Detection and Extraction of Fault Reports from Online User Reviews

With the advent of web 2.0 and social networks users and consumers share their opinions and thoughts about various products and services. It is very important for producers and vendors to capture and analyze these user generated contents and extract information such as pros and cons of their products as compared to products from competitors. Negative opinions and criticizing reviews help vendors to enhance their products or services and, hence improve their market shares. This research aims to extract opinions and reviews containing a report of failure or weakness in a product. A random forest initially classifies and filters opinion texts potentially containing a fault report. Later, an LDA (latent Dirichlet allocation) based topic modeler summarizes the reports by their source and severity. Experiments over a reasonably large set of reviews from Amazon web site reveal that random forests trained with as small set of labeled samples performs well in detecting failure cases. Summary results by LDA conceptually groups different failure types and identifies their sources and, as a byproduct, suggests widely used textual patterns in fault reports.