USING AUTOMATED LEXICAL RESOURCES IN ARABIC SENTENCE SUBJECTIVITY

International Journal of Artificial Intelligence & Applications (IJAIA), Vol. 5, No. 6, November 2014
DOI : 10.5121/ijaia.2014.5601 1
USING AUTOMATED LEXICAL RESOURCES IN
ARABIC SENTENCE SUBJECTIVITY
Hanaa Mobarz1
, Mohsen Rashown 2
, Ibrahim Farag 3
1,3
Department of Computer Science, Faculty of Computers and Information, Cairo
University
2
Department of Electronics and Communications, Faculty of Engineering, Cairo
University
ABSTRACT
A common point in almost any work on Sentiment analysis is the need to identify which elements of
language (words) contribute to express the subjectivity in text. Collecting of these elements (sentiment
words) regardless the context with their polarities (positive/negative) is called sentiment lexical resources
or subjective lexicon. In this paper, we investigate the method for generating Sentiment Arabic lexical
Semantic Database by using lexicon based approach. Also, we study the prior polarity effects of each word
using our Sentiment Arabic Lexical Semantic Database on the sentence-level subjectivity and multiple
machine learning algorithms. The experiments were conducted on MPQA corpus containing subjective and
objective sentences of Arabic language, and we were able to achieve 76.1 % classification accuracy.
KEYWORDS
Sentiment analysis, Lexical recourses, Opinion mining, Subjectivity Lexicon, Arabic opinion mining.
1. INTRODUCTION
Subjective text expresses opinions, emotions, sentiment and beliefs, while objective text generally
report facts. So the task of distinguishing subjective from objective text is useful for many natural
languages processing applications like mining opinions from product reviews [1], summarizing
different opinions [2], question answering [3], etc.
Sentiment analysis or opinion mining refers to the general method to extract subjectivity and
polarity from text, and semantic orientation refers to the polarity and strength of words, phrases,
or texts. Our concern is primarily with the semantic orientation of texts, but we extract the
sentiment of words and phrases towards that goal [4].
Lexical Resources for sentiment analysis is collection of words that express the subjectivity in
text. In the research literature, theses words are also known as sentiment words, opinion words,
opinion-bearing words, and polar words .There are two types of opinion words; positive opinion
words which used to express the desired state (for examples ""‫جميل‬" beautiful", "‫"رائع‬
"wonderful", "‫"جيد‬ "good", and "‫حب‬" "love" ), and negative opinion words which used to express
the undesired state (for examples " "‫رھيب‬" terrible", "‫"سئ‬ "bad", "‫كره‬ " "hate", and "‫قبيح‬" "ugly" ).
There are three main approaches for generating subjective lexicons; manual approach, dictionary-
based approach, and corpus-based approach. Manual approach is very time consuming ([5], [6],
[7], and [8]) and thus it is not usually used alone, but combined with automated approaches as the
final check because automated methods make mistakes. Also manual approach has a lack of

2
neutralism that the decision of classification may be affected by (culture, religion, ethics ... etc).
Dictionary based approach; one of the simple techniques in this approach is based on
bootstrapping using a small set of seed opinion words and a dictionary, e.g., WordNet [9]. The
strategy is to first collect a small set of opinion words manually with known orientations, and then
to grow this set by searching in the WordNet for their synonyms and antonyms. The newly found
words are added to the seed list. The next iteration starts. The iterative process stops when no
more new words are found. This approach is unable to find opinion words with domain specific
orientations. Corpus based approach; the methods here depend on a seed list of opinion words to
find other opinion words in a large corpus and syntactic or co-occurrence patterns. Using the
corpus-based approach alone to identify all opinion words is not as effective as the dictionary-
based approach because it is hard to prepare a huge corpus to cover all language words. However,
this approach has a major advantage that the dictionary-based approach does not have. It can help
find domain specific opinion words and their orientations if a corpus from only the specific
domain is used in the discovery process [10].
Numerous of researches are available on building sentiment lexicons in English and other
languages but there are a very few number of such resources in Arabic. There are some factors
that have contributed to reducing the research related to the construction of Arabic lexical
resources for sentiment analysis such as:-
• The lack of availability of Arabic semantic database, still consider one of the most
difficult problems that face researcher who try to generate an Arabic subjective lexicon.
• Arabic is a high inflectional and derivational language; often a single word has more than
one affix such that it may be expressed as a combination of prefix(es), lemma (stem), and
suffix(es). The prefixes are articles, prepositions, or conjunctions. The suffixes are
generally objects or personal/possessive anaphora.
• The same three-letter root can give rise to different words, different meanings with
different orientations. For example ( ‫ج‬,‫م‬,‫ل‬ ) may be ("‫"جميل‬ "beautiful" ,positive) or ("‫"جمل‬
"camel", objective) .For example ( ‫س‬,‫ھـ‬,‫ل‬ ) may be ("‫"سھولة‬ "easy" ,positive) or ("‫"سھل‬
"plain", objective). For example ( ‫و‬,‫ض‬,‫ع‬ ) may be ("‫"التواضع‬ " modesty" ,positive) or
("‫"الوضاعة‬ " rascaldom", negative).
• Arabic letters can be written with different shapes according to their position in the word.
E.g. "Alif" has four forms ( ‫ا‬,‫أ‬,‫إ‬,‫آ‬ ), "Yaa" has two forms( ‫ى‬,‫ي‬ ) and "Taa el marpouta and
el haa el marpouta " ( ‫ة‬,‫ه‬ ).
The main contribution of this work is the development of an Arabic lexical resource of sentiment
analysis by exploiting the relations available in the Arabic lexical semantics database. The
database archives approximately 150,000 Arabic words, 18,413 semantic fields, and 20 semantic
relations, such as synonyms, antonym, hyponymy and causality. In addition, we study the prior
polarity effects of each word using our Sentiment Arabic Lexical resource on the sentence-level
subjectivity and multiple machine learning algorithms.
The paper is organized as follows: Section 2 mentions related works of subjective lexicon and
sentence level subjectivity classification. Section 3 gives an over view of Arabic Lexical
Semantic Data Base –RDI which is part of The Research and Development International (RDI)
toolkit. It is a main resource for generating our subjective lexicon. Section 4 presents the
proposed algorithm for generating Sentiment Arabic lexical semantic database-RDI (SentiRDI)
and how to determine the term orientation and subjectivity. Section 5 describes the corpus that is
used in sentence subjectivity classifier. Section 6 focuses on tools used in classification and the
experimental results. Finally, Section 7 draws our conclusions and future work.

