OCR Patent Documents: File Histories and Prior Art

To OCR patent documents that arrive as image-only PDFs, split the file history into its dated papers, run each part through OCR, and export a searchable PDF for reading plus plain text for searching. Then read the claim amendments, formulas and drawing sheets against the page images, because markup and symbols are where recognized text goes wrong.

A patent matter runs on exact words. One claim term, and what the applicant said about it during prosecution, can carry a whole brief. That makes patent paper a poor place to trust recognized text blindly and a good place to have it, as long as every quotation is read back against the page.

Claim numbers are the index of a file history. Search the converted papers for each patent claim and for remarks made in patent prosecution, including any discussion of nonobviousness. Read a utility patent specification beside those claims before you quote either one in a patent infringement brief. Confirm the claim numbers and the mailing dates on the image.

Which patent records still arrive as image-only PDFs?

The patent records that reach you as pictures of pages are the prosecution file history, older patents, foreign references, non-patent literature such as journal articles and product manuals, and anything a client scanned from a paper folder. Test each file before you plan any work: search it for a word you can see on the first page.

The file history is the largest of these. The USPTO's Manual of Patent Examining Procedure, section 719, states that "the electronic file record in which the U.S. Patent and Trademark Office maintains the application papers is referred to as an image file wrapper". The name tells you what to expect from a downloaded copy, but your own copy is the thing to test, since a file history that passed through a vendor or a prior firm may already carry a text layer of unknown quality.

Granted patents split by age. The USPTO's own search tool draws on separate collections, and its Patent Public Search questions page describes them:

So an older United States patent may be searchable inside the Office's tool while the PDF on your desk is still a set of page pictures. The same page says that Patent Public Search "does not have access to foreign patent databases", so a foreign reference comes from somewhere else, in whatever form that source supplies.

A quick sort of an incoming patent folder:

• Has text already: a patent or published application saved from a full-text source. Leave it alone.
• Image only: the file wrapper, references attached to an information disclosure statement, scanned lab notebooks, old paper from the client's own records. These are the OCR candidates.
• Mixed: a PDF where some pages select and others do not. Treat it as image only and check the result page by page.

The post on why you cannot copy text from a PDF explains the difference between a page picture and a page with text behind it.

What does a prosecution file history contain, page by page?

A prosecution file history holds every paper exchanged between the applicant and the patent office, in the order it was entered: the application as filed, the office's actions, the applicant's responses and amendments, cited references, fee and form pages, and the allowance. The descriptions here follow common usage. Confirm each against the document list of your own file.

The papers do not convert equally well, so it helps to know what kind of page each one is before you read the output.

Section 719 of the USPTO's Manual of Patent Examining Procedure says of the image file wrapper that "the electronic file record is the official record of the application". For your purposes that means the page image is the authority and the recognized text is a finding aid. Cite the page, and quote only what you have read on it.

Split before you upload

Count the pages of your file history, then split it at document boundaries first, one PDF per paper, named with the date and the paper type, for example 2019-03-14 Office Action.pdf. Splitting does three things: it keeps each upload inside the size and page limits of whatever tool you use, it lets you spend your checking time on the papers that matter, and it gives every search hit a document name as well as a page.

Schema Law takes a PDF, PNG or JPEG of up to 20 MB, and the pricing page lists the page limit per document on the Free, Solo and Team plans. It reads each page into text and tables and shows the result page by page, so you can compare a claim listing with its scan before you export anything.

Claim amendments, underlining and strikethrough after OCR

Claim amendments are the pages where plain recognized text can mislead you, because the marks that show what changed are formatting and not characters. A struck-through word may come out as an ordinary word, as garbage, or not at all, and an underline may vanish or clip the letters it sits under.

The rule that produces these marks is 37 CFR 1.121(c)(2). For a claim that is being amended, it says: "The text of any added subject matter must be shown by underlining the added text. The text of any deleted matter must be shown by strike-through except that double brackets placed before and after the deleted characters may be used to show deletion of five or fewer consecutive characters." The same paragraph adds that deleted matter goes inside double brackets "if strike-through cannot be easily perceived".

Three consequences follow for a converted amendment.

• Deleted words may read as live words. If the strikethrough line is lost and the letters survive, the text shows the claim with the deleted language still in it. A search for the deleted term finds the claim, and nothing in the text tells you the term was removed.
• Struck words may read as noise. A line through the middle of each letter changes its shape. Expect substitutions such as an e read as a c or an o, and words that no search will find.
• Double brackets are characters. [[ and ]] are ordinary text, so they have a fair chance of surviving. Search for them. Every hit marks a short deletion.

Use the status identifiers

Paragraph (c) of the same rule requires that "the status of every claim must be indicated after its claim number" with one of a fixed set of identifiers: "(Original), (Currently amended), (Canceled), (Withdrawn), (Previously presented), (New), and (Not entered)". Those words are typed in clean text and make reliable search terms.

A working routine for each amendment paper:

1. Search the converted text for Currently amended. Each hit is a claim with markup. List the claim numbers.
2. Open the scan to each listed claim and read it against the text, word by word.
3. In your own notes, type the claim twice: once as it stood before the paper and once as it stands after it. Do not rely on the exported text for either version.
4. Record the paper's date and page beside both versions.
5. Search for Canceled and New and add those claims to your table.

Paragraph (c)(3) says claims that are not being amended appear "in clean version, i.e., without any markings". Those claims convert like any other typed paragraph, and they give you a useful check: a claim marked Previously presented should match the text you built from the prior amendment. If it does not, either the recognition or your notes are wrong, and the scan settles which.

The Word export is not a substitute for reading the scan here. Whether underlining and strikethrough carry into any converted file is something to test on a page from your own file, and the post on whether OCR keeps formatting explains which kinds of structure can survive.

