Using Domain-Specific Verbs for Term Classification 
 
Irena Spasić 
Computer Science 
University of Salford, UK 
I.Spasic@salford.ac.uk
Goran Nenadić 
Department of Computing 
UMIST, UK 
G.Nenadic@umist.ac.uk
Sophia Ananiadou 
Computer Science  
University of Salford, UK 
S.Ananiadou@salford.ac.uk
 
Abstract 
In this paper we present an approach to 
term classification based on verb com-
plementation patterns. The complementa-
tion patterns have been automatically 
learnt by combining information found in 
a corpus and an ontology, both belonging 
to the biomedical domain. The learning 
process is unsupervised and has been im-
plemented as an iterative reasoning pro-
cedure based on a partial order relation 
induced by the domain-specific ontology. 
First, term recognition was performed by 
both looking up the dictionary of terms 
listed in the ontology and applying the 
C/NC-value method. Subsequently, do-
main-specific verbs were automatically 
identified in the corpus. Finally, the 
classes of terms typically selected as ar-
guments for the considered verbs were in-
duced from the corpus and the ontology. 
This information was used to classify 
newly recognised terms. The precision of 
the classification method reached 64%. 
1 Introduction 
Basic notions used when describing a specific 
problem domain are concepts, classes and attrib-
utes (or features). The identification of concepts, 
linguistically represented by domain-specific terms 
(Maynard and Ananiadou, 2000), is a basic step in 
the automated acquisition of knowledge from tex-
tual documents. Textual documents describing new 
knowledge in an intensively expanding domain are 
swamped by new terms representing newly identi-
fied or created concepts. Dynamic domains, such 
as biomedicine, cannot be represented by static 
models, since new discoveries give rise to the ap-
pearance of new terms. This makes the automatic 
term recognition (ATR) tools essential assets for 
efficient knowledge acquisition.  
However, ATR itself is not sufficient when it 
comes to organizing newly acquired knowledge. 
Concepts are natively assorted into groups and a 
well-formed model of a domain, represented 
through terms and their relations, needs to reflect 
this property consistently. Dynamic domain mod-
els should be able to adapt to the advent of new 
terms representing newly discovered or identified 
concepts. In other words, newly extracted terms 
need to be incorporated into an existing model by 
associating them with one another and with already 
established terms preferably in an automated man-
ner. This goal may be achieved by relying on term 
clustering (the process of linking semantically 
similar terms together) and term classification (the 
process of assigning terms to classes from a pre-
defined classification scheme). In particular, classi-
fication results can be used for efficient and consis-
tent term management through populating and 
updating existing ontologies in expanding domains 
such as biomedicine. In this paper, we compare 
some of the term classification approaches and in-
troduce another approach to this problem.  
The paper is organised as follows. In Section 2 
we provide a brief overview of the existing term 
classification approaches and suggest the main idea 
of our approach to this problem. Section 3 de-
scribes the learning phase of our classification 
method. Further, Section 4 provides details on the 
classification algorithm. Finally, in Section 5 we 
describe the evaluation strategy and provide the 
results, after which we conclude the paper. 
2 Term Classification Approaches 
Similarly to general classification algorithms, the 
existing term classification approaches typically 
rely on learning techniques. These techniques are 
most often statistically based (e.g. hidden Markov 
models, naive Bayesian learning, etc.). Other tech-
niques include decision trees, inductive rule learn-
ing, support-vector machines (SVMs), etc. We, on 
the other hand, suggest the use of a genetic algo-
rithm as a learning engine for the classification 
task. Let us now discuss some approaches to the 
automatic classification of biomedical terms. 
Nobata et al. (2000) implemented a statistical 
method for term classification. In their approach, 
each class was represented by a list of (single) 
words. The first step was to estimate the condi-
tional probability P(c | w) of each word w being 
assigned to a specific class c, based on the assump-
tion that each word occurrence is independent of 
its context and position in the text. Further, yet an-
other strong restriction was made by assuming that 
there was one-to-one correspondence between 
terms and their classes. In addition, this approach 
is not applicable to “unknown” terms, i.e. terms 
containing words for which no classification prob-
