Context Repair with Large Language Models for Document-level Neural Machine Translation

Anonymous

Abstract: Current neural machine translation models generate translation output sentence-by-sentence, where each translation procedure is carried out independently. This sentence-level decoding strategy results in an inherent issue of incoherence. Consequently, considerable effort has been dedicated to document-level machine translation to mitigate the incoherence problem. In this work, we propose a simple and effective technique to repair document context by leveraging the power of large language models. The document-level translation task is decomposed into a sentence-level translation task and a contextual information repair task. we first employ a conventional sentence-level translation model to generate sentence-level translation outputs. Then, we pair these outputs with their corresponding translation references to create few-shot examples. Finally, we utilize a large language model along with these few-shot examples to perform context repair for the test sentences. Experimental results on the Bilingual Web Books test set demonstrate the effectiveness of the proposed approach in document-context translation. Besides, the approach also works with other methods. Further analysis and human evaluation results indicate that the proposed approach outperforms the baseline model in terms of human preference.

Paper Type: short

Research Area: Machine Translation

Contribution Types: Theory

Languages Studied: English,Chinese