Quote context and sources source check: a search guide

Exact-phrase search is the wrong tool for scanned material and almost everyone uses it anyway. A phrase in quotation marks, typed into a historical archive, asks the index for a character-perfect match against text a machine read off browned newsprint. The index often does not have it, the page often does, and the searcher concludes the line was never printed.

A quote context and sources source check is search work, and this page is about doing it so the machinery's failures never become your findings.

What to take away

The index is a transcription of the page, and transcriptions of old print are unreliable.
Short and dull beats long and memorable, every time.
Use the least common word that is also the most likely to be read correctly.
Sorting by date is the whole point. Relevance ranking is your enemy here.

What the index actually contains

Scanned pages are indexed by automated character recognition, and its accuracy depends on the typeface, the condition of the paper, the density of the column, and the quality of the original scan. Nineteenth-century newsprint is close to a worst case on all four.

OCR Error Types

Error type

Similar shapes: Letters swap
Ligatures: Old letterforms
Column breaks: Tight columns
Hyphenation: Line-break hyphen
Punctuation: Invented or deleted

Cause

Similar shapes: Plausible wrong spelling
Ligatures: Read as different letters
Column breaks: Words joined or split
Hyphenation: Neither half indexed
Punctuation: Apostrophes at risk

Search effect

Similar shapes
Ligatures
Column breaks
Hyphenation
Punctuation

The resulting errors are not random. They follow the shapes of letters, which is useful, because it means they are predictable.

Letters with similar shapes swap for each other, so a word acquires a plausible-looking wrong spelling.
Ligatures and old letterforms are read as different letters entirely.
Tight column breaks join or split words.
Hyphenation across a line break splits a word in two, and the index has neither half.
Punctuation is invented and deleted freely, so anything with an apostrophe is at risk.

An exact-match full-text search over that index will find only the instances the machine happened to read correctly. That may be most of them, or very few, and you have no way to tell from a null result.

Choosing what to search

The instinct is to search the memorable part of the line. That is the worst possible choice, twice over: the memorable part is what gets rewritten in transmission, and it is often the part with unusual punctuation.

Picking a Search Fragment

Pick three or four consecutive middle words
Prefer ordinary words in unusual order
Avoid apostrophes, hyphens, old letterforms
Avoid first and last few words
If nothing, cut to two words plus date

Search instead for the dullest distinctive fragment you can find.

Pick three or four consecutive words from the middle of the line.
Prefer ordinary words in an unusual order, over unusual words.
Avoid apostrophes, hyphens, and anything that might have been set in an old letterform.
Avoid the first and last few words, which are the most heavily edited in retelling.
If nothing comes back, cut to two words and add a date restriction instead.

Two- and three-word sequences are well understood for this. An n-gram of that length is short enough to survive transcription errors elsewhere and long enough to be rare.

Wording drifts in transmission, which is why the middle of a line is the stable part. Very short lines are the hard case: when no dull fragment can be picked, work from a probable date and publication and read the pages.

Working around the errors

Problem	Symptom	Workaround
One word misread	Phrase fails, shorter fragments work	Search around the bad word
Hyphenation split	Common word appears absent	Search the halves, or the neighbors
Punctuation invented	Possessives and contractions fail	Rephrase without them
Old letterforms	Whole passages missing	Search modernized and archaic spellings
Column bleed	Nonsense adjacent words	Browse the page image by date instead

Where the archive offers it, approximate matching does this work for you by tolerating a small number of character differences. Many historical archives do not offer it, which is why shortening the string is the universal fallback.

Sorting, not ranking

Relevance ranking is designed to surface what most people want, and what most people want is the famous instance. That is precisely the thing you are trying to see past.

Sort ascending by date, always. Then read the earliest hits rather than counting them. The question is never how many times the line appears; it is what the oldest appearance attaches to it. A line that shows up anonymously in a filler column, decades before anyone famous is named, has just told you most of what you needed to know.

Restricting the date range is the other half of this. Set an upper bound well before the line became well known, and see whether anything survives. That converts a vague search into a bounded one, and the bound is what makes the result reportable.

Reading the results

Look at what the line is doing on the page. A filler paragraph, a sermon extract, and an advertisement are three different kinds of evidence.
Read the surrounding column. The line's neighbors often reveal it as part of a longer passage.
Note whether the earliest instances name anyone. The pattern of the name appearing later than the line is close to decisive.
Where the page image is available, read it rather than trusting the transcription.
Record what you searched, not only what you found.

Coverage matters as much as technique. Knowing which titles and years an archive holds is what turns a null result into a statement. Chronicling America, the Library of Congress's digitized newspaper collection, and HathiTrust, which scans books and periodicals, both state what they hold; ProQuest Historical Newspapers, Gale, and Newspapers.com do the same for their own collections.

What a collection chose to scan also decides which lines are findable at all, so a coverage statement is worth more than a hit count.

Common questions

How many searches before a null result is meaningful?

At least three different fragments, in at least two archives, with the date range restricted. Fewer than that and you are reporting the index's mood rather than the record.

Should I search the whole line ever?

As a first try, since it costs nothing and occasionally works. Never as the basis for a conclusion.

What if the archive has no date sorting?

Then use date filtering in slices: search one decade at a time. It is slower and gives you the same information.

Is browsing page images practical?

For a bounded question, yes. If you have a probable date and a probable publication, reading the actual pages is often faster than fighting the index, and it is the only method that is immune to transcription error.