3
2. RELATED WORK
In this section, we will present the most notable research work on building English sentiment
lexicons and previous attempts to build Arabic sentiment lexicons. In addition, we will also
present research work performed earlier on subjectivity analysis has been applied on English and
Arabic.
Lexical Resources for opinion mining is used in determining term orientation (subjective
/objective) and term polarity (positive/negative) out of context by using cognitive knowledge of
human beings which called prior polarity lexicons. Various previous works ([11]; [12]; [13]) have
already proposed techniques for making dictionaries for those sentiment words. Several sentiment
lexicons are available for English such as SentiWordNet [13], Subjectivity Word List [14],
WordNet Affect list [15], Taboada’s adjective list [16]. SentiWordNet is the most widely used in
several applications such as sentiment analysis, opinion mining and emotion analysis.
SentiWordNet is an automatically constructed lexical resource for English that assigns a
positivity score and a negativity score to each WordNet synset. [17].
For Arabic sentiment lexicon there are few attempts. An Arabic lexicon from similarity graph (A
similarity graph is a type of graph where by two words or phrases have an edge if they are similar
in polarity or meaning) was created in [18]. Lexicon consists of two columns, the first is the word
and the second column represents the score of the word. An Arabic sentiment lexicon was created
in [19] that contained strong as well as weak subjective clues by manually translating the MPQA
lexicon [14].
A machine translation procedure is used to translate available English lexicons [20], including
SentiWordNet [13], which is the most famous and most widely used English polarity lexicon
[21], into Arabic. They retrieved 229,452 entries, including expressions commonly used in social
media. The authors reported having problems with both coverage and with the quality of some of
the entries. They also stated that they have not tested the system for the task of sentiment analysis.
An Egyptian dialect sentiment lexicon was created in [22]. The researchers identified a set of
lexico-syntactic patterns indicative of subjectivity, used a seed list of 380 manually constructed
words, and subsequently performed pattern matching on a data set collected from tweeter. The
incorrectly learned candidate terms were manually filtered. They retrieved 4,392 entries (193
compounds negative, 83 compound positive, 3,344 negative, and 772 positive).The work
addressed dialectical or slang terms for the Egyptian dialect, which makes it unsuitable for use for
other dialects.
Arabic sentiment lexicon that assigns sentiment scores to the words found in the Arabic WordNet
(Arabic version of WordNet) was created in [23]. The authors here used lexicon_based approach
for generating their lexicon by Starting from a small seed list contains 20 words (10 positive
words and 10 negative words); they used semi-supervised learning to propagate the scores in the
Arabic WordNet by exploiting the synset relations. The algorithm assigned a positive sentiment
score to more than 800, a negative score to more than 600 and a neutral score to more than 6000
words in the Arabic WordNet. The lexicon was evaluated by incorporating it into a machine
learning-based classifier.
The sentence level classification considers each sentence as a separate unit and assumes that
sentence should contain only one opinion. Sentence-level sentiment analysis has two tasks
subjectivity classification and sentiment classification.
Research work performed earlier on subjectivity analysis has been applied only on English and
mostly at document-level and word-level. Some methods such as ([24], [25], [26], [27], [28],

