Kadri Vider, Heili Orav
Estonian wordnet and Lexicography
1. Introduction
There has been the need for computer thesauri in Estonian lexicography for a long time. For
1997 it was clear, that besides morphological and syntactical analysis a lexical database
based on word semantics was needed. There was the WordNet (WN) created by G.A. Miller and others at Princeton University in the 1980-s, and we followed suit. Compilation of
the Estonian Wordnet started in 1997 and the work is still in progress. The work is funded
partly by Estonian Science Foundation and partly in the frames of Estonian National Programme of Language Technology. The existing Estonian wordnet contains nouns, verbs
and some adjectives and proper nouns.
2. WordNet
The design of a semantic database is faced with two fundamental questions:
(a) what kind of semantic information is stored: linguistic knowledge or world-knowledge
(b) how is this information stored: as a language internal system or using some meta-language
Many semantic databases are not explicit with respect to these distinctions and it is often
very difficult to keep them apart. In the case of traditional dictionaries you might expect to
find ‘linguistic knowledge’ telling you what semantic properties determine the exact usage
of a word, but dictionary definition very often also contain fragments of world-knowledge.
In the case of conceptual networks that try to capture our common-sense knowledge and
reasoning you might expect to only find ‘world knowledge’ but what is defined are often
the same words (Bloksma et al. 1996).
By focusing on historical evidence, the standard dictionaries neglected questions concerning the synchronic organization of lexical knowledge. Beginning with word association
studies at the turn of the 20th century and continuing down to the sophisticated experimental
tasks of the past twenty years, psycholinguists have discovered many synchronic properties
of the mental lexicon that can be exploited in lexicography (Miller et al, 1990).
In 1985 a group of psychologists and linguists at Princeton University undertook to develop
a lexical database along lines suggested by these investigations. G.A. Miller and others
create WordNet - a semantic database of English. Inasmuch as it instants hypotheses based
on results of psycholinguistic research, WordNet can be said to be a dictionary based on
psycholinguistic principles (Fellbaum 1998).
126
Kadri Vider, Heili Orav
In WordNet fundamental differences in the semantic organization of syntactic categories
can be clearly seen and systematically exploited. Nouns are organized in lexical memory as
topical hierarchies, adjectives are organized as N-dimensional hyperspaces, and verbs are
organized by a variety of entailment relations. The most ambitious feature of WordNet is its
attempt to organize lexical information in terms of word meanings, rather than word
forms. (Miller et al. 1993)
3. Terms
A word can have several meanings and mostly it has. As WordNet is based on word meaning, all of the words that can express a given sense are grouped together in a SYNONYM
SET or shortly called SYNSET.
For example if you are searching Estonian word ‘viis’ you can find at least four different
meanings, i.e. four synsets:
Viis_1, 5_1: number 5 (meaning is number five)
Viis_2, väga_hea_1: kõrgeim hinne eesti koolisüsteemis (meaning is five – the highest grade in
Estonian school system)
Viis_3, meloodia_1: helide järgnevus muusikas (meaning is melody – consequence of sounds in
music)
Viis_4, mood_1, stiil_1, laad_1: väljakujunenud, iseloomulik või äratuntav tegutsemislaad või
olemisviis (meaning is manner, style, mode – specific, known manner of performance)
WordNet is organised by semantic relations. Since a semantic relation is a relation between
meanings, and since meanings can be represented by synsets, it is natural to think of semantic relations as pointers between synsets.
The database is restricted to the relations suggested by the psycholinguistic data: synonyms,
antonyms, hyponyms, hyperonyms, meronyms, holonyms, causation etc. These and other
similar relations serve to organise the mental lexicon.
Some examples:
Hierarchical: koer (dog) HAS_HYPERONYM loom (animal)
Antonymy: halb (bad) ANTONYM hea (good)
Part/hole: koer (dog) HAS_MERONYM saba (tail)
Role/involved: õpetaja (teacher) ROLE õpetama (to teach)
Causation: tapma (to kill) CAUSE surema (to die)
As in EuroWordNet project is in Estonian Wordnet also more than 50 different semantic
relations, most of them are reciprocated:
If A HAS_HYPERONYM B, then B HAS_HYPONYM A.
