Energy-efficient lightweight models are becoming essential for advancing Green Artificial Intelligence (AI) in language technologies. With the increasing demand for natural language applications across devices, reducing computational overhead and energy consumption is critical for sustainable AI development. However, existing methods often rely on large-scale deep learning models that consume excessive resources, limiting deployment on edge devices and contributing to higher carbon footprints. These challenges hinder the broader accessibility and ecological scalability of language technologies. To address these issues, we propose a novel framework called Knowledge Distillation with Sparse Quantization (KD-SQ). This approach compresses large models into compact student models through knowledge distillation, while sparse quantization further reduces redundant parameters, ensuring energy efficiency without significant loss in accuracy. The proposed KD-SQ method is applied to machine translation tasks, enabling efficient deployment on mobile and low-resource platforms. It ensures high translation quality while significantly lowering computation and memory usage, aligning with the principles of Green AI. Findings demonstrate that KD-SQ achieves up to 45% reduction in energy consumption and 38% decrease in memory footprint, while maintaining competitive accuracy compared to existing models. This validates KD-SQ as a sustainable and scalable solution for eco-friendly language technologies.
Parmanand Yadav (Thu,) studied this question.