CivicXAI-Net: A Lightweight Multi-Output DistilBERT Framework for Explainable Civic Sentiment and Sarcasm Detection

Abstract

With growing reliance on real-time civic engagement systems in India's Smart Cities Mission, public opinion sentiment analysis as feedback has become most vital for responsive governance. But traditional sentiment analysis systems misinterpret sarcasm-laden discourse, especially in civic forums where rhetorical tone and passive aggression dominate citizen complaints. This paper introduces CivicXAI-Net, a lightweight, multi-output DistilBERT-based system that can sense sentiment polarity and sarcasm occurrence in parallel in civic comments. The model uses LIME and SHAP for token-wise explainability and offers interpretability in decision-making in Integrated Command and Control Centres (ICCCs).  Trained on a harmonized dataset of 1,000 civic statements with synthetic, Twitter including a self-annotated sarcasm corpus [5], CivicXAI-Net achieves stable sarcasm detection (~61% accuracy) and offers contextual insights into citizen sentiment. The architecture is optimized for edge deployment, XAI compliance, and domain adaptation for urban governance. A case study in Tumakuru Smart City ICCC demonstrates its feasibility in policy response automation, detection of service dissatisfaction zones, and improving participatory feedback loops.