Lexical-Semantic Relations: Theoretical and practical perspectives (Lingvisticæ Investigationes Supplementa)

Lexical-Semantic Relations

LINGVISTICÆ INVESTIGATIONES: SUPPLEMENTA

Studies in French & General Linguistics / Études en Linguistique Française et Générale This series has been established as a companion series to the periodical “LINGVISTICÆ INVESTIGATIONES”, which started publication in 1977.

Series Editors: Éric Laporte (Université Paris-Est Marne-la-Vallée & CNRS) Annibale Elia (Università di Salerno) Gaston Gross (Université Paris-Nord & CNRS) Elisabete Ranchhod (Universidade de Lisboa)

Volume 28 Petra Storjohann (ed.) Lexical-Semantic Relations. Theoretical and practical perspectives

Lexical-Semantic Relations Theoretical and practical perspectives Edited by

Petra Storjohann Institut für Deutsche Sprache, Mannheim

JOHN BENJAMINS PUBLISHING COMPANY AMSTERDAM/PHILADELPHIA

8

TM

The paper used in this publication meets the minimum requirements of American National Standard for Information Sciences — Permanence of Paper for Printed Library Materials, ANSI Z39.48-1984.

Library of Congress Cataloging-in-Publication Data Lexical-semantic relations : theoretical and practical perspectives / edited by Petra Storjohann. p.   cm. -- (Linguisticae investigationes. Supplementa ISSN; 0165-7569; v. 28) Includes bibliographical references and index. 1. Semantics. 2. Lexicology. 3. Relational grammar. I. Storjohann, Petra, 1972P325.L482   2010 401'.43--dc22 2010009954 ISBN 978 90 272 3138 3 (Hb: alk. paper) ISBN 978 90 272 8816 5 (Eb) © 2010 – John Benjamins B.V. No part of this book may be reproduced in any form, by print, photoprint, microfilm, or any other means, without written permission from the publisher. John Benjamins Publishing Co. • P.O.Box 36224 • 1020 ME Amsterdam • The Netherlands John Benjamins North America • P.O.Box 27519 • Philadelphia PA 19118-0519 • USA

Table of contents

Preface

vii

Introduction Petra Storjohann

1

Lexico-semantic relations in theory and practice Petra Storjohann

5

Swedish opposites: A multi-method approach to ‘goodness of antonymy’ Caroline Willners and Carita Paradis

15

Using web data to explore lexico-semantic relations Steven Jones

49

Synonyms in corpus texts: Conceptualisation and construction Petra Storjohann

69

Antonymy relations: Typical and atypical cases from the domain  of speech act verbs Kristel Proost An empiricist’s view of the ontology of lexical-semantic relations Cyril Belica, Holger Keibel, Marc Kupietz and Rainer Perkuhn The consistency of sense-related items in dictionaries: Current status,  proposals for modelling and applications in lexicographic practice Carolin Müller-Spitzer

95 115

145

Lexical-semantic and conceptual relations in GermaNet Claudia Kunze and Lothar Lemnitzer

163

Index

185

Preface

The availability of corpus-guided methods and the emergence of new semantic models, particularly cognitive and psycholinguistic frameworks, have prompted linguists to develop a range of immensely fruitful new approaches to sense relations. Not only does the field of sense relations have immediate relevance for the study of paradigmatic structures in lexicology, it is also a much discussed field for a variety of other application-oriented areas such as lexicography, Natural Language Processing and database engineering of lexical-semantic webs. It was in this context that the Institut für Deutsche Sprache in Mannheim (Germany) held an international colloquium from 5th–6th June 2008 on the subject of “Lexical-Semantic Relations from Theoretical and Practical Perspectives”. This event brought together researchers with an interest in semantic theory and experts with a more practical, application-based background looking at different languages. The papers in this volume derive from the colloquium and address specific semantic, lexicographic, computational and technological approaches to a range of meaning relations, particularly those which have traditionally been classified as “paradigmatic” sense relations, as well as exploring the construction, representation, retrieval and documentation of relations of contrast and meaning equivalence in a variety of languages including German, English and Swedish. This book provides specialists from different disciplines and areas with the opportunity to gain an insight into current cross-linguistic research in semantics, corpus and computer linguistics, lexicology, applied teaching and learning, and lexical typology as well as technological applications such as computational lexical-semantic wordnets. The overall aim of this book is to make up for some of the shortcomings of more traditional and often non-empirical studies, by providing an overview of current theoretical perspectives on lexical-semantic relations and presenting recent application-oriented research. Above all, its aim is to stimulate dialogue and revive discussion on sense relations in general, a subject which requires reappraisal in the light of recent semantic theories and which merits application via contemporary linguistic/lexicographic methods and procedures.

viii Lexical-Semantical Relations

I am appreciative of the help of all authors who have contributed to this book and who have clearly demonstrated the tremendous scope of the field and the importance of current trends in the study of paradigmatic structures. I also wish to thank some colleagues and friends for their criticisms, their help and support. My gratitude also goes to that handful of very special people for their understanding, their sense and sensitivity when making this book. Beyond these, a sincere thanks also goes out to the Institut für Deutsche Sprache Mannheim, for hosting the colloquium, thereby enabling semanticists, lexicographers, experts in Natural Language Programming and computer linguists to share their common interest in lexical-semantic relations. 

Petra Storjohann (Institut für Deutsche Mannheim)

Introduction Petra Storjohann

This collective volume focuses on what have traditionally been termed the “paradigmatics” or “sense relations” of a lexical unit. These include relations of contrast and opposition, meaning equivalence, hyponymy, hyperonymy etc., all of which have captured the interest of researchers from a range of disciplines. In the existing literature, studies on sense relations often just cover one specific phenomenon of one specific language, stressing specific semantic or methodological aspects. The present book covers different languages and different paradigmatic phenomena. It outlines the full complexity of the subject, combining linguistic and methodological elucidations with discussions on current practical and application-oriented research. The papers in this volume which are concerned with lexicological questions examine a range of linguistic models and semantic modelling, and the use of data such as corpora or the Internet as a lexical resource for information retrieval. Various authors demonstrate that research on language and cognition calls for evidence from different sources. They explain the nature of different lexical resources and the working methods associated with them, and they suggest some theoretical implications of a larger semantic model. The lexicological papers have as a common theme contextualised and dynamically constructed structures and look at the phenomenon as it is governed by conventional, contextual and cognitive constraints which favour specific choices. The semantic approaches used here concentrate on questions of mental representation, linguistic conventionalisation, cognitive processes and ideas of constructions, and they respond to the opportunities presented by methodologies from psycholinguistics or corpus studies. Moreover, the book explores recent developments in building large lexical resources as well as some lexicographic and text-technological aspects. These include for example elucidations on the structure and possible applications of reference databases such as GermaNet in natural language processing and computational linguistics. This database has been constructed in the style of its English counterpart WordNet and it is an integral part of EuroWordNet. This volume



Petra Storjohann

is also concern with sense-related items in dictionaries. Recent insights into the paradigmatics of a word from a lexical-semantic point make a compelling case for dictionary makers to include more appropriate and innovative descriptions of sense-related items in reference works. Additionally, new technical standards and text-technological facilities are still largely being ignored by lexicographers although they offer opportunities to enhance dictionaries and to make them more consistent. These too are some of the concerns of this book. Together, these papers not only give an impression of the scope of the field by looking at lexico-semantic relations from a range of positions and for different purposes, but they also demonstrate how cross-linguistic examinations benefit each other, how research areas fertilise and complement each other and how results in one field have an impact on research in other disciplines, each enriching the insights and developments of the others. In “Lexico-semantic relations in theory and practice” Petra Storjohann provides an overview on the subject of sense relations in different linguistic fields such as lexicology, corpus studies, lexicography and computer linguistics. The paper particularly focuses on the shift of approaches and methodologies to lexicosemantic relations in lexical semantics and concentrates on perspectives taken in German and in English linguistics. The paper is thought as a general discussion on the subject and it reveals some of the open questions and the challenges that need to be approached in the future. Both psycholinguistic and corpus-linguistic approaches are taken in the paper “Swedish opposites. A multi-method approach to goodness of antonymy” by Caroline Willners and Carita Paradis where the nature of English and Swedish antonymy and the degree of conventionalisation of antonymic word pairs in language and in memory are examined. Methodologically, their analyses are conducted on the basis of data from textual co-occurrence, and from judgement and elicitation experiments. Both types of evidence are used as a means of substantiating semantic theories. The paper not only examines differences in applying various methods, but also addresses the meanings of conventionalised canonical antonym pairings, including issues such as dimensional clarity, symmetry and contextual range. In terms of theoretical implications, it is argued that opposite meaning is construed and the authors show this with help of both highly conventionalised and less conventionalised binary opposites. Whether the Internet can be used as a valid corpus source for linguistic analyses is a question addressed by Steven Jones in “Using web data to explore lexicosemantic relations”. Taking English antonymy as an example, he explores how the web can be used as a lexical resource to reveal and quantify relational structures, and raises the question of whether prior semantic statements on canonicity can be based on such a methodology. He starts out from the assumption that specific



Petra Storjohann

theory and semantic models, they also propose an empirically-driven methodology specifically for the explorative examination of lexical-semantic relations, and they support the development of methods for empirical work with corpora and the detection of language structures in comprehensive linguistic data in more general terms. An interesting insight into the problems of inconsistent lexicographic information is provided in Carolin Müller-Spitzer’s paper “The consistency of senserelations in dictionaries. Current status, proposals for modelling and applications in lexicographic practice”. The paper reveals how inconsistent reference structures (e.g. ‘non-reversed reference’) are in German dictionaries of synonymy and antonymy. Problems such as bidirectional linking that is realised as unidirectional references are challenges for lexicographers as well as dictionary users. Although computational procedures are available to solve the problem, these have not been implemented so far. As Müller-Spitzer argues, a coherent lexicographic XMLbased modelling architecture is a prerequisite for an effective data structure. With the help of elexiko, a specific corpus-based, electronic reference work, she illustrates how text-technological methods can provide support for the overall consistency of sense-related pairings during the process of compiling a dictionary. Her discussion also outlines the technical requirements for achieving consistency in data-managing and data-linking. In their article “Lexical-semantic and conceptual relations in GermaNet”, Claudia Kunze and Lothar Lemnitzer discuss the relevance of some lexical as well as conceptual relations for a lexical resource of German. An overview of how these relations have been integrated in the construction and maintenance of GermaNet is given, and recent developments and their repercussions in terms of theoretical and application-oriented research are discussed. Other practical perspectives on the subject of meaning relations include the possible innovative and beneficial applications of GermaNet. It is beyond any doubt that this book remains a comparatively brief account of a complex field with a number of issues that are not even touched upon. Nevertheless, the intention is to stimulate further discussion and promote closer collaboration between the different fields. This collective volume attempts to show what research on lexical-semantic relations has to offer and to demonstrate, as Alan Cruse (2004: 141) asserts, “that sense relations are a worthwhile object of study”. As such, it is an invitation to scholars from every field whose interests involve words, meaning, the mind and/or language technology and who have a shared interest in lexical-semantic relations and approach the subject from various positions in philosophy, psychology, neuroscience, linguistics, lexicography, computer science, early childhood language acquisition and second language education, to name but a few.

Lexico-semantic relations in theory and practice Petra Storjohann

This paper provides a general overview of the treatment of lexico-semantic relations in different fields of research including theoretical and application-oriented disciplines. At the same time, it sketches the development of the descriptions and explanations of sense relations in various approaches as well as some methodologies which have been used to retrieve and analyse paradigmatic patterns.

1.

Lexicology: From structural to cognitive approaches

1.1

Structuralist approaches

From a lexicological point of view, the subject of sense relations has long been closely linked with several traditions of structural semantics and lexical field analysis, particularly within German linguistics. For decades, the theory of lexical field analysis was a very popular area of research, reaching its peak in the 1970s and 80s. Hence, it is automatically associated with the classical notion of the study of a language system, with atomised and isolated approaches, and the semantics of lexemes in terms of distinctive features. The emphasis is simultaneously on fixed and inherent semantic properties, componential meaning analysis and the idea that meaning can be neatly decomposed and described. The view was held that language is as an “externalized object” (Paradis 2009) with clearly recognisable structures. Sense relations were of particular interest since the basic assumption was that lexical meaning is constituted by the relations a lexeme holds with other lexemes in the same lexical-semantic paradigm. Structuralists not only made use of language as a system but also refered to lexical relations in terms of paradigmatic and syntagmatic structures implying strict distinctions between them.

. For an overview on the development of lexical field theory see Storjohann (2003: 25–40).



Petra Storjohann

Formalist linguists sought to define the meaning of lexical items by decompositional approaches, which worked well for modelling structural aspects such as phonology or syntax. But classical decompositional theories suffered from a number of problems, above all the belief that vocabulary has a definitional structure with distinct boundaries that can be precisely delimited. The traditional conception of sense relations was that of semantic connections between words and semantic interrelations among members of paradigmatic sets were viewed as stable and context-independent structures. Today, as a result, the phenomenon of sense relations is stigmatised and too closely linked to traditional or old-fashioned models. In German linguistics in particular, where once research on lexical-semantic structures flourished, the chapter on sense-related lexical terms was essentially closed by the works of Lutzeier (1981, 1985). His studies not only offered systematic examinations of lexical fields and their sense relations but they also made use of a stringent terminology and introduced the notion of contextual restrictions by bringing in key elements such as verbal context, syntactic category and semantic aspects. Particularly his later work pointed out the discrepancy between structuralist descriptions and textual structures in language use and prepared the ground for more empirical research on lexico-semantic relations in actual discourse. Nonetheless, the perception that the subject is obsolete persists to the present day and German semanticists have not further contributed to the general discussion on sense-related items in more recent contexts. In contrast, the situation with regard to semantics was never as bleak in the case of English linguistics, where scholars around the world were not so keen to avoid the subject of sense relations after the decline of the structuralist period. Lexical semantics in general has thrived in the UK, and its tradition is best exemplified by names such as John Lyons and Alan Cruse, both of whom have developed exhaustive definitions and descriptions of semantic relations. For them, the study of sense relations was central to the study of meaning. At the same time, the Firthian tradition developed which concentrated on syntagmatic relations. Collocations become the key notion and later the centre of attention to corpus linguists. Generally, more contextualised approaches to sense relations were encountered at that time with Cruse’s (1986) approach as a central piece of work in the tradition of the British Contextualism and consequently, it is studies on the English language which have succeeded in further advancing theories about lexico-semantic relations.

. Cf. Lyons (1968, 1977) and Cruse (1986). . “You shall know the meaning of a word by the company it keeps” (Firth 1957: 179).



1.2

Lexico-semantic relations in theory and practice

Cognitive approaches

As the notion of the lexicon started to be of interest to approaches to syntax which left behind the division between grammar and lexis, the nature of lexical semantics and the notion of the mental lexicon changed. New methodologies were introduced which looked at language from a usage-based perspective. However, corpus linguistics has largely focused its efforts on collocations and co-occurrences, and although linguistic theories have progressed, particularly in the area of cognitive linguistics, most semantic research has centred around issues such as polysemy and metaphor. And although the cognitive strand generally has had a major impact on lexical studies (cf. Geeraerts/Cuyckens 2007), the study of sense relations has not been a central component in the new semantic paradigm, and, as Cruse (2004: 141) concludes, “cognitive linguists, for the most part, have had very little to say on the topic”. Throughout his later work, Cruse himself has been concerned with bringing the cognitive aspect into his theory of meaning (Cruse 1992; Cruse/Togia 1995; Croft/Cruse 2004), unfortunately without incorporating new methodological approaches to substantiate his ideas. New guiding principles, assumptions and foundational hypotheses have become points of departure for semantic research in general, and they have gradually been transferred to the understanding of how sense relations are established in text and discourse. These concern how meaning is constructed. According to the cognitive school, meaning construction is equated with knowledge representation, categorisation and conceptualisation. Meaning is a process, it is dynamic, and it draws upon encyclopaedic knowledge and the subject of sense relations has started to be re-examined from a cognitive point of view. We now have a different understanding of how semantic relations are mentally represented and linguistically expressed, notions that are owed to the proliferation of research in the field of cognitive linguistics. Today, a number of linguists, mostly outside German linguistics, with a particular interest in lexical semantic relations, have reopened the chapter on sense relations offering new perspectives, employing new methodologies and using empirical evidence for their work. In particular, the Group for Comparative Lexicology has sought to advance theories around English and Swedish lexical relations. They have succeeded in showing how sense relations materialise in text and discourse. The question of whether sense relations are lexical relations, or rather conceptual-semantic relations, or relations among contextual construals, has been addressed. As a result, classical notions of the paradigmatics . Steven Jones (e.g. 2002), Lynne Murphy (e.g. 2003, 2006), Carita Paradis (e.g. 2005) and Caroline Willners (e.g. 2001) are particularly concerned with the study of English and Swedish opposites.





Petra Storjohann

of a lexical item have largely been abandoned. Recent semantic theories now account for lexical-semantic relations and are capable of accommodating all kinds of relations “ranging from highly conventionalized lexico-semantic couplings to strongly contextually motivated pairings” (Paradis forthcoming).

