* * *

Introduction

The EU-project SIMPLE, Semantic Information for Multifunctional Plurilingual Lexica, is a follow-up of PAROLE, Preparatory Action for Linguistic Resources Organisation for Language Engineering. It has aimed at the development of wide-coverage semantic lexicons for the 12 European languages. The SIMPLE lexicons add information on a semantic layer to the information on morphological and syntactic layers encoded in PAROLE lexicons. All the lexicons share the same formal representation model, which is Entity/Relationship model, and the same DTD (Document Type Definition) implemented in SGML. The SIMPLE's theoretical and formal model is documented in Linguistic Guidelines, (Lenci et al. 1998) and the former knowledge of the guidelines is presupposed here.

Språkdata, Göteborg University, has participated in both the projects. The task of building the PAROLE and SIMPLE lexicons for Swedish has been to a large extent supported by the lexical resources elaborated by Department of Swedish Language, Språkdata, Göteborg University. The machine readable monolingual dictionaries Svenska ord (1992), Nationalencyklopedins ordbok (1995), and the lexical database, Göteborgs lexikaliska databas, (GLDB), provided a considerable amount of information on lexical data concerning morphology, syntax and semantics. GLDB is a monolingual lexical database with focus on modern Swedish. It is the most exhaustive source of information on modern Swedish. The PAROLE corpora and the corpora available in the Swedish Language Bank, Språkbanken, have been also frequently consulted. Moreover, the lexicographic and lexicological research carried at Språkdata has shown to be supportive with regard to questions concerning morphological, syntactic and semantic information in the PAROLE and SIMPLE lexicons (Gellerstam 1988, Toporowska Gronostaj 1996, Järborg 1989, Malmgren 1988).

To provide some background knowledge for this documentation, we start with a short overview of the conceptual model underlying the SIMPLE project, which is followed by some introductory notes on the Swedish PAROLE lexicon. The core section of this report deals with the documentation of the Swedish SIMPLE lexicon (hereafter Swe_S). We conclude the core section by presenting ways to automatically extend the content of Swe-S and showing its usability for different NLP applications. We reprint in the Appendix A a paper on this subject that will be presented at the LREC 2000 conference in Athens (Kokkinakis et al, 2000). In the the Appendix B of this document we provide some examples from the PAROLE and SIMPLE lexicons for Swedish to give an insight in the description modes and format used in these lexicons for the encoding of morphology, syntax and semantic of nouns, verbs and adjectives.

0.1 The SIMPLE project

The notion of semantic type is central for the SIMPLE model and its ontology. It corresponds roughly to a word sense assigned to a lexical item. There are 139 semantic types distinguished in the SIMPLE ontology. Each semantic type is defined as a cluster of structured semantic information significant for a given word sense. Information on semantic class, domain, argument structure of predicative expressions and selectional restrictions on arguments as well as qualia roles constitute a relevant part of the semantic type specification; (Lenci at al. 1998, Calzolari (1999), Pedersen & Keson (1999)). The SIMPLE ontology is multidimensional as it is based on the principle of orthogonal inheritance (Pustejovsky 1995), and in this respect, it contrasts with the LexiQuest’s semantic class ontology which is based on a standard, monodimensional approach. The latter ontology includes 95 semantic classes. Both ontologies are hierarchically structured. The above mentioned ontologies were designed for the semantic typing of verbs and nouns. For the semantic typing of adjectives, an ontology consisting of 14 semantic types has been used.

0.2 The Swedish PAROLE lexicon

The Swedish PAROLE lexicon is a language engineering resource which provides access to the lexically determined morphological and syntactic information on 20.000 Swedish words. It follows the conceptual and formal model formulated within the PAROLE project (Guimier, Ogonowski, 1998). A subset of the entries in the Swedish PAROLE lexicon, that is 5.621 words, has been enriched with the semantic information encoded in Swe-S.

The encoding of morphosyntactic data for Swedish has followed the model presented in the PAROLE report (WP-4.2.2b), Morphosyntactic Description of Swedish (Danielsson, Järborg, (1996)). In the Swedish PAROLE morphological lexicon there are 19.985 morphological units described according to 259 morphological modes. The morphological modes capture types of inflectional relations a morphological unit can display. Their descriptions vary according to the part of speech of a morphological unit and the morphological properties associated to its morphosyntactic type. To account for these, a technical stem and a set of relevant affixes have been postulated for each morphological mode.

