To address the fragmented use of technical terms in vocational college English teaching and to reduce students’ grammatical errors, this study proposes a Bidirectional Long Short-Term Memory with Domain Adaptation (BiLSTM-DA) model, built upon the Long Short-Term Memory (LSTM) architecture. The model incorporates a domain adaptation mechanism and utilizes dynamic word embeddings to generate context-aware representations of specialized vocabulary. It also integrates an error-driven attention mechanism to accurately identify common grammatical mistakes. Based on this model, an intelligent English teaching framework tailored to the needs of vocational education is developed. In addition, a three-dimensional evaluation system is constructed, encompassing technological, pedagogical, and ecological dimensions. Experimental results show that: (1) Compared with mainstream models, BiLSTM-DA achieved an error coverage rate of 93.7%, exceeding Robustly Optimized Bidirectional Encoder Representations from Transformers for Education Data Understanding (RoBERTa-EDU) by 8.2%. For long-tail domain-specific term recognition, it reached an F1-score of 85.4%, representing a 14.6% improvement over Text Graph Convolutional Network with Global and Local Topology (TextGCN-GLT). Moreover, its response latency remained within 0.8 s, significantly lower than the 1.8 s of Decoding-Enhanced BERT with Attention Masking, Version 3 (DeBERTa-V3), thus more effectively supporting the real-time interaction requirements of vocational English classrooms. (2) The BiLSTM-DA-based teaching model enhances students’ recognition of technical terms, grammatical error detection, knowledge internalization, and skill transfer. This model also reduces teachers’ grading workload, alleviates students’ learning anxiety, and promotes educational equity. This study enriches the theoretical and methodological foundations of vocational English instruction and provides practical guidance for integrating artificial intelligence into vocational education.