Guest Speakerer

Effective Probabilistic Models for Syntactic and Semantic Disambiguation

Kristina Toutanova
Stanford University

Thurs., Mar. 3rd
3:30pm; 480 Dreese Labs
All interested parties are invited.
Refreshments will be served immediately preceding the talk.

Abstract:
This talk presents statistical classifiers for sophisticated syntactic and semantic natural language representations. In natural language processing, many important problems can be formulated as disambiguation problems. The goal is to disambiguate or classify an ambiguous natural language object in context. For solving high-level problems, such as question answering, the labels are complex structures - syntactic parse trees and semantic role structures. This entails new challenges for classification algorithms.

I advance the state of the art in several domains by (i) choosing representations that encode domain knowledge more effectively and (ii) developing machine learning algorithms that deal with the specific properties of linguistic disambiguation tasks - sparsity of training data and large, structured spaces of hidden labels.

I focus on three tasks. For syntactic disambiguation, I present a novel representation of parse trees that connects the words of the sentence with the hidden syntactic structure in a direct way and leads to natural definition of tree kernels via convolution of string kernels. For disambiguation of the semantic role structure of verbs, I describe a model that captures regularities of the label space and incorporates the knowledge that the semantic frame of a verb is a joint structure with strong
dependencies between arguments. Finally, I present a method for combination
of multiple knowledge sources for estimation of lexical distributions, based on construction of random walks in the space of words.

Host: Donna Bryon