Proper nouns carry the highest spelling-correction risk

What it is

Correction models pull rare strings toward common words in the lexicon. Proper nouns—people, brands, product codes, drug names, internal project names—are exactly those rare strings, and the “misspelling” is often the identity. Folding “Figma” into “figure,” a rare surname into a common character, a trade name into an ordinary noun, looks more correct to a language model and, in the task, deletes the target. Proper nouns carry the highest correction risk not because they are easier to mistype, but because the costs are asymmetric: a wrong correction points at another entity; a right correction points at spelling.

This leaf does not discuss how an index tolerates character differences. That is a matching-layer issue. Here the only claim is: once the correction module is about to rewrite the query, proper nouns are the class of token that should least often be rewritten automatically.

Why it happens

Correction asks “which more common word does this resemble.” Commonness comes from general corpora and site-wide popular queries; proper nouns live in the long tail. The prior labels a proper noun as “more likely a mistyped frequent word.” People type proper nouns as known items: the string is the key, and changing the key changes the lock. Ordinary words, wrongly corrected, land near a synonym; proper nouns, wrongly corrected, land on another person, product, or file.

Case, hyphens, and internal capitals (iPhone, alphanumerics in ticket IDs) are legal shapes of names; treating them as spelling noise shatters identity. Users are also more confident about proper nouns—that is the name they came to find—so they are slower to suspect a rewrite until they open a wholly unrelated result. The risk is prior stacked on task structure, not the edit operation itself.

Studying it

Sample proper-noun queries on their own; do not fold them into an overall spelling-accuracy number.

• Paradigms: spelling-correction UX studies stratified by token class (ordinary word / person / brand / identifier); known-item tasks using real proper nouns (including uncommon spellings) to see whether auto-correction reroutes; in logs, whether the entity clicked after correction is still the same object.
• Independent variables: case sensitivity, whether a name list or user lexicon can exempt tokens, whether the correction threshold is higher on proper nouns.
• Dependent variables: rewrite rate of proper nouns, whether the entity switches after rewrite, rate of noticing the rewrite.
• Methodological note: a single accuracy from a general spelling test set dilutes the high harm of proper nouns. Gold is “did the entity stay,” not “does it look more like a word.” Labs that use only famous brands make exemption lists look effective; long-tail internal codes are where production fails.

Where it stops holding

When the proper noun itself is misremembered (wrong drug name, wrong colleague), full exemption leads to zero results; offer a tappable nearby query rather than auto-rewriting. Ordinary words are sometimes names (Apple; in Chinese, a brand that is also a grain), so decide from whether the entity exists in the collection, not from POS tags alone. On multilingual sites, a common word in one language is a name in another; a global lexicon will injure one of them. User dictionaries, recently opened objects, and address books are safer exemption sources than general corpora.

Applying it

• For names, brands, identifiers, and case-sensitive tokens, do not auto-rewrite by default; at most offer “did you mean [common word]” on the result page while still running the original.
• Exempt using this collection’s titles, directory, and product names, not only “correct spelling” from a general dictionary.
• If a correction does change a proper noun, the banner should say an object may have changed (“switched to a more common word; the original name is still searchable”), not only “spelling was fixed.”
• Verify with ten real long-tail proper nouns that exist (including uncommon case and hyphens). Rewritten to a common word with a different entity means the risk is unhandled. If exemption covers only famous brands and internal codes still change, the lexicon is too thin.

Related

• Within the group: G3.04.1 Auto-correction must show the original query and allow reverting it · G3.04.3 Zero results should offer nearby queries rather than stop
• Adjacent: G1.08 Terminology consistency · G3.14 Fuzzy matching and approximate search · G3.02 Query formulation
• Search terms: proper noun · spelling correction · out-of-vocabulary