abilities had been determined. A special class, re-
ferring to “other”, was introduced to cover such 
words. Bearing in mind the increasing number of 
new terms, such an approach is bound to produce 
skewed results, where many of the terms would 
simply be classified as “other”. 
While Nobata et al. (2000) statistically proc-
essed the information found inside the terms, Col-
lier et al. (2001) applied statistical techniques to 
the information found outside the terms. A hidden 
Markov model based on n-grams (assuming that a 
term’s class may be induced from the previous n-1 
lexical items and their classes) was used as a theo-
retical basis for their classification method. The 
method relied on the orthographic features includ-
ing numerals, capital and Greek letters, special 
characters (such as `-`, `/`, `+`, etc.), parenthesis, 
etc. In the biomedical domain, such features often 
provide hints regarding the class of a specific term. 
Each unclassified term was assigned a class of the 
most similar (with respect to the orthographic fea-
tures) term from the training set. This approach 
encountered the minority class prediction problem. 
Namely, the best classification results in terms of 
recall and precision were achieved for the most 
frequent class of terms in their training corpus, 
while the worst results were those achieved for the 
least frequent class. 
Hatzivassiloglou et al. (2001) proposed a 
method for unsupervised learning of weights for 
context elements (including words as context con-
stituents and the corresponding positional and 
morphological information) of known terms and 
using these weights for term classification. Three 
well-known learning techniques were used: naive 
Bayesian learning, decision trees, and inductive 
rule learning. Simplified classification experiments 
in which a classification algorithm was choosing 
between two or three options respectively were 
conducted. The precision of binary classification 
was around 76% for all three learning algorithms, 
and the precision dropped to approximately 67% 
when choosing between three options. If the pro-
posed techniques were to be applied for general 
classification where the number of options is arbi-
trary, the precision is expected to decrease even 
further. 
Nenadic et al. (2003b) conducted a series of 
large-scale experiments with different types of fea-
tures for a multi-class SVM. These features in-
cluded document identifiers, single words, their 
lemmas and stems, and automatically recognised 
terms. The results indicated that the performance 
was approximately the same (around 60% in the 
best case) when using single words, lemmas or 
stems. On the other side, terms proved to be better 
(more than 90% precision) than single words at 
lower recall points (less than 10%), which means 
that terms as features can improve the precision for 
minority classes. The best results were achieved 
with document identifiers, but such features cannot 
be used on the fly in new documents.  
Spasic et al. (2002) used a genetic algorithm 
(GA) based on a specific crossover operator to ex-
plore the relationships between verbs and the terms 
complementing them. The GA performed reason-
ing about term classes allowed to be combined 
with specific verbs by using an existing ontology 
as a seed for learning. In this paper, we use the re-
sults of the proposed methodology as a platform 
for term classification. In the following section we 
briefly overview the method for the acquisition of 
verb complementation patterns. 
3 Verb Complementation Patterns 
By looking at the context of an isolated verb occur-
rence it is difficult to predict all term classes that 
can be combined with the given verb. On the other 
hand, the whole “population” of terms comple-
menting a specific verb is likely to provide a cer-
tain conclusion about that verb with respect to its 
complementation patterns. This was a primary mo-
tivation for Spasic et al. (2002) to use a GA as it 
operates on a population of individuals as opposed 
to a single individual. This fact also makes the ap-
proach robust, since it does not rely on every spe-
cific instance of verb-term combination to be 
correctly recognised.  
As not all verbs are equally important for the 
term classification task, we are primarily interested 
in domain-specific verb complementation patterns. 
In our approach, a complementation pattern of a 
domain-specific verb is defined as a disjunction of 
terms and/or their classes that are used in combina-
tion with the given verb. The automatic acquisition 
of these patterns is performed in the following 
steps: term recognition, domain-specific verb ex-
traction, and the learning of complementation pat-
terns. Let us describe each of these steps in more 
detail. 
 
