Research
meets practice: t-survey 2005
An online survey on terminology extraction and terminology management
(
Yamile Ramírez Safar
(ramirez.yamile@googlemail.com)
Linguistic Data Processing Section
Department 4.6 - Applied Linguistics
& Translating/Interpreting
Saarland University
Abstract
This paper reports the results of an online survey on terminology management and terminology extraction conducted by
the Linguistic Data Processing section of
With this survey we want to investigate the
relationship between research and practice in the area of terminology extraction and evaluate
if there is any need to reconcile both. Aimed at translators, terminologists,
interpreters and project managers, the main goals of the survey are to
investigate the dissemination and application of terminology management tools
(with a focus on terminology extraction tools) and to assess the demands on
today's terminology extraction tools.
1. Introduction
Terminology management has now been taught for several
years at universities and in translation-oriented education programmes. It
continues to be a lively research field involving applied linguists,
computational linguists, and computer scientists.
One
of the most active research areas in terminology management in the last fifteen
years has been terminology extraction
(TE). A surprising amount of applied research in this field has produced
several different approaches (linguistic, statistical, and hybrid) as well as
tools to facilitate the time-consuming process of terminology compilation. In
recent years, there has been an increasing demand for terminology extraction
software (c.f. e.g. Warburton, 2001). However, although a number of term
extractors are already available on the market today, it seems that they are
still not widely used nor do they meet the real needs of translators,
interpreters, and terminologists.
Today,
there is no doubt about the increasing importance of terminology in our
society. Terminology plays a very important role in many different fields such
as standardization, translation, technical documentation, and software
localization. Accurate and complete terminology improves the productivity of
translators and technical writers, and is a prerequisite for successful
communication. Consequently, terminology management dealing with the
preparation, processing, and documenting of a specialist vocabulary has become
an increasingly important activity.
In
the past 25 years, methods, applications and tools for professional terminology
work have changed and developed substantially. Despite these improvements, the
creation, organization, and management of terminology are still time consuming
and cost-intensive tasks. As has been pointed out by the LISA Terminology
Survey Report (2001), term identification and extraction are the tasks that are
the most frequently performed manually and that would significantly benefit
from automation.
Terminology
extraction can be defined as the operation of identifying term candidates (TC)
in a given text and should be terminologically distinguished from term recognition,
which is the operation of comparing the TC lists (the output of TE) with a
terminological database (TDB) in order to identify known/unknown terms.
According
to Lieske (2002), the term extraction process comprises four different tasks,
namely a) the compilation of a corpus (machine readable texts, that serve as a
basis for the TE), b) terminology extraction (identification and extraction of TCs),
c) the evaluation or validation of the results (determining the terminological
relevance of a TC), and finally d) the classification of terminology according
to classes and categories.
TE
has three major applications: terminology management, translation, and information
retrieval (Thurmair 2003), the first and second being the focus of our
investigation. In terminology management and translation, TE is used for the
creation, enlargement, and update of terminology. The main goals are to
facilitate the task of identifying terms and their translations and to increase
the productivity of translators, interpreters, and terminologists. With respect
to the number of languages involved in the extraction process, we can basically
distinguish between monolingual and bilingual TE. In the
translation process, monolingual TE is usually carried out before the translation
starts, either on the source texts or on reference texts. The goal is to
identify all relevant terminology in a text to be translated or - in the case
of terminology - to identify all terms of a certain (sub-) domain. Bilingual TE,
on the other hand, is mostly carried out on translated texts (parallel corpora
or translation memories). Its main goal is the recognition of potentially
equivalent terminological units in two languages (Thurmair, 2003). In both
cases the resulting TCs are often compared to existing terminological databases
in order to distinguish known terms from unknown terms (cf. Saß 2004).
2. Approaches to terminology
extraction
Depending
upon the type of information used as a basis to identify terms, approaches to
TE are usually classified as linguistic, statistical or hybrid. Linguistic and
statistical approaches can be further subdivided into term-based (intrinsic)
and context-based (extrinsic) methods (cf. Bourigault et al. 2001, Streiter et
al. 2003).
Terminology
extraction tools (TETs) following a linguistic approach try to identify terms
by their linguistic structure, e.g. morphological and syntactic structure. For
this purpose, texts are annotated with linguistic information with the help of
morphological analysers, part-of-speech taggers and parsers. Then TCs with a
certain tag structure are filtered from the annotated text by using pattern
matching techniques. Intrinsic methods try to filter TCs according to their
internal structure, e.g. according to the morphological structure (e.g. German "Zylinderabschaltung
(cylinder cutout)" ds=zylinder#ab_$schalten~ung ). Extrinsic methods try to identify
TCs by analysing the morpho-syntactic structure of a word or phrase, such as
looking for part-of-speech sequences like NP= noun + noun (e.g. printer menu).
Another technique is to filter TCs by looking for commonly used text structures
such as definitions and explanatory contexts like "X is defined as " or "X is
composed of " (cf. Pearson 1998, Saß 2004).
The
assumption underlying the statistical approaches to TE is that specialized documents
are characterized by the repeated use of certain lexical units or
morpho-syntactic constructions. TETs based on statistics try to filter out
words and phrases that appear with a frequency higher than a given threshold by
applying different statistical measures (see Manning & Schütze 1999 for an
overview). Term-based statistical methods try to compute the structure of TCs
e.g. by using n-grams. Often, the structure of existing terms is compared to
the structure of words and phrases in a corpus in order to filter out TCs with
a similar n-gram structure (example-based approach; cf. Streiter et al. 2003).
Another common method is to compare the frequency of words and phrases in a
specialized text to their frequency in general language texts assuming that
terms tend to appear more often in specialized texts than in general language
texts.
Different
evaluation criteria exist for TETs, involving among others accuracy, supported
file formats, languages, etc. (for a detailed list see Lieske 2002, Sauron
2002, Saß 2004). The most frequently used criteria are noise and silence, and
recall and precision. While noise refers to the ratio between discarded TCs and
the accepted ones, silence refers to the number of terms not detected by a TET.
Recall and precision are two measures frequently used in IR, the former being
defined as the ratio between the sum of correctly retrieved terms and the sum
of existing terms, the latter being defined as the ratio between correctly extracted
terms and the sum of proposed TCs (c.f. Zielinski 2002).
In
order to eliminate noise from the resulting TC lists, linguistic and
statistical approaches often make use of stop word lists. Stop word lists
contain "empty" words (cf. Carstensen et al. 2001) that are not of interest for
the terminologist, because they do not qualify as terminological units, but
that are often filtered as TCs because of their morpho-syntactical structure or
because of their frequency value (e.g. articles, pronouns, auxiliary verbs, and
prepositions).
