Generating Sense Inventories for Ambiguous Arabic Words

Abstract

The process of selecting the appropriate meaning of an ambigous word according to its context is known as word sense disambiguation. In this research, we generate a number of Arabic sense inventories based on an unsupervised approach and different pre-trained embeddings, such as Aravec, Fast text, and Arabic-News embeddings. The resulted inventories from the pre-trained embeddings are evaluated to investigate their efficiency in Arabic word sense disambiguation and sentence similarity. The sense inventories are generated using an unsupervised approach that is based on a graph-based word sense induction algorithm. Results show that the Aravec-Twitter inventory achieves the best accuracy of 0.47 for 50 neighbors and a close accuracy to the Fast text inventory for 200 neighbors while it provides similar accuracy to the Arabic-News inventory for 100neighbors. The experiment of replacing ambiguous words with their sense vectors is tested for sentence similarity using all sense inventories and the results show that using Aravec-Twitter sense inventory provides a better correlation value

Full text

446 The International Arab Journal of Information Technology, Vol. 18, No. 3A, Special Issue 2021
Generating Sense Inventories for Ambiguous
Arabic Words
Marwah Alian1 and Arafat Awajan1,2
1King Hussein School of Computing Sciences, Princess Sumaya University for Technology, Jordan
2Information Technology College, Computer Science Department, Mutah University, Jordan
Abstract: The process of selecting the appropriate meaning of an ambigous word according to its context is known
as word sense disambiguation. In this research, we generate a number of Arabic sense inventories based on an
unsupervised approach and different pre-trained embeddings, such as Aravec, Fasttext, and Arabic-News
embeddings. The resulted inventories from the pre-trained embeddings are evaluated to investigate their efficiency
in Arabic word sense disambiguation and sentence similarity. The sense inventories are generated using an
unsupervised approach that is based on a graph-based word sense inductionalgorithm. Results show that the Aravec-
Twitter inventory achieves the best accuracy of 0.47 for 50 neighbors and a close accuracy to the Fasttext inventory
for 200 neighbors while it provides similar accuracy to the Arabic-News inventory for 100neighbors. The experiment
of replacing ambiguous words with their sense vectors is tested for sentence similarity using all sense inventories
and the results show that using Aravec-Twitter sense inventoryprovides a better correlation value.
Keywords: Word sense induction, word sense disambiguation, arabic text, sense inventory.
Received February 25, 2021; accepted March 7, 2021
https://doi.org/10.34028/iajit/18/3A/8
1.
Introduction
Semantic similarity has an important role in different
applications of Natural Language Processing (NLP) [4].
Ambiguous words affect the semantic similarity
between two texts because the similarity score between
texts depends on the similarity of their context words to
determine if the two texts are similar or not [1, 23].
A single word that may have different meanings is
called an ambiguous word, and the process of detecting
the appropriate meaning of an ambiguous word is known
as Word Sense Disambiguation (WSD) [17]. The
context of an ambiguous word consists of the words
surrounding the ambiguous target word.
The ability to define what a word means with respect
to its meaning is one of the most difficult issues in NLP.
Ambiguity is common across all languages, but it has
greater challenges in Semitic languages such as Arabic
[5].
According to Alian et al. [6], WSD approaches can
be categorized into knowledge-based, supervised,
unsupervised, and hybrid approaches. In knowledge-
based approaches, the different meanings of an
ambiguous word are extracted from a dictionary or a
lexicon. Supervised approaches use training annotated
corpus and testing sets, unsupervised approaches have
no training set and instead use word context and
clustering algorithms, and hybrid approaches merge the
different methods.
One of the unsupervised approaches is word sense
induction, which represents words as a graph and then
uses a clustering algorithm to group similar words in
the graph. Each cluster is considered as a sense. Our
research uses one of the word sense induction
approaches to build a sense inventory for Arabic based
on pre-trained embeddings. The sense inventory is used
in WSD for sentences with ambiguous words from the
Arabic paraphrasing benchmark [3]. The sense
inventory is then evaluated using the retrieved senses in
terms of accuracy measure [2].
