Aiming at the semantic complexity and long text characteristics of traditional cultural literature, this study proposes an automatic keyword extraction method based on deep learning. A hybrid architecture combining the Bidirectional Encoder Representations from Transformers (BERT) pre-trained language model and the attention mechanism is designed to achieve efficient keyword extraction through input embedding, semantic encoding, attention focus and classification output. Data preprocessing uses word segmentation, stop word filtering and term frequency-inverse document frequency (TF-IDF) weighting to optimize feature representation; model training uses cross entropy loss and AdamW optimizer to ensure convergence and robustness. Experimental results show that the proposed model is significantly better than TF-IDF, TextRank and pure BERT models in precision, recall and F1 score, with an F1 score of 81.5%, and its adaptability in multi-topic literature is verified by heat maps and convergence curves. This method effectively solves the problems of semantic ambiguity and context dependence, provides an efficient tool for the digital protection of cultural heritage, and has important academic and application value.