Content uploaded by Andrejs Vasiļjevs
Author content
All content in this area was uploaded by Andrejs Vasiļjevs on Oct 13, 2014
Content may be subject to copyright.
Corpus based analysis for multilingual terminology entry
compounding
Andrejs Vasiljevs, Kaspars Balodis
Tilde, University of Latvia
Vienibas gatve 75a, Riga, LV-1004, Latvia
E-mail: andrejs@tilde.lv, kbalodis@gmail.com
Abstract
This paper proposes statistical analysis methods for improvement of terminology entry compounding. Terminology entry
compounding is a mechanism that identifies matching entries across multiple multilingual terminology collections. Bilingual or
trilingual term entries are unified in compounded multilingual entry. We suggest that corpus analysis can improve entry compounding
results by analysing contextual terms of given term in the corpus data.
1. Introduction
This paper addresses some of the problems in terminology
consolidation that are discovered in EuroTermBank
project. This section briefly describes terminology
consolidation challenge and EuroTermBank project. The
next section introduces concept of terminology entry
compounding. The following sections discus possibilities
for application of statistical methods for improving entry
compounding results as well as first experiments in this
field.
Globalization from the one side and growing language
awareness from the other side dictates the need to
consolidate different national terminology resources, to
harmonize international terminology, to provide online
access to reliable multilingual terminology. There are
number of terminology resources and databases provided
by different institutions or national terminology bodies.
However these databases are mostly limited in language
coverage or are subject specific.
EuroTermBank project (Auksoriute et al., 2006) has a
goal to collect, harmonize and disseminate dispersed
terminology resources through online terminology data
bank. The EuroTermBank was developed by 8 partners
from 7 European Union countries – Germany, Denmark,
Latvia, Lithuania, Estonia, Poland and Hungary.
Web-based terminology data bank
www.eurotermbank.com has been developed to provide
easy access to centralized terminology resources as well
as methodology for harmonization of terminology
processes.
Large number of terminology resources was acquired and
processed gathering in total 1.5 million terms from about
600 000 term entries, involving almost 30 languages.
To consolidate representation of multilingual terms an
automated terminology entry compounding mechanism
has been proposed (Vasiljevs, Rirdance, 2007) that
identifies matching entries across multiple terminology
collections.
2. Terminology entry compounding
Entry compounding solves the problem of unified
representation of multiple potentially overlapping term
entries that are present in a consolidation of a huge
number of multilingual terminology sources. Majority of
terminology resources that are available in Eastern
European countries are bilingual with a source language
mostly being English. Much smaller number of resources
is monolingual or has terms in three or more languages.
Since multiple terms in multiple languages can refer to the
same concept, the concept is the shared element that must
be used to link the terms together in a multidimensional
database (Wright, 2005).
EuroTermBank data structure is modeled according to
concept-oriented approach to terminology. Terminology
entry denotes an abstract concept that has designations or
terms as well as definitions in one or more languages. If
terminology bank contains entries coming from different
collections and designating the same concept we have an
obvious interest to merge them into one unified
multilingual entry.
For example, if we have term pair EN computer – LV
dators coming from Latvian IT terminology resource and
another term pair EN computer – LT kompiuteris from
Lithuanian IT terminology resource we may want to join
these two into unified entry EN computer – LV dators –
LT kompiuteris. Such multilingual entry allows to get
correspondence between language terms that are not
directly available in any terminology resource (in our
example new term pair LV dators – LT kompiuteris).
But merging entries just on the bases of matching term in
one language that is common for these entries will lead to
many erroneous term correspondences. For example, if
we have LV-EN entry stumbrs-stick and ET-EN entry
kang-stick, we may want to merge these entries into
compound entry LV-ET-EN stumbrs-kang-stick. But if
we would add to this alignment LV-EN entry
rokturis-stick it would lead to wrong LV-ET translation
rokturis-kang.
Such problems are obvious due to the frequent ambiguity
of terms among subject fields or rarer cases of ambiguity
in the context within one subject field. We can conclude
that the only error-free method for merging entries is
evaluating whether these entries denote the same concept.
Unfortunately in practice it is often impossible or very
expensive to make comparisons of cross-lingual
terminology concepts. There is a lack of experts with
sufficient knowledge of respective languages and subject
2318
fields. The task is considerably hindered by the fact that
majority of EuroTermBank terminology collections does
not have term definitions included.
In EuroTermBank, a practical solution is proposed by
introducing the terminology entry compounding
approach. Entry compounding is an automated approach
for matching terminology entries based on available data.
The most reliable indication for matching entries is
having unique and unambiguous concept identifiers. The
best example is terms from ISO terminology standards.
These term entries have an identifier in the form
[Standard_identifier].[term_number]. Accordingly, all
national standards share the same identifier for
corresponding entries and can be merged with a very high
degree of reliability. Another case of unique
internationally applied identification is the usage of Latin
names in medicine and biology (with a number of
exceptions with different Latin names designating the
same concept). If there is no unique identification for
concepts in collections, less precise matching criteria are
used, namely, the English term and the subject field.
