The so-called “crowd wisdom” phenomenon took on a whole new dimension with the advent of the web and the emergence of myriads of online product reviews written by both happy and (more often than not) angry customers. A thorough understanding of this feedback via automatic semantic analysis of these comments would be of great value to both companies (to better adapt products to user needs) and customers (to better inform purchase decisions).

The supervised joint aspect and sentiment model (SJASM) is a new probabilistic model that can be used to infer both the overall and feature-specific ratings of a given product from a set of user-generated reviews. This probabilistic model has been trained and tested on publicly available, manually annotated databases of game, compact disc (CD), and hotel reviews. Model fitting is performed using, for each training document (in 80 percent of the data), a set of pairs of opinion words and (positive or negative) sentiments, extracted using the natural language processing (NLP) Stanford Parser tool, and its user-provided general rating.

A comparative evaluation of SJASM to seven existing rating models is then performed using the remaining 20 percent of data, where SJASM is shown to outperform competitors on all key metrics: semantic aspect detection, aspect-level sentiment identification, and overall rating prediction. Even though this paper is not an easy read, requiring advanced skills in probabilistic inference mathematics, its practical importance cannot be downplayed for both companies interested in leveraging their customer input and researchers focused on providing them with ever-better knowledge extraction tools.