Both
approaches, linguistic and statistical, have their advantages and
disadvantages. TETs following a purely linguistic approach tend to produce too
many irrelevant TCs (noise), whereas those following a purely statistical
approach tend to miss TCs that appear with a low frequency value (silence; cf.
Clematide 2003). Linguistic-based TETs often provide better delimited TCs than
statistical-based ones. Furthermore they usually reduce TCs to their canonical
form (base form or lemma) and thus do not provide result lists with repeated
TCs as is the case for statistical-based tools. However, the disadvantage of
linguistically based TETs is that they are language-dependent and thus only
available for major languages. Statistical TETs, on the other hand, can also be
used for lesser-used languages that lack computational resources such as
minority languages (cf. Streiter et al. 2003).
On
their own, linguistic and statistical methods still fail to tackle many of the
basic inherent problems of TE, such as the reduction of TCs to their canonical
form, detection of multi-word terms, term variations, and terms in discontinuous
units (i.e. coordinations, juxtapositions). Since the nature of these problems
is so varied, it seems that only a combination of approaches will help to provide
efficient TETs. In fact, the only method which has been "recovered" by many
authors because of its still unexplored possibilities is the hybrid
approach. Authors such as Cabré (2001) claim that the only way to make progress
in the field of automatic terminology extraction is by combining statistical
and linguistic methods. Fortunately, this seems to be the approach current
investigations in the area of TE are taking (Thurmair, 2003).
Although
there are many investigations about the important role semantic information
plays in the performance of TETs (e.g. to find out the semantic relations
between terms), most TETs currently available on the market do not incorporate
a semantic component (Thurmair, 2003). So, from a practical point of view, it
can be stated that a term is normally considered by TETs as a word or
phrase with a particular structure and belonging to a specific category (noun or
nominal phrase) which additionally occurs with relatively high frequency (Thurmair,
2003).
3. Terminology extraction software -
a short market overview
In
this section we will give a short overview of TETs that are commercially
available on the market. We will briefly present the major tools and compare
their general features and functions, and classify them according to the
approaches outlined in the previous section. Tools developed in research, which
are in most cases prototypes of limited applicability and dissemination, will
not be taken into account. The interested reader is invited to refer to Cabré
et al. (2001) for an overview of such systems developed in the 1990s.
As
has been the case for terminology management systems, in the beginning many
different TETs came onto the market. Many companies tried to offer some term
extraction functionality, either integrated in their CAT-tools (e.g. Trados
TermExtract, now MultiTerm Extract) or as standalone versions (e.g.
TerminologyExtractor from Chamblon). Today it seems that only a few major tools
have asserted themselves on the market. Among these figure MultiTerm Extract
(Trados), SDL PhraseFinder, Xerox Termfinder, Terminology Wizard (Synthema),
and TerminologyExtractor (Chamblon).
Depending
upon the underlying technology (statistical or linguistic), supported languages,
file formats, license type (freelance, network) etc., the price for TETs varies
from around 100 to several thousand euro.
All
TETs provide lists of term candidates that can either be exported for external validation
(e.g. in *.txt, *.csv) or validated directly in the TET. During the validation
process, the user can usually edit TCs, add manually extracted terms, set
status flags (e.g. checked, unchecked), and add TCs to stop word or
abbreviation lists. While the first TETs only provided the user with very long
TC lists with nearly no additional information that could help to validate the
TCs, today's TETs usually link TCs to the context, mostly in form of KWIC
concordance windows. Linguistically based tools usually also provide
information about POS, and canonical form of the TC. Furthermore, all TETs
incorporate sorting functions. TCs can be sorted either according to frequency
or alphabetical or - as in the case of SDLPhraseFinder 2005 - according to a
"confidence index" in which TCs are ranked according to frequency in
combination with the term pattern type. Another feature that is intended to
speed up the validation process and that has been added recently to TETs is
term recognition. In this step TCs are compared to selected terminological
databases in order to distinguish known from unknown words.
In
order to allow the user the adaptation of the extraction algorithms, most
systems provide additional parameters such as the minimum or maximum length of
TCs, minimum frequency of TCs, and minimum or maximum of proposed translations
in bilingual mode. Trados MultiTerm Extract, for example, gives the user the
possibility to adapt the noise/silence ratio by the use of a slider. After the
validation process the accepted TCs can either be exported into one of the
standard exchange formats or - in the case of integrated TETs - into the
respective terminological database. The table below gives a comprehensive
overview over some of the major TETs and their properties.
|
|
Trados MultiTerm
Extract 7
|
SDL Phrase Finder
2005
|
Multitrans (Corpus
builder module)
|
Xerox TermFinder
|
Synthema Terminology
Wizard 3.0
|
Chamblon Terminology
Extractor 3.0
|
|
Term extraction |
mono- &
bilingual |
mono- &
bilingual |
mono- &
bilingual |
mono- &
bilingual |
mono- &
bilingual |
monolingual |
|
Approach |
statistical |
hybrid |
statistical |
hybrid (XELDA Technology) |
hybrid |
statistical |
|
Supported
languages |
137 SL & TL |
nl, de, en, fr, it, pt, es |
all |
de, fr, it, en, sv, es, ru, pt, no, hu, fi, nl, da
(total 15 languages) |
SL= de, en, fr, it, pt, TL=SL+af, da, ca, nl, pl,
pt, ru, es, cz, tr, hr |
en, fr, + ? |
|
Term
identification (single/multi word terms) |
SWU, MWU |
SWU, MWU |
SWU, MWU |
SWU, MWU |
SWU, MWU |
SWU, MWU |
|
Parameters |
Min_length,
max_length, max_num, QA-filter (slider), search for new translations,
max_trans, min_trans_freq |
Exclude_uppercase,
max_length |
Min_length,
max_length, list possible translations |
Handle unknown
words as nouns, convert all terms to lower case, minimal frequency |
Context width,
frequency filter, minimum frequency, text filter, |
|
|
Input |
One or more |
One or more |
One or more |
One or more |
One or more |
One or more |
|
File formats |
doc, html, jsp,
asp, aspx, xml, sgml, qsc, xtg, ttg, tag, rtf, tmx, tmw, ttx, bif, txt |
rtf, html, txt,
SDL TM, SDL itd |
doc, WordPerfect,
txt, Unicode text files, html, xml, pdf, TMX-compliant translation memories,
etc. |
html, sgml, xml,
txt, doc, rtf |
rtf, html, txt,
Trados TM, Transit |
doc, html, rtf,
txt |
|
Stop lists |
yes |
yes |
yes |
yes |
yes |
yes |
|
Term recognition
(existing TDBs) |
yes |
no documentation
available |
yes |
no |
yes |
no documentation
available |
|
Results |
no lemmatization |
no documentation
available |
no lemmatization |
|
lemmatized |
|
|
Additional
information |
context,
frequency, |
POS, root form, |
context,
frequency, source, |
frequency,
context |
context |
frequency,
context |
|
Validation of TCs |
editable list,
different status codes, (check, delete, add translation, generate sample
sentences, add manually extracted terms, etc. |
editable list comments,