1.3

Corpus material and language in use

As Paradis (forthcoming) notes, it is methodologies which have radically changed studies on meaning and semantic relations. The basis of investigations is now determined by corpus procedures, by different observational and experimental techniques and by computational facilities and these contribute profitably to insights into the nature of the paradigmatics. A particularly promising trend within the new linguistic context is the fact that recent theories have also brought lexical semantics, and thus the subject of lexical-semantic relations, much closer to language in use and thought. Through the use of corpora, for example, we gain a different notion of language as it emerges from language use. The central function of language as a means of natural communication and its role in social interaction are no longer ignored. Conclusions are drawn not on the basis of intuitive judgement, but from real data and on the basis of mass data which account for recurrence, variability and the distribution of patterns. Generally, semanticists from various schools of thought have for a long time proved to be immune to corpus methods, and it is only recently that some researchers have made a compelling case for incorporating methods of corpus linguistics into semantics. This is all the more astonishing since both cognitive linguists and corpus linguists share an interest in contextualised, dynamically constructed meaning and in the grounding of language use in cognitive and social-interactional processes. Language in natural communicative situations involving speakers and addressees has come to occupy the seat of honour in cognitive linguistic research and the combination of the theoretical and empirical developments has sparked new interest in research on lexico-semantic relations and their functions in language and thought.  (Paradis forthcoming)

In terms of empirical corpus studies, it is however predominantly the subject of English opposites that has attracted interest from a corpus-based perspective (e.g. Jones 2002; Murphy 2006), demonstrating how corpus evidence can be used to derive semantic models. Until now, corpus-oriented studies of sense relations have been rather few and far between. However, systematic corpus-guided investigations have shown that corpus methodologies have contributed greatly to the study of lexical-semantic paradigms, and yielded new insights into issues such as how these relational patterns behave and function in discourse.



2.

Lexico-semantic relations in theory and practice

Lexicography

The field where findings on semantic relations demand to be accounted for and where they are of potential utility is lexicography. Sense relations are documented in dictionaries of synonymy and antonymy or in onomasiological reference books such as a thesaurus. There is a striking clash between the findings of theoretical semantic research on the one hand, and the commercial and practical missions of dictionaries on the other hand. Dictionary entries provide lists of sense-related items which are treated as stable relations between words, often not even assigned to a specific sense. And however inappropriate and inconsistent the representations of the facts about a word and its relations might be, it seems impossible to make a reference book radically different. The pressure of a dictionary is to present definite answers and clear-cut definitions. Hence, often sets of discrete synonyms or antonyms are given for words without overlapping meanings. Although it is commonly agreed that the construction of lexico-semantic relations is flexible, lexicographers continue to offer only vague descriptions and struggle to present meaning, and hence sense relations, as context-dependent, variable and dynamic. In addition, although corpora have been available for some time now, the exploration of mass data and the use of corpus tools for lexicographic analysis are restricted to corpus-based investigations, leaving a pool of linguistic evidence to be used for acts of verification only. Corpus-driven methodologies, however, where the corpus is approached without any prior assumptions and where collocation profiles reveal insights into the use of sense-related items, are largely ignored. As a result, as Alan Cruse comments: No one is puzzled by the contents of a dictionary of synonymy, or by what lexicographers in standard dictionaries offer by way of synonyms, even though the great majority of these qualify neither as absolute nor as propositional synonyms.  (Cruse 2004: 156)

An analysis of dictionary consultations by Harvey and Yuill in 1994 showed that in 10% of cases, users were looking for meaning equivalent terms. In over 36% of these situations, users were left without answers, or the information given was not satisfactory. Information on contextual conditions and situational usage was lacking. No other type of search showed the same degree of dissatisfaction. Users do

. For further differences between corpus-based and corpus-driven methodologies see Tognini-Bonelli (2001). . Unpublished research paper quoted in Partington (1998: 29).



10

Petra Storjohann

have an intuition for contextual restrictions of synonyms and they need to know the precise circumstances in which a lexical item can be substituted by a similar item. Yet most synonymies do not go beyond providing information on style or regional restrictions, although there are good methods available for comparing collocational profiles. As Partington stresses: Concordances of semantically similar lexical items can be studied, and students will inevitably discover differences in use which are not contained in grammars and dictionaries.  (Partington 1998: 47)

In fact, lexicological insights about natural language, about synonyms and antonyms in use, should be reflected in lexicographic descriptions. And corpus methods may help to correct or supplement dictionary information to support users when deciding in what circumstances substitution is possible. Essentially, the new approaches and methodological opportunities to examine a word’s meaning and sense relations show that the lexicographer’s craft needs to be rethought. Another problem we are still facing today is that of persistent inconsistencies in the documentation of synonyms and antonyms even in computer and corpusassisted dictionaries. Although there are analyses of different dictionaries pointing out some of these inconsistencies (e.g. Storjohann 2006; Paradis/Willners 2007), they do not include ways of implementing methods and tools to equip a reference work with stable and consistent cross-referencing or alternatively bidirectional linking as found in electronic resources. This is a field of research where computational linguistics needs to come up with answers.

3.

Computational linguistics – lexical databases and corpus tools

One area where the subject of sense relations has a more encouraging track record is the vast field of natural language processing, machine translation, ontologies, lexical database projects and the development of tools for computational and corpus linguistics. These are areas where sense relations have been topical and prosperous over the past two decades. Building a large computational lexicon has for example been the aim of Princeton WordNet and versions of this in other languages (e.g. GermaNet). It is not their objective to describe the nature of lexico-semantic relations in language use nor is it their goal to create a reference system designed to be a psychologically realistic model of the mental lexicon as the Princeton Group initially aimed at. Such resources have been de. For further problems of inconsistencies in dictionaries and possible answers to avoid them see Müller-Spitzer in this volume.



Lexico-semantic relations in theory and practice

veloped to support automatic text analysis, to serve as a thesaurus application and to demonstrate the organisational principles of the lexicon by listing relational structures. Particularly paradigmatic relations play a central role, as so-called synsets are grouped and interlinked by means of conceptual-semantic and lexical relations such as direct and indirect antonymy. These synsets and their relations constitute a supposedly stable language system. Semantically vague lexemes and lexical gaps within the lexicon pose problems for its model and generally, from a psychological and lexicological point of view, the plausibility of some explanations and descriptions need to be critically questioned. As is the case for lexicography, greater collaborative research is also required between semanticists and IT specialists in order to construct computational resources able, on the one hand, to provide more objective descriptions of the lexicon and, on the other, to provide adequate research tools for the linguistics community. It is also the work of computational and corpus linguists as well as IT experts which furnishes theoreticians and lexicographers with machine-readable corpora and with the procedures and research methods required to access mass data in order to extract synonyms or antonyms and to confirm or revise prior assumptions. As a matter of fact, as Church et al. (1991: 116) point out, the human mind is incapable of discovering statistically relevant, typical patterns or even ordering them according to significance scores.

Recognising recurring structures is an essential goal of any linguistic interpretation. To all those retrieving, identifying and analysing paradigmatic relations, the application of various linguistic methods and tools have become indispensable for linguists and lexicographers with empirical pursuits. Irrespective of their view on semantic models, more and more linguists and lexicographers base their findings on corpus-analysing methods and hence on the employment of semantic and mathematical-statistical models. The work of those who develop and refine methods of analysis has therefore become increasingly important. But on the other hand, researchers in linguistics should participate more fully in the development of computational tools so that these can also meet more theoretical research needs. Despite the fact that corpus search and analysing tools are widely used in linguistics now, there are hardly any publications which examine the underlying ideas, methods and models behind most corpus applications. Essentially, for any empirical lexicological work, knowledge about underlying models is crucial . Cf. Miller et al. (1990). . See Willners/Paradis in this volume and Divjak/Gries (2008).

11

12

Petra Storjohann

in order to be able to analyse, evaluate and interpret the retrieved corpus data. Again, only much more collaboration between the different fields of linguistic research can provide greater theoretical as well as practical understanding.

References Church, Kenneth, Gale, William, Hanks, Patrick and Hindle, Donald. 1991. “Using statistics in lexical analysis.” In Lexical Acquisition: Using On-Line Resources to Build a Lexicon, Uri Zernik (ed.), 114–164. Hillsdale, NJ: Lawrence Erlbaum. Croft, William and Cruse, Alan. 2004. Cognitive Linguistics. Cambridge: Cambridge University Press. Cruse, Alan. 1986. Lexical Semantics. Cambridge: Cambridge University Press. Cruse, Alan. 1992. “Antonymy revisited: Some thoughts on the relationship between word and concepts.” In Frames, fields and contrasts: New essays in semantic and lexical organization, Adrienne Lehrer and Eva Feder Kittay (eds), 289–306. Hillsdale, NJ: Lawrence Erlbaum. Cruse, Alan. 2004. Meaning in Language. (2nd ed.) Oxford: Oxford University Press. Cruse, Alan and Togia, Pagona. 1995. “Towards a cognitive model of antonymy.” Journal of Lexicology 1: 113–141. Divjak, Dagmar and Gries, Stefan. 2008. “Clusters in the mind? Converging evidence from near synonymy in Russian.” The Mental Lexicon 3(2): 188–213. Firth, John R. 1957. Papers in Linguistics. London: Oxford University Press. Geeraerts, Dirk and Cuyckens, Hubert (eds). 2007. The Oxford Handbook of Cognitive Linguistics. New York: Oxford University Press. Harvey, Keith and Yuill, Deborah. 1994. “The COBUILD testing initiative: The introspective, encoding component”. Unpublished research paper. Cobuild/University of Birmingham. Jones, Steven. 2002. Antonymy: A corpus-based perspective. London: Routledge. Lutzeier, Peter Rolf. 1981. Wort und Feld. Wortsemantische Fragestellungen mit beson­derer Berücksichtigung des Wortfeldbegriffes. Tübingen: Niemeyer. Lutzeier, Peter Rolf. 1985. “Die semantische Struktur des Lexikons.” In Handbuch der Lexikologie, Christoph Schwarze and Dieter Wunderlich (eds), 103–133. Königstein/Ts: Athenäum. Lyons, John. 1968. Introduction to Theoretical Linguistics. Cambridge: Cambridge University Press. Lyons, John. 1977. Semantics. 2 vols. Cambridge: Cambridge University Press. Miller, George, Beckwith, Richard, Fellbaum, Christiane, Gross, Derek and Miller, Katherine. 1990. “Introduction to WordNet: An on-line lexical database.” International Journal of Lexicography 3(4): 235–244. Murphy, M. Lynne. 2003. Semantic relations and the lexicon. Cambridge: Cambridge University Press. Murphy, M. Lynne. 2006. “Is ‘paradigmatic construction’ an oxymoron? Antonym pairs as lexical constructions”. Constructions SV1. http://www.constructions-online.de/. Paradis, Carita. 2005. “Ontolgies and construals in lexical semantics.” Axiomathes 15: 541–573. Paradis, Carita. (forthcoming). “Good, better, superb antonyms: A dynamic construal approach to oppositeness.” In Proceedings of 19th International Symposium on Theoretical and Applied Linguistics (Aristotle University of Thessaloniki, April 2009).



Lexico-semantic relations in theory and practice

Paradis, Carita and Willners, Caroline. 2007. “Antonyms in dictionary entries: Selectional principles and corpus methodology.” Studia Linguistica 61(3): 261–277. Partington, Alan. 1998. Patterns and Meanings. Using Corpora for English Language Research and Teaching. Amsterdam/Philadelphia: John Benjamins. Storjohann, Petra. 2003. A Diachronic Contrastive Lexical Field Analysis of Verbs of Human Locomotion in German and English. Frankfurt: Peter Lang. Storjohann, Petra. 2006. “Korpora als Schlüssel zur lexikografischen Überarbeitung. Die Neubearbeitung des Dornseiff.” In Lexikographica. Internationales Jahrbuch für Lexikographie. 21/2005, Fredric F. M. Dolezal, Alain Rey, Herbert Ernst Wiegand, Werner Wolski and Ladislav Zgusta (eds), 83–96. Tübingen: Niemeyer. Tognini-Bonelli, Elena. 2001. Corpus Linguistics at Work. Amsterdam/Philadelphia: John Benjamins.

13

Swedish opposites A multi-method approach  to ‘goodness of antonymy’* Caroline Willners and Carita Paradis

This is an investigation of ‘goodness of antonym pairings’ in Swedish, which seeks answers to why speakers judge antonyms such as bra–dålig (good–bad) and lång–kort (long–short) to be better antonyms than, say, dunkel–tydlig  (obscure–clear) and rask–långsam (speedy–slow). The investigation has two main aims. The first aim is to provide a description of goodness of Swedish antonym pairings based on three different observational techniques: a corpus-driven study, a judgement experiment and an elicitation experiment. The second aim is to evaluate both converging and diverging results on those three indicators and to discuss them in the light of what the results tell us about antonyms in Swedish, and perhaps more importantly, what they tell us about the nature of antonymy in language and thought more generally.

1.

Introduction

In spite of the widespread consensus in the linguistic literature that contrast is fundamental to human thinking and that antonymy as a lexico-semantic relation plays an important role in organising and constraining the vocabularies of languages (Lyons 1977; Cruse 1986; Fellbaum 1998; Murphy 2003), relatively little empirical research has been conducted on antonymy, either using corpus methodologies or experimental techniques. No studies have been conducted using a combination of both methods. The general aim of this article is to describe a combination of methods useful in the study of antonym canonicity, to summarise the results and to assess their

* Thanks to Joost van de Weijer for help with the statistics, to Anders Sjöström for help with producing figures and to Simone Löhndorf for help with data collection.

16

Caroline Willners and Carita Paradis

various advantages and disadvantages for a better understanding of goodness of antonymy as a lexico-semantic construal. By combining methods, we hope to contribute to the knowledge about the nature of antonymy as a relation of binary contrast. A mirror study has been performed for English and is reported on in Paradis et al. (2009). Antonyms are at the same time minimally and maximally different from one another. They activate the same conceptual domain, but they occupy opposite poles/parts of that domain. Due to the fact that they are conceptually identical in all respects but one, we perceive them as maximally similar, and, at the same time, due to the fact that they occupy radically different poles/parts, we perceive them as maximally different (Cruse 1986; Willners 2001; Murphy 2003). Words that we intuitively associate with antonymy are adjectivals (Paradis and Willners 2007). Our approach assumes antonyms, both more strongly canonical and less canonical, to be conceptual in nature. Conceptual knowledge reflects what speakers of languages know about words, and such knowledge includes knowledge about their relations (Murphy 2003: 42–60; Paradis 2003, 2005; Paradis et al. 2009). Treating relations as relations between concepts, rather than relations between lexical items is consistent with a number of facts about the behaviour of relations. Firstly, relations display prototypicality effects, in that there are better and less good relations. In other words, not only is torr (dry) the most salient and wellestablished antonym of våt (wet), but the relation itself may also be perceived as a better antonym relation than, say, seg–mör (tough–tender). When asked to give examples of opposites, people most often offer pairs like bra–dålig (good–bad), svag–stark (weak–strong), svart–vit (black–white) and liten–stor (small–large), i.e. common lexical items along salient (canonical) dimensions. Secondly, just like non-linguistic concepts, relations in language are about construals of similarity, contrast and inclusion. For instance, antonyms may play a role in metonymisation and metaphorisation. At times, new metonymic or metaphorical coinages seem to be triggered by relations. One such example is slow food as the opposite of fast food. Thirdly, lexical pairs are learnt as pairs or construed as such in the same contexts. Canonicity plays a role in new uses of one of a pair of a salient relation. For a longer introduction to this topic, see Paradis et al. (2009). The central issue of this paper concerns ‘goodness of antonymy’ and methods to study this. Like Gross and Miller (1990), we assume that there is a small group of strongly antonymic word pairs (canonical antonyms) that behave differently from other less strong (non-canonical) antonyms. (Direct/indirect and lexical/ conceptual are alternative terms for the same dichotomy.) For instance, it is likely that speakers of Swedish would regard långsam–snabb (slow–fast) as a good example of canonical antonymy, while långsam–kvick (slow–quick), långsam–rask (slow–rapid) and snabb–trög (fast–dull) are perceived as less good opposites.



Swedish opposites

All these antonymic pairs in turn will be different from unrelated pairs such as långsam–svart (slow–black) or synonyms such as långsam–trög (slow–dull). As for their behaviour in text, Justeson and Katz (1991, 1992) and Willners (2001) have shown that antonyms co-occur in the same sentence at higher than chance rates, and that canonical antonyms co-occur more often than non-canonical antonyms and other semantically possible pairings (Willners 2001). These data support the dichotomy view of the Princeton WordNet and Gross and Miller (1990). The test set used in the present study consists of Swedish word pairs of four different types: Canonical antonyms, Antonyms, Synonyms and Unrelated word pairs (see Tables 4 and 5). The words in the Unrelated word pairs are always from the same semantic field but the semantic relation between them is not clear even though they might share certain aspects of meaning, e.g. het–plötslig (hot–­sudden). Synonyms and Unrelated word pairs were introduced as control groups. While it is not possible to distinguish the four types using corpus methodologies, we expect significant results when judged for ‘goodness of oppositeness’ experimentally and in the number of unique responses when the individual words are used as stimuli in an elicitation test. All of the word pairs included in the study co-occur in the same sentence significantly more often than chance predicts. An early study of ‘goodness of antonymy’ is to be found in Herrmann et al. (1979). They assume a scale of canonicity and use a judgement test to obtain a ranking of the word pairs in the test set. We include a translation of a subset of his test items in this study in an attempt to verify or disconfirm his results. The procedure is as follows. Section 2 discusses some methodological considerations before the methods used are described in detail in the sections that follow. Corpus-driven methods are used to produce the test set (Section 4) that is used in the elicitation experiment (Section 5) and the judgement experiment (Section 6). A general discussion of the results and an assessment of the methods are found in Section 7. Finally, the study is concluded in Section 8. Before going into details about our method and experiments, we give a short overview of previous work relevant to the present study.

