International Journal of Multidisciplinary Sciences and Advanced Technology

ISSN xxxx-xxx Volume 1, NO 2 (2020) pp. 167-184

©. http://www.ijmsat.com

Valence Shifter Rules for Arabic Sentiment Analysis

B.Ihnaini1, M. Mahmuddin2
1
Faculty f Business Administration, Department of Business Information Technology,
Emirates College of Technology (ECT), Abu Dhabi, United Arab of Emirates
E-mail: baha.ihnaini@ect.ac.ae
2
School of Computing, College of Art & Sciences, Universiti Utara Malaysia (UUM),
Sintok, Kedah, Malaysia
E-mail: ady@ahsgs.uum.edu.my

Abstract. Sentiment analysis refers as computational and natural language processing techniques used to
extract subjective information expressed in a text. In this sentiment analysis study, three main valence
shifter words are addressed: a) Negation, b) Intensification, and c) Contrast words. Handling valence shifter
words were not thoroughly addressed in Arabic sentiment analysis. Therefore, this paper aims to construct
a new valence shifter rules in order to enhance the performance of lexicon-based sentiment analysis on
Arabic text. To achieve the objective a set of rules are proposed to handle the valence shifters words by
applying rules to find the scope of words, and shifting value that is produced by these words. Several
datasets and lexicons from the literature are used to evaluate the proposed idea. Experimental results
indicate that applying the proposed valence shifter rules have increased overall performance of 5% on
average. Taking into consideration that the percentage of texts that include valence shifter words is less
than 18%, it is clear that the 5% of enhancement is a significant improvement of lexicon-based sentiment
analysis on Arabic text.

Keywords: (Arabic Sentiment Analysis), (Lexicon-Based Approach), (Valence Shifters).

1. Introduction

Sentiment analysis refers as computational and natural language processing techniques used to
extract subjective information expressed in a text. Sentiment analysis approaches can be classified into
either lexicon-based or machine learning. Applying the lexicon-based approach, which mainly implies the
sentiment by using the term`s sentiment value in the text, needs a set of rules called valence shifters to
handle phrases in order to improve the accuracy. Valence shifters are terms that change the sentimental
orientation of other terms such as negations, intensifiers, and contrasts. Negations are used to reverse the
semantic polarity of a particular term, while intensifiers are used to increase or decrease the degree to
which a term is positive or negative (Kennedy & Inkpen, 2006), whereas contrast is the mechanism in a
language which joins two or more smaller units with opposite properties into a bigger unit.

One important mutual feature in sentiment analysis and computational linguistics is negation. Little
work has been undertaken in Arabic in order to address the issue of negation, either in the negation
detection problem itself or the effect of negation in sentiment analysis (Alotaibi, 2015; Awwad &
Alpkocak, 2016; Abdul-Mageed, 2017; Ihnaini & Mahmuddin 2018; Ghallab & Mohsen 2020). Some of
the works deal with the negation by touching the basic idea during the sentiment analysis process by
flipping the negated words value (Duwairi, Ahmed, & Al-Rifai, 2015; Mataoui, 2016), or they are just
counted and added to the total score (Ibrahim et al., 2015; Al-Twairesh et al., 2016). Few researchers
considered the scope of the negation word, and the shifting values for these negation words (Assiri et al.,
2017). Besides, there are also other types of valence shifters that have not been considered thoroughly for
Arabic sentiment analysis, such as intensification words, and contrast words. Aimed at intensification,

167
 Valence Shifter Rules for Arabic Sentiment Analysis

some issues related to the scope of the intensifier words were located, since unlike English, in Arabic, an
intensifier word could intensify a word not next to it. While contrast words have not been handled by
Arabic researchers. Therefore, it is absolutely essential to further study the valence shifters in Arabic
sentiment analysis, and come up with suitable solutions. Hence, handling valence shifters is the objective of
this research paper.

The rest of the paper is organized as follows: Section 2 presents the Literature Review, Experimental
Design is presented in Section 3, while the Results and Evaluation are illustrated in Section 4. Conclusion
is presented in Section 5.

Valence shifter words are used to change sentiment, either as negation, intensification, or
contrast. In the following sections, each type will be discussed thoroughly from the literature.
2.1 Negation Words

In sentiment analysis, handling negation is very important, because one negation word could reverse
the sentence polarity from positive to negative, and vice versa. Following is two similar sentences with a
different sentiment, “this movie is good/ ‫”هذا الفيلم جيد‬, and “this movie is not good/ ‫”هذا الفيلم ليس جيد‬. There
are many types of negation that could appear in sentences. Direct negation where negation word followed
directly with the negated word, for example, “not good/‫”ليس جيد‬. Long distance negation where the negation
word and negated words are not neighbors, for example, “not very beautiful/ ‫ ”مش كثير حلو‬or “does not have
good engine/‫”ال تحتوي على محرك جيد‬.

Some studies in Arabic has been done for solving the negation problem, whether detecting the
negation issue, or the results of using negation in this analysis. Negation principle was put in the studies of
Elhawary and Elfeky (2010), by measuring the sentiment score in each word or phrase depending on the
Arabic lexicon. In this task, they applied the inversing to the negated words. However, there was a couple
of issues in it. The first is stating the use of twenty negation words but without mentioning it. The second is
the way of selecting the negated words or phrases in each sentence. These problems can have an influence
on sentiment analysis procedure because polarity might be changing.