3.1   Term Recognition 
 
First, a corpus is terminologically processed: both 
terms present in the ontology and the terms recog-
nised automatically are tagged. Terms already 
classified in the ontology are used to learn the 
classes allowed by the domain-specific verbs, 
while the new terms are yet to be classified based 
on the learnt classes. New terms are recognized by 
the C/NC-value method (Frantzi et al., 2000), 
which extracts multi-word terms. This method rec-
ognises terms by combining linguistic knowledge 
and statistical analysis. Linguistic knowledge is 
used to propose term candidates through general 
term formation patterns. Each term candidate t is 
then quantified by its termhood C-value(t) calcu-
lated as a combination of its numerical characteris-
tics: length |t| as the number of words, absolute 
frequency f(t) and two types of frequency relative 
to the set S(t) of candidate terms containing a 
nested candidate term t (frequency of occurrence 
nested inside other candidate terms and the number 
of different term candidates containing a nested 
candidate term): 
 





∅≠−⋅
∅=⋅
=−
∑
∈
)( if  ,))(
|)(|
1
)((||ln
)( if  ),(||ln
)(
)(
tSsf
tS
tft
tStft
tvalueC
tSs
 
 
Obviously, the higher the frequency of a candi-
date term the greater its termhood. The same holds 
for its length. On the other side, the more fre-
quently the candidate term is nested in other term 
candidates, the more its termhood is reduced. 
However, this reduction decreases with the in-
crease in the number of different host candidate 
terms as it is hypothesised that the candidate term 
is more independent if the set of its host terms is 
more versatile. 
Term distribution in top-ranked candidate terms 
is further improved by taking into account their 
context. The relevant context words, including 
nouns, verbs and adjectives, are extracted and as-
signed weights based on how frequently they co-
occur with top-ranked term candidates. Subse-
quently, context factors are assigned to candidate 
terms according to their co-occurrence with top-
ranked context words. Finally, new termhood esti-
mations (NC-values) are calculated as a linear 
combination of the C-values and context factors.  
Nenadic et al. (2003a) modified the C/NC-value 
to recognise acronyms as a special type of single-
word terms, and, thus, enhanced the recall of the 
method. On the other hand, the modified version 
incorporates the unification of term variants into 
the linguistic part of the method, which also im-
proved the precision, since the statistical analysis is 
more reliable when performed over classes of 
equivalent term variants instead of separate terms. 
 
3.2   Domain-Specific Verb Recognition 
 
Verbs are extracted from the corpus and ranked 
based on the frequency of occurrence and the fre-
quency of their co-occurrence with terms. A stop 
list of general verbs frequently mentioned in scien-
tific papers independently of the domain (e.g. ob-
serve, explain, etc.) was used to filter out such 
verbs. The top ranked verbs are selected and 
considered to be domain-specific. Moreover, these 
verbs are also corpus-specific (e.g. activate, 
bind, etc.). Table 3 provides a list of such verbs, 
which were used in the experiments. 
 
3.3   Complementation Pattern Learning 
 
In order to learn a verb complementation pattern 
for each of the selected verbs separately, terms are 
collected from the corpus by using these verbs as 
anchors. A GA has been implemented as an itera-
tive reasoning procedure based on a partial order 
relation induced by the domain-specific ontology.
1
 
In each iteration pairs of verb complementation 
patterns represented as sets of terms and term 
classes are merged. This operation involves the 
substitution of less general terms/classes by their 
more general counterparts, if there is a path in the 
ontology connecting them. Otherwise, the disjunc-
tion of the terms is formed and passed to the next 
iteration. Figure 1 depicts the process of learning a 
verb complementation pattern. 
Since the partial order relation induced by the 
ontology is transitive, the order in which terms are 
processed is of no importance. The final verb com-
plementation patterns are minimal in the sense that 
the number of terms in a verb complementation 
pattern and the depth of each individual term in the 
ontology are minimised. 
 
 
Figure 1. Learning the complementation pattern 
                    for the verb bind 
4 Term Classification Method 
The verb complementation patterns have been ob-
tained by running the GA on a set of terms some of 
which were present in an ontology, which is used 
                                                           
1
 The partial order relation is based on the hierarchy of 
terms/classes: term/class t
1
 is in relation with t
2
, if there is a 
path in the ontology from t
2
 to t
1
. In that case, we say that t
2  
is 
more general than t
1
.
 
during the learning process. The newly recognised 
terms (i.e. the ones not found in the ontology) will 
remain included in the final verb complementation 
patterns as non-classified terms, since at this point 
it is not known which classes could replace them. 
All elements of the final verb complementation 
patterns can be thus divided into two groups based 
on the criterion of their (non)existence in the on-
tology. The elements already present in the ontol-
ogy are candidate classes for the newly recognised 
terms. Let us now describe the classification 
method in more detail. 
Let V = {v
1
, v
2
, ... , v
n
} be a set of automatically 
identified domain-specific verbs.
 
