Plot text is very valuable supporting information in movie recommendations. It has several characteristics: 1) It is rich in content. Each movie often has a document of more than 200 words to describe it, which can give the movie a rich semantic meaning. 2) Objectivity. Plot texts are different from review information. A movie may have thousands of reviews with mixed and conflicting opinions. However, a film has only one plot text, which is fair in tone and does not take a position. Despite its appealing properties and potential for accurate movie portrayal, the lack of a building block for effectively mining plot semantics has led to the marginalization of plot text in the design of movie recommendation algorithms. Therefore, in this paper, we explore the application of the Transformer, currently the best natural language processing module, to learning movie plot texts to help achieve more accurate rating prediction. We propose the "Plot-Aware Transformer" model (PAT) to model the process of "user-movie" rating interaction. We test the PAT model on several movie datasets and demonstrated that the model is competitive. In all tasks, PAT achieves state-of-the-art performance compared to baseline experiments.

With the ever-growing dynamicity, complexity, and volume of information resources, the recommendation technique is proposed and becomes one of the most effective techniques for solving the so-called problem of information overload. Traditional recommendation algorithms, such as collaborative filtering based on the user or item, only measure the degree of similarity between users or items with single criterion, i.e., ratings. According to the experience of previous studies, single criterion cannot accurately measure the similarity between user preferences or items. In recent years, the application of deep learning techniques has gained significant momentum in recommender systems for better understanding of user preferences, item characteristics, and historical interactions. In this work, we integrate plot information as auxiliary information into the denoising autoencoder (DAE), called SemRe-DCF, which aims at learning semantic representations of item descriptions and succeeds in capturing fine-grained semantic regularities by using vector arithmetic to get better rating prediction. The results manifest that the proposed method can effectively improve the accuracy of prediction and solve the cold start problem.