Compositional expectation: a purely distributional model of compositional semantics

• Authors: Justin Washtell
• Venue: IWCS '11 Proceedings of the Ninth International Conference on Computational Semantics
• Year: 2011

The past year has witnessed a surge of interest in the issue of compositional semantics: modelling the meaning of complex phrases. To date, distributional approaches have successfully dealt only with the meaning of individual words in context. Recent attempts to address the more general case of compositional meaning have tended to focus either on mathematical models, which have yet to be demonstrated useful in a linguistic setting, or on syntactically-motivated approaches which do not yet permit application to unconstrained text. We present a purely distributional compositional model, based on the simple addition of expectation vectors. Expectation vectors (Washtell, 2010) are particularly appealing from a compositional standpoint as they are naturally sensitive to word-order alterations whilst being insensitive to the substitution of distributionally similar words. We explore the properties of these and two baseline models using datasets based upon human judgements of phrasal similarity. Whilst far from solving the problem of compositionality, our findings raise interesting questions and provide some useful ideas and benchmarks for those tackling this very current problem.