A statistical approach to term extraction.pdf

© Servicio de Publicaciones. Universidad de Murcia. All rights reserved. IJES, vol. 11 (2), 2011, pp. 159-182
ISSN: 1578-7044
International Journal
of
English Studies
IJES
UNIVERSITY OF MURCIA www.um.es/ijes
A statistical approach to term extraction
ROGELIO NAZAR*
Instituto Universitario de Lingüística Aplicada
Universidad Pompeu Fabra
Received: 4 June 2011 / Accepted: 3 December 2011
ABSTRACT
This paper argues in favor of a statistical approach to terminology extraction, general to all languages but with
language specific parameters. In contrast to many application-oriented terminology studies, which are focused on
a particular language and domain, this paper adopts some general principles of the statistical properties of terms
and a method to obtain the corresponding language specific parameters. This method is used for the automatic
identification of terminology and is quantitatively evaluated in an empirical study of English medical terms. The
proposal is theoretically and computationally simple and disregards resources such as linguistic or ontological
knowledge. The algorithm learns to identify terms during a training phase where it is shown examples of both
terminological and non-terminological units. With these examples, the algorithm creates a model of the
terminology that accounts for the frequency of lexical, morphological and syntactic elements of the terms in
relation to the non-terminological vocabulary. The model is then used for the later identification of new
terminology in previously unseen text. The comparative evaluation shows that performance is significantly
higher than other well-known systems.
KEYWORDS:
English technical terminology, terminology extraction, computational terminography, quantitative linguistics.
RESUMEN
Este artículo presenta argumentos en favor de una aproximación estadística a la extracción de terminología,
general a todas las lenguas pero con parámetros específicos para cada una de ellas. En contraste con la tendencia
general en terminología aplicada, que suele ser específica de una lengua y un dominio de especialidad, el
presente artículo adopta unos principios generales acerca de las propiedades estadísticas de la terminología
especializada y un método para obtener los parámetros correspondientes a una lengua en particular. Este método
se utiliza para la identificación automática de los términos en los textos, y su efectividad es evaluada en este
artículo mediante un estudio empírico en el caso de la terminología médica en inglés. El modelo requiere escasa
complejidad teórica y computacional, y no necesita recurrir a fuentes de conocimiento lingüístico u ontológico.
Este algoritmo aprende automáticamente a identificar términos durante una fase de entrenamiento en que se
utilizan conjuntos de ejemplos de unidades terminológicas y no terminológicas. Con estos ejemplos, el algoritmo
elabora un modelo de los términos teniendo en cuenta la frecuencia de elementos léxicos, morfológicos y
sintácticos en relación al vocabulario no terminológico. Sobre la base de este modelo, identificará luego nuevos
términos en nuevos textos. El estudio comparativo demuestra que el presente algoritmo tiene un desempeño
significativamente superior al de otros sistemas conocidos.
PALABRAS CLAVE:
Terminología especializada en inglés, extracción de terminología, terminografía computacional, lingüística
cuantitativa.
______________
*Address for correspondence: Rogelio Nazar. Institute for Applied Linguistics. Pompeu Fabra University, Roc
Boronat 138, 08018, Barcelona, Spain. Tel: 00 34 935421187; E-mail: rogelio.nazar@upf.edu

Rogelio Nazar
160
1. INTRODUCTION
Technical terms, or, more broadly speaking, specialized terminology, constitute the portion of
the vocabulary of a language which has a special meaning for the community targeted by the
texts where terms occur. This community shares a common area of interest and a variable
amount of knowledge in particular thematic domains. Abstracted from each particular text,
terms refer to specific concepts in this common knowledge and these concepts are interrelated
configuring a conceptual network necessary for a precise interpretation of the texts. This
paper adopts terms as its object of study and proposes a methodology to separate them from
the rest of the vocabulary of the texts with a comparatively high degree of precision.
As an object of study, terminology has attracted the attention of different professional
groups. Evidently, the most numerous group is formed by translators working with technical
literature. The greatest difficulty of this task is the proper selection of the terminology in the
target language. Most probably, the translator will find terminological units in the source
language which are not yet documented in terminographic resources. This never-ending need
is pressing terminologists to create and update terminographic resources at a very fast pace,
generating in turn the demand for computer-aided terminology management and terminology
extraction from corpus, also called computational terminography (Bourigault et al. 2001).
Parallel to the practical or application-driven research on terminology, terms have been
a topic of interest of national and international normalization organisms, which seek to
provide order and guidelines of proper use. These organisms are informed by a prescriptive
tradition represented by Wüster (1979) and followers (Felber, 1984; Arntz & Picht, 1995 and
others). Terminology is also an object of study of linguistics, and there exist numerous lines
of research presenting diverse theories of terminology as an object of study abstracted from its
language or domain (Cabré, 1999; Gaudin, 1993; Kageura, 1995; Sager, 1990; Temerman,
2000). Linguists working in this area have to be informed on developments in the field of
computational terminography because the techniques used to study corpora can be an
invaluable source of data for a linguistic approach to terminology. Conversely, by elaborating
a general theory of terms, linguists can contribute to the development of better terminology
extraction software. This is the point of view adopted in the present paper, which is intended
to explore a general theory of terms rigorous enough to have a predictive power, that is to say,
a theoretical construct capable of determining how likely it is that a given segment of text is a
terminological unit.
The fact that the method casts the problem of term extraction purely in statistical terms
and disregards external sources of knowledge can be interpreted as the most significant
difference with other approaches that can be categorized broadly as applied linguistics. In the
case of this discipline, the goal is not the advancement of knowledge on a theory of
terminology, but rather the production of high quality raw terminographic material for the
compilation of glossaries, and for that purpose, it is natural to use all sources of information