Older patents and foreign prior art references

Older patents and foreign references are harder inputs than a modern typed response: the type is worn or was copied several times, the layout is in columns, and the language may not be English. Convert them to make them findable, and treat every quotation as something to read from the page.

Older United States patents

Look at what you have before you convert. Where the printed patent sets its specification in two columns with line numbers running down the gutter between them, three things can go wrong:

• Lines from the left and right columns are joined across the gutter, producing sentences that exist nowhere in the patent.
• The line numbers (5, 10, 15 and so on) are read into the sentences beside them.
• Column numbers in the header are read as part of the first line.

If you cite the patent by column and line, the numbers matter, but they belong in your citation and not in the middle of the quoted sentence. The post on whether OCR can handle two-column documents covers reading order and how to test it. For pages set in typewriter faces, see the post on OCR for typewritten documents.

If the Office's USOCR collection already has text for the patent, remember how the USPTO describes it: scanned with optical character recognition, with "limited field-specific text searching". That text was produced by recognition too. It is a second reading to compare with yours and not a certified transcript.

Foreign references

A foreign patent or article raises two separate questions: can the characters be recognized, and can anyone on the team read the result. Recognition of a language depends on the tool and its settings, so run a sample page from your own reference and have someone who reads the language compare it with the scan. Schema Law's site makes no statement about which languages it reads. The post on OCR for foreign-language documents covers scripts, accents and the settings to look for.

Keep the translation question apart from the OCR question. A machine translation of recognized text stacks one source of error on another. Where a reference will be relied on in a filing, check your forum's rule on translations, and work from the translation that rule calls for.

A short record for each prior art reference keeps the set usable:

Do formulas, sequence listings and patent drawings survive OCR?

Mostly no. Formulas, sequence listings and drawings are not running prose, and general OCR is built for running prose. Expect a chemical structure to come out as nothing or as stray letters, a long sequence to pick up wrong characters, and a drawing to yield only its scattered reference numerals. Plan to work from the page image for all three.

Schema Law reads pages into text and tables. Its site makes no statement about formulas, drawings or sequence data, so treat those pages as unconverted until you have looked at them.

Formulas

A mathematical expression depends on position: superscripts, subscripts, fraction bars, symbols stacked over each other. Flattened into a line of text, x² can become x2, and a fraction can become two unrelated numbers on different lines. Chemical structures are drawings with letters at the corners, and recognition returns the letters without the bonds. Wherever a claim or a cited passage contains a formula, copy it by hand from the page and have a second person check it.

Sequence listings

For newer applications, the scanned pages may not be the right source at all. Under 37 CFR 1.831(a), an application filed on or after July 1, 2022 that discloses a sequence covered by the rule "must contain, as a separate part of the disclosure, a computer readable Sequence Listing in XML format". Where such a file exists, get it and use it. For applications filed before that date, the MPEP explains that the earlier sequence rules, 37 CFR 1.821 through 1.825, apply, and the listing may have been filed in a different form. Find out what was filed in your application before you spend time on the page images.

If page images are all you have, do not search recognized sequence text and trust a miss. A run of letters such as ATGGCC gives recognition no dictionary word to lean on, and one wrong letter defeats an exact search.

Drawings

A patent drawing sheet carries a figure label, reference numerals and lead lines. The drawing rule, 37 CFR 1.84(p)(1), requires that "reference characters (numerals are preferred), sheet numbers, and view numbers must be plain and legible", and paragraph (p)(3) sets a minimum height for them. That helps a human reader. It does not make the sheet into text. What you can hope for from OCR on a drawing sheet is modest:

• The figure label (FIG. 3) may be found, which lets you search for the sheet.
• Some numerals will be found, in no useful order, and a numeral crossed by a lead line may be misread.
• Rotated sheets may yield nothing until they are turned upright.

The practical use is a figure index: a list of which page of the file holds which figure. Build it by searching for FIG and then confirming each hit by eye. Engineering sheets in a construction dispute behave in a similar way, and the post on OCR for construction documents covers drawing title blocks, revision notes and specification sections.

Searching a converted file history for a claim term

Search a converted file history for a claim term by running the term and its variants across every paper, logging each hit with the paper's date and page, and then reading each hit on the scan. The log is the work product. A search that returns nothing proves only that the recognized text lacks the string.

Take a fictional example. The claim recites a "resilient latch member", and you want everything the applicant and the examiner said about it.

1. List the variants. resilient latch, latch member, resilient, latch, plus words the specification uses for the same part, such as clip or detent.
2. Add the misreadings. Recognition can turn latch into Iatch or 1atch, and a word broken at a line end into resil- and ient. Search for a fragment such as atch memb to catch them.
3. Search every paper, oldest first. With one PDF per paper, your PDF reader's folder search or a plain text search across the exported .txt files will do it.
4. Log each hit.

1. Read each hit on the page image. Confirm the words, and confirm who wrote them. A response quotes the examiner and an office action quotes the applicant, so the same sentence can appear in both.
2. Check the silent pages. Any paper with no hits that you would expect to discuss the term gets a manual skim. Remarks sections and interview summaries are the first candidates.

Which export to use depends on the job. The searchable PDF keeps the page image in front of you, which suits reading and citing. Plain text suits searching many papers at once and pasting a passage into a draft for comparison. The export formats listed on the homepage show what each one holds.

Before a passage goes into a claim chart or a brief, read it from the scan and cite the paper, its date and its page. The general method for comparing recognized text with its source is in the post on why OCR output is full of errors.

Unpublished applications, invention disclosures and client lab notebooks are confidential in a way a published file history is not. Get the client's position on outside services before any unpublished paper leaves the firm. Schema Law's homepage lists what it commits to on training, access, encryption and deletion. It claims no certification or compliance standard.