Tutorial@NLPCC 2022, Sept 23, 2022
Shujian Huang, Nanjing University, [email protected]
Wenhao Zhu, Nanjing University, [email protected]
One important change for machine translation in the deep learning era is that the translation knowledge are no longer represented in a symbolic way but embedded in the parameters of the neural networks. However, even large scale neural networks cannot learn all the knowledge in the training data, especially for the low frequency events. K-nearest-neighbor machine translation is a retrieval based technique. KNN-MT employs a translation datastore with symbolic translation knowledge to assist neural machine translation models, showing great potential in modeling low-frequenct events, fast adaptation, etc. This talk covers both the basis of KNN-MT and its recent advances, including dynamically integrate the symbolic knowledge into the neural system, how to control the size of the symbolic knowledge base and the interpretability of this framework.