Development of a machine translation system using federated learning architecture for low resource Indian languages

This paper presents a work on federated learning (FL) based machine translation considering three different language pairs i.e., English to Hindi, English to Assamese and English to Bodo. FL is the process of training individual models on its own dataset at clients’ nodes and updating all the models at the server end to get a global model. It is one of the first works for FL based machine translation for any pair considering English and an Indian Language (i.e., Hindi, Assamese etc.). Dataset is obtained from the national platform for language technology project, and English Assamese IndicNECorp1.0 from wmt2023. A standard transformer encoder decoder model is built for performing the machine translation training. Different FL architectures are built based on the number of clients, i.e., 1 (non federated), 2, 4, 6, 8. Experiments were carried out to see the affect of increasing the number of clients. Federated algorithm based on federated average technique is adapted by taking the mean of all the clients’ parameters value to create a global model at the server. Performance of the model is evaluated considering BLEU, CHrF and TER score, and it is found that increasing the number of clients decreases the model performance. The results show that non-federated learning outperforms FL at 20 rounds. Moreover, after 50 rounds, with fewer participating clients, the FL model exceeds performance of non-Federated Learning models.