A syntactic unit (Usyn) is a basic access point to the syntactic layer. In the Swedish PAROLE syntactic lexicon, there are 28.882 syntactic units linked to 19.985 morphological units. These syntactic units have their syntactic behaviour described by means of 438 syntactic descriptions. In the Swedish syntactic lexicon following aspects of lexically determined syntactic behaviour have been accounted for:

• subcategorization, including functions of subcategorized elements, information on their optionality, their position and morphosyntactic realisation, syntactic and semantic restrictions
• characteristics of the lexical unit when associated to a subcategorization frame
• control
• pronominalization

This lexically determined information has been considered as criterial for distinguishing different types of subcategorizations frames in Swedish, or in other words, for postulating 438 Description objects.

1 General design information

1.1 Lexicon population

Lexicon population in the Swe-S lexicon can be characterized as composed of two sets. The first one consists of 8.571 fully encoded lexicon items, which have been manually encoded. It covers nouns, verbs and adjectives. The second one consists of 25.000 partially encoded nouns, which have been automatically extracted from corpora, given 1.800 fully encoded words from the Swe-S as input. This set covers 22.000 compound- and 3.000 simplex-nouns. For details on the methodology used for extending the coverage of the Swe-S lexicon see Appendix A.

Overall statistics: covering the fully encoded Semus in Swe-S

Number of Semus linked to Usyns and Ums 8.571

Semu per category:

While populating the Swe-S lexicon manually, the preference was given to those lexical items which were considered as Swedish equivalents to the base concepts. The base concept set, common to all SIMPLE partners, is a selection of concepts derived from the EuroWordNet by means two operational criteria, namely, the number of relations (general or limited to hyponymy) and the high position of the concept in a hyperonym-hyponym hierarchy. Frequency was another significant factor taken into consideration.

As far as the population of the sample lexicon for Swedish is concerned, the selection of items included there aimed at:

1. presenting a representative sample of semantic types encoded in the Swe-S lexicon,
2. exhibiting semantic cohesion within the lexicon (e.g blomma 'to bloom', krokus 'crocus',
3. linking the syntactic and semantic information in the Swedish PAROLE and SIMPLE lexicons,
4. showing the distribution of nouns, verbs and adjectives among the semantic types.

1.2 Current lexicon contents

The fully encoded semantic data presented in the Swedish SIMPLE lexicon provide following information:

• semantic type, whose value is an element in the SIMPLE ontology list
  (e.g. SemU <katt> ’cat’: Earth_Animal).
   
• domain, whose value is an element in the LexiQuest's domain list
  (SemU <katt>:Zoology).
   
• semantic class, whose value is an element in the LexiQuest's semantic class list.
  (SemU <katt>: Mammal).
   
• glossa, a definition taken from GLDB.
   
• semantic argument structure, list of arguments assigned by the predicative expression.
   
• selectional restrictions/preferences on arguments, whose values are either semantic types or representations of SemUs. SemUs are chosen whenever the preference is restricted to a unique realisation, e.g. for the verb mjau ’miaow’ the first argument is specified as <katt>.
   
• status of the argument, the arguments can take one of the following values: true, default or shadow. The true value is chosen when the arguments are obligatorily realized; the default value is for semantically optional realisations and the shadow value is for those arguments which are incorporated in the meaning of a lexical item (Pustejovsky 1995).
   
• link to the syntactic unit (Usyn). The Usyns in the Swedish PAROLE lexicon are linked to the SemUs in the Swe-S lexicon, which is effected in a robust information block with a coherent and exhaustive morphological, syntactic and semantic description. The linking of these units is either one-to-one, one-to-many or many-to-one.
   
• link to a corresponding lexeme in GLDB, which not only provides access to all the lexical information encoded in GLDB, but also relates these two resources to each other.
   

2. Semantic encoding

2.0 Links from the Swe-S lexicon to GLDB

In the SIMPLE model, a set of templates provides guidelines concerning sense assignment, as each template is a structured set of semantic information, which highligts relevant aspects of word meaning.

