Transcript High-Precision Natural Language Interfaces: A Graph

High-Precision Natural Language
Interfaces: A Graph-Theoretic
Approach
Ana-Maria Popescu, Oren Etzioni, Henry Kautz
February 26th, 2002
Research Overview
Goal
Develop high-precision natural-language
interfaces to relational databases
Semantic Interpretation Challenge
Correctly identifying the database tokens
referred to by a given sentence
February 26th, 2002
Approach
Easy Questions
Natural subset of English accurately interpreted as
non-recursive Datalog clauses
What are the Chinese restaurants in Seattle ?
What is the population of Washington ?
What Texas jobs require 3 years of experience ?
February 26th, 2002
Approach
Semantic Interpretation: Graph Matching
Use max-flow as basis of sound, polynomial-time
procedure for semantic interpretation
Incorporate Syntactic Information
Use parser or tagger output to eliminate ambiguity
February 26th, 2002
Approach
Given a sentence q:
• Identify potential database elements
• Use semantic and syntactic constraints
to create semantic interpretation of q
• Generate final SQL query(queries)
February 26th, 2002
Database
Lexicon
Parser
Tokenizer
Matcher
Query Generator
Equivalence Checker
English Question
SQL Query + Answer Set
PRECISE Architecture
February 26th, 2002
DB Values
Value
Tokens
S
DB Attributes
Movies = what
Movies
What
Actor = Woody Allen
Actor
Woody Allen
Director = Woody Allen
Director
Attribute
Tokens
film
E
I
K
2
Paul Mazursky
Director = Paul Mazursky
The graph created by PRECISE for the question
“What are the Paul Mazursky films with Woody Allen ?”
February 26th, 2002
Experimental Results
Systems: PRECISE,Mooney,EnglishQuery
Dataset: 3 databases (Mooney et. all, 2001)
Results:
• PRECISE achieves the lowest precision error
rate amongst the three systems
• Current work focuses on improving the recall
failure rate
//add 2 slides containing result graphs
February 26th, 2002
Future Work
Directions
• Extend PRECISE to handle additional types
of user queries
• Use clarification dialogues
• Learn unknown words
• Add speech-input/output capabilities
February 26th, 2002