Negation was taken into consideration by Farra, Challita, Assi, and Hajj (2010) during their
experience of finding the Arabic text's sentiment. Their research has dealt with the negation manner by just
calculating negation words' frequency in a sentence while trying to create the sentence's semantic feature
relying on Arabic sentiment lexicon. Features that have been used in their study are the frequency positive,
negative, and neutral words. Researchers have not considered how negation words in the sentence effect on
other words. Therefore, a decrease in accuracy occurred comparing with other studies. Moreover, they did
not mention the negation words used in their study. Furthermore, depending on the simple representation,
could not measure all sentence's semantics and syntax, which can be helpful for sentiment classification.
Hamouda and El-taher (2013) created a sentiment analyzer for the Arabic new pages' comments on
Facebook. They did some comparison between different features and machine learning algorithms. For
example, they dealt with Arabic negation, and even though there are many negation words, they only
calculated five of them and did not count them in dialects. Just a negation words' percentage of comment or
post was added as a feature, and they did not take into consideration the negation influence on words. But
the main problem of this process is that it is only appropriate for the chosen domain by them, which is
comments and posts on news pages on Facebook.

Many other studies adopted the switch negation technique to handle negation, which is reversing the
next word to the negation word to an opposite value (Al-Kabi, Al-Qudah, Alsmadi, Dabour, & Wahsheh,
2013; Medhat, Hassan, & Korashy, 2014b; Refaee & Rieser, 2014; Narayanan, Arora, & Bhatia, 2013; Pak

168
 Valence Shifter Rules for Arabic Sentiment Analysis

& Paroubek, 2010; Mataoui, 2016), which fails when polarity word is not next to the negation word, such
as in “ ‫ ”ليس قرار جيد‬where the middle word is non-polarity word means decision, and the first is a negation
word, while the last is a polarity word means good, hence, the result should be negative in this sentence.

In another simple handling of negation, Al-Horaibi and Khan (2016) created a list of negation words
that are used frequently in the Arabic language; when such negation words are found in the classifier, a
score of (negative = 1) is given to the word in the list. While another study multiplies the weight of the
negation word by the sentiment orientation of the next word, and then add it to the total sentiment
orientation.

The negation was handled in Al-Twairesh et al. (2016) by checking if the positive or negative word
is along with the negation word. If it is positive, then no increment to the counter of positive words. But if
the word is negative, then this time do the increment to the counter of negative words. Besides, considering
the negation word`s score to be -1. Using this technique, the performance increased by (1% to 4%) across
all datasets.

Four rules to handle negation were proposed in Assiri et al. (2017). First, to switch the word with
polarity after the negation word. Second, only switch the first polarity word even if it appears two or three
words after the negation word. Third, if the term “‫ ال‬/no” is followed by a non-polarity word first then a
polarity word, it makes no change. Finally, if a negation word is followed by non-polarity words, then it is
considered as a negative tweet such as in “not a dialogue style”. These rules increased performance by 3%.
2.2 Intensification Words

Many researchers handled negation, but less handled intensifiers, Assiri et al. (2017), or discussed
them but without handling it (Duwairi, Ahmed, & Al-Rifai, 2015). Intensifiers refer to words that change
the sentiment of the neighboring non-neutral terms. They can be divided into two categories, namely
amplifiers, such as “‫جدا‬/very”, and downtoners, such as “‫قليل‬/little” that increase and decrease the intensity
of sentiment, respectively. The type of the lexicon, weather polarities are binary or the polarities as a score
restricts the way of handling intensification. In the work of Abdulla, Majdalawi, Mohammed, Al-Ayyoub,
and Al-Kabi (2014), they handled intensification by adding one if the type of intensifier is amplifier, for
example, “‫جيد جدا‬/very good” will have a value of (1+1), by deducting one if the type of intensifier is
downtoner, for example, “‫جيد قليال‬/slightly good” will have a value of (1-1). The usage of binary lexicon
forces them to use this method.

While El-Beltagy and Ali (2013) multiply the weight of the intensification by the weight of the
following word. Also, Al-Aziz, Gheith, and Eldin (2016) adopted the multiplication method to handle
intensification words. While Shoukry and Rafea (2012) didn`t handle the intensification, because of the
absence of the intensifiers lists for Egyptian dialect. Mataoui (2016) built a lexicon of intensifier words
from MSA and then added the equivalent from the Algerian dialect but without mentioning how he used
this lexicon.

In the work of Ibrahim et al. (2015), they only handled the amplifier words, by doubling the next
term to such word, whether the next word is positive or negative. Whereas Eskander and Rambow (2015)
created a list of amplifiers and downtoners by examining AraMorph (Buckwalter, 2002) glosses manually,
then according to the type of this term, they double or halve the score of the next term.

2.3 Contrast Words

Contrast is a linguistic mechanism that joins two smaller units having opposite characteristics into
one bigger unit, such as the word “‫لكن‬/but” and “‫رغم‬/although”. Despite its importance in determining the

169
 Valence Shifter Rules for Arabic Sentiment Analysis

overall polarity of a sentence, but nobody handled this kind of valence shifters for Arabic sentiment
analysis.

In English sentiment analysis, this kind of words has been used for expanding a seed of words, in
such a manner that if two words are joined by a contrast word, then both have opposite polarities
(Hatzivassiloglou & McKeown, 1997). This has been used to determine the dominance sentence of two
sentences linked by contrast word (Pang, Lee, & Vaithyanathan, 2002; Heerschop et al., 2011).

The work presented by Yuan (2016) described handling contrast words by dividing the sentence
before and after the contrast word to the main sentence and secondary sentence, and the overall polarity
will be the main sentence if it is not an objective sentence, otherwise, the overall polarity will be opposite
of the secondary sentence.

