Transcript Slide 1

Faculty of
Computer
Science
Tony Abou-Assaleh, Nick Cercone, and Vlado Kešelj
{taa,nick,vlado}@cs.dal.ca
INTRODUCTION
EVALUATION
Abstract
Motivation
Environment
– Head-driven Phrase Structure Grammar (HPSG) can represent
both syntax and semantics
– Question Answering (QA) plays an important role in Natural
Language Processing (NLP)
– Relaxed Unification enables HPSG to process data that is
inconsistent, incomplete, and uncertain.
Objectives
– Identify limitations of current QA systems
– Develop a framework to overcome these limitations
– Build a QA advanced prototype
– Objectively evaluate the prototype against other approaches
RELAXED UNIFICATION
Question answering is an important task of natural language
processing. Unification-based grammars have emerged as
formalisms for reasoning about natural text. Head-driven
Phrase Structure Grammar is a unification-based grammar that
is capable of representing syntax and semantics in a uniform
fashion allowing one to reason about syntax and semantics
simultaneously. Classical unification, on which the unificationbased grammars rely, is strict in the sense that it requires a
perfect agreement between the terms being unified. In practise,
data is seldom error-free and can contain inconsistent
information. The theory of relaxed unification is a new theory
that relaxes the constraints of the classical unification. Relaxed
unification tolerates errors and inconsistencies in the data and
facilitates reasoning under uncertainty. We use the relaxed
unification formalism in the head-driven phrase structure
grammar to build a question-answering advanced prototype
that addresses some of the pitfalls of the current questionanswering systems.
– Text REtrieval Conference (TREC) Question Answering Track.
– Open-domain, closed-class questions.
– Shallow semantic analysis is some times required.
– Restricted QA: only factoid and definition questions.
Approach
– Use an Information Retrieval (IR) search engine to locate
relevant passages.
– Use HPSG to parse the question and generate a query.
– Use HPSG to parse the passage.
– Use relaxed unification to locate a query match in the passage.
– Compute the correctness of the result using a metric to rank
the answers.
Input
Classical Unification
– Two terms, t and u, are unifiable
if and only if there exists a
substitution

such
that
t  = u .
HPSG AND QUESTION ANSWERING
Documents
Collection
– “How much could you rent a
Volkswagen bug for in 1966?”
Relevant Text
– “… you could rent a Volkswagen
bug for $1 a day.”
– two terms, t and u, are always
unifiable
with
a
unifying
substitution

such
that
t  = u .
– Application-specific correctness
measure evaluates the result.
Ranked
Answers
Question
Query
Relaxed Unification
– Relaxed Term consists of
variables and sets of functions,
where sets can be empty.
Function arguments are relaxed
terms.
E.g. { f ( { a } , { h ( { b } ) } ) }.
Output
Answer
– “$1 a day.”
Extraction
IR Search
Engine
HPSG
Parser
Relevant
Passages
Query
HPSG
Parser
Ranked
Matches
Evaluation
Parsed
Passages
Relaxed
Unification
Matches
References
Abou-Assaleh, Tony and Cercone, Nick and Keselj, Vlado. Towards the
Theory of Relaxed Unification. In Proceedings of the 14th
International Symposium on Methodologies for Intelligent Systems,
ISMIS 2003, volume LNAI 2871 of Lecture Notes in Computer Science,
Springer, Maebashi City, Japan, October 28--31, 2003.
Kešelj, Vlado. Question Answering using Unification-based
Grammar. In Advances in Artificial Intelligence, AI 2001, volume LNAI
2056 of Lecture Notes in Computer Science, Springer, Ottawa,
Canada, June 2001.