Neural Collaborative Embedding From Reviews for Recommendation

This paper mainly studies the personalized rating prediction task based on review texts for the recommendation. Recently, most of the related researches use convolutional neural networks to capture local context information, but it loses word frequency and global context information. In addition, they simply equate the user (item) embedding to review embedding, which brings some irrelevant information of the review text into user preference or item property. Moreover, they only consider the low-order interactions, which remain the fine-grained user-item interactions to be explored. To solve these problems, we propose a novel method neural collaborative embedding model (NCEM). We first adopt pre-trained BERT model, which has been proven to improve most of the downstream NLP tasks, to simultaneously capture the global context and word frequency information. In addition, a self-attention mechanism is introduced to learn the contribution of each review. Next, we develop a neural form of standard factorization machine, which can model first-order and second-order user-item interactions by stacking multiple layers. The extensive experiments on four public datasets showed that NCEM consistently outperforms the state-of-the-art recommendation approaches. Furthermore, the recommendation interpretability can be improved by showing users the high score reviews accompanied recommended item.