AStatistical Approach to Term Extraction
161
available. On the contrary, from the point of view of theoretical linguistics, exogenous
sources of knowledge are not considered relevant. There are, however, common interests in
this case for both disciplines, since a term extraction algorithm that does not need abundant
linguistic resources has the advantage of being easily adapted for languages that are less
documented, which is, unfortunately, the current situation of the vast majority of the
languages of the world.
In this context, the present paper explores a mathematical approach that could help us to
formalize the object of study, by formulating a model to separate the technical terms from the
rest of the vocabulary as signal from noise. It is a model based on universal principles, but
includes a methodology to obtain the specific parameters of the language, English in the case
of the empirical study presented in this paper. This mathematical model of terms is built
during a training phase that records the study of examples of both terminological units and a
reference corpus of general language text. Later, in the analysis phase, this model is used to
predict whether certain fragments of specialized texts, being single or multi-word units, are
most likely terms, and this is done taking into consideration the frequency of certain features
in contrast to the reference general language corpus. These features are lexical, morphological
and syntactic patterns, which are easier to collect and process than features such as the
collocational or distributional behavior of the analyzed units. While it is possible to study the
distributional behavior of terms using co-occurrence graphs (see Section 2), these techniques
demand more computational complexity. In comparison with these distributional proposals,
the general approach presented in this paper is very efficient and its performance remains
competitive. The algorithm is easy to implement and fast to execute, analyzing a million
words in a few seconds, which makes the present proposal an ideal approach for web based
applications, where speed of response is a critical variable. In comparison with more simple
statistical algorithms for terminology extraction which are based on frequency, an important
property of the present proposal is that it is not vulnerable to low frequency units. Statistical
analysis, in the present case, includes hapax legomena and dis legomena. This means that the
algorithm reaches high recall and, at the same time, that it can analyze minimal samples of
text, as is the case of the experiment reported here.
The rest of the paper is organized as follows: Section 2 is devoted to the review of
previous research on the subject of automatic terminology extraction. Section 3 will present
the approach adopted in this paper. Section 4 presents the results obtained in an experiment
with English medical terms and in Section 5 these results are evaluated and compared with the
results obtained with different term extraction systems on the same data. Section 6, finally,
draws some conclusions and lines of future work.

Rogelio Nazar
162
2. RELATED WORK
There have been reports on algorithms for terminology extraction for more than forty years, if
we take into account the literature produced in quantitative linguistics and in information
retrieval on the specific topic of « keyword extraction ». An example of such algorithms is the
still very popular TF.IDF (Term frequency – Inverse Document Frequency, Sparck Jones,
1972). This algorithm is used as a measure of term specificity, or in other words, a method to
determine how important or informative a given term is in relation to a particular document
and a particular corpus. In a sufficiently large and diverse corpus –comprising documents of
general and domain specific topics–, the fact that a particular term has a significantly high
frequency of occurrence in a relatively short number of documents can be interpreted as a sign
that this particular term has a specialized meaning. On the contrary, units with a frequency
that shows uniform distribution within the entire corpus can be considered units of the general
language, non-informative or non-interesting from the information retrieval and
terminological perspective. The algorithm does not answer a yes/no question, but it is useful
to rank the vocabulary of a given corpus according to the score each unit obtains, in order to
retain those best positioned and discard the rest. The work of Juilland and Chang (1964)
shows a somewhat similar intuition. In a large corpus organized by thematic domains, those
units which have similar frequency in all partitions of the corpus are likely to be the core
vocabulary of a language. Eliminating this vocabulary from the whole corpus, one is left with
the vocabulary used in each domain.
In the late eighties, with the advancement and wide spread use of technology,
publications in the field of terminology extraction experienced a steady growth, and strategies
diversified into three main lines of research: 1) statistical approaches, which are mainly
language-independent; 2) symbolic or rule-based language-specific approaches, and 3) hybrid
approaches, combining the first two.
Within the statistically-oriented stream, strategies can again be divided into those
focused on « unithood » and those on « termhood » (Kageura & Umino, 1998). The first term
refers to the degree in which multi-word terminological units show strong syntagmatic
association, and the second the degree in which a given term represents a specific concept in a
particular domain. Some authors have applied different measures of association to try to
determine the degree of unithood of multi-word terminological units, that is, applying
frequency based statistics to the single-word components of multi-word terms (Dagan &
Church, 1994; Daille, 1994; Enguehard & Pantera, 1994; Frantzi, 1997; Frantzi et al, 2000;
Pantel & Lin, 2001). A simple extension of this idea, still in the study of the syntagmatics of
multi-word terminology, is to observe how many distinct multi-word terms contain a
particular single term component (Nakagawa & Mori; 2002). If a component is productive,
that is, if it persistently appears as part of a limited number of word n-grams, this could be
interpreted as indicative that these n-grams are terminological.

163
With respect to the degree of termhood, there have been attempts to study it with
distributional statistics elaborating on Sparck Jones' (1972) idea. Aside from assuming that a
term will have high frequency in few documents, one could also expect that the set of
documents where the term occurs will have something in common, some kind of
quantitatively measurable deviation with respect to the average (Hisamitsu et al., 2000). Other
authors have suggested that distributional strategies can be used by comparing the frequency
of term candidates in specialized corpora with their frequency in reference corpora of general
language (Scott, 1997; Böhm et al, 2002; Drouin, 2003; Kilgarriff et al., 2004). Streiter et al
(2002) offer an approach with a philosophy similar to the present paper. These authors report
an experiment of terminology extraction based on examples instead of abstract rules. With the
motivation of working in less documented languages, they describe a supervised algorithm
trained with examples of terms –and not from reference corpora– which learns surface
features such as affix patterns, graphic patterns and length patterns. Affix patterns are, in this
case, the last letter of each component of the multi-word training set terms; graphic patterns
are sequences of initial upper or lower case letters in the same training set and length patterns
are the mean character length of the training set terms plus/minus three standard deviations. In
terms of precision, however, they report modest results.
In recent years, a new statistical approach to both termhood and unithood that has gained
much attention is the application of graphs of lexical co-occurrence. Inspired in graph theory
as used in social sciences for the study of social networks (Barnes & Harary, 1983; Watts &
Strogatz, 1998), these graphs characterize the profile of co-occurrence of a given unit and can
therefore be used to map the meaning (or meanings) of a given term from the textual corpus
and not from hand crafted lexical resources. Empirical research shows that is an effective
method for word sense disambiguation and induction, term extraction and even term
translation (Chatterjee et al. 2010; Madani & Yu, 2010; Matsuo et al, 2006; Nazar, 2010;
Véronis, 2004; Widdows & Dorow, 2002). In this line of research, a graph for each candidate
has to be elaborated from very large corpora such as the web, to afford a minimum recall.
Contrary to the approach of the present article, co-occurrence networks can be used to analyze
units without taking into consideration any aspect of the form of the term, but rather its
distributional behavior only. It can be used, for instance, to tell that a seemingly common
word, such as difference or mouse, can have a specialized use in certain domains, such as
mathematics or computer hardware. Graphs can be used to cluster the contexts of occurrence
of a given unit by spotting attractors or hubs, which are regions in the graph showing high
density and interconnection of nodes. In one of the senses of the term mouse, for instance, one
can see an attractor concentrating terms such as keyword, screen, click, double click, and so
on. As already noticed in the introduction, such models can be interesting from scientific
point of view, but, depending on the particular implementation, may also be computationally
too expensive for certain applications, such as a web based terminology extraction system.

