Ontology Based Arabic Opinion Mining

In Arabic language, studies in the area of opinion mining are still limited compared to that being carried out in other languages. In this paper, we highlight the problem for Arabic opinion mining techniques when analysing reviews having different features with different opinion strengths. The traditional works of opinion mining consider all features extracted from the reviews to be equally important, so they fail to determine the correct opinion of the review and make the review's sentiment classification less accurate. This research presents a technique based on an ontology that uses feature level classification to classify Arabic user-generated reviews by identifying the relevant features from the review based on the degree of these features in the ontology tree. Then, we exploit the important features extracted to determine the overall polarity of the review. Moreover, summarisation for each feature is done to determine which feature has satisfied or dissatisfied customers. To evaluate our work, we use public datasets which are hotels and books datasets. We used f-measure metrics to assess the performance and compare the results with other supervised and unsupervised techniques. Also, subjective evaluation is used in our method to demonstrate the effectiveness of feature and opinion extraction process and summarisation. We show that our method improves the performance compared with other opinion mining classification approaches, obtaining 78.83% f-measure in hotels domain and 79.18% in books domain. Furthermore, the subjective evaluation shows the effectiveness of our method by getting an average f-measure of 84.62% in hotels dataset and 86.31% in books dataset.