Mod-Guide: A New Approach to Enhancing Cultural Sensitivity of AI Content Moderation for Ethnic Minorities Using RAG Technology

Core Views

This article introduces the Mod-Guide system, which integrates the life experience narratives of ethnic minorities into the LLM content moderation process using Retrieval-Augmented Generation (RAG) technology to address the insufficient recognition of culturally insensitive speech by AI moderation systems.

Original Author & Source

• Original Author/Maintainer: arXiv authors
• Source Platform: arXiv
• Original Title: Mod-Guide: An LLM-based Content Moderation Feedback System to Address Insensitive Speech toward Indigenous Ethnic and Religious Minority Communities
• Original Link: http://arxiv.org/abs/2606.13397v1
• Publication Time: 2026-06-11T14:28:18Z

Keywords

Content Moderation, RAG, Retrieval-Augmented Generation, Ethnic Minorities, AI Ethics, Cultural Sensitivity, LLM, Restorative Justice

Epistemological Dilemma

Current AI moderation systems are trained based on the perspective of mainstream groups and lack understanding of ethnic minorities' cultural contexts, leading to two major issues:

1. Failure to Recognize Subtle Offenses: Some speech is harmless in mainstream contexts but has derogatory meanings for specific groups (due to the absence of ethnic minority voices in training data);
2. Risk of Misjudgment: Over-moderation suppresses legitimate expressions of minority groups, while under-moderation allows harmful speech to go unpunished.

Concealment of Culturally Insensitive Speech

Unlike obvious hate speech, culturally insensitive speech ignores the perspectives of marginalized groups through implicit erasure, misrepresentation, or normative frameworks, causing profound real harm.

Core Design

Mod-Guide is an LLM-based content moderation feedback system whose core innovation is integrating the "life experience narratives" of ethnic minorities into the process. The research targets the Hindu community (the largest religious minority) and the Chakma community (an indigenous ethnic minority) in Bangladesh.

Hermeneutic Inclusion

Ensuring that minority perspectives are understood and considered by AI is not a technical fix but a rethinking of the epistemological foundation of moderation—allowing the experiences of marginalized groups to become a reference for system decisions.

1. Co-Creation of a Culture-Rooted Corpus

Collaborate with community members to build a corpus that includes examples of culturally insensitive speech targeting the groups, using participatory methods to ensure it truly reflects daily experiences, covering a spectrum from subtle misunderstandings to obvious discrimination.

2. RAG-Enhanced Moderation Pipeline

Use Retrieval-Augmented Generation (RAG) technology to retrieve relevant background information and precedents from the community knowledge base during moderation, inputting them as prompts to the LLM to enhance cultural sensitivity without retraining the model.

3. Mixed-Method Evaluation

Invite participants from minority and mainstream groups to evaluate the accuracy of moderation decisions and the perception differences between groups.

Key Results

• RAG-enhanced moderation responses are significantly superior to baseline systems in contextual accuracy;
• Minority participants believe RAG outputs better capture the cultural meaning and potential harm of speech, while mainstream participants may underestimate offensiveness;
• Perception differences challenge the assumption of a "single objective standard" and highlight the need for diverse perspectives.

Restorative Justice

Content moderation should focus on repairing harm rather than just punishing or deleting; it needs to understand the context and nature of harm instead of just matching keywords.

Hermeneutic Inclusion

Break the hermeneutical injustice where marginalized groups lack resources to define their own experiences. By directly incorporating community experiences into technical systems, it provides a path for diverse perspectives to participate in knowledge production.

1. Community Participation: Technical solutions need direct collaboration with marginalized groups; otherwise, it is difficult to understand what truly constitutes "insensitive speech";
2. RAG Technical Path: Dynamically inject diverse perspectives as an alternative to pursuing completely unbiased training data (which may be unrealistic);
3. Diverse Evaluation: Single-group evaluation may mask potential harm to specific users, so diverse perspectives must be included.

Limitations

• Only covers two specific communities; the generalizability of results needs verification;
• The RAG method cannot solve all cultural understanding issues; deep knowledge may require model training or architectural adjustments.

Future Directions

• Expand to more communities and cultural contexts;
• Explore fine-grained cultural sensitivity modeling;
• Research privacy-preserving knowledge sharing mechanisms.