UnBias-Plus: Detect, Explain, and Rewrite Bias

UnBias-Plus is an open-source software toolkit designed to address the persistent challenge of bias in both human-written and AI-generated content. Unlike many existing moderation tools that focus primarily on binary safety classification, UnBias-Plus provides a granular, interpretable approach to bias analysis. It integrates four core functions: segment-level multi-class bias classification, biased span localization, neutral text rewriting, and automated reasoning for each detected bias.

The system operates via a multi-stage pipeline that processes input text through a fine-tuned language model (currently supporting Qwen3 variants). The model identifies specific biased segments, assigns them a severity score, provides a rationale for the classification, and suggests neutral replacements. The toolkit is designed for accessibility, offering multiple interfaces including a Python package, a command-line interface (CLI), a REST API, and an interactive web application. This flexibility allows researchers and practitioners to integrate bias review into diverse workflows, from automated content moderation to editorial support.

Validation against the BABE news media dataset demonstrates that UnBias-Plus effectively balances bias reduction with the preservation of neutral content. The authors highlight a trade-off between model variants: the Qwen3-8B model excels at conservative bias reduction and maintaining semantic fidelity, while the Qwen3.5-4B model offers higher precision in segment-level localization and rewrite quality. The project provides all necessary components—including the source code, fine-tuned models, and datasets—to support reproducibility and domain-specific adaptation.

By providing an open-source, transparent framework for bias characterization, UnBias-Plus lowers the barrier for developers and researchers to build fairness-aware systems. It moves beyond simple "flagging" to provide actionable explanations and rewrites, which is critical for domains like journalism and education where understanding the "why" behind a bias label is as important as the detection itself.