Hybrid Evaluation of Translation Quality in AI-assisted Language Learning: A Quantitative Approach Integrating BERT and Human Feedback
Wei Zheng · 2025
With the rapid advancement of artificial intelligence (AI), AI-assisted applications in translation have become increasingly prevalent. However, the quality of translations generated by these systems often falls short of desired standards. Consequently, the translation approach has gradually evolved from pure machine translation (MT) to a hybrid model of “MT plus post-editing.” Typically, post-editing is conducted by the original author or experts, introducing a degree of subjectivity. Currently, AI-assisted language processing tools are being leveraged to enhance post-editing quality. In this study, Google’s AI-assisted language learning tool was selected for translation tasks, revealing issues such as mistranslations, missing components, and omissions. From a linguistic perspective, these problems primarily stem from the challenge of accurately representing context in AI-assisted translations. To optimize translations, this research proposes a quantitative approach integrating BERT (Bidirectional Encoder Representations from Transformers) with human feedback. A correction model based on BERT and human feedback is established, which re-scores selected sentences by combining the output probabilities of a grammar error correction model, ultimately generating refined translation outputs. This hybrid method effectively improves translation quality in AI-assisted language learning.