OCR Glossary for Lawyers: Terms Vendors and Courts Use

The OCR terms a lawyer meets fall into six groups: kinds of recognition (OCR, ICR, OMR), kinds of file (image-only, searchable, native), scanning settings (dpi, bitonal, grayscale), quality measures (accuracy, confidence), layout terms and production terms. This OCR glossary defines each term in one line and links to the post that explains it in full.

Recognition terms: OCR, ICR, OMR, HTR and zonal OCR

These five terms name kinds of reading. OCR reads printed characters, ICR reads hand-printed ones, OMR detects marks such as ticked boxes, HTR reads running handwriting, and zonal OCR reads only chosen areas of a page. A product may do several of them under one name.

Where to read more: the post on PDF text recognition explains what OCR is, what goes in and what comes out. ICR vs OCR covers handprint and mark recognition. What zonal OCR is covers templates, what an OCR engine is covers the layers of a tool, and AI OCR for lawyers covers model-based reading.

The Sedona Conference publishes a reference glossary for discovery practice. Its Glossary of eDiscovery and Digital Information Management defines OCR as "a technology process that captures text from an image for the purpose of creating a parallel text file that can be associated with the image and searched in a database." Its entry for zone OCR describes a feature that "populates data fields by reading certain regions or zones of a document". Those definitions come from litigation support, where the text usually lives in a separate file beside the image.

Easily confused: "handwriting recognition" on a product page may mean ICR, HTR or both. Ask which. The limits of each are in the post on OCR for handwritten notes.

File terms: image-only, searchable, native, text layer, PDF/A

These terms describe what a file holds, which you cannot tell from its name. An image-only PDF holds pictures. A native PDF holds the characters its author typed. A searchable PDF holds pictures with a text layer behind them. PDF/A is a set of archival rules any of the three can be saved under.

Easily confused: "native" has two meanings. In discovery, native format means the file type the creating application uses, such as a spreadsheet produced as a spreadsheet. In everyday PDF talk, a native PDF is one made by software, not by a scanner. A PDF printed from a word processor is a native PDF and, in the discovery sense, a converted file. Ask which meaning a request intends.

What do the scanning terms dpi, bitonal, grayscale and deskew mean?

These are capture terms, the settings and repairs applied when paper becomes a picture. Dpi is resolution. Bitonal and grayscale are color modes. Deskew is a correction for crooked pages. They matter because recognition can only work with the picture the scanner gave it.

The Sedona glossary gives short definitions of three of these. Bitonal: "A bitonal image uses only black and white." Dots per inch is "a measure of the resolution of an image, where more dots in the linear inch indicates a higher resolution." And de-skewing is "the process of straightening skewed (tilted) images", which the glossary calls one of the image enhancements that can improve OCR accuracy.

Where to read more: the best DPI for OCR explains resolution, and grayscale vs black and white scanning explains color mode. The full set of settings for a scanning project is in the guide to digitizing law firm paper files.

Easily confused: dpi describes a scan of paper. A phone photo has pixel dimensions and no true dpi until you know how large the sheet was.

Quality terms: accuracy, error rate, confidence, ground truth

Quality terms describe how close recognized text is to the page. Accuracy and error rate are measured afterwards, against a checked transcript called ground truth. Confidence is the software's own estimate, made while reading, with nothing to compare against. The two kinds of number answer different questions.

Where to read more: how OCR accuracy is measured owns error rates and ground truth, and the OCR confidence score owns confidence. For the hands-on side, see proofreading OCR output against the original.

Easily confused: accuracy and confidence. A page can carry high confidence and still hold a wrong digit, because confidence is a guess the software makes about itself. And an accuracy percentage means little until you know whether it counts characters or words, and on whose documents it was measured. Treat any accuracy figure a vendor quotes as a statement about that vendor's test pages.

Which layout and output terms describe a page's structure?

Layout terms describe how software maps a page before and after it reads the characters: where the blocks are, what kind each is, and in what order they should be read. Output terms describe how that structure is carried into the file you receive.

The difference between a searchable page and a usable document is mostly layout. A two-column brief can have every character right and still paste as nonsense if the columns were read straight across. A table can be recognized perfectly and still be useless if its cells arrive as one long line.

Output format decides how much structure survives the trip. Schema Law exports eight formats, from Word and Excel to JSON and plain text, and the export formats section of the homepage says what each is for.

Easily confused: "zone" in layout analysis is a region the software found by itself. In zonal OCR, a zone is a region a person drew in advance.

Which OCR terms turn up in production requests and court rules?

Four terms do most of the work: text-searchable, extracted text, load file and unitization. Court filing rules use the first. Production specifications and ESI protocols use the other three, along with Bates number and form of production. Each has a practical meaning that differs a little from its everyday sense.

The Sedona glossary defines unitization and load file in terms worth quoting. Unitization is "the assembly of individually scanned pages into documents." Physical unitization relies on staples, paper clips and folders. Logical unitization has a person review the page images, using cues such as consecutive page numbering, report titles and matching headers and footers. A load file, in the same glossary, "indicates where individual pages or files belong together as documents, to include attachments, and where each document begins and ends."

A worked example ties the terms together. Suppose the fictional firm Halvorsen Pike receives a production letter in Adair v. Crestline Freight that says: "Hard-copy documents will be produced as images, logically unitized, with OCR text and a load file. Electronic documents will be produced with extracted text." Translated: the paper was scanned and split into documents by someone who looked at the pages; the words came from recognition, so expect errors; a separate file tells your software how the pieces fit; and the emails and office files carry exact text taken from the originals. The scanned portion is where your search terms are most likely to miss.

Easily confused: "searchable" in a production request does not always mean a searchable PDF. It can mean text files delivered beside the images. Read the specification.

Test a production letter on real pages. Take one exhibit and one pleading from the set. Find a party name inside the exhibit, and find a citation to a statute inside the pleading. Hits that land on the right sheet tell you the request and the file describe the same thing.

On tooling: Schema Law does not produce load files on any plan. Its exports are the eight formats listed for every plan on the pricing page. Check any term in this glossary against the rule, order or protocol in front of you, because the document that binds you may define it differently.