3. Research Methodology and Experimental Design

In this research paper, lexicon-based approach is adopted to conduct sentiment analysis on

Arabic text. The proposed experimental design starts with data collection, followed by text pre-
processing before polarity extraction from sentiment lexicons is conducted, then, applying
valence shifter rules, and finally, evaluation measurement. Based on these components, a series of
steps were outlined to form the experimental design as shown in Figure 1.

Data Collection

Pre-processing

Polarity Sentime
Extraction nt
Lexicons
Valence Valence Shifter Design
Figure 1: The Experimental
Shifter Rules
Lexicons

Performance
Evaluation

Several steps included in the lexicon-based experimental design. Data collection is the first
step in all approaches of sentiment analysis. For the purpose of evaluation, numerous datasets
have been used in this research from the literature in order to reduce the effect of the dataset on
the proposed valence shifter rules (Ihnaini & Mahmuddin, 2018). Namely, ArSenTD-LEV (Baly
et al., 2018), ASTD (Nabil, 2015), RR (Refaee & Rieser, 2014), AraSenti-Tweet (Al-Twairesh et

170
 Valence Shifter Rules for Arabic Sentiment Analysis

al., 2016). These datasets have been pre-processed in five different stages, data cleaning,
tokenization, stop words removal, normalization, and stemming.

Moreover, as shown in Figure 1, sentiment lexicons are required to perform lexicon-based

experiments. To eliminate the effect of the lexicon type on the proposed valence shifter rules
many lexicons from the literature have been selected to conduct the experiments. namely, AEL,
AHL, and DAHL by Mohammad et al., (2016a), NileULex by El-Beltagy (2016), ArSenL by
Badaro et al., (2014), SLSA by Eskander et al., (2015), AraSenti-Trans, and AraSenti-PMI by Al-
Twairesh et al., (2016), and Arabic Senti-Lexicon by Al-Moslmi et al., (2017). Furthermore,
lexicons that contain the valence shifter words are required. This type of lexicons is essential to
perform linguistic rules. There are three types of valence shifter lexicons, contrasts, intensifiers,
and negation. In order to build such lexicon, a manual process of collecting the Arabic valance
shifter words is held, and also these are manually classified into three types:

1. Negation words, for example; “‫ليس‬/not”, “‫مش‬/no”.

2. Intensifier words, which include two types, amplifier words, for example; “‫كثير‬/ a lot”,
and downtoner words, for example, “‫قليل‬/a little”.
3. Contrast words, which include two types, C1, where the main sentence is before the
contrast word. For example, “‫بالرغم‬/although”, and C2, where the main sentence is after
the contrast word. For example, “‫لكن‬/but”.

3.1 Construction of Valence Shifters Rules

Once one or more valence shifter words appear in the text, a set of proposed rules to be applied to
handle the presence of these words. These rules are implemented in the following order as shown in Figure
2. Next sections will explain in details each of the proposed rules to handle such words in order to enhance
the process of Arabic sentiment analysis.

Apply contrast rules if Apply negation rules if Apply intensifier rules if

contrast word is found negation word is found intensifier word is found

Figure 2: Order of Valence Shifter Rules.

171
 Valence Shifter Rules for Arabic Sentiment Analysis

3.1.1 Contrast Rules

Dealing with contrasts and annotating the sentence depends on the type of contrast word. There
are two types of contrast words, C1 and C2, where C1 indicates that the sentence is divided into two parts,
before C1 is the main sentence, while after C1 is the secondary sentence. An examples of contrast word of
type C1 is {“‫ رغم ان‬/ Although”, “‫ حتى‬/ Even”, “‫ حتى وان‬/Even if”, “‫ بدل‬/Instead of”, “‫ عالرغم‬/ Despite”, “ ‫على‬
‫ الرغم‬/Even though”, “‫ مع اني‬/Although I”, “‫ مع ان‬/ In spite”, “‫ مع انو‬/ In spite of the fact that”}. While when
using contrast word of type C2, the main sentence is after the contrast word, and the secondary sentence is
before the contrast word. An example of contrast word of type C2 is {“‫ بل‬/Yet”, “‫ مع ذلك‬/Still”, “‫ بالعكس‬/ On
the contrary”, “‫ لكن‬/But”, “‫ بس‬/But”, “‫ اال انه‬/Though”, “‫ اال اذا‬/Unless”, “‫ ورغم‬/Even that”, “‫رغم ذلك‬,
/Nonetheless”}.
Let C1 be a list of contrast words of type C1 and C2 be a list of contrast words of type C2 defined as:
C1 = {Set of contrast words of type C1}
C2 = {Set of contrast words of type C2}
If a word is found in the C1 list, then the polarity score of the opinion word after it is divided by ,
and if a word is found in the C2 list, then the polarity score of the opinion word before it is divided by as
follows:

(1)

where w is opinion words in a text, wb is words before contrast word c, wa is words after contrast
word c, α is a positive number greater than 1, and pol(w) is the polarity of word w in the presence of a
contrast word.
Example 1: ‫فيلم رائع رغم المقطع الصوتي السيء‬/Great movie despite the bad soundtrack.
Example 2: ‫مقطع صوتي رائع ولكن الفيلم سيء‬/Great soundtrack but the movie is bad.
In Example 1, the author`s opinion is considered positive, since the used contrast “despite” is from
C1. While the author`s opinion in Example 2 is considered negative since the used contrast “but” is from
C2. Therefore, a lower weight is given to the secondary sentence by dividing its polarity by α to make sure
the final result is mainly from the main sentence when summing both of the secondary and main sentences.

