Discovering Hidden Risks in Society with Multimodal Social Media Using Large Language Models

Identifying and promptly addressing social risk factors is of great importance for maintaining social stability. Social media posts provide an opportunity to uncover hidden risks related to group consciousness or underlying issues behind societal phenomena through public attention and discourse. However, these risks are diverse and multifaceted, with some being difficult to define in advance, which makes it challenging to analyze them using pre-set algorithms. Previous solutions often utilized predefined algorithm for known risk categories, which cannot handle diverse social risks. Encouragingly, the recent development of general-purpose language models has provided powerful tools for analyzing complex and unknown types of risks. This paper is based on the use of social media data to perform clustering analysis on complex multimodal event data. Then we employ large language models to conduct causal inference on the resulting clusters of events to discover potential social risks. Extensive experiments on social media data have been conducted to verify the effectiveness of the proposed approach, demonstrating its capability to analyze intricate risk factors that were previously difficult to address.