Rogelio Nazar
164
With respect to more linguistically-informed strategies, the vast majority has relied
heavily on syntactic patterns, defining sequences of Part-of-Speech tags that are likely
terminological (Bourigault, 1996; Jacquemin, 2001), searching for Greek and Latin roots in
the candidates (Ananiadou, 1994) or including ontologies (Vivaldi & Rodríguez, 2006) or the
Wikipedia (Vivaldi & Rodríguez, 2011). The latter two can be considered hybrid proposals,
since they combine both statistical and rule based systems. Justeson and Katz (1995) also
combined frequency information with handcrafted lists of noun phrase patterns, under the
assumption that frequent noun phrases in texts are likely terminological. Patry and Langlais
(2005) tried to automatically acquire such patterns directly from examples provided by a user
in a supervised manner. Different combinations of statistics with syntax and semantics can
also be found (Maynard & Ananiadou, 2001; Vivaldi, 2001; Vivaldi & Rodríguez, 2001).
In comparison with the above mentioned studies, and considering the balance between
cost of application and quality of the results, the method reported in Nazar and Cabre (2011)
on experimental terminology extraction in Spanish might be the most simple and effective
solution, and it is this methodology the one that will be replicated for the present experiments
in English (Section 3). This methodology has been recently adopted in the Terminus software
(Cabré & Nazar, forthcoming), which is now available online. Table 1 shows the URL and
references of different term extractors that are capable of analyzing English text and have
online demos available. These tools where used for the evaluation in Section 5. TermFinder
and Yahoo Api are proprietary algorithms. Fivefilters is a simple statistically based term
extractor, meant to be an open source alternative to the Yahoo Api. The rest of the systems
have already been mentioned in this section.
Software Reference URL
TerMine Frantzi et al., 2000 http://www.nactem.ac.uk/software/termine/cgi-bin/termine_cvalue.cgi
TermFinder Translated Labs,
online.
http://labs.translated.net/terminology-extraction/
Yahoo API YAHOO, online. http://developer.yahoo.com/search/content/V1/termExtraction.html
Termostat Drouin, 2003 http://idefix.ling.umontreal.ca/~drouinp/termostat_web/
Fivefilters Minoukadeh, online. http://fivefilters.org/term-extraction/
Terminus Cabré & Nazar,
forthcoming
Now hosted on a provisional URL: http://iula05v.upf.edu
When finished, the official URL will be: http://terminus.upf.edu/
Table 1: Online term-extraction services

165
3. METHODS AND MATERIALS
As already stated in the introduction, the point of departure of the terminology extraction
method presented here is a supervised statistical algorithm, with a training phase where
examples of both terminological and non terminological units are provided by a user. With
this training set, the algorithm will record the frequency of events that are produced on
lexical, morphological and syntactic levels. The idea is that if there are certain events that are
more likely to occur in the set of terms of the training data in comparison with the non-
terminological set, then these events can be considered intrinsic characteristics of the terms.
The following subsections explain the learning process at each of the above mentioned levels
in more detail and how the learned data are used in the analysis of new data. The training
material is the same on all levels: on the one side, a list of at least three thousand validated
terms from the analyzed domain and, on the other side, a reference corpus of general language
of at least two million words.
3.1. Training on the lexical level
The first level of the analysis is simply the frequency count of all single word types in both
corpus. Word types, in this case, are defined as case-insensitive orthographic words, i.e., units
between spaces or punctuation signs. Multi-word terminological units contain, thus, n word
types defined in this way. The result is a matrix with each word type as row and the relative
frequency of the type both in the specialized terms and the reference corpus as columns.
Different columns are created for the frequency of the word forms and of the lemmata.
3.2. Training on the morphological level
The training at the morphological level is very similar to the previous, the only difference
being that, at the morphological level, types are defined in a different manner. While in the
lexical level we count the frequency of words, in this level we count the frequency of
fragments of words, defined as sequences of characters both at the beginning and at the end of
each single word. The length of these segments can be considered an execution parameter.
Previous experiments suggest that appropriate lengths are 3, 4 and 5 letters. The segmentation
of the words is applied to both word forms and lemmata. The format of the matrices is similar
to 3.1, containing the relative frequency of prefixes and suffixes of different length.
It is important to note that it is type frequency and not token frequency which is counted
in this level. Type frequency is a more appropriate measure of the productivity of morphemes
(Baayen & Renouf, 1996), since in this way we count the number of different word types that
contain them. For instance, it would be incorrect to think that the prefix sai- is productive in
general English just because the form said is very frequent.

Rogelio Nazar
166
3.3. Training on the syntactic level
The idea of having the algorithm follow a syntactic training is that it can develop a model of
the syntax of the terminology of the domain. There is an interesting contrast here with respect
to mainstream extractors, as commented in Section 2, since there is no explicit information
about terminological syntactic patterns in English. The algorithm relies on POS-tagging,
however this does not mean that one can no longer claim language independence because, as
shown by Schmid (1994), POS-tagging can also be set up as a language independent problem
using supervised statistical algorithms, as is the case of the present proposal. Something
different is the amount of time that it takes to train a POS-tagger in a new language, but what
remains true is that learning is always based on examples and not on explicit rules.
Under the assumption that one can use a POS-tagger, as is of course the case in English,
the training procedure consists of tagging the example terms. The purpose of the training at
the syntactic level is to elaborate a matrix similar to the previous levels, now recording the
frequency of the syntactic patterns (sequences of POS tags) found in the example terms.
Typically, the most frequent syntactic pattern in English terminology will be the single noun;
followed by adjective+noun or noun+noun, noun+preposition+noun, and so on. Eventually, if
a real user applying this methodology is for some reason interested in excluding some
syntactic pattern, the list of patterns could be manually edited. Naturally, no manual editing of
this type was conducted for the purpose of evaluating this algorithm in the experiment
reported in Section 4. With respect to the syntactic patterns of the reference corpus, they are
considered not relevant and therefore are not recorded.
For technical reasons, one must present each term as a sentence to the tagger. These are
not the best conditions for the application of such devices, but the effects of tagging errors are
not so damaging at this phase. One can safely assume that a majority of the cases will be
tagged correctly, and since the important information are the most frequent patterns, on
average the model will approximate correctly to the actual syntax of the terms.
3.4. Using the training for the analysis of new data
Once the training process has finished, the analysis phase consists of evaluating the candidate
terms found in an analyzed document using the information gathered during the training
phase. If a given word or phrase in the text has a syntactic pattern which is frequently used in
terms, if it contains words frequently used in terms or, moreover, if its morphology is
typically terminological, then the algorithm would have enough clues to promote this unit as a
term candidate. The idea is that the algorithm combines and balances the information learned
during the training phase to rank candidate terms from the analyzed sample.
Methodologically, the first step is to segment the analyzed sample into sequences which
were recorded as frequent patterns in the syntactic training. In contrast to the training phase,

