Regina Barzilay to speak, Thursday, March 11th

[Joseph Turian]

When: 6:10 pm, Thursday March 11th
Where: 7th Floor Interschool Lab, Schapiro Engineering Research Lab (by
my office, 722)
Directions: http://www.cs.columbia.edu/resources/directions
Speaker: Regina Barzilay, MIT

Welcome back one of our own graduates!

            Probabilistic Approaches for Modeling Text Structure and
            their application to Text-to-Text Generation

                          Regina Barzilay (MIT)


   Text-to-text generation aims to produce a coherent text by
   extracting, combining and rewriting information given in input
   texts. Examples of its applications include summarization, answer
   fusion in question-answering and text simplification.  In this
   talk, I will present models of document structure that can be
   effectively used to guide content selection in text-to-text
   generation.  First, I will focus on unsupervised learning of
   domain-specific content models.  These models capture the topics
   addressed in a text, and the order in which these topics appear.
   I will present an effective method for learning these models from
   unannotated domain-specific documents, utilizing hierarchical
   Bayesian methods. Next, I will present a method for assessing the
   coherence of a generated text.  The key premise of our work is
   that the distribution of entities in coherent texts exhibits
   certain regularities. Our algorithms can learn these patterns from
   raw texts, without recourse to manual annotation or a predefined
   knowledge base. Finally, I will show how these models can be
   effectively integrated in text-to-text applications such as
   automatic generation of Wikipedia articles.