| 07 |
Choose the best translation from diverse candidatesGenerating diverse translation with perturbed kNN-MT 
|
|---|
If machine translation can output a variety of translation candidates, we can easily select the appropriate translation from them according to context or TPO. However, the previous method to generate translation candidates only outputs similar candidates, which cannot help users. To generate accurate and diverse translation candidates, we propose a method to generate candidates by searching a wide range of examples from large-scale data. Our method includes a feature of randomly selecting data from retrieved examples. Thus, our approach probabilistically considers a range of examples, leading to generating accurate and diverse translation candidates. In the future, our method will make it easy to edit translation errors by choosing the correct translation from candidates. It might also be possible to adapt the model to the specific domain by changing searched data for more accurate translation.
[1] Yuto Nishida, Makoto Morishita, Hidetaka Kamigaito, Taro Watanabe, “Generating Diverse Translation with Perturbed kNN-MT,” in Proc. The 29th Annual Meeting of ANLP (in Japanese), 2023.
Makoto Morishita
Linguistic Intelligence Research Group, Innovative Communication Laboratory