Estimating Reactions and Recommending Products with Generative Models of Reviews

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 783–791, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Estimating Reactions and Recommending Products with Generative Models of Reviews Jianmo Ni and Zachary C. Lipton and Sharad Vikram and Julian McAuley Department of Computer Science and Engineering University of California San Diego jin018, zlipton, svikram, jmcauley@ucsd.edu Abstract Traditional approaches to recommenda- tion focus on learning from large volumes of historical feedback to estimate simple numerical quantities (Will a user click on a product? Make a purchase? etc.). Natural language approaches that model informa- tion like product reviews have proved to be incredibly useful in improving the perfor- mance of such methods, as reviews pro- vide valuable auxiliary information that can be used to better estimate latent user preferences and item properties. In this paper, rather than using reviews as an inputs to a recommender system, we focus on generating reviews as the model’s output. This requires us to effi- ciently model text (at the character level) to capture the preferences of the user, the properties of the item being consumed, and the interaction between them (i.e., the user’s preference). We show that this can model can be used to (a) generate plau- sible reviews and estimate nuanced reac- tions; (b) provide personalized rankings of existing reviews; and (c) recommend ex- isting products more effectively. 1 Introduction Review text has been extensively studied in mod- ern recommender systems as a means of improv- ing the performance on traditional recommenda- tion tasks. Compared with conventional tech- niques that model simple numerical feedback (rat- ings, clicks, purchases, etc.), review text pro- vides valuable information about user and item attributes, and more importantly the interaction between them. Recent systems have adapted ideas from topic modeling and sentiment analysis to leverage the side-information contained in re- views; in essence, these approaches use language models as a form of ‘regularization,’ such that the model should explain user preferences and re- view text simultaneously (McAuley and Leskovec, 2013; Bao et al., 2014; Ling et al., 2014). In parallel, recent advances in generative text modeling have demonstrated the effectiveness of recurrent neural networks in capturing content, structure, and style in natural language. As a re- sult, several recent works have focused on learn- ing generative models of product reviews, either to generate reviews per se, or as a means of learn- ing user and item attributes (Lipton et al., 2015; Radford et al., 2017; Dong et al., 2017; Hu et al., 2017). In this paper, we address the problem of how to leverage both review text and implicit feed- back simultaneously in order to provide richer user experiences and more meaningful recommenda- tions. In particular, we focus on estimating latent user preferences through implicit feedback while simultaneously predicting the contents of the re- views themselves. Thus, given an item that a user hasn’t yet interacted with, our goals are to (a) gen- erate a plausible review, in order to estimate the user’s nuanced …

Content truncated to 500 words