home · mobile · calendar · colloquia · 2008-2009 · 

Colloquium - Schuler

Spoken Language Understanding Using an Interactive Language Model
University of Minnesota

Spoken language interfaces for applications like home organizers, reminder systems, or immersive design may need to allow users to create new entities with names not found in existing training corpora. In such cases an "interactive" language model can be used to estimate probabilities of successive words based on attributes and relations over entities that may be described by those words. These interpretations can be defined as vectors, corresponding to distributions over head words or sets of denoted individuals in a world model; then composed with relations, defined as matrices.

Unfortunately, semantic composition is typically understood to conform to syntactic phrase structure, and matrix multiplication is generally too expensive to be practical in conventional cubic-time chart parsers, used to hypothesize this phrase structure. This talk will therefore describe a method for exploiting human-like memory bounds to obtain a worst-case linear bound on the number of composition operations required, times a constant memory store depth, and a constant beam width.

The talk will first present corpus evidence that this constant memory store depth can be extremely small -- consisting of only three to four memory elements -- without sacrificing parsing coverage, in line with estimates of human short-term memory capacity. The talk will then describe an evaluation of a real-time spoken language interface that implements this interactive model, interpreting referents as sets of individuals from a world model domain consisting of 100 individuals, using a vocabulary of 1000 names, attributes, or relations.

Two articles (the former under review, the latter in press) describing the syntax and semantics of this model are available:

William Schuler graduated from the University of Pennsylvania in 2003 with a PhD in Computer and Information Science and is now an Assistant Professor of Computer Science at the University of Minnesota. He has been studying psycholinguistically-motivated models of spoken language understanding for nearly ten years, funded since 2005 by a National Science Foundation CAREER grant. In 2006 his research program was awarded a Presidential Early Career Award for Scientists and Engineers (PECASE) in a ceremony at the White House. In 2007 he was awarded a McKnight Land-Grant Professorship by the University of Minnesota, consisting of a one-year research leave and funds to promote his research.

Hosted by Martha Palmer.

Department of Computer Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
May 5, 2012 (14:13)