Preference in Eurotra

Directorate-General For Translation European Commission

Description

Information technology and telecommunications
Information - Documentation

Sujets

• Information processing
• Linguistics
• Machine translation
• Programming language
• Terminology

Informations

• Publié par: EUROPEAN-COMMISSION
• Nombre de lectures: 15
• Langue: English
• Poids de l'ouvrage: 1 Mo

Extrait

STUDIES IN MACHINE TRANSLATION
AND NATURAL LANGUAGE PROCESSING
Volume 3 Studies in machine translation
and natural language processing
Published by:
Office for Official Publications
of the European Communities Managing editor
Erwin Uilentini (CEC), Luxembourg
Editorial board
Jacques Durand (University of Salford, United Kingdom)
Frank van Eynde (Nationaal Fonds voor Wetenschappelijk Onderzoek,
België)
Tom C Gerhardt (CRP-CU/CRETA, Luxembourg)
Steven Krauwer (Rijksuniversiteit Utrecht, Nederland)
Bente Maegaard (Center for Sprogteknologi, Danmark)
Karsten Strørup, (CEC, Luxembourg)
Luxembourg: Office for Official Publications of the European Communities, 1993
ISSN 1017-6568
© ECSC-EEC-EAEC, Brussels · Luxembourg, 1993
Primed in Germany Volume 3
Preference in Eurotra
Edited by
Paul Bennett and Patrizia Paggio
Commission of the European Communities Volume 3
Preference in Eurotra
Editors
Paul Bennett
Patrizia Paggio
Contents
PAUL BENNETT, PATRIZIA PAGGIO
Introduction 7
PATRIZIA PAGGIO
The Eurotra preference mechanism Β
PAUL BENNETT
Some preference rules for English 27
PATRIZIA PAGGIO
Applying preference at the relational level 3
BOLETTE SANDFORD PEDERSEN
Lexical semantics and preference 49
HEINZ-DIETER MAAS
Preferences and complex words 65
LUCA DIM, GIOVANNI MALNATI
Weak constraints and preference rules 7
PAUL BENNETT
A comparative evaluation of the Eurotra preference mechanism 91 P. Bennett, P. Paggio: Introduction
PAUL BENNETT AND PATRIZIA PAGGIO
Introduction
This volume presents the results of research within the Eurotra project on preference.
The introduction begins by presenting a survey of work on preference, discusses some
general issues which the Eurotra preference theory addresses, and finally describes the
papers in the present volume.
The concept of preference was probably first used in computational linguistics by
Kimball (1973), who proposed, among a whole set of parse preferences, the principle of
right association, according to which the parser always attaches a new constituent to the
right of the lowest node in the parse tree. Frazier and Fodor (1978) observe that this
heuristic does not always provide the right parse of a syntactically ambiguous sentence,
and add to it the principle of minimal attachment, which chooses a high attachment if
this minimizes the number of nodes in the tree. To mediate between these two contra­
dictory principles, they advocate a parsing model based on a two-stage process, where
short sequences of about half a dozen words are parsed first, and intermediate results are
then delivered to a more 'intelligent' processor to attempt a complete parse. According
to this method, for which the authors adduce a psycholinguistic explanation, attach­
ments across long sequences of words are simply not seen by the parser. Both principles,
redefined slightly, have been recently advocated by Hobbs and Bear (1990), who,
however, admit that syntactic heuristics are not always sufficient means to achieve
correct disambiguation.
A different approach to preference can be found in Fass and Wilks (1983), or Wilks
(1985). According to this view, attachment ambiguities can be resolved only on the basis
of the semantic preferences linked to each word. Syntactic heuristics are rejected as
inadequate, as to each ambiguous sentence that a syntactic principle will be able to
disambiguate there corresponds at least one other showing the same syntactic structure
but different semantic priorities. Wilks (1985) also criticizes the system described in
Shieber (1983), where syntactic heuristics are replaced by a parsing strategy where
ambiguities of attachment are described and resolved in terms of conflicts between
different types of action in a shift-reduce parser. The results obtained by such a parser
would not be different from what the interaction of right association and minimal attach­
ment would produce.
Lexically induced preference is also the main disambiguation criterion invoked in
Ford et al. (1982). In this framework, however, lexical selectional restrictions are 8 Studies in MTandNLP, Milium· 3
expressed as syntactic rather than semantic constraints. Furthermore, lexical preference
is made to interact with general heuristic strategies. An example is the principle of final
arguments, which predicts a delay in the closure of the final argument of any given
lexical form, so that the parser, where possible, always attaches any following phrase
as a daughter to this final argument.
In Jackendoff (1985) the concept of preference is defined in a different way, and
receives a different application. This definition originates from the paradigm of cogni­
tive science, where preference is considered a complex cognitive process that governs
not only linguistic processing, but also other psychological processes, and in fact human
decision-making in general. In short, decisions are shown to be based on systems of
preference rules, i.e. on the interaction of graded judgement criteria. In this framework, e rules are used to resolve ambiguities, as well as to provide default assign­
ments in conceptual taxonomies. For example, 'flying' is a property which is generally
taken to define the category BIRD. However, OSTRICH is subordinate to BIRD,
although ostriches do not fly. Therefore, the assignment of the property of 'flying' to
subordinates of BIRD must be seen as the result of the application of a preference rule,
which the type OSTRICH can deviate from and still be a member of the category BIRD.
Jackendoff points out that the idea of default assignment is also relevant to the use of
preference heuristics in parsing, where incoming input is assigned a default structure on
the basis of preference criteria which, just as in the OSTRICH case, may well be over­
written.
Recently, attention has been given to statistical models that permit the automatic
acquisition of semantic preferences on the basis of the likelihood of occurrence of
certain patterns in real text. For example, Sekine et al. (1991) provide an algorithm
through which a system can train itself to compute the likelihood of occurrence of simple
word relations. In Basili et al. (1991), a similar strategy is used, where the reliability of
the associations found by the system, however, is improved by introducing a manual
semantic tagging of the relevant words. Certainly, analyses of this kind add empirical
foundation to the notion of semantic preferences.
The aim of this brief discussion of various approaches to preference is not to provide
a complete survey of the relevant literature, but only to create a conceptual background
against which the Eurotra preference mechanism can be seen. Thus, we have mentioned
only some of the most influential theories, as well as a few of the most recent contribu­
tions to the topic. A more comprehensive discussion of different preference mechanisms
in relation to the strategy adopted in Eurotra is provided by Bennett in the concluding
chapter of this volume.
In the light of the divergence of opinions found in the literature as to what factors a
preference theory should take into account — general and maybe contradictory heuristic
principles, lexical syntactic expectations, semantic selectional restrictions, even ref­
erential considerations, as in Hirst (1987) — the Eurotra project has chosen an approach
which allows the grammar developer to express preference statements based on vari­
ous criteria, the interaction of which is controlled by means of scoring.