The GLDB's approach to word sense discrimination is a relational one, which means that sense readings are internally structured and that kernel sense and its subsenses are related to each other by a set of lexical transfer rules. The lexical information conveyed by the hierarchical microstructure of a lexicon entry in GLDB is mapped onto the SIMPLE ontology and semantic information pertaining to a semantic type. In consequence, links between these two resources are established and the lexical data can be freely accessed. In the Swe-S lexicon, the links between the two systems are shown by means of the used marking convention for SemUs and glosses. While SemUs are marked consecutively, GLDB's lemma-lexeme-subsense numbers point to the readings in the GLDB and thus provide input to all the syntactic and semantic information linked to a reading there.

2.1 Criteria for syntax-semantic linking

Word sense discrimination has been assumed to be a point of departure for the linking of syntactic and semantic layers, as it is the word meaning that determines its syntactic behavior, argument structure, argument type (true, default or shadow) as well as selectional preferences and semantic roles of its arguments. This information, with the exception of semantic roles that are not specified there, is mapped in Swe-S onto positions at the syntactic layer and consequently it is linked to information on (i) number and optionality of positions in subcategorization pattern, (ii) syntactic functions and their morphosyntactic and lexical realisations, (iii) semantic preferences on arguments with regard to the feature 'animatness'. As a consequence, recurrent mapping patterns can be distinguished.

Let's turn now to the SGML objects. The linking between syntactic positions and semantic arguments is specified in the Correspondence object. Non-predicative nouns are linked by relating to one or more Semus to which a syntactic unit corresponds. Three types of links have been used in Swe-S:

• one-to-one   (e.g symboliserar D02P12, krokus DN0, medicinsk DNA)
• one-to-many (e.g.trea DN0)
• many-to-one (e.g.förklarar D01P12S, D01P12DO)

krokus   'crocus'

The one-to-many relation is illustrated below for the non-predicative polysemous noun trea 'three, third'. The syntactic unit DN0 is linked to six Semus, assigned to the following semantic types: 1-Human, 2-Building, 3-Vehicle, 4-Sign, 5-Unit_of_measurment and 6-Number.

trea   noun meaning 'three, third'

For verbs and deverbal nouns taking arguments, the correspondence is established between particular positions in the subcategorization frame and their arguments. Thus the two arguments a0 and a1 of the verb brevväxla 'correspond (with sb)', are linked to the subject and prepositional object positions respectively in the subcategorization frame for the syntactic unit D01P11POMED. The same verb is also linked to the syntactic unit D01PL, with one position occupied by a plural subject. In this case one can talk about reduced correspondence. These two show semantic affinity, and therefore are encoded with the same Semu number, except that the latter is indexed by 'a' (alternation).

brevväxla    'correspond (with sb)'

In the Swe-S all the verbs, deverbal and relational nouns have their positions linked to the arguments. In most cases these correspondences are isomorphic. Non-isomorphic cases occur for example when a shadow argument is part of a predicative representation, e.g. with weather verbs, or even in other constructions where the number of syntactic positions does not match the number of semantic arguments. As en example of the latter, one can mention a syntactic unit, DACVASUPR, which is a predicative construction with an adjective as head. The lack of correspondence follows from the fact that we have postulated an argument on the semantic layer in order to encode selectional preferences imposed by an adjectival head on the subject. These are needed in order to motivate the semantic type of an adjective in predicative constructions.

2.2 Criteria for assigning domain features

The hierachically structured domain list, elaborated by ERLI/LexiQuest, has been a point of departure for assigning domain information to the readings. This repository of specific domain values has been extended with the value General, to cover general, underspecified readings. While encoding domains, the attention has been paid to the following:

• distinction between general vocabulary and terms; the latter have been assigned a domain feature;
• domain's role for disambiguating readings; if the domain information was the only distinctive factor, it was always encoded;
• multiple domain assignment.

2.3 Criteria for assigning Semantic Class and Template Type

Template Type assignments were based on semantic specifications included in templates. The content of these specifications was compared to the information provided by glosses in GLDB. To make the assignment as exact as possible we have used both the core and recommended ontological types.

Since word senses differ in their lexical complexity and transparency, the assignment task varied accordingly. Concrete nouns could be assigned to the template types without problems in most cases, abstract nouns were more problematic. Verbs with transparent meaning were easier to assign than verbs with underspecified, vague meanings. Thus, the more abstract and vague the word senses were, the less obvious their assignment was. The same can be said about the semantic class assignments.

