Using Dependency Bigrams and Discourse Connectives for Predicting the Helpfulness of Online Reviews

Helpfulness prediction represents an interesting research topic with immediate practical applications both from a data mining and marketing perspective. In this study we evaluate the performance of two text-based features that have not been used in that context, namely (a) a variation of the bigram feature, utilizing grammatical dependencies and (b) discourse connectives. By treating helpfulness prediction as a binary classification task we show that both features contain valuable information but however they should be used with caution due to the restrictive experimental setup. The study serves as a ground for future work regarding the usefulness of the proposed features in review helpfulness prediction