Computing That Serves

Playing with Linguistic Structures


Thursday, April 2, 2015 - 11:00am


Vivek Srikumar


Eric Ringger

Colloquium presented by Vivek Srikumar, Assistant professor in the School of Computing at the University of Utah
Thursday, April 2, 2015 at 11:00 A.M. 
Location: 1170 TMCB


Natural language understanding requires programs that can automatically reason about the meaning of text. By representing the meaning of a document in a structured, symbolic form (such as a graph relating entities and events), we can use ideas from logical inference for reasoning about the text.
In the first part of this talk, I will discuss a new reading comprehension task of automatically answering difficult AP exam style questions based on text from a biology textbook. The questions test an understanding of the events and their relationships expressed in the text. I will show that representing text as a structured semantic graph gives substantial improvement in answering these complex questions.
In the second part of the talk, I will present a new formalism for representing structured output called DISTRO that accounts for the fact that outputs in NLP (e.g. sequences, trees or arbitrary graphs) are designed to encode rich, and often overlapping, semantics. And yet, at training time these objects are typically treated as discrete units of meaning. I will show that allowing semantically similar outputs to share parameters during training can lead to better predictions.


I am an assistant professor in the School of Computing at the University of Utah. Previously, I was a post-doc at Stanford University. I obtained my Ph.D. from the University of Illinois at Urbana-Champaign in 2013. My research lies in the areas of Machine Learning and Natural Language Processing. In particular, I am interested in research questions arising from the need to predict structured representations of text using little or incidental supervision. My work has been published in various NLP and machine learning venues and recently received the best paper award at EMNLP.