Four sense inventories are generated and tested using
Aravec-Twitter, Aravec-Wiki, Fasttext, and Arabic-
News pre-trained embeddings. Another experiment is
conducted to show the benifit of using the sense
representation by replacing the sense representaion in
the sentence and comparing this new representation of
the sentence by the use of pre-trained embedding of an
ambigous word in the sentence. Also, for these four
sense inventories, an experiment is performed by
replacing the vector of the retrieved sense instead of the
ambiguous word vector to represent the sentence
representative vector. Next, the similarity between
sentences is measured and compared to the human
evaluation using Pearson correlation.
This paper is organized in five sections as follows.
Section 2 reviews the previously proposed work related
to Arabic WSD. Section 3 explains the sense induction
algorithm used for constructing the sense inventory
while section 4 discusses the experiment and
results.Then, section 5 presents the conclusion.

Generating Sense Inventories for Ambiguous Arabic Words 447
2.
Related Work
Different approaches are proposed for Arabic WSD
using word representation. For example, Alian et al. [5]
used Wikipedia as a lexical resource and a Vector Space
Model as a representational approach to texts. Cosine
similarity is then used to measure the relatedness
between Wikipedia’s retrieved senses and the text that
has an ambiguous word.
Hadni et al. [13] utilized two external resources,
Arabic WordNet (AWN) and English WordNet (WN),
to translate terms that cannot be found in AWN using a
machine translation system. The nearest concept for the
ambiguous word is chosen based on the number of
relationships between concepts in the same local
context. The authors evaluated their approach using
naïve Bayesian and support vector machine. The
proposed approach achieves an accuracy of 0.732 using
Wu and Palmer’s [24] measure with support vector
machine.
Representing words as vectors in the distributional
space has attracted researchers in different NLP
applications and provided promising results. Arabic
WSD based on word embedding is one of these
applications. For example, Laatar et al. [16] proposed a
WSD method based on word embedding where the word
embedding is learned using the skip-gram model [19].
The similarity is measured between context vector and
sense vectors, where the context vector is computed
using word embeddings that appear in the context of an
ambiguous word. The definitions of senses are retrieved
from a dictionary, and the most similar definition vector
to the context vector is selected as the appropriate sense.
This approach achieves an accuracy of 0.78.
In addition, Alkhatlana et al. [7] utilized two
embedding methods, Word2vec and GloVe, to generate
global contexts of words and extract the synsets of
ambiguous words from AWN. They constructed a test
dataset to be used for the WSD task. The sense vector is
obtained based on the retrieved AWN synset and then
the cosine similarity between context vector and sense
vector is computed. The most similar sense vector is
considered as the correct sense for an ambiguous word.
Pelevina et al. [22] proposed an approach for learning
word sense embeddings. This approach provides a sense
inventory from word embeddings by applying a
clustering algorithm to an ego-network or word graph
with related relationships. They used two WSD
methods where the vectors for context words are taken
from the matrix of word embeddings or the matrix of
context representative vectors, the first method used the
probability of sense in a context while the second
method used the similarity between the sense
representative vector and the context vector.
Then, Chang et al. [10] introduced an efficient
graph-based approach for word sense induction which
constructs global non-negative vector embeddings.
Then they used clustering for the generated graph to get
senses of each ambiguous word. The experiment was
conducted using three datasets and the results show
similar or better sense clusters compared to other
methods such as Pelevina et al. [22].
Logacheva et al. [18] proposed a new unsupervised
WSD approach based on the work of Pelevina et al.
[22]. This approach depends on pre-trained embeddings
and does not need any external annotated corpus. In this
approach, a semantic graph is constructed for words in
the vocabulary of the pre-trained embedding model and
then the graph is clustered into subgraphs according to
the similarity between word vectors. Each subgraph
represents a sense. Next, a retrofitting approach is used
to make the sense vector in the direction of the
ambiguous word. The authors used Fasttext
embeddings to build sense inventories for 158
languages, including Arabic.
