Incremental Constraint-based Parsing: An Efficient Approach for Head-final Languages
Date
07/1997Author
Güngördü, Zelal
Metadata
Abstract
In this dissertation, I provide a left-to-right incremental parsing approach for Headdriven
Phrase Structure Grammar (HPSG; Pollard and Sag (1987, 1994)). HPSG is a
lexicalized, constraint-based theory of grammar, which has also been widely exploited
in computational linguistics in recent years. Head-final languages are known to pose
problems for the incrementality of head-driven parsing models, proposed for parsing
with constraint-based grammar formalisms, in both psycholinguistics and computational
linguistics. Therefore, here I further focusmy attention on processing a head-final
language, specifically Turkish, to highlight any challenges that may arise in the case
of such a language. The dissertation makes two principal contributions, the first part
mainly providing the theoretical treatment required for the computational approach
presented in the second part.
The first part of the dissertation is concerned with the analysis of certain phenomena in
Turkish grammar within the framework of HPSG. The phenomena explored in this part
include word order variation and relativization in Turkish. Turkish is a head-final language
that exhibits a considerable degree of word order freedom, with both local and
long-distance scrambling. I focus on the syntactic aspects of this freedomin simple and
complex Turkish sentences, detailing the assumptions Imake both to dealwith the variation
in the word order, and also to capture certain restrictions on that variation, within
the HPSG framework. The second phenomenon, relativization in Turkish, has drawn
considerable attention in the literature, all accounts so far being within the tradition of
transformational grammar. Here I propose a purely lexical account of the phenomenon
within the framework of HPSG, which I claim is empirically more adequate than previous
accounts, as well as being computationally more attractive.
The motivation behind the work presented in the second part of the dissertation mainly
stems from psycholinguistic considerations. Experimental evidence (e.g. Marslen-
Wilson (1973)) has shown that human language processing is highly incremental, meaning
that humans construct aword-by-word partial representation of an utterance as they
hear each word. Here I explore the computational effectiveness of an incremental processing
mechanism for HPSG grammars. I argue that any such processing mechanism
has to employ some sort of nonmonotonicity in order to guarantee both completeness
and termination, and propose a way of doing that without violating the soundness of
the overall approach. I present a parsing approach for HPSG grammars that parses a
string of words fromleft to right, attaching every word of the input to a global structure
as soon as it is encountered, thereby dynamically changing the structure as the parse
progresses.
I further focus on certain issues that arise in incremental processing of a “free”word order,
head-final language like Turkish. First, I investigate howthe parser can benefit from
the case values in Turkish in foreseeing the existence of an embedded phrase/clause
before encountering its head, thereby improving the incrementality of structuring. Second,
I propose a strategy for the incremental recovery of filler-gap relations in certain
kinds of unbounded dependency constructions in Turkish, which further enables one
to capture a number of (strong) preferences that humans exhibit in processing certain
examples with potentially ambiguous long-distance dependency relations.