Time: 2:15pm-3:30pm
Place: Room C415A (basement level), CUNY Graduate Center (365 5th Ave, between 34str. & 35str)

Speaker:  Yoav Artzi, University of Washington

Title: Weakly Supervised Learning of Semantic Parsers in Interactive Systems

Situated linguistic interactions provide many opportunities for autonomous language learning. 
For example, an agent can learn to improve its language facility through trial and error, by 
asking questions and imitating experts. In this talk, I will describe two approaches that use 
situated interactions to learn to map sentences to rich meaning representations. In both cases, 
we show accurate interactive learning while treating meaning as latent, thereby avoiding the 
high cost of data annotation typically required for such semantic analysis tasks. In the first 
instance, we show how a dialog system can learn from failures. Similarly, for an agent 
following instructions, we describe learning through trial and error with joint interpretation 
and execution of instructions. In both instances we induce a log-linear weighted CCG 
semantic parser, which includes both parsing parameters and a lexicon. Such approaches, 
when integrated into complete systems, promise the potential of continuous improvement 
through automatic learning.

Yoav Artzi is a Ph.D. candidate working with Luke Zettlemoyer in the Computer Science & 
Engineering department at the University of Washington, Seattle. His research studies the 
acquisition of grounded natural language understanding within interactive systems. His work 
focuses on modeling semantic representations and designing weakly supervised learning 
algorithms. Prior to that, He  completed a B.Sc. in Computer Science in Tel Aviv University. 
He is a recipient of the 2012 Yahoo Key Scientific Challenge award.