Veranstaltung am

The role of discreteness in gradient inference judgments: A case study of lexically triggered inferences


Similar to judgments of natural language strings' well-formedness, judgments of strings'inferential relationships display substantial amounts of gradience when aggregated across responses collected in formal experiments. One area where such gradience presents itself is among inferences that are classically analyzed to be conditioned by some form of discrete lexical semantic representation–e.g. some veridicality inferences, such as the inference from Jo {liked, didn't like} that Bo left to Bo left, may be driven by knowledge that like is an emotive factive predicate. Such gradience has recently led some researchers to argue that the classical picture is wrong: no such discrete lexical semantic representations exist, and we should rather seek explanations that rely mainly on pragmatic reasoning to generate the inferences on which these judgments are based.

In this talk, I argue that, while pragmatic reasoning clearly has an important role to play in modulating these inferences, the classical picture remains largely correct when augmented with a minimalistic, modular probabilistic semantics. In the first part of the talk, I present evidence that, when appropriately modeling nuisance variables' contributions to noise in inference judgments, something very close to classical classifications of predicates in terms of their lexically triggered inferences is revealed. In the second part of the talk, I present evidence that, when comparing inference judgment models that assume responses are driven by some discrete representation or process conditioned by the lexical item to models that assume the representation or process is fundamentally gradient, those models that assume discreteness fit judgment data reliably better in quantitative model comparison. I take these two pieces of evidence together to suggest that the role of pragmatic reasoning in modulating (at least these sorts of) lexically triggered inferences is to resolve indeterminacy among linguistic expressions that condition these inferences, rather than to generate the inferences themselves. 


Ort: Campus Augustusplatz, Hörsaalgebäude, HS 10


Weitere Termine im Kolloquium finden Sie auf der Veranstaltungsseite.