Constructing an Arabic Opinion Mining Model: With Special Reference to Telecommunication Companies and Hotel Reviews

Abstract

Due to the recent significant growth of e-commerce applications, most of the widely used products are marketed online. This triggered online assessment of products. As such, the success or failure of companies is partially measured by their ability to take assessments of their products seriously. Analysis of these assessments is necessary for ensuring continuous customer satisfaction and further improvements of current and future products. Naturally, understanding the preferences of customers is crucial for product manufacturer as it helps them in product development, marketing and consumer relationship management. On the other hand, customers use of reviews by other’s online assessments influence their decision as to whether or not they purchase a product. Expectedly, assessment are given in unstructured texts of a natural language. Thus, their processing requires appropriate knowledge in different domains that include, but are not limited to: database, information retrieval, information extraction, machine learning, and natural language processing. However, it becomes difficult for product manufacturers or dealers to keep track of large number of assessments, hence forth will be called opinions and/or sentiments. In the past few years, researchers looked at different ways of taking further advantage of opinions in what is now known as opinion mining or sentiment analysis. The scope of opinion and sentiment includes characteristic, functionality and features of product. This thesis is about novel methods that addresses challenges of opinion mining of Arabic texts. To that end, a set of Arabic language corpora from hotel and telecommunication companies has been collected. The set was developed for evaluating the proposed sentiment analysis methodologies. As well, Arabic Sentiment Classifier (ASC) has been implemented at the document-level. This research focuses on improvement of the effectiveness of feature selection using Information Gain . It then proposes a generic framework on for feature-based level analysis. The Arabic Sentiment Analyzer (ASA) framework consists of two main modules: a language resource construction and an opinion miner. For the language resource construction module, the first phase proposes constructing an opinion lexicon for Arabic opinion word. It is based on a bootstrapping process over an online dictionary. A few seed sentiment words have been used for bootstrapping based on the synonym and antonym structures of the dictionary. This method is simple and efficient as it IVgives reasonable results. During the second phase, features of objects are extracted based on frequent nouns, noun phrases, association rule mining and Natural Language Processing (NLP) techniques. This phase takes advantage of syntactic patterns to improve the accuracy of frequency- based techniques. Product features are stored in feature sets. After a language resource is constructed, the opinion mining module uses a novel information summarizing and visualization approach. The approach is based on NLP techniques for defining sentiment sentences, identifying orientations of features and summarizing results. The visualization module is aimed at providing users an effective way of browsing the set of feature according to the polarity expressed by each assessments. In piratical results reflect efficiency of the proposed system.