Seminar: Ivan Titov

Inducing meaning representations from text is one of the key objectives of NLP. Most of existing statistical techniques for tackling this problem rely on large human-annotated datasets, which are expensive to create and exist only for a very limited number of languages. Even then, they are not very robust, cover only a small proportion of semantic constructions appearing in the labeled data, and are domain-dependent. In this work, we investigate Bayesian models which do not use any labeled data but induce shallow semantic representations from unannotated texts. Unlike semantically-annotated data, unannotated texts are plentiful and available for many languages and many domains which makes our approach particularly promising. We evaluate our approach in two set-ups. First, we experiment with the PropBank corpus, where it achieves the best reported results among unsupervised approaches, then we evaluate it on a question-answering task for the biomedical domain, where it also shows competitive performance.  We also look into several extensions of the model, and specifically consider multilingual induction of semantics, where we show that multilingual parallel data provides a valuable source of indirect supervision for induction of shallow semantic representations.  

Joint work with Alexandre Klementiev.

Bio

Ivan Titov joined the Saarland University as a junior faculty and head of a research group in November 2009, following a postdoc at the University of Illinois at Urbana-Champaign. He received his Ph.D. in Computer Science from the University of Geneva in 2008 and his master's degree in Applied Mathematics and Informatics from the St. Petersburg State Polytechnic University (Russia) in 2003.

His current research interests are in statistical natural language processing (models of syntax, semantics and sentiment) and machine learning (structured prediction methods, latent variable models, Bayesian methods).