| Description |
Distributional semantics (DS) models have gained popularity in computational semantics for their ability to produce rich computational representations of concepts, derivable from observable language use. It has been argued, however, that DS does not quite yet deserve the name of \'semantics\'. Indeed, its focus on the conceptual level has made it mostly unsuitable for tackling traditional problems in the study of meaning -- in particular, the question of reference. In this talk, I will ask to what extent DS can provide a model of singular entities, and thereby connect with formal, set-theoretic semantics and correspondence theories of meaning. I will start with the classic case of proper names and propose that distributions of individuals can be built through a function acting on the conceptual level of standard distributional representations. This will be illustrated by an empirical study over the proper names of two novels, which shows that the resulting entities have several desired properties of individuals, from the way they instantiate relevant concepts to high semantic coherence. Preliminary results also indicate that such \'entity distributions\' can be integrated in a plausible model of plurality.
|
