What Is a Meronym?

What Is a Meronym?

A meronym is a word that names a constituent part of a larger whole. “Wheel” is a meronym of “car”; “keyboard” is a meronym of “laptop”. The part-to-whole relation is called meronymy, and its inverse — whole-to-part — is holonymy. In NLP, meronymy sits alongside hypernymy and synonymy as one of the lexical-semantic relations encoded in resources such as WordNet and used as a feature in entity extraction, ontology learning, knowledge-graph construction, and structured retrieval. Modern LLMs absorb meronymy implicitly through pre-training, and downstream evaluators test whether the model preserves part-whole relations correctly.

Why It Matters in Production LLM and Agent Systems

Most LLM applications never explicitly compute meronymy, but the relation is silently load-bearing. A RAG system answering “what components does the engine include?” needs the retriever to surface the meronyms of “engine” — pistons, crankshaft, valves — not just synonyms or category siblings. A knowledge-graph-backed agent reasoning about a hospital schedule must distinguish “ward” (a part of the hospital) from “hospital” itself when applying access policies. Get meronymy wrong and the agent answers the right question against the wrong scope.

The pain is most visible in three roles. The RAG engineer sees a low ContextEntityRecall score and traces it to a chunker that split a part-list across pages, breaking the part-whole link. The product manager sees an agent confidently list components that belong to a different product — wheels of a different car model. The compliance lead sees a redaction policy that masks the holonym (“hospital records”) but leaves the meronyms (“ward records”, “shift records”) exposed.

For 2026-era multi-modal agents the relation extends to images and 3D scenes. A vision-language model asked to identify the failed component of a machine part needs to ground meronymy across pixels and text. Without explicit meronym evaluation, vision-LLM regressions silently break diagnostic agents.

How FutureAGI Handles Meronymy

FutureAGI does not encode WordNet or maintain its own meronym table — that’s the role of your knowledge graph or ontology. What FutureAGI does is evaluate whether the model preserves meronymy at the boundary between retrieved context and generated response. The standard pattern is to attach ContextEntityRecall to a Dataset of question/answer pairs where the gold entities encode part-whole pairs, and check that the response retains the parts when the user asks about the whole.

A concrete workflow: an industrial-equipment knowledge agent uses a vector DB plus a domain ontology that encodes meronymy (“turbine” → [“blade”, “shaft”, “casing”]). FutureAGI’s Dataset.add_evaluation runs ContextEntityRecall with expected_entities set to the gold meronyms for each question. A CustomEvaluation wraps an LLM-as-a-judge that scores whether the response correctly maps user questions about the whole to the corresponding parts. Both evaluators feed an offline regression-eval gate so any change to the chunker, retriever, or response prompt is checked against meronym preservation before merge. In production, traceAI emits the retrieved chunk’s entity span list, and a sampled cohort of live traces is re-scored daily — when meronym recall drops below 0.85 the platform team is paged before users hit the regression.

How to Measure or Detect It

Meronymy preservation is measured at the retrieval/response boundary:

• ContextEntityRecall — fraction of expected meronym entities present in retrieved context; the canonical signal.
• Custom LLM-judge — wraps a rubric that asks “does the response correctly identify the parts of X?” and returns score + reason.
• Knowledge-graph coverage — fraction of part-whole edges in the source ontology that are exercised by the eval set; surfaces blind spots.
• Per-domain meronym recall — split the metric by document type or domain to see where chunking is breaking part-whole context.
• Offline-vs-online delta — gap between offline meronym recall and live traffic recall; widening means production data has drifted.

Minimal Python:

from fi.evals import ContextEntityRecall

recall = ContextEntityRecall(
    expected_entities=["blade", "shaft", "casing"]
)
result = recall.evaluate(
    input="What components does the turbine include?",
    output=response,
    context=retrieved_chunks,
)
print(result.score)

Common Mistakes

• Conflating meronymy with hyponymy. A car is a kind of vehicle (hyponym); a wheel is a part of a car (meronym). The retrieval strategies differ.
• Evaluating only on one whole-to-part direction. Test both directions — given the part, recover the whole — to catch retrievers that only learned one side.
• Chunking that splits a part-list across pages. The whole appears on page 3, the parts on page 4; meronymy is broken at the chunk boundary.
• Treating WordNet as ground truth for a domain. Domain ontologies encode meronymy general-purpose lexicons miss; use the domain graph for evals.
• Skipping meronym evals on model upgrades. A new base model can shift part-whole reasoning without changing aggregate accuracy.