The International Arab Journal of Information Technology, Vol. 17, No.

1, January 2020 45

Extracting Word Synonyms from Text using Neural

Approaches
Nora Mohammed
College of Engineering, Al-Qadisiyah University, Iraq

Abstract: Extracting synonyms from textual corpora using computational techniques is an interesting research problem in the
Natural Language Processing (NLP) domain. Neural techniques (such as Word2Vec) have been recently utilized to produce
distributional word representations (also known as word embeddings) that capture semantic similarity/relatedness between
words based on linear context. Nevertheless, using these techniques for synonyms extraction poses many challenges due to the
fact that similarity between vector word representations does not indicate only synonymy between words, but also other sense
relations as well as word association or relatedness. In this paper, we tackle this problem using a novel 2-step approach. We
first build distributional word embeddings using Word2Vec then use the induced word embeddings as an input to train a feed-
forward neutral network using annotated dataset to distinguish between synonyms and other semantically related words.

Keywords: Neural networks, semantic similarity, word representations, natural language processing.

Received April 17, 2017; accepted October 24, 2017

https://doi.org/10.34028/iajit/17/1/6

1. Introduction WordNet, or do not exist at all. Moreover, and despite

the massive volume of WordNet entries, it does not
Many attempts have been made in the literature to include many contemporary compound adjectives
answer the question: what does it mean for two words (e.g., hard-wired, industry-leading, etc.,).
to be synonyms? Jurafsky and Martin [5] state that the Computational synonymy discovery from textual
synonymy relation between two words holds if they can sources can assist automating the process of building
be substituted for each other in all situations and have similar lexical resources for other languages.
the same propositional meaning. With this definition, From a linguistic point of view, synonymy can be
there will only be a few examples that can be categorized as one of the paradigmatic relations which
categorized as perfect synonyms since different words relate two words that tend to occur in similar context.
that have identical meanings in all contexts are rare in Besides synonymy, other paradigmatic relations have
natural languages. For example, the pair big-large been recognized. These include antonymy (e.g., large-
would be considered as synonyms by most native small), hypernymy (e.g., car-vehicle), co-hyponym
English speakers. However, in the sentence “He's like a (e.g., cat-dog), and metonymy (e.g., face- nose). On
big brother to me”, “big” cannot be substituted by the other hand, two words exhibit a syntagmatic
“large”. Leeuwenberg et al. [7] consider two words to relation if they tend to co-occur in the same context
be synonyms if they denote the same concept and are [2]. For example, cup and coffee are syntagmatically
interchangeable in many contexts, with regard to one of related due the popularity of the expression “cup of
their senses. The latter definition is quite convenient for coffee”. In the psychological literature, this semantic
computational applications. relation is referred to as association [4]. Turney [13]
Synonyms discovery has been a long-standing referred to the two major relation categories as
exercise for linguists and lexicographers. Furthermore, functional similarity (words that are paradigmatically
the accessibility of large repository of synonymy pairs related) and domain similarity (words that are
can be crucial for many Natural Language Processing syntagmatically related).
(NLP) applications such as information retrieval, Recently, neural word embeddings techniques such
question answering, and machine translation evaluation as word2vec [9] and GLoVe [11] have been leveraged
systems. The largest lexical database that has been to build distributional semantic models that capture
manually built and maintained for English language is semantic similarity between word lexemes. However,
WordNet [10]. WordNet has been widely used as a it has been observed that these techniques cannot be
synonyms repository in many NLP applications. directly used to discover synonyms in textual corpora
However, manually building and maintaining such because context-based distributional models as well as
lexical resources is an expensive and labour-intensive similarity measures are biased towards context
task. Therefore, lexical databases such as WordNet for similarity, which is not identical to semantic similarity
other languages are either small, compared to English most of the times. For example, in a pilot experiment
46 The International Arab Journal of Information Technology, Vol. 17, No. 1, January 2020

we have conducted, the most similar word to “large” INPUT PROJECTION OUTPUT
was “small”, even though the two are actually exact
opposites of each other (antonyms pair). These
w(t-2)
observations motivate the need for novel computational
techniques that are tailored to capture synonymy
between word pairs and not any other semantic relation. w(t-1)
The rest of this paper is organized as follows: section
w(t)
2 presents the related work in the literature regarding
computational synonyms extraction. Section 3 w(t+1)
introduces the data and approach we have used to
construct the word embeddings, or distributional word
representations. Section 4 presents the neural network w(t+2)
we have used to build a classifier for synonym
Figure 1. Architecture of the Skip-Gram model for learning word
identification.
representations.