In the ERLI/LexiQuest classification there are 95 semantic classes to classsify the readings of nouns and verbs. Since this classification differs from the one postulated by SIMPLE ontology as to the number and distribution of categories, the mapping between categories is either one-to-one, one-to-many or many-to-one. For instance, the following semantic classes: Instrument, Apparatus, Measuring_instument or Musical_instrument map onto the template Instrument. We have taken advantage of such differences in order to nuance sense assignments.

As far as the task of assigning template types to adjectives is concerned, we have used all the 14 options provided by the ontology. It might be of interest to note that meaning components explicated in this ontology have had a double function, as they have served both as meaning components and as labels for semantic classes for adjective types. The list of semantic classes for adjectives has been extended with class categories taken from the set of semantic classes for nouns and verbs, e.g Attribute, Emotion, whenever these were semantically motivated.

2.3.1 Language specific typing

The set of 139 ontological types postulated in the guidelines has proved to be relevant for Swedish, as illustrated below (R stands for recommended ontological types):

General ontology for nouns and verbs

General ontology for adjectives

2.4 Classes derived from encoding

The semantic information encoded in the Swe-S lexicon can be automaticaly sorted, clustered and accessed in a number of different ways. We consider the following types of semantic classes to be of interesse for NLP and NLU applications, not to mention their relevance for checking the intern consistency in the SIMPLE lexicons:

• ontology or/and domain classes
• polysemous class
• semantic frame classes
• hyponym - hyperonym classes

The classes of semantic data derived from sorting according to the semantic type, domain or/and semantic class build semantic sublexicons can be easily enriched with other semantic, syntactic or morphological data available in the PAROLE -SIMPLE lexicon and GLDB. Examples of tripartite semantic classes are given below, with information on template type, domain and semantic type as input:

Such tripartite classes support semantic annotating, information extraction, machine translation and natural language understanding in an effective way.

Polysemy

The access to data on types of polysemous classes being instantiated in a language has shown to be indispensable for consistent encodning av meanings in the Swe-S lexicon. Below we have listed some frequent types of regular polysemous classes instantiated in the Swe-S lexicon. It should be observed that polysemy relations are not explicitly coded in the Swe-S lexicon, as they to a large extent can be automatically extracted from the GLDB. In the list below, the first two columns display pairs of semantic types which stand in a polysemous relation, the third one provides examples. The list is alphabetically sorted.

There is no doubt that reaching a consensus on how to deal with polysemy relations is of great relevance for: (i) checking consistency in lexicons, (ii) determining criteria for sense discrimination, (iii) structuring lexicon entries with regard to the distinction between core senses and its subsenses, (iv) text disambiguating, (v) multilingual linking of lexicons.

Semantic frame classes

A subset of words in the Swe-S lexicon carries information on their predicative representation and selectional restrictions on arguments. These representations create semantic frames whose selectional restrictions on arguments are specified by means of reference to semantic types. This information seems to be especially supportive for extending the coverage of the lexicon entries, as the new entries can by default inherit information on the semantic type assignment. To make the assignment as exact as possible, the information on semantic representation should be integrated with information on syntax. For example the verbs of type Natural transition show preference for a frame with a Living entity in the subject position, the same can be said about the Cooperative Speech act verbs. Thus their meanings determine semantic types of the arguments and this information can be reused for applications concerning annotating, disambiguating, lexical acquisition, NLU and automatic extending the coverage of the Swe-S lexicon.

Hyponym - hyperonym classes

In the SIMPLE lexicon model the hyponym-hyperonym relation is part of the qualia encoding, namely Formal. In the Swe-S lexicon this information has not been encoded, because it is already available for all the noun and verb entries in GLDB, through the application called SEMNET, which parses definitions and finds genus proximum. Access to this knowledge is indispensable for a number of NLP and NLU tasks.

2.5 Representation of predicative information

Having already hinted at the relevance of predicative representation for different application tasks, we give below an example showing how the selectional restrictions on verb arguments are encoded for the verb stimmar 1. be noisy, 2 shoal. Selectional restrictions on the arguments in a predicative representation in Swe-S are encoded either by pointing to one or more template type, as is the case in the examples below, or to a particular SemU. Whenever a predicative expression implies a specific referent, its reference is encoded by means of a SemU (i.e. bark, quack).

3 Usability of the semantic data encoded in Swe-S

The issue of usability of the semantic data encoded in Swe-S has been discussed in an article that will be published in the proceedings from LREC 2000 conference. The article is reprinted in Appendix A.