167
errors of the tagger in this phase can cause considerable damage, because at this point the
tagger acts as a filter: any unit that is not encapsulated as potentially terminological from the
syntactic point of view, will now be irremediably lost. If it is indeed the case that the tagger
misses a correct candidate by assigning wrong POS tags, this unit will most probably not be
recovered later, unless it is wrongly assigned a POS tag which is also considered
terminological, in which case the candidate would apply for the next steps (even tagged with a
wrong lexical category).
A resulting matrix records the relative frequency in the analyzed sample of this first
batch of candidates, along with the score provided by its syntactic pattern (defined as the
relative frequency of the pattern in the training set) and the score that the unit obtains in the
lexical and morphological level. The score is given in Equation 1: for each element j, the
algorithm simply takes the observed frequency fo(j) (defined in this case as the relative
frequency of element j in the training set of terms) divided by the expected frequency, fe(j)
(defined as the relative frequency of the same element in the reference corpus) plus a positive
constant s to avoid having an expected frequency with zero value when j does not occur in the
reference corpus.
w(j) = fo(j) / ( fe(j) + s) 1
As said before, a candidate j can be weighted in this way with respect to different
features. On a first run, a candidate term will obtain different scores at the lexical levels for
each of the words it contains, excluding function words. In the case of the lexical coefficient,
at this point it acts as a filter, discarding any candidate below an empirically determined
threshold (0.05), and this takes into account both the inflected form as well as the lemma. In
the case of multi-word expressions, this evaluation is conducted with its first and last
component, because a legitimate multi-word term is likely to have a high frequency word as a
central component but not in the extremes, as it is the case of the component of in the multi-
word term circle of Willis.
On a second run, new values are added in the same way for each of the segments of three
to five letters at the beginning and at the end of each word of the candidate. The final score
(T) of a candidate j is computed shown in Equation 2, where the factors are the observed
frequency of the candidate in the analyzed document –fo(j)–, the syntactic coefficient –syn(j)–
, the lexical coefficient –lex(j)– and, finally, the morphological coefficient –morph(j). Any
candidate resulting in a final score below a threshold (0.5) is automatically discarded but,
naturally, this is an adjustable parameter.
T(j) = fo(j) . syn(j) . lex(j) . morph(j) 2

Rogelio Nazar
168
4. RESULTS
The first step of the experiment has been to gather a reference corpus of general English and a
list of terms from the medical domain. A relatively small reference corpus of English was
used, taken from Quasthoff et al. (2006). This corpus comprises mainly press articles and has
en extension of around two million words. Better results are to be expected when training
with larger corpora. However, for the purpose of the present experiment, it is preferable to test
the performance in conditions of poor training for the sake of replicability in less documented
languages. With respect to the training with examples of terms, again a small amount of
training examples was used, in order to test the reliability of the algorithm in non-ideal
conditions. This was a random sample of 3000 medicine terms from the Mosby dictionary
(Anderson et al., 1998).
4.1. Results of the training phase
4.1.1. Results of training at the level of the lexicon
Tables 2 and 3 show, respectively, part of the result of the frequency count with the 15 more
frequent word forms in the reference corpus and in the term examples (the rest cannot be
shown for obvious space limitations). Function words were eliminated from Table 3 using the
Rank Word Frequency
1 the 118301
2 of 59408
3 to 56152
4 a 47661
5 and 43595
6 in 43184
7 for 20758
8 that 18849
9 is 17231
10 said 14940
11 on 14744
12 was 14172
13 with 11966
14 by 11812
15 from 10770
Table 2: Most frequent words in the general language
reference corpus
Rank Word Frequency
1 syndrome 74
2 disease 54
3 reflex 32
4 system 30
5 blood 26
6 therapy 26
7 health 24
8 hydrochloride 22
9 cancer 22
10 sodium 20
11 fracture 20
12 anesthesia 20
13 pressure 19
14 carcinoma 18
15 artificial 18
Table 3: Most frequent words in a random sample of 3000
medical terms

169
100 most frequent words obtained in Table 2 as a stoplist. Hapax legomena and dis legomena
are excluded from both tables. The lexical training finishes here.
4.1.2. Results of training at the morphology level
As already mentioned in section 3.1.2, the training at the morphology level is very similar to
lexical training, the only difference being that in this case we count the frequency of initial
and final fragments of words of variable length, both from the reference corpus and from the
term examples. Tables 4, 5 and 6 show the frequency of initial and final segments of words of
different length in the reference corpus. Tables 7 and 8, in turn, show the most frequent initial
and final segments in the examples of terms. Only a few examples are shown, but the
calculation is also performed on initial and final segments of 3 and 4 letters.
Rank Segment Frequency
1 comp- 1138
2 comm- 828
3 mill- 764
4 cont- 689
5 stat- 679
6 mark- 671
7 inte- 666
8 mont- 628
9 coun- 624
10 shar- 601
11 cent- 558
12 cons- 548
13 bill- 496
14 pres- 478
15 offi- 465
Table 4: Most frequent 4 letter-long segments at the
beginning of words in the reference corpus
1 -tion 1243
2 -ting 1049
3 -ding 699
4 -ring 623
5 -ther 604
Table 5: Most frequent 4 letter-long segments ending a
word in the reference corpus
1 -ing 5636
2 -ers 1947
3 -ion 1574
4 -ted 1396
5 -ter 1121
Table 6: Most frequent 3 letter-long segments ending a
word in the reference corpus
Observing tables 7 and 8, one finds striking how with only a limited number of terms it is
possible to capture the morphology that is typical of the domain. It is easy to collect affixes
using this technique in comparison with manual compilation. Adding to the difficulty of
manually elaborating a complete list of affixes of the domain, one has to consider that the
effort would have to be repeated with every new domain. Certainly, not all technical domains
show an extensive use of morphology as medical terms, but at least some measurable degree
of deviation can be expected. A definitive answer to this question will need further empirical
research on different domains.