4. EuroWordNet – a multilingual semantic database
Estonian Wordnet and Lexicography
127
In 1996 the European Union supported project EuroWordNet (EWN) started, which aimed
to develop a multilingual database with basic semantic relations between words for several
European languages (Dutch, Spanish, Italian and English; from 1998 also Czech, French,
German and Estonian). The wordnets are stored in a central lexical database system and the
word meanings are linked as in the Princeton WordNet. Each wordnet represents a unique
language-internal system of lexicalisations (Vossen 1998a).
The design of the EuroWordNet database is first of all based on the structure of the Princeton WordNet and specifically version WordNet1.5. The notion of a synset and the main
semantic relations has been taken over in EuroWordNet. However, some specific changes
have been made to the design of the database, which are mainly motivated by the following
objectives:
1) to create a multilingual database;
2) to maintain language-specific relations in the wordnets;
3) to achieve maximal compatibility across the different resources;
4) to build the wordnets relatively independently (re)-using existing resources.
To be able to maintain the language-specific structures and to allow for the separate development of independent resources, they make a distinction between the language-specific
modules and a separate language-independent module.
Each language-specific wordnet is structured along the same lines as WordNet synonyms
are grouped in synsets, which in their turn are related by means of basic semantic relations
such as hyponymy (between specific and more general concepts), meronymy relations
(between parts and wholes). By means of these relations all meanings can be interconnected, constituting a huge network or wordnet.
Each language module represents an autonomous and unique language-specific system of
language-internal relations between synsets. Nevertheless, some specific measures have
been taken to enlarge the compatibility of the different language-specific wordnets, as the
definition of a common set of so-called Base Concepts (BC). Base Concepts are meanings
(i.e. synsets) that play a major role in the wordnets (Vossen et al. 1997).
Equivalence relations between the synsets in different languages and WordNet1.5 are made
explicit in the so-called Inter-Lingual-Index (ILI). Each synset in the monolingual wordnets has at least one equivalence relation with a record in this ILI, either directly or indirectly via other related synsets. Language-specific synsets linked to the same ILI-record should
thus be equivalent across the languages. Synsets linked to the same WordNet1.5 synset are
supposed to be equivalent or close in meaning and can then be compared.
5. Estonian Wordnet
Compilation of Estonian wordnet (EstWN) started from 1997 and the work will continue.
Its relevant to mention that other partners in EWN project have created their wordnets automatically, but we have preferred manual work. One reason for this is absence of suitable
electronical material, and on the other hand, it gives more exact results.
128
Kadri Vider, Heili Orav
In addition to the general vocabulary also legal vocabulary as to facilitate the precise translation of legal texts is made. Table 1 gives an overview of characteristic numbers Estonian
wordnet is reached to.
Table 1: Number of synsets per different part-of-speeches in EstWN
Proper Nouns Nouns Verbs Adjectives Total
SYNSETS
446
6734
2757 307
10244
Meanings (variants)
471
11230 5746 518
17965
Meanings per synset
1,06
1,67
2,08
1,75
Different words and compounds 470
9518
3797 419
14204
Meanings per word
1
1,18
1,51
1,26
Semantic relations
474
15024 5596 538
21632
Semantic relations per synset
1,06
2,23
2,11
2,03
1,69
1,24
1,75
The major lexical information, what is so necessary for compiling thesaurus, comes from
monolingual explanatory and/or sense distinguishing dictionaries or synonyms dictionaries.
In Estonian there are not so many of them, more over we don’t have enough machinereadable dictionaries.
