Word Sense Disambiguation of Persian Homographs
F. Jani and A.H. Pilevar
Language Engineering Lab., Computer Engineering Dept., Bu Ali Sina University, Hamedan, Iran
Keywords: Thesaurus, Word Sense Disambiguation, Persian Language Corpus.
Abstract: This paper seeks to elaborate on the disambiguation of Persian words with the same written form but
different senses using a combination of supervised and unsupervised method which is conducted by means
of thesaurus and corpus. The present method is based on a previously proposed one with several differences.
These differences include the use of texts which have been collected by supervised or unsupervised method.
In addition, the words of the input corpus were stemmed, and in the case of those words whose different
senses have different roles in the sentence, the role of the word in the input sentence was considered for
disambiguation. Applying this method to the selected ambiguous words from “Hamshahri”, which is a
standard Persian corpus, we achieved to a satisfactory accuracy of 97 percent in the results, and evaluated
the presented method as a better and more efficient in comparison with the similar methods.
1 INTRODUCTION
There are words in some languages which cannot be
pronounced unless one knows their meaning in the
context. Such words are called homographs. For
converting a text to equivalent speech, the system
must be able to disambiguate homographs.
Therefore, according to the above definition, a word
with several senses is called homograph, and
disambiguating homographs is a matter of word
sense disambiguation (WSD).
WSD is related to attributing a proper sense to a
certain word in a text or speech (Gausted, 2004). In
contrast to other languages, and English above all,
for which many WSD systems have been designed,
it should be noted that, due to a lack of tagged
corpora, no such system has been so far developed
in a large scale and with a high accuracy for the
Persian language.
With regard to the strategies and approaches
used, there are three methods for solving the
problem of assigning the appropriate meaning to an
ambiguous word in context:
I. Using an explicit lexicon such as thesaurus,
with knowledge based method.
II. Disambiguation of word senses based on,
trained sense tagged corpus, for extracting
relevant information.
III. Combining the aspects of both I, and II
methodologies.
This paper seeks to elaborate on the
disambiguation of Persian homographs using a
combination of supervised and unsupervised method
which is conducted by means of corpus and
thesaurus. The method of paper (Ide and Veronis,
1998) is extended from different aspects and
presented in this article. These differences include
the use of texts which have been collected by
supervised or unsupervised method. In addition, the
words of the input corpus were stemmed and, in the
case of those words whose different senses have
different roles in the sentence, the role of the word in
the input sentence was considered for
disambiguation.
The organization of this paper is as follows: The
presented algorithm is discussed in Section 2. In
section 3, the corpus and thesaurus used in the
experiments are described. Section 4 provides a
report on the proposed method and its comparison
with the previous method as implemented on a
selection of homographs. Finally, section 5 offers an
analysis as well, the obtained conclusions.
2 THE PROPOSED METHOD
The method which is described in (Ide and Veronis,
1998) is used in this article and the algorithm is
extend for recognizing the ambiguous words. The
presented algorithm is based on conceptual
328
Jani F. and Pilevar A..
Word Sense Disambiguation of Persian Homographs.
DOI: 10.5220/0004031703280331
In Proceedings of the 7th International Conference on Software Paradigm Trends (ICSOFT-2012), pages 328-331
ISBN: 978-989-8565-19-8
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
categorization of context such that to be able to use
it for disambiguation of the word senses. Therefore,
in the first step we determine, to which category the
ambiguous word belongs. Then we construct the
related context discriminator and, since each word
sense belongs to one category, the correct sense can
be predicted (Ide and Veronis, 1998; Makki and
Homayounpour, 2008). The thesaurus is used for
determining the conceptual categories.
In this step, we extract sentences for each sense
which contain the ambiguous word and then begin to
disambiguate with regard to words that are
collocated with one of the senses of the ambiguous
word, but it must be noted that, due to the ambiguity
of the homograph, we cannot extract sentences
merely through the ambiguous word. Thus, we
should use thesaurus to find the synonymous words
for each sense of the ambiguous word and extract
the sentences containing those words for each
conceptual category from the input corpus.