Rogelio Nazar
170
1 trans- 19
2 inter- 16
3 hyper- 15
4 neuro- 13
5 micro- 13
6 psych- 13
7 lymph- 12
8 elect- 11
9 pseud- 11
10 osteo- 9
11 angio- 9
12 intra- 9
13 kerat- 8
14 laryn- 8
15 fibro- 8
Table 7: Most frequent 5 letter-long initial segments in
the sample of 3000 English medical terms
1 -ation 111
2 -ction 29
3 -tosis 24
4 -ology 20
5 -ative 18
6 -genic 15
7 -ional 14
8 -raphy 13
9 -ctomy 13
10 -ility 12
11 -lysis 12
12 -cular 11
13 -atory 11
14 -ement 11
15 -usion 10
Table 8: Most frequent 5 letter-long final segments in
the sample of 3000 English medical terms
4.1.3. Results of training on the syntactic level
The final training on the syntactic level is performed, as explained in 3.1.3, using a POS-
tagger, Schmid's (1994) TreeTagger in this case. What can be seen in Table 9 are the 10 most
frequent syntactic patterns found in the sample of terms that formed the training set.
Rank Syntactic Pattern Frequency
1 NN 1122
2 JJ NN 658
3 NN NN 435
4 JJ 168
5 JJ NN NN 79
6 NN NN NN 64
7 JJ JJ NN 57
8 VV NN 42
9 NN PO NN 32
10 NP PO NN 26
Table 9: The ten most frequent syntactic patterns found in the sample of terms

171
The most frequent pattern, as was to be expected, is the single-noun (NN); the second is the
adjective+noun pattern (JJ NN); the third is noun+noun (NN NN); and so on. Patterns with
frequency below 5 were not registered since their variability increases exponentially and thus
they are not useful for a model.
4.2. Results of the analysis phase
For this experiment, the sample of text to be analyzed is a semi-randomly selected specialized
document, the first pdf document served by Google with a randomly selected English medical
term, which happens to be a document of high degree of specialization. For convenience, only
a small fragment of the document constitutes the sample to be analyzed: just the 285 word
long abstract. From this fragment, and prior to its submission to the system, an observer
manually extracted the totality of the single and multi-word terms that, according to
terminological criterion, could be included as an entry in a medical glossary. This handmade
list will be used later for the evaluation of this and other term extractors (Section 5).
The fragment of text can now be submitted to the extraction algorithm, which will
produce, as a first step, the tokenization and POS-tagging of the text using TreeTagger. Table
10 shows a small fragment of the result of this first step.
TOKEN POS-TAG LEMMA
The DT the
neuroanatomic JJ neuroanatomic
substrate NN substrate
of IN of
cognitive JJ cognitive
deficits NNS deficit
in IN in
long JJ long
term NN term
survivors NNS survivor
of IN of
prematurity NN prematurity
with IN with
PVL NP PVL
is VBZ be
poorly RB poorly
... ... ...
Table 10: Fragment of the analyzed sample after POS-tagging with TreeTagger
Unfortunately, an important number of tagging errors are made by TreeTagger as a
consequence of the high degree of specialization of the document. One has to expect a higher
error rate in this kind of material because of the much higher number of Out-of-Vocabulary
units (OOV's) which tend to be tagged as nouns (NN) by TreeTagger. For instance, the

Rogelio Nazar
172
sequence parieto occipital white matter, which should be tagged as JJ+JJ+JJ+NN (three
adjectives followed by a noun) are instead tagged as NN+NN+JJ+NN, presumably because
the first two components of the term are not in the vocabulary of the tagger. The same occurs
with other sequences, such as pulvinar abnormalities, tagged as NN+NN when it should be
JJ+NN. Certainly, the POS-tagging of specialized corpora remains an open question, and it
could represent an interesting line of research to find the correct tagging when one expects
that a high proportion of the vocabulary will be unknown for the tagger. This may require a
completely different tagging strategy, and the issue is too complex to be addressed in this
paper. At the moment, errors of the tagger will have to be assumed until a better solution is
found.
Rank Candidate Frequency Syntax Lexicon Morphology Total
1 thalamus 6 1.00 1.00 74.71 448.33
2 abnormality 3 1.00 1.88 71.04 400.74
3 atrophy 1 1.00 2.82 135.79 383.09
4 cerebral 1 0.15 11.73 192.10 337.41
5 parieto 2 1.00 1.00 151.27 302.55
6 efferent myelinated 1 0.59 2.82 169.86 281.09
7 parieto occipital 2 0.39 1.00 133.67 103.66
8 method 1 1.00 3.23 30.49 98.41
9 cortex 1 1.00 1.91 47.79 91.31
10 microstructural abnormality 2 0.59 1.44 50.48 85.29
11 efferent 1 0.15 4.64 98.85 68.73
12 abnormal 1 0.15 4.28 80.59 51.64
13 diffusivity 1 1.00 1.00 48.81 48.82
14 neuronal loss 1 0.59 1.05 74.62 45.99
15 lambda 1 1.00 0.99 43.61 43.27
16 neuroanatomic substrate 1 0.59 1.00 64.95 38.09
17 cognition 1 1.00 1.00 34.98 34.99
18 neuronal 1 0.15 1.91 122.11 34.94
19 microstructure 1 1.00 1.00 30.72 30.73
20 pulvinar abnormality 1 0.39 1.80 40.30 28.13
Table 11: The 20 best rated candidates extracted from a 285 word long medical abstract (all syntactic patterns
together)
Once the POS-tagging of the analyzed sample is finished, the algorithm extracts
sequences of units with syntactic patterns which were registered as frequent during the

173
syntactic training phase. With this first batch of term candidates, the algorithm can produce
the values of each candidate using the information gathered during the training phase. The
best 20 candidates are shown in Table 11, irrespective of their syntactic pattern. Optionally,
the candidates can also be presented in a breakdown of patterns, as in the examples shown in
Tables 12, 13 and 14.
A close examination of Table 11 offers a first impression of the quality of the results
and the different weight contributed by each factor. Almost all the 20 best rated candidates are
indeed medical terms, with few exceptions. Possibly, candidates such as parieto (position 5),
efferent myelinated (6) and parieto occipital (7) might seem to be segmentation errors, as they
are parts of a term acting as modifiers instead of being well-formed phrases with a head.
Adjectives and modifiers represent frequent patterns within the set of training examples of
terms. The reason for including units with these syntactic patterns in a terminological
dictionary is that they are relevant from a terminological perspective. These elements are
recorded to account for their occurrence in different terms modifying different types of heads,
which is the purpose of the breakdown in syntactic patterns that will follow, where there are
lists of, for instance, adjectives, which are found as part of different terms. As mentioned in
Section 3.3, if in a real application a terminologist applying this method is not interested in
such patterns, the list of patterns that feeds the algorithm could eventually be edited after the
training. The relation between modifiers and heads, as the number and proportion of heads per
modifier or vice-versa, is still an open question and a very interesting line of research, similar
to Dagan and Church (1994) or Nakagawa and Mori (2002).
A candidate like method, in position 8, is instead terminologically less interesting, and
therefore in this experiment it is considered a false positive, even when the Mosby dictionary
reserves an entry for this unit. Other units that received high ranking by the algorithm are
statistical terms such as lambda, p or positive correlation. It may be disputable if these units
can be considered medical terms, since they are more properly terms from statistics. However,
they are frequently used in medical literature. Should they be included in a terminological
dictionary? Despite the statistical origin of these terms, most terminologists would agree. The
terms lambda, p-value, and correlation are also listed as entries in the Mosby dictionary.
With respect to the breakdown of the candidates in syntactic patterns, only a few examples are
shown. The classification is useful for practical purposes, but for the theoretical purpose of
this article, this information is of secondary importance. More important is the precision of
the selection. The cases of errors that are found are at times caused by previous tagging errors,
as is the case with the previously mentioned parieto tagged as a noun. Candidates such as
child, volume, reduction, effect or patient are less relevant for a terminologist and in this
experiment they are considered false positives as well, irrespective of the fact that, again, the
Mosby dictionary also includes these five examples as entries. Other clear false positives are,
for instance in the Adjective+Noun pattern, segmentation errors (efferent myelinated) and
genuine scoring errors (long term, significant reduction, overall volume, major finding).

