Rating-boosted abstractive review summarization with neural personalized generation

In this paper, we study abstractive summarization for product reviews in the recommender systems, which aims to generate condensed text for online reviews. The summary generation is not only relevant with the content of the review itself but should be fully aware of the intrinsic features of the corresponding user and product, i.e., personalization, which are helpful to identify the saliency information in the reviews. Therefore, we propose a Rating-boosted Abstractive Review Summarization with personalized generation (RARS). In our approach, we first propose a neural review-level attention model to effectively learn user preference embedding and product characteristic embedding from their history reviews. Then, we design a personalized decoder to generate the personalized summary, which utilizes the representations of the user and the product to calculate saliency scores for words in the input review to guide the summary generation process. In addition, the rating information can explicitly indicate the sentiment opinion, hence we jointly optimize the summary generation and rating prediction through a multi-task framework, where the two tasks inherently share user preference embedding and product characteristics embedding. Extensive experiments on four datasets show that our model can effectively improve the performance of both review summarization and rating prediction.