Qubut/patent-ate-hupd

patent-ate termhood on HUPD

This dataset is a ranked phrase table: each row is a multiword (or surface) key, a nested-frequency C-value, and how many filings contain that key (df). patent-ate corpus then score built it from the Harvard USPTO Patent Dataset (HUPD). It is not HUPD itself.

Start with the slice (3,393 rows, tens of kilobytes). The full table is 88,607,764 keys from 4,518,254 filings and the Parquet is 1.99 GB.

Source corpus

• Harvard USPTO Patent Dataset (HUPD), Suzgun, Melas-Kyriazi, Sarkar, Kominers, and Shieber (2022).
• Dataset card: huggingface.co/datasets/HUPD/hupd
• Paper: arXiv:2207.04043
• Site: patentdataset.org

HUPD is English-language US utility patent applications. This table was scored from a local tree of those JSON objects (claims, abstract, summary).

How the scores behave

Most nested C-value sits in short strings. Two-word keys hold about half of the positive mass, three-word keys most of the rest; unigrams barely register and phrases of six words or more are a thin tail.

C-value and raw document frequency still agree on legal headings. Side by side, the left ranking opens with claim fragments such as "least a portion"; the right ranking is section titles such as "detailed description".

patent-ate can multiply log(1 + C-value) by inverse document frequency (Sparck Jones, 1972; Lucene BM25-style). That product is zero when df * df exceeds the document count, so a phrase past √N ≈ 2,126 drops out of a single ranked list. Almost every key in the histogram sits at small df; the long tail past the line is those headings, about 8.65 million phrases (9.8%).

lithium ion battery (df 3,018) sits past that line and scores zero; the hyphenated sibling lithium-ion battery (df 1,756) keeps a combined rank of 71.5. The product is what ranks rare technical compounds instead of the headings: "rf network node" and "polymer-anticancer agent conjugate" rise once C-value is no longer used alone.

Named examples from the slice (C-value still loves legal headings; the product does not):

Phrase	C-value	df	product score
another aspect	2,080,434	694,564	0
detailed description	1,061,982	843,582	0
lithium ion battery	22,744	3,018	0
lithium-ion battery	9,028	1,756	71.5
sina molecule	86,557	317	108.7
interior panel	3,864	539	74.6
vehicle interior panel	773	64	74.2

The full parquet stores key, c_value, and df. The product lives in the patent-ate package as product_score_expr.

Slice first

File	Rows	Size	Role
slice/preview.parquet	3,393	~60 KB	Stratified preview
slice/preview.csv	3,393	~246 KB	Same table as CSV
slice/preview.duckdb	3,393	~524 KB	DuckDB with table termhood_slice
slice/preview.meta.json			Stratum counts and file sizes

The slice mixes C-value leaders, high-df legal headings, keys ranked highly by the in-package product, high-df keys with weak nested C, length bands, and a random draw. Extra columns n_words, score, and stratum are computed for the preview only.

Scoring used DuckDB as ephemeral scratch. The published .duckdb is this compact slice, not a spill file.

from datasets import load_dataset

preview = load_dataset('Qubut/patent-ate-hupd', 'slice', split='preview')

import polars as pl

pl.scan_parquet(
    'hf://datasets/Qubut/patent-ate-hupd/slice/preview.parquet'
).head(20).collect()

import duckdb

con = duckdb.connect('preview.duckdb')  # after downloading slice/preview.duckdb
con.sql("SELECT * FROM termhood_slice WHERE stratum = 'top_score' LIMIT 10")

Full table

File	Role
termhood.b936d1ae57054167aa63d8b3d2ced5b1.parquet	Immutable generation fact table
termhood.meta.json	Manifest (schema_version, generation, data_file, total_docs, n_keys)

Field	Value
generation	b936d1ae57054167aa63d8b3d2ced5b1
schema_version	2
total_docs	4,518,254
n_keys	88,607,764

Column	Type	Meaning
key	string	Phrase key (normalized form or surface spelling)
c_value	float64	Nested-frequency C-value (Frantzi, Ananiadou, and Mima 2000; JATE 3.3 formula)
df	int64	Number of filings that contain the key

from datasets import load_dataset

table = load_dataset('Qubut/patent-ate-hupd', split='termhood')

import duckdb

duckdb.sql("""
SELECT key, c_value, df
FROM read_parquet(
  'hf://datasets/Qubut/patent-ate-hupd/termhood.b936d1ae57054167aa63d8b3d2ced5b1.parquet'
)
ORDER BY c_value DESC
LIMIT 10
""")

The matching Python store is patent_ate.TermhoodStore. Software: github.com/Qubut/patent-ate

License

This derived table follows the HUPD license: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0). See the HUPD dataset card.

The patent-ate software is MIT. That license does not replace the HUPD terms for this table.

Citation

@article{suzgun2022hupd,
    title={The Harvard USPTO Patent Dataset: A Large-Scale, Well-Structured, and Multi-Purpose Corpus of Patent Applications},
    author={Suzgun, Mirac and Melas-Kyriazi, Luke and Sarkar, Suproteem K. and Kominers, Scott Duke and Shieber, Stuart M.},
    year={2022},
    publisher={arXiv preprint arXiv:2207.04043},
    url={https://arxiv.org/abs/2207.04043}
}

Also cite Frantzi, Ananiadou, and Mima (2000) for C-value, and patent-ate for the extraction and scoring job.