<p>The dynamic evolution of consumer demand is crucial in optimizing their product strategies. However, existing topic modeling approaches often face two challenges when processing consumer opinion texts: insufficient topic identification accuracy and inadequate topic evolution patterns. These challenges stem from the high-dimensional sparsity in feature space and unbalanced topic distribution in time dimension. To address these challenges, this study proposes a two-stage dynamic topic modeling (TDTM) approach, which integrates domain thesaurus construction, temporal-textual embedding, and LABIN spectral clustering to enable efficient dynamic topic identification and in-depth mining of demand evolution patterns. In the offline stage, the TDTM approach combines semantic network analysis with the PageRank algorithm to construct a high-quality domain thesaurus, enhancing the consistency of semantic representation. In the online stage, the TDTM approach extracts temporal semantic features through temporal-textual embedding and optimizes unbalanced topic distributions using LABIN spectral clustering to overcome the limitations of traditional clustering algorithms. Experimental results on two real datasets show that the TDTM approach improves topic coherence (TC) by 11.9% and topic diversity (TD) by 10.49%. Ablation experiments further validate the critical contributions of temporal-textual embedding and LABIN spectral clustering to the improved precision of topic modeling. A visual analysis of demand evolution across different animation series reveals commonalities and differences in consumer demands, providing valuable insights for developing new products in the animation industry. This study enhances the effectiveness of dynamic topic modeling, presents an innovative framework for analyzing dynamic consumer demands, and offers significant theoretical and practical contributions.</p>