Sparse Wiki Grounding

When an LLM says "Einstein invented the telephone," can you explain why that's wrong and who actually did?

This system provides interpretable knowledge grounding with explicit reasoning chains. Every association has a type (SCOPE, HISTORY, KNOWN_FOR) and a weight - auditable and correctable.

---

The Problem

LLMs encode world knowledge in opaque parameters. When they hallucinate, we observe the output is wrong but can't explain why.

Approach	Detects Errors	Explains Why	Provides Correction
LLM Confidence	Unreliable	No	No
Embedding Similarity	Sometimes	No	No
RAG Retrieval	Sometimes	Partial	Sometimes
This System	Yes	Yes	Yes

---

How It Works

from wiki_grounding import EntityStore, SpreadingActivation

store = EntityStore("data/entities_demo.db")
einstein = store.search("Albert Einstein")[0]
anchors = store.get_entity_anchors(einstein.entity.id)

for anchor_id, label, category, weight in anchors[:5]:
    print(f"  {category}: {label} (weight: {weight:.2f})")

# Output:
#   SCOPE: Physics (1.00)
#   SCOPE: Philosophy (1.00)
#   SCOPE: Quantum mechanics (1.00)
#   KNOWN_FOR: Relativity (0.70)

Einstein → Physics is an explicit SCOPE anchor with weight 1.0. No "telephone" anchor exists.

Spreading Activation

spreader = SpreadingActivation(store)
activated = spreader.spread(einstein.entity.id, use_anchors=True)

for result in activated[:3]:
    print(f"  {result.entity.entity.label}: {result.activation:.3f}")

# History of philosophy: 0.648 via anchor:Philosophy
# Game theory: 0.420 via anchor:Philosophy

The activation path is visible. Compare to embeddings where cosine_sim(Einstein, Philosophy) = 0.7 tells you nothing about why.

---

Data Coverage

Table	Count
Entities	10,082
Entity Links	38,941
Anchor Dictionary	15,433
Entity-Anchor Links	202,052

Wikipedia vital articles (levels 1-4) with typed semantic connections.

---

Performance

Operation	Latency
Links only	~20ms
With anchor layer	~76ms

The anchor layer adds latency but provides typed connectivity - you know why entities are connected.

---

Quick Start

git clone https://github.com/rohan-vinaik/sparse-wiki-grounding
cd sparse-wiki-grounding
pip install -e .

PYTHONPATH=src python examples/explore_entity.py "Marie Curie"

---

Failure Mode Detection

When an LLM claims "Einstein invented the telephone":

1. Look up Einstein's anchors → no "telephone" or "invention" anchor
2. Look up telephone's anchors → find "Alexander Graham Bell"
3. Provide correction with evidence

The explanation is the data structure itself.

---

Foundation

• Collins & Loftus (1975): Spreading activation in semantic memory
• Ranganathan: Multi-dimensional faceted classification
• Wierzbicka: Semantic primitives and decomposition

---

MIT License