1996 SALS-SIG Seminars
SALS-SIG Research Seminar
What I'll Probably Say and What I'll Probably Mean
MRI Language Technology Group, Department of Computing, Macquarie University
When: Tuesday, 4th June 1996
Where: Room E6A102, Macquarie University
This talk describes one of the major strands of my recently submitted doctoral thesis: the meaning distributions theory. It is aimed at a general audience and should be accessible with little or no knowledge of natural language processing research. For the experts, the meaning distributions theory advocates the use of probabilistic conceptual models in statistical language learning, that is semantic forms are assigned probabilities independent of grammar and lexis, and suggests a working approximation, namely that grammar and lexis are constrained by the conceptual representation but are otherwise uniformly distributed. For everyone else what this means is that existing work on automatically learning language has focussed on evaluating how likely a given _sentence_ is, while this work emphasises how likely a given _meaning_ is. After explaining the workings of existing statistical language learners and of the meaning distributions theory, I will describe an application of the theory to the problem of parsing noun compounds. I'll then report experiments on both human and machine performance on this task, the results of which support the theory. The moral of the story is: Learn to mean before you try to talk. Not exactly a gripping plot, but maybe the special effects will carry it at the box office.
|Last modified: July 1997|