Aviation Parser: A Knowledge-Guided Self-Evolving Optimization Framework with LLMs for NOTAM Understanding

ACL ARR 2025 February Submission7433 Authors

Abstract: Accurate parsing of Notices to Airmen (NOTAMs) constitutes a critical requirement for aviation safety, yet existing methods suffer from template rigidity that impedes effective handling of non-standard syntax, regional expression ambiguities, and the semantic-practice gap. This paper proposes a knowledge-guided self-evolving optimization framework that integrates Large Language Models (LLMs) with an Aviation Knowledge Graph (AviationKG) to achieve efficient structured NOTAM parsing. The framework comprises three innovative modules: 1) Knowledge-Enhanced Retrieval (KG-TableRAG), which resolves semantic ambiguities through binding of knowledge graph relations with infrastructure tables to constrain search spaces; 2) Self-Evolving Optimization (SEVO), employing dynamic preference alignment and error-driven curriculum learning to iteratively enhance complex instruction compliance; 3) Consensus Inference Engine (CIE), improving edge-case robustness via terminology-preserved input diversification and majority voting decoding. Experimental results demonstrate that our framework achieves a 30.4\% accuracy improvement over the base model within 3-5 iterations on a labeled dataset of 10,000 global NOTAMs, with ablation studies confirming the collaborative efficacy of modular components. This research establishes the first knowledge-driven, continuously optimized LLM solution for aviation text parsing, whose methodology demonstrates extensibility to other high-precision-demanding professional domains.

Paper Type: Long

Research Area: Information Extraction

Research Area Keywords: NLP Applications；Information Extraction

Languages Studied: English

Submission Number: 7433