matthewleechen/300YearsOfBritishPatents

Dataset Card for '300 Years of British Patents'

Dataset Description

300 Years of British Patents is a large-scale dataset of the universe of historical patent specifications published by the British Patent Office between 1617-1899. The dataset contains the titles, full texts of specifications, lists of tokenized words, and extracted named entities for 322,874 specifications.

Dataset Structure

The dataset is structured as a single compressed JSONL. Each line is a patent-level dictionary with the following keys:

Field Name	Type	Description
patent_id	str	Unique identifier for each patent
year	int	Year the patent specification was filed
patent_title	str	Title (or single-line description) of the patent
full_text	list[dict]	OCR text organized into page-level dictionaries
word_tokens	list[str]	Ordered list of processed words
front_page_entities	list[dict]	List of dictionaries containing extracted entities

Dataset Field Details:

• full_text: A list of dictionaries, where each dictionary represents a page and contains:

  • page_num: The number of the page (int)
  • page_text: The OCR-extracted text content for that page (str)

• word_tokens: We use spaCy's en_core_web_lg to tokenize the full text into a list of words, dropping numbers and punctuation

• front_page_entities: A list of dictionaries, where each dictionary represents an extracted entity and contains:

  • class: The entity class (i.e., either PER, ADD, OCC, FIRM, DATE, or COMM) (str)
  • entity_text: The extracted entity text (str)
  • start: The start index of the entity in the original text (int)
  • end: The end index of the entity in the original text (int)
  • person_id: A unique person identifier for a given patent applied to every detected PER entity, and associated ADD, OCC, FIRM entities (list[int])
  • inventor_id: A unique person identifier across all patents applied to every detected PER entity
  • latitude: Latitude (in degrees), for ADD entities only
  • longitude: Longitude (in degrees), for ADD entities only

Example:

An example of a patent-level dictionary:

{
    "patent_id": "GB187901368A",
    "year": 1879,
    "patent_title": "IMPROVEMENTS IN THE MANUFACTURE OF TIN PLATE AND BLACK PLATE, AND IN THE MACHINERY, APPARATUS, AND PROCESSES EMPLOYED IN SUCH MANUFACTURE",
    "full_text": [
      {
        "page_num": 1,
        "page_text": "LETTERS PATENT to Henry Bessemer, of Denmark Hill, in the County of\nSurrey, Civil Engineer, and (truncated...)"
      },
      {
        "page_num": 2,
        "page_text": "the purposes of our Invention, we therefore prefer to employ homogeneous malleable\niron or mild steel made (truncated...)"
      },
      (truncated...)
    ],
    "word_tokens": [
      "letters",
      "patent",
      "to",
      "henry",
      "bessemer",
      "of",
      "denmark",
      "hill",
      "in",
      "the",
      "county",
      "of",
      "surrey",
      "civil",
      "engineer",
      (truncated...)
    ],
    "front_page_entities": [
        {
            "class": "DATE",
            "entity_text": "1st July 1879",
            "start": 341,
            "end": 354
        },
        {
            "class": "DATE",
            "entity_text": "5th April 1879.",
            "start": 370,
            "end": 385
        },
        {
            "class": "PER",
            "entity_text": "Henry Bessemer",
            "start": 18,
            "end": 32,
            "person_id": [1],
            "inventor_id": 18364
        },
        {
            "class": "PER",
            "entity_text": "Alfred George Bessemer",
            "start": 96,
            "end": 118,
            "person_id": [2],
            "inventor_id": 14250
        },
        {
            "class": "ADD",
            "entity_text": "Denmark Hill, in the County of Surrey",
            "start": 37,
            "end": 74,
            "latitude": 51.4748,
            "longitude": -0.0594,
            "person_id": [1]
        },
        {
            "class": "ADD",
            "entity_text": "Oakfield, South Dulwich, in the same County",
            "start": 123,
            "end": 166,
            "latitude": 51.4491,
            "longitude": -0.0667,
            "person_id": [2]
        },
        {
            "class": "OCC",
            "entity_text": "Civil Engineer",
            "start": 76,
            "end": 90,
            "person_id": [1]
        }
    ]
},

Usage

The easiest way to use the dataset is with the HuggingFace `datasets' library.

from datasets import load_dataset

dataset_all_years = load_dataset(
    "matthewleechen/300YearsOfBritishPatents",
    data_files="all_patents.jsonl.gz"
)

We provide a Colab starter notebook here to inspect the data, convert to an inventor-patent CSV, and try out our named entity recognition models.

Language

English

Dataset Authors

Enrico Berkes (Maryland), Matthew Lee Chen (Harvard), Matteo Tranchero (UPenn)

License

CC BY-SA 4.0