Abstract: 【Purpose/significance】Through deep learning method, fine-grained book reviews are mined, and the similarity between
books is optimized based on the mining results.【Method/process】Firstly, book reviews are collected from online book review websites, and attribute lexicon is constructed by pos analysis. Then, reviews are typed based on attribute lexicon, and annotation data is learned by BERT-BiLSTM model to realize automatic classification of reviews. Finally, vector representation of classified reviews is carried out by BERT. The similarity between reviews is calculated by cosine similarity to represent book similarity【. Result/conclusion】The ac? curacy, recall and F1 values of BERT-BiLSTM review classification model constructed in this paper reach 0.922, 0.921 and 0.921 re? spectively, which can achieve good review classification results.The model divides reviews into five types: writing style, character iden? tity, plot, summary and reader's attitude, so as to calculate the similarity between books and get a more consistent result【. Innovation/limitation】Compared to other types of reviews, the effectiveness of calculating book similarity by character identity and plot type of re? views needs to be improved.A more fine-grained analysis of these two types of reviews can be performed later.