Hybrid NLP Architecture for Crisis-Aware Chatbot Integrating Emotion Classification and Retrieval-Augmented Legal Response Generation

Abstract

Survivors of domestic violence frequently face barriers to accessing timely legal and emotional support, underscoring the need for intelligent, accessible assistance systems. This paper presents a legally capable and emotionally aware conversational system designed to classify user-generated text and provide contextually rich responses in the domain of domestic violence. The system employs BERT-based transformer models to perform three layers of classification on pre-processed user input: crisis classification (safe, urgent, or emergency), intent classification (legal or emotional), and emotion classification across seven fine-grained psychological categories. A Retrieval-Augmented Generation (RAG) module then retrieves relevant content from a curated legal knowledge base to ground legal responses in verified documents, reducing hallucination and improving factual reliability. Experimental results demonstrate that the hybrid architecture supports reliable, safety-aware response generation and shows promise as a supportive tool for survivors in crisis situations.