3.1.2 Negation Rules

First of all, negation words are collected from both MSA, such as “‫ ”ليس‬and from DA such as
“‫”مش‬. Then, if a negation word appears in the text to be classified, two main subjects are considered to
form the rule of negation. Finding the scope of the negation word is the first subject, and the other is to find
the shifting value for each word in the scope. When a negation word appears in a text, the scope of this
negation word is found first. To find the scope there is a set of rules to follow. First, the scope will not
exceed a fixed number of words which is called window size. Second, the first word is always included in
the scope. Finally, the next word is included if its polarity is not opposite to the polarity of the previous
word.
In order to find the optimal window size of the negation’s scope, several experiments were
conducted. Each Lexicon-based experiment is conducted on a dataset formed by collecting negated text and
using different window sizes of negation each time, by using a window size of 2 words, until the window
size of 8 words. Table 1 shows the accuracies for each window size, where the accuracy is found using
Equation 3.2.

172
 Valence Shifter Rules for Arabic Sentiment Analysis

Table 1. Finding Window Size of the Negation`s Scope

Window Size Accuracy Rate
Window size of 2 words 69.06%
Window size of 3 words 70.59%
Window size of 4 words 67.56%
Window size of 5 words 66.94%
Window size of 6 words 65.63%
Window size of 7 words 65.22%
Window size of 8 words 64.85%

From Table 1, it is noticed that using the window size of 3 words produces the highest accuracy.
Therefore, this window size is adopted for the proposed negation technique. Equation 2 shows how to find
the negation`s scope using a window size of 3 words.

Let Neg be a list of negation words defined as:

Neg = {Set of negation words}

(2)

where pol(w+1) is the polarity of the first word after negation word, pol(w+2) is the polarity of the second
word after negation word, pol(w+3) is the polarity of the third word after negation word, and Sn is the
scope of the negation word.

To find the effect of negation words on their scope, it is noticed that the effect of negation on
positive polarities is different from negating negative polarities. Besides, shifting values are higher when
dealing with stronger polarities. In order to find these shifting values, two sources are used, first is from
hundreds of multi-words that have been formed when translating the English lexicons to Arabic while as in
Table 2. Such as the word “fearless”, when translated to Arabic it forms an expression of two words, one is
the negation prefix “‫بدون‬/less”, and the other word is “‫خوف‬/fear”, and from the English lexicon, the polarity
score of the word “fearless” is “+2”, and the polarity score of the word “fear” is “-3”. Therefore, the effect
of the negation word “‫بدون‬/without” is “+5”, and this is done for all words with +2, and the average of them
is considered their shifting value. The same is applied for the rest of the polarities between [-5, 5].

Table 2. Examples on Polarity Shifting of Negative Prefixes Words

Word with Negative
Word Polarity Polarity
Prefixes
Ethical ‫أخالقي‬ +4 Unethical ‫غير أخالقي‬ -2

Like ‫يعجب‬ +2 Dislike ‫ال يعجب‬ -1

Legal ‫قانوني‬ +1 Illegal ‫غير قانوني‬ -2
Interrupted -1 Uninterrupted ‫غير‬ +1
‫متقطع‬ ‫متقطع‬
Selfish ‫أناني‬ -3 Unselfish ‫غير أناني‬ +2

173
 Valence Shifter Rules for Arabic Sentiment Analysis

Through the above mentioned process, it is found that shifting value of the words with polarities
between [-5, -2] are almost the same, and so for the words with polarities between [-2, 0], [0, 2], and [2, 5].
Table 2 shows some examples of words with and without these negative prefixes along with the polarity for
each.
The second source is the manually collected words which are more than 100 words gathered from
the given lexicons for each zone of these four zones, and the negation effect on these words was manually
annotated by the two Arabic linguists. Some examples of these polarity words with its polarity scores from
the given lexicons and the manually annotated polarities of these words when negated by a negation word
are presented in Table 3.
Table 3. Examples of Polarity Shifting by Negation Word.
Polarity Word Polarity from Negated Word Manual
Lexicons Annotation
Great‫رائع‬ +4 Not great ‫غير رائع‬ -2

Agree ‫يوافق‬ +2 Not agree ‫ال يوافق‬ -1

Peace‫سالم‬ +1 No peace ‫ال سالم‬ -2

Torture‫عذاب‬ -4 Without torture ‫بدون‬ +1

‫عذاب‬
Sad‫حزين‬ -3 Not sad‫غير حزين‬ +2

Hard‫صعب‬ -1 Not hard‫ليس صعب‬ +1

By finding the average of the effect of these negation words on the manually annotated words for
each zone, with the average shifted values by negative prefixes, it is found that the negation shifting value
rules are as shown in Equation 3.

(3)

where w is opinion words in a text, Sn is the scope of the negation word, and pol(w) is the polarity of word
w.

3.1.3 Intensifier Rules

To form intensifier rules, a list of intensifier words is generated for both MSA, such as “‫بدقة‬/
precisely”, “‫قليالا‬/little”, “‫حقا ا‬/ really”, “‫تماما ا‬/ exactly”, and for PAL, such as “‫حبطرش‬/ a lot”, “‫قليالا‬/little”,
“‫يادوب‬/ barely”, and for each word in the list there is a modifier value as a percentage, some words are
called amplifiers that have a modifier values more than 100%, such as “‫ بالتاكيد‬/certainly”, and the other
words in the list are called downtoners that have modifier values less than 100%, such as “‫ بالكاد‬/barely”.

