Customer sentiment analysis with more sensibility

Abstract

Customers’ evaluations on products can be derived by analyzing online reviews using machine learning. Positive or negative responses can be sensed by words they write in reviews, and topics they compliment or complain about can be grasped by clustering reviews. Combination of those results is regarded as the customers’ sentiment analysis. When reviews are given as free-form text without scores, general-purpose dictionaries are used to recognize sentiment words. However, customers do not only use standard words to express their emotions, but they also use non-grammatical language such as internet jargon. Unfortunately, existing methods cannot capture those sentiment words. Moreover, combination of sentiment words with customer topics simply represents frequencies and does not indicate detailed evaluation patterns. In this study, we propose a customer sentiment analysis method consisting of sentiment propagation and customer review analysis. It works more sensibly by expanding sentiment words from dictionary to those varying words as mentioned above. To implement this, semi-supervised learning is employed to a word graph that is constructed by a word embedding algorithm. Using this more sensible word graph, customer review analysis is conducted. Reviews are grouped into major complaint topics. Meanwhile, an index for customer dissatisfaction is designed by composition of ‘controversy’ and ‘complaint’. The former stands for ‘coverage of dissatisfaction’ while the latter indicates ‘degree of dissatisfaction’. The proposed method was applied to 3,11,550 reviews across five automobiles from ten internet communities. Case study illustrates which parts of automobiles lead to customer dissatisfaction, and therefore where investment and examination are required.