Efficiently Adapting to New Domains via Resizable Datastores in Machine Translation

Domain adaptation, a fundamental task in deep learning, aims to mitigate performance degradation caused by domain shifts. Domain adaptation has been a significant focus in the field of neural machine translation, and ongoing research is dedicated to addressing this challenge. The existing kNN-MT-based domain adaptation methodologies are inefficient because of the large size of the datastore. Subsequent studies attempted to reduce the size of the datastore to shorten the search time; however, a trade-off between size and performance existed. In this study, we propose adaptively resizing datastore for k-nearest neighbor machine translation (ARK-MT), a methodology that adaptively adjusts the size of the datastore according to the confidence scores of the translation model. The experimental findings demonstrate that, compared with existing methods, the ARK-MT model exhibited superior performance across all four domains examined. Moreover, the results validated that ARK-MT reduces the inference time through efficient computational operations.