4
[29], [30], [31] and [32]) which concentrated at sentence-level explored the effects of adjective
orientation and gradability on sentence subjectivity. Authors in [29] judge the sentence
subjectivity based on the adjectives used in the sentence. The idea was extended in [1] by using
the polarity of the previous sentence to break classification ties.
Bootstrapping is another approach used in [27] is considered the state of art in sentence
subjectivity detection. The idea is to use a high-precision classifier to automatically label a large
un-annotated data. The labelled data is then given to a pattern supervised learning algorithm to
generate a set of linguistically rich extraction patterns. The learned patterns are used to identify
more subjective statements and the bootstrapping process continues. A similar method to
automatically construct a corpus of HTML documents with polarity labels was used in [30].
Similar work involving self-training is described in [33] and [34]. A comprehensive survey of
subjectivity recognition using different clues and features was presented in [35].
In Arabic, There are few works addressing subjectivity classification problem on sentence level
([18], [35]). Authors in [18] hit subjectivity analysis task on the three levels of document, phrase,
and word and also provide polarity for mined sentences. They use lexicon based similarity graph
algorithm to obtain sentiment polarities. A two-stage classification approach for subjectivity and
polarity classification on sentence level was adapted in [35]. Authors use SVM classifier with a
linear kernel to achieve these tasks.
3. ARABIC LEXICAL SEMANTIC DATA BASE –RDI
Arabic Lexical Semantic database (RDI-ArabSemanticDB) is part of The Research and
Development International (RDI) toolkit. The Research and Development International (RDI)
toolkit mainly consists of Arabic RDI-ArabMorpho-POS tagger [37] and RDI-ArabSemanticDB
tool [38]. Arabic Lexical Semantic database of semantic fields [38] based on number of important
references ( "‫العربية‬ ‫األفعال‬ ‫تصريف‬ ‫معجم‬"-"‫الوسيط‬ ‫المعجم‬"–"‫الكبير‬ ‫المكنز‬ ‫معجم‬" ) ("Glossary of
Arabic verbs conjugation" – "the intermediary dictionary" – "dictionary thesaurus great")
Applying 20 semantic relations between fields (synonym, antonym, part of, kind of, causality,
hyponym ….etc) and morphological analysis of each word into its type, prefix, form root and
suffix. Around 2000 general concept are managed under 18,413 semantic fields, and 150,000
words based on stems (multiplied given prefix and suffix). The Database is flexible to be updated
specifically for certain domains such as news and broadcasting domains. Some examples of
semantic relations in RDI-ArabSemanticDB are given in Figure 1. Arabic Lexical Semantics
Database RDI toll can do four functions [38]; first, retrieving the Arabic words that fall under a
certain semantic field belongs to the category of specific knowledge of semantic fields previously
public; known as placement front. Second, attributing any word to one or more of the semantic
fields referred to in advance; known as posterior placement. Third, retrieving the semantic
relations between any two pairs of semantic fields. Forth, drawing semantic fields associated with
any semantic relationship with any particular semantic field.

5
Synonym
‫الديمقراطية‬,‫التشاور‬,‫التعددية‬,‫االنتخابية‬,........‫إلخ‬
Democracy, consultation, pluralism, electoral … etc.
Synonym
‫تحتوي‬ ‫التي‬ ‫المواضع‬ ‫كل‬ ‫متضمنة‬ ‫البحث‬ ‫نتيجة‬ ‫لتأتي‬ ‫بالمترادف‬ ‫البحث‬ ‫أختيار‬ ‫و‬ "‫"اإلرھاب‬ ‫كلمة‬
‫كلمة‬"‫اإلرھاب‬"‫يرادفھا‬ ‫ما‬ ‫أو‬"‫التطرف‬-....‫إلخ‬
The word "Terrorism" and select searching synonyms to come as a result
of search, including all positions that contain the word "terrorism" or
what synonyms "extremism -.... etc.
Include
‫اليوم‬ ‫ھذا‬ ‫يشمله‬ ‫ما‬ ‫مواضيع‬ ‫متضمنة‬ ‫البحث‬ ‫نتيجة‬ ‫لتأتي‬ ‫باالشتمال‬ ‫البحث‬ ‫أختيار‬ ‫و‬ "‫"اإلرھاب‬ ‫كلمة‬
‫من‬"‫التفجير‬-‫التدمير‬-‫التحريم‬-....‫إلخ‬
The word "terrorism" and select searching inclusion comes as a result of
the search, including what topics covered in this day of "bombing-
Destruction- prohibition -.... etc.
Figure1. Examples of semantic relation
4. BUILDING AN ARABIC SUBJECTIVE LEXICON (SENTIRDI)
SentiRDI is a Sentiment Arabic lexical semantic database-RDI which determines both
subjectivity and orientation for all semantic fields (18,413) and all words (150,000) which are
covered by Arabic Lexical Semantic database. The algorithm determines both subjectivity and
orientation for each semantic field. Our algorithm begins with chosen seed sets. Initially, we used
the seed list defined in D.Turney and L. Littman [39]. The seed list contained 14 words {good,
nice, excellent, positive, fortunate, correct, superior, bad, nasty, poor, negative, unfortunate,
wrong, inferior}.We translated them into Arabic and used the 14 translated words in the initial
run but it couldn't reach to all the semantic fields. The seed sets were extended by randomly
choosing new words from the semantic fields that were unreachable by the previous seed sets and
by adding these words to them. Also, we added an objective seed set in order to reach to all
semantic fields in Arabic lexical semantic database. Figure 2, presents the three seed sets of our
algorithm (positive, negative, and objective).
Positive Seed
Set (S_pos)
[ "‫األلفة‬", "‫التواضع‬", "‫الشجاعة‬", "‫الشھامة‬", "‫م‬َ‫الكر‬", "‫الجودة‬", "‫العطاء‬",
"‫الضحك‬", "‫الجمال‬", "‫الصدق‬", " ‫السھ‬‫ولة‬ ", "‫الذكاء‬", "‫االجتھاد‬", "‫التھذيب‬",
"‫األمل‬"]
["familiarity," "humility," "courage," "magnanimity", "generosity",
"quality", "tender", "laugh", "Beauty", "honesty", "easy,"
"intelligence," "diligence, "" polite, "" hope "]
Negative Seed
Set (S_neg)
["‫الخوف‬", "‫الفقر‬", "‫التمرد‬", "‫التشاؤم‬", "‫الضعف‬", " ‫الفس‬‫ا‬‫د‬ ", "‫الحزن‬", "‫الكره‬",
"‫البخل‬", "‫اإلبكاء‬", "‫التسول‬", "‫الذنب‬", "‫القبح‬", "‫الجھل‬", "‫الكدر‬"]
["Fear," "poverty," "rebellion," "pessimism," "weak," " Corruption,"
"sadness," "hatred," "greed," " crying," "beg," "guilt," "ugliness, "
" ignorance, "" Chagrin "].
Objective Seed
Set(S_obj)
[" "‫اآللة‬ , "‫األرض‬","‫الرجل‬","‫الصوت‬" ]
["Machine", "Earth", "man", "sound"].
Figure2. Positive, negative, and objective seed sets

