The traditional methods have some limitations when dealing with the complexity and diversity of natural language, which leads to the limitation of classification accuracy. LSTM model can automatically learn the semantic features of text, characterize sequence dependence and time interaction without manual feature extraction, which improves the efficiency and accuracy of classification. It is devoted to the application research of LSTM model in corpus classification, which effectively improves the quality of corpus and provides a solid foundation for the subsequent model training. In feature extraction, advanced word vector model and innovative text vectorization method are used to create favorable conditions for the model to accurately learn semantic features. Based on the algorithm design process, the experiment is carried out on the open IMDB film review dataset and the self-constructed foreign literature corpus, and the LSTM-based algorithm, SVM and naive Bayes classifier are compared and analyzed. The results show that compared with other algorithms, LSTM-based corpus classification algorithm has obvious advantages and can better adapt to the needs of corpus classification tasks.