Where neural machine translation and translation memories meet
Sik Feng Cheong · 2024
Hong Kong Government Press Releases ( https://www.w3.org/1999/xlink" xlink:href=" https://www.info.gov.hk/gia/ "> https://www.info.gov.hk/gia/ ) by different bureaux and departments serve as an important channel of communication between the government and the public, and they are usually available in both English and Chinese, the two official languages of Hong Kong. The publication of bilingual press releases could benefit from translation technology, especially with recent advances in neural machine translation (NMT). However, it would be suboptimal to use popular online NMT platforms as they are often designed for general texts and could have issues with terminology and style in the target language. In this chapter, the author proposes the method “pre-translation with translation memories (TMs)” to adapt general NMT engines to the translation of the press releases from English into Chinese. Unlike conventional TMs for human translators, our TMs complement neural translation models, featuring length-ranked bilingual sentences and sub-sentential units for pre-translating the source text prior to NMT. Our experimental results show that our method outperforms the NMT-only baseline in terms of the BLEU score, suggesting that TMs offer a simple solution to the adaptation of general NMT to domain-specific translation tasks, without the need of fine-tuning existing models.