A comparison between the previously discussed
approaches that works with Arabic is given in Table 1.
The comparison includes the authors, publication year,
category of approach, corpus or dataset used, sense
inventory, evaluation metric and results.
In this research, we apply the approach of Logacheva
et al. [18] who used pre-trained embeddings of
Aravec. WSD is then applied to 86 sentences from the
Arabic paraphrasing benchmark using the senses
extracted from sense inventories. The results are
compared to the results of the senses retrieved from
the Fasttext inventory.
Table 1. Comparison between approaches used for WSD.
Ref
Year
Approach category
Dataset/Corpus/embeddings
Sense Inventory
Similarity
measure
Evaluation metric
Results
Alian et al. [5]
2016
Knowledge based
External dataset
Wikipedia
Cosine similarity
N/A
-
Hadni et al. [13]
2016
Knowledge based
Essex Arabic Summaries
Corpus (EASC)
AWN, WN.
Wu and Palmer's
Accuracy
0.732
Laatar et al. [16]
2017
Semi-supervised
Historical Arabic Dictionary
Corpus,
Almu-Jam-Alwasit
dictionary
Cosine similarity
Precision
0.78
Alkhatlana et al.[7]
2018
Semi-supervised
collected from Arabic News
AWN
Cosine similarity
N/A
-
Logacheva et al. [18]
2020
Unsupervised
Fasttext pre-trained
embeddings
Created from
embeddings
Cosine similarity
N/A for Arabic
-
3.
Word Sense Disambiguation
Polysemy is defined as the existence of many possible
meanings for a word. These meanings are called senses,
and the word will be called a multi-sense word. It is one
drawback of word representation, and it can be solved

448 The International Arab Journal of Information Technology, Vol. 18, No. 3A, Special Issue 2021
using sense representation techniques. Our work will be
based on context and sense representation to
disambiguate a word and find the similarity.
WSD is used to find the proper meaning of a word
with an uncertain meaning given its context [14, 22].
Owing to the ambiguity in human languages, a word
may represent different meanings in two contexts. These
meanings are identified in sense inventories in separate
units, known as senses [15]. For example, the word “
Hkym” in the following sentences has different
meanings according to its context:
The boy went to a doctor for treatment.
The man was wise in his choice.
The different senses of the word “Hkym” in Arabic
are a person with wisdom  , doctor , and
philosopher . In the first sentence, the word “
Hkym” means “doctor ” while in the second
sentence it means “wise  .” The context “for
treatment” helps identify that the sense of “Hkym”
is a doctor not a wise man.
The process of disambiguation of polysemy words
depends on building a sense inventory to retrieve the
senses of the ambiguous word and then selecting the
most similar sense to the context of an ambiguous word.
We benefit from the work of Logacheva et al. [18] in
building an Arabic sense inventory using two different
Aravec pre-trained embeddings. One is trained on
Twitter and the other is trained on Wikipedia.
The algorithm of Logacheva et al. [18] consists of
two main concepts: the first relies on graph-based word
sense induction, and the second is graph filtering using
vector operations for word vectors.
3.1. Word Sense Induction
Word sense induction depends on finding a list of
nearest neighbors for word embedding in the
distributional space. The method of constructing the
semantic graph is as follows:
For each word win the vocabulary:
 Construct the set of N-nearest neighbors (S) for the
target word w. Let S members be {s1, s2,…,sn}.
 Construct the set of N-anti-neighbors (∆) that consists
of words that are not similar to the corresponding
nearest neighbors of w, where the vectors of these
words is computed as the subtraction between the
vector of word w and its neighbor s: (w −si).
 Construct the set 
={
,
 ,…,
}that consists of the
most similar words to the vectors in ∆, but the result
may be the target word w [17].
 The set of anti-pairs consists of (si,
) but not the
target word w(
   .
