Vocabulary Inheritance in CAVEAT

Principle

Each unreliability mode in the CAVEAT hierarchy may have an associated vocabulary file (linked via caveat:lexiconFile). A mode’s effective vocabulary is the union of its own terms and all ancestor terms, collected by traversing the rdfs:subClassOf chain from the mode to the root.

Example

For caveat:BiofieldEnergyHealing, the inheritance chain is:

BiofieldEnergyHealing
  └── lexicon: engagement/biofield-energy-healing.yaml
  └── parent: VitalistEnergyMedicine
        └── lexicon: engagement/vitalist-energy-medicine.yaml
        └── parent: Pseudoscience
              └── lexicon: rejection/pseudoscience-base.yaml
              └── parent: PremiseLevelFailure (no lexicon)
                    └── parent: UnreliabilityMode (no lexicon)

The effective vocabulary for BiofieldEnergyHealing is:

  1. All terms from engagement/biofield-energy-healing.yaml
  2. Plus all terms from engagement/vitalist-energy-medicine.yaml
  3. Plus all terms from rejection/pseudoscience-base.yaml

Resolution Algorithm

def resolve_vocabulary(mode_uri, graph):
    """Collect all lexicon files from mode to root."""
    lexicon_files = []
    current = mode_uri
    while current:
        lexicon = graph.value(current, CAVEAT.lexiconFile)
        if lexicon:
            lexicon_files.append(str(lexicon))
        parents = list(graph.objects(current, RDFS.subClassOf))
        # Follow only the CAVEAT hierarchy (skip BFO parents)
        current = next(
            (p for p in parents
             if (p, RDFS.subClassOf, CAVEAT.UnreliabilityMode) in graph
             or p == CAVEAT.UnreliabilityMode),
            None
        )
    return lexicon_files

The script scripts/resolve_vocabulary.py implements this and outputs a flat merged vocabulary for any given mode.

Term Precedence

When the same term appears at multiple levels (e.g., a term defined at the parent level and redefined at the child level), the more specific (child) definition takes precedence. In practice, this should be rare — child lexicons should add terms, not redefine parent terms.

Design Rationale

Top-down inheritance ensures that:

  1. Shared terms are defined once. Rejection terms like “pseudoscience” apply to all pseudoscience subtopics and are defined in pseudoscience-base.yaml, not repeated in every subtopic lexicon.

  2. Specificity increases with depth. The root provides broad terms; leaves provide narrow, corpus-specific terms. This mirrors how the unreliability modes themselves become more specific with depth.

  3. New subtopics get a useful vocabulary immediately. A newly added pseudoscience subtopic inherits all parent and grandparent terms before any subtopic-specific terms are added.