인공지능을 이용한 조약문 자동번역의 가능성과 한계

초록

Translation of treaty text can be said to be one of the fields most sensitive to linguistic differences. While States parties to a treaty that use languages other than those in the authentic texts generally produce official translations, there are instances where some ‘officially’ authentic texts are actually ‘de facto’ translations. In either case, treaty texts have traditionally been translated by humans; however, with the advancement of automatic translation technology, it appears we have reached a point where the potential of machine translation should be explored. Machine translation has evolved from the early ‘rule-based’ stage to the ‘data-driven’— ‘statistical’ followed by ‘neural’—stage, and based on a mechanism derived from this process, large language models (LLMs) have been developed. In this article, we reviewed the translation of treaty texts by some automatic translators, i.e. Google Translate and DeepL, as well as some large language models, i.e. chat GPT and Gemini. For translation, we selected some articles of the 1945 ICJ Statute and of the 1982 United Nations Convention on the Law of the Sea. In order to compare with the translations of these articles, we also tested the translation of specific provisions of the 1972 Convention on International Liability for Damages Caused by Space Objects. Since this experiment did not utilize all automatic translators or large language models, nor attempted to translate all treaties, we cannot draw definitive conclusions; however, we obtained at least the following two results. First, while large language models such as chat GPT and Gemini generally appear to provide more accurate translations of legal terms than automatic translators like Google Translate or DeepL, even these models sometimes produce inappropriate translations of legal terms, which require consideration of a broad range of contexts and comparison with other authentic texts. Second, there are instances where large language models translate terms other than legal ones inappropriately. In particular, the so-called problem of translation inconsistency is revealed, where the same term in an authentic text is translated differently.

키워드