Appearance
When volunteer transcriptions come back inconsistent, the fault is nearly always in the guidelines, not the volunteers. The fastest fix is to find the specific rules people interpret differently, replace abstract instructions with cropped image-and-answer examples, and resolve the four perennial trouble spots — abbreviations, dates, unclear text and deletions — with one explicit rule each. This guide diagnoses the common failures and gives the fix for each.
Why do volunteers transcribe the same thing differently?
Run this diagnosis: pull 20 completed tasks and find words transcribed two or more ways. Each cluster of disagreement points to one weak guideline. The root cause is usually an abstract rule ("expand contractions where appropriate") that every reader resolves differently. The fix is to make the rule shown, not told:
text
WEAK: "Transcribe abbreviations sensibly."
STRONG: [crop of "Wm" in the manuscript] → type exactly: Wm
(do not expand to William)A gallery of ten such image-to-text examples eliminates more variance than ten paragraphs of prose.
How do you handle the abbreviation problem?
This is the most common ambiguity. Decide once between diplomatic (keep as written) and expanded (silently complete), and never mix them:
| Approach | Volunteer types | Best when |
|---|---|---|
| Diplomatic | recd | Research needs the original form; reproducible |
| Expanded | received | Output is for general readers only |
For most archives, diplomatic transcription is the safer default because expansion can always be added later, but a wrong expansion is hard to undo. State your choice in the first three lines of the guidelines.
How do you stop date and number errors?
Dates wreck structured datasets faster than anything else. Give one unambiguous rule with an example:
text
RULE: Transcribe the date exactly as it appears. Do not standardise.
"3d March 1841" → type: 3d March 1841
"blank" → leave the field empty, do not guessIf you need ISO dates, do that conversion in post-processing where it can be audited — not in the heads of 40 different volunteers.
How do you mark unclear and missing text?
Provide fixed tokens so the absence of a reading is itself machine-findable:
- Illegible:
[illegible] - Uncertain reading: a trailing query, e.g.
Thornton[?] - Damaged or torn:
[torn] - Blank field: leave empty (state this explicitly)
Without fixed tokens, volunteers improvise — "can't read", "???", "..." — and you cannot reliably count or filter gaps afterwards.
How long should the guidelines be?
One screen. The relationship between length and compliance is inverse: past about a page, each extra paragraph lowers the share of volunteers who read to the end. Structure it as rules first, reference second:
- Top: the 4-6 must-follow rules, each one line.
- Middle: the examples gallery (image crops + correct text).
- Bottom: a link to a fuller reference for rare cases.
Anything a volunteer needs constantly belongs at the top; anything they need once belongs in the linked reference.
How do you treat the forum as a debugging tool?
A question asked twice is a guideline defect. Keep a running list of forum questions; when one recurs, do three things: add a worked example to the guidelines, pin a short answer in the forum, and tell existing volunteers what changed. This closes the loop and stops the same error propagating through hundreds of pages. Track your "repeat question" count as a health metric — it should fall toward zero as the project matures.
Key Takeaways
- Inconsistency is a guideline bug; diagnose it from 20 sample tasks.
- Replace abstract rules with cropped image-to-text examples.
- Choose diplomatic vs expanded transcription once and state it first.
- Mandate exact dates and convert in auditable post-processing.
- Use fixed tokens for illegible, uncertain, damaged and blank text.
- Keep guidelines to one screen and treat repeated forum questions as defects to fix.
Frequently Asked Questions
Why are my volunteers transcribing the same word differently?
Almost always because the guideline is abstract rather than shown. Replace the rule with a cropped image of that exact word next to its correct transcription, and the variation usually disappears.
Should guidelines tell volunteers to expand abbreviations or keep them literal?
Pick one rule and state it once, prominently. For most historical research, transcribe diplomatically (keep abbreviations as written) and let an algorithm or editor expand later; mixing approaches ruins searchability.
How long should transcription guidelines be?
One screen of must-follow rules plus an examples gallery. If your guidelines run several pages, volunteers will not read them and your error rate will rise, not fall.
How do I handle dates and numbers consistently?
Give one explicit format with a worked example, e.g. 'transcribe dates exactly as written; do not convert'. Ambiguity over dates is the single biggest source of unusable structured data.
What do I do when the same question keeps appearing in the forum?
Treat a repeated question as a guideline bug, not a volunteer failure. Add the answer to the guidelines with an example, then post it pinned in the forum so it is findable.