Latent topic recognition is a technique for extracting hidden themes or patterns from textual data. And it plays an important role in popular application fields such as text mining and information retrieval. It can improve search accuracy, achieve personalized recommendations and automate document classification. At present, traditional natural language processing methods suffer from problems such as long processing time on large-scale datasets and insufficient consideration of the correlation between words in text, resulting in inaccurate and unstable recognition of potential themes and internal keywords in document sets. To address the aforementioned issues, this paper proposes an improved latent topic recognition model called TR-LDA. It aims to combine the TF-IDF model with the LDA model. At the same time, by integrating the TextRank algorithm, a semantic analysis module is constructed to consider the co-occurrence relationships of words within the text. The experimental results on the THUCNews Dataset indicate that the average F1 score increased from 0.4263 with the original LDA to 0.5127 with TR-LDA. Semantic consistency improved by 54.20%, and perplexity was reduced by 98.74%. Furthermore, there was a time saving of 69.08%.