A Joint Model for Extracting Latent Aspects and Their Ratings From Online Employee Reviews

Abstract

The personal description of a company associated with job satisfaction, company culture, and opinions of senior leadership is available on workplace community websites. However, it is almost impossible to read all of the different and possibly even contradictory reviews and make an accurate overall rating. Therefore, extracting aspects or sentiments from online reviews and the corresponding ratings is an important challenge. We collect online anonymous employees’ reviews from Glassdoor.com which allows people to evaluate and review the companies they have worked for or are working for. Here, we propose a joint rules-based model which combines the numerical evaluation reflected in the form of 1–5 stars, and the reviewed context to extract aspects. The model first inputs the five aspects with the initial word sets that are manually screened, and expands the aspect keyword sets through bootstrapping semi-supervised learning, and then uses latent rating regression to obtain the aspect score and aspect weight to update the corresponding score. Our experimental evaluation has shown better results as compared with an unsupervised learning of the latent Dirichlet allocation. The results could not only help companies understand their strengths and weaknesses, but also help job seekers apply for companies.