The Mansfield projectA "Did you mean" corrector for every town in England
Find the real names one edit away from a typo, then decide which to suggest, and measure the cost of each choice.
The learner starts with the ONS list of built-up areas in England from the 2021 census: 6,406 names, each with a population. Every name turns out to be unique, because the ONS adds a district in brackets where two places share a name. For any typed text, the program generates every string one edit away: one letter deleted, inserted or changed, or two neighbouring letters swapped. Any of those that is a real place becomes a candidate. "Mansfeild", "mansfiled" and "mansfeld" all lead to Mansfield; "mansfield woodhose" leads to Mansfield Woodhouse; "forrest town" to Forest Town; "market warsup" to Market Warsop.
Usually there is only one candidate, but not always, and that is where the real design choice sits. The program can pick the first candidate alphabetically, which treats every place equally, or pick the candidate with the largest population, on the reasoning that more people are likely to be searching for it. To compare the two fairly, the learner makes 3,000 test typos with one random edit each, in two ways: once choosing the intended place completely at random, and once choosing it in proportion to population, which is closer to what real searches look like.
How often the corrector suggests the intended place, 3,000 one-edit typos each way, our Python run, 29 September 2026| Who is typing | First candidate alphabetically | Largest-population candidate |
|---|
| Intended place chosen at random | 96.7% | 96.9% |
| Intended place chosen in proportion to population | 96.1% | 98.9% |
| Typos with more than one candidate (random / by population) | 3.9% | 5.6% |
The popularity rule looks like a clear win: on realistic searches it lifts accuracy from 96.1% to 98.9%. The first row tells a quieter story. When every place is equally likely to be the one somebody wants, the rule barely helps, 96.9% against 96.7%, because it fixes some typos and breaks others: it helped on 30 and hurt on 25. A typo of Cloughton, a village of 715 people, is corrected to Loughton, population 33,345, every single time. A handful of typos are impossible to fix at all: 5 of the first 3,000 turned out to be the exact name of a different real place.
One local test fails completely. "Kirkby in ashfeild" finds no candidate, because the official name is Kirkby-in-Ashfield: two missing hyphens plus the swapped letters make three edits, beyond the program's reach. Treating hyphens and spaces as the same before comparing fixes it at once. Tidying text before matching, called normalisation, often matters as much as the clever part.
Ages 8 to 11
Spot the misspelt town names on a list and explain which letter went wrong.
Ages 11 to 15
Write a Python function that lists every word one edit away from a typo.
Ages 15 and up
Compare tie-breaking rules on thousands of test typos and report who each one fails.
ONS names, our corrector
The place names and populations come from the Office for National Statistics 2021 census tables for built-up areas. The corrector, the test typos and every percentage above are our own work.