|
editable list |
editable list |
editable list,
different status codes, add to dictionary, stop list, abbreviation list |
no |
|
Sorting of TCs |
frequency,
alphabetical |
frequency,
alphabetical, confidence index |
frequency,
alphabetical, |
no documentation
available |
frequency,
alphabetical |
frequency,
alphabetical |
|
Concordance
window |
yes (add as context
sentence) |
yes |
yes |
yes |
yes |
yes |
|
|
Trados MultiTerm
Extract 7
|
SDL Phrase Finder
2005
|
Multitrans (Corpus
builder module)
|
Xerox TermFinder
|
Synthema Terminology
Wizard 3.0
|
Chamblon Terminology
Extractor 3.0
|
|
Export format |
MultiTerm, MultiTerm
XML, txt, |
SDL TermBase
online, tab delimited |
no documentation
available |
OLIF II |
eif, xls |
txt |
|
Exportable info |
user definable |
user definable |
no documentation
available |
no documentation
available |
lemmatized term,
translation, frequency, status, example sentence or segment |
terms, example
sentences (not both at the same time), with/Without frequency |
|
Export filter |
yes |
yes |
yes |
no documentation
available |
yes |
yes |
|
Component or
single product |
part of Trados
CAT-Tools, integrates with MultiTerm, Workbench, Workbench |
integrates with
SDL family of products, SDLX,
TermBase online |
integrates with
Multitrans |
part of Xerox
Terminology Suite, integrates with TermOrganizer |
standalone |
standalone |
|
Price |
Freelance (1055 ) |
? |
568.00 - 2060.00
|
negotiable |
500 (mono-), 800 (bilingual) |
~ 100 |
Table 1: TETs and their properties
It
should be added that several evaluations of TETs have been conducted by
researchers and practitioners. Even if it is not our intention to discuss them
here, we would like to refer the interested reader to some of these evaluations
that may serve as a starting point for further reading: Saß (2004), Lieske
(2002), Sauron (2002), and the report of the Schweizer Bundeskanzlei (2001).
4. t-survey 2005 - Description of
the survey's design and goals
t-survey was made available on the Internet
in English, French, German and Spanish from mid-May 2005 until September 2005. Its
target group includes translators, interpreters, terminologists and project managers
who work on a freelance basis or as employees. Therefore, in order to reach as
many language professionals as possible, the survey was promoted in many major
CAT mailing lists, through several terminology portals (TermNet, DTT), and by
translator and interpreter associations (BDÜ, ITI). Over 400 professionals all
over the world have responded to the questionnaire.
Unlike
other terminology surveys carried out before (e.g. the LISA terminology survey (2001)), the purpose of our survey was to examine
the relationship between research and practice in the area of terminology
extraction and to evaluate if there is any need to reconcile the two. In
comparison with t-survey, the LISA terminology survey had completely
different purposes, and its research area was differently defined. The
Localization Industry Standards Associations (LISA), responsible for providing
practical advice, business guidelines and standards information for translation
and localization workflow tools, designed a series of surveys in the area of
terminology management. These surveys were carried out in 2001 and aimed mainly
at the investigation of the methods, tools and practices for managing
terminology in the localization industry (Warburton, 2001:1-5).
Our
t-surveys main goals are to
investigate the dissemination and application of terminology management tools
with particular focus on terminology extraction tools (TETs), and to assess the
demands on todays TETs. Analysis of the users responses should allow us to
determine whether there are differences between the requirements of various
professional groups working with terminology, i.e. between translators and
terminologists, and, if so, what these differences are. Furthermore, it should
enable us to summarize the required functionalities and to formulate design
criteria for TETs that fit the different user profiles. In summary, the purpose
is to obtain useful information which properly used will have an impact both on
research as well as on software development.
The
survey was designed in a way that no previous knowledge about terminology
extraction was required by the users. For maximum user-friendliness and minimum
frustration that may result from lengthy forms with partly irrelevant
questions, the forms were designed to change dynamically according to the
answers. The surveys design was based on a task model of translation where, for
example, the translation of documents was viewed as a combination of tasks to
include analysis, terminology research, translation, term extraction, and term
management. The total time estimated to answer the questions was approximately
10 minutes.
t-survey is divided into five sections. The
first section, called Personal
Information, is aimed at identifying the users profiles, such as by age, profession,
type of employment, professional experience, and type of education/formation.
The
second section, Working Environment,
focuses on the working conditions and resources that professionals use. This
refers, for example, to information such as company size (number of employees),
available computer facilities (type of operating systems) and availability of Internet
resources (type of Internet connection).
Working
languages, areas of specialization (such as technology, law, economics or
others), text sorts (e.g. operating instructions, patents, software
documentation, commercial reports, web pages, business correspondence, laws,
certificates, scientific publications, among others), text size (calculated in
standard pages) and file formats (*.txt, *.html, *.doc, *.xls, *.pdf, etc.) are
some of the types of information requested in the surveys third section called
Translation, the main goal of which
is to gather information about the translation process itself.
In
the fourth section, Terminology Management,
the central point of interest deals with the process of compilation and storage
of terminology. Questions are related to the types of terminology work (e.g.
monolingual, bilingual, multilingual) and the frequency in which it is carried
out, the method of managing terminology (index cards, Word/Excel tables,
relational databases, terminology management systems), the types of information
that are stored in the terminological data base (for example, nouns, noun
phrases, verbs, collocations, etc.), the criteria for recognizing terms,
additional information gathered together with terms (such as administrative
information, contexts, definitions, grammatical and semantic information,
graphics, etc.) and the estimated time to search terminology during the process
of translation.
The
fifth and last section, Terminology
Extraction, is aimed at obtaining concrete information on the use of TETs
by the respondents, such as the type of tools used or tested (e.g. MultiTerm Extract
(Trados), SDLPhraseFinder (SDL), TermFinder (Xerox), TermFinder (Acrolinx) and
Autoterm (IAI)), the purpose of carrying out terminology extraction (extraction
of terminology from source texts, from reference texts, from translation
memories), the performance of the TETs tested or used, and finally the desired
functions for TETs and the amount of money available to be spent on a TET with
the desired functions included.
Following
the design of the survey and viewing translation of documents as a combination
of tasks, as previously mentioned, it can be deduced that sections 3-5 of the
survey cover three stages of the translation process, namely text analysis,
terminology identification and terminology management.