4. Statistics

In the following, we show the distribution of nouns, verbs and adjectives with respect to template types. We list only those template types that have more than 10 words encoded per template.

5. Examples

A sample lexicon that covers 100 fully described words is enclosed with the documentation. The morphological and syntactic information encoded in the Swedish PAROLE lexicon is enriched with the semantic information encoded in the Swe-S lexicon. Three entries from that sample are reprinted in the Appenix B: blomma 'flower', kastanj 'chestnut' and liten 'little'

Bibliography

Calzolari, N. 1999. SIMPLE: Harmonised Semantic Lexicons for the European Languages. In Proceedings of the SIGLEX-99 Workshop: "Standardizing Lexical Resources", Maryland, USA

Gellerstam, M., 1988, Verb Syntax in a Dictionary for Second-Language Learning. In Computer-Aided Lexicology, Almquist & Wiksell; Stockholm.

Guimier, E., Ogonowski, A., 1998. Report on the Syntacic layer, ftp://parole:yaooakgn@www.erli.fr

Holmes, Ph., Hinchliffe, I., 1994. Swedish, London: Routledge Grammmar.

Järborg, J., 1989. Betydelseanalys och betydelsebeskrivning i Lexikalisk databas. Göteborg: Inst. för språkvetenskaplig databehandling

Järborg, J., 1990. Användning av SynTag. Göteborg: Inst. för språkvetenskaplig databehandling.

Nationalencykopedins ordbok, 1995-96. Utarbetad vid Språkdata, Göteborgs Universitet. Höganes: Bra böckers.

Svenska ord – med uttal och förklaringar. 1992. Stockholm:Nordstedts.

Lenci, A., F. Busa, N. Ruimy, E. Gola, M. Monachini 1999, N. Calzolari, A. Zampolli, 1999. Linguistic Specifications, D2.1.

Malmgren, S. (1988). ’On Regular Polysemy in Swedish’, in: Studies in Computer-Aided Lexicography, Almquist & Wiksell, Stockholm.

Pedersen, B.S. and Keson, B. (1999). SIMPLE - Semantic Information for Multifunctional Plurilingual Lexica: Some Examples of Danish Concrete Nouns. In Proceedings of the SIGLEX-99 Workshop: "Standardizing Lexical Resources", Maryland, USA

Pustejovsky, J., 1995. The Generative Lexicon. Cambridge: MA. The MIT Press.

Toporowska Gronostaj, M. 1996. Integrerad valensbeskrivning. Mot ett formaliserat verbvalenslexikon. Göteborgs universitet: Inst. för svenska språket.

Appendix A

Annotating, Disambiguating & Automatically Extending

the Coverage of the Swedish SIMPLE Lexicon

Kokkinakis D., Toporowska Gronostaj M., Warmenius K.

Språkdata, Göteborg University
Box 200, SE-405 30,
Sweden
{svedk, svemt, svekws}@svenska.gu.se

Abstract

During recent years the development of high-quality lexical resources for real-world Natural Language Processing (NLP) applications has gained a lot of attention by many research groups around the world, and the European Union, through the promotion of the language engineering projects dealing directly or indirectly with this topic. In this paper, we focus on ways to extend and enrich such a resource, namely the Swedish version of the SIMPLE lexicon in an automatic manner. The SIMPLE project (Semantic Information for Multifunctional Plurilingual Lexica) aims at developing wide-coverage semantic lexicons for 12 European languages, though on a rather small scale for practical NLP, namely less than 10,000 entries. Consequently, our intention is to explore and exploit various (inexpensive) methods to progressively enrich the resources and, subsequently, to annotate texts with the semantic information encoded within the framework of SIMPLE, and enhanced with the semantic data from the Gothenburg Lexical DataBase (GLDB) and from large corpora.

1. Introduction

During recent years there has been an increased interest to acquire, on a large-scale, high-quality semantic lexicons, McKeown & Hatzivassiloglou (1993), Dorr & Jones (1996), Hearst & Schütze (1996), Takunaga et al. (1997), Roventini et al. (1998), Viegas et al. (1998). The methodology behind these approaches is usually corpus-driven. It is based on the (re-)use of machine readable resources of various types, and the application of cost effective ways to eliminate the acquistion bottleneck, such as derivational morphology, customization of off-the-shelf resources and statistical techniques. The approach adopted here for the extension task is in line with the methodologies mentioned.