2.

Methodological considerations

In various previous studies, we explored antonymy using corpus-based as well as corpus-driven approaches (e.g. Willners 2001; Jones et al. 2007; Murphy et . In current empirical research where corpora are used, a distinction is made between corpus-based and corpus-driven methodologies (Francis 1993; Tognini-Bonelli 2001: 65–100;

17

18

Caroline Willners and Carita Paradis

al. 2009; Paradis et al.). Corpus data are useful for descriptive studies since they reflect actual language use. They provide a basis for studying language variation, and they also often provide metadata about speakers, genres and settings. Another, very important property of corpus data is that they are verifiable, which is an important requirement for a scientific approach to linguistics. Through corpus-driven methods, it is possible to extract word pairs that share a lexical relation of some sort. However, there is no method available for identifying types of relation correctly. For instance, it is not possible to tell the difference between antonyms, synonyms and other semantically related word pairs (in this case word pairs from the same dimensions, which co-occur significantly at sentence level, but are neither antonyms, nor synonyms, e.g. klen (weak) – kort (short). The answer(s) to the types of question we are asking are not to be found solely on the basis of corpus data. As Mönnink (2000: 36) puts it The corpus study shows which of the theoretical possibilities actually occur in the corpus, and which do not.

The questions we are asking call for additional methods. A combination of corpus data, elicitation data and judgement data is valuable in order to determine if and how antonym word pairs vary in canonicity. It also sheds light on different aspects of the issue. Like Mönnick (2000), we believe that a methodologically sound descriptive study of linguistics is cyclic and preferably includes both corpus evidence and intuitive data (psycho-linguistic experimental data).

3.

Data extraction

3.1

Method

Antonyms co-occur in sentences significantly more often than chance would predict and canonical antonyms co-occur more often than contextually restricted Storjohann 2005; Paradis and Willners 2007). The distinction is that the corpus-based methodology makes use of the corpus to test hypotheses, expound theories or retrieve real examples, while in corpus-driven methodologies, the corpus serves as the empirical basis from which researchers extract their data with a minimum of prior assumptions. In the latter approach, all claims are made on the basis of the corpus evidence with the necessary proviso that the researcher determines the search items in the first place. Our method is of a two-step type, in that we mined the whole corpus for both individual occurrences and co-occurrence frequencies for all adjectives without any restrictions, and from those data we selected our seven dimensions and all their synonyms.



Swedish opposites

Table 1.  Observed and expected sentential co-occurrences of 12 different adjective pairs (from Willners 2001: 72) Word1

Word2

N1

N2

Co

Expected Co

Ratio

P-value

bred djup gammal hög kall kort liten ljus långsam lätt lätt tjock

smal grund ung låg varm lång stor mörk snabb svår tung tunn

  113   117 1050   760   102   262 1344    84    55   225   225    53

   55    17   455   333   102   604 2673   126   163   365   164    85

   2    1   47   47   12   21 111    7    4    5    7    4

  0.12   0.04   8.84   4.68   0.19   2.93 66.48   0.20   0.17   1.52   0.68   0.08

17.39 27.17   5.32 10.04 62.32   7.17   1.67 35.82 24.11   3.29 10.25 47.98

0.0061 0.036 0 0 0 0 0 0 0 0.020 0 0

antonyms (Justeson and Katz 1991; Willners 2001). This knowledge helps us to decide which antonyms to select for experiments investigating antonym canonicity. Willners and Holtsberg (2001) developed a computer program called Coco to calculate expected and observed sentential co-occurrences of words in a given set and their levels of probability. An advantage of Coco was that it took variation of sentence length into account, unlike the program used by Justeson and Katz (1991). Coco produces a table which lists the individual words and the number of individual occurrences of these words in the corpus in the four left-most columns. Table 1 lists 12 Swedish word pairs that were judged to be antonymous by Lundbladh (1988) from Willners (2001): N1 and N2 are the number of sentences respectively in which Word1 and Word2 occur in the corpus. Co is the number of times the two words are found in the same sentence and Expected Co is the number of times they are expected to co-occur in the same sentence if predicted by chance. Ratio is the ratio between Observed and Expected co-occurrences and P-value is the probability of finding the actual number of co-occurrences that was observed or more under the null hypothesis that the co-occurrences are due to pure chance only. All of Lundbladh’s antonym pairs co-occurred in the same sentence significantly more often than predicted by chance. Willners (2001) reports that 17% of the 357 Swedish adjective pairs that cooccurred at a significance level of 10–4 in the SUC were antonyms. The study . Stockholm-Umeå Corpus, a one-million-word corpus compiled according to the same principles as the Brown Corpus. See http://www.ling.su.se/staff/sofia/suc/suc.html.

19

20 Caroline Willners and Carita Paradis

Table 2.  The top 10 co-occurring adjective pairs in the SUC, sorted according  to rising p-value Swedish antonyms

Translation

höger–vänster kvinnlig–manlig svart–vit hög–låg inre–yttre svensk–utländsk central–regional fonologisk–morfologisk horisontell–vertikal muntlig–skriftlig

right–left female–male black–white high–low inner–outer Swedish–foreign central–regional phonological–morphological horizontal–vertical oral–written

included all adjectives in the corpus. When the same data were (quite unorthodoxly) sorted according to rising p-value, antonyms clustered at the top of the list as in Table 2. Most of the antonym word pairs were classifying adjectives with overlapping semantic range, e.g. fonologisk–morfologisk (phonological–morphological) and humanistisk–samhällsvetenskaplig (humanistic–of Social Sciences). Among the 83% of the word pairs that were not antonyms were many other lexically related words. Furthermore, Willners (2001) compared the co-occurrence patterns of what Princeton calls direct antonyms and indirect antonyms. Both types co-occur significantly more often than chance predicts. However, there is a significant difference between the two groups: while the indirect antonyms co-occur overall 1.45 times more often than would be expected if predicted by chance, the direct antonyms co-occur 3.12 times more often than expected. The hypothesis we are testing in this study is that there are good and bad antonyms (cf. canonical and non-canonical). Coco provides a data-driven method of identifying semantically related word pairs. We used Coco to suggest possible candidates for the test set. However, since we wanted a balance between Canonical antonyms, Antonyms, Synonyms and Unrelated word pairs in the test set, human interference was necessary and we picked out the test items manually from the lists produced by Coco.

3.2

Result

Using the insights from previous work on antonym co-occurrence as our point of departure, we developed a methodology for selecting data for our experiments. To start with, we agreed on a set of seven dimensions from the output of



Swedish opposites

Table 3.  Seven corresponding canonical antonym pairs in Swedish and English Dimension

Swedish antonyms

Translation

speed luminosity strength size width merit thickness

långsam–snabb mörk–ljus svag–stark liten–stor smal–bred dålig–bra tunn–tjock

slow–fast dark–light weak–strong small–large narrow–wide good–bad thick–thin

the corpus searches of sententially co-occurring items that we perceived to be good candidates for a high degree of canonicity and identified the pairs of antonyms that we thought were the best linguistic exponents of these dimensions (see Table 3). For cross-linguistic research we made sure that the word pairs also had well-established correspondences in English. The selected antonym pairs are all scalar adjectives compatible with scalar degree modifiers such as very. Using Coco, we ran the words through the SUC. All of them co-occurred in significantly high numbers at sentence level and these pairs were set up as Canonical antonyms. Next, all Synonyms of the 14 adjectives were collected from a Swedish synonym dictionary. All the Synonyms of each of the words in each antonym pair were matched and run through the SUC in all possible constellations for sentential co-occurrence. This resulted in a higher than chance co-occurrence for quite a few words for each pair. We extracted the pairs that were significant at a level of p < 0.01 for further analysis. Using dictionaries and our own intuition, we then categorised the word pairs according to semantic relations. Finally, we picked two Antonyms, two Synonyms and one pair of Unrelated adjectives from the list of significantly co-occurring word pairs for dimension. Table 4 shows the complete set of pairs retrieved from the SUC: 42 pairs in all. We also included eleven word pairs from Herrmann et al.’s (1979) study of ‘goodness of antonymy’ (see Table 5). From his ranking of 77 items, we picked every sixth word pair, translated them into Swedish and classified them according to semantic relation: Canonical antonym (C), Antonym (A) and Unrelated (U). None of the pairs from Herrmann et al. (1979) were judged to be synonymous. The word pairs as well as the individual words in Table 4 and Table 5 were used as the test set in the psycholinguistic studies described below.

. Strömbergs synonymordbok 1995. Alva Strömberg, Angered: Strömbergs bokförlag.

21

22

Caroline Willners and Carita Paradis

Table 4.  The test set retrieved from the SUC. See Appendix A for translations Canonical antonyms

Antonyms

Synonyms

Unrelated

långsam–snabb

långsam–flink tråkig–het vit–dunkel melankolisk–munter lätt–muskulös senig–kraftig obetydlig–kraftig liten–väldig smal–öppen trång–rymlig dålig–god ond–bra4 genomskinlig–svullen fin–grov

långsam–släpig snabb–rask ljus–öppen mörk–svart svag–matt stark–frän stor–inflytelserik liten–oansenlig smal–spinkig bred–kraftig dålig–låg bra–god tunn–spinkig tjock–kraftig

het–plötslig

ljus–mörk svag–stark liten–stor smal–bred dålig–bra tunn–tjock

dyster–präktig flat–seg klen–kort liten–tjock fin–tokig knubbig–tät

Table 5.  Test items selected from Herrmann et al. (1979) Word1

Word2

Translated from

Herrmann’s score

Semantic relation

ful smutsig trött lugn hård irriterad sparsmakad overksam förtjusande framfusig vågad

vacker fläckfri pigg upprörd böjlig glad spännande nervös förvirrad Hövlig sjuk

beautiful–ugly immaculate–filthy tired–alert disturbed–calm hard–yielding glad–irritated sober–exciting nervous–idle delightful–confused bold–civil daring–sick

4.90 4.62 4.14 3.95 3.28 3.00 2.67 2.24 1.90 1.57 1.14

C A C A A A A U U A A



. Due to sparse data, this item was added despite the fact that it did not meet the general criterion of being over the limit of 0.01. We chose ond–bra (evil–good) because we expected interesting results for the English counterpart in the mirror study. Ond–bra (evil–good) is included in the test set, but is not included in the result discussions.



Swedish opposites

4.

Elicitation experiment

This section describes the method and the results of the elicitation experiment.

Stimuli and procedure The test set for the elicitation experiment involves the individual adjectives that were extracted as co-occurring pairs from the SUC and translations of selected word pairs from Herrmann et al.’s (1979) list of adjectives perceived by participants as better and less good examples of antonyms (see Table 4 and Table 5). Some of the individual adjectives occur in more than one pair, i.e. they might occur once, twice or three times. For instance, långsam (slow) occurs three times and snabb (fast) occurs twice. All second and third occurrences were removed from the elicitation test set, which means that långsam (slow) and snabb (fast) occur once in the test set used in the elicitation experiment. Once this was done, the adjectives were automatically randomised and printed in the randomised order. All in all, the test contains 85 stimulus words. All participants obtained the adjectives in the same order. The participants were asked to write down the best opposites they could think of for each of the 85 stimuli words in the test set. For instance:

Motsatsen till LITEN är ‘The opposite of SMALL is’ Motsatsen till PRÄKTIG är ‘The opposite of DECENT is’

The experiment was performed using paper and pencil and the participants were instructed to do the test sequentially, that is, to start from word one and work forwards and not to go back to check or change anything. There was no time limit, but the participants were asked to write the first opposite word that came to mind. Each participant also filled in a cover page with information about name, sex, age, occupation, native language and parents’ native language. All the responses were then coded into a database using the stimulus words as anchor words.

Participants Twenty-five female and 25 male native speakers of Swedish participated in the elicitation test. They were between 20 and 70 years of age and represented a wide range of occupations as well as levels of education. All of them had Swedish as their first language, as did their parents. The data were collected in and around Lund, Sweden.

23

24

Caroline Willners and Carita Paradis

Predictions Our predictions are as follows: – The test items that we deem to have canonical antonyms will elicit only one another. – The test items that we do not deem to be canonical will elicit varying numbers of antonyms – the better the antonym pairing, the fewer the number of elicited antonyms. – The elicitation experiment will produce a curve from high participant agreement (few suggested antonyms) to low participant agreement (many suggested antonyms).

4.1

Results

We will start by reporting the general results in Section 5.1.1 and then go on to discuss the results concerning bidirectionality in Section 5.1.2. We performed a cluster analysis, the results of which are presented in Section 5.1.3.

4.1.1 General results The main outcome of the elicitation experiment is that there is a continuum of lexical association of antonym pairs. In line with our predictions, there was a number of test words for which all the participants suggested the same antonym: bra (good) – dålig (bad), liten (small) – stor (large), ljus (light) – mörk (dark), låg (low) – hög (high), mörk (dark) – ljus (light), sjuk (ill) – frisk (healthy), smutsig (dirty) – ren (clean), stor (large) – liten (small), and vacker (beautiful) – ful (ugly). All the elicited antonyms across the test items are listed in Appendix A. Appendix  A also shows that there is a gradual increase of responses from the top of the list to the bottom of the list. The very last item is sparsmakad (fastidious), for which 33 different antonyms were suggested by the participants (including a non-answer). The shape of the list of elicited antonyms across the test items in Appendix A strongly suggests a scale of canonicity from very good matches to test items with no clear partners. While Appendix A gives all the elicited antonyms across the test items, it does not provide information about the scores for the various individual elicited responses. The three-dimensional diagram in Figure 1 is a visual representation of how some stimulus words elicited the same word from all participants. Those are the maximally high bars found to the very left of the diagram (e.g. bra (good), liten (small), ljus (light), etc.). Then four words follow for which 49 of the participants suggested the same antonym while another opposite was suggested in the 50th case. These four stimulus words were dålig (bad), svag (weak), stark (strong), and



Swedish opposites

Figure 1.  The distribution of Swedish antonyms in the elicitation experiment. The Y-axis gives the test items, with every tenth test item written in full. The X-axis gives the number of suggested antonyms across the participants given on the Z-axis

ond (evil). Forty-nine of the participants suggested bra (good) as an antonym of dålig (bad), stark (strong) for svag (weak), svag (weak) for stark (strong) and god (good) for ond (evil). The ‘odd’ suggestions were frisk (healthy) for dålig (bad), klar (clear) for svag (weak), klen (feeble) for stark (strong) and snäll (kind) for ond (evil). Since there are two response words for each of the four stimuli in these cases, there are two bars, one 49 units high at the back representing the most commonly suggested antonym and one small bar, only one unit high, in front of the big one, representing the single suggestions frisk (healthy), klar (clear), klen (feeble) and snäll (kind). The further we move towards the right in Figure 1, the more diverse the responses. In fact, the single suggestions spread out like a rug covering the bottom of the diagram as we move towards the right. However, there is usually a preferred response word which most of the participants suggested. There are some stimuli for which two response words were equally popular choices or which at least were both suggested by a considerable number of participants. For example, for lätt (light/easy), 29 participants suggested tung (heavy) and 20 svår (difficult); het (hot) elicited the responses kall (cold) (24) and sval

25

26 Caroline Willners and Carita Paradis

(chilly) (20); and for god (good), participants suggested ond (evil) (20) and äcklig (disgusting) (19). A common feature of these stimulus words is that they are associated with different strongly competing meaning dimensions or salient readings. Some other examples are framfusig (bold): tillbakadragen (unobtrusive) (20) and blyg (shy) (16); trång (narrow): rymlig (spacious) (17) and vid (wide) (15); fläckfri (spotless): fläckig (spotted) (17) and smutsig (dirty) (15); grov (coarse): fin (fine) (17) and tunn (thin) (14). Like Appendix A, Figure 1 indicates that there is a scale of canonicity with a group of highly canonical antonyms to the left and a gradual decrease of canonicity as we move towards the right in the diagram. The stimulus words on the lefthand side of Figure 1 cannot be said to have any good antonyms at all.

