ITSC 2025 Paper Abstract

Paper TH-EA-T22.5

Xin, Shiwei (Air Traffic Management Research Institute, (Nanyang Technologica), LI, HU (School of Mechanical and Aerospace Engineering (Nangyang Technol), Yang, Zeru (School of Electrical and Electronic Engineering (Nangyang Techno), Zhou, Yehong (Nanyang Technological University), Dong, Liang (Air Traffic Management Research Institute, (Nanyang Technologica), Chen, Chun-Hsien (School of Mechanical and Aerospace Engineering (Nangyang Technol)

LLM-Based Triple Extraction with Class-Constrained Structural Prompting for Phase-Aware Knowledge Graphs in Aviation Incident Analysis

Scheduled for presentation during the Invited Session "S22b-Emerging Trends in AV Research" (TH-EA-T22), Thursday, November 20, 2025, 14:50−14:50, Coolangata 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Keywords Real-time Coordination of Air, Road, and Rail Transport for Incident Management, AI, Machine Learning and Predictive Analytics for Traffic Incident Detection and Management, Data Analytics and Real-time Decision Making for Autonomous Traffic Management

Abstract

Runway incursions and excursions remain a leading cause of aviation accidents, yet timely risk assessment is hindered by the free-form style of incident reports. Existing frameworks such as Human Factors Analysis and Classification System (HFACS) depend on manual annotation, limiting scalability and consistency in rapidly changing operational contexts. To close this gap, a Class-Constrained Structured Prompting (CCSP) pipeline converts raw PDF narratives into subject–relation–object triples, verifies them through a dual channel that fuses semantic similarity with Large Language Models (LLM) factual checks, and embeds the validated knowledge in a phase-aware sequential knowledge graph (KG). Results demonstrate statistically significant improvements in triple extraction quality across factual accuracy, structural coherence, and groundedness metrics with schema-aware and attribute-augmented prompting. Furthermore, the phase-aware mapping sequence consolidates 100 lexical variants into a controlled vocabulary covering the nine The International Civil Aviation Organization (ICAO) flight phases, yielding a standardised table that reduces sparsity and enables fine-grained temporal modelling. These findings significantly enhance structured information extraction from unstructured aviation safety narratives, enabling more effective real-time analysis of runway-related incidents.