A CLUSTERING AND WORD SIMILARITY BASED APPROACH FOR IDENTIFYING PRODUCT FEATURE WORDS

Product designers need to capture feedback from customers in order to assess how the product performs and is perceived in the market. One such example of publicly available source of customer's feedback is the online reviews in an e-commerce website. Two main difficulties in dealing with the reviews are finding relevant words related to a product and grouping different words that represent the same product feature. To overcome these difficulties, both lexical and distributional approaches are utilized in the paper. Using distributional information, words are embedded into real vector space using word2vec and then clustered. Using lexical information from WordNet, the head word for each cluster is identified by considering the similarity with the head words of other clusters. A comparison is made between using X-means and iterative c-means clustering with added word similarity information when breaking a cluster. In the case study of wearable technology products, starting from a large number of words, the approach is shown to identify relevant product feature words.