Semi-Automated Knowledge Engineering and Process Mapping for Total Airport Management

This paper proposes a semi-automated framework that combines symbolic Knowledge Engineering with generative Large Language Models to construct a traceable, domain-grounded Knowledge Graph for Total Airport Management, effectively overcoming data silos and semantic inconsistencies while demonstrating that document-level processing enhances the recovery of complex procedural dependencies.

The Gist

Imagine an airport as a massive, bustling orchestra. You have pilots, air traffic controllers, baggage handlers, fuel trucks, and gate agents. Each group speaks its own dialect, follows its own sheet music, and often doesn't know what the others are doing until it's too late.

This paper is about building a universal translator and a master conductor for this orchestra, using a special kind of "smart robot" (AI) that doesn't just guess, but actually knows where it got its information.

Here is the breakdown of the paper in simple terms:

1. The Problem: The "Lost in Translation" Airport

Airports are full of complex documents (manuals, regulations, logs) written in heavy jargon.

• The Silo Effect: The baggage team's data is locked in one system, the pilots' data in another, and the fuel team's in a third. They don't talk to each other.
• The Jargon Trap: If a pilot says "OK" and a controller hears "Takeoff," disaster can happen (like the famous Tenerife crash).
• The AI Risk: We want to use modern AI (Large Language Models) to read all these documents and create a clear map of how the airport works. But, standard AI is like a confident storyteller who sometimes makes things up (hallucinations). In aviation, you can't have a robot making things up about when a plane should take off.

2. The Solution: The "Scaffolded" Robot

The authors built a system called LangExtract to solve this. Think of it like building a house:

• The Blueprint (Knowledge Engineering): Before the robot starts, the experts (humans) draw a strict blueprint. This is a "Knowledge Graph"—a digital map that defines exactly what things are (e.g., "Baggage Handler," "Refueling") and how they connect.
• The Scaffolding: Instead of letting the robot run wild, they put up "scaffolding" around it. They give the robot the blueprint and say, "Only build things that fit this blueprint."
• The Result: The robot reads the messy documents and fills in the blueprint with facts, but it must point to the exact sentence in the original document where it found that fact. It's like a student doing a research paper who has to cite every single quote.

3. The Big Surprise: Reading the Whole Book vs. Reading Page-by-Page

Usually, when you ask an AI to read a very long document (like a 16-page airport manual), it tends to get confused in the middle. It's like trying to remember a 100-page story; you might forget the beginning by the time you reach the end. This is called the "Lost in the Middle" problem.

• The Experiment: The team tested two ways of reading:
  1. Page-by-Page: Reading one page at a time (Short Context).
  2. The Whole Book: Reading all 16 pages at once (Long Context).
• The Twist: They expected the "Whole Book" method to fail. Instead, it worked better!
• Why? Airport procedures are like a chain reaction. To understand why a plane is delayed, you need to see the fuel truck, the baggage, and the pilot all in the same context. By reading the whole document, the AI could see the connections between the beginning and the end, fixing errors that happened when it only looked at small chunks.

4. The Output: The "Swimlane" Map

Once the robot builds this perfect, verified map of facts, the system automatically turns it into a Swimlane Diagram.

• What is it? Imagine a flowchart where the "lanes" are the different teams (Pilot Lane, Ground Crew Lane, Controller Lane).
• The Magic: The system draws arrows showing exactly who does what and when. If a step is missing or out of order, the map shows it clearly.
• The Safety Net: Every single arrow and box in this map is linked back to the original sentence in the manual. If a manager asks, "Why did you draw this step?" the system can instantly show the source text.

5. Why This Matters

This isn't just about making pretty charts. It's about safety and trust.

• No Black Boxes: Unlike other AI tools that give you an answer and say "trust me," this system says, "Here is the answer, and here is the exact page in the manual that proves it."
• Total Airport Management (TAM): This helps the whole airport run as one unit. If the weather changes, the system can instantly see how that ripples through the fuel, baggage, and pilot schedules, helping everyone react faster.

In a Nutshell

The authors created a "Smart Librarian" for airports. This librarian reads all the confusing manuals, organizes them into a clear, step-by-step map of who does what, and—most importantly—puts a sticky note on every single fact pointing to the original source. This ensures that when the airport runs on this new system, it's not just efficient; it's safe, transparent, and 100% verifiable.

Technical Summary

1. Problem Statement

The paper addresses the critical challenge of data silos and semantic inconsistency within Total Airport Management (TAM). Airport operations involve multiple stakeholders (airlines, ground handlers, ATC) who generate vast amounts of unstructured data in technical reports, manuals, and logs.

