ZettlemoyerTalk110510 < Main

---+++ The Treehouse presents:
---++++ Luke Zettlemoyer, University of Washington CSE
---+++++ November 5, 2010 12:30-1:20 pm

Recent research has focused on the problem of learning to map natural<br /> language sentences to semantic representations of their underlying<br /> meaning. Most of this work has focused on analyzing sentences in<br /> isolation. In this talk, I will present a new approach for learning<br /> context-dependent mappings from sentences to logical form.<br /> <br /> In this setup, the training examples are sequences of sentences<br /> annotated with lambda-calculus meaning representations. I will present<br /> an algorithm that maintains explicit, lambda-calculus representations<br /> of salient discourse entities and uses a context-dependent analysis<br /> pipeline to recover logical forms. The method uses a hidden-variable<br /> variant of the perception algorithm to learn a linear model used to<br /> select the best analysis. Experiments on context-dependent utterances<br /> from the ATIS travel-planning corpus show that the approach recovers<br /> fully correct logical forms with 83.7% accuracy.

This topic: Main > WebHome > QuarterlySchedule > ZettlemoyerTalk110510
Topic revision: r1 - 2010-11-03 - ebender