Bias Detection and Optimization of Recruitment Texts Based on BERT

This study proposes an end-to-end framework for automatically detecting and mitigating implicit social bias in recruitment texts, providing an effective technical solution for fair and intelligent information processing in digital recruitment systems. As trustworthy semantic analysis and automated decision support become increasingly important for intelligent communication environments and information transmission applications, improving the neutrality and interpretability of textual content is also valuable for broader engineering-oriented information systems. Based on BERT, a bias detection model integrating BiLSTM and a hierarchical attention mechanism is developed to achieve fine-grained bias classification through word-level and sentence-level attention analysis. Furthermore, a T5-based bias optimization model is fine-tuned using biased–unbiased parallel corpora to accomplish semantic-preserving text rewriting. Experimental results demonstrate that the proposed detection model achieves an accuracy of 0.882 and an F1 score of 0.842 on the test set, outperforming the strongest baseline DeBERTa-v3 by 0.021 in F1 score. Ablation studies further verify that the BiLSTM and hierarchical attention modules improve the F1 score by 0.023 and 0.017, respectively. The optimization model obtains a neutrality score of 4.7 and an information retention score of 4.5 in human evaluation, while achieving BLEU and ROUGE-L scores of 0.65 and 0.71, respectively. The proposed framework provides a reliable and interpretable approach for automated fairness correction of recruitment texts and offers practical guidance for intelligent text processing and trustworthy information management in engineering applications.