In order to solve the problems of ambiguous word translation boundary, insufficient utilization of context constraints, and unstable differentiation of candidate terms in scientific and technological texts, this paper proposes a translation discrimination model based on multi-granularity attention mechanism, which integrates word-level, phrase-level, and sentence-level semantic information into a unified discrimination framework, and combines candidate matching and result correction strategies to improve the accuracy of translation selection. Experiments are carried out based on 42 600 bilingual sentence pairs and 18 240 ambiguous word examples in scientific texts. The results show that the Accuracy, Macro-F1 and MRR of the model reach 91.8%, 90.9%and 0.946, respectively, and the average reasoning time is 20.2 ms. The overall performance is better than that of the comparison methods, indicating that the model has good discrimination ability and application value in complex scientific and technological contexts.
