Automatic Social Risk Identification in Infrastructure Megaprojects: A Hybrid Retrieval-Augmented Generation Approach

Abstract

Social risks are those potentially undermining social stability or social order. Social risk identification is critical in infrastructure megaprojects because its occurrence would lead to broad societal impacts. Yet managers often misjudge social risks due to cognitive biases and difficulty in processing all available information. This study develops a hybrid retrieval-augmented generation model that identifies social risks from multisource projects and policy documents, and automatically generates causal chains linking risk sources to events, consequences, and responses. Evaluation results indicate that the model produces more complete risk chains than expert reports and outperforms DeepSeek on the same risk identification tasks.