Unlearning bias in text diffusion models based on decoupled adapter
Abstract
Social biases in text generation models severely compromise technical fairness and social inclusiveness. Existing studies mostly focus on debiasing at the word embedding level, ignoring dynamic biases in generated text, with limitations like low training efficiency and poor generalization. To address this, we propose Debias-Adapter, a bias mitigation method based on text diffusion models. Built on the Latent Diffusion Model (LDM) framework, it introduces a decoupled cross-attention mechanism and adds an adapter module to the Transformer-Encoder, optimizing only
Fields of Study & Subjects
Ethics and Social Impacts of AI
Topic Modeling
Hate Speech and Cyberbullying Detection
Publication details
- Journal / source
- Advances in Engineering Innovation
- Volume
- 17
- Issue
- 1
- Pages
- 80–94
- Publication date
- 14 January 2026