<p>Vocabulary in language often has multiple meanings, and this semantic ambiguity can easily lead to ambiguity in the translation process, affecting the detection of translation errors. Therefore, in order to improve the quality of text translation and enhance the recognition effect of translation errors, this study designs an intelligent detection method for text translation errors based on multi-level textual features and maximum entropy model. Firstly, preprocess the translated text through text segmentation and part of speech tagging; Then, from the aspects of phrase tree structure, word collocation, and context, Latent Dirichlet Allocation (LDA) technique combined with transfer learning is used to extract multi-level textual features of the translated text; Finally, by combining the Transformer architecture with the maximum entropy model, which has the ability to flexibly integrate multi-level textual features, a comprehensive text translation error detection model is constructed. The model includes an embedding layer, an encoder, and a maximum entropy decoder. Under known constraints, the model maximizes its entropy to accurately determine the correct translation among multiple possible translation results, achieving intelligent recognition of translation text errors. The experimental results show that after applying this method, BLEU score of translated texts remain between 93.2 and 95.3, and the recall rate of text translation error detection results is higher than 95.6%, indicating that this method has achieved good application results.</p>