Rogelio Nazar
174
Rank Candidate Frequency Lexicon Morphology Total
1 thalamus 6 1.00 74.71 448.33
2 abnormality 3 1.88 71.04 400.74
3 atrophy 1 2.82 135.79 383.09
4 pulvinar 1 1.50 45.36 10.19
5 parieto 2 1.00 151.27 302.55
6 myelinated 1 1.50 140.51 4.88
7 method 1 3.23 30.49 98.41
8 cortex 1 1.91 47.79 91.31
9 diffusivity 1 1.00 48.81 48.82
10 lambda 1 0.99 43.61 43.27
11 cognition 1 1.00 34.98 34.99
12 microstructure 1 1.00 30.72 30.73
13 child 3 1.30 7.14 27.95
14 injury 1 3.06 8.01 24.53
15 term 2 0.42 25.85 21.84
Table 12: The 15 best rated nouns
1 efferent myelinated 1 2.82 169.86 281.09
2 microstructural abnormality 2 1.44 50.48 85.29
3 pulvinar abnormality 1 1.80 40.30 28.13
4 neuronal loss 1 1.05 74.62 45.99
5 neuroanatomic substrate 1 1.00 64.95 38.09
6 secondary effect 1 2.88 8.32 14.09
7 cognitive deficit 1 0.79 21.85 10.06
8 microstructural damage 1 0.70 17.37 7.13
9 visual processing 1 0.79 14.46 6.72
10 positive correlation 1 1.19 8.85 6.17
11 preterm child 1 1.15 7.80 5.28
12 long term 1 0.28 25.12 4.12
13 extensive interconnection 1 0.85 7.95 3.95
14 significant reduction 1 0.79 6.75 3.13
15 overall volume 1 0.97 5.27 3.00
Table 13: The 15 best candidates with the pattern Adjective + Noun

175
1 cerebral 1 11.73 192.10 337.41
2 efferent 1 4.64 98.85 68.73
3 abnormal 1 4.28 80.59 51.64
4 neuronal 1 1.91 122.11 34.94
5 axonal 1 1.00 134.50 20.14
6 neuroanatomic 1 1.00 122.95 18.41
7 atrophic 1 1.50 64.06 2.06
8 cognitive 2 0.99 34.82 10.35
9 pulvinar 1 1.50 45.36 10.19
10 periventricular 1 1.50 37.62 1.21
11 afferent 1 1.00 52.09 7.80
12 secondary 1 4.46 9.00 6.02
13 quantitative 1 2.73 13.97 5.71
14 microstructural 1 1.00 29.93 4.48
15 visual 1 0.91 26.81 3.65
Table 14: The 15 best rated Adjectives.
5. EVALUATION AND COMPARISON WITH OTHER SYSTEMS
The evaluation of a new methodology needs more than simply to run an experiment and
record precision and recall. The evaluation has to be carried out with reference to other
parameters, usually a baseline algorithm representing a classic or basic solution to the same
problem and, when possible, a direct comparison with the results of other algorithms on the
same data set. The manual extraction of the terms from the sample conducted at the first
phases of the experiment (Section 4.2) is now used as a gold-standard for the comparison with
the selection of terms proposed by each algorithm. The systems listed in Table 1 of Section 2
plus a baseline algorithm that will be described below, are used to extract term candidates
from the same sample of text.
The manual extraction from the sample resulted in 49 different terminological units.
With this list, one can calculate precision as the proportion of candidates in the ranked list
proposed by each algorithm which are also in the list of manually extracted terms; as well as
recall, as the proportion of manually extracted terms found also in the ranked list proposed by
each algorithm. This evaluation method can possibly be criticized for relying on the judgment
of a single observer, when a specialist –or, better, a group of them– could do a more precise
selection of the terminology. However, it is to be expected that a single observer will be able

Rogelio Nazar
176
to identify at least the majority of the terms correctly, which should be enough to evaluate a
term extractor.
The first step of the evaluation is the comparison with the baseline algorithm, which is
the frequency rank of case-insensitive n-grams from the sample, undoubtedly the most simple
method for extracting terminology from text. In this case, the baseline includes unigrams,
bigrams and trigrams, and applies a stoplist consisting of the 100 most frequent words in the
reference corpus. Any n-gram including a member of the stoplist at the initial or final position
is eliminated from the ranking. Hapax legomena and n-grams with punctuation signs are also
filtered out. In general, these kinds of strategies tend to have low recall since, following a
Zipfean distribution, a large proportion of the terminological units of a corpus have low
frequency.
Using the ranking produced by the algorithm described in this paper and the one
produced by the baseline algorithm, it is possible to plot Figure 1, with precision at the
vertical axis and recall at the horizontal axis. Ideally, an algorithm should produce both high
precision and recall, thus it can be interpreted that the surface behind the curve is a measure of
the quality of the output. Normally, however, an algorithm will not produce 100% recall.
Some systems are prone to offer few candidates which are most surely terminological, rather
than offering many less reliable units. It is, in any case, a matter of balance between noise and
silence. As already mentioned, silence or low recall in the case of the baseline algorithm is a
consequence of Zipf law. Depending on the size, one would expect that between a third and
close to a half of the terms in a specialized document will be hapax legomena.
Figure 1: Precision and recall plot comparing the results of this experiment with the
baseline algorithm