2. Related Work Computational techniques to extract synonyms

from raw text have been inspired by the classical
Word embeddings (also known as distributional word
distributional hypothesis “if two words have almost
representations) are vector representations for words
identical environments, we say that they are
that are usually constructed from raw text based on
synonyms” [3]. Since word embeddings that have
linear context (words that occur in the neighbourhood
been constructed using linear contexts encode
of a target word). In these representations, each word is
information about the environments of a target word,
converted into a vector of numerical values or real
they can be used to the synonyms extraction task. A
values. However, in other cases some grammatical notable word in this direction is that of Van der Plas
information are utilized in the process of constructing
and Tiedemann [14] which used distributional word
word vectors. An example of the latter approach is the
similarity to find synonyms in Dutch. Their measure
work of Levy and Goldberg [8] which constructed
of distributional similarity was constructed using a
vectors based on grammatical context (using
monolingual Dutch corpus. The authors observed that
dependency relations) instead of linear context.
using this approach other word relation categories
In the last few years, several neural network-based
(e.g., antonyms, hyponyms, etc.,) would occur more
models for constructing word embeddings have been
than synonyms. To tackle this problem, the authors
proposed. Among these, the most popular approach is
have used different parallel corpora and word
the Continuous Bag-of-Words model (CBoW) and the
alignment to obtain translation context. The idea is to
Skip-Gram model (SG) developed by Mikolov et al.
harness the lexical gap that exists in some languages
[9]. The popularity of these two models is due to the
or the fact that words which have similar meaning
fact that they require less time for constructing the translate to only word in another target language.
vectors compared to other neural-based approaches. It
Using both distributional similarity and translational
has been shown that embeddings constructed by these
similarity, the authors were able to find more true
two approaches encode many valuable information
synonyms that using distributional similarity in
about the relation between a pair of words (e.g.,
isolation. Nevertheless, using translational similarity
similarity, relatedness, morphological information,
to extract synonyms for one language requires the
etc.,). Word embeddings constructed by the Skip-Gram
availability of several translations of that language to
model, for instance, have shown to perform very well
other languages, which is not always available
for the analogy task. A typical example of the analogy
specially for least resourced languages.
task is figure out the question: “man” is to “woman” as
Perhaps the most similar work to the research
“king” is to “?”, which is not a trivial task for
conducted in this paper in the recent literature for
computers to perform. Using word embeddings, it has
synonymy discovery is the work of Leeuwenberg et
been reported that the distance between the vector
al. [7]. The authors presented a minimally supervised
representations of the words “king” and “queen” is very
approach to extract synonyms using distributional
close to the distance between the words “man” and
word vectors for English and German. Their aim was
“woman”. The Skip-Gram model of word2vec is
to improve evaluation systems for machine translation
illustrated in Figure 1.
by adding more synonymous words to the lexical
repository. The main contribution of their work is that
they proposed a custom similarity measure tailored to
the synonymy discovery task which they referred to as
relative cosine similarity. It has been shown in their
work that the conventional cosine similarity is a bad
measure for the synonymy identification task. To
Extracting Word Synonyms from Text using Neural Approaches 47

