Human-in-the-Loop Approaches to Improving Machine Translation

Abstract

This paper provides an overview of various human-in-the-loop methodologies employed to improve machine translation, including interactive translation, post-editing, and active learning. Interactive translation systems enable translators to collaborate with machine translation models in real-time, allowing them to provide feedback, corrections, and suggestions during the translation process. This iterative interaction not only improves the immediate translation output but also enhances the underlying model through continuous learning and adaptation. Post-editing involves human translators refining machine-generated translations to improve their accuracy, fluency, and coherence. By incorporating human expertise, post-editing helps refine machine-generated translations, making them more suitable for specific contexts, domains, or linguistic nuances. Active learning strategies leverage human feedback to iteratively improve machine translation models. By strategically selecting samples for human annotation based on uncertainty or model performance, active learning accelerates the learning process and maximizes the impact of human input on model improvement.