Concerning
the specific goals of this survey, we primarily want to identify the profiles
of those participants who carry out terminology work and those who do not. The
idea was to have information from primary sources about the use of TETs in
practice and the reception these tools have among language professionals.
With
regard to those participants who claimed to be involved in terminology
activities, we are especially interested in the following points:
(1) to what professional groups do they
belong,
(2) what are their qualifications,
(3) what is their knowledge level
regarding computer technology and terminology,
(4) what are the frequency and type of their
terminology work,
(5) for what purposes do they perform
terminology work,
(6) what importance is given to
terminology tasks,
(7) with what languages and language
combinations do these professionals work most of the time, and
(8) what TETs are mostly used and
preferred.
With
regard to those participants who do not carry out terminology work, we also
want to determine their user profiles and the reasons they give for not taking
into consideration terminology tasks in their work. Moreover, we consider it
important to learn if these participants have tested some TETs before and if
they would be willing to use them, provided that the tools would satisfy their
needs. Finally, since the economic aspect was also considered to be important,
all participants were asked to indicate the price they would be willing and
able to pay for a TET which satisfies their needs.
Based
on this detailed overview of the users profiles it should be possible to accurately
determine the users requirements, which should be the leitmotif for research in the field of TE for the development of TETs.
5. Results and analysis
We will now present the results of the survey
and their respective analysis section by section. A total of 451 participants
filled in the survey, expending an average time of 15 minutes[2].
The high number of participants is a sign of the great interest translators,
interpreters and terminologists have in issues related to translation and
terminology management and developments in the area of CAT tools.
5.1 Personal
information
Especially translators seemed to be very
interested in terminology extraction and terminology extraction tools. This can
be concluded from the fact that the majority of the professionals who
participated in the survey were translators (371). 81 of the participants were
interpreters, 56 project managers and 61 terminologists. Since more than one
profession could be selected, 250 participants indicated working exclusively as
translators, 5 as interpreters, 12 as project managers and 29 exclusively as
terminologists (see figure 1). 86 participants indicated having additional
professions, which means that a large number of the participants could be
considered highly qualified. To the question regarding the working status, 292 of
the total participants indicated working as freelancers, 126 as employees and
only 24 both as freelancers and employees (see figure 2).
Figure 1 Figure 2
Combining information on professions and
working status, it turns out that the majority of translators and interpreters
work as freelancers whereas the majority of terminologists and project managers
work as employees. Of all professional groups, freelance translators were the largest
one (see figure 3).
Figure 3
In general, participants were between 20 and 75
years old. For all participants, the average age is around 40 years and the
average working experience is 5 years or more (see figure 4).
Figure 4
Regarding qualifications, 63% (235) of all
translators have a degree as translators; 77% (62) of all interpreters as
interpreters; and only 16% (10) of all terminologists as terminologists. 35%
(160) of the total participants mentioned other degrees in fields like
technology, natural sciences and humanities. Based on these figures, it can be
said that the majority of the professionals working as translators and the
majority of professionals working as interpreters have a degree in their
special field, whereas this does not seem to be the case for terminologists (see
figure 5). This may be related to the fact that, although there are
universities and educational centres which include terminology courses in their
study programs, very few of them offer a degree in terminology.
Figure 5
Many participants had additional qualifications
such as certifications as translators or interpreters or have attended
specialization courses: 29% were certified translators, 43% certified
interpreters and 6% both certified translators and interpreters (see figure 6).
The number of additional qualifications underlines, as mentioned previously,
that the majority of participants are highly qualified.
Figure 6
5.2 Working environment
61% of all participants answered that they work
in companies with 1 to 5 people; 3% in companies with 5 to 10; 4% in companies
with 10 to 20; 21% in companies with 20 or more people. The large number of
participants working in companies from 1 to 5 people is related to the fact
that the majority of the participants being either freelance translators or
interpreters. From the figures it can be assumed that the majority of employees
work in companies with 20 or more people.
Regarding the use of operating systems, it
turned out, as expected, that 85% of all participants work with one of the
Microsoft Windows operating systems; 6% with Linux and 5% with Mac. The
predominance of Microsoft Windows operating systems is maybe due to the fact
that nearly all CAT tools are only available for this platform. As far as the
Internet is concerned, 96% of all participants indicated using the Internet for
their daily work. From the types of Internet connections, xDSL turned out to be
the leader with 72%, followed by ISDN (18%) and modem (10%). 75% of all
participants have a flat rate whereas only 25% of them have a connection with
download or time limit. These results show that translators, interpreters and
terminologists make extensive use of the Internet for their work, be it as a
source of knowledge (information on specialized areas, reference texts), as a
source of linguistic information (terminology, usage examples, contexts), or as
an absolutely indispensable tool when it comes to project management
(reception, processing and delivery of translation orders, etc.). In sum, these
figures confirm - as is common knowledge - that the Internet is one of the
most important tools for translation and terminology work.
5.3 Translation
The results of the third section show that all
participants work with a wide variety of languages, mainly the official
languages of the European Union, with English, German, French, Spanish and
Italian predominating (see figure 7). When interpreting these figures, it is
important to keep in mind that our survey was available in four of the
predominating languages and that it was announced mainly in European mailing
lists and Internet portals.
Figure 7
The average number of working languages per
participant lies between 3 and 4. These numbers can certainly also be explained
by the high percentage of freelance translators who participated in the survey.
Concerning the language combinations of the five most frequently used languages,
the pair English-German has shown to be the most common one, followed by
English-French, German-French, English-Spanish and German-Spanish (see figure
8).
Figure 8
From the areas of specialization, technology is shown to be the most
important area with 274 answers, followed by economy with 222 and law
with 178 answers. Among other specialization areas mentioned by the
participants, medicine and pharmacology, arts, culture, music and politics were the most frequent ones. When comparing the three main
specialization areas with the five major working languages, it seems that the
proportions between technology, economy and law are equal throughout the languages, which confirms their
relevance and predominance in comparison with other areas. It is also worth
noting that in all languages technology
is always the predominating specialization area, followed by economy and law (see figure 9).
Figure 9
The text types participants most frequently
work with are manuals, operating instructions, web pages, business correspondence, software
documentation, commercial reports,
and training material (see figure
10). The text types from manuals to
commercial reports are the ones
with the highest demand on translation and all come from the specialization
areas technology and economy, whereas the text types with the
lowest demand on translation come from the specialization area law. This confirms the results shown in
figure 9 where law is less frequent
than technology and economy.