In this paper, we focus on ways to extend and enrich, as far as possible, automatically the coverage of the Swedish semantic lexicon by taking into consideration compounding, a distinctive feature of the Swedish language, and semantic similarity in noun phrases of enumerative type. With the support of semantic data from the Swedish SIMPLE lexicon (Semantic Information for Multifunctional Plurilingual Lexica, LE4-8346), Gothenburg Lexical DataBase (GLDB) and large corpora both raw and exposed to shallow parsing, we enhance the incorporation of new semantic entries into the SIMPLE lexicon. We expect to be able to extend the 6,000 entries in the Swedish SIMPLE lexicon to over 120,000 entries. Our assumption is based on the results obtained from the tests carried out so far on input data of 1,000 entries, which became 25,000 (22,000 through compounding and 3,000 through semantic similarity).

Furthermore, we semantically annotate texts with all the available material, and we apply Machine Learning techniques for the disambiguation of ambiguous readings. The annotation task provides an excellent opportunity to evaluate the usability of the semantic information encoded in SIMPLE.

This paper is organized as follows: first we give a brief presentation of the SIMPLE project and particularly of the Swedish lexicon; then we present how compounding and semantic similarity in enumerative phrases (under certain conditions) can contribute to the augmenting and enrichment of the lexicon, when subjected to compound segmentation and shallow parsing; we continue by describing a practical application of the semantic lexicon, namely semantic annotation and disambiguation; we then give some general remarks on the usability of the SIMPLE model, while conclusions end the presentation.

2. The SIMPLE Project

The EU-financed SIMPLE project aims at developing wide-coverage semantic lexicons for 12 European languages. The Swedish SIMPLE lexicon (hereafter Swe-S) is one of these. All lexicons share a common semantic model and a common encoding formalism in SGML. The semantic data in the SIMPLE lexicons is being linked to the morphological and syntactic data in their respective PAROLE lexicons, developed within the EU project PAROLE, (Preparatory Action for Linguistic Resources Organisation for Language Engineering). Out of the 20,000 words in the PAROLE lexicons, a subset of about 6,000 words, or approximately 10,000 senses, has been enriched with semantic descriptions in the SIMPLE counterpart. The content and the design of the SIMPLE model are documented in Lenci et al. (1998).

The notion of semantic type is central for the SIMPLE model and its ontology. It corresponds to a word sense assigned to a lexical item. There are 139 semantic types distinguished in the SIMPLE ontology. Each semantic type is defined as a cluster of structured semantic information significant for a given word sense. Information on semantic class, domain, argument structure of predicative expressions and selectional restrictions on arguments as well as qualia roles constitute a relevant part of the semantic type specification; (Calzolari (1999), Pedersen & Keson (1999)). The SIMPLE ontology is multidimensional as it is based on the principle of orthogonal inheritance (Pustejovsky 1995), and in this respect, it contrasts with the LexiQuest’s semantic class ontology which is based on a standard, monodimensional approach. The latter ontology includes 95 semantic classes. Both ontologies are hierarchically structured.

3. The Swedish SIMPLE Lexicon

The theoretical and formal design of the Swe-S lexicon is conformant to the SIMPLE's linguistic guidelines presented by the specification group, Lenci et al. (1999). In the Swe-S lexicon, there are about 10,000 semantic units (hereafter Usems) encoded, comprising 7,000 noun, 2,000 verb and 1,000 adjective Usems. These 10,000 units are mapped onto 6,000 entries. Usems are described with respect to the following information:

• semantic type, whose value is an element in the SIMPLE ontology list
  (e.g. Usem <katt> ’cat’: Earth_Animal).
   
• domain, whose value is an element in the LexiQuest's domain list
  (Usem <katt>:Zoology).
   
• semantic class, whose value is an element in the LexiQuest's semantic class list.
  (Usem <katt>: Mammal).
   
• glossa, a definition taken from GLDB.
   
• semantic argument structure, list of arguments assigned by the predicative expression.
   
• selectional restrictions/preferences on arguments, whose values are either semantic types or representations of Usems. Usems are chosen whenever the preference is restricted to a unique realisation, e.g. for the verb mjau ’miaow’ the first argument is specified as <katt>.
   
• status of the argument, the arguments can take one of the following values: true, default or shadow. The true value is chosen when the arguments are obligatorily realized; the default value is for semantically optional realisations and the shadow value is for those arguments which are incorporated in the meaning of a lexical item (Pustejovsky 1995).
   