achieve their goal, the authors conducted an intrinsic general public in written form, and the verb
and extrinsic evaluation on their approach. For intrinsic “book/VB” which is synonymous to the verb
evaluation, two human evaluators accessed whether or “reserve”. To ensure a good quality of the constructed
not each word pair is synonymous. For extrinsic word embeddings, we only considered the words with
evaluation, the authors used their system in machine frequency higher than 25. This was followed by a
translation evaluation task and observe an improvement normalization step in which words were lower-cased
on the evaluation metric of the machine translation. It (e.g., “The/DT” to “the/DT”) and digits were replaced
worth mentioning that Leeuwenberg et al. [7] favoured by a wildcard (e.g., “546/CD” to “DIGIT/CD”). The
minimally supervised system so their approach can be final vocabulary size was 197,361 word types. Our
extended to other least resourced languages which may pre-processed corpus contained 24.7 million sentences
not supported by rich lexical databases or sophisticated and around 602 million word tokens.
NLP tools. The only source of annotation that was used In order to construct vectors from the POS-tagged
in that research is part-of-speech tagged corpus. corpus, we used the neural word embedding model
In this paper, we use a similar approach to the work (word2vec). Using word2vec, vectors of similar
proposed by Leeuwenberg et al. [7], but we add another lexical items would be grouped together in the
supervised step that requires a small set of annotated semantic vector-space. The Skip-Gram model was
data. We show that synonymy identification can be used, number of dimensions was set to 300, and
tackled as a supervised machine learning task where the context window size was set to 5. Then, we used the
features are actually the word embeddings constructed cosine similarity measure to obtain a cluster of similar
using a generic technique such as Word2Vec. Contrary words for nouns, verbs, adjectives, and adverbs.
to previous works in NLP and script recognition which
relies on manually engineered features (for example, 3.2. Qualitative Analysis
the work of Bahashwan et al. [1] and Khan et al. [6]),
To perform a simple (qualitative) evaluation of the
our work relies on hidden representative features
aforementioned approach, we manually inspect the 5
discovered by a neural network.
most similar words to a given set of target words
(Table 1).
3. Constructing Word Embeddings Many interesting observations can be made from
An intuitive simple solution to this problem is to Table 1. For the word “police”, none of the top 5
convert each lexical item in the text corpus into a vector words can be recognized as a direct synonym.
representation, then computing the similarity between However, these words exhibit domain similarity to the
the vectors using a similarity measure (e.g., cosine target word. Since the corpus used in this experiment
similarity). Given vector representations for two words, is a news corpus crawled from the open web, one can
namely w1 and w2, the cosine similarity can be conclude that these words occur in crime-related
expressed as inner product space which measures the reports. In these contexts, the phrase “police
cosine of the angle between them. The higher the investigators” can be regarded as a synonym to the
similarity between two vectors, the higher the chance phrase “police detectives”. More interestingly, other
that the lexical items they represent are synonyms. This paradigmatic relations can be recognized within the
might be followed by a classification step to identify top 5 words themselves. For example, “police
whether the similarity is actually due to synonymy investigators” and “police officers” are co-hyponyms
(positive instance) or other type of of the hypernym “employees at the police
similarity/association (negative instance). In this department”.
section, the approach that was used to build the word Table 1. A few target words (with their POS tag) and their 5 most
embeddings is explained in details. similar words.
Target word (POS) Top 5 similar words
3.1. Data and Approach authorities, officers, investigators, detectives,
police (NN)
eyewitnesses
In this paper, we used NewsCrawl 2014 corpus from large (JJ) small, huge, sizable, massive, big
WMT workshop. We removed the headlines from the social-networking photo-sharing, video-sharing, micro-blogging,
corpus by removing sentences that do not end with a (JJ) on-demand, twitter-like
state-funded, government-funded, non-federal,
punctuation mark. After that, the corpus was tagged funded (JJ)
non-university, bursary
using Stanford Part-Of-Speech (POS) tagger developed scientific (JJ)
peer-reviewed, empirical, neuroscientific
by Toutanova et al. [12]. The advantage of using POS laboratory-based, scholarly
cheap, off-the-shelf, affordable, cost-effective,
tagged data is that homographs with different POS tags inexpensive (JJ)
nontoxic
can be distinguished if the POS tag is appended to the murdered (VBD) raped, abducted, hanged, stabbed, killed
word. For example, the two main senses of the word treated (VBN)
cared, discharged, drugged, hospitalized,
readmitted
“book” would be represented as two different words, quickly (RB) soon, swiftly, fast, easily, slowly
that is, the noun “book/NN” which refers the
instrument that is used to convey information to the
48 The International Arab Journal of Information Technology, Vol. 17, No. 1, January 2020

