In the field of natural language processing, how to extract and understand semantic information from large-scale corpora has become a core challenge in semantic analysis. To solve this problem, this paper proposes a corpus-based semantic analysis and computational model, combining a variety of innovative technologies to improve the accuracy and efficiency of semantic processing. First, this paper uses a pre-trained language model based on the BERT (Bidirectional Encoder Representations from Transformers) architecture for text representation, and deeply extracts the potential semantic relationships in the text. Then, combined with a grammatical model based on dependency syntactic analysis, it captures the grammatical structure and semantic role relationships in the text, further improving the understanding of complex language structures. Finally, this paper introduces a hierarchical attention mechanism to enable the BERT model to focus on the key information in the text more effectively. Experimental results show that in the semantic matching task, the accuracy of the BERT model is stable at more than 0.85 under multiple similarity thresholds, and the F1 value remains at around 0.87, which is significantly better than the LSTM (Long Short-Term Memory) and Transformer models; in the semantic reasoning task, BERT remains above 0.75 in high-difficulty tasks; in the long text processing task, BERT's accuracy remains above 0.75 when the text length is 1000 characters; in terms of computational efficiency, BERT's training time is long (500 s), and its reasoning time and memory consumption are large, but its semantic understanding ability is strong. The model proposed in this paper provides an effective method for processing complex semantic tasks and provides optimization directions for future research.