A zero-shot LLM framework for multimodal grievance classification, urgency scoring, and abuse detection in civic feedback systems

Abstract

A unified model is presented for civic grievance redressal, integrating multimodal complaint intake, zero-shot semantic routing, sentiment-derived urgency estimation, and behavior-sensitive abuse detection within a scalable microservice architecture. The framework consolidates components that are typically handled independently by combining transformer-based text processing, CTC-enabled speech transcription, affective-intensity modeling, and longitudinal user-behavior analysis into a coherent decision pipeline. Typed and spoken complaints are projected into a shared semantic representation using a MobileBERT zero-shot classifier, while a recurrent neural network trained with Connectionist Temporal Classification (CTC) provides robust transcription of multilingual and dialect-rich voice submissions. Urgency indicators obtained from lexicon-based sentiment analysis are incorporated into time-aware escalation logic, and abuse mitigation integrates toxicity scores with a repetition-weighted behavioral model to identify and regulate systematic misuse. The platform operates as a containerized microservice ecosystem with WebSocket-enabled real-time updates and AES-encrypted data storage. Experiments conducted on a 1000-sample multimodal dataset show consistent performance, including 92.4% routing accuracy, 0.041 MAE in urgency estimation, 96.2% toxicity precision, 96.8% SLA compliance, and sub-150 ms end-to-end latency. These outcomes indicate suitability for deployment in linguistically diverse and resource-constrained civic environments. Planned extensions include enhanced multilingual ASR, adversarially robust toxicity modeling, and incorporation of image-based grievance modalities.