Figure 10
Nearly all participants are used to working
with the standard file formats. Since the majority of participants use
Microsoft Windows operating systems (as shown in section 5.2), it is not
surprising that *.doc is by far the most frequently used file format, followed
by *.pdf, *.xls, *.ppt, *.txt, *.html and *.xml. Other formats such as *.rtf,
*.fm, *.qxd or *.indd are significantly less frequent (see figure 11). Apart
from *.pdf and *.ppt formats, most of the commercial TETs support all these
standard file formats (see table 1 in section 3).
Figure 11
5.4 Terminology
Management
91% of all participants carry out
terminological work. This is surprising because the majority of participants are
freelance translators, and these professionals normally work around the clock,
under pressure and have little time left to manage their terminologies. These
results, in turn, underline the importance of terminology management that has
been outlined in section 1. As can be seen in figure 13 this is true both for
translators and interpreters. 45% of the participants indicated that they
always carry out terminological work, 45% that they do it only if time permits.
Finally, 10% do terminological work only on demand (see figure 12). Participants
who do not carry out terminological work said, among others, that terminology
work is too time-consuming, that there is no need for it, that terminology is
usually delivered by the customer or by a terminology service provider or ones
own terminology department.
Figure
12 Figure 13
Regarding the type of terminology work, 63% of
all participants stated that they perform bilingual terminology work, 23% that
they perform multilingual terminology work and 14% that they perform
monolingual terminology work (see figure 14).
Figure 14
When comparing the different types of
terminological work in relation to the professions, for translators and
interpreters, the percentage for bilingual terminology work is significantly
higher than the percentage for mono- and multilingual terminology work. On the
other hand, in the case of terminologists, the percentage of multilingual
terminology work is higher than mono- and bilingual terminology work. It is
worth noting that for terminologists the percentage of monolingual terminology
work is significantly higher than in the case of translators and interpreters.
This is probably due to the job profile of terminologists, since their main goal
is not translation, but primarily management and standardization of
terminology.
Figure 15
Terminology management is mostly carried out by
using appropriate terminology management systems (TMS) (71%). Word and Excel
tables (48%) and the storage of reference texts as source of terminology (37%)
also play an important role. Relational data bases (10%) and index cards (6%)
are far behind (see figure 15). Again this confirms the trend among translators
to be up-to-date when it comes to advanced technologies (cf. section 5.1).
Figure 16
When comparing the use of TMS among the
professions, more terminologists (82%) use TMS than translators (69%) and
interpreters (53%). Word and Excel tables are used more by translators (51%)
and interpreters (66%) than by terminologists (see figure 16). These figures
can be explained by the fact that terminologists are very specialized when
managing terminology and thus require more specific terminology management
solutions. One possible explanation for the relatively high percentage of Word
and Excel tables used by translators and interpreters may be that word
processor and spreadsheet software have a wider dissemination than TMS, and
that translators and interpreters are more familiar with the use of this standard
software. Another possible explanation might be that translators and
interpreters are not as familiar with the principles and methods of terminology
management as terminologists are.
Figure 17
When it comes to TMS, TRADOS Multiterm is the
undisputed market leader. 74% of all the participants who answered the question
on the use of TMS work with TRADOS Multiterm. Termstar, Déjà Vu and SDLX
Termbase are far behind (see figure 17). Some participants mentioned that they
carry out terminology work using the glossary and terminology management
components of other translation memory systems like Wordfast (8%), Across (1%)
and Multitrans (1%). 3% of the respondents work with TMS developed on their
own. It is worth noting that the predominance of TRADOS may be, among others, related
to the fact that it was one of the first commercial terminology management systems
on the market.
Regarding the identification of terminology,
translators, interpreters and terminologists indicated that they do not only store
nouns and noun phrases (96%) in their terminology databases, but also verbs
(82%) and collocations (78%).
When it comes to the criteria used to determine
the terminological relevancy of a word or word group, pragmatic criteria
predominate clearly. 69% of the participants include unknown words in their
terminology (familiarity criteria), 67% consider frequency an identification
criterion for terms, 35% also use formal criteria - such as similarity to known
terms - and 22% indicate other criteria (see figure 18).
Figure 18
The other term identification criteria indicated
by the participants can be clustered as follows: 65% is related to semantics
(context, expert knowledge, etc.) and 30% to pragmatics (e.g. time spent on terminology
research). 5% of the participants indicated that they identify terms
intuitively. The number of participants who did not answer this question at all
is surprisingly high (137). This leads us to the assumption that many
participants have no conscious but rather intuitive knowledge of how to
identify terms.
When asked which type of information they save
in their databases, 22% of the participants indicated that they do not save any
additional information apart from the term and its translation (see figure 19;
nothing). The most frequently saved types of information are contexts (68%;
cont), definitions (68%; def), information on usage restrictions (60%; uinfo),
subject codes (59%; scode), administrative information such as status, last
modified by (40%; adm), grammatical information (33%; gram) and semantic
information (synonyms, antonyms) (32%; sem). Less frequently saved types of information
include semantic relations (part-whole, hyponym-hypernym) (15%; trel) and figures
and graphics (14%; illu). Comparing the types of information stored by
translators, interpreters and terminologists, it becomes clear, as expected,
that the different information types are more often stored by terminologists
than by the other professionals. This can be easily recognized in figure 19,
particularly for semantic information, semantic relations, definitions,
illustrations, and grammatical information.
Figure 19
There is a surprisingly high percentage of participants
who indicated that they get terminology from their customers (80%), whereas
only 1% always get terminology from their customers, 49% sometimes and 50%
rarely. 20% do not get any terminology from their customers at all. When asked
if the delivered terminology was complete or not, 59% answered 'no', 2%
answered 'yes' and 39% answered 'sometimes'. Though the majority of freelance
translators said that their customers frequently provide terminology, most of
the terminology is incomplete. This means that, in the end, translators
definitely have to spend time on terminology work (investigation, validation,
etc.).
Figure 20
In fact, as indicated by translators and
interpreters, the average estimated time expended on terminology work is about
30% of the total time expended on a translation order. This is certainly a
significant and by no means negligible amount. When carrying out highly
specialized translations, this percentage may reach up to 70% or more (see
figure 20). Therefore, automation in terminology management would probably be
very welcome by translators and interpreters.
5.5 Terminology
Extraction
In order to determine the dissemination of
TETs, participants were asked if they have heard of TE before, and if they had
any experience with TE tools. While the majority (71%) of all translators, interpreters
and terminologists have already heard of terminology extraction, 29% have never
heard of it before. Only 15% of all participants that answered the questions in
this section use TETs for their daily work; the majority (85%) do not use them.
People who work as terminologists use TETs significantly more frequently (40.9%)
than translators and interpreters (15.2% and 13.9% respectively). Nevertheless,
the majority of the terminologists still do not use TETs (59.1%) (see figure
21).
