Exploiting Target Language Data for Neural Machine Translation Beyond Back Translation

Neural Machine Translation (NMT) suffers 001 from the challenges of translating in new do-002 mains and low-resource languages. To address 003 these challenges, researchers have proposed 004 methods to incorporate additional knowledge 005 into NMT, including the integration of transla-006 tion memories (TMs). However, finding TMs 007 that closely match the input sentence remains 008 difficult, particularly for specific domains. In 009 contrast, monolingual data is widely available 010 in most languages and back-translation is be-011 lieved as a promising method to utilize target 012 language data. But, it still needs additional 013 training. In this paper, we propose Pseudo-014 kNN-MT, a method that exploit target language 015 data during the inference phase, without train-016 ing the NMT model. Also, we further inves-017 tigate the assistance of large language model 018 (LLM) in NMT. Experimental results show that 019 our method can improve translation quality by 020 a great margin. Interestingly, LLMs are found 021 to be helpful for strong NMT systems. 022 1 Introduction 023 Neural Machine Translation (NMT) has witnessed 024 significant advancements with the introduction of 025 deep learning techniques(Sutskever et al., 2014;