During the phase 
of learning verb complementation patterns, each of 
these verbs is associated with a set of classes and 
terms it co-occurs with. Let C
i
 = {c
i,1
, c
i,2
, ... , c
i,m
i
} 
denote a set of classes assigned automatically to 
the verb v
i
 (1 ≤ i ≤ n) by a learning algorithm based 
on the information found in the corpus and the 
training ontology. As indicated earlier, we define 
such set to be a verb complementation pattern for 
the given verb.  
 
4.1   Statistical Analysis 
 
As we planned to use verb complementation pat-
terns for term classification, we modified the origi-
nal learning algorithm (Spasic et al., 2002) by 
attaching the frequency information to terms and 
their classes. When substituting a less general class 
by its more general counterpart,
2
 the frequency 
information is updated by summing the two 
respective frequencies of occurrence. In the final 
verb complementation pattern, each class c
i,j
 has 
the frequency feature f
i,j
, which aggregates the fre-
quency of co-occurrence with v
i
 (1 ≤ i ≤ n; 1 ≤ j ≤  
m
i
) for the given class and its subclasses. The fre-
quency information is used to estimate the class 
probabilities given a verb, P(c
i,j
 | v
i
): 
 
∑
=
=
l
m
l
li
ji
ji
f
f
p
1
,
,
,
 
                                                           
2
 The ontology used for learning allowed multiple inheritance 
only at the leaf level, that way incurring no ambiguities when 
substituting subclass by its superclass. The multiple inheri-
tance at the leaf level was resolved by mapping each term to 
all its classes, which were then processed by a GA. 
Unclassified terms remain present in the final 
verb complementation patterns, and, like classes, 
they are also assigned the information on the fre-
quency of co-occurrence with the given verb. 
When classifying a specific term, this information 
is used to select the verb based on whose pattern 
the term will be classified. Precisely, the verb the 
given term most frequently co-occurs with is cho-
sen, as it is believed to be the most indicative one 
for the classification purpose. 
 
4.2   Term Similarity Measure 
 
A complementation pattern associated with the 
chosen verb typically contain several classes. In 
order to link the newly recognised terms to specific 
candidate classes, we used a hybrid term similarity 
measure, called the CLS similarity measure. It 
combines contextual, lexical and syntactic proper-
ties of terms in order to estimate their similarity 
(Nenadic et al., 2002).  
Lexical properties used in the CLS measure re-
fer to constituents shared by the compared terms. 
The rationale behind the lexical term similarity 
involves the following hypotheses: (1) Terms shar-
ing a head are likely to be hyponyms of the same 
term (e.g. progesterone receptor and oes-
trogen receptor). (2) A term derived by modi-
fying another term is likely to be its hyponym (e.g. 
nuclear receptor and orphan nuclear re-
ceptor). Counting the number of common con-
stituents is a simple and straightforward approach 
to measuring term similarity, but it falls short when 
it comes to single-word terms and those introduced 
in an ad-hoc manner. Thus, properties other than 
lexical need to be included.  
We use syntactic properties in the form of spe-
cific lexico-syntactical patterns indicating parallel 
usage of terms (e.g. both Term and Term). All 
terms used within a parallel structure have identi-
cal syntactic features and are used in combination 
with the same verb, preposition, etc., and, hence, 
can be regarded as similar with high precision. 
However, patterns used as syntactic properties of 
terms have relatively low frequency of occurrence 
compared to the total number of terms, and in or-
der to have a good recall, a large-size corpus is 
needed. In order to remedy for small-size corpora, 
other contextual features are exploited.  
Context patterns (CPs) in which terms appear 
are used as additional features for term compari-
son. CPs consist of the syntactic categories and 
other grammatical and lexical information (e.g. 
PREP NP V:stimulate). They are ranked ac-
cording to a measure called CP-value  (analogue to 
C-value for ATR). The ones whose CP-value is 
above a chosen threshold are deemed significant 
and are used to compare terms. Each term is asso-
ciated with a set of its CPs, and contextual similar-
ity between terms is then measured by comparing 
the corresponding sets. Automatically collected 
CPs are indeed domain-specific, but the method for 
their extraction is domain independent. 
 
