posted on 2023-06-08, 20:34authored byDaoud Clarke, Bill Keller
This paper explores theoretical issues in constructing an adequate probabilistic semantics for natural language. Two approaches are contrasted. The first extends Montague Semantics with a probability distribution over models. It has nice theoretical properties, but does not account for the ubiquitous nature of ambiguity; moreover inference is NP hard. An alternative approach is described in which a sequence of pairs of sentences and truth values is generated randomly. By sacrificing some of the nice theoretical properties of the first approach it is possible to model ambiguity naturally; moreover inference now has polynomial time complexity. Both approaches provide a compositional semantics and account for the gradience of semantic judgements of belief and inference.
Funding
A Unified Model of Compositional and Distributional Semantics: Theory and Applications; EPSRC; EP/I037458/1
History
Publication status
Published
File Version
Accepted version
Journal
Proceedings of the 11th International Conference on Computational Semantics