Improving aspect extraction using aspect frequency and semantic similarity-based approach for aspect-based sentiment analysis

Abstract

Identifying the targets of users’ opinions, referred as aspects, in aspect-based sentiment analysis, is the most important and crucial task. A large number of approaches have been proposed to accomplish this task. These approaches identify a huge amount of potential aspects from customer reviews. But not all the extracted aspects are interesting and include terms which are not related to the product and these irrelevant terms affect the performance of the aspect extraction approaches. Therefore, in this paper, we are proposing a two-level aspect pruning approach to eliminate irrelevant aspects. The proposed approach performs the task of aspect pruning in two steps: (a) by calculating the frequency of each word and selecting the most frequent aspects; and (b) by calculating the semantic similarity of non-frequent words and eliminate aspects which are not semantically related to the product. Our experimental evaluation has shown a significant improvement of the proposed approach over the compared approaches.