Discuss Your Challenge
Alignment failure in patent translation is domain-specific, claim-specific, and compounding. The diagnostic starts with your corpus.
"A fluent output is not a compliant output. NMT systems optimizing for perplexity minimize semantic drift on average — but patent prosecution operates on the tail of that distribution, where a single hallucinated term or inverted article can trigger a § 112 rejection or narrow enforceable scope. This engagement is designed to locate and correct that tail."
Domains currently covered by the alignment corpus:
Submit a Request
Input required
A claim set, a rejected office action, or a domain description. Raw NMT output with known errors is the highest-value input.
Jurisdiction scope
English → French translation for EPO (EPC Art. 65), FR national phase, and CH/BE/LU validation procedures.
Engagement output
Annotated error taxonomy, corrected Gold Set segments, and a domain-specific alignment protocol document.
Response SLA
Initial response within 48h (business days, CET). Diagnostic call scheduled within the following week.
Confidentiality
All submitted claim sets and corpus materials are treated as attorney-client privileged. NDA available on request.
Engagement Parameters
The following table defines the input/output structure of a standard alignment diagnostic. Use it to assess fit before submitting a request.
| Parameter | Specification |
|---|---|
| Minimum corpus | 1 independent claim + dependent chain (typically 10–20 claims) |
| Source format | Raw NMT output preferred; human draft accepted for comparison baseline |
| Error taxonomy | L1 Hallucination · L2 Terminological drift · L3 Consistency failure · C1–C4 Structural non-compliance |
| Annotation tool | Label Studio (NER) + dependency parse validation |
| Deliverable format | Annotated TMPE Gold Set (.xlsx) + alignment protocol (.pdf) |
| Language pair | EN → FR (EPO / FR / CH / BE / LU jurisdictions) |