Probabilistic approaches to syntax and semantics
Instructor: Shalom Lappin
Time: 11:00 - 12:30
Location: Tydings, Room 1114
Course Description: In this course I will present current research on probabilistic representations of syntactic and semantic knowledge. The first part of the course will focus on the use of enriched language models to track speakers' judgements of grammaticality. In these models grammaticality is not directly reduced to probability of utterance. Degree of well-formedness is computed by a normalising function on probability distributions, where this function filters out certain properties of the probability distribution, while highlighting others. We will also look at experimental work on determining whether grammaticality is a binary categorical or a gradient property.
The second part of the course will consider a rich probabilistic type theory as the basis for a compositional semantics for natural language. A central element of this theory is that the meaning of a declarative sentence is modeled as a judgement that a situation is of a specified type, where this judgement has a probability value. Probabilities are distributed over a range of situation types. These types are intensional, but neither intensions nor modality is characterised in terms of possible worlds. The probabilistic type theory provides the interface between Bayesian learning of classifiers for perceptual judgements and the interpretation of sentences in natural language.