Content uploaded by Julio Gonzalo
Author content
All content in this area was uploaded by Julio Gonzalo
Content may be subject to copyright.
Terminology Retrieval: towards a synergy between
thesaurus and free text searching
Anselmo Peñas, Felisa Verdejo and Julio Gonzalo
Dpto. Lenguajes y Sistemas Informáticos, UNED
{anselmo,felisa,julio}@lsi.uned.es
Multilingual Information Retrieval usually forces a choice between free text
indexing or indexing by means of multilingual thesaurus. However, since they
share the same objectives, synergy between both approaches is possible. This
paper shows a retrieval framework that make use of terminological information
in free-text indexing. The Automatic Terminology Extraction task, which is
used for thesauri construction, shifts to a searching of terminology and becomes
an information retrieval task: Terminology Retrieval. Terminology Retrieval,
then, allows cross-language information retrieval through the browsing of
morpho-syntactic, semantic and translingual variations of the query. Although
terminology retrieval doesn’t make use of them, controlled vocabularies
become an appropriate framework for terminology retrieval evaluation.
Introduction
The organization of information for later retrieval is a fundamental area of research
in Library/Information Sciences. It concerns to understand the nature of information,
how humans process it and how best to organize it to facilitate use. A number of tools
to organize information have been developed, one of them is the information retrieval
thesaurus. A thesaurus is a tool for vocabulary control. Usually it is designed for
indexing and searching in a specific subject area. By guiding indexers and searchers
about which terms to use, it can help to improve the quality of retrieval. Thus, the
primary purposes of a thesaurus are identified as promotion of consistency in the
indexing of documents and facilitating searching. Most thesauri have been designed
to facilitate access to the information contained within one database or group of
specific databases. An example is the ERIC1 thesaurus, a gateway to the ERIC
documents database containing more than 1.000.000 abstracts of documents and
journal articles on education research and practice. Via the ERIC interface, one can
navigate the thesaurus and use the controlled vocabulary for more accurate and
fruitful search results. Thesaurus were a resource used primarily by trained librarians
obtaining good performance. However nowadays on-line database searching is carried
out by a wider and less specialized audience of Internet users and recent studies
(Hertzberg 1999) claim that most end-users obtained poor results, missing highly
relevant documents. Nevertheless there is a strong feeling in the documentalist field
1 http://www.ericfacility.net/extra/pub/thessearch.cfm
that the use of a thesaurus is a central issue for raising the quality of end-users results,
(Trigari 2001) specially in a multilingual context where natural language ambiguity
increases, producing additional problems for translingual retrieval. A multilingual
thesaurus guarantees the control of the indexing vocabulary, covering each selected
concept with a preferred term, a descriptor, in each language, and ensuring a very
high degree of equivalence among those terms in different languages. However,
multilingual thesauri construction and maintenance is a task with a very high cost,
which motivates the exploration of alternative approaches based on free text indexing
and retrieval.
In this paper, on one hand, we show how NLP techniques have a part to play both
in thesaurus construction and in free text searching in specialized collections; on the
other hand, we describe an evaluation framework for an NLP-based full-text
multilingual search system where a thesaurus resource is used as a baseline. The
structure of the paper is the following: Section 2 reports the developed methodology
implying NLP techniques to support the construction of the European Schools
Treasury Browser (ETB2) multilingual thesaurus in the field of education. This
methodology easily shifts to a new strategy with IR shared objectives: terminology
retrieval. Section 3 introduces Website Term Browser (WTB) (Peñas 2001a), a
browsing system that implements this strategy for searching information in a
multilingual collection of documents. The system helps the user to cross language
barriers including terminology variations in different languages. In order to assess the
performance of WTB we have designed an evaluation framework for the terminology
retrieval task, that takes profit of the ETB multilingual thesaurus. Section 4 presents
this evaluation framework and shows the results obtained. The conclusion points out a
direction in which NLP techniques can be a complement or an alternative to thesaurus
based retrieval.
