Latent-variable learning for natural language parsing and translation
Thursday, 21st May, 2015 at 11:00AM
Abstract: In this talk, I will discuss two pivotal aspects of natural language processing: syntax and semantics.
For syntax, I will show how latent-variable modeling can be used to improve the accuracy of syntactic modeling of language, as well as machine translation. Latent-variable learning usually introduces non-convex optimization problems (such as log-likelihood maximization), but I will describe an algorithm which reduces the problem of latent-variable learning to a convex optimization problem. The algorithm is based on the method of moments, and can be shown to yield consistent parameter estimates.
For semantics, I will describe a problem in NLP, that of lexical event ordering. The goal is to order a set of events, given as a bag of events, so that the order describes a temporal relation. I will suggest the domain of cooking recipes as an excellent testbed for such a structured-prediction problem.