Abstract: In this paper, under the theory of dual-route model, the impact on book review helpfulness exerted by five aspects of its content, namely informativeness, structure, linguistic style, argument and subjectivity, is investigated. Thus, the significant content features can be confirmed and used for review helpfulness modeling. Two models are involved in the study. The interpretation one, built by employing a feature selection algorithm, is used for identifying the content features impacting on review helpfulness significantly; while the tree-based regression model is used for predicting review helpfulness and rank. For interpretation model, the research result indicates that informativeness, structure and argument related features are much more significantly related with review helpfulness. As for prediction model based the optimal features, its R² achieves 78% and the error index MSE is less than 0.001. Specifically, the predictive rank is basically in line with vote based ranking for reviews with higher score helpfulness. Overall, all results indicate the helpfulness of a review can be predicted quite accurately according to its content only, which means the study contributes to find out feasible solutions for the review quality control and effective utilization.

Key words: online review, prediction, review helpfulness, text mining