Probabilistic Models for Learning a Semantic Parser Lexicon

Explore a 42-minute seminar on probabilistic models for learning a semantic parser lexicon, presented by J. Krishnamurthy at the Paul G. Allen School. Dive into the world of lexicon learning, a crucial first step in training semantic parsers for new application domains. Discover how the proposed probabilistic models, trained directly from question/answer pairs using EM, offer significant improvements over existing heuristic methods. Learn about the simplest model's concave objective function, which guarantees EM convergence to a global optimum. Examine the experimental evaluation on 4th grade science questions, showcasing impressive error reductions and efficiency gains compared to prior work. Explore the competitive results achieved on Geoquery without dataset-specific engineering. The seminar covers various topics, including semantic parser formalism, lexicon building, alignment problems, context-free grammar, and different model approaches such as the Independent Model, Coupled Model, and Pal Model. Gain insights into end-to-end evaluation, parse examples, and semantic types, providing a comprehensive understanding of this innovative approach to semantic parser lexicon learning.

Introduction
Semantic Parser
Diagram Questions
Training
Lexicon
Building a Semantic Parser
Semantic Parser Formalism
A Good Lexicon
Prior Lexicon Learning
Outline
Problem specification
Alignment problem
Contextfree grammar
Summary
Approach
Independent Model
Coupled Model
Pal Model
Standard Application
Parse Example
Semantic Type
Experiments
Example Questions
EndtoEnd Evaluation
Geo Query Results
Summarize
Syntax Categories