mempalace.kg_predicate_norm

Source: mempalace/kg_predicate_norm.py

Predicate normalization for the AGE knowledge graph (issue #50).

The LLM triple extractor emits ~1000+ distinct relation_type strings. Three classes of contamination bloat the predicate vocabulary far past the underlying semantic relation count:

1. Code tokens treated as predicates (appendchild, createelement, executemany, setattribute, getelementbyid) — JS/Python API method names the extractor pulled out of source-code drawers. These should be dropped.
2. Near-synonyms not collapsed (is / is_a / is_an_instance_of; was_a / is_a_kind_of) — canonicalized to one relation type.
3. Grammatical fragments with negation/punctuation glued in (don't_adapt, aren't_merged, 'doesn't_appear') — apostrophes and quotes are stripped and the negation polarity is denormalized into a not_<base> form rather than left as an arbitrary contraction.

This is a pure module — no DB, no AGE imports, no network. The single public entry point is :func:normalize_predicate, which returns the canonical predicate string or None to signal "drop this triple". That makes it trivially unit-testable and safe to run as a read-only dry-run pass over the live vocabulary without touching the graph.

Wiring this into the write path is the daemon's choice and must be opt-in; this module never mutates anything on its own.

Functions

normalize_predicate

def normalize_predicate(raw: str) -> Optional[str]

Normalize a raw extractor predicate to its canonical form.

Returns the canonical predicate string, or None to signal that the triple should be dropped (code token, empty, or punctuation-only input).

Pipeline:

1. fold — lowercase, snake_case, strip quotes/apostrophes (class 3 prep)
2. drop if empty or a known/heuristic code or shell token (class 1)
3. drop if a content-free function word / modal (class 1c, issue #45)
4. strip negation prefix, remember polarity (class 3)
5. canonicalize the base via the synonym map (class 2)
6. re-apply not_ prefix if it was negated

The negation prefix is applied after canonicalization so that doesn't_appear and does_not_appear both land on not_appear, and a negated synonym (is not a part of → base a_part_of → part_of) collapses to not_part_of.

Note the base is whatever remains after the negation prefix is peeled: isn't_a strips isnt_ leaving base a (not in SYNONYM_MAP), so it yields not_a, not not_is_a.