4.1.2 Bidirectionality In addition to the distribution of the responses for all the test items across all the participants, we also investigated to what extent the test items elicit one another in both directions. For instance, 50 participants gave dålig (bad) as an antonym of bra (good) and ful (ugly) for vacker (beautiful), but the pattern was not the same in the other direction. This is part of the information in Appendix A and Figure 1, but it is not obvious from the way the information is presented. For the test items that speakers of Swedish intuitively deem to be good pairs of antonyms, this strong agreement held true in both directions, although not at the level of a oneto-one match, but one-to-two or one-to-three. While 50 participants responded with dålig (bad) as the best opposite of bra (good), two antonyms were suggested for dålig (bad): bra (good) by 49 participants and frisk (healthy) by one participant. This points to the possibility that there is a stronger relationship between bra (good) and dålig (bad) than between frisk (healthy) and dålig (bad). In other words, Figure 1 shows that the more canonical pairs elicit only one or two antonyms, while there is a steady increase in numbers of ‘best’ antonyms the further we move to the right-hand side of the figure. 4.1.3 Cluster analysis In order to shed light on the strength of the lexicalised oppositeness, a cluster analysis of strength of antonymic affinity between the lexical items that co-occurred in both directions was performed. It is important to note that only items that were also test items were eligible as candidates for participation in bidirectional relations. This means that some of the pairings suggested by the participants were not included in the cluster analysis. For instance, tung (heavy) was considered the best antonym of lätt (light) by 29 of the participants (as compared to 20 for svår (difficult)), but since neither tung nor svår were included among the test items, the pairings were not measured in the cluster analysis. The results of



Swedish opposites

the cluster analysis are, however, comparable to the results of sentential co-occurrence of antonyms in the corpus data and the results of the judgement experiment, since the same word pairs are included. To this end, a hierarchical agglomerative cluster analysis using Ward amalgamation strategy (Oakes 1998: 119) was performed on the subset of the data that were bidirectional. Agglomerative cluster analysis is a bottom-up method that takes each entity (i.e. antonym paring) as a single cluster to start with and then builds larger and larger clusters by grouping together entities on the basis of similarity. It merges the closest clusters in an iterative fashion by satisfying a number of similarity criteria until the whole dataset forms one cluster. The advantage of cluster analysis is that it highlights associations between features as well as the hierarchical relations between these associations (Glynn et al. 2007; Gries and Divjak 2009). Cluster analysis is not a confirmatory analysis but a useful tool for exploratory purposes. Figure 2 shows the dendrogram produced on the basis of the cluster analysis. The number of clusters was set to four to match the four conditions on the basis of which we retrieved our data from the sententially co-occurring pairs in the first place (Canonical antonyms, Antonyms, Synonyms and Unrelated). Figure 2 shows the hierarchical structure of the clusters. There are two branches. The leftmost branch hosts Cluster 1 and Cluster 2 and the right-most branch Cluster 3 and Cluster 4. The closeness of the fork to the clusters indicates a closer relationship. The tree structure reveals that there is a closer relation between Cluster 3 and Cluster 4 than between Cluster 1 and Cluster 2. Figure 2 gives the actual pairings in the boxes at the end of the branches. There are fewer pairs at the end of the left-most branches than at the end of the branches on the right-hand side. Five of the word pairs in Cluster 1 were included in the test set as Canonical antonyms: långsam–snabb (slow–quick), ljus–mörk (light–dark), svag–stark (weak–strong), bra–dålig (good–bad) and liten–stor (small–large) (subscripted with c in Figure 2). The other two word pairs in Cluster 1 were vit–svart (white–black) from the luminosity dimension and tjock–smal (fat–thin) from thickness. The rest of the word pairs in Cluster 1 were not included as pairs in the experiment. In Cluster 2, there are four word pairs featured in the test set as Canonical: tunn–tjock (thin–thick), bred–smal (wide–narrow), vacker–ful (beautiful–ugly) and trött–pigg (tired–alert). The rest of the word pairs in Cluster 2 are intuitively good parings. They were, however, not among the parings that we deemed canonical in the design of the test set, e.g. upprörd–lugn (upset–calm), väldig–liten (enormous–small), fin–ful (pretty–ugly), nervös–lugn (nervous–calm), ond–god (evil–good) and rymlig–trång (spacious–narrow).

27

28

Caroline Willners and Carita Paradis

Figure 2.  Dendrogram of the bidirectional data

It is not obvious what the systematic differences are between the degrees of oppositeness in Clusters 3 and 4. As the dendrogram above shows, they are in fact associated. However, they do not correspond to the Synonyms and Unrelated word pairs in the test set, since the cluster analysis is based on the results of the elicitation experiment where the participants were asked to provide the best antonym.

5.

Judgement experiment

This section describes the methodology of the judgement experiment in which the participants were asked to evaluate word pairings in terms of how good they thought each pair was as a pair of antonyms. The experiment was carried out online. The design of the screen is shown in Figure 3.



Swedish opposites

Figure 3.  An example of a judgement task in the online experiment  (translated into English)

As Figure 3 shows, the participants were presented with questions of the form: Hur bra motsatser är X–Y? (How good is X–Y as a pair of opposites?) and Hur bra motsatser är Y–X? (How good is Y–X as a pair of opposites?) The question was formulated using bra (good) (not dålig (bad)) in order for the participants to understand the question as an impartial how-question, since Hur dåliga motsater är fet–smal? (How bad is fat–lean as a pair of opposites?) presupposes ‘badness’. The end-points of the scale were designated with both icons and text. On the left-hand side there is a sad face (very bad antonyms), while there is a happy face on the right-hand side (excellent antonyms). The task of the participants was to tick a box on a scale consisting of eleven boxes. We were also interested in whether the ordering of the pairs had any effect. Our predictions were as follows. – The nine test pairings that we deem to be canonical will receive 11 on the scale of ‘goodness’ of pairing of opposites. – The order of presentation of the Canonical antonyms as well as the Antonyms will give rise to significantly different results. Word1–Word2 will be considered better pairings than Word2–Word1. – There will be significant differences between the judgements about Canonical antonyms, Antonyms, Synonyms and Unrelated pairings.

Stimuli The same test set as in the elicitation experiment was used (see Table 4 and 5), but while the pairing of the antonyms was not an issue in the first experiment, it was essential to the judgement test. The stimuli were presented as pairs and the test items were automatically randomised for each participant. Half of the participants were given the test items in the order Word1–Word2, while the other half were presented with the words in reverse order, i.e. Word2–Word1.

29

30

Caroline Willners and Carita Paradis

Procedure The judgement experiment was performed online using E-prime as experimental software. E-prime is a commercially available Windows-based presentation program with a graphical interface, a scripting language similar to Visual Basic and response collection. E-prime conveniently logged the ratings as well as the response times in separate files for each of the participants. The participants were presented with a new screen for each word pair (see Figure 3). The task of the participants was to tick a box on a scale consisting of eleven boxes. The screen immediately disappeared upon clicking which prevented the participants from going back and changing their responses. Between each judgement task there was a blank screen with an asterisk, and when the participants were ready for the next task they signalled that with a mouse-click. Before the actual test started, the participants were asked to give some personal data (name, age, sex, occupation, native language and parents’ native language). There then followed some instructions such as how to do the mouse-clicks and information about the fact that the test was self-paced. Each participant had two test trials before the actual judgement test of the 53 test items. The purpose of the study was revealed to the participants in the instructions. As has already been mentioned, the judgement experiment was divided into two parts: 25 participants were given the test set as Non-Reverse (Word1–Word2, e.g. långsam–snabb (slow–fast)) and 25 participants were given the test set in the reverse order: Reverse (Word2–Word1, e.g. snabb–långsam (fast–slow)). This was done to measure whether the order of the sequence influenced the results in any way. Participants Fifty native speakers of Swedish participated in the judgement test. None of them had previously participated in the elicitation test. Twenty-nine of the participants were women and 21 were men between 20 and 62 years of age. All of them had Swedish as their first language. 5.1

Results

This section reports on the results of the judgement experiment. We start reporting on the results concerning sequencing in Section 6.1.1 since they affect the treatment of the data reported in the section on strength of canonicity (Section 6.1.2).

5.1.1 Sequencing As has already been pointed out, the test was performed in such a way that half of the participants were presented with the test items in the order: Word1–Word2,



Swedish opposites

and the other half in reverse order, Word2–Word1. We assumed that the order would have an impact on the results, at least for the Canonical antonyms. A subject analysis and an item analysis were performed. The factors involved were directionality, category (Canonical antonyms, Antonyms, Synonyms and Unrelated) and the interaction between directionality and category. In the subject analysis, each participant was the basic element for analysis. All judgements for the individual participants were averaged within each of the four conditions, yielding four numbers per participant. Then a repeated measures ANOVA analysis of variance (Woods et al. 1986: 194–223) was performed on both data sets. In the item analysis, each item (i.e. word pair) was the basic element for analysis. The judgements given by each participant on each condition were averaged, resulting in four numbers for each item, and a Univariate General Linear Model analysis was performed. Finally, Bonferroni’s post hoc test (Field 2005: 339) was used to compare the differences between the categories. The same procedure was used for the response times. The statistical analysis shows that the order of sequence does not have any effect on the results: F1[1,48] = 1.056, p = 0.309, F2[1,98] = 0.206, p = 0.651. The interaction between the sequence and category does not have an effect either: F1[3,144] = 0.811, p > 0.05, F2[3,98] = 0.069, p = 0.976. Category, on the other hand, does have an effect: F1[3,144] = 1777.991, p < 0.001, F2[3,98] = 138.987, p < 0.001. Figure 4 shows that the two test batches (marked with REV = 0 and

Figure 4.  Sequential ordering: there is no significant difference between the mean  answers of the two test batches

31

32

Caroline Willners and Carita Paradis

Table 6.  Mean responses for each of the word pairs in the test set,  both directions included Word pair

Mean response

Semantic category

ljus–mörk långsam–snabb liten–stor svag–stark trött–pigg dålig–bra ful–vacker smal–bred tunn–tjock

10.92 10.88 10.84 10.80 10.76 10.68 10.64 10.60 10.40

C C C C C C C C C

fin–grov trång–rymlig dålig–god smutsig–fläckfri lugn–upprörd melankolisk–munter liten–väldig framfusig–hövlig hård–böjlig långsam–flink ond–bra irriterad–glad senig–kraftig obetydlig–kraftig vit–dunkel lätt–muskulös liten–tjock tråkig–het smal–öppen sparsmakad–spännande

10.32 10.20   9.84   9.36   9.28   9.04   8.52   8.40   7.84   7.80   6.84   6.56   5.88   5.44   5.40   4.44   4.08   3.68   3.20   2.76

A A A A A A A A A A A A A A A A U A A A

bra–god dyster–präktig overksam–nervös fin–tokig förtjusande–förvirrad långsam–släpig svag–matt stark–frän stor–inflytelserik knubbig–tät genomskinlig–svullen

  2.52   2.00   1.92   1.88   1.84   1.80   1.76   1.76   1.68   1.68   1.68

S U U U U S S S S U A



Word pair snabb–rask bred–kraftig flat–seg dålig–låg ljus–öppen klen–kort liten–oansenlig het–plötslig mörk–svart tjock–kraftig vågad–sjuk smal–spinkig tunn–spinkig

Swedish opposites

Mean response

Semantic category

  1.60   1.60   1.56   1.56   1.48   1.48   1.44   1.44   1.40   1.40   1.32   1.28   1.24

S S U S S U S U S S A S S

REV = 1) follow the same pattern. Since the order of the sequence did not have an impact on the results, the data for the two directions will be treated as one batch and will not be separated in the analyses that follow.

5.1.2 Strength of canonicity The mean response for each word pair in the test set is presented in Table 6. The mean responses for the Canonical antonyms vary between 10.40 and 10.92. None of the word pairs have a response mean of 11, which we expected for the Canonical antonyms. They do, however, top the list. The means for the Antonyms vary greatly, from 10.32 for fin–grov (fine–course) to 1.68 for genomskinlig–svullen (transparent–swollen). Below 2.52, a mix of unrelated and synonymous word pairs are found and the word pair that was judged to be the ‘worst’ antonym pair was tunn–spinkig (thin–skinny) (1.24). The overall mean responses for the four categories are presented in Table 7. The Canonical antonyms have a mean response of 10.72, close to the maximum, 11. The standard deviation is also small for this category, 0.6, which reflects high consensus among the participants. The Antonyms have a significantly lower mean of 6.82, but with a large standard deviation, 3.37. This indicates a lower degree of consensus among the participants. The response for the Synonyms is 1.61, with a standard deviation of 1.33, and for the Unrelated it is 1.92, with a standard deviation of 1.55. There is no significant difference between the last two categories. The results in Table 7 are also illustrated in Figure 5. We performed a repeated measures ANOVA and the differences between the Canonical antonyms and Antonyms as well as between Antonyms and the two other categories (Synonyms and

33

34

Caroline Willners and Carita Paradis

Table 7.  Mean responses for Canonical antonyms, Antonyms, Synonyms  and Unrelated word pairs Category

Mean

Std. deviation

Canonical antonyms Antonyms Synonyms Unrelated

10.724   6.824   1.609   1.920

0.6084 3.3728 1.3279 1.5502

Figure 5.  Mean responses for Canonical antonyms, Antonyms, Synonyms and Unrelated word pairs

Unrelated) were significant both in the subject analysis (F1[3,147] = 1784.874, p  or (relational partner) and not according to typographic presentation such as “italic” or “bold”. It is also the case that the

Figure 5.  Parts of the XML structure (elexiko entries ehemalig and früher)



The consistency of sense-related items in dictionaries 153

modelling is very fine-granular. Every single piece of lexicographic information has its corresponding XML-tag, so every unit is (individually) accessible by computer. For the presentation, the XML data are transformed by an XSLT stylesheet into a HTML-based browser view. Consequently, the presentation of the lexicographic information is defined separately from its content. So the main question concerns the problem of how necessary reference information is recorded in the entries. If the lexicographer is working on one entry (in this case ehemalig) and a synonym relation with another word (in this case früher) is detected, for example, in the corpus, s/he has to list this term in its corresponding place within the XML structure of the entry. In this case, it is the tag within the parent element <synonymie>, where s/he has to insert a form of identification, a number assigned to the targeted entry () and in certain cases also the aforementioned signpost of an individual targeted sense (). The ID of an entry is a numeric code greater than zero. The ID of a specifically linked sense of a synonym is the corresponding guide word. The attribute “0” is only given in cases where the linked synonym has not yet been lexicographically described in full, due to the fact that elexiko is not yet a complete dictionary but an ongoing project. The problem, however, is the lack of technical assistance. When the lexicographer edits an entry and the XML instance is being parsed against the DTDs, there is no technical support in elexiko which signals whether the inserted item could be related to a specific sense or alternatively, if the reference needs to be set to the lexeme in general. If the latter is the case, the lexicographer chooses the attribute “0” instead of the right lesart-ID (the right guide word). So if a lexicographer is working on the entry früher at a later point than on its corresponding sense-related lexemes, s/he might forget to go back to a certain synonym, for example ehemalig. This step is, however, essential in order to add the necessary identification of the guide word and its corresponding sense to realise the link as a sense-related reference (). Therefore, when starting the lexicographic work on one entry, there is a need for an automatic message with information on all other existing entries which contain data references with respect to the entry that is currently being edited. Again, this is nothing new for software developers working in (commercial) lexicographic context. However, in this context, the problem is discussed on another level, namely the modelling level. Solving this problem for elexiko is difficult at the moment because from a strictly formal perspective, all reference structures in elexiko go from x to y and – at best – from y to x. This bidirectionality is not currently being checked

. Elexiko is a corpus-based dictionary.

