Consumer Review Analysis Using Topic Modelling

Consumers put a lot of information about the products that they are using in the form of reviews in several review sites. These reviews are valuable but human reader won't be able to read them all. Hence, a machine learning technique is required to extract some meaningful idea out of those reviews to help decision makers to take informed decisions. The current study tries to address this issue with topic modelling.

For the said purpose, for three competing mobile phones (Samsung Galaxy Note, Xperia Z and iPhone 5s), reviews were collected through web crawling and subsequent web scraping. After data pre-processing, latent semantic analysis was done to find out optimum number of topics. Once the number of topics was extracted, correlated topic modelling was applied to extract topics from texts for subsequent analysis. Random forest classification was used to see if the topics could differentiate reviews of one brand from another brand.

The work clearly shows how different smart phones were reviewed based on different dimensions. The extracted topics were found to discriminate the smart phones with almost 60% accuracy having Cohen's kappa value around 0.37 suggesting fair performance. The themes extracted suggested that consumers had given different levels of importance to different features while writing their reviews on different phones.

Output of topic modelling cannot be used directly as text summarisation. Hence, subjective judgements are required to interpret the results. This is where domain knowledge becomes very important.

This is an original work of the researchers representing use of text mining tools for review analysis.