The data for the compilation of the Estonian WordNet are got from the following sources:
(a) to get the word meanings, explanations and examples the Estonian Explanatory Dictionary,
which is unfortunately not a completely machine readable dictionary, is used
(b) in case of synonymy and antonym relations the dictionaries of synonyms and antonyms are
used
(c) to link concepts with Inter Lingual Index bilingual dictionaries are used
(d) other tools of Estonian language technology are used. For example word frequency records are
compiled on the basis of the Corpus of Estonian Literary Language (CELL, 1 mln. words) and
the materials of the Corpus are also used to define the different meanings of a word and quotations
from the Corpus are used as examples.
6. Weaving Estonian wordnet
In general, the wordnet was built in two major cycles as indicated in Figure 1. Each cycle
consisted of a building phase and a comparison phase. For creation of database itself we use
the EWN Database Editor Polaris (Louw 1998).
129
Estonian Wordnet and Lexicography
Add new synsets
How existing synsets
cover language usage
Figure 1. Building cycles of EstWN.
Word-frequency
lists
Find all
meanings of
a word
Monolingual
resources
Join synonymous
word-meanings into
synset
Link synsets via
different semantic
relations
Multilingual
resources
Link synsets via ILI
relations
6.1. Words from texts
EstWN is supposed to cover the Estonian base vocabulary in its initial version. Besides
Base Concepts translation from Princeton Wordnet was avoided, main emphasis was given
on frequency lists of word forms. The base vocabulary was determined by statistical analysis of Corpus of Estonian Literary Language.
Lexical entries in EstWN are presented nominal singular form for nouns and supine form
for verbs. In real texts the words are mostly in their full richness of forms. For example,
there are fourteen cases in Estonian, it makes 28 forms in every noun paradigm. On average
45% of the word forms are morphologically ambiguous in Estonian texts and this fact causes irrelevant data in lexicon word-list.
The number of compounds in Estonian is indefinite. It is quite easy for a user to make up
compounds that are not found in any dictionary, but are still understandable by other users.
It is not reasonable to enter all compounds that one can imagine into wordnet. The difference can be made by analysing the components of a compound: if the components are of
low frequency as independent words in corpora, then the compounds they are building up
should be added into lexicon.
Estonian is a flective language with a free word order and that makes complicated to figure
out all phrases from corpus texts. The elements of phrase can be scattered around sentence
in an unpredictable order. Multi-word expressions are included into EstWN if they build up
a conceptual unit and are commonly used as lexical units.
130
Kadri Vider, Heili Orav
6.2. A set of synonymous word meanings
One of the problems concerned dividing meanings of a word into different synsets in
EstWN. Due to lexicographical sources what we use, over-differentiation (word meaning
marked as too specific) and over-generalisation (word meaning marked as too general) may
occur. Traditional dictionary entries are centred to word and all syntactically or pragmatically specific meanings are represented there, but semantically they are not always significant.
Which new words or meanings should be concentrated on to upgrade the EstWN? It is
essential that words actually used in text will be added.
Results of word sense disambiguation (WSD) of corpus texts turned to be a good way to
add missing synsets and senses into our wordnet. There were significant inconsistencies in
opinions of these people, who disambiguated the texts. This shows us the most problematic
entries in EstWN, the need to reconsider the borders of meaning some concepts.
6.3. Language-dependent lexicalisation vs Inter-Lingual Index
As said, wordnets of different languages are linked via Inter Lingual Index in EWN. But
lexicalisation features or borders of one concrete word meaning are different in language by
language in many cases and finding exact equal relation (EQ_SYNONYM) is often impossible. Fortunately, it is possible to use other eq-relations for describing more exactly what is
difference between source language and ILI synsets.
The Estonian language as Finno-Ugric language is different from most Indo-European
languages as in EuroWordNet. Some specific lexicalisation patterns appear, when we tried
to link Estonian synsets to ILI.