154 Carolin Müller-Spitzer

Figure 6.  Bidirectional references in elexiko: current status

or tested for systematically by any tool. At the moment, two reference points between which a bidirectional relation exists are just two references without really being in connection with one another, as illustrated in Figure 6. This is the problem most dictionaries face. As things stand, bidirectional references are two independent references from one point out of the hierarchical tree (which is an XML document) to another. Additionally, the two poles between which a relation holds have a different scope and they are technically independent of each other. Improving this situation is quite straightforward, as demonstrated in Figure 6. Provided that we want to base the modelling of reference structures on a standard, the XML-connected standard XLink (XML Linking Language) is one possible option. Regarding the application of such a standard, XLink is not currently implemented in many tools, so it could be argued that it is not necessary to map the modelling onto this standard. However, in my view, it is very useful to look at this standard, simply because the standard reflects a large number of considerations about reference structures in general, which can only be of benefit (cf. Nordström 2002). However, the following proposals for modelling reference structures could also be implemented in tailor-made XML DTDs or schemas. XLink has been established to allow “elements to be inserted into XML documents in order to create and describe links between resources” (XLink: 2). The view is taken here that it is a slim as well as an adequately complex format, enabling users to model simple but also more complex linking structures with XLink. In the introduction of the XLink specification, it is stated that: „XLink provides a . For the description of the XLink Standard see http://www.w3.org/TR/xlink/



The consistency of sense-related items in dictionaries 155

framework for creating both basic unidirectional links and more complex linking structures. It allows XML documents to: – assert linking relationships among more than two resources, – associate metadata with a link, – express links that reside in a location separate from the linked resources”. It is especially important to be able to associate metadata with a link and to build a link database or (abbreviated) linkbase (which is meant for storing links separately from the linked resources). Before going into detail as to why this is important, some notes on reference structures in general are necessary. Reference structures can be classified into unidirectional and bidirectional relations. For instance, a reference from a dictionary entry to an illustration, a corpus sample or an external encyclopaedia is a unidirectional link. It points in one direction only, i.e. from the target resource there is no reverse reference, which means that referring back to the original source is neither intended nor useful. In the majority of cases, the target resource is outside the lexicographer’s responsibility, as it is outside the lexicographic database. In the context of paradigmatic structures, it is bidirectional references which are of particular interest. Unlike the aforementioned unidirectional reference, creating bidirectional references is part of the lexicographer’s compiling responsibilities. It is only then that two resources may function as a source on the one hand and as the target resource on the other. For example, references to Wikipedia are always given in one direction only (cf. Müller-Spitzer 2007a and b: 169). In this section, the modelling concept is looked at more closely. The first and most general guideline of the concept for modelling cross-reference structures is to model bidirectional references as extended links (in the terminology of XLink) and to store them in a linkbase. This approach has the following advantages: – someone can work on the links or change them externally without touching the dictionary entries themselves and – a link database supports the management of the cross-reference structures. Looking at the current model of references in elexiko again, the difference should become obvious. The model idea which is pursued here is one where all links are stored on an individual level, separated from the entries (a draft of this is shown in Figure 7). What does this mean for the elexiko example? The lexicographic information relating to references should be imported automatically into the linkbase file. In Figure 8, we can see that every piece of important information relating to the crosslinking of the given sense-related items is transferred into XLink and its specific elements and attributes. The process is as follows: any information relat-

156 Carolin Müller-Spitzer

Figure 7.  Bidirectional references in elexiko: use of a linkbase

Figure 8.  Part of the XLink-linkbase

ing to reference structures is stored in this linkbase. This procedure is performed automatically during the process of checking the entry file into the underlying dictionary database system. In this model, the connections from one synonym to another and vice versa are not two independent references any longer, but form one complex link object. This has the following advantages: Firstly, precise addressing of source and target resource is possible. In the data model which is currently being used, one reference goes from the (so from one lexicographic piece of information in the entry) to the target entry or, alternatively, to a specific sense as a whole. Coming from the



The consistency of sense-related items in dictionaries 157

related item and looking in the other direction, it is just the other way around. So we have an incorrect addressing structure because the synonym relation is – with regard to the content – valid between two contextual instantiations as represented by senses. Now the storage in a linkbase enables us to address the starting point and the target resource independent of the position where the lexicographic information on these references is given in the entries. We can also see this in the part of the linkbase in Figure 8. The attribute @label addresses the sense as a whole as the starting or finishing point of the relation. We might ask whether this is really important. It is assumed here that it is, because if we think of quite different ways of presenting lexicographic data (for example a network of all semantically related senses), it is very important to identify the resources precisely. This is also of particular importance for accessing lexicographic data, for example, if one wishes to present a search word together with its sense-related items. The second main advantage of using a linkbase is the ability to associate metadata with a link. This is explained further by introducing some general features of XLink. One crucial point of extended links in XLink is that “the extended-type element may contain a mixture of the following elements in any order, possibly along with other content and markup: – locator-type elements that address the remote resources participating in the link – arc-type elements that provide traversal rules among the link’s participating resources – title-type elements that provide human-readable labels for the link – resource-type elements that supply local resources that participate in the link.” [XLink: 11] The option to add a human-readable title attribute and to specify traversal rules may seem interesting features in this context. Traversal rules are rules which define how to follow a link and can be specified in extended links as follows: An extended link may indicate rules for traversing among its participating resources by means of a series of optional arc elements. The XLink element for arc is any element with an attribute in the XLink namespace called type with the value ‘arc’. [XLink: 16] […] The arc-type element may have the traversal attributes from and to […], the behavior attributes show and actuate […] and the semantic attributes arcrole and title […]. The traversal attributes define the desired traversal between pairs of resources that participate in the same link, where the resources are identified by their label attribute values. The from attribute defines

158 Carolin Müller-Spitzer

resources from which traversal may be initiated, that is, starting resources, while the to attribute defines resources that may be traversed to, that is, ending resources. The behavior attributes specify the desired behavior for XLink applications to use when traversing to the ending resource. [XLink: 17]

Although it seems to be very interesting to define traversal rules and titles here in the linkbase together with the data model, it has to be noticed that in present-day technology one would prefer to separate the data structure from the application logic. This is an important point of criticism of the XLink-Standard. Looking again at the example (cf. Figure 8), we can see that in the presented part of the linkbase there are two objects modelled as resources, which are elements with the XLink attribute “locator”. They are identified by their labels, in this case the IDs, together with the guide words which are used for identifying senses in elexiko. At the bottom of the middle column in Figure 8, the first arc (an element with the XLink-specific attribute @arc) with the name <synonym_zu1-2>” appears. The second arc is termed <synonym_zu2-1>. All these elements are parts of the complex link object synonymierelation. The behaviour of attributes of XLink can generally speaking be characterised as follows: these attributes refer to an application while traversing the links. As has been pointed out before, these attributes should be defined separately from the linkbase. What impact do these general explanations have when they are related to the concrete examples of sense-related items in elexiko? By specifying a model for a linkbase, it is possible to define a fixed type of extended link for the relation type synonymy. This model may be presented as follows. The synonym link object always connects two remote resources with one another. These resources are entries or individual senses respectively. The corresponding arcs between these resources could be specified by traversal rules. In the case of synonymy, it would be useful to specify the value of show into new. That is, when a user clicks on a synonym, the targeted entry has to be opened in a new frame. In this way, both windows can be arranged next to each other on the computer screen and the two entries can then be received simultaneously. The value of actuate is probably to be assigned to onRequest. This means that the user has to click on the given relation partner (i.e. sense-related item) in the entry in order to follow the link. This abstract model is then applicable to each concrete synonym relation. This kind of model allows us to specify different presentations of different kinds of lexical semantic relation. For example, in the case of lexemes which are connected by hyperonymy (i.e. superordination) or hyponymy (subordination), users might like to look these up one after the other, whereas in the case of lexemes which are



The consistency of sense-related items in dictionaries 159

connected by synonymy or antonymy, users might compare both by simultaneously looking up both entries. Being able to add metadata to a link object can also be used in an extended way. It may be that a relationship between two words might be more significant for lexeme a than it is for lexeme b. When a collocation analysis of both terms individually is performed, synonym a might rank higher in its significance to b than the other way around. This is a regular observation with polysemous lexical items. Such an observation could also be given as metadata, so the XLink-feature of adding metadata could be used here as well. It is clear that this new modelling concept is much more powerful than the one that is currently being used, for example in elexiko and in other dictionaries as well. The final question that remains open is what the consequences might be of the employment and integration of such a linkbase into an editing system for the practical lexicographic work in elexiko (and similar projects). The course of the working process can be roughly sketched as follows: a lexicographer checks an entry out of the database system in order to work on it. Then, as a first step, the linkbase verifies whether the chosen entry is registered as a target resource. If it is, the lexicographer receives an automatic message informing him/her which resource (that is which entry) is the starting point of the relation and which type of relation exists between the items. So the lexicographer can have this in mind while studying the corpus results. Then the entry is lexicographically edited in an XML editor, including all references. Next, another check routine is added. While checking the edited entry-file back into the database system, the corresponding information is imported into the linkbase and – and this is the most important thing – a query is initiated as to whether the information about one reference structure is consistent with the information about the other. If one connection runs from the start resource to the target resource but not back again, the lexicographer receives an error message. In elexiko, it is important from a pragmatic point of view to keep the writing practice in the XML editor as it is. Alternatively, in other projects, it may be better practice to source all linking information out of the entries and store them in the linkbase only. Not only does this form of modelling reference structures allow for typing and assigning attributes as well as for new ways of presenting reference structures, but, most importantly, it also allows lexicographers to make entries more consistent. At this point, it should be stressed again that the modelling concept presented in this paper does not necessarily imply the use of XLink. A tailor-made XML-structure, based on the same guidelines, is able to perform the same tasks.

160 Carolin Müller-Spitzer

5.

Future perspectives

As we have seen, a consistent, content-based and fine-grained way of structuring lexicographic data allows for new ways of presenting sense-related items in dictionaries. In this paper, again the case of the dictionary elexiko is taken for the purpose of demonstration. Here, sense-related items are presented in a specific part of the dictionary. This presentation is well arranged and the screen is not overcrowded. The disadvantage of this kind of presentation is that it is difficult to see which sense-related items are given for one headword in all of its senses at a glance. Instead, the user has to click from one sense to another and it is therefore very difficult to compare the given sense-related items of the individual senses. Provided that all the information about sense-related items is consistently structured, it can be presented in a different way. In Figure 9, an alternative presentation illustrates the paradigmatics of früher in all its senses collectively. The sense-related items of the entry früher could be loaded into any tool which is able to generate a graph or net. The only prerequisite is strictly hierarchically structured data (as shown in Figure 9) in the underlying dictionary entry. Throughout this paper, the relation of synonymy serves as an example, but in this entry, we encounter numerous other types of paradigmatic relation. The

Figure 9.  elexiko entry früher with its sense-related items in different senses  presented as a net



The consistency of sense-related items in dictionaries 161

headword itself is positioned in the middle, and around it are its senses, labelled “LA” for “Lesart”, followed by the types of sense relation such as synonymy or incompatibility. The purpose of the graph is to show that as well as more traditional lexicographic presentations, there are other ways of presenting sense-related connections in a dictionary. What needs emphasising at this point is the fact that different ways of presentation can be made without any change to the underlying data. The preferred way of structuring data enables us to do this at the touch of a button, if the appropriate tool is available. Moreover, if the necessary data are structured according to the new modelling concept, much more elaborate ways of presentation are possible. To conclude, providing and supporting consistency of sense-related items and reference structures in dictionaries in general should not be seen as an irrelevant hobby of dictionary reviewers and text-technologists. Providing consistency is crucial for lexicographic practice and it can have an effect on qualitative enhancements. At the same time, a consistent way of data modelling and structuring is a prerequisite for developing innovative forms of presenting lexicographic data in an electronic medium such as the Internet. The benefits of consistency in general lie therefore on both sides, for the lexicographer who is compiling the dictionary and for the user who needs to make successful look up procedures.

References Blumenthal, Andreas, Lemnitzer, Lothar and Storrer, Angelika. 1988. “Was ist eigentlich ein Verweis? Konzeptionelle Datenmodellierung als Voraussetzung computergestützter Verweisbehandlung.” In Das Wörterbuch. Artikel und Verweisstrukturen (Jahrbuch 1987 des Instituts für deutsche Sprache), Gisela Harras (ed.), 351–373. Düsseldorf/Bielefeld: Pädagogischer Verlag Schwann-Bagel and Cornelsen-Velhagen u. Klasing. Cruse, D. Alan. 2004. Meaning in Language. (2nd ed.) Oxford: Oxford University Press. Engelberg, Stefan and Lemnitzer, Lothar. 2001. Lexikographie und Wörterbuchbenutzung (Stauf­ fenburg Einführungen, vol. 14). Tübingen: Stauffenburg. Joffe, David and De Schryver, Gilles-Maurice. 2004. “TshwaneLex – Professional off-the-shelf lexicography software.” In DWS 2004 – Third International Workshop on Dictionary Writing Systems, 17–20. http://tshwanedje.com/publications/dws2004-TL.pdf (last visited on 2009/07/03). Lew, Robert. 2007. “Linguistic semantics and lexicography: A troubled relationship.” In Language and meaning. Cognitive and functional perspectives, Malgorzata Fabiszak (ed.), 217– 224. Frankfurt am Main: Peter Lang. Litkowski, Ken C. 2000. “The Synergy of NLP and Computational Lexicography Tasks”. Technical Report 00-01. Damascus, MD: CL Research. Müller-Spitzer, Carolin. 2007a. “Vernetzungsstrukturen lexikografischer Daten und ihre XMLbasierte Modellierung.” Hermes 38/2007: 137–171.

162 Carolin Müller-Spitzer

Müller-Spitzer, Carolin. 2007b. Der lexikografische Prozess. Konzeption für die Modellierung der Datenbasis (Studien zur deutschen Sprache 42). Tübingen: Gunter Narr. Nordström, Ari. 2002. “Practical XLink.” In XML 2002, Proceedings by deepX. www.idealliance.org/papers/xml02/dx_xml02/papers/06-00-11/06-00-11.html (last visited June 2008/06/01). Storjohann, Petra. 2005. “elexiko: A Corpus-Based Monolingual German Dictionary.” Hermes 34/2005: 55–73. XML Linking Language (XLink) Version 1.0 (2001). W3C Recommendation 27 June 2001 http://www.w3.org/TR/xlink/ (last visited on 2009/07/03).

Reference works Duden. 2007. Das Synonymwörterbuch, 4th edition. Mannheim/Leipzig/Wien/Zürich: Dudenverlag. elexiko. 2003ff. In OWID – Online Wortschatz-Informationssystem Deutsch, Mannheim: Institut für Deutsche Sprache, www.owid.de/elexiko_/index.html (last visited on 2009/07/03). Merriam Webster Online. 2009. (www.Merriam-Webster.com). entry Consistency. http://www. merriam-webster.com/dictionary/consistency (last visited at 2009/07/02). Wahrig. 2006. Synonymwörterbuch, 5th edition. München/Gütersloh: Wissen Media Verlag. Wikipedia.com. 2009. (www.wikipedia.com). entry Consistency: http://en.wikipedia.org/wiki/ Consistency (last visited on 2009/07/03). Stanford Encyclopedia of Philosophy. (http://plato.stanford.edu/entries/logic-classical/). entry Consistency http://plato.stanford.edu/search/searcher.py?query=consistency (last visited at 2009/07/02).

Lexical-semantic and conceptual relations in GermaNet Claudia Kunze and Lothar Lemnitzer

GermaNet is a lexical resource constructed in the style of the Princeton  WordNet. Lexical units are grouped in synsets which represent the lexical  instantiations of concepts. Relations connect both these synsets and the lexical units. In this paper, we will describe the kinds of relations which have been  established in GermaNet as well as the theoretical motivation for their use.

1.

Lexical-semantic resources in natural language applications

Digital lexical resources such as machine-tractable dictionaries (MTD) and lexical knowledge bases (LKB) are extensively used in natural language processing and computational linguistics. Basic application scenarios include: – – – – – –

word sense disambiguation; information retrieval and information extraction; linguistic annotation of language data on several layers of description; text classification and automatic summarisation; tool development for language analysis and language generation; machine translation.

Lexical-semantic wordnets, being considered as “lightweight linguistic ontologies”, have become popular online resources for a number of different languages since the success of the original Princeton WordNet (cf. Miller 1990; Fellbaum 1998). Encoding different kinds of lexical-semantic relations between lexical units and concepts, wordnets are valuable background resources not only in computational linguistic applications, but also for research on lexicalisation patterns in and across languages.

164 Claudia Kunze and Lothar Lemnitzer

2.

The relational structure of lexical-semantic wordnets

Wordnets which are structured along the lines of the Princeton WordNet (cf.  Miller 1990; Fellbaum 1998) encode the most frequent and most important concepts and lexical units of a given language, as well as the sense relations which hold between them. In wordnets, a word or lexical unit is represented as a conceptual node, with its specific semantic links to other words and concepts in that language. For example, Stuhl (chair) is represented with its superordinate term Sitzmöbel (seating furniture) as well as its subordinate terms Drehstuhl (swivel chair), Klappstuhl (folding chair), Kinderstuhl (high chair) etc. Furthermore, the superordinate concept Sitzmöbel (seating furniture) is connected to the concepts Lehne (armrest), Sitzfläche (seat) and Bein (chair leg), which represent generic parts of a piece of seating furniture (see Figure 1). A concept is characterised not only by its representational node, but also by its web of semantic relations to other words and concepts. In wordnets, the basic unit of representation is the so-called synset which groups equivalent or similar meaning units, the synonyms, in a common conceptual node. Thus, wordnets do not conflate homonyms, but disambiguate word senses. As pointed out in the introduction, word sense disambiguation is a necessary precondition for numerous applications in natural language processing. Wordnets project natural-language hierarchies, in contrast to formal ontologies which constitute language-independent or domain-specific conceptual networks. In the next section, we focus on describing the German wordnet GermaNet and the semantic relations it encodes (Kunze 2004) as well as its integration into a polylingual wordnet architecture connecting eight European languages (EuroWordNet, see Vossen 1999).

Figure 1.  Subtree Sitzmöbel (seating furniture) from the GermaNet hierarchy



2.1

Lexical-semantic and conceptual relations in GermaNet 165

GermaNet – a German wordnet

With GermaNet, a digital semantic lexicon has been built as an important contribution to a German language resource infrastructure. The German wordnet has adopted the database format and the main structural principles of the Princeton WordNet 1.5, which pioneered numerous language-specific initiatives in relation to building wordnets. GermaNet is not merely a translation of WordNet 1.5, but pursues its own core themes in conceptual modelling, by including artificial concepts (see below) and assuming a taxonomic approach for representing not just nouns but all different parts of speech. GermaNet was built from scratch, taking into account different lexicographical resources such as Deutscher Wortschatz (Wehrle and Eggers 1989) and Brockhaus-Wahrig (Wahrig et al. 1984), as well as the frequency lists of various German corpora. GermaNet models lexical categories such as nouns, verbs, and adjectives. The synset as central unit of representation supplies the set of synonyms for a given concept, as for example, {Streichholz, Zündholz} (match), {fleißig, eifrig, emsig, tüchtig} (busy) and {vergeben, verzeihen} (to forgive, to pardon). GermaNet encodes semantic relations either between concepts (synsets) or lexical units (single synonyms in the synsets). Currently, GermaNet contains some 58,000 synsets with nearly 82,000 lexical units, most of them nouns (43,000), followed by verb concepts (9,500) and adjectives (5,500). The coverage of the German wordnet is still being extended with text corpora being the basis of these extensions. GermaNet contains only a few multi-word items such as gesprochene Sprache (spoken language) or Neues Testament (New Testament). Proper nouns are primarily taken from the field of geography, e.g. city names, and are specifically labelled.