Figure 21
Some of the reasons given by participants for
not using TETs were the following in order of importance:
1. The performance of terminology extraction
tools is not good enough (98)
2. The results lists of the terminology
extraction contain too many wrong term candidates (85)
3. Terminology extraction tools find very few
terms (71)
4. The time spent checking the list of term
candidates is too long (55)
5. Based on the results lists of the
terminology extraction I cannot decide whether a word is a term or not (47)
6. Terminology extraction tools are too
expensive (15)
7. I have never heard about terminology
extraction tools (9)
TETs are mainly used to extract terminology
from source texts, translation memories and reference texts. Figure 22 shows
the proportions. The difference between the results regarding terminology
extraction from source texts (43) vs. terminology extraction from translation
memories (32) can be explained with the working scenario of translators and
terminologists. Source texts are always available for translators and
terminologists, whereas translation memories often do not exist for specific
domains (sub-domains). Reference texts are often unavailable as well and it can
be a very time-consuming and difficult task to collect them.
Figure 22
Compared to the actual use of TETs, the participants'
potential interest in using these tools is expressed in the following figures:
253 (85%) would use TETs to extract terminology from reference texts, 179 (61%)
from translation memories and 169 (57%) from source texts. These figures are
interesting, since they are exactly the opposite of the actual usage of TETs.
In other words, the participants who work with TETs use them to extract
terminology mostly from source texts and less frequently from reference texts,
while the potential users of TETs would like to use them to extract terminology
primarily from reference texts and secondarily from source texts (see figure
23).
Figure 23
Concerning the TETs used, TRADOS also turned
out to be the undisputable market leader with MultiTerm Extract (56%); far
behind comes SDLPhraseFinder (12%). The high dissemination of MultiTerm
Extract, in comparison with other considered TETs, may be explained by the fact
that MultiTerm Extract fully integrates within the most frequently used CAT
software package (cf. section 5.4), easily available on the market, and which
supports the most file formats and languages (see table 1, section 3). In
addition, its price is certainly more affordable when compared to other tools incorporating
linguistic knowledge. These tools are significantly more expensive due to their
high development costs (see table 1, section 3). Finally, it is also
interesting that 12% of the participants use TETs developed on their own and
that 7% use concordance programs (e.g. WordSmith) to identify terminology in
texts (see figure 24). This confirms - as has been said earlier - that existing
TETs still do not meet the users' requirements.
Figure 24
In order to determine the dissemination of
TETs, participants who do not use TETs were asked whether they have at least
tested them. While 33% of the participants answered that they have tested TETs,
67% answered that they have not. Regarding professions, interpreters who do not
work with TETs turned out to be the group which has tested them least (77%),
followed by translators (73%). As expected, the percentage of terminologists
(73%) that do not use TETs but have tested them is much higher than that of translators
(28%) and interpreters (23%). This fact underlines, once again, that
terminologists have more specialized needs (see figure 25).
Figure 25
Among the TETs participants have tested,
MultiTerm Extract has the highest percentage (55%), followed by others (20%),
which interestingly are mostly concordance programs used for TE. This leads us
to the supposition that concordance tools might incorporate useful functions
for terminology research. 10% of the participants have also tested
SDLPhraseFinder, whereas the percentages of other TETs tested (Term Finder
(Xerox), Term Finder (Acrolinx), Autoterm), vary between 3% and 4%. This is
certainly related to their minor dissemination, which can be explained by their
language dependence, higher prices, etc. (see figure 26).
Figure 26
In order to understand which tools tend to be the
preferred among the ones tested, we compared the figures of tested TETs with
the number of used TETs (see figure 27). Only 30% of those who tested MultiTerm
Extract use it for their daily work. The percentage of the people who
tested/used the other tools is significantly higher even if these tools are
much less disseminated (SDLPhraseFinder 60%, 71% Xerox Term Finder, and 50%
Term Finder (Acrolinx) and Autoterm). One possible explanation for these
figures might be that language professionals prefer to work with linguistically
based term extraction tools instead of statistical-based ones. Unlike MultiTerm
Extract, all other TETs include linguistic knowledge.
Figure 27
As shown at the beginning of this section, for the
majority of the participants the reason for not using TETs is dissatisfaction with
the performance of these tools. Since we wanted to make their assessment
measurable, we asked the participants to judge the performance of TETs on a
scale from 1 (excellent) to 6 (insufficient). According to their answers, the
average performance of TETs turned out to have a value of 3.75, which is still far
from being excellent (see figure 28).
Figure 28
In order to get information about the users'
requirements, we asked participants what functionalities they desire in a TET. The
majority of the participants (278) want TCs to be linked directly with their
contexts. This is probably because context usually plays a very important role
in the determination of the terminological relevancy of a lexical unit and thus
for the identification of term candidates. However, it is worth noting that
only a few participants indicated context as criterion for selecting terms (see
section 5.4). This is further support for the assumption that many language
professionals identify their terminology intuitively and do not dispose of well
defined criteria to identify terms.
Figure 29
A high number of participants (229) also indicated
they would like their TETs to communicate with the terminological database in
order not to propose already known terms as TCs, a functionality which is still
not available in all TETs (see section 3). Other desired functionalities were,
for example, that TETs propose synonyms and antonyms (208), or that they show
semantic relations for the terms contained in the database and/or in the text
(137), both functionalities clearly in connection with semantic criteria (see
figure 29).
Ideas about the price for TETs that provide the
functions described above vary from 0 to 5000 according to professions. The
majority of the participants (85) are willing to pay between 100 and 250, 54
persons would pay between 250 and 500, 34 between 500 and 750. Only a
minority is willing to pay more than 750. The average price that participants
would pay for such a TET is 353. One trend is very obvious: the more expensive
a TET is, the less probable it is that users would buy it (see figure 30).
Figure 30
The information on prices is very interesting
when compared to the actual prices of TETs available on the market (cf. table
1, section 3). According to table 1 in section 3, Chamblon Terminology
Extractor seems to be the only TET to fit the price desired by the majority of
the participants (around 100), followed by Synthema Terminology Wizard
(between 500 and 800) and MultiTerm Extract from TRADOS (1055 for the
freelance edition). The other TETs available are by far more expensive.
Figure 31
Finally, we compared how much the different
professionals would spend on a TET. It turned out that terminologists (674) are
willing to pay more than project managers (457), interpreters (284) and
translators (274). As mentioned previously, this is certainly due to their
more specific needs when working with terminology. For translators and
interpreters, these results seem to be coherent since they work mostly
freelance and cannot afford to invest much money on a TET. In fact, the prices
translators and interpreters would be willing to pay lie under the overall
average price expectation (353). In the case of project managers, the price
that these professionals are willing to pay for a TET is higher than the
average. This may be due to the fact that they normally work in a company which
can probably afford to pay more for CAT tools than freelancers can.
