Explainable AI (XAI) has emerged as an important concept in AI, bridging the gap between complex machine learning models and human understanding.
In this context, the launch of adaptNMT, an innovative open-source application designed to simplify the development and deployment of neural translation models (NMT), marks a step towards achieving explainable neural machine translation (XNMT) while also addressing the growing concern of environmental sustainability in AI model development.
As described by Séamus Lankford, Haithem Afli, and Andy Way from ADAPT in a recent research paper, adaptNMT is a tool tailored for both technical and non-technical users in the field of machine translation (MT). It is built upon the widely adopted OpenNMT framework and offers a platform for creating, training, and deploying RNN and Transformer NMT models.
The system is designed to run either in the cloud as a Colab instance using Google Cloud or using local infrastructure. Models are trained using parallel corpora and real-time monitoring is enabled through visualization and extensive logging.
The architecture supports the development of models using vanilla RNN-based NMT, Transformer-based approaches, and soon transfer through fine-tuning. Translation and evaluation can be carried out using either single models or ensembles.
One of its strengths lies in its modular approach, breaking down the complex stages of the typical NMT process — such as environment setup, dataset preparation, training of subword models, parameterizing and training of main models, evaluation and deployment — into independent, manageable steps.
This modularity not only enhances the comprehension of complex processes but also provides an accessible gateway for those new to the field. “It is hoped that such work will be of particular benefit to newcomers to the field of Machine Translation (MT) and in particular to those who wish to learn more about NMT,” said the researchers.
The user-friendly interface of adaptNMT empowers users to visualize model training progress, customize hyperparameters, and deploy models seamlessly. This not only makes complicated NMT models easier to understand but also helps researchers and experts improve their models in a practical way.
Furthermore, adaptNMT emphasizes environmental sustainability in AI research. It introduces a “green report” that calculates and presents information about the power consumption and carbon emissions associated with model development.
Lankford, Afli, and Way emphasized that “it is also a very cost-effective option for those working in the domain of low-resource languages since developing smaller models require shorter training times.”
Exploring LLM Fine-Tuning
In addition, they presented future plans for adaptNMT, including integrating modern learning methods like zero-shot and few-shot learning, similar to those used in advanced models like GPT-3 and Facebook LASER. Additionally, they intend to develop a separate application called adaptLLM, focused on fine-tuning large language models (LLMs) for low-resource languages, demonstrating a commitment to addressing language technology challenges.
The researchers also plan to develop the green report feature further, incorporating an improved user interface (UI) and user recommendations for developing greener models. “As an open-source project, we hope the community will add to its development by contributing new ideas and improvements,” they concluded.
Note: adaptNMT is freely available on GitHub at http://github.com/adaptNMT.
Your information will not be shared with third parties. No Spam.