However, in reality, some words in thesaurus are not
only ambiguous in themselves, but also they cannot
often replace the ambiguous word in the sentence;
therefore, in this method, a very limited number of
words in the thesaurus were considered, and to
increase accuracy, some sentences were extracted
from the corpus in a supervised manner. More
precisely, a sentence in which the ambiguous word
or its synonym resides is extracted, because by
considering a window around the ambiguous word
we may go beyond the scope of one sentence and it
should be noted that two contiguous sentences do
not necessarily share a conceptual connection. In
fact, our aim is to extract the words which are
collocated with the ambiguous word in the sentence
as well as their probability of occurrence.
Another difference of this method with the
previous one is the use of texts whose words have
been stemmed. The reason is that in the collected
texts, the same word may occur in different
morphological forms and in an unstemmed mode a
separate probability is calculated for each form. If
the input sentence is the same, every time a word is
likely to appear in one of its forms, it has been
considered as a stem for it. After identifying the
categories, similar to (Ide and Veronis, 1998) the
discriminators are constructed in the following steps:
1. For the conceptual categories, determine the
representatives of the contexts.
2. In contexts, calculate the weights of the
words
3. For new contexts, use the calculated weights
in step 2, for predicting sense of ambiguous
word.
In the proposed algorithm, the above steps are
developed as follows:
2.1 Determine the Representatives of
the Contexts
In the supervised mode, the sentences containing
ambiguous words were extracted from the
Hamshahri Corpus and added in separate conceptual
categories according to the meaning of the
ambiguous word. In the unsupervised mode, the
sentences containing a synonym of the ambiguous
words in different conceptual categories were
extracted. Thus, for each conceptual category
several sentences were obtained in a classified form.
2.2 Calculate the Weights of the Words
For every word (w) in the collected sentences the
probability Pr(TCat|w) is calculated for each
category (TCat) of the ambiguous word by use of the
law of conditional probability. Similar method is
proposed in (Makki and Homayounpour, 2008;
Yarowsky, 1992), the salient words with larger
probabilities (Formula 1) are selected, and their
logarithms are considered as weights.
Pr(w|TCat)/Pr(w)
(1)
In (Ide and Veronis, 1998) this probability as
well as its logarithm is considered for all of the
words in the collected texts. The presented article is
doing the same, but because Pr(w) is equal for all
conceptual categories, only Pr(w|TCat) is considered
in the proposed method, which is caused a higher
speed in the system.
2.3 Predicting Sense of Ambiguous
Words
For any newly entered ambiguous word, the system
assigns a score value. This score value calculated
based on Formula 2 which adds the words category
weights.
Score(TCat)=
∑ log( (Pr(w|TCat)*Pr(TCat)) / Pr(w))
(2)
For each ambiguous word, the highest assigned
category score is selected as its proper sense. Also
those words of the test context that have not
appeared in any of the collected contexts, for a
certain category, scores are calculated by Formula 3.
Pr(w|TCat)= log(1 / N(TCat) )
(3)
Word Sense Disambiguation of Persian Homographs
329
In formula 3, N is the frequency or number of
context, and TCat is the category of the corpus. The
relation between probability and frequency is
reverse, because for the unknown conceptual
categories, there is a small chance for some words to
occur in the collected contexts.
3 THE IMPLEMENTED CORPUS
AND THESAURUS
In this paper, Hamshahri Corpus was used; the
corpus is made up of Hamshahri Newspaper’s
archive in Iran, and containing 190209 texts with a
wide range of topics such as political, cultural, sport,
arts, etc. For the supervised mode, due to the lack of
a proper sense-tagged corpus in Persian, some parts
of the corpus were tagged manually. Furthermore, a
part of the intended texts was extracted through an
unsupervised method and by means of a thesaurus.
A Persian thesaurus which contains words and
expressions is used. The utilized Persian thesaurus
is very similar to Roget’s thesaurus (Fararooy, 1997;
Fararooy, 2004), with the same number of heads,
sections, classes, and indexes.
4 EXPERIMENTS AND RESULTS
We have applied our system for fifteen different
Persian homographs. In presented method, we
implement a hybrid technique because, for each
homograph a category is extracted from the
thesaurus, and for the texts collection corpus is used.
The accuracy of the proposed method is shown in
Table 1.