6. Conclusions
The survey showed that even if many
translators, interpreters, terminologists and project managers have already
heard about terminology extraction, and a relatively large number of them have had
some experiences with TETs, TETs are still not widely disseminated. Only few
language professionals, mainly terminologists, use TETs for their daily work
(15% against 85% that do not use them), which has been explained by their
working scenario and their special needs in terminology management.
Based on the survey's results, we know that for
many translators and interpreters the basic concepts of terminology work (what
is a term, how to recognize it, etc.) are often not clear or systematized. In
fact, the majority of professionals apply pragmatic criteria in order to
identify terms such as the TC frequency in a text (67%) or the TCs' similarity
to other known terms (22%). Concerning terminology management in general, a
relatively high percentage of respondents admitted to managing their
terminology within CAT systems like WordFast and Déjà Vu, translation memory
systems that additionally offer some very basic functionalities for
terminology management (mostly glossaries). Regarding TETs, it has been
documented that they still do not seem to fulfil the requirements of
translators and interpreters, the main reasons being the noise and silence
problem, the lack of information for the validation of TCs, and, finally, the
high price. The lack of knowledge about the principles of terminology work and -
according to the users - the poor performance of TETs could in turn explain why
translators and interpreters mostly do not use TETs or instead use translation
memory systems for terminology purposes.
Nevertheless, the survey documents that there
is a lively interest among all interested professional groups in terminology
extraction issues. This interest can be explained by the large part of
terminology work involved in translation and the importance of terminology in
general. Therefore, many translators, interpreters, and terminologists might be
future potential users of TETs.
Translation-oriented
studies
Concerning the impact on translation-oriented
studies, the results of our survey show that, for example, participants do not
seem to be completely sure about the properties of a term and about how terms
can be identified. Therefore, they mostly stick to pragmatic criteria to determine
if a lexical unit is a term or not. Given the importance of terminology work in
the translation process (see 5.4), courses on terminology and terminology
management need to become a fixed component in translator and interpreter
education programs - which is still not always the case. Furthermore, given
that the average age of participants is around 40 years, and that 15 years ago
there was nearly no terminology module in translation-oriented study courses,
there is also a need for further education courses in terminology work. This
should not be understood as merely training courses for the different tools
offered by many tools specialists or by the software companies themselves, but rather
as courses that focus both on the theoretical and practical side of terminology
work and that give an overview of the recent developments in this area. This
latter type of course is still very rare.
In sum, the importance of terminology work
documents the possibility of an increasing specialization on terminological
issues that could result in an increasing need for terminologists and/or
terminology service providers. This would advocate for the creation of appropriate
terminology-oriented study courses.
For all types of courses that deal with
terminology, professional teachers and trainers are still needed. The same
holds true for didactical approaches and methods to train these trainers.[3]
Research
The survey has shown that the majority of
participants do not seem to have concrete ideas concerning the semantics of
terms. In most cases the semantics of terms was determined intuitively, using
structural linguistic knowledge of terms or pragmatic criteria, such frequency.
This is, among other reasons, probably due to the lack of detailed semantic
descriptions of terms. Therefore, we would suggest that research should further
investigate the properties of terms, above all their semantic properties, in
order to deliver accurate descriptions of terms that are the prerequisites for
the definition of more accurate and practical criteria for term identification.
The results of the survey strongly indicate a
trend towards the preference for linguistically-based TETs. These types of TETs
require computational resources and tools for high-quality linguistic analysis
involving e.g. robust full parsing and semantic tagging. Since such tools are
still under development and are not yet applicable in industrial contexts,
research still has to make progress in these areas of language technology (see
also section 2). This is true for both- major and minor languages.
Another research field besides terminology and
computational linguistics that could contribute to the improvement of TETs is corpus linguistics. Many participants
indicated that they save contexts and definitions as additional information in
their terminological databases. Finding useful contexts and definitions is undoubtedly
a task that is at least as labour-intensive and time-consuming as term
identification itself and thus could also significantly benefit from
automation. This requires criteria for determining what makes a context of a
term useful, as well as for identifying such contexts. The same is valid for
definitions and for techniques for automatic definition extraction.
Software development
Based on the results of the survey it can be
determined that many terminologists, translators and interpreters are still not
satisfied with the overall performance of TETs and that they demand high-quality
tools. One of the direct consequences for the development of TETs is that these
tools should be based on linguistic knowledge, implementing language technology
in order to fulfil the users' expectations (lemmatized TCs, properly delimited
TCs, duplicate TCs, term recognition, term variant recognition, etc.). Since
linguistically-based TETs are language dependent, it is obvious that, in the
beginning, they can only be developed for the most frequently used working
languages. Only if computational resources for minor languages are available
can the product palette be gradually extended. In the meantime the only
practical solution lies in the (additional) implementation of statistical approaches
- knowing that these will not completely fulfil the user requirements. Since
translators, interpreters and terminologists mainly work with 3 to 4 languages,
the adoption of linguistic approaches to terminology extraction seems
justified.
In a few words, the language professionals under
investigation do not need tools that function for all languages, but tools that
work well for their working languages. Table 2 summarizes the gathered
information related to the users' requirements and profiles.
|
|
Terminologists |
Translators |
Interpreters |
|
Professional
status |
employees (66%) |
freelancers (76%) |
freelancers (84%) |
|
Company
size |
10-20 (14%), >20 (64%)
people |
1-5 (76%) |
1-5 (80%) |
|
Internet
connection |
available (flatrate) |
available (flatrate) |
available (flatrate) |
|
Working
languages |
2-4 |
2-5 |
2-4 |
|
Specialization
Areas |
tech, law, econ |
tech, econ, law |
law, tech, econ |
|
Text
types |
web pages, laws,
regulations, manuals, scientific |
manuals, operating
instructions, web pages, |
no data |
|
File
formats |
doc, pdf, html, txt, xls,
ppt, xml |
doc, pdf, xls, ppt, txt,
html, xml |
doc, pdf, xls, ppt, txt,
html |
|
Term
recognition |
yes |
yes |
yes |
|
Integration
into CAT systems |
yes |
yes |
yes |
|
Type
of TE |
multi, bi, mono |
bi, multi, mono |
bi, multi, mono |
|
Sources
for TE |
source texts, TM, reference
texts |
source texts, TM, reference
texts |
no data |
|
|
Terminologists |
Translators |
Interpreters |
|
Linguistic
categories |
nouns, verbs, NP/VP
phrases |
nouns, verbs, NP/VP phrases |
nouns, verbs, NP/VP phrases |
|
Term
identification |
frequency, familiarity,
similarity, context |
frequency, familiarity,
similarity, context |
frequency, familiarity,
similarity |
|
Required
information |
cont, def, uinfo, adm, gram,
sem, trel, illu |
cont, def, uinfo, adm |
cont, def, uinfo |
|
Price
idea |
~ 670 |
~ 280 |
~ 280 |
Table 2: Information
related to users' requirements and profiles
One aspect of TETs that has been criticized by
participants is the validation process. Many of them indicated that the
validation of TCs often is a very time-consuming task: on the one hand, because
too many TCs are extracted, and on the other hand, because it is difficult to
determine the terminological relevancy of TCs based on the information provided
in the results lists. Many participants claimed that additional information
such as contexts and definitions are required for TC validation. Again,
automatic context extraction and definition extraction could be useful
functions. Reducing the time needed for validation seems a necessary
prerequisite for the acceptance of TETs.
