How many Tamil words have a one-letter twin?

In our sample of Tamil Wikipedia, 1,150,070 of the 14,850,390 word occurrences that contain one of the letters ல, ள, ழ, ர, ற, ந, ண or ன (7.7%) are of a word that has a one-letter twin, where both the word and its twin occur at least 10 times. Counting every word occurrence in the sample, the figure is 6.6%. The percentage depends heavily on the threshold: 16.2% at 3 times, 1.8% at 100.

Read this as a count of vocabulary, not of confusion. At 10 times, 73.8% of the occurrences counted are of a word that is at least 20 times as common as every one of its twins, and in 193 of the 1,330 pairs one word is at least 20 times as common as the other. Often the other word is a variant spelling, a name or a repeated typing mistake. The figure does not say how often a writer might pick the wrong word, and this is a study of Tamil Wikipedia, not a check of your writing: having a twin does not make a word wrong.

What is a twin?

Two written forms are twins when they are the same except for one of these letters at one place, as மலை and மழை are, and both occur often enough in the text. We did not check the forms against a dictionary, so a pair can include a spelling variant, a transliteration variant, a name, a fragment or a typing mistake that happens to be repeated. A doubled letter changes as a whole and counts as one place: பல்லி and பள்ளி differ in two letters written in the same place, and we count that as one change. A form in which only half of a doubled letter changed is not counted as a twin of the doubled form. The word-pair page lists ல, ள and ழ pairs that the Tamil editor chose, with a meaning and a sentence for each.

The counts

The text is a sample of Tamil Wikipedia that we use to build and test our tools: 1,598,428 sentences and 17,457,202 word occurrences (1,213,516 different written forms). 1,099,888 of those forms contain at least one of the eight letters, and their occurrences are 85.1% of all occurrences. A small set of sentences that we keep aside for testing is not included. The snapshot is not published and the date of the Wikipedia dump was not recorded when it was made; its checksum is in the summary file.

At 10 times there are 1,330 pairs. They split by the letters that differ like this:

What this does not say

• It does not say how often writers confuse these words. A twin is a fact about written forms in a text, not about mistakes.
• It does not say which twin a writer meant in a given sentence. The counts are of forms, not of meaning in context.
• The pair list is a list of candidates, not of confusable words. Many pairs are a common word and a rare variant or typing mistake of it. It also counts an inflected form and its base form as separate pairs, and it includes names and fragments.
• Some twins are two accepted spellings of the same word, or two ways to write a borrowed word. They are not wrong forms of each other.
• Wikipedia is one kind of text: edited, written in an encyclopaedic style, and heavy with names and place names. Other kinds of writing, such as news, fiction, letters and speech, may give different figures. Long articles and automatically made articles can also raise a count.
• Counting forms in which only half of a doubled letter changed would raise the figures; most of them are typing mistakes, so we left them out.
• The percentage is not a share of all text, and it cannot be compared with a figure for another corpus or another set of letters; we chose the eight letters.
• A figure is only meaningful with its threshold: the share moves from 16.2% to 1.8% as the threshold goes from 3 to 100, and it is 4.4% at 30 times.

How the counting was done

1. Take as words the unbroken runs of characters from the Tamil block of Unicode, so a token with Latin letters or digits is cut at them, and normalise them to Unicode NFC. Nothing is stemmed or merged, so each inflected form is its own form. Count how often each form occurs.
2. Within one group of letters (ல ள ழ; ர ற; ந ண ன), replace one group letter in a form with a placeholder. Two forms that give the same placeholder form are twins if each occurs at least the threshold number of times. Letters from different groups are never swapped, so ல and ர do not make a pair.
3. Count each pair once. A trio such as வால், வாள் and வாழ் gives three pairs.
4. The share is the number of occurrences of forms that have a twin, divided by the number of occurrences of forms that contain at least one of the eight letters.

No sentence is kept: only single written forms and their counts. One command in the Thiruttham code makes every number on this page and a rerun on the same snapshot gives the same files. The snapshot is not published, so the counts can be checked against the files below but not recomputed from scratch by a reader.

The data

Three files: the 1,330 pairs at 10 times, with the group, both words, how often each occurs and the letter at the place they differ (pairs as CSV); every figure on this page with its threshold (summary as JSON); and a note with the licence and credit line (read me).

The counts are derived from Tamil Wikipedia text, which its contributors released under Creative Commons Attribution-ShareAlike licences. We release the counts files under CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/). If you use them, credit Tamil Wikipedia (ta.wikipedia.org) and Thiruttham, link the licence, say whether you changed the files, and share any changed version under the same licence. Suggested credit: "Word-pair counts derived from Tamil Wikipedia (CC BY-SA 4.0), compiled by Thiruttham, thiruttham.com/research/one-letter-twins". The licence covers the counts files, not the text of this page.

To cite: Thiruttham (2026-11-02). One-letter twins in Tamil: a data study. https://thiruttham.com/research/one-letter-twins. The page and files are versioned by this date.

• Tamil words that differ by ல, ள or ழ
• Tamil spelling rules, one page each
• How it works: rules or AI?