Sentiment analysis in a hybrid hierarchical classification process

As more and more reviews are generated online, sentiment analysis has been widely studied and developed both in academia and industry recently. In this paper, we propose a novel approach to tackle two complementary sub-tasks of sentiment analysis on review texts, i.e., the Attribute Detection (AD) task and the Sentiment Orientation (SO) task, in a Hybrid Hierarchical Classification Process (HHCP). Specifically, the HHCP approach employs a linear Fisher classifier to achieve the AD task in an ontology-based hierarchical classification process. As evidences show that common statistical classifiers that have superior performances on semantic classifications do not necessarily work well on classifying sentiment information, we did not continue to use the linear Fisher classifier in the SO task. Instead, we turn to a rule-based heuristic classification method on performing sentiment orientation for attributes identified from the AD task. The proposed HHCP approach is empirically analyzed in extensive experiments. Experiments conducted for performance comparison not only show that our proposed HHCP approach outperforms the other three baseline methods, but also address all the concerns raised before experiments. Further experiments on analyzing the impact of dimensionality d of the input vector space confirm that the conclusions drawn from performance comparison hold very well as d varies. Experiments of studying computational efficiency demonstrate that compared with the existing HL-SOT approach our proposed HHCP approach is more efficient in computation.