In this paper, we first construct a corpus system that can automatically crawl, clean and categorize business texts from the Internet. Then, by introducing the HowNet lexical annotation algorithm and probabilistic sentence alignment model, the corpus is deeply semantically related and structurally aligned to Chinese and English, making it a structured bilingual teaching resource. Based on this, training strategies are designed for oral and written expressions, such as contextualized quiz, role-playing, imitation writing training, and mind mapping-assisted writing, etc., so as to transform the static corpus into an interactive teaching path. A semester-long comparative teaching of 124 students found that the experimental class with AI corpus-assisted instruction had significantly higher overall business English proficiency than the traditionally taught control class, with posttest mean scores of 88.52 and 81.10, and the mean speaking score of students in the experimental class increased from 12.17 to 17.43, far exceeding that of the control class, which was 14.62. The mean score for written expression jumped from 12.27 to 17.29, again significantly higher than the 14.71 of the control class. Statistical analysis of all p-values of 0.000 confirms the significance of the differences. The questionnaire survey shows that more than 80% of the students affirmed that the model is helpful in improving their speaking and writing skills, and more than 90% of the students think that this way of learning is more interesting and easy to learn.
