Abstract
Logical metonymy (The author began the book) has often been explained in terms of a type-clash between an eventsubcategorizing metonymic verb (begin) and an entity-denoting object (book), triggering the recovery of a covert event (reading). Experiment 3 in Traxler, Pickering and McElree (2002) is often cited as a well-known correlate of this clash, as the authors find a significant interaction between type of verb (metonymic vs. non-metonymic) and type of object (entitydenoting vs. event-denoting). Type clashes and selectional restriction violations are widely invoked in linguistic theory, but more recent work in psycholinguistics suggests that these phenomena might be better captured via graded notions such as typicality and thematic fit. For example, some of the materials of Traxler and colleagues’ seem to favor event-denoting items via a better thematic fit (The pastor finished/prepared the funeral /sandwich).
Similarity-based frameworks such as Distributional Memory and ECU described in Lenci (2011) lend themselves particularly well to modeling the graded effect of thematic fit and expectation about upcoming arguments in sentence processing. We contrast three ECU-based models: verb-only, sum and product. Verb-only exploits expectations coming from the verb to yield the thematic fit for a filler in a given object argument position (e.g. for sandwich and prepare); the other two models do the same by combining expectations from subject and verb (e.g. for sandwich and <pastor, prepare>) by means of two different functions (sum and product). We use the models to (1) mirror the results from Traxler and colleagues; (2) contrast expectations coming from the verb and expectations coming from the composition of subject and verb, in order to evaluate whether the type-clash effect might be due to a cueing effect of the subject; (3) suggest an alternative account of logical metonymy interpretation which re-defines the binary notion of type-clash in more graded terms, i.e. as thematic fit. The models successfully replicated the pattern of results of the psycholinguistic experiment, yielding the lowest thematic fit for metonymic verbs combined with entity-denoting objects, and a main effect of object type (entity-denoting vs. event-denoting; **sum, ***product, ***verb-only), and a significant verb-object interaction (**sum, **verb-only). All models yielded a better thematic fit for event-denoting than for entity-denoting objects on the whole dataset (**sum, ***product, ***verb-only), for metonymic verbs (***sum, **product, ***verb-only), but not for non-metonymic verbs. The pattern of results of the psycholinguistic experiment was replicated not only by models incorporating the subject (sum and product), but also by the verb-only model, thus suggesting the measured effect is not ascribable to an effect of subject only, but is indeed due to the verb-object interaction. The models’ success in replicating the results from the psycholinguistic experiments shows that similarity-based models are an adequate tool to model phenomena such as selectional preferences and logical metonymy, suggesting that they can both be accounted for in terms of thematic fit.
Alessandra Zarcone
Professor of Language Technologies and Cognitive Assistants
Computational linguist with a background in NLP and in psycholinguistics, working on AI, NLP and human-machine interaction.