The second target word is the adjective “large”. The similar behaviour. This observation suggests that fine-
top 5 similar words are all adjectives that can be used to grained POS tags might help distinguishing different
describe size. Among these, three words are senses of homographs as well as different senses of
synonymous to the word large (huge, massive, and big). the verbs. Finally, the last entry in Table 1 shows the
However, the most similar word is “small”, which adverb “quickly” and its top 5 similar words.
actually an antonym to the target word. The top 5
words seem to exhibit functional similarity to the target 3.3. Does Distributional Similarity Indicate
word. Synonymy?
We have observed the presence of many
contemporary compound adjectives in the corpus that To further understand the nature of the distributional
are made of multi-words separated by a hyphen. An similarity between words, we conduct an investigation
example of these adjectives is social-networking as in on a set of target words that compiled by the author.
“social-networking/JJ website/NN”. The top 5 similar To narrow the scope of our investigation, we only
words to this token were also compound adjectives, and considered adjectives for this investigation. We looked
interestingly, all of them are within the web 2.0 in depth into the most similar words to a set of 100
domain. However, these adjectives exhibit functional words. We used WordNet as a reference to discover
similarity as well. For example, if we replace “social- the relation between the similar words if they are
networking” by “micro-blogging” in the sentence linked somehow in this lexical database. If the two
“Twitter is a popular social-networking platform”, the words are not linked in WordNet, we perform a
meaning of the sentence would not change. In addition, manual evaluation to categorize the relation between
“photo-sharing”, “video-sharing”, and “micro- the two words. The result of this investigation is show
blogging” are different ways of creating user-generated in Table 2. Moreover, we were able to find the
content on the web, which manifests the co-hyponymy following categories of words similarity/relatedness in
relation. This observation motivates an appealing our investigation.
research direction that aims to discover mulit-word  WordNet synonyms: words which are recognized as
adjectives and their domain of usage (e.g., technology, synonyms in WordNet.
bio-sciences, politics. etc.,). A similar trend can be  WordNet antonyms: words which are recognized as
observed with the adjective “funded”. The two most antonyms in WordNet.
similar lexical items are actually a specification of the  WordNet similar to: words which are recognized as
adjective, namely “state-funded” and “government- similar to each other in WordNet.
funded”.  WordNet see also: words which are connected to
Because the technique used in this paper is based on each other in WordNet under the see also category.
linear context (that is, context words that are present in  Nearly synonyms: words which are recognized as
the neighbourhood of a target word), adjectives that synonyms.
tend to modify the same set of words would have high
 Specification: A word is a specific case of another
cosine similarity. For example, the most similar word to
(more general) word.
the target word “scientific” is “peer-reviewed”, which
 Domain Similarity: words that occur in the same
can be justified by the fact that these two adjectives
domain or topic (topically similar words).
modify similar word (e.g., “scientific journal” vs.
“peer-reviewed” journal or “scientific evidence” vs  Contrasting: words that contrasting but do not
“empirical evidence”). Perhaps the only direct example qualify as opposites or antonyms.
of the feasibility of distributional similarity to discover  Association: words that are associated in most
synonyms is present in the target word “inexpensive”, context.
where the most similar words can considered as  Shortening: A word is an orthographic short form
synonyms (or nearly synonyms) by most native English of another full word.
users (i.e., cheap, off-the-shelf, affordable, cost-  Similar: The words are somehow similar, but they
effective). are neither synonyms nor antonyms. They do not
For verbs, we observed a similar behaviour to those exhibit any relatedness.
already discussed. For example, the top 5 similar words WordNet contains a lot of words that are connected by
to the verb “murdered” in the past simple tense (tagged the synonymy or antonyms relation. However, some
as VBD) are all verbs that describe criminal events, adjective pairs might be considered as synonyms by a
only one of them can be considered as a synonym human evaluator even though WordNet does not
(“killed”). On other hand, the verb “treated” in the past recognize them as direct synonyms. For example, the
participial tense (tagged as VBN) was similar to verbs word “appealing” would be considered as synonym to
that describe events which usually take place at a the word “attractive” by most English speakers. But
hospital or medical institution (the medical sense of the WordNet puts “appealing” under the category similar
verb treat). Interestingly, the verb “treated” in the to the word “attractive”. The same can be said about
simple past tense (tagged as VBD) does not show a
Extracting Word Synonyms from Text using Neural Approaches 49

