• KD used for compression and knowledge transfer in MT, shaping supervision and translation quality. • Survey covers 105 papers up to Oct 2025, providing comprehensive landscape of KD4MT. • Categorizes advances by methodological contributions and practical applications, revealing common trends. • Highlights gaps: no unified evaluation, limited benchmarks, and underexplored risk factors. • Offers practical guidelines for selecting KD methods and warns of hallucination, bias amplification. • Discusses impact of large language models reshaping KD4MT research and future directions.
Article Summaries:
- A new survey, “KD4MT: A Survey of Knowledge Distillation for Machine Translation,” reviews 105 papers on the use of knowledge distillation (KD) in MT up to October 2025. The authors explain KD and MT basics, then outline standard KD techniques and categorize advances by methodology and practical application. Their analysis highlights trends, identifies gaps-particularly the lack of unified evaluation standards-and warns of risks such as hallucination and bias amplification. The paper also discusses how large language models are reshaping KD4MT and provides a public database and glossary to aid researchers in selecting and comparing KD methods.
Sources: