Unsupervised tag recommendation for popular and cold products

The rapid expansion of the Internet and its connectivity has given tremendous growth to e-commerce sites. Product reviews form an indispensable part of e-commerce sites. However, it is challenging and laborious to go through hundreds of reviews. In this paper, we address the problem of summarizing reviews by means of informative and readable tags. We present a novel unsupervised method of generating tags and rank them based on relevance. We refine the generated tags using NLP syntactic rules to make them more informative. Our proposed Tagging Product Review (TPR) system takes into consideration the opinions expressed on the product or its aspects. We also address the problem of tag generation for cold products, which have only a limited number of reviews and that too, with very short content. We use transfer learning to build a tag cloud from popular product reviews and use it to identify good tags from cold product reviews. We evaluate our proposed system using online reviews of twelve products of varying popularity, collected from Amazon.com. Our result demonstrates the effectiveness of our approach at generating relevant tags compared to three popular baseline methods. Our proposed approach gives an average tag relevance score (NDCG) of around 79% for popular products and 85% for cold products. Our approach also gives an average precision of 89% for identifying correct tags. The results suggest that our TPR system successfully summarize reviews by means of tags.