6
The expansion algorithm is divided into two parts; first, the expansion algorithm of objective seed
set which explained in (Figure 3). Second, the expansion algorithm of positive and negative seed
sets which explained in (Figure 4).
Figure3. Expansion algorithm of objective seed set
Figure 3 explains the Procedure Expand_Objective_Seedset. The procedure takes objective seed
set (Figure 2), set of semantic relations were needed to expand the objective seed set, the Lexical
Semantic database RDI, and the Current_level variable which initialize by zero and incremented
by one in each iteration. Queue is initialized by semantic fields in objective seed set. Before
putting any semantic field in Obj_list, we ensure that it wasn't added before, set its level with the
value of Current_level, set its repetition value equal one. If the semantic field putted in Obj_list
before, its repetition value increment by one. All objective relations were applied over each new
semantic field in order to expand the set. The function stops when there isn't any new semantic
fields are added (queue is empty).
Figure4 explains the Procedure Expand_Pos_Neg_Seedset. The procedure takes positive and
negative seed sets (Figure 2), set of semantic relations were needed to expand two seed sets, the
Lexical Semantic database RDI, and the Current_level variable which initialize by zero and
incremented by one in each iteration. This procedure is the same as the pervious but it has two
differences; the initialization of two queues and types of semantic relations are needed to expand
positive and negative seed sets. Positive and negative queues are initialized by the synonyms of
words in positive and negative seed sets. For the relations; there are two types of relations;
relations are applied to the semantic field, the resulted semantic fields have the same orientation
{Causality, Causative, hyponymy, and Hypernym} and relations are applied to the semantic field,
the resulted semantic fields have the opposite orientation {Antonym}.
Function: Expand_Objective_Seedset
Input:
- S_obj: objective seed set
- Obj_Rel{ Totality, Part_of, Inclusion_K, KindOf Circumstantial_Place , Locality_Place,
Circumstantial_Time , and Locality_Time}
- Q: an object contains the Lexical Semantic database RDI.
- Current_level: number of current iteration ; initialized by 0
Output:
- Obj_list : list of objective semantic fields
Begin
For each Semantic field in S_obj
Add Semantic field in obj_ queue // initialize obj_queue
While (obj_queue is not empty)
Current_level Current_level+1
Cur_Node_obj remove the top node from obj_queue
If Cur_Node_obj(not found) in Obj_list //to ensure that the added node isn't visited
Cur_node_obj.level Current_level
Cur_node_obj.repetation 1
Add Cur_node_obj to Obj_list
Else
Cur_node_obj.repetation Cur_node_obj.repetation +1;
For each R in Obj_Rel search Q with Cur_node_obj to get Related_Nodes
For each N in Related_Nodes
If (N not found in obj_queue AND N not found in Obj_list)
Add N to Obj_queue
End

7
Figure4. Expansion algorithm of positive and negative seed sets
After more than hundred trials found that number of terms in seed set are thirty four (15 positive
terms, 15 negative terms and 4 objective terms) which cover 98.64% of semantic fields of
Database [40]. The rest of semantic fields (250) were annotated manually. Number of positive
Function: Expand_Pos_Neg_Seedset
Input:
- S_pos , S_neg: positive and negative seed set
- PN_Same_Rel { Causality, Causative, hyponymy, and Hypernym }
- PN_Opposite_Rel { Antonym}
- Q: an object contains the Lexical Semantic database RDI.
- Current_level: number of current iteration ; initialized by 0
Output:
- Pos_list : list of positive semantic fields
- Neg_list :list of negative semantic fields
Begin
For each Word in S_pos Get Synonyms
Add Synonyms in pos_ queue // initialize pos_queue
For each Word in S_neg Get Synonyms
Add Synonyms in neg_ queue // initialize neg_queue
While ((pos_queue is not empty) OR (Neg_queue is not empty) )
Current_level Current_level+1
Cur_Node_pos remove the top node from pos_queue
If Cur_Node_pos (not found) in Pos_list //to ensure that the added node isn't visited
Cur_node_pos.level Current_level
Cur_node_pos.repetation 1
Add Cur_node_pos to pos_list
Else
Cur_node_pos.repetation Cur_node_pos.repetation +1;
Cur_Node_neg remove the top node from neg_queue
If Cur_Node_neg( not found) in neg_list
Cur_node_neg.level Current_level
Cur_node_neg.repetation 1
Add Cur_node_neg to neg_list
Else
Cur_node_neg.repetation Cur_node_neg.repetation +1;
For each R in PN_Same_Rel search Q with Cur_node_pos to get Related_Nodes
If (N not found in pos_queue AND N not found in Pos_list)
Add N to pos_queue
For each R in PN_Opposite_Rel search Q with Cur_node_pos to get Related_Nodes
If (N not found in neg_queue AND N not found in Neg_list)
Add N to neg_queue
For each R in PN_Same_Rel search Q with Cur_node_neg to get Related_Nodes
If (N not found in neg_queue AND N not found in Neg_list)
Add N to neg_queue
For each R in PN_Opposite_Rel search Q with Cur_node_neg to get Related_Nodes
If (N not found in pos_queue AND N not found in Pos_list)
Add N to pos_queue
End