From Terminology Extraction to Terminology Retrieval
Thesaurus construction requires collecting a set of salient terms. For this purpose,
relevant sources including texts or existing term lists have to be identified or
extracted. This is a task combining deductive and inductive approaches. Deductive
procedures are those analysing already existing vocabularies, thesauri and indexes in
order to design the new thesaurus according to the desired scope, structure and level
of specificity; inductive approaches analyse the real-world vocabularies in the
document repositories in order to identify terms and update the terminologies. Both
approaches can be supported by automatic linguistic techniques. Our work followed
the inductive approach to provide new terminology for the ETB thesaurus, starting
with an automatic Terminology Extraction (TE) procedure (Peñas 2001b). Typically,
TE (or ATR, Automatic Terminology Recognition) is divided in three steps
(Bourigault 1992), (Frantzi 1999):
1. Term extraction via morphological analysis, part of speech tagging and
shallow parsing. We distinguish between one word terms (mono-lexical
2 http://www.eun.org/etb
terms) and multi-word terms (poly-lexical terms), extracted with different
techniques detailed in (Peñas 2001b)
2. Term weighting with statistical information. The weight is a measure of the
term relevance in the domain.
3. Term selection, ranking and truncation of terminological lists by thresholds of
weight.
These steps require a previous one in which relevant corpora is identified,
automatically collected and prepared for the TE task.
After collecting terms, documentalists need to decide which ones are equivalent,
which are finally selected and which other terms should be introduced to represent
broad concepts or to clarify the structure of semantic relations between terms. The
main semantic relations are hierarchical (represented as BT and NT) and RT to
express an associative relationship. To support documentalists decisions, a web-based
interface making use of hyperlinks was provided. Through this interface, access to
candidate terms contexts as well as their frequency statistics were provided.
This was the methodology employed to term extraction task and thesaurus
construction. However, while the goal in the Terminology Extraction is to decide
which terms are relevant in a particular domain, in a full text search users decide
which are the relevant terms according to their information needs, i.e. the user query
gives the relevant terms. In this case, the automatic terminology extraction task
oriented to texts indexing should favour recall rather than precision of the extracted
terms. This implies:
1. Terminology list truncation is not convenient.
2. Relaxing of poly-lexical patterns is possible.
And also suggests a change of strategy. From a thesaurus construction point of
view, TE procedure shifts to term searching becoming a new task: terminology
retrieval.
From a text retrieval perspective, retrieved terminology becomes an intermediate
information level which provides document access bridging the gap between query
and collection vocabularies even in different languages.
This framework, shared for both tasks, needs:
1. A previous indexing to permit phrase retrieval from query words.
2. Expansion and translation of query words in order to retrieve terminology
variations (morpho-syntactic, semantic and translingual).
This strategy has been implemented in the Website Term Browser described in the
next section.
Website Term Browser
The system, Website Term Browser (WTB) (Peñas 2001a), applies NLP techniques
to perform automatically the following tasks:
1. Terminology Extraction and indexing of a multilingual text collection.
2. Interactive NL-query processing and retrieval.
3. Browsing by phrases considering morpho-syntactic, semantic and translingual
variations of the query.
Terminology Extraction and Indexing
The collection of documents is automatically processed to obtain a large list of
terminological phrases. Detection of phrases in the collection is based on syntactic
patterns. Selection of phrases is based on document frequency and phrase
subsumption. Such processing is performed separately for each language (currently
Spanish, English, French, Italian and Catalan). We reused, in a relaxed way, the
terminology extraction procedure originally meant to produce a terminological list to
be used by documentalists in a thesaurus construction process (Peñas 2001b).
Query processing and retrieval
Cross-language retrieval is performed by translating the query to the other
languages in the collection. Word translation ambiguity can be drastically mitigated
by restricting the translation of the components of a phrase into words that are highly
associated as phrases in the target language (Ballesteros 1998). This process is
generalized in the Website Term Browser as follows:
1. Lemmatized query words are expanded with semantically related words in the
query language and all target languages using the EuroWordNet lexical database
(Vossen 1998) and some bilingual dictionaries.
2. Phrases containing some of the expanded words are retrieved. The number of
expansion words is usually high, and the use of semantically related words (such
as synonyms or meronyms) produce a lot of noise. However, the retrieval and
ranking of terms via phrasal information discards most inappropriate word
combinations, both in the source and in the target languages.
3. Unlike batch cross-language retrieval, where phrasal information is used only to
select the best translation for words according to their context, in this process all
salient phrases are retained for the interactive selection process.
4. Documents are also ranked according to the frequency and salience of the
relevant phrases they contain.