2.1.1 Lexical relations in GermaNet The richness of lexical-semantic wordnets derives from the high number of semantic links between the lexical objects. A principal distinction is made between lexical relations and conceptual relations: – Lexical relations are bi-directional links between lexical units (word meanings) such as synonymy (equivalent meanings for synset partners such as Ruf . Cf. http://www.sfs.uni-tuebingen.de/GermaNet. . However, the first computational large-scale semantic network based upon content words of dictionary definitions was developed by Quillian (1966) in order to model semantic memory in Artificial Intelligence. . Cf. Hamp and Feldweg (1997), the differences between GermaNet and Princeton WordNet are listed in Lemnitzer and Kunze (2002).

166 Claudia Kunze and Lothar Lemnitzer

und Leumund (reputation)), and antonymy, holding between, for instance, Geburt (birth) and Tod (death), glauben (to believe) and zweifeln (to doubt), schön (beautiful) and hässlich (ugly). – Conceptual relations such as hyponymy, hyperonymy, meronymy, implication and causation hold between concepts, and thus apply to all synset variants. Hyponymy and hyperonymy form converse pairs: while Gebäude (building) constitutes the hyperonym of Haus (house), Haus constitutes the hyponym of Gebäude. The most important structural principle in semantic networks is the hyponymy relation, as between Rotkehlchen (robin) and Vogel (bird), yielding the taxonomical structure of the linguistic ontology. For nouns, deep hierarchies are possible, e.g. the concept Kieferchirurg (oral surgeon) has a taxonomic chain which comprises 15 elements (known as its “path length”). The GermaNet data model also applies the taxonomic approach to verbs and adjectives. With regard to adjectives, the Princeton WordNet and several of its successors prefer a model based upon antonymy (between central pairs of antonym representatives such as good and bad which build clusters with similar concepts respectively grouped around them like satellites (the “satellite approach”)). GermaNet overcomes this unsatisfactory treatment of adjectives and accounts for the Hundsnurscher/Splett classification of adjectives. The meronymy relation (“part-whole relation”) is assumed only for nouns. Dach (roof) cannot be appropriately classified as a kind of Gebäude (building), but is part of it. Part-whole relations also pertain to abstract structures, as regards the membership of a certain group, such as Vorsitzender (chairman) of a Partei (party), or the substance in a composition, such as Fensterscheibe (window pane) which is made of Glas (glass). Typically, the link between lexical resultatives such as töten (to kill) and sterben (to die) or öffnen (to open) and offen (open) is specified as a causal relation. The causal relation can be encoded between all parts of speech. Little use is currently being made of the implication relation, the so-called entailment, which applies for example between gelingen (to succeed) und versuchen (to attempt). The meaning of a lexical unit is characterised by the sum of relations it holds to other lexical units and concepts. Furthermore, GermaNet encodes pertainymy as a kind of semantic derivational relation (such as finanziell (financial) and Finanzen (finance)), and associative links (see also) between concepts which cannot be captured by a . For further details cf. http://www.sfs.uni-tuebingen.de/GermanNet and Lemnitzer and Kunze (2002).



Lexical-semantic and conceptual relations in GermaNet 167

Figure 2.  Partial tree from the GermaNet hierarchy for öffnen (to open)

standard semantic relation (such as Weltrangliste (world ranking list) and Tennis (tennis) or Talmud (Talmud) and Judentum (Jewishness)). Figure 2 depicts the causative verb öffnen (to open) and its semantically related concepts. Synsets and lexical units are presented with their respective reading numbers from the GermaNet database. The connection of the synset öffnen_3, aufmachen_2 (to open) with its superordinate term wandeln_4, verändern_2 (to change) is represented by the upwards arrow, the correlation with its three hyponyms – aufstoßen_2 (to push open), aufbrechen_1 (to break open) and aufsperren_1 (to unlock) – by downwards arrows. The causal relation to the intransitive concept öffnen_1, aufgehen_1 (to open) is represented by the arrow with the dashed line. The lexical units are related to different antonyms: öffnen_3 (to open) is related to its antonym schließen_7 (to close), and aufmachen_2 (to open) to its antonym zumachen_2 (to shut). The bi-directionality of the antonymy relation is represented by a left-right arrow.

2.1.2 Cross-classification of concepts and artificial concepts A concept like Banane (banana), as well as a number of other fruits, can be classified both as Pflanze (plant) and as Nahrungsmittel (food) and can thus be assigned to different semantic fields. In order to access this kind of information, cross-classification of such concepts in different hierarchies is applied (see Figure 3).

. In terms of the major distinction being made, pertainymy is a lexical relation, whereas the associative relation is classed as a conceptual relation.

168 Claudia Kunze and Lothar Lemnitzer

Figure 3.  Cross-classification in GermaNet

Figure 4.  Use of artificial concepts

Wordnets are expected to represent only lexical units and concepts of a given language. In contrast to this, GermaNet also makes use of artificial concepts in order to improve the hierarchical structure and to avoid unmotivated co-hyponymy. Following Cruse’s approach (cf. Cruse 1986), co-hyponyms (concepts which share a common mother node) should be incompatible with one another. For example, the concepts Säugling (baby), Kleinkind (toddler), Vorschulkind (preschooler) and Schulkind (schoolchild) embody subordinate terms of Kind (child), which are mutually exclusive. The word field lehrer (teacher) contains hyponymic terms such as Fachlehrer (specialist subject teacher), Grundschullehrer (primary school teacher) and Konrektor (deputy head teacher), which cannot be represented meaningfully on a common hierarchy level. In order to model the partial network more symmetrically, and therefore more adequately, two artificial nodes have been created and introduced into the natural language hierarchy, namely ?Schullehrer (teacher in a certain type of school) and ?hierarchischer_Lehrer (teacher with a hierarchical position), as can be seen in Figure 4. Furthermore, GermaNet encodes subcategorisation frames for describing the syntactic complementation patterns of verbal predicates. As we are focusing in



Lexical-semantic and conceptual relations in GermaNet 169

this paper on the description of semantic relations, we refer the reader to the GermaNet homepage for details on verbal frames and example sentences for these frames (see http://www.sfs.uni-tuebingen.de/GermaNet/).

2.2

EuroWordNet – a polylingual wordnet

The GermaNet base vocabulary, comprising some 15,000 synsets, has been integrated into the polylingual EuroWordNet (http://www.hum.uva.nl/˜ewn/), which connects wordnets for eight European languages in a common architecture (see Vossen 1999). EuroWordNet models the most frequent and important concepts of English, Spanish, Dutch, Italian, French, German, Czech and Estonian. The Interlingual Index (ILI) serves as the core component of the database architecture, to which monolingual wordnets are linked. As a language-independent module, the ILI contains an unstructured list of ILI-records (taken from WordNet and therefore biased towards American English) which are labelled with a unique identifier. Concepts from different languages are related to suitable equivalents from the ILI via equivalence links. Matching of specific language pairs applies via the ILI, for example guidare – conducir (Italian – Spanish) regarding the concept to drive (fahren) in Figure 5. Language-independent EuroWordNet modules also include the so-called Top Ontology with 63 semantic features and the Domain Ontology which supplies semantic fields. All language-specific wordnets contain a common set of so-called Base Concepts, which consists of 1,000 nouns and 300 verbs. The base concepts function as the central vocabulary of polylingual wordnet construction, ensuring compatibility between single language-specific wordnets. Base Concepts

Figure 5.  EuroWordNet architecture

170 Claudia Kunze and Lothar Lemnitzer

are characterised by semantic features or feature bundles of the Top Ontology. For example, Werkzeug (tool) is characterised by the features artefact, instrument, object. Base Concepts dominate a number of nodes and/or a hierarchical multilevel chain of subordinates, or they constitute frequently occurring concepts in at least two languages. Base Concepts are on the one hand more concrete than the semantic features of the Top Ontology such as dynamic, function and property, but on the other hand more abstract than Rosch’s basic level concepts, such as Tisch (table) and Hammer (hammer) (see Rosch 1978). The adequate level of abstraction for Base Concepts is realised by superordinate concepts of Rosch’s Basic Level Concepts, for example Möbel (furniture) for Tisch (table) und Werkzeug (tool) for Hammer (hammer). After mapping the inventory of Base Concepts to the ILI, the Top Concepts and first-order hyponyms have been linked, thus yielding a first subset of some 7,500 concepts. The construction of language-specific wordnets could then be carried out independently, particularly as the Top Ontology inherits semantic features, allowing for a balanced coverage of different semantic fields across wordnets. Due to diverging lexicalisation patterns across languages arising from linguistic and cultural variation, and as a result of gaps in Princeton WordNet (as a basic resource feeding the ILI), it is not always possible to find matching equivalents for all language-specific concepts. Therefore, the EuroWordNet data model also supplies non-synonymous equivalence links, as well as combining them for the assignment of appropriate transfer terms. For the German concept Sportbekleidung, no synonymous target concept sports garment is provided by the ILI. Alternatively, two equivalence links can be established, one with the hyperonym garment (Kleidung), the other with the holonym sports equipment (Sportausrüstung). The international cooperation involved in constructing a polylingual wordnet architecture has brought about a quasi standard for wordnet development. It also functions as a model for adding further languages. In the year 2000, the Global WordNet Association was founded for fostering common research on wordnets. Several polylingual architectures go back to the EuroWordNet ILI, such as BalkaNet for some South Eastern European languages (cf. Tufiş et al. 2004) or CoreNet for Chinese, Korean and Japanese (see http://bola.or.kr/CoreNet_Project/).



Lexical-semantic and conceptual relations in GermaNet 171

3.

Adding syntagmatic relations

3.1

Motivation

It has recently been acknowledged that one of the shortcomings of wordnets is their relatively small number of syntagmatic relations, in terms of both types and individual instances. Some relations which cross the borders between the parts of speech are encoded, but they are few and far between, and in most cases they encode morphosemantic information. The prevalence of paradigmatic relations might be due to the origin of wordnets in the cognitive sciences, where paradigmatic relations are central and syntagmatic relations are only marginal. However, wordnets have grown out of their cognitive science context and are widely used in the field of natural language processing. It has recently been stated that in such application contexts, wordnets suffer from the relatively small number of relation instances between their lexical objects. It is assumed that applications in Natural Language Processing (NLP) and Information Retrieval (IR), in particular those relying on word sense disambiguation, can be boosted by a lexical-semantic resource with a higher relational density and, consequently, shorter average paths between the lexical objects. This situation also applies to the German wordnet. The lexical objects in GermaNet are connected by only 3,678 paradigmatic lexical relations between lexical units and 64,000 paradigmatic conceptual relations between synsets. Even if we count lexical objects which are related indirectly through an intermediate node, the network is not very dense, and most of the paths between the synsets and lexical units are very long. We have therefore decided in the context of a semantic information retrieval project, in which GermaNet plays a crucial role as a lexical resource, to extend the German wordnet with two types of syntagmatic relation. The first relation holds between verbs and the nominal heads of their subject noun phrases, and the second between verbs and the nominal heads of their direct object noun phrases. We decided to use terms for these relations which do not express or

. Morato et al. (2003) give an overview of applications of wordnets in general. . See, for instance, Boyd-Graber et al. (2006) and Fellbaum (2007). . “Semantic Information Retrieval” (SIR), a project funded by the German Science Foundation. Gurevych et al. (2007) and Gurevych (2005) describe the use of GermaNet in this project.

172 Claudia Kunze and Lothar Lemnitzer

imply a commitment to any syntactic theory. We therefore call the former relation “Arg1” and the latter relation “Arg2”. In the following, we report on our work on the acquisition of these two new types of (syntagmatic) relation. The sources which we have been using for this task are two large, syntactically annotated German corpora: a newspaper corpus and the German Wikipedia. Special attention was directed to the question of how the insertion of instances of this relation into GermaNet affects the neighbourhood of the nodes which are connected by an instance of the new relation. In particular, we observed whether there was a significant decrease in the sum total of path lengths which connect the newly related nodes and the nodes which are in the neighbourhood of these nodes. In the following section we will: (a) give an overview of research relating to syntagmatic relations in connection with wordnets; (b) describe in detail the acquisition and filtering process which leads to the extraction of relevant word pairs; (c) present the results of our experiments with path lengths over the GermaNet graph, and (d) draw some conclusions and outline future research.

3.2

Related work

Research into the (semi-)automatic detection and integration of relations between synsets has proliferated in recent years. This activity can be seen as a response to what Boyd-Graber et al. (2006: 29) identify as a weakness of the Prince­ ton WordNet: WordNet, a ubiquitous tool for natural language processing, suffers from sparsity of connections between its component concepts (synsets).

Research into the (semi-)automatic acquisition and integration of new synsets aims to reduce the amount of time-consuming and error-prone manual work required without these methods. Snow et al. (2006) and Tjong Kim Sang (2007) present highly efficient means of carrying out this task. They exploit the fact that taxonomic relations between lexical objects are reflected in the distributional patterns of these lexical objects in texts. These efforts are, however, directed towards paradigmatic relations, the hyperonymy/hyponymy relation in particular. They are therefore less relevant for our acquisition task, though we might use them in the future to extend GermaNet with more instances of these paradigmatic relations. Some effort has already been made to introduce non-classical, cross-category relations into wordnets. Boyd-Graber et al. (2006) introduce a type of relation



Lexical-semantic and conceptual relations in GermaNet 173

which they call “evocation”. This relation expresses the fact that the source concept as a stimulus evokes the target concept as a consistent human response. In other words, this is a mental relation which cuts across parts of speech. This approach nevertheless differs from ours, as we use corpus data instead of human response data and we acquire what is in the texts rather than what is supposed to be in the human mind. The relation we introduce is syntactically motivated, which is not the case in the experiment on which Boyd-Graber et al. report. Amaro et al. (2006) attempt to enrich wordnets with predicate-argument structures, where the arguments are not real lexical units or synsets but rather abstract categories such as instrument. Their aim is a lexical-semantic resource which supports the semantic component of a deep parser. This motivates their introduction of a highly abstracted categorisation of these arguments. This is not what we intend to do. Our data might lend themselves to all kinds of abstraction, as we will explain later on, but our primary intention is to capture phenomena which are on the surface of texts. Yamamoto and Isahara (2007) extract non-taxonomic, in particular thematic relations between predicates and their arguments. They extract these related pairs from corpora by using syntactic relations as clues. In this respect, their work is comparable to ours. Their aim, namely to improve the performance of information retrieval systems with this kind of relation, is also comparable to ours. However, they do not intend to include them in a lexical-semantic resource. Closest to our approach is the work of Bentivogli and Pianta (2003). Their research is embedded in the context of machine translation. Seen from this perspective, the almost exclusive representation of single lexical units and their semantic properties is not sufficient. They therefore propose modelling the combinatory idiosyncrasies of lexical units by two means: (a) the phraset as a type of synset which contains multi-word lexical units, and (b) syntagmatic relations between verbs and their arguments as an extension of the traditional paradigmatic relations. Their work, however, focuses on the identification and integration of phrasets. They only resort to syntagmatic relations where the introduction of a phraset would not otherwise be justified. We take the opposite approach, in that we focus on the introduction of instances of the verb-argument relation and resort to the introduction of phrases only in those cases where it is not possible to ascribe an independent meaning to one of the lexical units (see below).

3.3

The corpora

The acquisition of the new word pairs and their relation is based on two large German corpora: (a) the Tübingen Partially Parsed Corpus of Written German

174 Claudia Kunze and Lothar Lemnitzer

Figure 6.  Parse tree of a TüPP-D/Z sentence

(TüPP-D/Z), and (b) the German Wikipedia. The first corpus contains approximately 11.5 million sentences and 204 million lexical tokens, and the second corpus contains 730 million lexical tokens. Both corpora have been linguistically annotated using the cascaded finite-state parser KaRoPars which had been developed at the University of Tübingen (cf. Müller 2004) and a modified version of the BitPar PCFG parser (cf. Schmid 2004 and Versley 2005). The results of the automatic linguistic analysis, however, have not been corrected manually, due to the sheer size of the corpora. To illustrate the explanations of the linguistic annotation, we present two example sentences. The first one, displayed in Figure 6, is from the newspaper corpus, and the second, in Figure 7, is from Wikipedia. The first sentence translates as “We need to sell the villas in order to pay the young scientists” where the accusative object Villenverkauf of the verb brauchen (to need) means “sale of the villas”. It is a complex noun and not very frequent in either corpus. The sentence in Figure 7 translates as “He gets the most inspiration from his father” where the accusative object Inspiration occurs in front of the subject er (he). The parsers analyse and mark four levels of constituency.

. We are aware of the vagueness of the term German Wikipedia. The Wikipedia is a moving target. To be more precise, we are using a downloadable snapshot of the Wikipedia which was available on the pages of the Wikimedia Foundation in September 2008. Since we are interested in recurrent lexical patterns only, our experiments are not sensitive to the subtle changes in the database which are caused by the permanent editing processes.



