PatSnap/novelty-search-bench

PatSnap Novelty Search Bench

A Bench for evaluating patent novelty search (prior art search) systems. Each sample contains a query patent publication number (PN) along with ground truth X-type (novelty-destroying) prior art references identified by patent examiners.

The Bench is deliberately designed as a 50% public release of an internal full evaluation set that combines two complementary sample types — cross-jurisdiction family-expanded samples and single-jurisdiction non-expanded samples — so that retrieval systems can be probed on the most informative prior-art scenarios alongside a standard non-expanded baseline.

📄 Companion paper: Research on Evaluation Methods for Patent Novelty Search Systems and Empirical Analysis — Shu Zhang, LiSha Zhang, Kai Duan, XinKai Sun · arXiv 2508.17782 (Aug 2025).

⚠️ Note on the paper. The dataset has been updated since the paper (arXiv, Aug 2025) was published; a revised paper is in preparation.

🐙 GitHub mirror: This dataset is also published as part of the patsnap/patent-bench monorepo, where you can find the reference metric scripts (novelty_metrics.py), evaluation examples, and additional sub-Benches.

Dataset Overview

Property	Value
Total samples (public release)	569
Family-expanded (type=family)	170 — 50% of internal full (340)
Non-expanded (type=public)	399 — ~50% of internal full (≈800)
Source	Real patent examination records
Jurisdictions	CN / US / EP / WO (PCT) / AU
Languages	English (55.0%), Chinese (45.0%)
IPC coverage	All 8 sections (A–H), balanced
Ground truth	Examiner-cited X-type references; for family-expanded samples, also includes X references cited by examiners of the query patent's cross-jurisdiction family members
License	CC BY-NC 4.0

Quick Start

from datasets import load_dataset

ds = load_dataset("PatSnap/novelty-search-bench", split="test")
print(f"Total samples: {len(ds)}")
print(ds[0])

# Build ground truth and evaluate your retrieval system
def build_gt(sample, collapsed=False):
    if collapsed:
        return set(sample["pn_x_family"]) | set(sample["pn_family_x_family"])
    return set(sample["pn_x"]) | set(sample["pn_family_x"])

for sample in ds:
    gt = build_gt(sample, collapsed=False)
    query_text = sample["description"]                 # ← bundled, no external lookup
    # ranked = your_search_system(query_text)
    # hits = sum(1 for r in ranked[:100] if r in gt)
    # ...

How to Use the Query

✅ Query input is bundled in the description field of every sample. Feed it straight into your retrieval system — no external lookup needed.

The description is the full specification text of the query patent, in its original lang. If the system under test has an input character limit, extracting only the Summary of Invention (发明内容) section is usually a workable fallback.

The PNs used here are in PatSnap standardized format and are resolvable by all major public patent databases if you want to cross-reference. Production system: PatSnap Eureka.

Data Fields

Field	Type	Description
id	int64	Sample identifier (1–569)
pn	string	Publication number of the query patent (PatSnap standardized PN)
apdt	int64	Application date (YYYYMMDD)
ipc	list[string]	IPC classification codes
ipc_main	string	Main IPC section (A–H)
lang	string	Language of the query patent (CN/EN)
jurisdiction	string	Filing jurisdiction (CN/WO/US/EP/AU)
type	string	Sample type: family (cross-jurisdiction family-expanded) or public (single-jurisdiction non-expanded)
pn_x	list[string]	X-type prior art reference PNs (directly cited by the examiner)
pn_x_family	list[string]	X-type references expanded to simple patent family member PNs
pn_family_x	list[string]	type=family only. X-type references cited by examiners of the query patent's cross-jurisdiction family members. Empty for type=public.
pn_family_x_family	list[string]	type=family only. Simple patent family expansion of pn_family_x. Empty for type=public.
description	string	Full description text of the query patent in its original lang. Ready to feed straight into a novelty-search system without any external lookup. If the system under test has an input character limit, extracting only the Summary of Invention (发明内容) section is usually a workable fallback.

Ground Truth Construction

Result format	type=family	type=public
Non-collapsed (individual patents)	pn_x ∪ pn_family_x	pn_x
Collapsed by family	pn_x_family ∪ pn_family_x_family	pn_x_family

The same formula pn_x ∪ pn_family_x applies uniformly to both sample types — for type=public, the union with an empty set degenerates to pn_x.

Evaluation Metrics

Metric	Description
top@K	% of samples with ≥1 GT hit in top K (K = 1, 3, 5, 10, 20, 50, 100)
Recall@K	GT hits in top K / total GT refs

The reference metric scripts (with strict / leaderboard mode by default, PN canonicalization, and ranked-list de-duplication) are available in the patsnap/patent-bench GitHub repo.

Distribution

By Sample Type

Type	Count	Percentage
family	170	29.9%
public	399	70.1%

By Jurisdiction

Jurisdiction	Count	Percentage
CN	223	39.2%
WO (PCT)	179	31.5%
US	89	15.6%
EP	74	13.0%
AU	4	0.7%

By Language

Language	Count	Percentage
English (EN)	313	55.0%
Chinese (CN)	256	45.0%

By IPC Section

IPC	Description	Count
G	Physics	75
C	Chemistry; Metallurgy	73
A	Human Necessities	72
B	Operations; Transport	72
H	Electricity	71
F	Mechanical Engineering	70
D	Textiles; Paper	68
E	Fixed Constructions	68

Limitations

• Retrieval-only Bench: Evaluates the search/retrieval step only.
• GT based on examiner citations: Not exhaustive prior art coverage.
• Temporal snapshot: Newer prior art after examination not included.
• Language bias: Only Chinese and English patents.
• Public release is a 50% sample of the internal full evaluation set.

Citation

@article{zhang2025novelty,
  title   = {Research on Evaluation Methods for Patent Novelty Search Systems and Empirical Analysis},
  author  = {Zhang, Shu and Zhang, LiSha and Duan, Kai and Sun, XinKai},
  journal = {arXiv preprint arXiv:2508.17782},
  year    = {2025},
  url     = {https://arxiv.org/abs/2508.17782}
}

@dataset{patsnap_novelty_search_bench_2026,
  title  = {PatSnap Novelty Search Bench},
  author = {PatSnap},
  year   = {2026},
  url    = {https://huggingface.co/datasets/PatSnap/novelty-search-bench},
  note   = {Open dataset accompanying Zhang et al. 2025 (arXiv:2508.17782)}
}

License

Released under CC BY-NC 4.0 — research and non-commercial evaluation purposes only.

Try the Production System

Experience the PatSnap Novelty Search AI Agent — the commercial system referenced in this Bench.

🔗 Try it on PatSnap Eureka