← Back to Portfolio

Structural Compliance — Case Study Library

4 Critical Morpho-Syntactic Constraints in Patent Translation

Generic LLMs optimize for linguistic naturalness. Patent law demands rigid structural formalism. These case studies document the binary constraints that generic models violate — and the alignment protocols that enforce legally compliant output.

Constraint 1

Verb Nominalization

Method Claims with "comprising"

In French patent practice, when a method claim uses "comprising" (comprenant), subsequent steps must be nominalized (turned into nouns), not left as infinitive verbs. Generic NMT models default to infinitive constructions because they are statistically more common in non-patent French, creating legally invalid claim structures that risk rejection under EPC Art. 84.

3 cases Photonics · LiDAR
View Case Studies →
Constraint 2

Article Distinction

Antecedent Basis & Referential Integrity

Patent claims function as closed loops where article choice is binary legal logic. "A/an" introduces new elements; "the/said" references previously defined elements. Generic models treat articles as low-value tokens, frequently deleting definite pointers in complex quantifiers and introducing "New Matter" that renders claims unenforceable under 35 U.S.C. § 112.

3 cases Manufacturing · Medical Devices
View Case Studies →
Constraint 3

Genitive Forms

Class vs. Instance — Prepositional Scope

The choice of preposition (de, du, or Zero-Marker) defines the engineering relationship between components. Generic NMT models harbor a "Definiteness Bias," habitually inserting specific articles to improve fluency. This creates false antecedent dependencies or anglicized calques that fail French technical grammar requirements.

3 cases Telecommunications 5G · Medical · Particle Physics
View Case Studies →
Constraint 4

Syntactic Linearity

Complex Noun Phrases & Modification Scope

English technical nomenclature is Head-Final (Modifiers + Noun), whereas French is Head-Initial (Noun + Modifiers). Generic NMT models suffer from "Linearity Bias," processing tokens sequentially without performing the necessary syntactic inversion, producing lists of unrelated items rather than single compound terms and creating legally indefinite claims.

4 cases Cardiology · Semiconductors · Medical Devices
View Case Studies →