Semantic Knowledge Graph

Extract knowledge graphs from text using Semantic Encoding — interpretable, Wikipedia-grounded, and fast.

Semantic graph extracted from text about cognition, storytelling, and ADHD. Node size = mention frequency, colors = EPA affect values.

What Makes This Different

Approach	Limitation
Dependency Parsing (spaCy)	Captures grammar, not meaning. "Bank" treated same in "river bank" vs "savings bank"
Embeddings (TransE)	Opaque similarity scores. Can't explain why concepts relate
LLM Extraction (GPT)	Expensive, slow, hallucinates, no grounding

Semantic Encoding approach:

• Position IS meaning — 8192-dim sparse ternary vectors where each dimension = semantic primitive
• Wikipedia grounding — Concepts linked to entities with multi-dimensional positions
• Interpretable — Every relationship traceable: "Einstein → Physics via SCOPE anchor, → 20th century via TEMPORAL dimension"

Installation

git clone https://github.com/rohanvinaik/semantic_knowledge_graph.git
cd semantic_knowledge_graph
pip install -e .

Quick Start

from gse_graph import encode_text, extract_graph, resolve_entities

text = """
Einstein discovered relativity. His work transformed physics
and our understanding of space and time.
"""

# Encode → Extract → Resolve
encodings = encode_text(text)
graph = extract_graph(encodings, text)
graph = resolve_entities(graph)

# Export for visualization
output = graph.to_json()

Cassette System

Swap extraction strategies based on your needs:

Text → encode_text() → TokenEncodings
                            ↓
              ┌─────────────┴─────────────┐
              ↓                           ↓
      ┌──────────────┐            ┌──────────────┐
      │   Grammar    │            │   Semantic   │
      │   Cassette   │            │   Cassette   │
      │              │            │              │
      │ • SVO edges  │            │ • Wiki link  │
      │ • Co-occur   │            │ • Spreading  │
      │ • Fast ~1ms  │            │   activation │
      └──────────────┘            │ • Anchors    │
                                  │ • Rich ~500ms│
                                  └──────────────┘
                            ↓
                    SemanticGraph
                 (nodes, edges, 3D viz)

# Fast structural analysis
graph = extract_graph(encodings, text, cassette="grammar")

# Rich semantic grounding (default)
graph = extract_graph(encodings, text, cassette="semantic")

Output Format

graph.to_json()
# Returns:
{
    "nodes": [
        {
            "id": "Q_Einstein",
            "label": "Albert Einstein",
            "x": 12.5, "y": -3.2, "z": 8.1,
            "color": "rgb(200,180,220)",
            "banks": {"MENTAL": 0.4, "TEMPORAL": 0.3},
            "mentions": 2
        }
    ],
    "edges": [
        {
            "source": "Q_Einstein",
            "target": "Q_Physics",
            "label": "shared:Relativity",
            "type": "MENTAL",
            "weight": 0.49
        }
    ]
}

Visualization

Output designed for 3d-force-graph:

import ForceGraph3D from '3d-force-graph';

ForceGraph3D()(document.getElementById('container'))
    .graphData(gseOutput)
    .nodeColor(n => n.color)
    .nodeVal(n => n.mentions);

See visualization_borges.html for a live interactive example.

Semantic Banks

Bank	Detects	Examples
SUBSTANTIVES	Entities	people, objects, places
ACTION	Processes	create, move, transform
MENTAL	Cognition	think, believe, understand
TEMPORAL	Time	before, after, during
SPATIAL	Location	above, inside, near
LOGICAL	Causation	because, therefore
EVALUATORS	Judgment	good, important

License

MIT