Publication

Abstract : Often customers refer to product specifications to know if a product fulfills their requirement, but in case they do not find answer to their queries they turn to the online reviews. A knowledge-based system tries to answer user queries using product specifications alone, but specifications available can be insufficient. An idea is introduced of combining question answer and review datasets to find the most relevant reviews corresponding to the question using semantic and statistical similarity measures and optimizing their weights. This is accomplished by using a large volume of already answered queries to find weights of four different similarity metrics (Cosine, BM25, WordNet and Word Embedding) used to find similarity between question and reviews. Similarity measure weights are optimized using PSO-based weight optimization technique where the fitness function is evaluated in terms of how best the sentiment extracted from top reviews agrees with answer of question in Q/A dataset. Results achieved surpass baseline in three out of four domains.