Assessing and predicting review helpfulness

Online customer reviews represent one of the most popular and accessible source of product/service information. E-commerce platforms enable users to vote for review for their helpfulness, which act as indicator of the review’s reliability for other readers. While numerous scientific publications have focused on the topic of predicting review helpfulness, several questions are yet to be addressed. Moreover, the current literature is highly heterogeneous, leading to inconsistent and contradictory results. Our aim with this study is to synthesize and critically assess the state of the art in research on what makes a review helpful and on predicting review helpfulness. Our primary findings reveal the use of highly varying datasets; a huge plethora of distinct features, including some which are counter-intuitive, as the count of n-letters words or of line breaks; the lack of benchmarks for comparing and assessing algorithms’ performance in predicting review helpfulness or the application of machine learning techniques overlooking the statistical characteristics of the data resulting in flawed results. We propose several research directions to overcome these gaps and advance the state of the art, such as a standard features set and algorithms to be used as benchmark for assessing future research. We also propose new approaches based on recent innovations in argumentation mining and deep learning as well as more advanced statistic techniques, such as lasso/ridge regression.