Adaptive Machine Translation with Large Language Models

Adaptive Machine Translation with Large Language Models Yasmin Moslem ADAPT Centre School of Computing Dublin City University Dublin, Ireland yasmin.moslem@adaptcentre.ie Rejwanul Haque ADAPT Centre Department of Computing South East Technological University Carlow, Ireland rejwanul.haque@adaptcentre.ie John D. Kelleher ADAPT Centre School of Computer Science Technological University Dublin Dublin, Ireland john.kelleher@adaptcentre.ie Andy Way ADAPT Centre School of Computing Dublin City University Dublin, Ireland andy.way@adaptcentre.ie Abstract Consistency is a key requirement of high- quality translation. It is especially important to adhere to pre-approved terminology and adapt to corrected translations in domain- specific projects. Machine translation (MT) has achieved significant progress in the area of domain adaptation. However, real-time adaptation remains challenging. Large-scale language models (LLMs) have recently shown interesting capabilities of in-context learning, where they learn to replicate certain input-output text generation patterns, without further fine-tuning. By feeding an LLM at inference time with a prompt that consists of a list of translation pairs, it can then simulate the domain and style characteristics. This work aims to investigate how we can utilize in-context learning to improve real-time adaptive MT. Our extensive experiments show promising results at translation time. For example, LLMs can adapt to a set of in-domain sentence pairs and/or terminology while translating a new sentence. We observe that the translation quality with few-shot in- context learning can surpass that of strong encoder-decoder MT systems, especially for high-resource languages. Moreover, we investigate whether we can combine MT from strong encoder-decoder models with fuzzy matches, which can further improve translation quality, especially for less supported languages. We conduct our experiments across five diverse language pairs, namely English-to-Arabic (EN-AR), English-to-Chinese (EN-ZH), English-to- French (EN-FR), English-to-Kinyarwanda (EN-RW), and English-to-Spanish (EN-ES). © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. Figure 1: Evaluation results for GPT-3.5 zero-shot, and few-shot translation with random context or fuzzy matches. Average scores across EN-AR, EN-ES, EN-FR, and EN-ZH language pairs. While using a random context outperforms zero-shot translation, using fuzzy matches reveals the best results. 1 Introduction Adaptive MT is a type of machine translation that utilizes feedback from users to improve the qual- ity of the translations over time. Feedback usually includes corrections to previous translations, ter- minology and style guides, as well as ratings of the quality of the translations. This can be partic- ularly useful for domain-specific scenarios, where baseline MT systems may have insufficient rele- vant data to accurately translate certain terms or phrases. There are still several challenges to ef- fectively incorporate user feedback into the trans- lation process, especially at inference time. In this work, we use a relatively wide definition of adap- tive MT to refer to learning from similar transla- tions (fuzzy matches) found in approved transla- tion memories (TMs) on the fly (Farajian et al., 2017; Wuebker et al., 2018; Peris and Casacuberta, 2019; Etchegoyhen et al., 2021), as well as real- time terminology-constrained MT (Hokamp and Liu, 2017; Post and Vilar, 2018; Dinu et al., 2019; Michon et al., …

Content truncated to 500 words