8
semantic fields are 3156, negative semantic fields are 4169 and objective semantic fields are
10,839.
4.1 Experiments analysis
The evaluation of semantic fields in SentiRDI is based on the exact match between identified
polarity of semantic fields and the manually annotated semantic fields according to three
properties (positive, negative, and objective). We used precision, recall and F-measure in
evaluation.
Testing SentiRDI is done by two ways; the first method by taking a subset of Arabic semantic
fields as a “gold standard” and annotated it manually. The subset contained 7612 semantic fields
was annotated manually by 5 annotators. The second method, by translating the Micro-WNOp
[41] (by Google translator) gold standard that is used to evaluate the SentiWordNet which
contains 1.105 synsets .
Table 1 shows the results of two tests that explained in the previous section. For each test, we
calculate Recall, Precision, and F_measure for all semantic fields (All_SF), objective semantic
fields (Obj_SF), negative semantic fields (Neg_SF), and positive semantic fields (Pos_SF) were
included in each test. The results of first test are better than the second one in terms of recall and
precision, and F-measure for all and objective semantic fields. The results of second tests are better
than the first in terms of Recall, Precision, and F-measure for negative semantic fields.
Table1. Testing SentiRDI Results
From results , we conclude that the direct translation from one language (English) to another
language (Arabic) doesn't give accurate results due to the drawbacks of machine translation
including the loss of polarity sentiments of some words when translated to other language (the
same term may have a lot of meaning with different polarities). Table 2 shows some terms that
are found in the Micro-WNOp [41] gold standard and give different meaning with different
polarities.
Table2. Samples in Micro-WNOp gold slandered.
Word
Different Meaning
Positive Negative Objective
Ball
‫نزھة‬a picnic
Shindig ‫راقصة‬ ‫حفلة‬
‫ـــــــ‬
a game of ball games ‫الكرة‬ ‫ألعاب‬ ‫من‬ ‫لعبة‬
‫الكرة‬Football
a bullet ‫رصاصة‬
Shark ‫ـــــ‬
Crook ‫نصاب‬
Swindler ‫محتال‬
Shark ‫قرش‬ ‫سمك‬
Sea dog ‫البحر‬ ‫كلب‬
Humble
Humble ‫متواضع‬
Meek ‫وديع‬
Lowly ‫وضيع‬
Servile ‫ذليل‬
‫ـــــ‬
Cure Healing ‫ﺷفاء‬ ‫ـــــــ‬
Curing fish ‫السمك‬ ‫تقديد‬
Salting meat ‫تمل‬‫اللحم‬ ‫يح‬
First Test Second Test
All_S
F
Obj_S
F
Neg_S
F
Pos_SF All_SF Obj_S
F
Neg_S
F
Pos_SF
Recall 87.32 89.3 83.16 86.44 84.67 89 85 79,43
Precision 87.72 93.09 79.95 74.76 84.95 87.93 82,45 80
F-
measure
87.52 91.16 81.52 80.18 84.81 88.46 83.7 79,71

9
5. SUBJECTIVITY CLASSIFIER
5.1 Polarity determination of An Arabic word using SentiRDI
The simplest case that input Arabic word (from text) may be exact matching with one of pre-
defined polarity of lexicon's semantic fields. However, in most cases , input word might be
derivational or fall under a certain semantic field ,so, Arabic RDI-ArabMorpho-POS tagger and
RDI-ArabSemanticDB tool were aiding to get stem , root, and synonyms (one or more semantic
field ) of the word in order to find the corresponding pre-defined polarity of lexicon semantic
fields. If the input word not found, we can classify it manually and add it to our lexicon.
Figure5. Determining polarity of an Arabic word algorithm
Figure5 explains how to determine the prior polarity of an input Arabic word using SentiRDI.
The procedure takes two inputs word and SentiRDI. First, stem and root of input word are gotten
by using RDI tools. Search for Word, root and stem in SentiRDI if found set the polarity of word
with the founded polarity then exit (step 2). Else, get the synonyms of the word and the polarity
of each one. Count number of positive synonyms (P), number of negative synonyms (N), and
Function: determining polarity of An Arabic Word
Input:
- Word
- Subjective lexicon (SentiRDI)
Output:
- Pol_word : word polarity (positive, negative, objective)
Begin
1- Inputs are word, its stem and its root (get stem and root using RDI which uses A*A*A*A* search
algorithm to find the correct root for the word and perform its morphological analysis
with 96.8% accuracy).
2- Search for the inputs in the SentiRDI, if it is found then get its polarity (pol_word) and
exit.
3- Else get the synonyms of the input word and the polarity of each one.
4- Get number of positive synonyms (P), number of Negative synonyms (N) and number of
objective synonyms (O).
5- Get M=max( P,N,O) // Get maximum of three number
{
if (M=P) then pol_word= positive
if (M=N) then pol_word= negative
if (M=O) then pol_word= objective
}
else if (N=P) then pol_word= objective
else if (N=O) then pol_word= negative
else if (P=O) then pol_word= positive
6- If the word couldn't be found in semantic database in this case we determine its
polarity manually and add it with its set polarity to SentiRDI.
End

