An efficient semantic recommender method forArabic text
| Date | 01 April 2019 |
| DOI | https://doi.org/10.1108/EL-12-2018-0245 |
| Pages | 263-280 |
| Published date | 01 April 2019 |
| Author | Bilal Hawashin,Shadi Alzubi,Tarek Kanan,Ayman Mansour |
An efficient semantic
recommender method for
Arabic text
Bilal Hawashin,Shadi Alzubi and Tarek Kanan
Al-Zaytoonah University of Jordan, Amman, Jordan, and
Ayman Mansour
Tafila Technical University, Tafila, Jordan
Abstract
Purpose –This paperaims to propose a new efficient semantic recommendermethod for Arabic content.
Design/methodology/approach –Three semantic similaritieswere proposed to be integrated with the
recommender system to improve its ability to recommend based on the semantic aspect. The proposed
similaritiesare CHI-based semantic similarity, singular value decomposition(SVD)-based semantic similarity
and Arabic WordNet-based semantic similarity. These similarities were compared with the existing
similaritiesused by recommender systems from the literature.
Findings –Experiments show that the proposed semantic method using CHI-based similarity and using
SVD-based similarity are more efficient than the existing methods on Arabic text in term of accuracy and
executiontime.
Originality/value –Although many previous works proposed recommender system methods for
English text, very few works concentrated on Arabic Text. The field of Arabic Recommender Systems is
largely understudied in the literature. Aside from this, there is a vital need to consider the semantic
relationships behind user preferences to improve the accuracy of the recommendations. The
contributions of this work are the following. First, as many recommender methods were proposed for
English text and have never been tested on Arabic text, this work compares the performance of these
widely used methods on Arabic text. Second, it proposes a novel semantic recommender method for
Arabic text. As this method uses semantic similarity, three novel base semantic similarities were
proposed and evaluated. Third, this work would direct the attention to more studies in this
understudied topic in the literature.
Keywords User preferences, Collaborative filtering, Machine learning,
Arabic recommender systems, Semantic similarity, Arabic language content, User interests
Paper type Research paper
1. Introduction
Recommender systems are used to suggestitems to users based on their interests. They are
used in various domains includingresearch papers recommendation, book recommendation,
product recommendation and many more. In this paper,the concentration is on the domain
of online movie stores, even though the proposed system can be used in other domains as
well.
Although many works have proposed recommender system methods in the literature,
very few works concentrated on recommender system methods for Arabic text. Arabic
content on the Web has significantly developed in the past few decades due to various
reasons and, in consequence, it is equally important to have an efficient recommender
system for such Arabic content.
Semantic
recommender
method
263
Received15 December 2018
Revised26 January 2019
7March 2019
Accepted18 March 2019
TheElectronic Library
Vol.37 No. 2, 2019
pp. 263-280
© Emerald Publishing Limited
0264-0473
DOI 10.1108/EL-12-2018-0245
The current issue and full text archive of this journal is available on Emerald Insight at:
www.emeraldinsight.com/0264-0473.htm
Aside from this, many existing recommender systems do not consider the actual user
interests. Any user would give an item a high rating for many reasons. In Hawashin et al.
(2015), a user interest extractor in recommendersystems was proposed to extract the actual
user interests. For example, a user interest could be horror movies, movies by the director
James Cameron, movies of the Eighties, movies of the actress Julia Roberts and so on.
However, in some cases, these interests could be limited, which would negatively affect the
accuracy of the recommendersystem. Moreover, in other cases, the item description may not
contain the user interest itself but one of its semanticallyrelated terms. For example, a user
interest could be ﺭﻉﺏ (horror), which is semanticallycorrelated to ﺥﻭﻑ (fear).In this case, if a
movie description does not contain the user interest itself, and without using the semantic
relationships among terms, this itemwould not be recommended to the user even though a
semantically related term to the user interest exists in the item description. From this
example, it can be concluded that there is a need to extend user interests by including the
semantically relatedterms to these user interests.
In this work, various existing recommender system methods on Arabic text are
compared. These methods were proposed originally for English text, and have never been
tested on Arabic content.Next, a new semantic recommender system method for Arabictext
is proposed. This method considers the semantic similarities between user interests and
other terms, and for this sake, three novelefficient semantic similarities were proposed.
As for the comparison of the existing methods, the performances of Item-Item content-
based similarity, Item-Item content-based correlation, Item-Item rate similarity, Item-Item
content similarity usinglatent semantic indexing (LSI) reduced space and User Interest-Item
similarity are compared on Arabic text. Unfortunately, no Arabic data set was available to
evaluate these methods. Therefore, a synthesized data set that is relatively similar to the
commonly used English MovieLens data set is used. This data set contains some noisy
information that makes it close to realworld data. As for the novel semantic method, and to
find the semantic relationships, three base methods are proposed; CHI-based semantic
similarity, singular value decomposition (SVD)-based semantic similarity and WordNet-
based semantic similarity. Their performance is compared with the existing methods. The
contributionsof this work are as follows:
comparing various widely used English recommender system methods on Arabic
text;
proposing an efficient semantic recommender method for Arabic text;
proposing three semantic similarities for Arabic recommender systems; and
insisting on the importance of having more research about Arabic recommender
systems.
The rest of this paper is organized as follows. Section 2 is a literature reviewof the related
works in this field. Section 3 provides descriptions of the compared existing recommender
system methods. Section 4 describes the proposedsemantic method for Arabic text and the
three-base novel semanticsimilarities. Section 5 is the experimental part and the discussion,
and Section 6 is the conclusion and futurework.
2. Literature review
Many recommender systems have been proposed in the literature for English text.
Collaborative filtering (Chien and George, 1999;Goldberg et al.,1992;Herlocker et al.,2004;
Pavlov and Pennock, 2003;Resnick et al.,1994;Rich, 1979;Sarwar et al.,2001) uses user–
user similarity and suggests items that were highly rated by similar users. In details, the
EL
37,2
264
To continue reading
Request your trial