4.3   Term-Class Similarity 
 
The CLS similarity measure applies to pairs of 
terms. However, in case of multiple choices pro-
vided by the verb complementation patterns, we 
need to compare terms to classes. In order to do so, 
we use the similarity between the given term and 
the terms belonging to the classes. The selection of 
terms to be compared is another issue. One possi-
bility is to use the full or random set of terms (be-
longing to the given class) that occur in the corpus. 
Alternatively, some ontologies provide a set of 
prototypical instances for each class, which can be 
used for comparison of terms and classes.
3
 More 
formally, if c is a class, e
1
, e
2
,..., e
k
 are terms repre-
senting the class, and t is a term, then the similarity 
between the term t and the class c is calculated in 
the following way: 
∑
=
∈
=
k
j
j
i
ki
etCLS
etCLS
ctEx
1
2
},...,1{
),(
),(
max),(  
 
This example-based similarity measure maxi-
mises the value of the CLS measure between the 
term and the instances representing the class. In 
addition, the values of the CLS measure are 
mapped into the interval (0,1) by performing vec-
tor normalisation in order to make them compara-
ble to the class probability estimations.  
 
4.4   Term Classification 
 
Finally, given the term t and the verb v
i
 it most  
frequently co-occurs with, a score is calculated for 
                                                           
3
 For example, in the UMLS ontology each class is assigned a 
number of its prototypical examples represented by terms. 
each class c
i,j
 from the set C
i
 according to the fol-
lowing formula: 
 
),()1(),(
,,, jijiji
ctExapactC ⋅−+⋅=    (1) 
 
where a (0 ≤ a ≤ 1) is a parameter, which balances 
the impact of the class probabilities and the simi-
larity measure.
4
 A class with the highest C(t, c
i,j
) 
score is used to classify the term t. Alternatively, 
multiple classes may be suggested by setting a 
threshold for C(t, c
i,j
). 
At this point, let us reiterate that the final verb 
complementation patterns are minimal in the sense 
that the number of terms in a verb complementa-
tion pattern and the depth of each individual term 
in the ontology are minimised. The latter condition 
may cause the classification to be crude, that is – 
new terms will be assigned to classes close to the 
root of the ontology. For more fine-grained classi-
fication results, the classes placed close to the root 
of the ontology should be either removed from the 
initial verb complementation patterns, thus being 
unable to override the classes found lower in the 
hierarchy or in other way prevented from substitut-
ing less general terms. The depth up to which the 
terms are to be blocked may be empirically deter-
mined. 
5 Experiments and Evaluation 
5.1   Resources 
 
The resources used for the experiments include an 
ontology and a corpus, both belonging to the do-
main of biomedicine. We used an ontology, which 
is a part of the UMLS (Unified Medical Language 
System) knowledge sources (UMLS, 2002). 
UMLS integrates biomedical information from a 
variety of sources and is regularly updated. 
Knowledge sources maintained under the UMLS 
project include: METATHESAURUS linking term 
variants referring to the same concepts; 
SPECIALIST LEXICON providing syntactic informa-
tion for terms, their component words, and general 
                                                           
4
 Note that when a = 0, the classification method resembles 
the nearest neighbour classification method, where the exam-
ples are used as a training set. On the other hand, when a = 1, 
the method is similar to naive Bayesian learning. However, in 
both cases the method represents a modification of the men-
tioned approaches, as the classes used in formula (1) are not 
all classes, but the ones learned by the GA. 
English words; and SEMANTIC NETWORK contain-
ing information about the classes to which all 
METATHESAURUS concepts have been assigned. 
The knowledge sources used in our term classi-
fication experiments include METATHESAURUS 
and SEMANTIC NETWORK. As the number of terms 
in  METATHESAURUS was too large (2.10 million 
terms) and the classification scheme too broad 
(135 classes) for the preliminary experiments, we 
made a decision to focus only on terms belonging 
to a subtree of the global hierarchy of the 
SEMANTIC NETWORK. The root of this subtree re-
fers to substances, and it contains 28 classes. 
The corpus used in conjunction with the above 
ontology consisted of 2082 abstracts on nuclear 
receptors retrieved from the MEDLINE database 
(MEDLINE, 2003). The majority of terms found in 
the corpus were related to nuclear receptors and 
other types of biological substances, as well as the 
domain-specific verbs extracted automatically 
from the corpus in the way described in Section 3.  
 