English was chosen as the most popular language in term
resources.
EuroTermBank uses Eurovoc as a subject field
classification. A number of terminology resources use
only top classification levels of Eurovoc but there are
many resources with detailed classification using Eurovoc
sublevels of different depth. For this reason it was decided
to take into account only the top classification level for
entry compounding. This means that sublevels are
equalized to the top classification level.
It is important to understand that entry compounding is a
data representation method that does not propose to create
new terminology entries. It is a visualization aid that
displays matching entries across collections in a
consolidated way. Matches are determined by applying a
number of criteria and as such cannot be error-free.
As majority of terminology resources integrated in
EuroTermBank are bilingual (Table1), we would like to
transform data representation from number of separate
bilingual entries to unified multilingual record.
Entry languages
Number of
entries
Percentage from
total
monolingual
11230
2%
bilingual
398854
68%
3-lingual
45497
8%
4-lingual
69134
12%
5-lingual
48761
8%
>5-lingual
12216
2%
Table 1 Multilinguality of EuroTermBank source records
Entry compounding solves the problem of visual
representation of multiple potentially overlapping term
entries that are present in a consolidation of a huge
number of multilingual terminology sources. At present,
the EuroTermBank database contains over 585,711 term
entries with more than 1,500,500 terms. When applying
entry compounding, over 135,000 or 23% of entries get
compounded. Hence entry compounding is a considerable
aid for the user in finding the required term, for example,
in the translation scenario between language pairs for
which term equivalence is not established in existing
collections.
Unfortunately abovementioned criteria for entry
compounding are insufficient and generate too much
incorrect alignments. High recall rate lead also to
relatively low precision although we currently do not have
exact precision evaluation figures.
If our term entries would include term definitions then we
could compare these by human review or by applying
automated analysis methods. But because large majority
of Eastern European terminology resources do not include
definitions we need to look for other sources to depict
meaning of terms.
We suggest to use multilingual text corpus as a source
were to look for term usage patterns and try to
disambiguate its meaning. Of course it is impossible to get
term definition from the regular text corpus. But we can
intuitively assume that term meaning is related to the
context where term usually appears in. This intuition has
also some rational basis. For cost and time saving many
institutions dealing with terminology creation are not
preparing definitions for new terms but instead include in
term database several typical examples of usage context.
We can assume that term t in language L1 and s in
language L2 are matching (or denoting the same concept)
if t and s have similar context patterns in L1 corpus and
L2 corpus respectively. By the context pattern we mean
characteristic collocates frequently appearing in
proximity of term. Because terminology is related to
special language (special language uses specific words
with specific, preferably unambiguous meaning, in
contrast to general language with wide lexicon of usually
very ambiguous words) we are interested in those
collocate words that are terms from the same subject field.
This is also based on common intuition that term in
specific subject field should be best described by other
terms from this subject field.
3. Proposed method
In the proposed method we try to grasp the intuition that if
two terms in different corpora have similar context
patterns then they might denote the same concept and
more frequent collocations have more impact on term
context pattern than less frequent ones.
Let’s assume that we have applied simple term
compounding for bilingual terminology resources as
described previously. For language L1 term t we have
several translation candidates s1, s2, ..., sn in language L2.
Our task is to select the most probable from these
candidates by analyzing context patterns of these terms.
Let’s denote frequency of term t in language L1 corpus
with count(t).
2319
Frequency of s1, s2, ..., sn in L2 corpus will be denoted
with count(s1), count(s2), ..., count(sn).
We denote collocations of term t with coll1(t), coll2(t), …,
collm(t) and respective frequency of these
collocations in proximity with t with
count(t, coll1(t)), count(t, coll2(t)), ...,
count(t, collm(t)).
We will select those collocations of the term t in language
L1 whose frequency is higher than certain threshold p.
This means that we will select
)(tcoll j
, where
p
tcount
tcolltcount j
)(
))(,(
.
For every such collocation we will find translation
candidate
k
xxx ,...,,21
in language L2. For every
candidate translation si of the term t:
if
p
scount
xscountxscountxscount
i
kiii
)(
),(...),(),( 21
then we will add to the score of this candidate the lowest
from the numbers
)(
),(...),(),( 21
i
kiii
scount
xscountxscountxscount
and
)(
))(,(
tcount
tcolltcount j
.
Now let’s do the same calculation from reverse side – for
every translation candidate si in language L2 we will
select collocations whose frequency is higher than certain
threshold p.
This means that we will select
)( ij scoll
, where
p
scount
scollscount
i
iji
)(
))(,(
.
For every such collocation we will find translation
candidates
k
xxx ,...,,21
in language L1.