As has been seen, translators, interpreters and
terminologists are not only interested in extracting and saving terms and their
equivalents, they also want to gather further information such as contexts,
definitions, or additional semantic information (e.g. semantic relations).
Therefore, it would be very welcome if the providers of TETs would take a step
forward towards the development of such functionalities.
To come back to our initial question, namely
how research and practice are related in the area of TE and if there is any
need to reconcile both, the results of our survey have shown that research and
practice do not completely go separate ways. The existing TETs fulfil the needs
of today's users only partially. Theoretically, TETs based on deeper linguistic
knowledge, including functions like the ones outlined above, can certainly be
developed. We think that a focus on this direction could not only satisfy the
needs and expectations of today's users, but also attract new users.
One possibility to discover the exact needs of
these potential new users of TETs and to bridge the gaps between research,
development, and practice is to encourage and enhance communication between
these groups, e.g. through discussion groups and roundtables, or through
surveys like the one reported in this paper. In this context we want to refer
to the role of terminological networks such as TermNet that help to underline
the importance and relevance of terminology work for language professionals,
customers, public institutions and industries. They also serve to describe the
needs and requirements of language professionals and to make customers, etc.
aware of this. Furthermore, they are also a means for informing both groups
about recent developments and trends in terminology work. In this way all
parties will benefit: First, the industry could develop better targeted tools
and consequently sell more, second, translators, interpreters and
terminologists would benefit from better tools and consequently increase their
productivity. In turn, this is of benefit to the customers.
Finally, we want to underline the importance of
the cooperation triangle between education, research
and practice. Only a close
cooperation among these three areas can guarantee the development of higher quality
TETs. First, education in terminology management is fundamental for better-aimed
research in terminology related issues. Second, the feedback between research
and practice is absolutely indispensable. While research contributes to
differentiate between the possible-doable from the not impossible-not doable,
practice brings in the already accumulated experience and the economical
resources indispensable to determine what is really needed. And third,
education and practice contribute to an exchange of ideas and experiences in
terminology management to adapt to each others' needs and demands. Only if all
these conditions are met can progress towards the optimization of terminology
management issues be achieved.
Acknowledgements
We would
like to thank Uwe Reinke and Karl-Heinz Freigang of the Linguistic Data
Processing Section of Dept. 4.6 at
7. References
Bourigault, D., Jacquemin, C. and L'Homme M-Cl, editors. (2001): Recent Advances in Computational Terminology. John Benjamins,
Cabré Castellvì, M. Teresa; Estopà Bagot, R.; Vivaldi Palatresi, J. (2001): Automatic term detection: A review of current systems. In: Bourigault, D.; Jacquemin, C.; L'Homme, M-Cl, (Editors) (2001), (53-89).
Carstensen, K.-U.; Ebert, C.; Endriss; C.; Jekat, S.; Klabunde, R.;
Langer, H. (Hrsg.) (2001): Computerlinguistik und Sprachtechnologie (Eine
Einführung). Heidelberg, Berlin: Spektrum Akademischer Verlag.
Clematide, S. (2003): Automatische Termextraktion (TE): Utopie, (kommerzielle) Wirklichkeit, nähere Zukunftt. Gastreferat Nachdiplomkurs Terminologie der ZHW.
Last checked: 28.06.05. URL: http://www.cl.unizh.ch/siclemat/talks/termext03/
Daille, B., Habert,
B., Jacquemin, C., and Royauté, J. (1996): Empirical observation of term
variations and principles for their description. Terminology, 3(2),
197-258.
Heid, U. (2001): Verfahren zur Extraktion von Termkandidaten aus
Texten: Ein Überblick. In: Mayer, Felix (Hrsg.): Dolmetschen & Übersetzen
Der Beruf im Europa des 21. Jahrhunderts.
Jacquemin, C.; Bourigault,
D. (2003): Term Extraction and Automatic Indexing. In: Mitkov, R. (editor): The
Oxford Handbook of Computational Linguistics.
Lieske, C. (2002): Pragmatische Evaluierung von Werkzeugen für die
Term-Extraktion. In: eTerminology - Professionelle Terminologie im Zeitalter
des Internet. Akten des Symposiums. Köln, 12.-13. April 2002. Köln: Deutscher Terminologie-Tag e.V. (109-131)
Manning, C.D.;
Schütze, H. (1999): Foundations of statistical natural language processing.
Pearson, J. (1998):
Terms in context.
Saß, R. (2004): Vergleichende Untersuchung von Terminologie-Extraktions-Tools. Eine computerlinguistische Arbeit mit Englisch und Deutsch. Saarbrücken: Fachrichtung 4.6 - Angewandte Sprachwissenschaft sowie Übersetzen und Dolmetschen - Universität des Saarlandes (Saarbrücker Studien zu Sprachdatenverarbeitung und Übersetzen, Band 21)
Sauron, V., 2002: Tearing out the Terms: Evaluating Terms Extractors. Proc. ASLIB 2002, London
Schweizer Bundeskanzlei, Sektion Terminologie / Arbeitsgruppe
Extraktoren (2001): Evaluation der Extraktionsprogramme von Xerox, Trados und
U. Mügge, Oktober 2001. Last
checked 28.9.2005 URL:
http://www.admin.ch/ch/i/bk/termdat/d/tworking/extract/extract.pdf
Streiter, O.;
Zielinski, D.; Ties,
Thurmair, G. (2003):
Making Term Extraction Tools Usable. Comprendium
Warburton, K.
(2001): LISA Terminology Survey Report 2001. Last checked: 28.06.05 URL:
http://www.lisa.org/2001/termsurveyresults.html
Zielinski, D. (2002): Computergestützte Termextraktion aus
technischen Texten (Italienisch), Saarbrücken: Saarland University. [Diploma thesis]. Last checked:
28.06.05. URL: http://fr46.uni-saarland.de/index.php?id=433