Browsing by phrases
Figure 1 shows the WTB interface. Results of the querying and retrieval process
are shown in two separate areas: a ranking of phrasal expressions that are salient in
the collection and relevant to the user's query (on the left part) and a ranking of
documents (on the right part). Both kinds of information are presented to the user,
who may browse the ranking of phrases or directly click on a document.
Phrases in different languages are shown to users ranked and organized in a
hierarchy according to:
1. Number of expanded terms contained in the phrase. The higher the number of
terms within the phrase, the higher the ranking. In the monolingual case, original
query terms are ranked higher than expanded terms.
2. Salience of the phrase according to their weight as terminological expressions.
This weight is reduced to within-collection document frequency if there is no
cross-domain corpus to compare with.
3. Subsumption of phrases. For presentation purposes, a group of phrases containing
a sub-phrase are presented as subsumed by the most frequent sub-phrase in the
collection. That helps browsing the space of phrases similarly to a topic
hierarchy.
Fig. 1. Website Term Browser interface.
Figure 1 shows an example of searching. The user has written the English query
words ”adult education” in the text box. Then, the system has retrieved and ranked
related terminology in several languages (Spanish, English, French, Italian and
Catalan). This terminology was extracted automatically during indexing, and now has
been retrieved from the query words and their translations. In the example, the user
has selected the Spanish tab as target language where there are three different top
terms (folders): ”formación de adultos”, ”adultos implicados en el proceso de
enseñanza” and ”educación de adultos”. The second one (”adultos implicados en el
proceso de enseñanza”) is not related to the concept in the query, but the term
browsing facility permits to discard it without effort. Top term folders contain
morpho-syntactic and semantic variations of terms. For example, the preferred
Spanish term in the ETB thesaurus is “educación de adultos”. However, in this case,
besides the preferred term, WTB has been able to offer some variations:
• Morpho-syntactic variations: ”educación permanente de adultos”,
“educación de personas adultas”.
• Semantic variations: ”formación de adultos” or ”formación de personas
adultas”.
In the example, the user has expanded the folder “educación de adultos” and has
selected the term ”educación de las personas adultas”, obtaining (on the right
handside) the list of documents containing that term.
Evaluation
The usefulness of term browsing versus document ranking was already evaluated
in (Peñas 2001a). Now the evaluation is aimed to establish the system coverage for
translingual terminology retrieval compared with the use of a multilingual
handcrafted thesaurus for searching purposes.
The second main point of this evaluation aims to study the dependence between the
quality of our results, the quality of used linguistic resources and the quality of WTB
processing.
While NLP techniques feed Terminology Extraction and thesaurus construction,
now a thesaurus becomes a very useful resource to give feedback and evaluate the
linguistic processes in a retrieval task.
Evaluation framework
The evaluation has been performed comparing the WTB terminology retrieval over
a multilingual web pages collection, with the European Schools Treasury Browser
(ETB) thesaurus. The multilingual collection comprises 42,406 pages of several
European repositories in the educational domain (200 Mb) with the following
distribution: • Spanish, 6,271 documents.
• English, 12,631 documents.
• French, 12,534 documents.
• Italian, 10,970 documents.
The ETB thesaurus alpha version used in the evaluation has 1051 descriptors with
its translations to each of the five considered languages (English, Spanish, French,
Italian and German). German hasn’t been considered in the evaluation because no
linguistic tools were available to us for that language.
Each ETB thesaurus descriptor has been used as a WTB query. The thesaurus
preferred translations have been compared with the WTB retrieved terms in each
language. In such a way, precision and recall measures can be provided.
Approximately half of the thesaurus descriptors are phrases (poly-lexical terms)
which can be used to evaluate the WTB terminology retrieval. Thesaurus mono-
lexical terms permit the coverage evaluation of linguistic resources used in the
expansion and translation of query words.
Qualitative evaluation
Figure 2 shows the interface for the qualitative evaluation. This interface is aimed
to facilitate inspection on the system behaviour, in order to detect errors and suggest
improvements on WTB system. The first column contains the thesaurus terms in each
language (in the example, therapy, terapia, thérapie and terapia). Each of them are
the preferred terms, or descriptors, in the thesaurus and have been used as WTB
queries. The retrieved terms in each target language are shown in the same row.
Fig. 2. Interface for qualitative evaluation of terminology retrieval (mono-lexical terms).
