Opinion mining from Arabic comparative sentences

Abstract

This paper discuses the problem of identifying comparative opinion sentences in Arabic text. Most works in the field of opinion mining concentrate on extracting knowledge from direct opinions. Directly mining positive or negative opinions on a product review or its features is only one form of opinion mining; comparing a product review with some other competitive products is another form. Comparisons focus on mining opinions from comparative sentences, ie, to determine which entities in products are preferred by its author. There are some work in this area in English language. This is the first in Arabic. Mining from comparative text can be divided into three tasks,. The first task is to identify comparative statement from non-comparative ones. In this task, we used method that depends on linguistic classification, where we got f-measure of 63.73%. Then we used three machine learning methods where we got better performance which is about 86.63% in the best case. Finally, for this task, we used combined approach of linguistic and machine learning where we got fmeasure of 88.87%. In the second task we generated a set of rules to characterize three types of comparative statements. We left the third task for future work