stratify

stratify(mappings: Iterable[SemanticMapping], *, accept_unspecified: bool = True) → Stratification[source]

Stratify mappings into a positive, negative, and predicted (positive) set.

Parameters:

  • mappings – A collection of semantic mappings

  • accept_unspecified – Whether to consider mappings that do not have an explicit justification (i.e., using semapv:UnspecifiedMatching) as having been manually curated

Returns:

A stratification tuple

Semantic mappings are stratified as predicted versus curated based on their mapping justification. Predicted semantic mappings have one of the following:

  • semapv:LexicalMatching

  • semapv:LexicalSimilarityThresholdMatching

  • semapv:LogicalReasoning

  • semapv:SemanticSimilarityThresholdMatching

  • semapv::StructuralMatching

Manually curated semantic mappings have one of the following justifications:

  • semapv:ManualMappingCuration

  • semapv:UnspecifiedMatching (when opted in with the accept_unspecified flag)

Remaining mapping justifications in the Semantic Mapping Vocabulary (SEMAPV) can’t be easily categorized. Semantic mappings are then subcategorized as positive or negative (i.e., when the predicate modifier is set to Not). Note, there are typically no negative predicted semantic mappings because software focuses on producing positive semantic mappings.

If needed, negative mappings can be sampled using techniques based on the open world assumption (OWA) or local closed world assumption (LCWA). The PyKEEN graph machine learning library has detailed documentation on these processes. However, SSSOM-Pydantic focuses on evaluations that don’t consider predicted negative mappings.