 These anti-pairs are words that should not be
connected in the graph unless both words si
are
members in the set of N-nearest neighbors (S).
 Construct the set of vertices of the graph (V) by
adding words from the set (S) and their anti-pair from
(
) only if the word and its anti-pair are part of the
set (N) of the target word w. In other words, only add
to the set (V) words that may benefit in separating
different senses of w.
 Construct the set of edges (E): For each word si in the
nearest neighbors (S), create a set of nearest
neighbors (S’)={c1,c2…,cn} and add the edge
between word si and the nearest neighbor cj if cj is
not an anti-pair of si.
There is no edge between a word si and its anti-neighbor
in the graph because they belong to different senses.
Then a clustering algorithm is used to get the senses
of an ambigous word. These steps are shown in Figure
1, while clustering is described in the following section.
Figure 1. Steps of word sense induction and graph clustering.
3.2. Clustering
The constructed graph is clustered into subgraphs where
each subgraph represents a sense of the target word. The
average of the word embeddings in each subgraph
represents the vector of the sense. Retrofitting is also
applied to the sense vector.
Each cluster represents a sense of the target word,
and the computed sense vector represents the keyword
of that sense. Each sense of the target word with its
keyword and cluster is saved to the sense inventory.
3.3. Disambiguation
Sense vectors are used for WSD in Arabic text by
extracting the senses of the ambiguous word from the
sense inventory and then computing the context vector
by averaging the vectors of context words that are most
similar to the ambiguous target word. The cosine
similarity is computed between the sense vector and the
context vector. Then the most similar sense will be
selected as the correct sense.
4.
Experiment and Results
To evaluate the disambiguation approach, four sense
inventories are generated and tested on sentence
similarity. The first one is generated from Twitter
Aravec, the second one uses Wiki Aravec, the third
sense inventory is generated using Arabic-News vectors
trained by Altowayan and Tao [9], and the last one is

Generating Sense Inventories for Ambiguous Arabic Words 449
generated by Logacheva et al. [18] from pre-trained
Fasttext vectors. The experiments are conducted to build
sense inventories based on word semantic graphs with
N-neighbors. N is tested for 50, 100, and 200.
4.1. Dataset
Arabic paraphrasing benchmark [20] is used in the WSD
experiment. This benchmark is constructed based on the
transformation rules for Arabic [8, 11], such as
permutation, deletion, and addition. These rules are
applied to the structure of a sentence to produce a new
sentence. The benchmark consists of 1010 sentence
pairs labeled for similarity and paraphrasing. In our
experiment, we used 86 sentences containing ambiguous
words.
4.2. Aravec Pre-Tained Embeddings
AraVec [20] is an Arabic distributed word embedding
model that is trained using different resources and is
available online with different dimensions. The word
embeddings are obtained using the Word2vec skip-gram
and Continuous Bag Of Words (CBOW) models [19].
Aravec-Twitter model is trained on Arabic tweets
with a vocabulary size of 145,428 and dimensions of 100
and 300 for each word vector. The document size is
66,900,000.
Aravec-Wiki model is trained using 1,800,000
documents from World Wide Web pages with Arabic
content. It has vector dimensions of 100 and 300 and a
vocabulary size of 662,109.
4.3. Fasttext Pre-Trained Embeddings
Arabic Fasttext embeddings are provided by Grave et al.
[12]. These embeddings result from training on
Wikipedia and Common Crawl corpus. They use an
extension of the Fasttext model with subword
information. This model is available online and has a
dimension of 300 for word vector.
4.4. Arabic-News Pre-Trained Embeddings
Altowayan and Tao [9] build their corpus from news
articles with Arabic content based on local Arabic
newspapers and the international Arabic news from
CNN and BBC. They trained the corpus using the
Word2vec CBOWmodel to learn word embeddings with
a window size of 10 and a vector dimension of 300. The
vocabulary size of the learned embeddings is 159,175.
