Analyzing the Impact of Carbon Emission in Training Neural Machine Translation Models: A Case Study

As the field of machine learning grows rapidly, a lot of attention has been paid to how training complex models affects the environment. Carbon emissions caused by the computing needs of machine learning algorithms are becoming a big concern. This is because these models need a lot of computing power and energy. The goal of this paper is to find out how training Neural Machine Translation models affects the environment in terms of carbon footprint and to look into ways to reduce that effect. Machine translation, which automatically translates from source to target, is an area of natural language processing where researchers have been actively working for a long time. The performance of Neural Machine Translation (NMT) is enhanced by exploiting Artificial Neural Networks (ANN) in its model implementation. However, NMT is highly data-hungry and it requires longer training time. In this paper, an attempt has been made to estimate carbon emission when the different NMT models are trained in low-resource language pairs such as English to Hindi and English to Bengali language pairs on different hardware configurations. Finally, different alternatives have also been suggested to reduce this carbon emission and thereby its adverse impact on the environment can be minimized.