Skip to content

Vocabulary Direction Semantics Analysis

Classification of 30 Builtin Types

OUTWARD (from → to) - 27 types

Causation (4 outward): - CAUSES: A directly produces B - ENABLES: A makes B possible - PREVENTS: A blocks B from occurring - INFLUENCES: A affects B

Causation (1 inward): - RESULTS_FROM: A results from B (B causes A) ← INWARD

Logical (4 types): - IMPLIES: A being true makes B necessarily true - CONTRADICTS: A and B cannot both be true (bidirectional?) - PRESUPPOSES: A assumes B is true - EQUIVALENT_TO: A and B express the same thing (bidirectional)

Structural (5 types): - PART_OF: A is component of B (component → whole) - CONTAINS: A includes B (container → contained) - COMPOSED_OF: A is made from B as material (whole → material) - SUBSET_OF: All A are B (subset → superset) - INSTANCE_OF: A is example of category B (instance → category)

Evidential (3 outward + 1 ambiguous): - SUPPORTS: A provides evidence for B - REFUTES: A provides evidence against B - EXEMPLIFIES: A serves as concrete example of B - MEASURED_BY: A quantified by B (measured ← measurer) - could be INWARD

Similarity (4 types - all bidirectional?): - SIMILAR_TO: A and B share properties - ANALOGOUS_TO: A maps to B metaphorically - CONTRASTS_WITH: A and B differ meaningfully - OPPOSITE_OF: A is inverse of B

Temporal (3 types): - PRECEDES: A happens before B - CONCURRENT_WITH: A and B happen simultaneously (bidirectional) - EVOLVES_INTO: A transforms into B

Functional (4 types): - USED_FOR: A's purpose is to achieve B - REQUIRES: A needs B to function - PRODUCES: A generates B as output - REGULATES: A controls B's behavior

Meta (2 types): - DEFINED_AS: A's meaning is B - CATEGORIZED_AS: A belongs to category B

Proposed Classification

DIRECTION_SEMANTICS = {
    # OUTWARD (from → to): 24 types
    "outward": [
        "CAUSES", "ENABLES", "PREVENTS", "INFLUENCES",
        "IMPLIES", "PRESUPPOSES",
        "PART_OF", "CONTAINS", "COMPOSED_OF", "SUBSET_OF", "INSTANCE_OF",
        "SUPPORTS", "REFUTES", "EXEMPLIFIES",
        "PRECEDES", "EVOLVES_INTO",
        "USED_FOR", "REQUIRES", "PRODUCES", "REGULATES",
        "DEFINED_AS", "CATEGORIZED_AS",
        "COMPLEMENTS",  # From ADR-048
    ],

    # INWARD (from ← to): 2 types
    "inward": [
        "RESULTS_FROM",  # A results from B (B causes A)
        "MEASURED_BY",   # A measured by B (B measures A)
    ],

    # BIDIRECTIONAL (no inherent direction): 4 types
    "bidirectional": [
        "SIMILAR_TO",
        "ANALOGOUS_TO",
        "CONTRASTS_WITH",
        "OPPOSITE_OF",
        "EQUIVALENT_TO",
        "CONTRADICTS",  # A contradicts B = B contradicts A
        "CONCURRENT_WITH",
    ],
}

Benefits

  1. Simple: One property per type, three possible values
  2. Stored once: In vocabulary metadata (both table and :VocabType node)
  3. Prompt clarity: LLM sees grouped types with clear direction
  4. Validation: Can detect reversed relationships
  5. Auto-correction: Can flip edges if direction wrong
  6. Extensible: New custom types get direction from categorization

Example Corrections

Before (GPT-OSS error): 

{
  "from": "false_sense_of_personal_identity",
  "to": "language_and_thought",
  "type": "ENABLED_BY"  // ← Wrong type! Not in vocabulary
}

With direction semantics: 1. Fuzzy matcher finds closest: "ENABLES" (outward) 2. Direction check: Should be "language ENABLES identity" (language → identity) 3. Auto-correct: Flip edge or reject relationship

Implementation Steps

  1. Migration 016: Add direction_semantics column
  2. Seed data: Update 30 builtin types with direction
  3. VocabularyCategorizer: Assign direction to new custom types (heuristic or default to "outward")
  4. Prompt update: Group types by direction in extraction prompt
  5. Validation: Add direction checking in ingestion pipeline