Let Iamp be a list of intensifier words of type amplifiers and I down be a list of intensifier words of
type downtoners represented as:

Iamp = {list of amplifier words}

Idown = {list of downtoner words}

174
 Valence Shifter Rules for Arabic Sentiment Analysis

If an intensifier word is found in a text, then the polarity of the words in the scope of the intensifier
word is computed as in Equation 4 (Taboada et al., 2011):

(4)

where Si is scope of intensifier word, w is a polarity word in the intensifier word`s scope, wa is an
amplifier word, wd is a downtoner word, pol(w) is the polarity of word w, Mod(wa) is the modifier value of
amplifier word wa, and Mod(wd) is the modifier value of amplifier word wd.

In order to find the words in the scope of the intensifier, several rules to be followed:

1) Next polarity word is in the scope of the intensifier word.

2) If the next word is a non-polarity word, then the polarity word of up to two next words is
checked, the closest polarity word is considered the scope of the intensifier word.
3) If the intensifier word is the last word, then the word before it is the scope of the intensifier
word if it is a polarity word, otherwise it is ignored.
4) If two intensifiers appear next to each other, then both of them are multiplied to each other,
then it is dealt with as a one intensifier word.

Example 3: “‫إنه فعلا يبدو رائعا ا عليك‬/It really looks beautiful on you”.

In Example 3, the word “beautiful” is considered as the scope of the intensifier word “really”
based on rule number 2, since the next word “look” has no polarity.
Example 4: “ ‫فعلا جداا فيلم سيء‬/really very bad movie”.

In Example 4 following rule number 4, the weight of the first two intensifier words “really” and
“very” are multiplied first, then it is multiplied with the sentiment value of the word "bad", as following:
(150%) * (150%) * (-2) = -4.5.

4. Results and Evaluation

This section presents the evaluation results of the proposed valence shifter rules, and how it has
led to improving the performance measurements. In order to perform this evaluation, four of the most
widely used performance measurements in the literature are calculated for each, namely, accuracy,
precision, recall, and F-score (Al-Kabi et al., 2013; Thelwall & Prabowo, 2009; Shoukry & Rafea, 2012).

Several experiments are performed to evaluate the proposed valence shifter rules. In order to
overcome the effect of the lexicon type on the obtained results, all of the available lexicons are used in this
experiment, namely, AEL, AHL, and DAHL by Mohammad et al., (2016a), UWOM and NileULex by El-
Beltagy (2016), ArSenL by Badaro et al., (2014), SLSA by Eskander et al., (2015), AraSenti-Trans, and
AraSenti-PMI by Al-Twairesh et al., (2016), and Arabic Senti-Lexicon by Al-Moslmi et al., (2017).
Besides, the rule is proposed for Arabic dialects in general, therefore, several Arabic dialects are used in
these experiments, namely, ArSenTD-LEV, ASTD, RR ,and AraSenti-Tweet.

175
 Valence Shifter Rules for Arabic Sentiment Analysis

With the purpose of evaluating the proposed valence shifter rules, first, the proposed negation rule
is evaluated in comparison with the existing rule which is switch negation, and also in comparison with not
handling negation. Table 4, shows the average F-score, precision, and recall when applying the proposed
negation rule, switch negation, and without handling negation on all datasets. Second, applying contrast
rules are evaluated using all lexicons on all datasets as shown in Table 5. Then, applying intensification
rules are evaluated as shown in Table 6. Finally, combining all valence shifter rules together are evaluated
as shown in Table 7.

In the first experiment, the results obtained are presented in Figure 3, and Table 4. From Figure 3,
where accuracies of the presented negation rules in compare with switch negation, and also in comparison
with not handling negation, it is concluded that there is also an improvement in the accuracies when using
the proposed negation rule by almost 2% from the existing switch negation rule, and 3% from not handling
negation. Taking into consideration the fact that the percentage of texts with negation words is almost 12%
of the whole texts, hence, to obtain 3% enhancement to total texts means a 25% enhancement to the texts
with negation, hence, the results of the proposed negation rules are considered promising over standard
baselines.

70.0
60.0
50.0
40.0
30.0
20.0
10.0
0.0

No Negation Switch Negation Proposed Negation

Figure 3: Accuracy Rates without using Negation Rules, using Switch Negation, and using Proposed
Negation Rules

In Table 4, it is clear that the performance of the proposed negation rules outperforms the switch
negation rule by almost an average of 2% in terms of precision, recall and F-score. Moreover, the results of
applying the proposed negation rules also outperform the results of not handling negation by almost 3% in
terms of precision, recall, and F-score. These increments in performance by the proposed negation rules are
considered valuable results, especially when combined with other valence shifter rules.

176
 Valence Shifter Rules for Arabic Sentiment Analysis

Table 4. Results obtained without using Negation Rules, with using Switch Negation, and with using
Researcher`s Negation
Average without Average of Switch Average of the
Lexicon Handling Negation Negation Proposed Rule
P R F P R F P R F
AEL 42.5 42.8 42.6 44.4 44.7 44.5 46.2 46.3 46.2
AHL 55.2 54.9 55.0 56.4 56.2 56.3 58.3 57.8 58.0
ArSenL 27.6 27.8 27.7 28.7 28.8 28.7 30.5 30.4 30.4
NileULex 38.6 38.6 38.6 40.1 40.2 40.1 41.9 41.8 41.8
DAHL 55.1 55.0 55.0 55.0 55.0 55.0 56.8 56.6 56.7
AraSenti-PMI 66.6 66.2 66.4 67.0 66.8 66.9 68.8 68.4 68.6
SLSA 40.0 39.9 39.9 41.6 41.5 41.5 43.4 43.1 43.2
AraSenti-Trans 59.1 58.8 58.9 60.6 60.4 60.5 62.4 62.0 62.2
Average 47.1 47.0 47.0 48.2 48.2 48.2 50.0 49.8 49.9
 Note: P is Precision, R is Recall, and F is F-score.