• link to the syntactic unit (Usyn). The Usyns in the Swedish PAROLE lexicon are linked to the Usems in the Swe-S lexicon, which is effected in a robust information block with a coherent and exhaustive morphological, syntactic and semantic description. The linking of these units is either one-to-one, one-to-many or many-to-one.
   
• link to a corresponding lexeme in GLDB, which not only provides access to all the lexical information encoded in GLDB, but also relates these two resources to each other.
   

In the course of building the Swedish SIMPLE (and PAROLE) lexicons we have, to a large extent, reused lexical data from GLDB which is the most comprehensive source of lexical information on contemporary Swedish, and information from the SO (1992) and NEO (1996).

4. Extending the Coverage of the Swe-S

The Swe-S resources are not quantitatively sufficient for realistic, large-scale Natural Language Processing (NLP) tasks, such as semantic annotation, and need to be extended. For this particular task, we take advantage of the productive compounding characteristic of Swedish and the use of raw and partially parsed corpora.

We assume that a considerable number of casual, or on the fly created compounds in Swedish can inherit relevant parts of semantic information provided on their heads by the Swe-S lexicon and thus, can be incorporated into the lexicon. By relevant parts, we mean in the first place the information concerning semantic type, domain and semantic class. To avoid errors, we exclude the information on argument structure from the inheritance, as the argument structure can undergo alternations in the process of compounding. This is the case when verbs and verbal nouns build compounds with either an obligatory or optional argument in the non-head position. The occurrence of an adjunct in the non-head position does not usually alter the predicative structure.

4.1 Compounding

The fact that over 70%, or approximately 80,000, of all the entries in the SAOL (1998) are compound forms casts light not only onto an immense lexical repository, which is available for this particular extension task, but also on the need to design effective tools and routines for compound segmentation, as new, casual compounds are created constantly in Swedish. Most of these casual compounds are relatively transparent, which implies that their meaning is a function of the meaning of its components being related to each other by an implied predicative functor. For instance, brödkniv brödXknivY ’bread knife’ implies ’Y for (cutting) X’ and bärsaft bärXsaftY ’juice from berries’ implies Y which contains X. In Swedish, compounds are written as single orthographic units and nouns are the most frequent modifiers occurring in non-head positions.

A combination of various heuristic methods is used for the extension. Compound segmentation is applied to compound noun tokens on large corpora and lists of new nouns are produced. To maintain quality assurance and compatibility with the rest of the data in the lexicon, new heuristics are applied to the content of the noun lists produced. To avoid generation of incorrect data, these heuristics inspect the modifying component of a compound in order to distinguish its characteristics, such as its part-of-speech and semantic category (if any). These characteristics of the modifier, when enriched with the corresponding characteristics of the compound head, provide data for a preliminary estimation of the correctness of the heuristics. Few examples will illustrate this point.

If the part-of-speech of the compound modifier is an adjective, a new Swe-S entry, which will not cause semantic anomalies in the derived lexical set, can be created with great confidence. The inheritance criterion applies here and the compounds are hyponyms to the head. For example, the lemma klocka ’bell/watch’ can be extended with compounds of type [ADJ-MODIFIER]+HEAD: [digital]klocka, [guld]klocka, [lill]klocka, [silver]klocka, [stor]klocka, where the adjectival modifying part in these examples are ’digital, gold, little, silver’ and ’big’. Similar results are obtained if the modifying part is a proper noun. For instance, anhängare ’supporter’ with modifiers such as: Berisha, Hammarby, Hitler, Likud and Mobutu, signal unambiguous compounds.

It is well known that the heuristics have a variable degree of performance on different types of compounds, and that some simple constraints are needed to exclude segmentation and interpretation errors. Particularly in the case where the part-of-speech of the modifying part of a compound is a noun (e.g. NOUN-MODIFIER[kultur]fråga ’cultural question’) or verb (e.g. VERB-MODIFIER[betal]teve ’pay-TV’). These constraints are formed by means of subroutines which impose checking of derived compounds against different lists to eliminate incorrect data. The lists with bound morphemes or lexicalized compounds, extracted from the GLDB allow exclusion of such compounds from the derived sets. Such constraints have proven to be a cheap way to automatically constrain the overgeneration of new entries in the lexicons.

