Optimization of Intelligent English Translation Quality Using Artificial Neural Networks for College English Teaching Applications

This study proposes a cross-modal consistency-enhanced translation model that constructs cross-modal encoder–decoder collaboration within the transformer framework. On the encoder side, text and visual objects are aligned at the entity level, incorporating cross-attention and a shared representation space. Threshold-based multimodal graph filtering identifies key objects and establishes word–object edges. On the decoder side, consistency verification is performed based on the semantic similarity between generated words and image labels, along with graph constraints, to suppress omissions and erroneous references. The model was evaluated on the 2016 and 2017 Workshop on Machine Translation datasets and the Microsoft Common Objects in Context datasets. The results demonstrate that effective integration of vision and language can greatly enhance translation accuracy and semantic consistency, offering a new direction for multimodal intelligent translation.