Latent Dirichlet Allocation For Estimating the Probabilities of Bilexical Dependencies
Speaker: Joel Lang
Abstract:
In the talk I will address the problem of estimating the probability of bilexical dependencies – dependencies that hold between two lexical items occurring in natural language text. Obtaining good estimates for such dependencies is central to a range of natural language processing tasks including lexicalised parsing, lexicon induction and analysing selectional preferences. I propose a method making use of Latent Dirichlet Allocation for estimating bilexical probabilities. Compared to previously formulated methods, the proposed method results in significantly lower perplexity scores on test data and yields better results in various evaluation tasks.