Product developers often rely on customer feedback when developing and improving their products. One source of such feedback are Web stores that enable customers to review and rate products and services. Given the large number of Web stores and reviews, the process of gathering and analyzing the feedback information is usually tedious and slow. It is often impossible to examine all the available reviews and the results of their analysis provides us with too much details. The later derails us from seeing the big picture or the main message of feedback that customers wanted to communicate with us.
The thesis applies machine learning methods to the task of processing customer reviews available in the Web stores with a purpose of giving predictions of their corresponding numerical ratings. In particular, we developed a solution that combines natural language processing algorithms with machine learning methods to predict customer ratings. The central thesis goal is to empirically prove that machine learning methods are more accurate than simple sentiment analysis and statistical methods when predicting customer ratings from natural language reviews.
Results of empirical evaluation show that on the selected prediction task machine learning methods perform better than simple sentiment analysis and statistical treatment with respect to multiple evaluation metrics.
|
