Extracting the Legal Description From a Property Deed

#ai #llm #machinelearning #programming

A legal description is not a description of the property. It is the instrument by which the property is identified, and two documents that describe the same parcel in different words are two different legal descriptions. That is why the extraction target is a block of text, not a set of coordinates.

Three systems, one field

Almost every American legal description is written in one of three conventions, and knowing which one you are looking at tells you what the rest of the string has to contain.

Metes and bounds. A perimeter walked as a sequence of calls, each a bearing and a distance from a starting monument: “thence North 12 degrees 34 minutes 20 seconds East, 150.00 feet to an iron pin”. Bearings are almost always quadrant bearings — a north or south reference, an angle under 90 degrees, and an east or west direction — rather than azimuths. The description begins at a point of beginning, which is frequently reached from a separate point of commencement, and it must return to the point of beginning. A description that does not close is a description with a missing or garbled call.
Lot and block. The parcel is a numbered lot on a recorded subdivision plat: a lot, a block, the subdivision name, and — the part that is often lost — the plat’s own recording reference and county. “Lot 7, Block 3” without the plat book and page and the county names nothing; the same lot and block exist in every county in the state.
Public Land Survey System. Aliquot parts of a section: “the Northeast Quarter of the Southwest Quarter of Section 14, Township 3 North, Range 5 West”, followed by the principal meridian, which is the part naive extractions drop. A township is nominally six miles square and holds 36 sections numbered in a serpentine order, and because the survey has to reconcile a rectangular grid with a curved earth, sections along the north and west edges of a township are irregular. Those irregular pieces are described as numbered government lots, not as quarters, which is why a PLSS description sometimes contains a lot number and is still not lot-and-block.

A single deed can carry more than one. A parcel assembled from an old farm and an adjoining platted lot will be described in PLSS for the first tract and lot-and-block for the second, joined by “AND” or numbered “Tract I / Tract II”. A schema with one string field per deed loses the second tract; a schema with an array of tracts does not.

The verbatim block is the deliverable

The temptation is to parse the calls into numbers on the way in: a list of bearings and distances, or a polygon. Resist it as the primary output. The legal description is operative text. If it is later quoted in a title commitment, a corrective deed or a complaint, the string that gets quoted has to be the string that is on the recorded page, punctuation and all. A pipeline that stores a reconstruction — even a faithful one — has stored a paraphrase of a legal instrument.

So the first extracted field is the description as written, preserved with its line breaks, its capitalisation and its abbreviations, and stored alongside the page image and the character offsets it came from. That gives a reviewer something to compare against, which is the only meaningful check available: there is no checksum on a metes and bounds call, no authority to look it up in, and no way to tell a correct bearing from a plausible wrong one without reading the page.

{
  "description_verbatim": "BEGINNING at an iron pin at the Southeast corner ...",
  "description_system": "metes_and_bounds",
  "source_page_range": [3, 4],
  "source_bbox_by_page": [[72, 410, 540, 720], [72, 96, 540, 210]],
  "derived": {
    "calls": [
      { "bearing": "N 12-34-20 E", "distance_ft": 150.0, "monument": "iron pin" }
    ],
    "derivation": "model-parsed, not authoritative"
  }
}

What a parsed description is actually for

The parsed form is still worth producing, but for a different job than the verbatim block. Its value is that it fails. A call list that does not close back to the point of beginning is evidence of a transcription error, and it is one of the very few automatic checks available on this document. Sum the calls as vectors; if the closure error is large relative to the perimeter, something in the string is wrong — a digit dropped from a distance, a bearing whose quadrant letter was misread, a whole call lost at a page break.

Treat that as a flag, not a correction. Real recorded descriptions fail to close for reasons that have nothing to do with your OCR: old surveys, curve calls described by arc length and radius that a naive summation treats as straight lines, and calls that run “along the centerline of the creek” where the boundary is a natural monument with no bearing at all. Surveyors resolve conflicts between calls by a documented order of priority in which natural monuments outrank artificial monuments, which outrank courses and distances, which outrank stated acreage. That ordering is a surveying and legal doctrine, not an extraction rule, and a pipeline that starts applying it has stopped extracting and started opining.

Where transcription breaks

The failures here are unusually concentrated in a few characters, all of which appear in every metes and bounds description and nowhere else in the document.

The degree sign, read as a zero, a lowercase o, a superscript or nothing at all. “N 12 34 20 E” with the symbols gone is unparseable and, worse, sometimes parses wrongly.
Minute and second marks — a prime and a double prime — read as an apostrophe and a quotation mark, then normalised into curly quotes by a downstream text cleaner. That normalisation is invisible in a diff viewer and destroys the field.
Feet expressed with the same prime character, so 150.00' and 150 minutes are one glyph apart in context.
Fraction glyphs in PLSS descriptions: a one-quarter character, the digits 1/4, and the spelled word “Quarter” all occur, and typewritten deeds sometimes use NE1/4SW1/4 with no separators at all.
Descriptions that continue onto a second page or onto an attached “Exhibit A”, which is frequently a different scan quality, a different typeface and occasionally a different orientation. The exhibit, not the body, is usually the operative text where both exist.

Most of these are properties of the scan rather than of the model, so they are addressed upstream: see the OCR pipeline page for the recognition side and PDF parsing for how text layers and image layers disagree on documents like these.

The clause everyone drops

The single most consequential omission in this field is an exception. A description that ends “LESS AND EXCEPT that portion conveyed to the County for road right-of-way by instrument recorded in Book 1102, Page 88” describes a smaller parcel than the same description without those words. So does “SAVE AND EXCEPT”, “EXCEPTING THEREFROM”, and a reservation of mineral rights. Conversely “TOGETHER WITH” adds something — typically an appurtenant easement — and is equally load bearing.

These clauses sit after the closing call, which is exactly where a model that has been told to extract “the legal description” decides it has finished. Give the extraction an explicit end condition instead: the description runs to the end of the descriptive text, including every exception, reservation and appurtenance, and stops at the next operative clause of the deed — the habendum, the warranty covenants, or the “subject to” list. Then add a separate boolean for whether any exception language was found, so that a missing exception on a parcel that has one shows up as a disagreement between two fields rather than as silence. Pair it with a per-field confidence signal and route anything ambiguous to a human who can look at the page.