patccatClaims: Product vs Process U.S. Patent Claim Classifier

A rule-based classifier that labels U.S. patent claims as product or process and extracts claim-structure flags (means-plus-function, Jepson, Beauregard, Markush, single-line, and more) from udpipe part-of-speech tags. It is the core classifier of the patccat pipeline, packaged for standalone use.

Getting started

Load the POS model once per session with patccat_load_model(); it downloads and caches the pinned udpipe English-EWT model on first use.
Classify a data frame of claim lines (one row per line) with fn.patccat(), or its idiomatic alias patccat_classify().

Key functions

fn.patccat, patccat_classify: Classify claims; return claimType and the claim-structure flags.
patccat_load_model: Download and load the pinned udpipe POS model.
patccat_defaults: Inspect the validated internal parameters and word lists (and use them as a starting point for the override arguments).
fn.benchmarking: Score classifier output against a gold benchmark.

Claim types

claimType is 0 (uncategorized), 1 (process / method), or 2 (product). Product-by-process claims are folded into product (2) and flagged separately in the prodByProcess column.

Reproducibility

On its default (no-override) path the classifier reproduces the committed AMT validation baseline – accuracy 0.9972248 and coverage 0.9897253 on the 9,830-claim multi-line sample. Supplying the override arguments of fn.patccat changes the classification and leaves that baseline behind.

Part-of-speech model

Claim text is tagged with the udpipe English-EWT model, pinned to build ud-2.5-191206 so classifications stay reproducible. The model file is a third-party artifact and is not shipped with the package; it is downloaded once into a per-user cache by patccat_load_model. Credit: the udpipe R package (Jan Wijffels) and the Universal Dependencies English-EWT treebank.

Acknowledgements

The R-package scaffolding, documentation, and repository setup were done with assistance from Anthropic's Claude (Cowork). The classifier and its validation are the authors' own work.

Author

Bernhard Ganglmair (maintainer, b.ganglmair@gmail.com) and W. Keith Robinson.