4.5. Results and Discussion
The retrieved senses are evaluated by an expert who
provides each selected sense with a label as correct or
incorrect. For ambiguous words that have no sense in the
sense inventory, an unknown label is given. The number
of target words to be disambiguated is 139.
Accuracy is measured as the correct senses from the
total senses, where the unknown senses are excluded
from the total number of senses as in Equation (1):
  

Tables 2, 3, and 4 compare the results of each sense
inventory for 50, 100, and 200 neighbors, respectively,
in terms of correct, incorrect, unknown senses, and
accuracy.
Table 2. Results of sense inventories for 50-Neighbors.
Sense Inventory
Correct
Incorrect
Unknown
Accuracy
Aravec-twitter
45
49
45
0.479
Aravec-Wiki
22
30
87
0.423
Fasttext
56
68
15
0.451
Arabic-News
36
79
24
0.313
Table 3. Results of sense inventories for 100-Neighbors.
Sense Inventory
Correct
Incorrect
Unknown
Accuracy
Aravec-Twitter
36
83
20
0.303
Aravec-Wiki
38
70
31
0.352
Fasttext
42
77
20
0.353
Arabic-News
36
83
20
0.303
Table 2 shows that the accuracy of the sense
inventory that is constructed based on Aravec-Twitter
pre-trained embedding provides the best accuracy value
of 0.48.
Table 3 shows that the accuracy of using Fasttext is
better than that of Aravec-Twitter inventory for 100
neighbors, but it is similar to the accuracy achieved by
the Aravec-Wiki inventory, while the accuracy of
Arabic-news inventory is similar to that of Aravec-
Twitter inventory.
The results of 200 neighbors in Table 4 show that the
accuracy of the Aravec-Twitter-based inventory
provides the least number of unknown senses. Fasttext-
based inventory provides a very similar accuracy value
as the Aravec-Twitter inventory. Arabic-news pre-
trained embeddings provides the lowest accuracy value
of 0.255.
Table 4. Results of sense inventories for 200-Neighbors.
Sense Inventory
Correct
Incorrect
Unknown
Accuracy
Aravec-twitter
60
69
10
0.465
Aravec-Wiki
53
72
14
0.424
Fasttext
54
62
23
0.466
Arabic-News
28
82
29
0.255
4.6. Word Sense Disambiguation
We use 86 sentences with ambiguous words from the
Arabic paraphrasing benchmark [3] to evaluate WSD
with the retrieved senses from the sense inventory.
We apply the algorithm of Logacheva et al. [18] to
the pre-trained word embeddings,that have been used to
construct the Arabic sense inventories, for the
disambiguation process.
The two Aravec models are Twitter and Wikipedia.
The Twitter model has 1,476,715 vocabularies, but
there is a limit for the vocabulary used in the experiment
of Logacheva et al. [18] . The Wiki model has 662,109
vocabularies. The experiments show that the number of
(1)

450 The International Arab Journal of Information Technology, Vol. 18, No. 3A, Special Issue 2021
vocabulary affects the clusters of each sense.
Sense inventories are generated from Aravec, Arabic-
News embedding vectors with N-neighbors (50, 100,
and200). The keyword of each sense cluster is generated
as in the approach of Logacheva et al. [18] by
determining the centroid of the cluster as the mean of
word vectors that belong to the sense cluster. The
resulted vector is shifted via a retrofitting approach to be
in the direction of the word vector.
We compute the scores for sentence similarity after
replacing ambiguous words with the selected sense
vectors retrieved from the generated 200-neighbors
Aravec sense inventory.
The four generated sense inventories are tested on
sentence similarity: Aravec-Twitter, Aravec-Wiki,
Arabic-News, and Fasttext inventory. Table 5 shows the
correlation of replacing an ambiguous word with its
retreived sense from each inventory to measure the
sentence similarity compared to human annotations.
Table 5. Correlation results of WSD from different sense inventories.
