yongminyoo91/patentmatch-temporal-clean-benchmark

PatentMatch Temporal and Component-Clean Extension

Status

Private research preview. Patent text files have not yet been uploaded.

Source

This benchmark is derived from PatentMatch: A Dataset for Matching Patent Claims with Prior Art.

• Paper: https://arxiv.org/abs/2012.13919
• Official project: https://hpi.de/naumann/s/patentmatch
• Source repository: https://github.com/julian-risch/PatentMatch

License

The PatentMatch paper states that the dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Users must cite the original PatentMatch work and preserve attribution.

Label semantics

• 1: X citation — a passage considered novelty-prejudicial or highly relevant to the novelty/inventive-step assessment.
• 0: A citation — a passage representing technological background or the general state of the art.

These labels represent different degrees of examiner-assessed relevance. They should not be described as arbitrary positive and random-negative pairs.

Primary task

Binary claim–prior-art passage classification.

Input:

claim + prior-art passage

Output:

X citation or A citation

Auxiliary task

Pairwise discrimination within groups sharing identical claim text.

This is an auxiliary evaluation, not a full-corpus first-stage retrieval task. Most exact-claim groups contain approximately one X passage and one A passage.

Clean split

Split	Pair rows
Train	16,928
Development	3,652
Test	3,650
Quarantine	1,046

The retained dataset contains 24,230 pairs. Application identities, cited documents, claim texts, and passage texts do not overlap across the train/development/test splits.

The splits are temporally ordered by application filing date.

Temporal ranges

• Train: 2012-07-04 to 2016-09-21
• Development: 2016-09-28 to 2017-05-10
• Test: 2017-05-17 to 2018-01-31

Multi-date exact-text groups

There are 71 exact-claim-text groups containing rows from multiple filing dates.

• They remain valid for the primary row-level classification task.
• They remain in overall exact-text pairwise analysis.
• They are excluded from year-specific pairwise analysis.

Current automatic baselines

The following automatic baselines have been evaluated:

• TF-IDF cosine similarity
• BM25
• zero-shot MiniLM bi-encoder
• zero-shot PatentSBERTa
• train-supervised MS MARCO MiniLM cross-encoder

The current results show that lexical, zero-shot dense, patent-domain dense, and train-supervised cross-encoder baselines remain close to random under the temporal and component-clean split.

Primary metrics:

• Accuracy
• Macro-F1
• ROC-AUC
• Average Precision

Auxiliary metrics:

• Pairwise Accuracy
• MRR
• nDCG

Because most auxiliary groups contain only two candidates, MRR and nDCG should always be interpreted together with their random baselines.

Human annotation extension

A 100-pair double-blind claim-element decomposition pilot package has been generated. Human annotation is not yet complete.

Planned human annotations include:

• claim-element decomposition;
• element-level support labels;
• evidence spans;
• all-elements completeness;
• missing-element identification;
• annotator confidence.

Limitations

1. This is currently a text-pair classification benchmark, not a full-corpus first-stage retrieval benchmark.
2. The auxiliary pairwise task is constructed from identical claim-text groups rather than a guaranteed shared application-level query ID.
3. Family metadata is not currently available.
4. Human claim-element and evidence-span annotations are pending.
5. An A citation should not be interpreted as universally irrelevant in all legal or technical contexts.

Citation

Users should cite the original PatentMatch paper:

Risch, Julian, Nicolas Alder, Christoph Hewel, and Ralf Krestel. "PatentMatch: A Dataset for Matching Patent Claims with Prior Art." 2021.

Official Split Audit

We audited the official PatentMatch ultra-balanced train/test split using the same structural identities employed by the XART construction pipeline. The official split contains no repeated claim IDs across training and test. However, 20 of 1486 structural supercomponents span both partitions, affecting 448 cleaned official test pairs, or 8.86% of the cleaned official test set.

The audit identifies 32 shared cited documents, 17 repeated exact claim texts, and 34 repeated exact passage texts across the official training and test partitions. XART retains 3650 of the 5056 cleaned official test pairs in its final test set, corresponding to 72.19%. All XART test pairs originate from the official PatentMatch test partition.

The audit code is available at scripts/audit_official_patentmatch_split.py. Aggregate results are available under results/audit, and the full report is available at reports/official_split_audit_report.md. Patent claim and passage text are not redistributed.