Level 2: Phrase-Level Accuracy
Multi-Word Technical Expressions
Technical terms often function as indivisible semantic units that lose meaning under word-by-word translation. Generic models break compound expressions into components and translate them independently, destroying the technical relationship that defines the concept. A "gate oxide layer" is a unified semiconductor concept, not three words that can be recombined arbitrarily.
Corpus status:
3 cases
— Active. New cases added as documented from ongoing HITL review.
| ID | Error Type | Case Summary | |
|---|---|---|---|
| L2-001 | Collocation | Collocational Precision & Register — "Performance" in Surgery. The term "performance" is polysemous. In medical patent claims ("performance of the surgery"), generic NMT defaults to exécution, which implies a mechanical or administrative task (robotic register). Standard medical terminology requires réalisation (carrying out). The alignment protocol uses Contextual Word Embeddings to enforce this high-register professional collocation, eliminating the "machine accent" from surgical claim language. | View PDF |
| L2-002 | Metonymy | Metonymy Resolution (Action vs. Object) — "Tamponade". English medical patents often use "Tamponade" metonymically to refer to the device itself. Generic NMT translates the literal cognate as tamponnement (the action) or tamponnade (the medical condition). This creates a "Ghost Object" error where an abstract action is described as having physical parts (e.g., "comprising a wall"), potentially invalidating the claim. The alignment enforces explicit classification, expanding the term to dispositif de tamponnement (tamponade device) to restore the physical apparatus category. | View PDF |
| L2-003 | Polysemy | Domain Polysemy (Anti-Hallucination) — "Tuning" in Diabetes Management. The patent describes a "Tuning module" used to adjust insulin therapy parameters based on patient data. Generic NMT selected the most common dictionary vector for "Tuning," yielding module d'accord — a musical or legal register ("Agreement Module", "Chord Module") wholly alien to a medical control-systems context. The correct domain-standard term is module de réglage (parameter adjustment). The alignment protocol applies a Domain Locking Protocol: it detects the medical/control-systems context via collocation keywords (insulin, glucose, processor, parameters), explicitly bans the music/legal vector space for the token "Tuning," and forces the mapping to réglage or ajustement whenever "Tuning" co-occurs with {module, parameter, PID, loop}. | View PDF |