2025 Agenda

Aviation safety analysis faces a structural problem, not a data shortage. Critical signals are dispersed across thousands of occurrence reports, written in multiple languages and shaped by inconsistent terminology. Analysts are expected to detect correlations, identify precursors, and anticipate emerging risks, yet the underlying evidence remains fragmented.

Each occurrence combines structured taxonomy fields with unstructured narrative text, but occurrences are largely isolated from one another at the dataset level. The result is a collection of documents, not a connected body of safety knowledge.

Semantic search improves recall by identifying conceptually similar reports, but similarity is not understanding. It cannot trace causal relationships, follow event sequences, or reason across contributing factors, outcomes, and timelines. Moving from retrieval to analysis requires explicit structure: knowledge graphs grounded in domain ontologies.

This talk presents a GraphRAG approach that integrates AI-driven information extraction with knowledge graphs aligned to the ECCAIRS taxonomy. We demonstrate a system that:

• Understands both exploratory questions and precise domain-specific language without relying on exact terminology

• Enables a general-purpose LLM to operate reliably in a highly regulated, technical safety domain

• Connects facts across reports through entity resolution, graph traversal, and relationship-based reasoning

• Supports temporal reasoning over event progressions and response timelines

Using a demonstration based on synthetic European occurrence data, we illustrate how this approach transforms aviation safety occurrences into a navigable semantic space. This enables new forms of exploration, analysis, and sense-making in a domain where complexity, uncertainty, and context are central, reshaping how safety investigations can be conducted at scale.