Knowledge Graphs for Cross-Domain Communication
Rosetta Stone
A 2nd-century BCE granodiorite stele inscribed with the same decree in three scripts—Egyptian hieroglyphs, Demotic, and Ancient Greek—whose discovery in 1799 unlocked the secrets of hieroglyphic writing.
Knowledge Graphs
A 21st-century digital illustration depicting a central knowledge graph hub interconnected with four domains—Human Language, Mathematics, Programming, and AI Systems—whose conceptual clarity positions it as a modern Rosetta Stone, enabling translation and communication across natural, formal, computational, and artificial intelligence languages.
From Graph to Narrative: Location Hierarchy
Understanding spatial relationships is key. This knowledge graph illustrates a "located-in" hierarchy for the Eiffel Tower, mapping its location from Paris within France, then Europe, and finally, on Earth. Each connection clearly defines the relationship, like "is in" or "is on," with node sizes and colors visually representing scale from local to global.
The Story of Connected Places
Prompt to LLM:
Can you write a little story that includes all the information stored in the graph?
Can you write a little story that includes all the information stored in the graph?
LLM Output:
On the pale blue planet Earth, a great landmass known as Europe stretches across the northern hemisphere. Within this continent lies the storied nation state of France, famed for its art, cuisine, and history. At the very heart of France sits Paris, a vibrant city of boulevards and cafés. And rising above that cityscape is the Eiffel Tower, the iron landmark that has come to symbolize the spirit of Paris itself.
On the pale blue planet Earth, a great landmass known as Europe stretches across the northern hemisphere. Within this continent lies the storied nation state of France, famed for its art, cuisine, and history. At the very heart of France sits Paris, a vibrant city of boulevards and cafés. And rising above that cityscape is the Eiffel Tower, the iron landmark that has come to symbolize the spirit of Paris itself.
The LLM enriches the knowledge graph by adding information learned during pretraining. By extracting and validating these new entities and relationships, we enhance our knowledge graph, empowering AI with deeper insights and stronger reasoning abilities.
Yet, knowledge graphs aren’t exhaustive. For instance, the concept "nation state," accurately describing France in the text, was omitted. Whether we choose to include terms like "nation state" depends entirely on the intended purpose of the graph. Just as humans selectively use language, we deliberately exclude certain details from our knowledge graphs to keep them relevant, focused, and effective.
Ancient Navigational Knowledge Graphs
Marshall Islands stick charts are remarkable examples of early knowledge representation. These intricate wooden grids, known as "Rebbilib," "Meddo," or "Mattang," were not maps in the traditional sense, but sophisticated navigational tools. They encoded complex information about ocean swells, wave patterns, and island locations, crucial for long-distance canoe voyages across the Pacific.
By Cullen328 - Own work, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=46844500
Each stick and shell on the chart represented specific relationships and entities – islands as shells, and sticks depicting wave patterns or currents. This rich "graph" of oceanic knowledge was often combined with oral histories and songs, transforming abstract data into vivid narratives that guided mariners across vast stretches of water. They illustrate how complex spatial knowledge can be encoded and transmitted.
The Tabula Peutingeriana: A Roman Road Map
The Tabula Peutingeriana is an extraordinary surviving copy of an ancient Roman road map, detailing the vast network of roads, cities, and significant landmarks of the Roman Empire. It works very much like a modern transit map. Unlike modern maps, it was a long parchment scroll, distorted to emphasize routes and distances rather than geographical accuracy. It served as a vital knowledge graph, connecting key locations and facilitating travel and logistics across the empire.
This remarkable artifact encoded complex spatial and logistical information. Cities, settlements, rivers, and mountains are represented as nodes, while the roads connecting them serve as edges, illustrating how the Romans visualized their interconnected world for efficient communication and movement of goods and armies.
Decision Trees as Knowledge Graphs
Decision Trees are an illustration of how knowledge graphs can contain reasoning information. Each node represents an entity or a decision point, such as "Applicant," "Credit Score," or "Loan Approved." The branches, or edges, signify the relationships and conditions that lead from one state to the next, like "has score > 700" or "income > $50K." This inherent structure makes them a natural fit for knowledge graph representation, providing a clear, interpretable model for automated reasoning.
Consider this simple loan approval tree as an example of how decision logic maps directly to interconnected nodes and edges within a knowledge graph.
Automated decision‐tree construction algorithms represents an early form of “old‐school” AI–driven knowledge‐graph generation, where nodes correspond to attribute tests and edges to branch outcomes—structuring data into an explicit graphical model.
In contrast, modern approaches span a much broader spectrum: from rule‐mining algorithms and embedding‐based link prediction to large‐language‐model–powered extractors that can ingest unstructured text and output richly annotated, semantically linked graphs. These advanced techniques build on the same principle—automating the discovery of structure in data—but now leverage deep learning and contextual understanding to generate knowledge graphs far more flexibly and with nuanced relationships than simple threshold splits.
In contrast, modern approaches span a much broader spectrum: from rule‐mining algorithms and embedding‐based link prediction to large‐language‐model–powered extractors that can ingest unstructured text and output richly annotated, semantically linked graphs. These advanced techniques build on the same principle—automating the discovery of structure in data—but now leverage deep learning and contextual understanding to generate knowledge graphs far more flexibly and with nuanced relationships than simple threshold splits.