Lexical-semantic and conceptual relations in GermaNet 175

Figure 7.  Parse tree of a Wikipedia sentence

The lexical level. For each word in both examples, the part of speech is specified using the Stuttgart-Tübingen-Tagset (STTS, cf. Schiller et al. 1999) which is a de facto standard for German. In the examples, some words are marked as heads of their respective chunks (“HD”), e.g. Villenverkauf and Inspiration. The chunk level. Chunks are non-recursive constituents and are therefore simpler than phrases. The use of chunks makes the overall syntactic structure less complex in comparison to deep parsing. In Figure 7 we have two noun chunks (labelled NCX) and one verb chunk (labelled VXVF). These are the categories which we need for our acquisition experiments. The functional specification of the noun chunks is of the utmost importance. NPs are, with regard to the predicate of the sentence, marked as subject NPs (in the nominative case, ON) or direct object NPs (in the accusative case, OA). The level of topological fields. The German clause can structurally be divided into the verbal bracket, i.e. one part of the verb in second position and the other part at the end of the clause, while the arguments and adjuncts are distributed more or less freely over the three fields into which the verbal bracket divides the clause: Vorfeld (VF), in front of the left verb bracket, Mittelfeld (MF), between the two brackets, and Nachfeld (NF), following the right bracket. The clausal level. The clause in Figure 7 is labelled as a simplex clause (SIMPX). In the example in Figure 6, “S” is the label at the root of the sentence. The parse tree which is generated by KaRoPars (i.e. the example in Figure 6) has a flat structure. Due to limitations of the finite-state parsing model, syntactic relations between the chunks remain unspecified. Major constituents, however,

176 Claudia Kunze and Lothar Lemnitzer

are annotated with grammatical functions in both cases. This information, together with the chunk and head information, is sufficient to extract the word pairs which we need for our experiments.

3.4

The acquisition experiment

From the linguistically analysed and annotated corpora, which we have described above, we extracted two types of syntactically related word pairs: (a) verb-subject (e.g. untersuchen, Arzt (to examine, doctor)) and (b) verb-direct object (e.g. finden, Weg (to find, way)). While the spectrum of possible subjects of a verb turned out to be very broad and heterogeneous, verb-object pairs were more readily identifiable and recurrent. Even in scenarios in which associations are derived on the basis of evocation, it is the case that a higher number of associations are arrived at by humans between verbs and their direct objects than between verbs and their transitive or intransitive subjects.10 We therefore focused our work on the analysis of verb-object pairs, taking only few verb-subject pairs into account. In order to rank the word pairs, we measured their collocational strength, which we consider to be a good indicator of their semantic relatedness. Two common measures – mutual information (MI), cf. Church et al. (1991), and log-likelihood ratio, cf. Dunning (1993) – are used and compared in our experiments. Mutual information can be regarded as a measurement of how strongly the occurrence of one word determines the occurrence of another; it compares the probability of, for example, two words occurring together with the probability of observing them independently of one another. Log-likelihood ratio compares expected and observed frequencies as might be expressed in a contingency table, a two by two table where the four cell values represent the frequency of word x occurring with word y, x and not y, not x and y and finally not x and not y, i.e. the number of observations where neither word appears. Mutual information seems to be the better choice for the extraction of complex terminological units, due to the fact that it assigns a higher weighting to word pairs where the partners occur infrequently throughout the corpus. Log-likelihood ratio, in contrast, is not sensitive to low occurrence rates of individual words and is therefore more appropriate for finding recurrent word combinations. This coincides with the findings which are reported by Kilgarriff (1996) and, for German, by Lemnitzer and Kunze (2007).

10. Such observations are reported by Sabine Schulte im Walde, cf. Schulte im Walde (2006).



Lexical-semantic and conceptual relations in GermaNet 177

Before encoding the new relations in GermaNet, we had to remove word pairs which we did not want to be inserted into the wordnet from the lists, for various reasons: – Pairs with wrongly assigned words due to linguistic annotation errors were removed. As has been stated above, the linguistic annotation was performed automatically and therefore produced errors. Some of these errors are recurrent and lead to word pairs which are highly ranked. For example, the auxiliary verb werden is very often classified as a full verb.11 So, from a sentence Er wird eine Aussage machen (He will make a statement), the pair werden – Aussage is erroneously extracted, instead of the correct pair machen – Aussage. – Word pairs which are fixed expressions or parts of them were also removed. It is well known that many idioms expose syntactic as well as semantic irregularities. In particular, it is impossible to assign a standard or literal meaning to the individual words. The idiom den Löffel abgeben (to hand in the spoon), for example, has nothing to do with cutlery, but is a colloquial expression for sterben (to die). Some of the word pairs are ambiguous. For example, the word pair spielen – Rolle can refer literally to somebody acting on a stage, or idiomatically to something which is important. The literal meaning can and should be captured by a relation connecting the word pair, whereas in the idiomatic meaning, this pair has to be treated as a single, complex lexical unit. – Support verb constructions were also discarded. It has been argued convincingly by Storrer (2006) that some types of support verb constructions and their verbal counterparts, e.g. eine Absage erteilen (literally to give a rejection) and absagen (to reject), show subtle differences in their use, and support verb constructions therefore merit an independent status as complex lexical units. Besides, it is often very difficult to assign a meaning to the verbal part of the construction, as a result of the mere supporting function of this element which is chosen for this role with some arbitrariness.12 For these reasons, we consider it to be inappropriate to represent (semi-) fixed expressions by relating their elements. They are very close to what Bentivogli and Pianta call “phrasets” (cf. Bentivogli and Pianta 2003). We see the need to also encode these complex multi-word expressions in GermaNet, but a discussion of a good strategy for doing this is beyond the scope of this article.

11. Which, in rare cases, it is. 12. For some constructions, it might be possible to assign a meaning to the supporting verb, e.g. Aufnahme finden (literally to find acceptance), but these cases are rare and border on free construction.

178 Claudia Kunze and Lothar Lemnitzer

Table 1.  Cumulative path length reduction, average of 100 word pairs  for both MI and G2 Method

Average PR value

MI G2

  2762.04 15867.38

In order to design a manageable experimental setting by which the impact of the new relations can be measured, we decided to insert the 100 top-ranked of the remaining word pairs manually. Inserting the word pairs involved a manual disambiguation where all words were mapped to the correct synsets. Semi-automatic insertion of the new relation instances would require reliable word sense disambiguation which is not yet available for German. In the following, we report on experiments in which we calculated the local impact of the new relation instances.13

3.5

Measurable local effects of the new relations

In our experiments,14 we wanted to measure the impact of newly introduced relations between verbs and nouns on the path lengths between the related words and the words which are in their neighbourhood. First, we introduced the relation instances which connect the lexical objects which we had acquired through our collocational analyses. We inserted these relations one by one. The settings for measuring the impact of each new relation are as follows: let s1 and s2 be two synsets and R (s1, s2) the new relation instance connecting the two synsets. Further, let SPb be the shortest path between s1 and s2 before the insertion of R (s1, s2) and let SPa be the shortest path between s1 and s2 after the insertion of R (s1, s2). By definition, the length of the shortest path between s1 and s2 after the insertion of (SPa) is 1 (see Figure 8). We calculate the path reduction PRs1,s2 the result of SPb – SPa. We now take S1 and S2, the sets of all synsets which are in the two sub-trees rooted by s1 and s2 respectively; in other words, we take all the hyponyms, the hyponyms of these hyponyms and so forth. 13. We also measured the global impact of the new relations, i.e. the impact of these links on the overall reduction of path length between any two nodes. There are, however, no visible effects and the selected word pairs do not have any impact which is different from the impact of the same number of relations inserted between randomly chosen lexical objects. For details, cf. Lemnitzer et al. (2008). 14. We are grateful to Holger Wunsch and Piklu Gupta, who have run some of the experiments which we report here.



Lexical-semantic and conceptual relations in GermaNet 179

Figure 8.  Local path reduction between two synsets s1 and s2. The dashed path is the old path. The new relation R (s1, s2) is depicted by the thick line between s1 and s2

We calculate the path reduction PRsm,sn for each pair sm ∈ S1, sn ∈ S2. The sum of all path reduction values is the local impact caused by the new relation instance. We calculated the sum total of the path reduction values for the 100 most highly ranked pairs according to the mutual information and the log-likelihood statistics. Table 1 shows the average cumulative path reduction value for both statistics. From these figures we can infer that: (a) there is a considerable local impact of the new relation instances, which is what we wanted to achieve, and (b) the impact of the word pairs extracted by log-likelihood ratio is much higher than that of the pairs extracted by mutual information. This confirms our assumption about the superiority of log-likelihood ratio for our acquisition task. By the new relations, we connect pairs of words in the wordnet which exhibit a certain kind of closeness through their frequent co-occurrence in texts. This is in contrast to the traditional paradigmatic relations. These relations connect words which are supposed to be related in the mental lexicon, but which seldom co-occur in texts. We assume that the syntagmatically related word pairs also have an organising function in the mental lexicon, but this is not our primary concern. The important point is that a verb like verschreiben (to prescribe), when occurring in a text, triggers a whole set of words which denote medical objects, but only when used in one particular reading. This is important for NLP tasks such as word sense disambiguation and anaphora resolution.

3.6

From collocation to semantic preference

It is collocations, or, in the terms of British contextualism, colligations, which have so far been encoded. We have linked verbal predicates with the nominal heads of their arguments. Collocations as well as colligations are also part of the

180 Claudia Kunze and Lothar Lemnitzer

Figure 9.  Collocates of the verbal predicate beseitigen

description of the meaning of a lexical item as proposed by John Sinclair (1996). Another important aspect of the meaning of a lexical item is, according to Sinclair and his followers’ theory of meaning, the semantic preference of the lexical item: the term semantic preference refers to the (probabilistic) tendency of certain units of meaning to co-occur with items from the same semantic subset, items which share a semantic feature. (Bednarek and Bublitz 2007: 122)

It is possible, from the collocation we have encoded in GermaNet, to arrive at the more abstract level of semantic preference. The example presented in Figure 9 should illustrate this point. In the centre is the lexical unit beseitigen (to remove). We have grouped four of the many direct objects around this verbal predicate: Fehler (mistake, error), Mangel (fault), Verschmutzung (pollution) and Gefahrenquelle (source of danger). These few examples can be grouped along the semantic feature abstract (Fehler, Mangel) vs. concrete (Verschmutzung, Gefahrenquelle). With these few examples we have identified one of the semantic dimensions which distinguish the collocating objects of beseitigen, i.e. unpleasant (abstract) states and (concrete) objects. Currently, we have encoded 9,400 instances of the new relations: 1,800 Arg1-relations and 7,600 Arg2-relations. We have not yet performed this abstraction step on a broader scale, but we are optimistic that such a step can be based on the syntagmatic relation data we have acquired and encoded into GermaNet.



Lexical-semantic and conceptual relations in GermaNet 181

References Amaro, Raquel, Chaves, Rui Pedro, Marrafa, Palmira and Mendes, Sara. 2006. “Enriching wordnets with new relations and with event and argument structures.” In Computational Linguistics and Intelligent Text Processing – 7th International Conference, CICLing-2006, LNCS 3378, Alexander Gelbukh (ed.), 28–40. Berlin/Heidelberg: Springer. Bednarek, Monika and Bublitz, Wolfram. 2007. “Enjoy! – The (phraseological) culture of having fun.” In Phraseology and Culture in English, Paul Skandera (ed.), 109–135. Berlin/New York: de Gruyter. Bentivogli, Luisa and Pianta, Emanuele. 2003. “Extending WordNet with Syntagmatic Information.” In Proceedings of the Second International WordNet Conference – GWC 2004, Masaryk University Brno, Czech Republic, Petr Sojka, Karel Pala, Pavel Smrz, Christiane Fellbaum and Piek Vossen (eds), 47–53. Boyd-Graber, Jordan, Fellbaum, Christiane, Osherson, Daniel and Schapire, Robert. 2006. “Adding Dense, Weighted Connections to WordNet.” In Proceedings of the Third International WordNet Conference, Jeju Island, Korea, Petr Sojka, Key Sun-Choi, Christiane Fellbaum and Piek Vossen (eds), 29–36. Church, Kenneth, Gale, William, Hanks, Patrick and Hindle, Donald. 1991. “Using Statistics in Lexical Analysis.” In Lexical acquisition: exploiting on-line resources to build a lexicon, Uri Zernik (ed.), 115–164. Hillsdale, NJ: Laurence Erlbaum. Cruse, Alan. 1986. Lexical Semantics. Cambridge: Cambridge University Press. Dunning, Ted. 1993. “Accurate methods for the statistics of surprise and coincidence.” Computational Linguistics 19(1): 61–74. Fellbaum, Christiane. 1998. WordNet: An Electronic Lexical Database. Cambridge, MA: MIT Press. Fellbaum, Christiane. 2007. Wordnets: Design, Contents, Limitations. http://dydan.rutgers.edu/ Workshops/Semantics/slides/fellbaum.pdf. Gurevych, Iryna. 2005. “Using the structure of a conceptual network in computing semantic relatedness.” In Proceedings of the 2nd International Joint Conference on Natural Language Processing (IJCNLP’2005), Jeju Island, Republic of Korea, 767–778. Gurevych, Iryna, Müller, Cristof and Zesch, Thorsten. 2007. “What to be? – electronic career guidance based on semantic relatedness.” In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics, Prague, Czech Republic, 1032–1039. Association for Computational Linguistics. Hamp, Birgit and Feldweg, Helmut. 1997. “GermaNet – a Lexical-Semantic Net for German.” In Proceedings of the ACL/EACL-97 workshop on Automatic Information Extraction and Building of Lexical-Semantic Resources for NLP Applications, Madrid, Piek Vossen, Nicoletta Calzolari, Geert Adriaens, Antonio Sanfilippo, and Yorick Wilks (eds), 9–15. Kilgarriff, Adam. 1996. “Which words are particularly characteristic of a text? A survey of statistical approaches.” In Proceedings of AISB Workshop on Language Engineering for Document Analysis and Recognition, 33–40. Falmer: Sussex. Kunze, Claudia. 2004. “Lexikalisch-semantische Wortnetze.“ In Computerlinguistik und Sprach­ technologie: eine Einführung, Kai-Uwe Carstensen, Christian Ebert, Cornelia Endriss, Susanne Jekat and Ralf Klabunde (eds), 386–393. Heidelberg/Berlin: Spektrum Verlag.

182 Claudia Kunze and Lothar Lemnitzer

Lemnitzer, Lothar and Kunze, Claudia. 2002. “Adapting GermaNet for the Web.” In Proceedings of the first Global WordNet Conference, Central Institute of Indian Languages, 174–181. Mysore, India. Lemnitzer, Lothar and Kunze, Claudia. 2007. Computerlexikographie. Tübingen: Gunter Narr Verlag. Lemnitzer, Lothar, Wunsch, Holger and Gupta, Piklu. 2008. “Enriching GermaNet with verbnoun relations – a case study of lexical acquisition.” In Proceedings LREC 2008, Marrakech, Marokko. Miller, George. 1990. “Special Issue: WordNet – An on-line lexical database.” International Journal of Lexicography 3(4). Morato, Jorge, Marzal, Miguel, Llorens, Juan and Moreiro, José. 2003. “WordNet Applications.” In Proceedings of the Second International WordNet Conference – GWC 2004, Masaryk University Brno, Czech Republic, Petr Sojka, Karel Pala, Pavel Smrz, Christiane Fellbaum and Piek Vossen (eds), 270–278. Müller, Frank H. 2004. Stylebook for the Tübingen Partially Parsed Corpus of Written German (TüPP-D/Z). Sonderforschungsbereich 441, Seminar für Sprachwissenschaft, Universität Tübingen. Quillian, M. Ross. 1966. Semantic Memory. Unpublished Ph.D. thesis, Carnegie Institute of Technology. Rosch, Eleanor 1978. “Principles of Categorization.” In Cognition and Categorization, Eleanor Rosch and Barbara B. Lloyd (eds), 27–48. Hillsdale, NJ: Lawrence Erlbaum. Schiller, Anne, Teufel, Simone, Thielen, Christine and Stöckert, Christine. 1999. Guidelines für das Taggen deutscher Textcorpora mit STTS (Kleines und großes Tagset). IMS, Universität Stuttgart. (http://www.ims.uni-stuttgart.de/projekte/corplex/TagSets/stts-1999.pdf) Schmid, Helmut. 2004. “Efficient Parsing of Highly Ambiguous Context-Free Grammars with Bit Vectors.” In Proceedings of the 20th International Conference on Computational Linguistics (COLING 2004), Geneva, Switzerland. Schulte im Walde, Sabine. 2006. “Can Human Verb Associations Help Identify Salient Features for Semantic Verb Classification?” In Proceedings of the 10th Conference on Computational Natural Language Learning, New York City, NY. Sinclair, John. 1996. The search for units of meaning. Textus 59(IX), 75–106. Snow, Rion, Jurafsky, Dan and Ng, Andrew. 2006. “Semantic taxonomy induction from heterogenous evidence.” In Proceedings of the 21st International Conference on Computational Linguistics and the 44th annual meeting of the ACL, Morristown, NJ, 801–808. Association for Computational Linguistics. Storrer, Angelika. 2006. “Funktionen von Nominalisierungsverbgefügen im Text. Eine korpusbasierte Fallstudie.“ In Von Intentionalität zur Bedeutung konventionalisierter Zeichen. Festschrift für Gisela Harras zum 65. Geburtstag, Kristel Proost and Edeltraud Winkler (eds.), 147–178. Tübingen: Gunter Narr. Tjong Kim Sang, Erik. 2007. “Extracting hypernym pairs from the web.” In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics. Companion Volume Proceedings of the Demo and Poster Sessions, Prague, Czech Republic, 165–168. Association for Computational Linguistics. Tufiş, Dan, Cristea, Dan and Stamou, Sofia. 2004. “BalkaNet: Aims, Methods, Results and Perspectives.” Romanian Journal of Information Science and Technology 7(1–2): 9–45. Versley, Yannick. 2005. “Parser Evaluation across Text Types.” In Proceedings of the Fourth Workshop on Treebanks and Linguistic Theories (TLT 2005), Barcelona, Spain.



