Co-Extracting Feature and Opinion Pairs From Customer Reviews Using Hybrid Approach

The Internet and WWW have turned today's world into global village. Over the years technology has significantly changed the way people communicate. How they feel, what they like or dislike, people like to share it online. Proper mining and analysis of public feelings or opinions can bring a miracle to existing business profit. For the same purpose instead of exclusive public opinion polling, machine learning tools are preferred. Few researchers are trying to interpret it with the help of linguistic approach. But the reviews or opinion of the customers' also known as user-generated content has the characteristics such as incomplete, grammatically incorrect, weakly structured etc. This causes degradation in the performance of the opinion mining task. Whereas, few researchers are trying to apply statistical measures for the same. Both the methods have their own pros and cons. In this paper, we have proposed a hybrid approach by integrating both the linguistic as well as statistics approach into a single unified framework. The key objective of our task is to co-extract the feature and opinion pair from customers reviews. Let's view the unstructured corpus a combination of a set of grammatically correct and a set of grammatically incorrect sentences. For the grammatically correct sentence, use syntactic pattern-based rules to extract the features and opinion pairs. For the set of sentences not interpreted by the syntactic parser, apply word alignment technique to extract probable feature opinion pair. Further, we determine the aspect of each probable feature and calculate the opinion association among these words. Finally, we calculate the soundness of extracted feature opinion pair with the document. For experimentation purpose, we used customer review dataset of five different products. Through the analysis, a high accuracy is achieved by the hybrid approach.