Synthesizing Natural Language Explanations for Recommendations

Previous research has shown convincing results that explaining the machine-generated recommendations can help customers make more accurate decisions and improve their satisfaction. Many E-commerce applications allow users to leave reviews expressing their experience and sentiments in addition to numerical ratings. These sentimental correlated reviews serve as a perfect resource of explanations for their corresponding ratings. Some works have jointly modeled the recommendation and review generation, but their objective is mostly memorizing the reviews verbatim. In our definition, explanations shall describe features of a given item in an aligned sentiment to defend the recommendation result. Under this insight, we propose a novel neural architecture which not only predicts personalized ratings for recommendation, but also generates supportive explanations describing customized features in a consistent sentiment towards the predicted ratings. On top of a multi-task model, we introduce a rating gate and feature gate to fuse the sentimental representation among the two tasks while specifically emphasizing the sentiment and feature in the generated explanations. Moreover, a content rating supervisor is added in combination with a Gumbel sampler to align the sentiment between end explanations and predicted ratings. To further regularize the generation quality, we adopt adversarial training which can smoothly integrate with the existing Gumbel sampler. Extensive experiments show our model exceeds the baselines in both rating prediction and text generation. Furthermore, the sentimental alignment of generated explanations is significantly improved through our method.