Sense inventory
Pearson Correlation
Twitter Aravec
0.399
Wiki Aravec
0.222
Fasttext
0.318
Arabic-News
0.29
5.
Conclusions
This paper presents a disambiguation approach for
Arabic words that uses the word sense induction
approach to build a sense inventory for Arabic words.
An evaluation for three sense inventories is provided,
where these inventories are based on four different pre-
trained embeddings, namely, Aravec-Twitter, Aravec-
Wiki, Fasttext, and Arabic-News embeddings.
In the experiment of 50 neighbors sense inventory,
the Aravec-Twitter sense inventory achieves the best
accuracy of 0.47, whereas in the 100 neighbors
experiment, the Fasttext sense inventory provides better
accuracy value.
In the case of 200 neighbors, the Aravec-Twitter and
Fasttext sense inventories achieve very similar accuracy
values.
Similarity between sentences is measured after
replacing the ambiguous word vector with the retrieved
sense vector and then the results are evaluated using
Pearson correlation. However, in the case of
paraphrasing identification task, the polysemy problem
still has to be studied. This task requires more analysis
of semantic similarity and material resources to evaluate
the effect of WSD.
References
[1] Alian M. and Awajan A., “Semantic Similarity
Approaches- Review,” in Proceedings of The
International Arab Conference on Information
Technology, Werdanye, pp. 1-6, 2018.
[2] Alian M. and Awajan A., “Sense Inventories for
Arabic Texts,” in Proceedings of The
International Arab Conference on Information
Technology, Giza, pp. 1-4, 2020.
[3] Alian M., Awajan A., Al-Hasan A., and Akuzhia
R., “Towards building Arabic paraphrasing
benchmark,” in Proceedings of The 2nd
International Conference on Data Science, E-
learning and Information Systems, Dubai, pp. 1-
5, 2019.
[4] Alian, M. and Awajan A., “Semantic Similarity
for English and Arabic Texts: A Review,” Journal
of Information and Knowledge Management, vol.
19, no. 4, 2020.
[5] Alian M., Awajan A., and Al-Kouz A., “Word
Sense Disambiguation for Arabic Text Using
Wikipedia and Vector Space Model,”
International Journal of Speech Technology, vol.
19, no. 4, pp. 857-867, 2016.
[6] Alian M., Awajan A., and Al-Kouz A., “Arabic
Word Sense Disambiguation-Survey,” in
Proceedings of The International Conference on
New Trends in Computing Sciences, Amman, pp.
236-240, 2017.
[7] Alkhatlana A., Kalita J., and Alhaddad A.,
“Word Sense Disambiguation for Arabic
Exploiting Arabic WordNet and Word
Embedding,” in Proceedings of The 4th
International Conference on Arabic
Computational Linguistics, Dubai, pp. 50-60,
2018.
[8] AlKouli M., Transformation Rules for Arabic
Language ( qwAEd tHwylyAh llgAh AlErbyAh),
Dar Al-Falah, 1999.
[9] Altowayan A. and Tao L., “Word Embeddings
for Arabic Sentiment Analysis,” in Proceedings of
the International Conference on Big Data (Big
Data), Washington, pp. 3820-3825, 2016.
[10] Chang H., Agrawal A., Ganesh A., Desai A.,
Mathur V., Hough A., and McCallum1 A.,
“Efficient Graph-based Word Sense Induction by
Distributional Inclusion Vector Embeddings,”
arXiv preprint arXiv:1804.03257, 2018.
[11] Chomsky N., Syntactic Structure, The Hague
Mouton Publishers, 1957.
[12] Grave E., Bojanowski P., Gupta P., Joulin A., and
Mikolov T., “Learning Word Vectors for 157
Languages,” in Proceedings of The International
Conference on Language Resources and
Evaluation, 2018.
[13] Hadni M., El Alaoui S., and Lachkar A., “Word
Sense Disambiguation for Arabic Text
Categorization,” The International Arab Journal
of Information Technology, vol. 13, no. 1A, no.