In the case of verbs there is a quite regular discrepancy between English and Estonian in the
expression of causativity. In Estonian causative counterparts of noncausative verbs are
regularly lexicalised, also different lexical entries, and the opposition is productive. In addition, there exist cases where in English there are verbs which have only causative or only
noncausative meaning, but in Estonian both meanings are lexicalised and used productively. For instance, the English verb ‘to spend (money)’ is translated as ‘kulutama’, but there
is also noncausative ‘kuluma’ (also about money) for which no exact English lexicalised
counterpart exists (the dictionary translation is ‘be spent’ but the Estonian ‘kuluma’ has no
passive meaning). The same helds e.g. about ‘levitama’ (to disseminate, distribute) and
noncausative ‘levima’. Problem rises from ILI synset, there is no causativity/noncausativity
marked, and so it is impossible to use nor eq_causes non eq_is_caused_by relation for linking source language and ILI synset.
For example ‘to move’ equivalents in Estonian are ‘liikuma’ and ‘liigutama’. The last is
causative. In Estonian are these differences systematically lexicalized variously from other
languages (Vider 2002).
Also onomatopoetic-descriptive words are more detailed in Estonian. One example: Estonian bear ‘mõmiseb’ and lion ‘möirgab’. In English they both ‘roar’.
Estonian Wordnet and Lexicography
131
7. Conclusion
The Estonian wordnet can be used, among others, for monolingual and cross-lingual information retrieval, which is enable to search by meanings and concepts. Also wordnet-type
thesaurus is helpful for translation’s tools or for compiling other, different kind of lexicons.
For Estonian language technology it is very important to have such lexical-semantic resource.
The present paper pointed out some difficulties, which appeared during practical compiling
of Estonian wordnet-type thesaurus.
References
Bloksma, Laura, P. Díez-Orzas, Piek Vossen 1996. User requirements and functional specification of
the EuroWordNet project, EuroWordNet (LE-4003) deliverable D001, University of Amsterdam.
Available at http://www.illc.uva.nl/EuroWordNet/docs.html
Fellbaum, Christiane 1998. Introduction. In: Fellbaum, Christiane (ed.) WordNet: An Electronic
Lexical Database, MIT Press, pp. 1-19.
Louw, M.1998. Polaris User's Guide.The EuroWordNet Database Editor. EuroWordNet (LE-4003),
Deliverable D023D024, Lernout & Hauspie - Antwerp, Belgium. Available at
http://www.illc.uva.nl/EuroWordNet/docs.html
Miller, George A., Richard Beckwith, Christiane Fellbaum, Derek Gross and Katherine J. Miller
1990. „Introduction to WordNet: an on-line lexical database.” In: International Journal of Lexicography 3 (4), 1990, pp. 235 - 244.
Miller, George A., Richard Beckwith, Christiane Fellbaum, Derek Gross and Katherine J. Miller
1993. “Introduction to WordNet: an on-line lexical database.” Available at .
ftp://ftp.cogsci.princeton.edu/pub/wordnet/5papers.ps
Vider, Kadri 2002. Notes about labelling semantic relations in Estonian WordNet In: D. N. Christodoulakis, C. Kunze, L. Lemnitzer (eds) Proceedings of Workshop on Wordnet Structures and
Standardisation, and how these Affect Wordnet Applications and Evaluation; Third International
Conference on Language Resources and Evaluation (LREC 2002). ELRA, Las Palmas de Gran
Canaria, pp. 56-59
Vossen, Piek, Laura Bloksma, Horacio Rodriguez, S. Climent, Nicoletta Calzolari, Adriana Roventini, Francesca Bertagna, Antonia Alonge, Wim Peters. 1997. The EuroWordNet Base Concepts
and Top Ontology. EuroWordNet (LE 4003) Deliverable D017, D034, D036. University of Amsterdam. Available at http://www.illc.uva.nl/EuroWordNet/docs.html
Vossen, Piek 1998. EuroWordNet: Building a Multilingual Database with Wordnets for European
Languages. In: K. Choukri, D. Fry, M. Nilsson (eds), The ELRA Newsletter, Vol3, n1, 1998.
Vossen, Piek 1998a. Introduction to EuroWordNet. Computers and the Humanities, 32 (2-3), pp. 7389