• Barriers: Information is often locked in proprietary systems, uses conflicting terminology, and lacks interoperability.
• Consequences: This fragmentation leads to miscommunication, which can have catastrophic safety implications (e.g., the 1977 Tenerife disaster).
• Limitations of Current Tech:
  • Traditional Knowledge Engineering (KE) is too slow and expensive due to manual curation.
  • Standard Large Language Models (LLMs) act as "black boxes," lacking data provenance (traceability to source text) and deterministic rigor required for safety-critical aviation environments.
  • Existing LLMs often suffer from the "lost-in-the-middle" phenomenon in long contexts, degrading performance on complex, non-linear procedural workflows.

2. Methodology

The authors propose a Scaffolded Symbolic Fusion framework that combines expert-curated symbolic knowledge with generative AI. The pipeline consists of four main stages:

A. Knowledge Graph (KG) Schema Design

• Foundation: The schema is built upon the NASA Air Traffic Management (ATM) ontology and Changi Airport Petri maps.
• Structure: Implemented in Protégé, the schema defines classes (e.g., Procedure, Sequenced Item, Stakeholder) and object properties (e.g., hasNext, hasStakeholder).
• Strategy: A modular disclosure strategy is used to handle proprietary data, sharing only high-level structures initially and expanding as needed.

B. The LangExtract Framework

The core extraction engine uses LangExtract (a Google library) to bridge unstructured text and structured KGs.

• Scaffolded Fusion: Expert-curated KG structures are used to generate few-shot prompts and system instructions. This constrains the LLM's probabilistic outputs to adhere to the formal domain schema.
• Context Window Analysis: The study compares two inference modes on the 16-page EUROCONTROL A-CDM manual:
  1. Short-Context (Page-level): Processing single pages to maximize attention density.
  2. Long-Context (Document-level): Processing the entire document to capture global dependencies.
• Deterministic Anchoring: To prevent hallucinations, the system fuses probabilistic LLM outputs with deterministic string alignment algorithms (text.find() and difflib.SequenceMatcher). This ensures every extracted triple is anchored to a specific sentence in the source text.

C. Operational Artifact Synthesis

• Automated Swimlane Diagrams: An algorithmic pipeline converts the populated KG into visual process maps.
• Algorithm: Uses a modified Breadth-First Search (BFS) topological sort to determine node depth (vertical positioning) and maps nodes to stakeholder lanes (horizontal positioning).
• Traceability: Every node in the generated diagram is linked back to its source sentence coordinates, creating an auditable trail.

3. Key Contributions

1. Scaffolded Symbolic Fusion Strategy: A novel method where expert ontologies guide LLM prompts, ensuring semantic alignment and reducing hallucinations in complex domains.
2. Empirical Insight on Context Windows: Contrary to the prevailing belief that long contexts degrade LLM performance ("lost-in-the-middle"), the study demonstrates that document-level processing significantly improves the recovery of non-linear procedural dependencies in airport workflows.
3. Provenance-First Architecture: The framework guarantees absolute traceability by anchoring every extracted knowledge triple to its ground source text, addressing the "black-box" limitation of generative AI in safety-critical sectors.
4. Automated Process Visualization: The ability to automatically generate stakeholder-attributed swimlane diagrams directly from unstructured text, bridging the gap between raw data and actionable operational tools.

4. Experimental Results

The framework was evaluated on the EUROCONTROL A-CDM Manual (16 pages, ~10k tokens) using the gemini-2.5-flash model.

• Performance Metrics:
  • Short-Context (Page-level): Precision (0.961), Recall (0.971), F1 (0.966).
  • Long-Context (Document-level): Precision (0.967), Recall (0.982), F1 (0.975).
• Key Findings:
  • Document-level processing outperformed page-level processing, reducing False Negatives (FN) by 38.5% (from 13 to 8).
  • Reason: The expanded context allowed the model to resolve logical inversions (where effects are described before causes) and maintain coherence across process boundaries that span multiple pages.
  • Provenance: 100% of extracted triples were successfully mapped to source sentences with verifiable audit trails.
  • Alignment: "Match Fuzzy" (semantic mapping) had a slightly higher false positive rate than "Match Exact," but "Match Lesser" produced zero false positives, indicating the model is conservative with partial mappings.

5. Significance and Future Work

• Significance: This work provides a practical, auditable pathway for integrating LLMs into safety-critical aviation operations. It moves beyond conversational AI to create machine-readable, verifiable operational models that support the Total Airport Management (TAM) initiative.
• Future Directions:
  • Real-time Verification: Integrating multimodal data (video analytics, transponder signals) to detect procedural deviations against the KG in real-time.
  • Entity Resolution: Using Named Entity Recognition (NER) and Hidden Markov Models (HMM) to map low-level sensor data to high-level KG entities.
  • Feedback Loops: Using live operational data to continuously refine and update the KG and documented procedures.

In summary, the paper presents a robust, semi-automated framework that leverages the scalability of LLMs while enforcing the rigor and traceability required for complex, multi-stakeholder airport management.