For example, when searching WTB with therapy (English term), in the first
column, the system retrieves terapeutico, terapia y terapéutica, in Spanish (same row,
second column); it also retrieves therapy y treatment in English (same row, third
column); etc.
Quantitative evaluation
If the preferred term in the thesaurus has been retrieved by WTB, then it is counted as
a correctly retrieved term. Then, precision and recall measures can be defined in the
following way:
• Recall: number of correctly retrieved terms divided by the number of terms
in the thesaurus.
• Precision: number of correctly retrieved terms divided by the number of
retrieved terms.
Figures 2 and 3 show that there are correct terms retrieved by WTB different from
the preferred terms in the thesaurus. Hence, the proposed recall and precision
measures are lower bounds to the real performance.
Fig. 3. Interface for qualitative evaluation of terminology retrieval (poly-lexical terms).
For example, Figure 3 shows how, among the retrieved terms by the English query
“adult education”, only the Spanish term “educación de adultos” adjusts to the
preferred term in the thesaurus. However, there are some morpho-syntactic variations
(“educación de adultas”, “educación de los adultos”), semantic variations
(“formación de adultos”), and related terms (“formación básica de las personas
adultas”) which are correctly retrieved terms by WTB but not counted as such.
Results
WTB retrieved terms have been directly extracted from texts and, for that reason,
recall will depend on the domain coverage of thesaurus descriptors in the test
collection. Although the test collection is very close to the thesaurus domain, it’s not
possible to guarantee the presence of all thesaurus terms in all languages in the
collection. Indeed, thesaurus descriptors are indexes to abstract concepts, which are
not necessarily contained in the texts being indexed.
Table 1 shows the coverage of thesaurus descriptors in the test collection where
exact matches have been considered (including accents).
Coverage Spanish English French Italian
Mono-lexical descriptors found in the collection 84.3% 81.9% 82.3% 81.1%
Poly-lexical descriptors found in the collection 56.5% 57.5% 54.2% 42.6%
Table 1. Coverage of the terms in the test collection with regarding to the thesaurus descriptors.
Mono-lexical term retrieval
Since mono-lexical term expansion and translation only depend of lexical resources,
potential retrieval capabilities can be evaluated independently of the collection, just
counting the mono-lexical thesaurus descriptors present in the lexical resources used
(EuroWordNet lexical database and bilingual dictionaries). Table 2 shows presence of
thesaurus descriptors in the lexical resources (monolingual case, in diagonal) and their
capability to go cross-language. The first column corresponds to the source languages
and the first row corresponds to the target languages. The cell values correspond to
the percentage of mono-lexical thesaurus descriptors recovered in the target language
from the source language descriptor.
Recall Spanish English French Italian
Spanish 91.6% 83.7% 60.9% 64.3%
English 80.4% 97.2% 63.9% 63.9%
French 66.3% 61.8% 85.5% 55.9%
Italian 67.9% 62.2% 53.9% 96.7%
Table 2. Potential recall of mono-lexical thesaurus descriptors in the WTB used lexical
resources.
Comparison with thesaurus give and idea of domain coverage by lexical resources.
Table 2 shows that recall for the Spanish/ English pairs is significantly higher than the
rest. The reason is that Spanish and English languages have been complemented with
bilingual dictionaries while French and Italian only use EuroWordNet relations. Since
monolingual cases show a good coverage, numbers point out that there is a lack of
connections between different language hierarchies in EuroWordNet. In conclusion,
with the currently used resources, we can expect a poorer behaviour of WTB
translingual retrieval implying French and Italian.
Poly-lexical term retrieval
WTB poly-lexical term retrieval depends of the previously extracted phrases from the
document collection and therefore, depends on the coverage of thesaurus descriptors
in the test collection.
Coverage of thesaurus descriptors in the test collection in the monolingual case
(Table 1, last row), gives an upper bound for recall in the translingual cases. Table 3
show WTB recall for each pair of languages in percentage over this upper bound for
the target language.
Recall Spanish English French Italian
Spanish 63.1% 45.8% 19.9% 16.3%
English 40.2% 66.5% 14.7% 7.4%
French 12.5% 15.6% 40.3% 7.8%
Italian 17.1% 17.2% 8.9% 39.3%
Table 3. WTB recall in % respect collection coverage (poly-lexical terms).