10
number of objective synonyms (O) (step 4) then get the maximum of the three numbers and set it
to (M). Set the polarity of input word with the polarity of (M) (step 5). If the input word isn't
covered by the semantic database, in this case set the polarity of word manually and add it to
SentiRDI.
5.2 Corpus
Corpus1
is Translated MPQA corpus provided in [42] containing subjective and objective
sentences of Arabic language .for English, MPQA corpus2
consisting of 535 English-language
news articles from 187 different foreign and U.S., manually annotated for subjectivity [32]. The
corpus consists of 9700sentences, 55% of them are labelled as subjective, while the rest are
objective.
5.3 Text Pre-processing
Simple text pre-processing was executed in order to remove special characters and non Arabic
characters. More advanced text pre-processing was executed in order to prepare it for input to
different learning algorithm.
• First step was extracting names from documents: this step reduced calculations for
determining the prior polarity for tokens because all named entities are objective.
Technique used in this step was Arabic Named Entity Recognition (ANER) [43] which
was an integration approach between two machine learning techniques, namely
bootstrapping semi-supervised pattern recognition and Conditional Random Fields (CRF)
classifier as a supervised technique. This technique extracted approximately 59,772
named entity from subjective corpus and 50,787 named entity from objective corpus.
• The Second step that was assigning Part Of Speech tags (POS). The tool that used to
assign POS tags was the Research and Development International (RDI) tool.
• Third step to obtain the prior polarity of words by using SentiRDI (Section 5.1).
Table3. Statistics from the MPQA Arabic opinion corpus
1 http://www.cse.unt.edu/~rada/downloads.html#msa
2 http://www.cs.pitt.edu/mpqa
5.4 Features selection
Feature execration involves extracting tokens that are relevant to detecting sentiment in the
sentence. The features proposed here were similar as in [44]:-
1. POS is the Part of Speech (POS) tagging of the word. RDI-ArabMorpho-POS tagger was
used [38] and by using Prior polarity lexicon (SentiRDI) for determining polarity of each
word.
Subjective corpus Objective corpus
Total tokens 129.169 90,357
Total sentences 5335 4365
Total NER 59772 50787

11
1.1 Number of positive noun.
1.2 Number of positive verb.
1.3 Number of negative noun.
1.4 Number of negative verb.
2. Average Polarity of sentence =
Where n is number of words in sentence, Pwi polarity of word i in sentence that is specified
before from prior polarity database (SentiRDI) such that
3. Average Term Frequency: Inverse Sentence Frequency (TF-ISF) for sentence (Si) can be
computed by the following equation:-
Where TF presents the number of occurrences of each term within the sentence and can be
normalized by dividing it by size of sentence.
Where Nt,s is the number of occurrences of term t in sentence S. ||S|| is the number of words in
sentence S. ISF is used for terms that appear in the small number of sentences. This factor is
useful because numbers of subjective terms are small compared with neutral (objective) ones.
6. EXPERIMENTAL ANALYSIS
We use more than thirty classifiers with different types (Bayes, Rules, Tress and Function). The
experimentation was executed using 10 Cross-Fold-Validation with two prior polarity subjective
lexicons (MPQA subjectivity lexicon which is proposed by [19] and SentiRDI Lexicon).
The evaluation is based on the exact match between identified subjectivity sentence and the
manually annotated true subjective sentence. We used Precision, Recall and F-measure to
evaluate.
Subjectivity classifiers results by using SentiRDI and MPQA subjectivity lexicon is presented in
Table 4. From our experiment we find that most classifiers give better results with SentiRDI

12
except Naïve Bayes that gives higher precision with MPQA subjective lexicon. FT classifier
gives the best result with selective features that is give accuracy 76.1% with using SentiRDI as a
prior polarity subjectivity lexicon.
Table4. Subjectivity classifiers results
7. CONCLUSION AND FUTURE WORK
This paper investigates the method for generating Sentiment Arabic lexical Semantic Database by
using dictionary based approach. Also, we study the prior polarity effects of each word using our
Sentiment Arabic Lexical Semantic Database on the sentence-level subjectivity and multiple
machine learning algorithms. Our main contribution is to acquire the sentiment Arabic lexical
semantic database (SentiRDI) having very large number of Arabic words with different
derivational forms and part of speech tags. The created lexicon is context independent and it can
be used in any opinion corpora.
In the future, we are going to extend the database depending on further analysis of exiting opinion
mining Arabic corpora. Also, we are going to try more classifiers to improve results more than
this and apply contextual polarity to get good examples can be added to our lexicon and improve
the classification results.
REFERENCES
[1] Minqing Hu and Bing Liu 2004." Mining opinion features in customer reviews," Proceedings of the
National Conference on Artificial Intelligence, 755–760.
[2] Kavita Ganesan , ChengXiang Zhai and Jiawei Han,2010. "Opinosis:a graph-based approach to
abstractive summarization of highly redundant opinions," Proceedings of the 23rd International
Conference on Computational Linguistics,340–348
[3] Alexandra Balahur Ester Boldrini , Andrés Montoyo and Patricio Martínez-Barco, 2009. "Opinion
and Generic Question Answering systems: a performance analysis," Proceedings of the ACL-IJCNLP
2009 Conference Short Papers, 157–160.
[4] M. Taboada, J. Brooke, M. Tofiloski, K. Voll, and M. Stede, "Lexicon-based methods for sentiment
analysis," Computational linguistics 37, no. 2, pp. 267-307, 2011.
[5] S. R. Das and M. Y. Chen, “Yahoo! for Amazon: Sentiment extraction from small talk on the Web,”
Management Science, vol. 53, pp. 1375–1388, 2007.
Classifier Name
MPQA Subjectivity
Lexicon
SentiRDI
NaiveBayes
F_measure 60.8 70.55
Precision 71.9 70.6
recall 65.5 70.5
Ridor
F_measure 71.2 73.82
Precision 71.7 75.4
recall 71.8 72.3
FT
F_measure 72.4 76.1
Precision 72.5 76.2
recall 72.6 76
LMT F_measure 72.4 76
Precision 72.6 76
recall 72.6 76

