Personalized Explainable Recommendation based on BERT

Abstract

With the proliferation of text information on the Internet, the rapid evolution of artificial intelligence, and the wide application of machine learning and deep learning, text emotion analysis has been widely concerned by the academic community. Personalized recommendation system has emerged as the circumstances demanded, and has quickly gained interest in both industry and academic circles. It has gradually become an extremely important part of people's life and work in many fields, such as e-commerce, short video content, take-out service, online advertising push and so on. First, the Chinese comment text of an item is analyzed in this paper. Since the absolute standard of binary text sentiment analysis cannot meet the need of the recommendation system to recommend the item to users for interpretation, a ternary text sentiment analysis method based on BERT model is used to combine with the idiosyncrasies of text data is proposed to solve the problems of poor Chinese text representation, low exactitude, and inability to precisely comprehend the semantic information expressed in the text, which are caused by polysemy of Chinese version. The proposed method can generate interpretable recommendations for items that users are interested in. The relevant properties of the text are captured by the Transformer encoder in the BERT model, meanwhile the attention framework is used to weight the information recovered from the pattern to highlight the hinge information in the comment on the text. Second, the SoftMax function is used for categorizing the text aspect data of users' reviews of items, as well as finally the recommendation system recommends interested items to users and produces emotional and coloration reasons for recommendation that are accorded with users' justifications. The method is applied to real datasets, and the results show that text breakdown effect has been achieved, which greatly improves the interpretability of recommendation system, which is more in line with users' ideas.