177
In the case of the experiment reported in the present paper, one can see that the
algorithm is able to maintain 75% precision at 80% recall, which is a good balance
considering the human effort necessary to post-process the ranked lists. In comparison,
FIVEFILTERS (Figure 2) performs significantly better than the baseline, but it is still closer
to the baseline than to the performance achieved in this experiment. YAHOO, in turn, shows
better precision than FIVEFILTERS but, comparatively, exhibits lower performance than the
results reported here (Figure 3). TERMFINDER (Figure 4) is between FIVEFILTERS and
YAHOO. TERMINE (Figure 5) is, from all the evaluated algorithms, the one that has the best
performance in precision if we take into consideration only the segment of best rated
candidates. Precision in TERMINE decreases rapidly, however, in relation to recall. Recall, in
any case, is below 40%. TERMOSTAT (Figure 6) follows the baseline, which is possibly a
consequence of being a statistical approach not designed for working with samples of reduced
size.
Considering the balance between precision and recall, the results reported in this paper
are clearly superior to those produced by the baseline and the rest of the algorithms, and this
is said only taking into account the results produced by each system. Other parameters for
evaluation could have also been included, such as the time and effort required to implement
each solution for each language and domain, especially when they are based on linguistic and
ontological knowledge. In the case of this experiment, the whole process (training + analysis)
took only a few minutes. This is assuming that a POS-tagger is already trained for that
language and that the rest of the training material (the list of terms and the reference corpus) is
available. Fortunately, in the present days it is easy to find this material on the web.
Figure 2: Precision and recall plot in comparison to FIVEFILTERS

Rogelio Nazar
178
Figure 6: Precision and recall plot in comparison with
TERMOSTAT
6. CONCLUSIONS AND FUTURE WORK
This paper has reported the application of a language and domain independent statistical
methodology for terminology extraction in the particular case of English medical terms.
Leaving aside the savings in effort of manually coding linguistic or specialized knowledge,
the fact that the method does not rely on external sources has two advantages: 1) it can be
used for other, less documented languages and domains; 2) the automatic recollection of the
needed (implicit) knowledge avoids the risk of subjective bias that can be found in knowledge
coded in the form of symbolic rules. The statistical inference performed by this algorithm
seems, in contrast to rule-based algorithms, more objective and systematic. Other positive
attributes of the proposal is that it is computationally efficient and inexpensive and that it is
also easy to understand and implement.
TERMINE
TERMFINDER
YAHOO

179
A question that remains open for future versions of this algorithm is whether to develop a
fully language independent POS-tagger. The current problem with using TreeTagger is that a
large sample of annotated data is needed to train the program in a new language and, in
practice, this is difficult to produce in a short period of time. Unsupervised methods, however,
could be explored. There have been reports on solutions to the problem of inferring the lexical
categories without previous knowledge (Biemann, 2010). In this scenario, there is no need to
tag the words with a proper lexical category name. Instead, lexical categories are identified by
arbitrary codes -i.e., numbers- as a result of a process of clustering of words based on
distributional properties. If an attempt in this line proves successful, it would considerably
reduce the effort of the user supervising the process.
As for future work, the most immediate task would be to improve on some technical
issues of the experiments, apart from the strictly mathematical aspect. A useful
implementation needs to meet diverse problems such as to develop a careful text-handling
procedure, preserving the integrity of the texts after they have been converted from different
file formats to plain text, guessing character encodings and identifying titles, paragraphs,
proper nouns and so on.
Better evaluation figures could be obtained finding new term extractors to repeat the
experiments, preferably with larger samples of text. A group of terminologists and specialists
in the domain should also be part of the experiment, instead of a single observer doing the
manual extraction. In this new experimental design, one can also check agreement statistics
among the observers, including a group of random extractors as reference. Clustering the
observers according to their selection in a dendrogram, one should see, ideally, that specialists
and terminologists are clustered together with those that can be considered good term
extraction systems, while those less reliable should lie apart from them, closer to the random
extractors. This could be, possibly, the most objective measure for the evaluation of term
extraction algorithms.
ACKNOWLEDGEMENTS
This research was funded by a contract as technician by the Spanish Ministry of Science and Innovation in the
Institute for Applied Linguistics of Pompeu Fabra University (Barcelona), in the scope of Project RICOTERM-
3. « Enriquecimiento de ontologías para la extracción de información de discursos de especialidad », lead by
Mercè Lorente (HUM2007-65966-C02-00). The author would also like to thank the anonymous reviewers for
their comments.
REFERENCES
Ananiadou, S. (1994). A Methodology for Automatic Term Recognition. Proceedings of
Coling 1994, 15th International Conference on Computational Linguistics (pp. 1034-
1038). Kyoto, Japan.

Rogelio Nazar
180
Anderson, K. N., Anderson, L. E. & Glanze, W. D. (1998). Mosby’s medical, nursing, and
allied health dictionary (5th ed.). St. Louis, MO: Mosby.
Arntz, R., Picht, H. (1995). Introducción a la terminología, Fundación Germán Sánchez
Ruipérez. Madrid: Pirámide.
Baayen, H., Renouf, A. (1996) Chronicling the Times: Productive Lexical Innovations in an
English Newspaper. Language, 72(1), 69–96.
Barnes, J.A. & Harary, F. (1983). Graph Theory in Network Analysis. Social Networks, 5,
235–244.
Biemann, C. (2010). Unsupervised Part-of-Speech Tagging in the Large. Research on
Language and Computation, 7(2), 101–135.
Böhm, K., Heyer, G., Quasthoff, U. & Wolff, C. (2002). Topic Map Generation Using Text
Mining. Journal of Universal Computer Science, 8(6), 623–633.
Bourigault, D. (1996). LEXTER, a Natural Language tool for terminology extraction.
Proceedings of Seventh EURALEX International Congress (pp. 771-779). Göteborg,
Sweden.
Bourigault, D., Jacquemin, C. & L'Homme, M-C. (Eds.) (2001). Recent Advances in
Computational Terminology. Amsterdam: John Benjamins.
Cabré, M. T. (1999). La Terminología: Representación y Comunicación. Barcelona: Institut
Universitari de Lingüística Aplicada.
Cabré, M. T. & Nazar, R. (forthcoming). Terminus: a Workstation for terminology and corpus
management. Proceedings of TOTH 2011 Conference. Annecy, France.
Chatterjee, D., Sarkar, S., Mishra, A. (2010). Co-occurrence Graph Based Iterative Bilingual
Lexicon Extraction From Comparable Corpora. Proceedings of the 4th Workshop on
Cross Lingual Information Access, COLING 2010 workshop (pp. 35-42). Beijing,
China.
Dagan, I., Church, K. (1994) Termight: identifying and translating technical terminology.
ANLC '94 Proceedings of the fourth conference on Applied natural language processing
(pp. 34-40). Association for Computational Linguistics.
Daille, B. (1994). Approche mixte pour l'extraction automatique de terminologie : statistiques
lexicales et filtres linguistiques. Doctoral Dissertation. Université Paris 7.
Drouin, P. (2003). Term extraction using non-technical corpora as a point of leverage.
Terminology, 9(1), 99–117.
Enguehard, C. & Pantera, L. (1994). Automatic Natural Acquisition of a Terminology. Journal
of Quantitative Linguistics, 2(1), 27–32.
Felber, H. (1984). Terminology Manual. Paris: Unesco/Infoterm.
Frantzi, K.T. (1997). Incorporating context information for extraction of terms. Proceedings
of the Association for Computational Linguistics (ACL/EACL) (pp. 501-503). Madrid,
Spain.
Frantzi, K., Ananiadou, S. & Mima, H. (2000). Automatic recognition of multi-word terms.
International Journal of Digital Libraries, 3(2), 117– 132.
Gaudin, F. (1993). Pour une socioterminologie: Des problèmes pratiques aux pratiques
institutionnelles. Université de Rouen.
Hisamitsu T., Niwa, Y, Nishioka, S., Sakurai, H., Imaichi, O., Iwayama, M. & Takano, A.
(2000). Term Extraction Using A New Measure of Term Representativeness.
Proceedings of the Second International Conference on Language Recources and
Evaluation (LREC 2000). Workshop Proceedings on: Terminology Resources and
Computation (pp. 13–20). Athens, Greece.
Jacquemin, C. (2001) Spotting and Discovering Terms through Natural Language Processing.
MIT Press.

