Extracting Inventor and Assignee Information From a Patent Filing

The front page of a granted patent is one of the few internationally standardised documents you will ever extract from. Every field is tagged with a two-digit number in brackets, and the number means the same thing in Tokyo, Munich and Alexandria even when the label beside it does not.

The numbers in brackets

Those numbers are INID codes — Internationally agreed Numbers for the Identification of (bibliographic) Data — defined in WIPO Standard ST.9 and printed by patent offices worldwide. The standard is published by the World Intellectual Property Organization, which maintains the ST-series standards that govern how patent documents are presented and exchanged.

The practical value is that they make a front page machine-readable without natural-language matching. A German document says “Erfinder” and a Japanese one prints its own label, but both print (72). An extraction keyed on the bracketed number works across offices with one schema; an extraction keyed on the word “Inventor” works on one office’s English-language output and silently returns nothing on everything else — which is the failure mode that looks like an empty field rather than an error.

So the instruction to the model is not “find the inventors”. It is “return the content of each bracketed INID field on this front page, keyed by its number, transcribed as printed”, followed by mapping to your own schema in code. That also means new or unexpected codes arrive in your data as unmapped keys rather than being dropped.

{
  "inid": {
    "21": "16/987,654",
    "22": "2020-08-07",
    "45": "2023-05-16",
    "54": "Thermally regulated enclosure for a modular sensor array",
    "71": "Northwind Instruments GmbH, Aachen (DE)",
    "72": [
      "Rivera, Alex, Aachen (DE)",
      "Okafor, Ngozi, Leuven (BE)"
    ],
    "73": "Northwind Instruments GmbH, Aachen (DE)",
    "74": "Whitfield & Associates LLP",
    "30": "10 2019 123 456.7, 2019-08-09 (DE)"
  },
  "office": "US",
  "labels_seen_in": "en"
}

The codes you actually need

The full standard is long; a handful of codes carry almost all of the bibliographic weight, and these are the ones worth mapping explicitly rather than by pattern.

• (21) application number and (22) filing date — the identity of the application, as distinct from the publication number.
• (30) foreign application priority data and (60) related national application data — the two blocks that together form the priority chain discussed in extracting claims and priority dates.
• (51) international patent classification and (52) national or cooperative classification — controlled vocabularies, revised on a schedule, so store the symbol and the scheme version you resolved it against.
• (54) title, (57) abstract, (56) references cited — the last of which is a whole extraction of its own, covered in extracting cited prior art.
• (71) applicant, (72) inventor, (73) assignee, (74) attorney or agent, and on some older United States grants (76), used where the inventor and the applicant are the same person.

Applicant, inventor and assignee are three things

These three fields are routinely collapsed into “who owns it”, and they answer different questions.

The inventor is a natural person, always — inventorship is a factual determination about who conceived the invention, and a company cannot be an inventor. If your extraction produces a company name in field (72), either the document is unusual or the extraction has drifted a field. That is a cheap type-level validation to run.

The applicant is whoever applied. Historically in the United States that had to be the inventor, which is why so many older documents print the inventors as applicants or use the combined (76); since the America Invents Act took effect in 2012 an assignee can be named as the applicant, so a modern US front page may show a company at (71) and individuals at (72). In most other offices a company applicant has always been ordinary.

The assignee is the party to whom rights had been assigned as recorded at the time the document was printed. It is frequently identical to the applicant and occasionally absent altogether, which does not mean the patent is unowned — it means no assignment was recorded before printing.

The printed assignee is a snapshot

This is the single most important caveat on the page, because it is the one that produces confidently wrong answers to the question people actually ask.

The assignee printed at (73) is fixed at publication. Patents are assigned afterwards — sold, transferred in acquisitions, pledged as security, reassigned during corporate reorganisations — and none of that changes the printed page. A twelve-year-old patent whose front page names a company that no longer exists is entirely normal, and the current owner may be four transactions away.

The recorded assignment history is held separately: in the United States by the patent assignment database, searchable at the USPTO’s assignment search, with equivalent registers at other offices. So the correct field name for what you extracted is assignee_at_publication, and any product that presents it as the owner should either join it to the assignment register or say plainly what it is. Recording that distinction in the schema is the difference between a dataset that ages gracefully and one that quietly becomes misinformation.

Names, order and the layout underneath

Each inventor entry carries a name and a residence — typically city and country, sometimes a state — and the entries are a list, not a string. Splitting on commas is wrong immediately: the printed form is frequently “Family name, Given name, City (Country)”, so a comma split produces three fragments per person, and a document with six inventors becomes eighteen fields. Split on the entry structure, not on punctuation.

Name order is not consistent across offices, and family-name-first ordering in the printed field is common. Store the name as printed plus, where the format is unambiguous enough to say so, separate given and family fields with a flag for how they were determined. Do not silently reorder; a reordered name that is also transliterated is effectively unmatchable against any other source.

Physically, the front page is a two-column layout with the bibliographic block on the left, the abstract on the right, and a representative drawing below or beside them. The bibliographic block is dense, small and full of short lines, which is precisely the geometry that a naive reading order interleaves with the abstract. The fix is the same coordinate clustering discussed in two-column reading order, and the INID codes give you an unusually good validation signal afterwards: the codes appear in ascending order down the block, so an extraction that returns them out of order has read across the columns rather than down them.