[4Xin2] Poster session 2

[4Xin2-93] Context-Aware Few-Shot Learning for Business Dialogue Translation

〇Masaya Ueda¹, Kengo Taketani¹, Noritaka Okamoto¹, Kei Kikuiri¹ (1.NTT DOCOMO, INC.)

Keywords:Machine Translation, Few-Shot Learning, Context-Aware Translation

In business meetings, there may be opportunities for information exchange among different languages speakers. To facilitate smooth communication in such situations, there are growing needs for machine translation (MT) for dialogue. MT for dialogue is important to understand contextual information because dialogue tends to contain many pronouns and abbreviations, as well as proper nouns. In this paper, we study MT using the Large Language Models for Japanese-English translation in business meetings. In order to evaluate the performance of MT in real meetings, we created a unique dataset by transcribing and translating speech data in real meetings. In order to effectively include contextual information in prompts, we propose contextual information selection methods that focus on semantic similarity and coreference resolution of source language sentences. The results of the evaluation showed that the proposed and existing methods improved the translation quality for each language direction.