As shown in Table ,3 English/ Spanish pairs show better behaviour than other pairs
of languages. The reason for this relies in that poly-lexical term retrieval is based in
the combination of mono-lexical terms, and this depends on the lexical resources
used. Again, just in the case of English/ Spanish pairs, EuroWordNet has been
complemented with bilingual dictionaries and, for that reason, these pairs of
languages present the best behaviour in both mono and poly-lexical term retrieval.
However, differences between mono and poly-lexical terms recall need further
consideration. While mono-lexical terms correspond to nouns, which are well covered
by EuroWordNet hierarchies, most poly-lexical terms include adjective components
which aren’t covered so well by EuroWordNet. This lack has been also corrected only
for English/ Spanish pairs using bilingual dictionaries and this is an additional factor
for a better recall.
The best recall is obtained for Spanish as source language. The reason relies in
that, for this language, WTB uses a morphological analyser which gives all possible
lemmas for the query words. All these lemmas are considered in expansion,
translation and retrieval. In this way, possible lemmatisation errors are avoided both
in query and texts, and increases the number of possible combinations for poly-lexical
term retrieval.
However, the recall values are quite low even in monolingual cases and thus, a
broader study explaining loss of recall is needed. As said, WTB poly-lexical term
retrieval depends of the previous extracted phrases and thus, not only depends of the
test collection. It depends also of phrase extraction, indexing and retrieval
procedures. Table 4 shows the loss of recall due to phrase extraction and indexing
procedures. There are several factors which lead to a loss of recall:
1. Phrase extraction procedure. Loss of recall due to not exhaustive
syntactic patterns and wrong part-of-speech tagging. The loss of recall
due to a wrong phrase extraction procedure is represented by the
differences between first and second rows and oscillates between 2.8% for
Spanish and 17.3% for French (1% to 12% in absolute values).
2. Phrase indexing. Loss of recall due to wrong phrase components
lemmatisation. The loss of recall due to wrong indexing (mainly wrong
lemmatisation of phrases components in texts) oscillates between 2% for
English and 34% for French (1% to 15% in absolute values).
3. Phrase retrieval. Loss of recall due to wrong lemmatisation, expansion
and translation of query words, and wrong discarding in phrase selection
and ranking of terms. WTB discards retrieved terms with document
frequency equal to 1 in order to improve precision in the terms shown to
users. This fact produces a loss of recall between 12.9% for Spanish and
36.7% for Italian (5% to 10% in absolute values).
4. Mismatching caused by accents and case folding. WTB doesn’t need to
separate documents in different languages. For this reason the loss of
recall due to accents mismatching is difficult to quantify here because it
produces a big confusion between languages. For example, there are lots
of terms in English equal to the French ones without accents. Similar
occurs between Italian and Spanish.
All this factors show that not only lexical resources must be improved, but also
linguistic processing tools as lemmatisers and part-of-speech taggers.
Poly-lexical descriptors Spanish English French Italian
found in the collection 56.5% 57.5% 54.2% 42.6%
found among extracted phrases
(loss of recall due to phrase extraction) 54.9%
(-2.8%) 50.1%
(-12.9%) 44.8%
(-17.3%) 40.0%
(-6.1%)
retrieved with WTB
(loss of recall)
(loss of recall due to phrase indexing)
40.9%
(-27.6%)
(-25.5%)
49.1%
(-14.6%)
(-2%)
29.2%
(-46.1%)
(-34.8%)
26.4%
(-38%)
(-34%)
retrieved with WTB discarding df=1
(loss of recall)
(loss of recall due to phrase selection)
35.6%
(-36.9%)
(-12.9%)
38.2%
(-33.5%)
(-22.1%)
21.8%
(-59.7%)
(-25.3%)
16.7%
(-60.7%)
(-36.7%)
Table 4. Loss of recall in WTB poly-lexical term retrieval by steps in the processing.
Regarding precision, Table 5 shows the lower bound for the mono-lexical terms
retrieval.
