Leveraging ChatGPT and Multilingual Knowledge Graph for Automatic Post-Editing
Min Zhang, Xiaofeng Zhao, Yanqing Zhao, Hao Yang, Xiaosong Qiao, Junhao Zhu, Wenbin Ma, Chang Su, Yilun Liu, Yinglu Li, Minghan Wang, Song Peng, Shimin Tao, Yanfei Jiang · 2023
Recently, ChatGPT has shown promising results for Machine Translation (MT).However, how to apply ChatGPT for Automatic Post-Editing (APE) remains as an open question.In this paper, we propose a novel zero-shot APE method by leveraging ChatGPT and Multilingual Knowledge Graph (MKG).In this method, we use MKG to find incorrectly translated entities, and then generate APE prompts for ChatGPT with these entities and their correct translations provided in MKG, aiming to have ChatGPT automatically correct the mistranslations.To evaluate our method, we construct two test datasets from WMT19 English-Chinese (En-Zh) and English-German (En-De) news translation shared task.Preliminary experimental results demonstrate that our APE method improves the translation accuracy of entities significantly (+29.1% and +7.3% absolute points for En-Zh and En-De respectively) and achieves a 4.2 BLEU improvement on the En-Zh dataset, showing that our method is effective.However, there is a 7.3 BLEU drop on the En-De dataset, for which we will conduct further research.