Whereas Table 5 shows the effect of applying the contrast rules on simple lexicon-based approach
by presenting precision, recall, and F-score with and without applying the contrast rules. Results reveal
that applying the contrast rules enhances performance by almost 2% on average. Moreover, Figure 4
shows that using the contrast rules always enhances the performance for all lexicons in the lexicon-
based approach.
Table 5. Results of not Applying Rules, and Results of Applying Contrast Rules
Average without Average with Applying
Applying Contrast Rules Contrast Rules
Lexicon
P R F P R F

AEL
42.5 42.8 42.6 44.2 44.6 44.4
AHL
55.2 54.9 55.0 57.3 57.0 57.1
ArSenL
27.6 27.8 27.7 28.9 29.1 29.0
NileULex
38.6 38.6 38.6 40.2 40.2 40.2
DAHL
55.1 55.0 55.0 57.3 57.1 57.2
AraSenti-PMI
66.6 66.2 66.4 69.2 68.8 69.0
SLSA
40.0 39.9 39.9 41.7 41.6 41.6
AraSenti-Trans
59.1 58.8 58.9 61.4 61.0 61.2
Average
47.1 47.0 47.0 49.0 48.9 49.0

177
 Valence Shifter Rules for Arabic Sentiment Analysis

 Note: P is Precision, R is Recall, and F is F-score.

70.0
60.0
50.0
40.0
30.0
20.0
10.0
0.0

No Rules Contrast Rules

Figure 4: Accuracy Rates of no Rules against Applying Contrast Rules

While Figure 5 shows the effect of using intensification rules when applied on lexicon-based
approach, it is noticed that it slightly improves the results by 0.7% only, due to the fact that intensification
words do not change the polarity of the words, unlike negation or contrast.

70.0
60.0
50.0
40.0
30.0
20.0
10.0
0.0

No Rules Intensification Rules

Figure 5: Accuracy Rates of no Rules against Applying Intensification Rules

Table 6 shows a comparison between lexicon-based performances with and without using the
intensification rule. It is concluded that the precision, recall, and F-score are always better when using the
intensification rule.

178
 Valence Shifter Rules for Arabic Sentiment Analysis

Table 6. Results without Applying Rules, and Results with Intensification Rules
Average without Applying Average with Applying
Lexicon Intensification Rule Intensification Rule
P R F P R F
AEL
42.5 42.8 42.6 43.3 43.5 43.4
AHL
55.2 54.9 55.0 56.1 55.8 55.9
ArSenL
27.6 27.8 27.7 28.2 28.4 28.3
NileULex
38.6 38.6 38.6 39.0 39.0 39.0
DAHL
55.1 55.0 55.0 55.5 55.5 55.5
AraSenti-PMI
66.6 66.2 66.4 67.6 67.2 67.4
SLSA
40.0 39.9 39.9 40.5 40.4 40.4
AraSenti-Trans
59.1 58.8 58.9 59.6 59.3 59.4
Average
47.1 47.0 47.0 47.8 47.7 47.7

 Note: P is Precision, R is Recall, and F is F-score.

Finally, the last experiment was conducted to evaluate combining all valence shifter rules together. In
this experiment, all lexicons are tested on all datasets with applying negation rules, contrast rules, and
intensification rules altogether. Table 7 shows the evaluation results of simple lexicon-based approach,
besides, a lexicon-based approach with applying all valence shifter rules. Results demonstrate that
implementing valence shifter rules has improved the lexicon-based approach by almost 5% on average in
terms of precision, recall, and F-score. Figure 6 shows the accuracy enhancement when applying all
valence shifter rules, which is also 5% on average when compared with applying simple lexicon-based
approach.

179
 Valence Shifter Rules for Arabic Sentiment Analysis

Table 7. Results without Applying Rules, and Results with Applying all Valence Shifter Rules
Average without Applying Any Valence Average of Applying All Valence
Lexicon Shifter Rule Shifter Rules
P R F P R F

AEL
42.5 42.8 42.6
48.0 48.2 48.1
AHL
55.2 54.9 55.0
61.0 60.5 60.7
ArSenL
27.6 27.8 27.7
32.0 31.9 31.9
NileULex
38.6 38.6 38.6
43.6 43.5 43.5
DAHL
55.1 55.0 55.0
59.9 59.6 59.8
AraSenti-PMI
66.6 66.2 66.4
71.9 71.5 71.7
SLSA
40.0 39.9 39.9
45.6 45.3 45.4
AraSenti-Trans
59.1 58.8 58.9
65.3 64.8 65.0
Average
47.1 47.0 47.0
52.3 52.1 52.2

 Note: P is Precision, R is Recall, and F is F-score.

