A tailings dam safety question answering framework integrating a knowledge graph with retrieval augmented generation

Tailings dam accidents pose serious threats to public safety and economic stability, creating a strong demand for intelligent information systems that can support emergency response and safety management. To address this challenge, this study proposes TSA-KA, a domain-specific question-answering framework for tailings dam safety that integrates structured knowledge representation with retrieval-augmented response generation. The framework first constructs a knowledge graph from emergency response plans and related domain documents to provide standardized and reliable knowledge support. To improve query understanding, an intent recognition model based on RoBERTa-TextCNN with an attention mechanism is developed to capture fine-grained semantic patterns in user questions. In addition, a retrieval-augmented generation module is introduced, in which ChatGLM3-6B is fine-tuned using low-rank adaptation to generate contextually appropriate and domain-consistent responses. The novelty of TSA-KA lies in the integration of a tailings-dam-safety knowledge graph, intent recognition, evidence retrieval, and retrieval-augmented generation into a unified domain-specific question-answering framework. The constructed knowledge graph contains 21 entity categories, 23 relation categories, 2,911 entities, and 20,146 relations/triples, and the question-answering experiments were conducted on 2,346 manually constructed question-answer pairs. The RoBERTa-TextCNN-Attention intent recognition model achieved an Aggregate F1 score of 94.12%, while the LoRA-adapted ChatGLM3-6B configuration achieved RAGAS scores of 0.89, 0.82, and 0.87 for answer relevancy, faithfulness, and context precision, respectively, on the common held-out comparative evaluation set. These results suggest that integrating structured domain knowledge with retrieval-augmented generation is a feasible way to improve domain relevance, evidence grounding, and contextual consistency in tailings dam safety question answering.