If these translations appear in context with t frequently
enough passing our threshold p:
)(
),(...),(),( 21
tcount
xtcountxtcountxtcount k
> p,
then we will add to the score of this candidate the lowest
from the numbers
)(
),(...),(),( 21
tcount
xtcountxtcountxtcount k
and
)(
))(,(
i
iji
scount
scollscount
.
We will assume that translation candidate si with the
highest resulting score is the most probable equivalent of
term t in language L2.
4. Experimental results
To test the proposed method we carried out experiment on
compounding of Latvian and Lithuanian terms. For this
experiment we used JRC-Acquis Multilingual corpus v3.0
which is the largest publicly available source of corpus
data for Latvian and Lithuanian (Steinberger et al., 2006).
Latvian corpus contains 22 906 documents with
27 592 514 words. Lithuanian corpus contains 23 379
documents with 26 937 773 words.
It could be asked why not to use well proven statistical
alignment methods to align terms from these corpora as
these are highly parallel texts mostly being translations
from the same source (English). But as we want to find a
method for more general case of lack of parallel
in-domain data, we split this corpus in 2 parts. For Latvian
corpus we used the first part and for Lithuanian – the
second. In such a way we got sufficiently large corpus of
un-parallel texts for Latvian and Lithuanian.
For experiment we selected 27 Lithuanian terms and 80
corresponding Latvian term candidates. Only terms with
at least 50 occurrences in corpus were selected and only
Lithuanian terms for which there were at least one correct
and one incorrect Latvian term were selected. Correct
translation was depicted by human terminologist.
Every Lithuanian term had from 2 to 8 candidate
translations in Latvian from which only 1 to 4 were
correct.
We implemented the proposed method and made
experiments with different settings of threshold parameter
p.
The size of window for collocations was 10 words to the
left and right of the term occurrences. As Latvian and
Lithuanian are highly inflected languages, morphological
normalization was applied.
To measure the usefulness of our method we chose the
value of threshold parameter p = 0.002.
We say that our method gives correct result on Latvian
term s (which is translation candidate of Lithuanian term
t):
If s is a correct translation of t and its score is at
least 5% higher than for every other incorrect
translation candidate of t.
If s is not a correct translation of t and its score is
at least 5% lower than for any other correct
translation of t.
2320
If scores of correct translation and an incorrect
translation of t differ by less than 5% then we say
that there is not enough difference in score.
Otherwise we say that our method gives wrong
result.
Examples of results are in Figure 1 and Figure 2. On X
axis there are different values of threshold p and on Y axis
are the scores for term pair.
Results of experiment showed that our method gave
correct answer in 61% of cases (for 49 out of 80 Latvian
terms).
Figure 1 Correct Latvian term virsraksts for Lithuanian
term antraštė achieved significantly higher score than
wrong translation priekšnieks
For 21% (17 out of 80 Latvian terms) there was not
enough difference in score.
Figure 2 Example of insufficient difference in score for
Latvian term jauda and Lithuanian term laipsnis.
For 18% (14 out of 80 Latvian terms) the method gave
wrong result.
5. Conclusions and Future Work
Proposed method of statistical corpus analysis of term
context demonstrates promising results to improve
automated terminology entry compounding.
These results encourage further research for different
language pairs and in different domains.
6. Acknowledgements
This work was supported by European Social Funds.
7. References
Auksoriute A. 2006 Towards Consolidation of European
Terminology Resources. Experience and
Recommendations from EuroTermBank Project.,
ISBN 9984-9133-4-1, Riga
Henriksen L., Povlsen C., Vasiljevs A. 2005.
EuroTermBank – a Terminology Resource based on
Best Practice. In Proceedings of LREC 2006, the 5th
International Conference on Language Resources and
Evaluation, Genoa, on CD-ROM , May 2006
Vasiļjevs A., Skadiņš R. 2005. Eurotermbank
terminology database and cooperation network. In:
Proceedings of the Second Baltic Conference on
Human Language Technologies, Tallinn, pp. 347-352.
Vasiljevs A., Rirdance S. 2007. Consolidation and
unification of dispersed multilingual terminology data,
International Conference RANLP 2007 (Recent
Advances in Natural Language Processing), Borovets,
Bulgaria, 2007
Wright, Sue Ellen 2005. A Guide to Terminological Data
Categories – Extracting the Essentials from the Maze.
In Proceedings of TKE 2005, the 7th International
Conference on Terminology and Knowledge
Engineering. Copenhagen, pp. 63-77.
Steinberger R., Pouliquen B., Widiger A., Camelia Ignat
C., Erjavec T., Tufiş D., Varga D. 2006, The
JRC-Acquis: A Multilingual Aligned Parallel Corpus
with 20+ Languages, In Proceedings of the 5th Intl.
Conf. on Language Resources and Evaluations, LREC
2006, May 2006, Genoa, Italy
antraštė
priekšnieks:
antraštė
virsraksts:
laipsnis
jauda:
2321
