ADAPT researchers Dr Séamus Lankford and Prof. Andy Way of DCU, and Dr Haithem Afli of MTU, recently published a groundbreaking study in the journal ‘Information’ as part of a special edition on ‘Machine Translation for Conquering Language Barriers’. Titled ‘adaptMLLM: Fine-Tuning Multilingual Language Models on Low-Resource Languages with Integrated LLM Playgrounds’, the research focuses on natural language processing and introduces ‘adaptMLLM’, a new application with the potential to transform machine translation (MT) for low-resource languages, a domain that has remained largely unexplored despite the rapid progress in Multilingual Language Models (MLLMs) and Large Language Models (LLMs).
The research focuses on the development of adaptMLLM, an open-source application dedicated to fine-tuning MLLMs and managing the complete MT workflow for low-resource languages. This tool is designed to be user-friendly, catering to developers, translators, and users who are engaged in MT. It offers a streamlined environment for configuration, easy customisation of hyperparameters, and provides a range of metrics for model evaluation.
Speaking about the research, Dr Séamus Lankford said: “adaptMLLM not only boosts performance but also conserves time and computational resources, opening new horizons for research and practical applications in the field of machine translation.”
The paper underscores the potential of LLMs in revolutionising various fields including language translation, content generation, and more creative applications like music and art generation. The authors highlight the No Language Left Behind (NLLB) project, which focuses on enhancing translations for low-resource languages across platforms and contributing to Wikipedia’s language diversity.
The adaptMLLM application’s architecture involves a modular approach for fine-tuning MLLMs, incorporating stages such as setting up the environment, preparing datasets, parameterisation, and deployment. Additionally, the application includes a “green report” to calculate carbon emissions, promoting responsible and sustainable AI development.
Designed as a platform as a service (PaaS) cloud computing application, adaptMLLM is implemented in Google Colab and offers intuitive Graphical User Interface (GUI) controls, making it accessible for educational and research settings.
The paper concludes with a discussion on future work, including exploring the effects of fine-tuning larger MLLMs, hyperparameter optimisation, and applying adaptMLLM to new shared tasks and competitions. The integration of GPT-3, GPT-4, and BARD playgrounds into adaptMLLM is also anticipated.
Read the full paper here: https://www.mdpi.com/2078-2489/14/12/638 / PDF Version: https://www.mdpi.com/2078-2489/14/12/638/pdf
The application can be freely downloaded on Github at the following link: https://github.com/adaptNMT/adaptMLLM/