5.2   Evaluation Framework 
 
When retrieving terms found in the context of do-
main-specific verbs (see Section 3 for details) both 
terms found in the ontology and terms recognised 
on the fly by the C/NC-value method should be 
extracted. However, for the purpose of evaluation, 
only terms classified in the ontology were used. In 
that case, it was possible to automatically verify 
whether such terms were correctly classified by 
comparing the classes suggested by the classifica-
tion method to the original classification informa-
tion found in the ontology. 
During the phase of retrieving the verb-term 
combinations, some of the terms were singled out 
for testing. Namely, for each verb, 10% of the re-
trieved terms were randomly selected for testing, 
and the union of all such terms formed a testing set 
(138 terms) for the classification task. The remain-
ing terms constituted a training set (1618 terms) 
and were used for the learning of complementation 
patterns.  
 
5.3   Results 
 
Based on the training set, domain-specific verbs 
were associated with the complementation patterns 
given (see Table 1 for examples). Then, each term 
from the training set was associated with the verb 
it most frequently co-occurred with. The comple-
mentation pattern learnt for that verb was used to 
classify the term in question.  
 
Verb Complementation pattern 
activate
bind
Immunologic Factor 
Receptor   
Enzyme 
Hormone 
Organic Chemical 
Hazardous or Poisonous Substance 
Pharmacologic Substance 
Table 1. Learnt verb complementation patterns 
 
Since the UMLS ontology contains a number of 
prototypical examples for each class, we have used 
these class representatives to compare unclassified 
terms to their potential classes as indicated in Sec-
tion 4. Table 2 shows the results for some of the 
terms from the testing set and compares them to 
the correct classifications obtained from the ontol-
ogy. 
  
Term Suggested 
class 
Correct classes 
4 hydroxy-
tamoxifen
Organic   
   chemical 
Organic chemical 
benzoic
acid
Organic  
   chemical 
Organic chemical 
Pharmacologic      
   substance 
testoster-
one
Pharmacologic 
   substance 
Steroid  
Pharmacologic 
   substance 
Hormone 
Table 2. Examples of the classification results 
 
Note that in UMLS one term can be assigned to 
multiple classes. We regarded a testing term to be 
correctly classified if the automatically suggested 
class was among these classes. Table 3 provides 
information on the performance of the classifica-
tion method for each of the considered verbs sepa-
rately and for the combined approach in which the 
verb most frequently co-occurring with a given 
term was used for its classification. The combined 
approach provided considerably higher recall 
(around 50%) and a slight improvement in preci-
sion (around 64%) compared to average values 
obtained with the same method for each of the 
verbs separately. The classification precision did 
not tend to very considerably, and was not affected 
by the recall values. The recall could be improved 
by taking into account more domain-specific verbs, 
while the improvement of precision depends on 
proper tuning of: (1) the module for learning the 
verb complementation patterns, and (2) the similar-
ity measure used for the classification. Another 
possibility is to generalize the classification 
method by relying on domain-specific lexico-
syntactic patterns instead of verbs. Such patterns 
would have higher discriminative power than verbs 
alone. Moreover, they could be acquired automati-
cally. For instance, the CP-value method can be 
used for their extraction from a corpus (Nenadic et 
al., 2003a). 
 
Verb Recall Precision F-measure 
activate 19.28 66.59 29.90 
bind 29.30 66.53 40.68 
compete   3.58 63.16   6.78 
conserve   2.41 61.82   4.64 
inhibit 16.62 62.81 26.28 
interact 13.16 64.31 21.85 
mediate 11.68 62.75 19.69 
modulate 10.44 64.13 17.96 
repress   6.18 62.91 11.25 
stimulate   9.39 63.25 16.35 
Average: 12.20 63.83 20.48 
Combined: 49.88 64.18 56.13 
Table 3. The performance of the classification  
                  method 
 
