Quantitative Assessment of Machine-Generated English Translations

In the foreseeable future, machine English translation evaluation will evolve across multiple dimensions. Current mainstream methods extract subword-level semantic representations from system outputs and references using pre-trained models, concatenate them post-pooling, and feed them to feedforward networks for quality prediction. However, this approach neglects deep word–word and sentence–sentence interactions. This study proposes deep differential feature interaction (DDFI), a novel evaluation method that explicitly models cross-system differences within a shared semantic space. DDFI first captures interactions via attention mechanisms, then amplifies these differences to generate deep differential interaction representations, enriching semantic expression. Comprehensive comparisons between DDFI and existing baseline methods are conducted on general-purpose German–English, English–Chinese, and Chinese–English translation tasks, validating the effectiveness of DDFI. Extensive ablation tests are conducted to confirm the reliability of the experimental results.