Compositional Distributional Models of Meaning

The problem of representing natural language meaning in a computationally tractable way has been approached in various and often mutually incompatible ways. Two such orthogonal classes of semantic representation are distributional and symbolic models of meaning. Distributional models of meaning exploit the co-occurrence of other terms with the term being modelled to determine its semantic content, applying Firth's well-known dictum that "You shall know a word by the company it keeps"; while the symbolic approach usually exploits grammatical features of sentences to model relations between entities in the world, thereby expressing the truth conditions of sentences.

These approaches seem theoretically orthogonal, and hence incompatible in implementation. Distributional models of meaning are quantitative and express the semantic relation between terms but offering no immediately obvious way of modelling the contribution of sentence structure to meaning; while typically the semantics of individual words in qualitative symbolic approaches is left "mysterious" and ill-defined.

This project seeks to reconcile these apparently orthogonal representations, guided by the intuition that lexical co-relation and grammatical roles both correspond to different ways of "knowing how to use" language, thus both aspects provide relevant information when we understand the meaning of an expression, and hence they are viable components for a new representation modelling natural language semantics. Category theoretic methods from quantum information theory provide the building blocks for a general framework in which to combine symbolic and distributional approaches to expressing and comparing the meaning of sentences.

In bringing together expertise from the fields of formal logic, philosophy of language, computational linguistics, pure mathematics and theoretical physics, this project aims to produce a new model of natural language semantics for use in the development of more sophisticated text processing applications, document retrieval systems, intelligent agents, etc.

• Stephen Clark (Cambridge)