Knowledge Transfer in Incremental Learning for Multilingual Neural Machine Translation

• DOI: 10.18653/v1/2023.acl-long.852
• Publication Date: 2023-08-04T20:57:42Z
• Authors: Kaiyu Huang; Peng Li; Jin Ma; Ting Yao; Yang Liu
• Keywords: Knowledge Transfer; Transfer of learning; Language Understanding; Training set

In the real-world scenario, a longstanding goal of multilingual neural machine translation (MNMT) is that single model can incrementally adapt to new language pairs without accessing previous training data. this studies concentrate on overcoming catastrophic forgetting while lacking encouragement learn knowledge from incremental pairs, especially when not related set original languages. To better acquire knowledge, we propose transfer method efficiently MNMT models diverse pairs. The flexibly introduces an external into models, which encourages completing procedure transfer. Moreover, all parameters are frozen ensure qualities degraded. Experimental results show our meanwhile maintaining performance outperforming various strong baselines in learning for MNMT.