80.0
70.0
60.0
50.0
40.0
30.0
20.0
10.0
0.0

No Rules All Valence Shifter Rules

Figure 6: Accuracy Rates of no Rules against applying All Valence Shifter Rules

180
 Valence Shifter Rules for Arabic Sentiment Analysis

5. Conclusion

In this paper, new valence shifter rules are proposed as in Section 3.1. While in section 4 an
evaluation of these rules have been conducted to measure the performance in term of accuracy,
precision, recall, and F-score, by testing negation rules first, then testing contrast rules, finally,
testing the intensification rules, where the results demonstrate an improvement in the performance
of the lexicon-based approach by 3% when applying the proposed negation rules, and by 2%
when applying contrast rules, besides, intensification rules improved the performance
measurements by almost 0.7% on average, finally, applying all rules together produced a
significant result, it has improved the performance measurements of the lexicon-based approach
by almost 5% on average. This enhancement of 5% is considered worthy of considering due to
the main reason that texts having valence shifters is almost 17% of the whole datasets, therefore,
5% of this 17 % of the datasets is considered more than 30% of performance enhancement.

References

Abdulla, N., Majdalawi, R., Mohammed, S., Al-Ayyoub, M., & Al-Kabi, M. (2014). Automatic lexicon
construction for arabic sentiment analysis. Proceedings - 2014 International Conference on Future
Internet of Things and Cloud, FiCloud 2014, 547–552. https://doi.org/10.1109/FiCloud.2014.95

Abdul-Mageed, M. (2017). Modeling Arabic subjectivity and sentiment in lexical space, Information
Processing and Management. http://dx.doi.org/10.1016/j.ipm.2017.07.004

Al-Ayyoub, M., Essa, S. B., & Alsmadi, I. (2015). Lexicon-based sentiment analysis of Arabic tweets.
International Journal of Social Network Mining, 2(July 2016), 101–114.
https://doi.org/10.1504/IJSNM.2015.072280

Al-Aziz, A. M. A., Gheith, M., & Eldin, A. S. (2016). Lexicon based and multi-criteria decision making
(MCDM) approach for detecting emotions from Arabic microblog text. Proceedings - 1st
International Conference on Arabic Computational Linguistics: Advances in Arabic
Computational Linguistics, ACLing 2015, (Mcdm), 100–105.
https://doi.org/10.1109/ACLing.2015.21

Al-Horaibi, L., & Khan, M. B. (2016). Sentiment analysis of Arabic tweets using text mining techniques.
International Journal of Computing & Information Sciences, 12(2), 100111F.
https://doi.org/10.1117/12.2242187

Al-Kabi, M., Al-Qudah, N., Alsmadi, I., Dabour, M., & Wahsheh, H. (2013). Arabic/English Sentiment
Analysis: An Empirical Study. The Fourth International Conference on Information and
Communication Systems (ICICS 2013), (October 2015).

Al-Kabi, M., Gigieh, A., Alsmadi, I., Wahsheh, H., & Haidar, M. (2013). An Opinion Analysis Tool for
Colloquial and Standard Arabic. The Fourth International Conference on Information and
Communication Systems (ICICS 2013), (April).

181
 Valence Shifter Rules for Arabic Sentiment Analysis

Al-Moslmi, T., Albared, M., Al-Shabi, A., Omar, N., & Abdullah, S. (2017). Arabic senti-lexicon:
Constructing publicly available language resources for Arabic sentiment analysis. Journal of
Information Science, 0165551516683908. https://doi.org/10.1177/0165551516683908

Al-Twairesh, N., Al-khalifa, H., & Al-salman, A. (2016). AraSenTi : Large-Scale Twitter-Specific Arabic
Sentiment Lexicons. Proceedings of the 54th Annual Meeting of the Association for
Computational Linguistics (ACL 2016), 697–705.

Assiri, A., Emam, A., & Al-Dossari, H. (2017). Towards enhancement of a lexicon-based approach for
Saudi dialect sentiment analysis. Journal of Information Science, 016555151668814.
https://doi.org/10.1177/0165551516688143

Badaro, G., Baly, R., & Hajj, H. (2014). A Large Scale Arabic Sentiment Lexicon for Arabic Opinion
Mining. Arabic Natural Language Processing Workshop Co-Located with EMNLP 2014, Doha,
Qatar, 176–184.

Duwairi, R. M., Ahmed, N. A., & Al-Rifai, S. Y. (2015). Detecting sentiment embedded in Arabic social
media - A lexicon-based approach. Journal of Intelligent and Fuzzy Systems, 29(1), 107–117.
https://doi.org/10.3233/IFS-151574

El-Beltagy, S., & Ali, A. (2013). Open issues in the sentiment analysis of Arabic social media: A case study
in Information Technology (IIT), 1–6.

El-Beltagy, S. R. (2016). NileULex: A Phrase and Word Level Sentiment Lexicon for Egyptian and
Modern Standard Arabic. To Appear in Proceedings of LREC 2016, (April), 2900–2905.

Elhawary, M., & Elfeky, M. (2010). Mining Arabic business reviews. Proceedings - IEEE International
Conference on Data Mining, ICDM, 1108–1113. https://doi.org/10.1109/ICDMW.2010.24

Elsahar, H., & El-Beltagy, S. R. (2014). A fully automated approach for Arabic slang lexicon extraction
from microblogs. Lecture Notes in Computer Science (Including Subseries Lecture Notes in
Artificial Intelligence and Lecture Notes in Bioinformatics), 8403 LNCS(PART 1), 79–91.