181
Juilland, A., Chang-Rodríguez, E. (1964). Frequency Dictionary of Spanish W
ords. The
Hague, Mouton.
Justeson J., Katz, S. (1995). Technical terminology: some linguistic properties and an
algorithm for identification in text. Natural Language Engineering, 1(1), 9–27.
Kageura, K. (1995) Toward the theoretical study of terms. Terminology, 2(2), 239–257.
Kageura, K. & Umino, B. (1998). Methods of Automatic Term Recognition. Terminology,
3(2), 259–289.
Kilgarriff, A., Rychly, P., Smrz, P., Tugwell, D. (2004). The Sketch Engine. Proceedings of
Euralex (pp. 105-116). Lorient, France.
Madani, O. & Yu, J. (2010). Discovery of numerous specific topics via term co-occurrence
analysis. Proceedings of the 19th ACM international conference on Information and
knowledge management (CIKM '10) (pp. 1841-1844). ACM, New York, NY, USA.
Matsuo, Y., Sakaki, T., Uchiyama, K. & Ishizuka, M. (2006). Graph-based word clustering
using a web search engine. Proceedings of the 2006 Conference on Empirical Methods
in Natural Language Processing (EMNLP '06). Association for Computational
Linguistics (pp. 542-550). Stroudsburg, PA, USA.
Maynard, D. & Ananiadou, S. (2001). TRUCKS: a model for automatic term recognition.
Journal of Natural Language Processing, 8(1), 101–125.
Minoukadeh, K. (online). FIVEFILTERS.org. http://fivefilters.org/term-extraction/
Nakagawa, H., Mori, T. (2002). A Simple but Powerful Automatic Term Extraction Method.
Proceedings of the second International Workshop on Computational Terminology
(COMPUTERM 02) (pp. 29–35). Morristown, USA.
Nazar, R. (2010). A Quantitative Approach to Concept Analysis. Doctoral Dissertation,
Universitat Pompeu Fabra.
Nazar, R. & Cabré, M.T. (2011) Un experimento de extracción de terminología utilizando
algoritmos estadísticos supervisados. Revista Debate Terminológico, 7, 36–55.
Pantel, P. & Lin, D. (2001). A Statistical Corpus-Based Term Extractor. Proceedings of the
14th Biennial. Conference of the Canadian Society on Computational Studies of
Intelligence (pp. 36-46). London, UK: Springer-Verlag.
Patry, A. & Langlais, P. (2005). Corpus-Based Terminology Extraction. Proceedings of the 7th
International Conference on Terminology and Knowledge Engineering (pp. 313-321),
Copenhagen, Danemark.
Quasthoff, U., Richter, M. & Biemann, C. (2006). Corpus portal for search in monolingual
corpora. Proceedings of the LREC 2006 (pp. 1799-1802). Genoa, Italy.
Sager, J. (1990). A Practical Course in Terminology Processing. Amsterdam/Philadelphia:
John Benjamins.
Schmid, H. (1994). Probabilistic Part-of-Speech Tagging Using Decision Trees. Proceedings
of the International Conference on New Methods in Language Processing (pp. 44-49).
Manchester, UK.
Scott, M., (1997). PC Analysis of Key Words and Key Key Words. System, 25(1), 1–13.
Sparck Jones, K. (1972). A Statistical Interpretation of Term Specificity and its Application in
Retrieval. Journal of Documentation, 28(1), 11–21.
Streiter, O., Zielinski, D., Ties, I. & Voltmer, L. (2002) Example-based Term Extraction for
Minority Languages: A case-study on Ladin. Paper presented at Sociolinguistics and
Language Planning, Ortisei, December 12-14.
Temmerman, R. (2000). Towards New Ways of Terminology Description: the Socio-Cognitive
Approach. Amsterdam: John Benjamins.
Véronis, J. (2004). HyperLex: Lexical Cartography for Information Retrieval. Computer
Speech & Language, 18(3), 223–252.

Rogelio Nazar
182
Vivaldi, J. (2001/2004). Extracción de Candidatos a Término Mediante Combinación de
Estrategias Heterogéneas, Barcelona: IULA, Sèrie Tesis 9.
Vivaldi, J.; Rodríguez, H. (2001). Improving term extraction by combining different
techniques. Terminology, 7(1), 31–48.
Vivaldi, J. & Rodríguez, H. (2011). Extracting terminology from Wikipedia. Procesamiento
del lenguaje natural, 47, 65–73.
Watts, D. J., Strogatz, S. (1998). Collective dynamics of 'small-world' networks. Nature, 393
(6684), 440–442.
Widdows, D. & Dorow, B. (2002). A Graph Model for Unsupervised Lexical Acquisition.
Proceedings of the 19th International Conference on Computational Linguistics (pp.
1093-1099), Taipei.
Wüster, E. (1979). Introduction to the General Theory of Terminology and Terminological
Lexicography. Vienna: Springer.

A statistical approach to term extraction.pdf

More Related Content

Similar to A statistical approach to term extraction.pdf (20)

More from Jasmine Dixon (20)

Recently uploaded (20)

A statistical approach to term extraction.pdf