1A, pp. 215-222, 2016.
[14] Ide N. and Véronis J., “Word Sense
Disambiguation: The State of the Art,”
Computational Linguistics, vol. 24, no. 1, pp. 1-
40, 1998.

Generating Sense Inventories for Ambiguous Arabic Words 451
[15] Jurgens D., “An Analysis of Ambiguity in Word
Sense Annotations,” in Proceedings of the 9th
International Conference on Language Resources
and Evaluation, Reykjavik, pp. 3006-3012, 2014.
[16] Laatar R., Aloulou C., and Belguith L., “Word
Sense Disambiguation of Arabic Language With
Word Embeddings As Part of The Creation of A
Historical Dictionary,” in Proceedings of the
International Workshop on Language Processing
and Knowledge Management, Sfax, 2017.
[17] Laatar R., Aloulou C., and Bilguith L., “Word
Sense Disambiguation of Arabic Language With
Word Embeddings As Part of The Creation of A
Historical Dictionary,” in Proceedings of The 8th
International Conference on Computer Science
and Information Technology, Amman, 2018.
[18] Logacheva V., Teslenko D., Shelmanov A.,
Remus S., Ustalov D., Kutuzov A., Artemova E.,
Biemann C., Ponzetto S., and Panchenko A.,
“Word Sense Disambiguation for 158 Languages
using Word Embeddings Only,” arXiv preprint
arXiv:2003.0665, 2020.
[19] Mikolov T., Sutskever I., Chen K., Corrado G.,
and Dean J., “Distributed Representations of
Words And Phrases and Their Compositionality,”
Neural Information Processing Systems, pp. 3111-
3119, 2013.
[20] Mohammad A., Eissa K., and El-Beltagy S.,
“Aravec: A Set of Arabic Word Embedding
Models for Use in Arabic Nlp,” Procedia
Computer Science, vol. 117, pp. 256-265, 2017.
[21] Navigli R., “Word Sense Disambiguation: A
Survey,” ACM Computing Surveys, vol. 41, no. 2,
pp. 1-69, 2009.
[22] Pelevina M., Arefyev N., Biemann C., and
Panchenko A., “Making Sense of Word
Embeddings,” in Proceedings of The 1st
Workshop on Representation Learning for NLP,
Berlin, pp. 174-183, 2016.
[23] Srivastava S. and Govilkar S., “A Survey on
Paraphrase Detection Techniques for Indian
Regional Languages,” The International Journal
of Computer Applications, vol. 163, no. 9, pp.
0975-8887, 2017.
[24] Wu Z. and Palmer M., “Verb Semantics and
Lexical Selection,” in Proceedings of The 32nd
Annual Meeting of the Associations for
Computational Linguistics, Stroudsburg, pp. 133-
138, 1994.
Marwah Alian is a PhD candidate in
Princess Sumaya University for
Technology since 2015. She received
her B.Sc. degree in Computer
Science from Hashemite University
in 1995 while her MS.c degree was
received in Computer Science in
2007 from Jordan University. Her research interest is in
the fields of e-learning systems, data mining and
Natural language processing.
Arafat Awajan is a Full Professor
and the president of Mutah
University. He was teaching at
Princess Sumaya University for
Technology (PSUT). He received his
PhD degree in Computer Science
from the University of Franche -
Comte, France in 1987. He has held various
administrative and academic positions at the Royal
Scientific Society and Princess Sumaya University for
Technology. Head of the Department of Computer
Science (2000 -2003) Head of the Department of
Computer Graphics and Animation (2005 -2006) Dean
of the King Hussein School for Information Technology
(2004 - 2007) Director of the Information Technology
Center, RSS (2008 -2010) Dean of Student Affairs
(2011 - 2014) Dean of the King Hussein School for
Computing Sciences (2014 -2017) He is currently the
vice president of the university (PSUT). His research
interests include: Natural Language Processing, Arabic
Text Mining and Digital Image Processing.