By Greentech Apps Foundation (GTAF) Research Team
How should an Islamic app respond to user reviews at scale while maintaining adab, humility, and cultural authenticity?
This question motivated our latest research work, ADAB: A Culturally-Aligned Automated Response Generation Framework for Islamic App Reviews, which has been accepted at the NeurIPS 2025 Workshop on Muslims in Machine Learning (MusIML).
👉 Accepted papers list:
https://www.musiml.org/events/2025-NeurIPS/accepted_papers.html
Why This Research Matters
Modern app stores contain tens of thousands of user reviews. While timely responses are known to improve user trust and engagement, most automated review-response systems are culturally neutral at best and culturally harmful at worst.
For Islamic applications, this gap is especially problematic.
Generic AI responses often:
- Ignore Islamic etiquette (adab al-hiwār)
- Miss religious and cultural context
- Sound transactional rather than sincere
- Fail to acknowledge worship-related sensitivities
Our research shows that cultural alignment is not a “nice to have”. It measurably improves perceived accuracy, relevance, and trust.
What Is ADAB?
ADAB is a culturally aligned AI framework designed specifically for Islamic app review responses. Rather than proposing a new language model, ADAB integrates several state-of-the-art techniques into a pipeline that respects Islamic values.
At a high level, ADAB combines:
- Aspect-Based Sentiment Analysis (ABSA)
To understand what the user is talking about and how they feel (e.g. audio quality, prayer times, Quran recitation). - Hybrid Retrieval-Augmented Generation (RAG)
Combining sparse search (BM25), dense embeddings, and FAISS HNSW indexing to retrieve relevant Islamic etiquette guidance and app-specific documentation. - Agentic Chunking
Instead of naïve token-based chunking, ADAB dynamically segments knowledge into semantically and culturally coherent units (e.g. greetings, manners, worship-related guidance). - Etiquette-Aware Prompt Engineering
Prompts explicitly instruct the model to follow Islamic norms: greetings, gratitude (shukr), patience (ṣabr), humility, and respectful closure.
The result is a system that responds not just correctly, but appropriately.
Evaluation: Humans vs LLMs
To test whether cultural alignment actually matters, we conducted a rigorous evaluation using real Islamic app reviews from GTAF’s ecosystem .
Key findings:
- ADAB responses were preferred in 40% of direct comparisons, versus 15.3% for a baseline LLM.
- Overall performance improved by 9.9%, with the largest gain in:
- Application Specificity: +30.39%
- Improvements in accuracy, relevancy, and specificity were statistically significant.
- Interestingly, LLMs acting as judges performed poorly, showing low agreement with human Muslim evaluators.
This confirms an important insight:
Current LLMs struggle to reliably judge culturally and religiously nuanced content.
Human-centered evaluation remains essential.
Contribution to Islamic Computing
This work contributes to a growing body of Islamic Computing research by demonstrating that:
- Cultural alignment can be systematically engineered
- Islamic values can be embedded without hard-coding theology
- Faith-sensitive AI can be evaluated empirically, not just normatively
We also introduce an open Islamic app review dataset to support future research in this space.
Looking Ahead
Future directions include:
- Scaling human evaluation to a larger portion of the dataset
- Multilingual support (Arabic, Bengali, Urdu, Bahasa Indonesia)
- Adapting the framework for other faith-based and culturally specific applications
- Improving automated evaluation methods for culturally sensitive AI
At GTAF, this work directly informs how we design AI systems for Quran, Hadith, Seerah, and broader Islamic education tools—always prioritising khidmah (service) to the Ummah.
Publication Details
This research was a collaborative effort between the Islamic University of Technology (IUT) and Greentech Apps Foundation (GTAF), conducted as a student thesis project at IUT and co-supervised by the GTAF research team.
Paper title:
ADAB: A Culturally-Aligned Automated Response Generation Framework for Islamic App Reviews by Integrating ABSA and Hybrid RAG
Authors:
K. M. Tahlil Mahfuz Faruk, Mushfiqur Rahman Talha, H. M. Kawsar Ahamad, Mohammad Galib Shams, Nabil Mosharraf Hossain, Syed Rifat Raiyan, Md Kamrul Hasan, Hasan Mahmud, Riasat Islam
Venue:
NeurIPS 2025 – Workshop on Muslims in Machine Learning (MusIML)
📄 Full paper available via the NeurIPS MusIML programme:
https://openreview.net/pdf?id=PnWmDdwTXE
Leave a Reply