For instance, when using large corpora, over 40 compounds with feber ’fever’, as head, could be extracted. However, it became evident that not all of them belong to the semantic class of Illness, e.g. resfeber ’excitement before a journey’. Thus, in some cases, additional inspection seems unavoidable, if we want to restrain automatic incorporation of lexicalised compounds with idiomatic, metaphoric or metonymic meanings. This inspection can be performed automatically by simply checking whether a given compound is included as a separate entry in GLDB. If this is the case, it means that the compound is lexicalised and should not be subjected to automatic inheritance. The manual inspection is needed, only if the derived compound shows diverging semantic and/or morphological patterns and the word is neither in a bound morpheme list, nor in the lexicalised compound list.

Moreover, the content of the Swe-S has been used as a means of bootstrapping the process. For instance, glas ’glass’, can be extended with compounds having Substance as a modifier in the compound form. Consequently the [NOUN-MODIFIER{Substance }]+HEAD compounds [vatten]glas, [vin]glas, [öl]glas, [likör]glas all have Substance as the modifier part, namely ’water, wine, beer’ and ’liqueur’.

A large number of already disambiguated compounds has been also extracted from GLDB, since the Swe-S entries are linked to the various senses and sub-senses in GLDB, and subsequently to the morphological examples of every entry (alias compounds). For instance, Swe-S encodes the non-compound lemma ämne (as having four senses, marked with 1/1-1/4), which are disambiguated here by means of their assignment to the following semantic types and semantic classes:

Each of these senses is exemplified in GLDB with a number of compounds, comprising totally 26 compounds with ämne as the head. Some of these are listed in the right column of table (1). Since there is only one compound with that head in the Swe-S lexicon (grundämne ’element’), incorporating new, disambiguated compounds was straightforward.

Table 1: ämne in Swe-S, and GLDB compounds with ämne as head.

4.2 Heuristic Incorporation of New Entries through Shallow Parsing

So far we have addressed the problem of the acquisition of compound nouns based on the content of the Swe-S lexicon, by applying heuristics, filters, and manual inspection, in some cases, in order to guarantee consistency. But how can we cope with the rest of the vocabulary?

Wilson and Thomas (1997:55-57) argue that one of the conditions that a semantic system should satisfy is that is should be able to account exhaustively for the whole vocabulary in the corpus, not just for a part of it. We have experimented with a corpus-based approach, using a cascaded finite-state syntactic parser (CASS-SWE), based on work done by Kokkinakis & Johansson Kokkinakis (1999), which seems a plausible way of progressively enriching the Swedish semantic resources.

An advantage of CASS-SWE is its ability to identify with high accuracy noun phrases, a property that we consider here as crucial for aiding the "discovery" of new semantic entries. Essentially the approach, which has similarities to naive clustering, is as follows. Gather large corpora (here 13 million tokens), part-of-speech tag, and then parse with CASS-SWE (the parser uses part-of-speech annotated input); from the resulted analyzed forest of chunks we filter out long noun phrases, namely those containing three or more common nouns. Finally, the overlap between the nouns in the NPs produced and the entries in Swe-S is measured. If at least two of the nouns (a figure arbitrarily taken) are also entries in the Swe-S, with the same semantic class, then there is a strong indication that the rest of the nouns are co-hyponyms, and thus semantically similar with the two already encoded in Swe-S. Accordingly, we take advantage of the transitivity aspect of hyponymy, and of the fact that two lexical items X and Y are co-hyponyms if: (i) they are disjuncts and therefore complementary; and (ii) have a common superordinate, e.g. animal is superordinate of cat, dog, horse and camel, cf. Sanfilippo et al. (1999).

Similarity plays an important role in word acquistion, and preliminary results have shown that the simple overlap works fairly well for the majority of the cases examined. However, the noise which is produced can be eliminated, if the semantic tags of all the words in a phrase are compared. Caution should be taken for cases where different semantic classes are involved in an enumerative NP, e.g.:

The unclassified husdjur in the first example, should not be assigned to a class Bio since there is another class involved in the same NP, namely Furniture. Similarly, no action should be taken in the second example, since two semantically ambiguous words with distinct classes are involved.

During classification an unseen example X, a test instance, is presented to the system and a distance metric D between the instances in the memory Y and X is calculated, D(X,Y). Various implemented algorithms (variants of the k-nearest neighbour algorithm) try to find the nearest training instance for X and create a class as prediction for the class of the test instance.