the representation of a noun like book includes telic properties (the purpose of the entity, e.g. read) and agentive properties (the mode of creation of the entity, e.g. write). The predicate-argument type mismatch triggers the retrieval of a covert event from the object noun qualia roles, thereby producing a semantic representation equivalent to begin to write the paper (see also the discussion in Traxler et al. (2002)). On the one hand, the lexicalist explanation is very appealing, since it accounts for the existence of default interpretations of logical metonymies (e.g. begin the book is typically interpreted as begin reading/writing the book). On the other hand, Lascarides and Copestake (1998) and more recently Zarcone et al. (2014) show that qualia roles are simply not flexible enough to account for the wide variety of interpretations that can be retrieved. These are in fact affected by the subject choice, the general syntactic and discourse context, and by our world knowledge. 1 Consider the classical example from Lascarides and Copestake (1998): My goat eats anything. He really enjoys