Mitigating Text Mislabeling in Neural Machine Translation with Multitask Learning Techniques

Text mislabeling is a prevalent issue in neural machine translation (NMT) systems, frequently resulting in inaccurate translations and diminished system performance. This paper introduces a novel approach to alleviate text mislabeling in NMT through the application of multitask learning (MTL) techniques. By concurrently training the translation model on multiple related tasks, such as sentence-level classification and token alignment, we aim to enhance the quality of translated text and mitigate the adverse effects of mislabeling. Experimental results on benchmark datasets reveal that the MTL-based model surpasses traditional NMT models, particularly in managing noisy or mislabeled data. These findings underscore the potential of MTL to boost translation accuracy and robustness in practical applications.