13
[6] S. Morinaga, K. Yamanishi, K. Tateishi, and T. Fukushima, “Mining product reputations on the
Web,” Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining
(KDD), pp. 341–349, 2002.
[7] R. M. Tong, “An operational system for detecting and tracking opinions in on-line
discussion.”Proceedings of the Workshop on Operational Text Classification (OTC), 2001.
[8] YI, J., NASUKAWA, T., BUNESCU, R. AND NIBLACK, W. 2003. Sentiment analyzer: Extracting
sentiments about a given topic using natural language processing techniques, In Proceedings of the
3rd IEEE International Conference on Data Mining, 427-434.
[9] C. Fellbaum, ed., Wordnet: An Electronic Lexical Database. MIT Press, 1998.
[10] B. Liu, "Sentiment Analysis and Subjectivity", Handbook of Natural Language Processing. Second
Edition, (editors: N. Indurkhya and F. J. Damerau), 2010.
[11] Pang Bo, Lee Lillian, and Vaithyanathan Shivakumar,2002. "Thumbs up? Sentiment classification
using machine learning techniques" In Proceedings of EMNLP, pages 79–86.
[12] Janyce Wiebe and Rada Mihalcea , 2006. "Word sense and subjectivity," In Proceedings of
COLING/ACL-06. Pages 1065--1072.
[13] Esuli Andrea and Sebastiani Fabrizio,2006. SentiWordNet:" A publicly available lexical resource for
opinion mining," In Proceedings fLREC.
[14] Wilson Theresa, Wiebe Janyce and Hoffmann Paul ,2005. "Recognizing Contextual Polarity in
Phrase-Level Sentiment Analysis," In Proceedings of HLT/EMNLP 2005, Vancouver,Canada.
[15] Carlo Strapparava and Alessandro Valitutti ,2004. "WordNet-Affect: an affective extension of
WordNet," In Proceedings of LREC 2004, pages 1083 – 1086, Lisbon, May .
[16] Kimberly Voll and Maite Taboada., 2007". Not All Words are Created Equal: Extracting
SemanticOrientation as a Function of Adjective elevanc," In Proceedings of the 20th Australian Joint
Conference on Artificial Intelligence. pages. 337-346.
[17] Amitava Das and Sivaji Bandyopadhyay,2010. "Towards The Global SentiWordNet," In the
Workshop on Model and Measurement of Meaning (M3), PACLIC 24, November 4, Sendai, Japan.
[18] Elhawary, M., and Elfeky, M. (2010). “Mining Arabic Business Reviews.” IEEE International
Conference on Data Mining Workshops
[19] M. Elarnaoty, S. AbdelRahman, and A. Fahmy. A Machine Learning Approach For Opinion Holder
Extraction Arabic Language. CoRR, abs/1206.1011, 2012.
[20] Muhammad Abdul-Mageed, Mona Diab, 2012. Toward building a large-scale Arabic sentiment
lexicon, Proceedings of the 6th International Global WordNet Conference.
[21] M. Abdul-Mageed, M. Korayem, and A. YoussefAgha. “Yes we can?”: Subjectivity annotation and
tagging for the health domain. In Proceedings of the International Confrence Recent Advances in
Natural Language Processing RANLP, Hissar, Bulgaria, 2011.
[22] Samhaa R. El-Beltagym, Ahmed Ali, 2013. Open issues in the sentiment analysis of Arabic social
media: a case study, Proceedings of 9th International Conference on Innovations in Information
Technology (IIT), pp. 215–220.
[23] Mahyoub, F. H., Siddiqui, M. A., & Dahab, M. Y. (2014). Building an Arabic Sentiment Lexicon
Using Semi-Supervised Learning. Journal of King Saud University-Computer and Information
Sciences.
[24] Vasileios Hatzivassiloglou and Janyce M. Wiebe, 2000." Effects of adjective orientation and
gradability on sentence subjectivity," in Proceedings of the International Conference on
Computational Linguistics (COLING).
[25] Janyce Wiebe and Theresa Wilson,2002 . "Learning to disambiguate potentially subjective
expressions," in Proceedings of the Conference on Natural Language Learning (CoNLL), pp. 112–
118.
[26] Hong Yu and Vasileios Hatzivassiloglou, 2003. "Towards answering opinion questions: Separating
facts from opinions and identifying the polarity of opinion sentences,” in Proceedings of the
Conference on Empirical Methods in Natural Language Processing (EMNLP).
[27] Ellen Riloff and Janyce Wiebe ,2003. "Learning extraction patterns for subjective expressions, "in
Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP).
[28] Pang, B. , and Lee, L. ,2004. "A sentimental education: Sentiment analysis using subjectivity
summarization based on minimum cuts," in Proceedings of the Association for Computational
Linguistics (ACL), pp. 271–278.
[29] Kim, S.M. , and Hovy, E. ,2005. "Automatic detection of opinion bearing words and sentences, "in
Companion Volume to the Proceedings of the International Joint Conference on Natural Language
Processing (IJCNLP).