The values for precision and recall provided in 
Table 3 refer to the classification method itself. If 
it were to be used for the automatic ontology up-
date, then the success rate of such update would 
also depend on the performance of the term recog-
nition method, as the classification module would 
operate on its output. We used the C/NC-value 
method for ATR; still any other method may be 
used for this purpose. We have chosen the C/NC-
value method because it is constantly improving 
and is currently performing around 72% recall and 
98% precision (Nenadic et al., 2002). 
6 Conclusion 
Efficient update of the existing knowledge re-
positories in many rapidly expanding domains is a 
burning issue. Due to an enormous number of 
terms and the complex structure of the terminol-
ogy, manual update approaches are prone to be 
both inefficient and inconsistent. Thus, it has be-
come absolutely essential to implement efficient 
and reliable term recognition and term classifica-
tion methods as means of maintaining the knowl-
edge repositories. In this paper, we have suggested 
a domain independent classification method as a 
way of incorporating automatically recognised 
terms into an existing ontology. For the prelimi-
nary experiments, we used the UMLS ontology in 
the domain of biomedicine, but the method can be 
easily adapted to use other ontologies in any other 
domain.  
The classification method makes use of the 
contextual information. Not all word types found 
in the context are of equal importance in the 
process of reasoning about the terms: the most in-
formative are verbs, noun phrases (especially 
terms) and adjectives. The presented term 
classification approach revolves around domain-
specific verbs. These verbs are used to collect 
unclassified terms and to suggest their potential 
classes based on the automatically learnt verb 
complementation patterns.  
Note that not every term appearing in a corpus 
is guaranteed to be classified by the proposed clas-
sification method due to the fact that a term need 
not occur as a complement of a domain-specific 
verb. Still, for a large number of terms the classifi-
cation method is expected to obtain the classifica-
tion information, as it is highly probable (though 
not certain) for a term to occur in a context of a 
domain-specific verb. The main goal of the method 
is to provide aid for the automatic ontology update 
by populating newly recognised terms into an ex-
isting ontology, rather than classifying arbitrary 
term occurrences in the corpus.  
The presented classification method can be eas-
ily modified to use lexical classes other than verbs 
as a criterion for classification. Even more, it can 
be further generalised to use a combination of lexi-
cal classes, which can be specified as a set of 
lexico-syntactic patterns. Further experiments with 
the generalisation of the classification method by 
basing it on a set of domain-specific lexico-
syntactic patterns instead of domain-specific verbs 
are expected to demonstrate better performance in 
terms of recall and precision. These facts suggest 
that our classification approach, in combination 
with the C/NC-value method, could be reliably 
used as a (semi)automatic ontology maintenance 
procedure.  

References 
Nigel Collier, Chikashi Nobata and Junichi Tsujii. 2001. 
Automatic Acquisition and Classification of Termi-
nology Using a Tagged Corpus in the Molecular Bi-
ology Domain. Journal of Terminology, John 
Benjamins. 
Katerina Frantzi, Sophia Ananiadou and Hideki Mima. 
2000. Automatic Recognition of Multi-Word Terms: 
the C-value/NC-value Method. International Journal 
on Digital Libraries 3(2):115-130. 
Vasileios Hatzivassiloglou, Pablo Duboue and Andrey 
Rzhetsky. 2001. Disambiguating Proteins, Genes, 
and RNA in Text: A Machine Learning Approach. 
Bioinformatics, 1(1):1-10. 
Diana Maynard and Sophia Ananiadou. 2000. Identify-
ing Terms by their Family and Friends. Proceedings 
of COLING 2000, Saarbrucken, Germany, 530-536.  
MEDLINE. 2003. National Library of Medicine. Avail-
able at: http://www.ncbi.nlm.nih.gov/PubMed/ 
Goran Nenadic, Irena Spasic and Sophia Ananiadou. 
2002. Automatic Acronym Acquisition and Term 
Variation Management within Domain-Specific 
Texts. Proceedings of LREC-3, Las Palmas, Spain, 
2155-2162.  
Goran Nenadic, Irena Spasic and Sophia Ananiadou. 
2003a. Automatic Discovery of Term Similarities Us-
ing Pattern Mining. To appear in Terminology. 
Goran Nenadic, Simon Rice, Irena Spasic, Sophia 
Ananiadou and Benjamin Stapley. 2003b. Selecting 
Features for Text-Based Classification: from Docu-
ments to Terms. Proceedings of ACL Workshop on 
Natural Language Processing in Biomedicine, 
Sapporo, Japan. 
Chikashi Nobata, Nigel Collier and Junichi Tsujii. 2000. 
Automatic Term Identification and Classification in 
Biology Texts. Proceedings of the Natural Language 
Pacific Rim Symposium (NLPRS’2000), 369-375. 
Irena Spasic, Goran Nenadic and Sophia Ananiadou. 
2002. Tuning Context Features with Genetic Algo-
rithms. Proceedings of 3
rd
 International Conference 
on Language,  Resources and Evaluation, Las Pal-
mas, Spain, 2048-2054. 
UMLS. 2002. UMLS Knowledge Sources. National 
Library of Medicine, 13
th
 edition.  
