Extracting Witness and Notary Information From a Notarised Document

The notarised part of a notarised document is not the document. It is a short block of prescribed wording at the end, and that wording states which of two very different notarial acts was performed. Extract the block without extracting its type and you have extracted a name and a date with no idea what they attest to.

The notarial certificate is a separate document

A notary completes a certificate — a self-contained block, sometimes on the same page as the signatures and sometimes on a loose page stapled to the back. It has its own structure, and the structure is the same across US jurisdictions even though the exact wording is prescribed state by state:

State of  Oregon        )
County of Washington    )  ss.

On this 14th day of April, 2026, before me personally appeared
JANE Q. SAMPLE, known to me or satisfactorily proven to be the
person whose name is subscribed to the within instrument, and
acknowledged that she executed the same for the purposes therein
contained.

                            ____________________________
                            Notary Public for Oregon
                            My commission expires: 2028-03-01
                            Commission No. 000000            [SEAL]

Everything above the first line is the instrument being notarised — a deed, a power of attorney, an affidavit. Everything from State of down is the certificate. Model them as two objects with a link, not as a flat set of fields, because they have different dates: an acknowledgement can be taken days after the instrument was signed, so instrument_date and notarisation_date differing is normal and not a discrepancy to flag.

The first two lines are the venue. It is not the notary’s address and it is not the signer’s address — it is where the notarial act took place, and it is the field most often extracted into the wrong slot because it looks like an address block. The ss. is a scribal abbreviation that belongs to the venue and is not data.

Acknowledgement and jurat are not the same

These are the two certificates you will see most, and conflating them is the substantive error on this page.

• An acknowledgement says the signer appeared before the notary and acknowledged that the signature is theirs and that they executed the document willingly. The signer does not have to sign in the notary’s presence. Nothing is sworn. The characteristic wording is “personally appeared … and acknowledged that”.
• A jurat says the signer signed in the notary’s presence and swore or affirmed that the contents are true. The characteristic wording is “Subscribed and sworn to (or affirmed) before me”. An affidavit carries a jurat; a deed normally carries an acknowledgement.

The difference is downstream-consequential: a recorder, a court or a lender may accept one and reject the other for a given instrument. So certificate_type is a required enum in the schema — acknowledgement, jurat, copy_certification, signature_witnessing, other — and its value is derived from the verb phrase, not from the document type. Detecting it is a text-classification job over a short block of near-boilerplate, which is exactly the case where a small set of few-shot examples of each wording outperforms an elaborate instruction.

Watch for the case where both appear. A document with an affidavit attached to a deed can carry a jurat on one page and an acknowledgement on another, with two different notaries and two different dates. The certificate is per-act, so the extracted object is a list.

The fields, and where each one hides

The extractable fields are split awkwardly between typed text and the seal image, and they overlap.

• Notary name. Usually printed below the signature line and also embossed or stamped in the seal. Frequently only in the seal.
• Commission number. In the seal on most state designs, typed on some.
• Commission expiration. Typed as “My commission expires” and often also in the seal. These two can disagree, and when they do the typed one is the one someone wrote by hand.
• Jurisdiction. The state of commission, in the seal; distinct from the venue state, though they normally match.
• Date of the act and signer name, in the certificate body, both typically handwritten into blanks.

The seal is the hard part and it is hard for reasons specific to seals. A rubber stamp is applied by a human, so it is rotated, often overlapping the typed text beneath it, sometimes double-struck, and frequently in an ink that scans faint. An embossed seal has no ink at all — it is a raised impression that a flatbed scanner renders as almost nothing, which is why many states require an inked stamp in addition. Treat the seal as an image region: locate it, crop it, and read it separately at a higher resolution rather than hoping a whole-page pass picks circular text out of a paragraph. Circular text is also where general OCR is weakest, so expect to keep a review sample on seal fields long after the rest of the document is running unattended.

The commission expiry check

The one arithmetic check this document type gives you: a notary’s authority is time-bounded, and the bound is printed on the same page as the act. If notarisation_date is after commission_expiration, something is wrong — a mis-OCRed year, a stamp that was not replaced when the commission was renewed, or an act performed without authority. All three are worth a human looking at, and the check costs one date comparison.

Make it robust before you make it strict. Both dates are frequently handwritten, both are commonly ambiguous between day-first and month-first order, and the expiry is often written as 03/01/28 with a two-digit year. Normalise to ISO with an explicit assumed order recorded alongside, treat any date whose day-and-month components are both twelve or under as ambiguous, and only assert an expiry violation when the gap is unambiguous under both readings. A validator that flags a quarter of your documents because it guessed the date order will be turned off within a week.

The second useful check is cheaper: the venue county should be one the notary is authorised to act in, and the venue state should match the seal’s state. A mismatch is not automatically invalid — rules on acting outside the commissioning county vary by state — but it is a reliable signal that one of the two fields was read from the wrong part of the page.

Witnesses are not the notary

A document can carry both, they sit near each other on the page, and they mean different things. Witnesses attest that they saw the principal sign; they are usually two, they sign and print their names, and on some instruments they must also give addresses. The notary performs a notarial act and is not a witness to the underlying document.

The extraction consequence is that witnesses and notary are separate fields with separate cardinality, and a model given a single signatories array will merge them — producing a contract apparently signed by four people when it was signed by two and witnessed by two. Give each role its own array and require a role label on every extracted person.

Two further roles appear in the certificate wording itself and are easy to mistake for witnesses to the document. A credible witness is someone who swears to the identity of a signer the notary does not personally know and who has no acceptable identification. A subscribing witness signs the instrument on behalf of someone who cannot sign, at their direction. Both appear inside the notarial certificate rather than in the signature block, and both should be extracted with the certificate, not with the signatories. Remote online notarisation adds a third artefact: an audio-video recording reference and an electronic seal, which usually means the “seal” on the page is a vector graphic with selectable text and, for once, trivially readable.