linguistic-semantics

Lexical and frame semantics for the target language: WordNet/OMW coverage gaps, FrameNet/PropBank-style SRL guidance, MWE/PARSEME handling, and semantic-equivalence eval for cross-lingual RAG and retrieval.

Overview

MWE handling is the dominant low-resource MT failure mode — "kick the bucket" translated literally produces nonsense. WordNet/OMW coverage for many low-resource languages is 5–30K synsets versus English's 117K. PropBank frames are not 1:1 across languages. linguistic-semantics makes these gaps explicit and recommends targeted tools for each.

Pipeline Position

Phase: Analyze (Phase 2)

Before this skill: linguistic-scope (language identity for resource lookup), linguistic-morph / linguistic-syntax (prerequisite structure analysis)

After this skill: linguistic-eval (semantic-equivalence metrics), linguistic-annotate (sense annotation projects)

When It Activates

• Need WordNet/OMW coverage for the target language
• Building/evaluating SRL or frame-semantics annotation
• Diagnosing MWE-related MT failures (idioms mistranslated literally)
• Sense-equivalence eval for retrieval/RAG grounding
• Adding semantic-grounded eval to an LLM-quality pipeline

When NOT to use: Purely surface-level eval (BLEU, chrF) → linguistic-eval. POS/dep parsing → linguistic-syntax. Pure annotation methodology → linguistic-annotate.

What It Does

WordNet / OMW Coverage

Per-language synset count varies 10–20× vs English. Languages like Yoruba have ~5K synsets vs English 117K. RAG grounding queries that depend on synset coverage silently fail in target languages when this gap is not accounted for.

Frame Semantics / SRL

PropBank-style SRL frames are NOT 1:1 across languages. English "GIVE" frame ≠ Spanish "DAR" frame structure exactly. For cross-lingual SRL: alignment via Predicate Matrix or MultiFrameNet is required.

FrameNet availability: Berkeley FrameNet for English; per-language FrameNet projects exist for ~20 languages. For others: Predicate Matrix bridges via English with coverage gaps.

MWE / PARSEME Handling

MWEs (multi-word expressions, idioms, light verbs) must be treated as single semantic units. Pre-tokenize MWEs before tokenization when possible using PARSEME-tagged corpora. Build a per-target MWE catalog if not available — even 500 idioms helps.

Never treat "kick the bucket" as 3 unrelated content words.

Semantic-Equivalence Eval

For RAG/retrieval/cross-lingual semantic similarity:

• COMET-style learned metric > BLEU for semantic equivalence
• LaBSE / SONAR cross-lingual embedding cosine for retrieval-style eval
• COMET-22 coverage varies — check per-language before reporting; a "missing" language gets random numbers

Inputs & Outputs

Example Usage

Language: Swahili (swa), task: cross-lingual MT eval

Semantics Analysis: Swahili
- WordNet/OMW: ~10K synsets (9% coverage vs Princeton WN)
    Gap: technical/abstract vocabulary largely absent
- FrameNet: absent for Swahili; use Predicate Matrix bridge via English
- PropBank-style SRL: stanza morph + custom frame alignment
- MWE strategy: no PARSEME corpus; build 500-idiom pilot catalog
- Semantic-equivalence eval: COMET-22 (check Swahili coverage);
    LaBSE cosine as supplementary
- RAG grounding: flag OMW gap — 91% of English synsets unavailable

Related Skills

• linguistic-scope — language identity for resource lookup
• linguistic-annotate — sense annotation project design
• linguistic-eval — COMET/LaBSE as eval metrics
• linguistic-bitext — MWE handling in parallel corpus