14
[30] Kaji N. , and Kitsuregawa, M., 2006. "Automatic construction of polari-tagged corpus from HTML
documents," in Proceedings of the COLING/ACL Main Conference Poster Sessions.
[31] Wilson , T. , Wiebe, J. , and Hwa, R. ,2006. "Just how mad are you? Finding strong and weak opinion
clauses," in Proceedings of AAAI, pp. 761–769, 2004. (Extended version in Computational
Intelligence, vol. 22, no. 2, pp. 73–99.
[32] Breck, E. , Choi, Y. , and Cardie, C. ,2007. "Identifying expressions of opinion in context," in
Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI), Hyderabad,
India.
[33] Janyce Wiebe and Ellen Riloff, 2005. "Creating Subjective and Objective Sentence Classifiers from
Unannotated Texts," In Proceeding of CICLing-05, International Conference on Intelligent Text
Processing and Computational Linguistics, pages 486–497, Mexico City, Mexico.
[34] Riloff, E. , and Wiebe, J. , and Wilson, T., 2003. "Learning subjective nouns using extraction pattern
bootstrapping, "in Proceedings of the Conference on Natural Language Learning (CoNLL), pp. 25–
32.
[35] Wiebe, J. , Wilson, T. , Bruce, R. , Bell, M. , and Martin, M. ,2004. "Learning subjective language,
"Computational Linguistics, vol. 30, pp. 277–308.
[36] M. Abdul-Mageed and M. T. Diab. Subjectivity and sentiment annotation of modern standard arabic
newswire. In Proceedings of the 5th Linguistic Annotation Workshop, LAW V ’11,pages 110–118,
2011.
[37] M.Attia, and M.Rashwan (2004). “A Large Scale Arabic POS Tagger Based on a Compact Arabic
POS Tag Set and Application on the Statistical Inference of Syntactic Diacritics of Arabic Text
Words”, NEMLAR.
[38] M.Attia, M.Rashwan, A.Ragheb, M.A.Al-Basoumy, and S.Abdou, 2008."A Compact Arabic Lexical
Semantics Language Resource Based on the Theory of Semantic Fields ," Proceedings of the 6th
international conference on Advances in Natural Language Processing.
[39] Peter D. Turney, Michael L. Littman, 2002. Unsupervised learning of semantic orientation from a
hundred-billion-word corpus, Technical Report EGB-1094, National Research Council Canada.
[40] Hanaa B. Mobarz, , Mohsen Rashwan, and Samir AbdelRahman, 2011, "Generating lexical
Resources for Opinion Mining in Arabic language automatically," The Eleventh Conference on
Language Engineering ESOLEC', Cairo-Egypt, http://esole-eg.org/index.php/en/conferences, Sept.
[41] Cerini, S., Compagnoni, V., Demontis, A., Formentelli, M., and Gandini, G. ,2007. Language
resources and linguistic theory: Typology, second language acquisition, English linguistics
(Forthcoming), chapter Micro-WNOp: "A gold standard for the evaluation of automatically compiled
lexical resources for opinion mining,"Franco Angeli Editore, Milano, IT.
[42] J. Wiebe and E. Riloff, 2005. "Creating Subjective and Objective Sentence Classifiers from
Unannotated Texts," In Proceeding of CICLing-05, International Conference on Intelligent Text
Processing and Computational Linguistics, pages 486–497, Mexico City, Mexico.
[43] Samir AbdelRahman, Mohamed Elarnaoty, Marwa Magdy and Aly Fahmy, 2010. "Integrated
Machine Learning Techniques for Arabic Named Entity Recognition, "IJCSI International Journal of
Computer Science, pp. 1694-0784.
[44] Samir AbdelRahman ,Hanaa B. Mobarz, , Mohsen Rashwan, and Ibrahim Farg, 2014, " Arabic
Phrase-Level Contextual Polarity Recognition to Enhance Sentiment Arabic Lexical Semantic
Database Generation," IJACSC International Journal of Advanced Research in Artificial Intelligence,
Vol 5,No.10.

USING AUTOMATED LEXICAL RESOURCES IN ARABIC SENTENCE SUBJECTIVITY

More Related Content

What's hot (19)

Viewers also liked (8)

Similar to USING AUTOMATED LEXICAL RESOURCES IN ARABIC SENTENCE SUBJECTIVITY (20)

Recently uploaded (20)

USING AUTOMATED LEXICAL RESOURCES IN ARABIC SENTENCE SUBJECTIVITY