Analyzing Thematic Evolution in Literary Texts via Chain-of-Thought Prompting and Dual-Channel Adversarial Modeling

Accurately modeling the topic evolution process in literary texts is a core challenge for achieving automated literary understanding and structured semantic representation. Traditional topic modeling methods, such as LDA and its variants, struggle to capture the deep rhetorical structures and frequent semantic migrations inherent in literary narratives, often leading to fragmented semantic chains and weak structural perception. This paper proposes a literary topic evolution analysis model that integrates a Chain-of-Thought (CoT) prompt chain mechanism with a dual-channel adversarial architecture to address these challenges. The model constructs a layer-by-layer semantic reasoning chain to guide the structured extraction of topic nodes while designing an "original semantic flow" and a "focused prompt flow" to enhance sensitivity to semantic transformations and topic migration paths. An inter-channel adversarial loss and semantic migration contrast mechanism are further introduced to improve explanatory power and modeling robustness. Experiments conducted on four classic Chinese and English literary corpora show that the proposed model achieves an average improvement of 9.5% in Thematic Coherence, and improvements of 8.2% and 10.1% in Transition Consistency and Structural Interpretability, respectively, outperforming mainstream models such as LDA, DTM, and Prompt-BERT. These results demonstrate the effectiveness and versatility of the proposed method across multi-style literary contexts.