The first column shows the homograph, the
second column shows the accuracy of the basic
method in (Ide and Veronis, 1998), and the two
right-hand side columns depict the accuracy of our
proposed method.
Having stemmed the words in the collected
sentences and calculated the probability of the
occurrence of each word in different conceptual
categories, a brief inspection discloses that the
probabilities have gained more reasonable and
appropriate values and the probability of index
words has increased due to their different forms
being homogenized. Moreover, regarding the score
which is finally given to different conceptual
categories, the appropriate sense in the proposed
method differs large from other sense as compared
with the previous method.
In addition, removing the ambiguous words from
the thesaurus causes information to be gathered
which is conceptually more relevant to the intended
concept.
As another factor, considering the role of the
word in the sentence in cases where the different
senses of the word have different roles has led to an
Table 1: comparison of the basic method with the
proposed role based and none role based.
Ambiguous
words
Role based %
None role
based %

(sabk /sabok)
100
97

(nafs / nafas)
95
95

(khord / kharad)
100
95.2
Average
98.33
95.73
accuracy of 100 percent. This method can also be
used to eliminate morphological ambiguities. Every
system of natural language processing needs
morphological units since it may encounter with
words whose morphological structure is impossible
to determine without knowing their meaning, for
instance: word “ (saram)”. It has two meanings:
the first one is “ ” (“my head”) which has a
morphological structure consisting of // (possessive
pronoun) and // (noun). The second meaning,
however, entails another structure, namely a single
noun “ (serom)” (“a medical drip”). The basis of
such methods lies on the highest probability or
calculating the similarity according to training data.
Therefore, its efficiency is extremely sensitive to the
number of training instances.
Training instances are the representative of the
ambiguous words. Therefore, the training data must
contain sufficient number of samples from various
texts.
5 CONCLUSIONS
This paper proposed a method for word sense
disambiguation of homographs in Persian language,
based on the method proposed in (Ide and Veronis,
1998). Basically the method is: sentences containing
the ambiguous word are collected from the corpus,
the appropriate sentences are gathered for each
conceptual category, synonymous words are
obtained from the thesaurus, and then the sentences
containing these synonyms are extracted. In the
basic method there is no need for sense-tagged data,
but for increasing accuracy and with the aim of
ICSOFT 2012 - 7th International Conference on Software Paradigm Trends
330
evaluating the efficiency of the basic method as
compared with the supervised method a number of
the intended texts here were gathered in a supervised
manner. In order to increase the accuracy, the words
in the obtained corpus were stemmed after extracting
the sentences, and it caused a noticeable increase in
efficiency and accuracy.
After obtaining the sentences in a classified form
for each conceptual category, the probability of
occurrence of each word was calculated in that
category. These probabilities lead to a score for the
input test sentence in each category which is, in fact,
indicative of the probability of each sense. Finally,
the conceptual category with a higher probability is
taken as the output of the system. Moreover, an
analysis of the corpus showed that regarding some
ambiguous words different senses may appear in
different roles in the sentence; therefore, by
considering the role of the ambiguous word in the
input test sentence we can produce an accuracy of
100 percent in such cases.
Also, the results showed that the modifications
applied to the basic algorithm in (Ide and Veronis,
1998) improved the accuracy of word sense
disambiguation by 6 percent. The proposed method
can also be used for morphological disambiguation.
REFERENCES
Fararooy, J., Thesaurus and Electronic transfer of Persian
language content, In: 2nd workshop on Persian
language and computer, Tehran, Iran, 2004.
Fararooy, J., Thesaurus of Persian Words and Phrases,
1999.
Gausted, T., Linguistic Knowledge and Word Sense
Disambiguation, PhD dissertation, Groningen
University, 2004.
Ide, N., Veronis, J., Introduction to the Special Issue on
Word Sense Disambiguation: The State of the Art,
Computational inguistics 24(1), 1–40, 1998.
Makki, R., Homayounpour, M., Word Sense
Disambiguation of Farsi homographs Using Thesaurus
and Corpus, Amirkabir University of Technology,
Tehran, Iran, 2008.
Word Sense Disambiguation of Persian Homographs
331