Multi-grained Aspect Fusion for Review Response Generation

Review response generation (RRG) aims to automatically generate responses to customer reviews. Responding to reviews in a right manner is important to online customer experience. However, most previous research on RRG focused on exploring coarse review information and ignored fine-grain aspects within reviews, especially those with negative sentiment. As a result, the generated responses are usually not targeted to users' real concerns in their reviews. To this end, we proposed a multi-grained aspect fusion model (MGAF) model to improve the targeting of generated responses. In particular, we first enhance the targeting ability by performing sentence-level aspect selection and response script learning. Then we integrate aspect-level keywords with sentiment information to further improve the diversity of generated responses. Experimental results on both Chinese and English datasets show that our proposed model outperforms the state-of-the-art models available, demonstrating the importance of fusing multi-grained aspect information for targeted response generation.