Fecha: 28 de abril de 2011
Lugar de celebración: Sala 2.24, Facultad de Psicología, UNED
This tutorial will introduce a compositional distributional model of meaning and will explain the mathematical concepts behind it. The model, jointly developed in Oxford and Cambridge, provides a compositional theory of meaning that computes meanings of sentences as vectors and is able to measure their similarity. The ingredients include a type-logic for formalizing the grammatical structure (called a pregroup), and a vector space semantics for lexical meaning. I will introduce the key concepts of each setting and go through simple fun examples to illustrate their applicability. I will also work through sample computations which parse the sentence, assign meaning to words, then compose them to obtain meaning of sentences.
Edward Grefenstette: Empirical Validation of Compositional Distributional Models of Meaning (12h30 - 14h00)
Modelling compositional meaning for sentences using empirical distributional methods has been a challenge for computational linguists. We implemented the abstract categorical model discussed in the previous talk using data from the BNC and evaluated it. In this talk, we will present the implementation, based on unsupervised learning of matrices for relational words and applying them to the vectors of their arguments. We'll also discuss the results of the evaluation, based on the word disambiguation task developed by Mitchell and Lapata (2008) for intransitive sentences, and on a similar new experiment designed for transitive sentences. We'll show our model to match the results of its competitors in the first experiment, and better them in the second. The general improvement in results with increase in syntactic complexity showcases the compositional power of our model.