Research on Semantic Analysis Model for News Keyword Extraction and Topic Modeling

This paper proposes an innovative semantic analysis model (SKTAM), which combines pre-trained language model and GNN to achieve deep integration of keyword extraction and topic modeling. High-dimensional semantic representation is generated through BERT embedding, and semantically related keywords are extracted by multi-head attention mechanism. At the same time, GNN is used to construct word co-occurrence graph to capture the potential semantic relationship between keywords. Experiments are conducted on CNN/Daily Mail and 20 Newsgroups datasets. The results show that the SKTAM model improves F1-score by 18% and recall rate by 15 % in keyword extraction tasks. In topic modeling tasks, the topic consistency index is improved by 21 % compared with LDA. In addition, the training time of SKTAM is shortened by 25 % compared with traditional methods, showing significant efficiency advantages. Studies have shown that the SKTAM model can effectively improve the accuracy of keyword extraction and the quality of topic modeling, and provides strong support for downstream tasks such as news clustering and recommendation. Future research will explore optimization solutions for multilingual expansion and ultra-large-scale corpus processing.