Ensuring both occupational and process safety in the iron and steel industry is a critical challenge. To pursue this objective, the analysis of past accidents can make a significant contribution. However, manual investigation of extensive incident records to uncover causes and their interrelations is often impractical, underscoring the need for automatic tools capable of extracting actionable insights for developing safety recommendations. Moreover, incident descriptions are frequently short and noisy, requiring expert intervention to infer meaningful information. In this work, we introduce a framework for gathering and processing incident data to identify significant causes, features, and patterns. Based on hierarchical clustering, our solution integrates domain-expert interaction to implement a human-in-the-loop learning framework. Experiments conducted on a real-world case study validate the effectiveness of the proposed approach.