Improving Bias Detection in NLP Models through Multitask Learning with Cloze and Random Mask Tasks
Kexin Weng · 2024
This paper applies a multitask learning approach designed to enhance the understanding of textual bias in natural language processing (NLP) models. By combining Cloze and Random Mask (RM) tasks, the study aims to train models that can recognize and bias better within context. The WinoBias dataset is used to verify proposed method, which includes both stereotypical and counter-stereotypical sentences. Applying model under the multitask learning framework allows it to understand diverse contextual scenarios, and then improving its ability to detect biases that might be overlooked. Experimental results evaluated on classification task indicate that this approach significantly improve the model’s performance in bias detection. The study offers a foundation for future work on reducing bias in various text analysis applications. The findings highlight the importance of developing methodologies for detecting bias in automated text processing systems, which ensures more accurate and fair outcomes in real-world applications.