Lexical-semantic and conceptual relations in GermaNet 183

Vossen, Piek. 1999. EuroWordNet: A mutlilingual database with lexicalsemantic networks. Dordrecht: Kluwer Academic Publishers. Wahrig, Gerhard, Krämer, Hildegard and Zimmermann, Harald. 1980–1984. Brockhaus­Wahrig deutsches Wörterbuch. 6 vols. Wiesbaden/Stuttgart: Deutsche Verlags-Anstalt. Wehrle, Hugo and Eggers, Hans. 1989. Deutscher Wortschatz. Ein Wegweiser zum treffenden Ausdruck. Stuttgart: Ernst Klett. Yamamoto, Eiko and Isahara, Hitoshi. 2007. “Extracting word sets with nontaxonomical relation.” In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics. Companion Volume Proceedings of the Demo and Poster Sessions, Prague, Czech Republic, 141–144. Association for Computational Linguistics.

Index

A abduction  117 agent classificatory assessment agent  121f. spontaneous-associative agent  121f. analysis of variance  31 anchor word  23 antonymy ancillary antonymy  50f., 60 antonym / antonymic pair  15ff., 49ff. antonymous groups  103, 112 comparative antonymy  52 coordinated antonymy  51, 55, 60f. direct antonyms  20 distinguished antonymy  52 extreme antonymy  52 goodness of antonymy  15ff., 21, 38f., 41f. idiomatic antonymy  52 indirect antonyms  20 interrogative antonymy  52 negated antonymy  52 transitional antonymy  52 association  24, 81f., 120f. associative link  166 B base concept  see basic level concept basic level concept  169f. Belica, Cyril  124, 127, 130, 133, 138, 142 bidirectionality  26, 153 bidirectional evidence  36 bidirectional link  165 bidirectional relation  26, 146, 155 unidirectional evidence  36

British National Corpus   see corpus C canonicity  21, 24ff., 33ff., 54 canonical antonyms  16ff., 20ff., 31ff., 44, 54f. canonical status  35 non-canonical antonyms  16f., 38, 40, 42, 54, 96 categories of being  118 causality  75ff. causal / causative relation  75ff., 166f. causation  75ff., 166 cause-effect  75ff. CCDB  see corpus chunk  175 Church, Kenneth G.  11, 120, 176 cluster analysis  26ff. coarse-grained structure   see structure Coco  see corpus tool cognitive approach  7 language structures  119 mechanism / principle  70, 92f. processing  122 co-hyponymy  168 collocation  6, 115ff., 120ff., 127ff., 179f. see also co-occurrence collocational context  120 collocation algorithm  124 collocational profile  10, 72, 126ff. higher-order collocation  124f., 127, 130 statistical collocation  120, 124

communicative attitude  96f. comparative function  53 complementarity  95, 102ff. computational linguistics  10, 163 concept  73ff., 101, 164ff. conceptual closeness  79, 82 domain  16, 102 implication  74ff., 84, 90f. modelling  165 operation  75 relation  72, 76, 165f., 171 conceptualisation  73ff. concordance  56f., 61, 64 concordancer  see corpus tool concgram  124 condition  78ff., 103 conditionality  78ff. conditional relation  79 connotation  118ff., 133 consistency  145ff. construal  7, 16, 71 construction  7, 49ff., 60ff., 69f., 85ff. ancillary construction  64 construction grammar  61, 65 dynamic contextual construction  73 lexico-syntactic construction  see frame context  120ff. global context  120f., 124, 127, 129f., 130ff., 137ff., 141f. local context  121, 124, 127 situational context  120 contrariety  112 control group  17 co-occurrence  7, 19ff., 49ff., 59f., 115ff., 124ff., 179 see also collocation

186 Lexical-Semantic Relations

sentential co-occurrence  19, 21, 27, 35, 44 coordinated framework / structure  see frame corpus  8ff., 15ff., 49ff., 71, 115ff., 124ff. British National Corpus  60 CCDB  127ff. corpus-based  17f. corpus-driven  17f., 71f., 118ff. German Reference Corpus / DeReKo  127f. corpus tool  9, 19ff., 56, 64, 127ff. concordancer  56f., 64 Coco  19ff. cross-reference  152ff. Cruse, Alan D.  6f., 15f., 69ff., 73, 75, 77, 82, 89, 93, 95, 102, 168 D data data-driven  20 data model  152ff., 169 data structure  152ff. database  50, 56, 71, 155f., 165, 167, 169 denotation  118, 120 DeReKo  see corpus deviation  33f., 39 dichotomy  16f., 39 dichotomous view  39 dictionary  9f., 21, 96, 145ff. corpus-based dictionary  147, 153 editing system  159 electronic dictionary  152ff. elexiko  145, 147ff., 160ff. inconsistency  144ff. lexicographic database / resource  155ff., 165ff. lexicographic practice  9f., 145ff. machine-tractable dictionary  163 printed dictionary  147f., 151 usage  146ff. discourse  49ff., 85ff., 119ff., 127 function  50ff., 85ff. situation  110f.

domain ontology  169 DTD  153f. E elexiko  see dictionary elicitation test  17, 23ff. Elman, Jeffrey L.  119, 123 emergence  118f., 123 emergentist perspective  118f. emotive attitude  98 empiricism empiricist perspective  116 entailment  see implication entrenchment  118 epistemic attitude  96, 98f., 105 EuroWordNet  see wordnet evaluation  101, 106ff., 110f. evaluative attitude  98 event type  see speech act evidence-based theory  117 extended link   see text-technology F factivity  99 falsification  117 Fellbaum, Christiane  15, 85, 163f. field lexical-semantic field  5f., 96, 101ff., 150, 167, 169f. topological fields  175 fine-grained structure   see structure Firth, John R.  119f. frame  50ff., 85ff. co-ordinated frame  51ff., 86ff. lexico-semantic frame  51, 85ff. lexico-syntactic frame  54, 65, 85ff. subordinated frame  89ff. syntactic frame / sequence  85ff. fuzzy similarity  133 G GermaNet  see wordnet general resource situation type  see speech act

generalisation  122 gradability  95ff. gradable antonyms  95ff., 102ff. Gross, Derek  16f., 54, 65 Gruber, Tom  116 H Harras, Gisela  96f., 111 Herrmann, Douglas J.  17, 21ff., 35, 38f., 41, 54, 65 hierarchical structure  27, 168 high-dimensional space  128f. Hoey, Michael  119f. Hofmann, Thomas R.  95, 102 homonymy  164 Hopper, Paul J.  118f. hypo-/hyperonymy  75, 82f., 166, 172 hyperonym  75, 82f., 166, 170 hyponym  167f., 170 I implication  71ff., 166 entailment  72, 75ff., 95, 102, 166 unilateral entailment  77 inconsistency  see dictionary indeterminacy  77, 80 induction  117 information retrieval  163, 171ff. institutional setting  101 interrelation  80, 89 interrogative function  53 J Jones, Steven  8, 17, 49ff., 65, 69, 85, 87f. judgement experiment  15ff., 28ff. Justeson, John S.  17, 19, 49, 54, 85 K Katz, Slava M.  17, 19, 49, 54, 85 Keibel, Holger  118, 121, 124, 127, 130, 133 knowledge conceptual knowledge 16, 73ff. encyclopaedic knowledge  7



implicit knowledge  73ff., 120 metalinguistic knowledge  41, 73 Kupietz, Marc  118, 121, 124 L Lang, Ewald  95, 102, 112 language competence  117, 123 convention  119, 122 experience  15ff., 71ff., 118ff., 133, 137, 140 use  8, 18, 49ff., 69ff. 115ff. Lehrer, Adrienne  37, 95, 102, 168 lexical association  24 field  see field lexicalisation 72f., 96f., 101, 103, 106f., 132 lexico-semantic frames   see frame linkbase (link database)  155ff. Löbner, Sebastian  95, 102 logical relation  70, 78f. Lutzeier, Peter R.  6, 70, 111 Lyons, John  6, 15, 70, 77, 102 M markedness theory  37f. meaning connotational meaning  118ff., 133 construed meaning  119 context-dependent meaning  118 denotational meaning  118, 120 representation of meaning  115f., 123 theory of meaning  115f., 120, 180 meaning equivalent   see synonymy mental lexicon  7, 10, 179 mental representation  71 see also representation meronymy  166 metaphorisation  16 metonymisation  16

Index 187

Mettinger, Arthur  54, 96 Miller, George, A.  54, 163f. Miller, Katherine, J.  16f. modelling  145ff., 155, 159, 165, 173 concept  155, 159 parsing model  175 de Mönnink, Inge  18, 39f. Murphy, M. Lynne  8, 15ff., 49ff., 61f., 65f., 69f., 72, 83, 85, 93

Proost, Kristel  95f., 103 proposition  83, 96, 111, 118 propositional attitude  96ff., 105ff. content  96ff., 101, 104ff. prototypicality effect  16 psycholinguistic  23ff., 34ff. approach  69 experimental data  18 technique  34 purpose-orientation  75, 81f., 92

N natural language processing  10, 163f., 171f. near synonyms  see synonymy negation  107ff.

Q quantitative linguistics  130 temporal reference  97f., 101, 104, 106, 108ff.

O online resource  148, 163 CCDB  see corpus DeReKo  see corpus elexiko  see dictionary lexical knowledge base  163 lexical-semantic resource  163ff., 171ff. ontology  115ff., 166, 169f. domain ontology  169f. top ontology  169f. opposition  see also antonymy binary opposition  51 good opposites  16, 38 word-internal oppositeness  110f. overlapping semantic range  20 P Paradis, Carita  5, 8, 10, 16, 18, 35, 42, 49f., 53, 55f., 65, 119 parse tree  174f. Partington, Alan  10, 69, 82, 91, 120 pattern  see frame perception  71, 73ff., 121 plesionymy  see synonymy polysemy  133, 136 pragmatic function  50 predicate  173, 175, 180 re-reactive predicate  100 presupposition  96ff., 108ff. priming  30, 55, 62ff., 141

R reference  74, 80, 83 database  see GermaNet corpus  see DeReKo / corpus link reference  151ff. structure   see text-technology temporal reference  97 relation syntagmatic relation  6, 171ff., 180 taxonomic relation  172 thematic relation  173 relational coercion  61 repeated measures  31, 33 representation  7, 71, 85, 92f., 101, 164f. response word  24ff., 35, 41 see also stimulus word S scale  17, 24ff., 37f., 51f., 54f., 59, 70, 95, 102 self-organising lexical feature maps / SOMs  133 semantic categorisation  42, 165ff. dimension  35 element / feature  8, 70, 96ff. 118, 180 field  see also field inclusion  75, 77 scale  51f., 54

188 Lexical-Semantic Relations

sense disambiguation  123, 163f., 171, 178f. sequencing  30 significance level  19 similarity  69ff., 121ff., 129ff. see also synonymy Sinclair, John  119f., 180 situational roles  110f. situation type  see speech act speaker’s intention  96f., 99, 104ff., 109 speech act general resource situation type  96f. event type  97f., 101, 104, 106, 108ff. situation type  96f., 104, 106ff. speech act verb  95ff., 99ff., 106ff. stimulus word  23ff., 35ff., 45, 55f. Storjohann, Petra  10, 49, 65, 69, 85, 137, 149 structure coarse-grained structure  116f., 122f. fine-grained structure  116f., 122f., 133 structuralism  5f., 70

superordination   see hyperonymy synonymy  68ff., 130ff. coordinated synonymy  85f. meaning equivalents  69ff., 84ff., 136ff., 147f. near-synonyms  91, 136ff. part-of-whole synonymy  75, 84 plesionymy  137 see also near-synonyms subordinated synonymy  85, 89 synonym cluster  85, 89 synset  164ff. systematisation, systematicity  121, 123 T taxonomic structure  133 Taylor, John  91, 137 template  see frame text-technology extended link  157f. reference structure  159 target resource  155ff., 159 traversal rule  157f. XLink  154ff. XML-structure  152ff. theory of meaning   see meaning

U usage-based  7, 54, 71, 115, 118f. V Vachková, Marie  121, 133, 138, 142 value descriptive value  116 explanatory value  117 functional value  116 W web-as-corpus approach  54 Willners, Caroline  10, 16f., 19f., 49f., 53, 55f. Winkler, Edeltraud  96, 182 wordnet EuroWordNet  169f. GermaNet  10, 161ff., 171f. language-specific wordnets  169f. WordNet  10, 163ff., 169f., 172 X XLink  see text-technology XML  see text-technology

In the series Lingvisticæ Investigationes Supplementa the following titles have been published thus far or are scheduled for publication: 28 Storjohann, Petra (ed.): Lexical-Semantic Relations. Theoretical and practical perspectives. 2010. viii, 188 pp. 27 Fradin, Bernard: La raison morphologique. Hommage à la mémoire de Danielle Corbin. 2008. xiii, 242 pp. 26 Floricic, Franck (dir.): La négation dans les langues romanes. 2007. xii, 229 pp. 25 Bat-Zeev Shyldkrot, Hava et Nicole Le Querler (dir.): Les Périphrases Verbales. 2005. viii, 521 pp. 24 Leclère, Christian, Éric Laporte, Mireille Piot and Max Silberztein (eds.): Lexique, Syntaxe et Lexique-Grammaire / Syntax, Lexis & Lexicon-Grammar. Papers in honour of Maurice Gross. 2004. xxii, 659 pp. 23 Blanco, Xavier, Pierre-André Buvet et Zoé Gavriilidou (dir.): Détermination et Formalisation. 2001. xii, 345 pp. 22 Salkoff, Morris: A French-English Grammar. A contrastive grammar on translational principles. 1999. xvi, 342 pp. 21 Nam, Jee-Sun: Classification Syntaxique des Constructions Adjectivales en Coréen. 1996. xxvi, 560 pp. 20 Bat-Zeev Shyldkrot, Hava et Lucien Kupferman (dir.): Tendances Récentes en Linguistique Française et Générale. Volume dédié à David Gaatone. 1995. xvi, 409 pp. 19 Fuchs, Catherine and Bernard Victorri (eds.): Continuity in Linguistic Semantics. 1994. iv, 255 pp. 18 Picone, Michael D.: Anglicisms, Neologisms and Dynamic French. 1996. xii, 462 pp. 17 Labelle, Jacques et Christian Leclère (dir.): Lexiques-Grammaires comparés en français. Actes du colloque international de Montréal (3–5 juin 1992). 1995. 217 pp. 16 Verluyten, S. Paul (dir.): La phonologie du schwa français. 1988. vi, 202 pp. 15 Lehrberger, John and Laurent Bourbeau: Machine Translation. Linguistic characteristics of MT systems and general methodology of evaluation. 1988. viii, 240 pp. 14 Subirats-Rüggeberg, Carlos: Sentential Complementation in Spanish. A lexico-grammatical study of three classes of verbs. 1987. xii, 290 pp. 13 Vergnaud, Jean-Roger: Dépendance et niveaux de représentation en syntaxe. 1985. xvi, 372 pp. 12 Hong, Chai-Song: Syntaxe des verbes de mouvement en coréen contemporain. 1985. xv, 309 pp. 11 Lamiroy, Béatrice: Les verbes de mouvement en français et en espagnol. Etude comparée de leurs infinitives. 1983. xiv, 323 pp. 10 Zwanenburg, Wiecher: Productivité morphologique et emprunt. 1983. x, 199 pp. 9 Guillet, Alain et Nunzio La Fauci (dir.): Lexique-Grammaire des langues romanes. Actes du 1er colloque européen sur la grammaire et le lexique comparés des langues romanes, Palerme, 1981. 1984. xiii, 319 + 58 pp. Tables. 8 Attal, Pierre et Claude Muller (dir.): De la Syntaxe à la Pragmatique. Actes du Colloque de Rennes, Université de Haute-Bretagne. 1984. 389 pp. 7 Taken from program. 6 Lightner, Ted: Introduction to English Derivational Morphology. 1983. xxxviii, 533 pp. 5 Paillet, Jean-Pierre and André Dugas: Approaches to Syntax. (English translation from the French original edition 'Principes d'analyse syntaxique', Québec, 1973). 1982. viii, 282 pp. 4 Love, Nigel: Generative Phonology: A Case Study from French. 1981. viii, 241 pp. 3 Parret, Herman, Leo Apostel, Paul Gochet, Maurice Van Overbeke, Oswald Ducrot, Liliane Tasmowski-De Ryck, Norbert Dittmar et Wolfgang Wildgen ess.: Le Langage en Contexte. Etudes philosophiques et linguistiques de pragmatique. Sous la direction de Herman Parret. 1980. iv, 790 pp. 2 Salkoff, Morris: Analyse syntaxique du Français-Grammaire en chaîne. 1980. xvi, 334 pp. 1 Foley, James: Theoretical Morphology of the French Verb. 1979. iv, 292 pp.