Eskander, R., & Rambow, O. (2015). SLSA: A Sentiment Lexicon for Standard Arabic. Proceedings of the
2015 Conference on Empirical Methods in Natural Language Processing (EMNLP 2015),
(September), 2545–2550. Retrieved from http://aclweb.org/anthology/D15-1304

Farra, N., Challita, E., Assi, R. A., & Hajj, H. (2010). Sentence-level and document-level sentiment mining
for arabic texts. Proceedings - IEEE International Conference on Data Mining, ICDM, 1114–1119.
https://doi.org/10.1109/ICDMW.2010.95

Ghallab, A., Mohsen, A., & Ali, Y. (2020). Arabic Sentiment Analysis: A Systematic Literature Review.
Applied Computational Intelligence and Soft Computing, 2020.

Hamouda, A. E. A., & El-taher, F. E. (2013). Sentiment Analyzer for Arabic Comments System.
International Journal of Advanced Computer Science and Applications, 4(3), 99–103.

Hatzivassiloglou, V., & McKeown, K. R. (1997). Predicting the semantic orientation of adjectives.
Proceedings of the 35th Annual Meeting on Association for Computational Linguistics,174-181.

182
 Valence Shifter Rules for Arabic Sentiment Analysis

Heerschop, B., Goossen, F., Hogenboom, A., Frasincar, F., Kaymak, U., & de Jong, F. (2011). Polarity
analysis of texts using discourse structure. Proceedings of the 20th ACM International Conference
on Information and Knowledge Management - CIKM ’11, 1061.

Ibrahim, H. S., Abdou, S. M., & Gheith, M. (2015). sentiment analysis for modern standard arabic and
colloquial. 2015 IEEE 2nd International Conference on Recent Trends in Information Systems,
ReTIS 2015 - Proceedings, 4(2), 353–358. https://doi.org/10.1109/ReTIS.2015.7232904

Ibrahim, H. S., Abdou, S. M., & Gheith, M. (2016). Automatic expandable large-scale sentiment lexicon of
modern standard Arabic and colloquial. Proceedings - 1st International Conference on Arabic
Computational Linguistics: Advances in Arabic Computational Linguistics, ACLing 2015, (April),
94–99. https://doi.org/10.1109/ACLing.2015.20

Ihnaini, B., & Mahmuddin, M. (2018). Lexicon-Based Sentiment Analysis of Arabic Tweets: A
Survey. Journal of Engineering and Applied Sciences, 13(17), 7313-7322.

Ihnaini, B., & Mahmuddin. M. (2018). An expandable and up-to-date lexicon for sentiment analysis of
Arabic tweets. Compusoft, 7, no. 11, 2884-2891.

Kennedy, A., & Inkpen, D. (2006). Sentiment classification of movie reviews using contextual valence
shifters. Computational Intelligence, 22(2), 110–125. https://doi.org/10.1111/j.1467-
8640.2006.00277.x

Mataoui, M. (2016). A Proposed Lexicon-Based Sentiment Analysis Approach for the Vernacular Algerian
Arabic. Research in Computing Science, 110(April 2016), 55–70. Retrieved from
http://rcs.cic.ipn.mx/2016_110/A Proposed Lexicon-Based Sentiment Analysis Approach for the
Vernacular Algerian Arabic.pdf

Medhat, W., Hassan, A., & Korashy, H. (2014). Sentiment analysis algorithms and applications: A survey.
Ain Shams Engineering Journal, 5(4), 1093–1113. https://doi.org/10.1016/j.asej.2014.04.011

Mohammad, S. M., Salameh, M., & Kiritchenko, S. (2016). How translation alters sentiment. Journal of
Artificial Intelligence Research, 55(January), 95–130. https://doi.org/10.1613/jair.4787

Narayanan, V., Arora, I., & Bhatia, a. (2013). Fast and accurate sentiment classification using an enhanced
Naive Bayes model. International Data Engineering and Automated Learning, Lecture Notes in
Computer Science, 8206, 194–201. https://doi.org/10.1007/978-3-642-41278-3_24

Pak, A., & Paroubek, P. (2010). Twitter as a Corpus for Sentiment Analysis and Opinion Mining. In
Proceedings of the Seventh Conference on International Language Resources and Evaluation,
1320–1326. https://doi.org/10.1371/journal.pone.0026624

Pang, B., Lee, L., & Vaithyanathan, S. (2002). Thumbs up?: sentiment classification using machine
learning techniques. Empirical Methods in Natural Language Processing (EMNLP), 10(July), 79–
86. https://doi.org/10.3115/1118693.1118704

Refaee, E., & Rieser, V. (2014). An Arabic twitter corpus for subjectivity and sentiment analysis.
Proceedings of the Language Resources and Evaluation Conference, (spring 2013), 2268–2273.
Retrieved from http://www.lrec-conf.org/proceedings/lrec2014/pdf/317_Paper.pdf

183
 Valence Shifter Rules for Arabic Sentiment Analysis

Shoukry, A., & Rafea, A. (2012). Sentence-level Arabic sentiment analysis. Proceedings of the 2012
International Conference on Collaboration Technologies and Systems, CTS 2012, 546–550.
https://doi.org/10.1109/CTS.2012.6261103

Taboada, M., Brooke, J., & Tofiloski, M. (2011). Lexicon-based methods for sentiment analysis.
Computational Linguistics, 37(September 2010), 267–307. Retrieved from
http://www.mitpressjournals.org/doi/abs/10.1162/COLI_a_00049

Thelwall, M., & Prabowo, R. (2009). Sentiment analysis: A combined approach. Journal of Informetrics,
3(2), 143–157. https://doi.org/10.1016/j.joi.2009.01.003

Yuan, B. (2016). Sentiment analysis of twitter data (Vol. 2016).

184