Precision Spanish English French Italian
Spanish 30.69% 25.93% 30.82% 26.64%
English 32.28% 37.49% 36.22% 30.12%
French 24.96% 25.15% 48.55% 27.82%
Italian 28.77% 27.89% 32.47% 47.06%
Table 5. WTB precision in mono-lexical terms retrieval
Precision Spanish English French Italian
Spanish 16.90% 15.30% 14.77% 12.95%
English 16.88% 23.47% 12.64% 8.44%
French 10.96% 10.35% 22.66% 10.77%
Italian 12.54% 10.57% 10.37% 30.24%
Table 6. WTB precision in poly-lexical terms retrieval
Table 6 shows the lower bound for poly-lexical terms retrieval. In the worst case,
there is one preferred descriptor in average among ten retrieved terms, and three in the
best case. Term discrimination is an easy and very fast task which is helped in the
WTB interface through the term organisation into hierarchies. Table 7 shows that, in
fact, about 70% of the retrieved relevant descriptors are retrieved in the top level of
the hierarchies. This is a good percentage to ensure fast discrimination of retrieved
terms.
Spanish English French Italian
Spanish 74.35% 72.22% 71.18% 94.73%
English 74.39% 73.17% 81.39% 82.35%
French 74.57% 65.78% 77.55% 94.44%
Italian 87.77% 80.76% 79.24% 65.38%
Table 7. Recall in top level in percentage over recall for poly-lexical terms retrieval
Conclusions
Terminology Retrieval gives a shared perspective between terminology extraction
and cross-language information retrieval. From thesaurus construction point of view,
the Automatic Terminology Extraction procedures shift to term searching. From text
retrieval perspective, retrieved terminology becomes an intermediate information
level which provides document access and cross the gap between query and collection
vocabularies even in different languages. This strategy has been implemented in the
Website Term Browser.
The evaluation framework for terminology retrieval has been established in this
paper, being crucial to improve all steps in the processing. While NLP techniques
feed Automatic Terminology Extraction for thesaurus construction, now, in a retrieval
framework, a thesaurus provides a baseline for terminology retrieval evaluation and
gives feedback on the quality, coverage and use of the linguistic tools and resources.
The qualitative evaluation interface shows that WTB is able to retrieve a
considerable amount of appropriate term variations not considered in the thesaurus.
Thus, terminology retrieval becomes a very good complement to thesauri in the
multilingual retrieval task.
The quantitative evaluation results are a lower bound of the real recall and
precision values because correct term variations, different from the preferred
thesaurus descriptors, are not taken into account. Results show a high dependence of
WTB terminology retrieval with respect to the used linguistic resources showing that
EuroWordNet relations between different languages must be improved. Results also
show the loss of recall due to phrase extraction, indexing and retrieval process. Future
work must study the loss of recall due to accent mismatching. We conclude that,
when appropriate resources and linguistic tools are available, WTB show a reasonable
good behaviour, although there is place for improvement.
Future work
Future work will refine the evaluation framework and include the study of infrequent
thesaurus descriptors (especially those not found in the collection). For these
purposes, the construction of a new test collection is planned querying Internet search
engines with the thesaurus descriptors. The crawling of the listed documents will
ensure a thesaurus coverage of 100% and will permit separate processing and
evaluation for each language. In this way, the evaluation of loss of recall due to accent
mismatching will be possible ensuring a better data consistency.
References
Ballesteros, L. and Croft W. B. Resolving Ambiguity for Cross-Language
Information Retrieval. Proceedings of the 21st Annual International ACM
SIGIR Conference on Research and Development in Information Retrieval.
1998; 64-71.
Bourigault, D. Surface grammatical analysis for the extraction of terminological noun
phrases. Proceedings of 14th International Conference on Computational
Linguistics, COLING'92. 1992; 977-981.
Frantzi, K. T. and S. Ananiadou. The C-value/NC-value domain independent method
for multiword term extraction. Journal of Natural Language Processing.
1999; 6(3):145-180.
Hertzberg, S. and Rudner L. The quality of researchers' searches of the ERIC
Database. Education Policy Analysis Archives. 1999.
Peñas, A. Gonzalo J. and Verdejo F. Cross-Language Information Access through
Phrase Browsing. Applications of Natural Language to Information Systems,
Proceedings of 6th International Workshop NLDB 2001, Madrid, Lecture
Notes in Informatics (LNI), Series of the German Informatics Society (GI-
Edition). 2001a; P-3:121-130.
Peñas, A. Verdejo F. and Gonzalo J. Corpus-based Terminology Extraction applied to
Information Access. Corpus Linguistics 2001; Lancaster University, UK.
2001b.
Trigari, M. Multilingual Thesaurus, why? European Schools Treasury Browser. 2001.
Vossen, P. Introduction to EuroWordNet. Computers and the Humanities, Special
Issue on EuroWordNet. 1998.
