Deep Semantic Analysis and Thematic Evolution Prediction of American Literary Texts Based on BERT and Knowledge Graph

When analyzing literary texts, the lack of computational tools that can accurately identify plots affects the accuracy of genre identification. To this end, this paper proposes a method based on the combination of BERT pre-trained model and knowledge graph, aiming to conduct deep semantic analysis of American literary texts and predict their thematic evolution. By leveraging the powerful language representation ability of BERT and the structured knowledge of knowledge graph, this paper constructs a comprehensive framework to enhance the understanding of complex semantics and themes in literary works. This paper studies the linking of text data with knowledge graph entities at the knowledge layer to form a sentence tree, which is converted into an input format understandable to the model through the embedding layer. In addition, the visibility matrix is introduced to restrict the mutual visibility of words in the sentence tree, and the self-attention mechanism of the Mask-Transformer decoder layer is used to focus on the visible area of each word with other parts of the text and related knowledge graph information. Finally, this paper also introduces the application of the Latent Dirichlet Allocation (LDA) model in modeling text topic distribution, and updates the topic and word distribution through Gibbs sampling. Experimental results show that this method can effectively analyze the deep semantics of literary works and predict the evolution trend of themes, providing a new perspective for literary research and knowledge discovery.