the word “beautiful”, which is put under the category necessarily indicate synonymy. It is quite rare that the
see also to the word “attractive”. Therefore, we added most similar word is actually a synonym to the target
these relations to Table 2. word. This raises the question: given a vector
From Table 2, one can observe that only 6 adjective representation of two words, can we build a system
pairs out of the 100 pairs are actually WordNet that classifies whether the two words are synonyms or
synonyms. Interestingly, 21 of the pairs are WordNet not? If yes, how to obtain labelled data to train the
antonyms. However, many of the pairs that are system?
identified as synonyms by manual investigation are We addressed the synonymy identification
actually within the similar to category in WordNet. We problem as a classification task. To make the problem
also identified other pairs that are not directly linked in simpler and doable within the time-frame of the
WordNet to be synonyms (we refer to them as nearly research, only adjectives were considered for the
synonyms). We found 23 pairs that can be qualified as classification. We used a feed-forward neural network
nearly synonyms. These observations could justify the with backward propagation as a learning algorithm.
low precision of the work of Leeuwenberg et al. [7] The classifier architecture is shown in Figure 2. To
when evaluation was performed against the synonyms obtain labelled training data, we extracted synonyms
in WordNet compared to the manual evaluation which pairs from SimLex-999 similarity lexicon with
gave much more optimistic results. Perhaps the most similarity score > 6:5. SimLex-999 is a gold-standard
interesting example in Table 2 is (natural-liquefied), resource for evaluating distributional semantic models
since the similarity between the two words seems to be [4]. According to authors’ own words: “Simlex-999
unknown at first. We manually checked the corpus for explicitly quantifies similarity rather than association
instances of the two adjectives, we found that the two or relatedness”. Simlex-999 was produced using
words co-occur in many contexts in the corpus due to crowd-sourcing of 500 paid native speakers who were
the wide use of the expression liquefied natural gas in asked to rate the similarity, as opposed to association,
newswire corpora such as the one we have used in these of different concepts provided a visual interface. Table
experiments. We chose to refer to this particular 3 shows some examples from the data set.
instance of similarity as association. It seems that the
Skip-Gram model of Word2Vec capture this kind of
association with high similarity score.
Table 2. Counts per category for 100 target adjective with their most
similar words.
Category Most similar (out of 100) Example
WordNet synonyms 6 Possible-potential
WordNet antonyms 21 Unsuccessful- successful
WordNet similar to 21 Responsible-accountable
WordNet see also 3 careful-cautious
Nearly-synonyms 23 limited-minimal
healthy-
Specification 5
heart-healthy Figure 2. Feed-Forward Neural Network Architecture for
scientific- synonymy classification.
Domain similarity 4
peer-reviewed
personal- Table 3. Examples from the SimLex-999 lexicon. Similarity scale
Contrasting 3
work-related ranges from 10 (most similar) to 0 (least similar).
natural-
Association 2
liquefied Word1 Word2 POS Sim-score
Shortening 1 professional-pro old new A 1.58
Similar 11 proper- adequate smart intelligent A 9.2
plane jet N 8.1
woman man N 3.33
Finally, we found 11 adjective pairs where the word dictionary N 3.68
semantic similarity somehow exists but does not create build V 8.48
get put V 1.98
qualified to be exactly synonyms and cannot be keep protect V 5.4
categorized under any of the aforementioned
categories. These pairs also are not directly linked in Considering the examples given in Table 3, one can
WordNet and we simply refer to this category as conclude that SimLex-999 was designed to reflect
similar. semantic similarity due to synonymy relation (as
shown in the similarity score of smart and intelligent)
4. Neural Network for Synonymy more than any other relation. For example, even
Identification though the words “man” and “woman” denote similar
concepts (i.e., co-hyponyms of the hypernym person
From the analysis in the previous section, it is clear that or human), they received a relatively low similarity
similarity (as measured by cosine similarity) does not score. In addition, and despite the strong functional
50 The International Arab Journal of Information Technology, Vol. 17, No. 1, January 2020

similarity between “old” and “new”, they received a antonyms, (co) hyponyms, etc.,). This task can be
low similarity scores. Therefore, the SimLex-999 can tackled as a multi-class classification in a similar
be used to perform a quantitative evaluation of the fashion to that we used to address the synonymy
effectiveness of semantic models to reflect synonymy identification task.
with high confidence.
Then, we expanded the dataset by linking the pairs 6. Conclusions
to the WordNet lexicon. For example, for a given pair
(good, great), we searched in the synsets of “good” In this paper, we addressed the problem of extracting
where “great” is one of the synonyms and included all synonyms from text corpus using word embeddings
other synonyms within the same synset. To obtain and supervised neural network. We used word2vec to
negative examples, antonyms in the same synset were construct word embeddings and performed a
extracted and antonym pairs were formed. Using this qualitative evaluation of the most similar words to a
method, we obtained 128 synonymy pairs (SYN) and few target words. Our investigation showed that
91 antonymy pairs (ANT). We added 90 pairs of words distributional similarity does not always indicate
that are neither synonyms nor antonyms (ELS) which synonymy but the similarity might be due to other
have been annotated manually. functional similarity (e.g., antonym) or domain
We performed several experiments or both 2-way similarity (e.g., association). Then, we showed that
classification (SYN/ANT) and 3-way classification embeddings constructed using word2vec can be used
(SYN/ANT/ELS). We used the vector representations as features to feed a neural network for synonymy
(the output of word2vec) for each pair as an input to the classification task. For future work, we suggested
neural network. In each of the classification runs, the extending our approach to sense relation identification
data were split into 75:25 training and testing, instead of synonymy discovery which can be tackled
respectively. The results are reported in Table 4. as multi-class classification task.

Table 4. Results of the classification experiments. References

Experiment # Hidden units Accuracy
20 0.76 [1] Bahashwan M., Abu-Bakar S., and Sheikh U.,
2-Class “Efficient Segmentation of Arabic Handwritten
134 0.74
25 0.60 Characters Using Structural Features,”
3-Class
136 0.63 The International Arab Journal of Information
Technology, vol. 14, no. 6, pp. 870-879, 2017.
Even though the training dataset was small, the [2] Crystal D. The Cambridge Encyclopedia of
results for the classification experiments are Language, Cambridge: Cambridge University
encouraging. In both cases (2-class and 3-class), the Press, 1987.
accuracy is way above random guessing, which indicate [3] Harris Z., “Distributional Structure,” Word, vol.
that supervised learning is feasible for the synonym 10, no. 2-3, pp. 146-162, 1954.
identification task using word embeddings as features. [4] Hill F., Reichart R., and Korhonen A., “Simlex-
999: Evaluating Semantic Models with (genuine)
5. Future Work Similarity Estimation,” Computational
Linguistics, vol. 41, no. 4, pp. 665-695, 2016.
The work that has been presented in this paper can be [5] Jurafsky D. and Martin J., Speech and Language
extended in a few potential directions. In our work, we Processing, Pearson, 2014.
used the word embeddings as features to feed the neural [6] Khan K., Baharudin B., and Khan A.,
network. Thus, the features of the words reflect the “Identifying Product Features from Customer
linear environments in which they tend to occur. For Reviews Using Hybrid
future work, we intend to enrich the features by adding Patterns,” The International Arab Journal of
additional features that are inferred from translational Information Technology, vol. 11, no. 3, pp. 281-
context to help distinguish synonymy from other sense 286, 2014.
relations. Translational context can be obtained either [7] Leeuwenberg A., Vela M., Dehdari J., and van
by either using parallel corpora with word alignment or Genabith J., “A Minimally Supervised Approach
word-to-word translation dictionaries. Nevertheless, For Synonym Extraction with Word
using parallel corpora means that one has to deal with Embeddings,” The Prague Bulletin of
the morphological variations of lexical items in Mathematical Linguistics, vol. 105, no. 1, pp.
morphology-rich languages. 111-142, 2016.
Another direction is to extend the work presented in [8] Levy O. and Goldberg Y., “Dependency-Based
this paper is to expand the problem into sense relations Word Embeddings,” in Proceedings of the 52nd
identification. That is, given two words that exhibit Annual Meeting of the Association for
distributional similarity, the task would be to identify
the sense relation of the two words (e.g, synonyms,
Extracting Word Synonyms from Text using Neural Approaches 51

Computational Linguistics, Baltimore, pp. 302-

308, 2014.
[9] Mikolov T., Chen T., Corrado G., and Dean J.,
“Efficient Estimation of Word Representations in
Vector Space,” in Proceedings of the 1st
International Conference on Learning
Representations, Scottsdale, pp. 1-12, 2013.
[10] Miller G., “WordNet: A Lexical Database for
English,” Communications of the ACM, vol. 38,
no. 11, pp. 39-41, 1995.
[11] Pennington J., Socher R., and Manning C.,
“Glove: Global Vectors for Word
Representation,” in Proceedings of the
Conference on Empirical Methods in Natural
Language Processing, Doha, pp. 1532-1543,
2014.
[12] Toutanova K., Klein D., Manning C., and Singer
Y., “Feature-Rich Part-of-Speech Tagging with A
Cyclic Dependency Network,” in Proceedings of
the Conference of the North American Chapter of
the Association for Computational Linguistics on
Human Language Technology, Edmonton, pp.
173-180, 2003.
[13] Turney P., “Domain and function: A Dual-Space
Model of Semantic Relations and
Compositions,” Journal of Artificial Intelligence
Research, vol. 44, pp. 533-585, 2012.
[14] Van der Plas L. and Tiedemann J., “Finding
Synonyms Using Automatic Word Alignment and
Measures of Distributional Similarity,” in
Proceedings of The COLING/ACL on Main
Conference Poster Sessions, Sydney, pp. 866-
873, 2006.

Nora Mohammed is currently a

researcher at the college of
engineering, Al-Qadisiyah
University, Iraq. Her main research
interest is in the field of natural
language processing and
computational linguistics. She has
worked on mining natural language text for information
retrieval, relation extraction, and synonymy discovery.
She received her Master degree in Computer Science
and Engineering from Osmania University, India.