seanghay/khmer-character-confusions

Khmer Character Confusions

Which Khmer characters people substitute for which, measured from live typing.

When a user is unsure of a spelling they swap a single character and search again. Each such swap is one row here: the character replaced, the character tried instead, and how often. No query text appears in this dataset at all — only character pairs and counts.

The result is an empirical confusability matrix for the Khmer script. It recovers the expected structure without being told about it: first/second series consonant pairs (ន/ណ, ត/ទ, ក/គ, ច/ជ, ប/ព), vowel height and length (ិ/ី, ុ/ូ, េ/ើ), liquids (រ/ល, ល/ឡ) and diacritics (៍/័).

Columns

column	description
from_char	the character that was replaced
to_char	the character tried instead
count	observed substitutions
clients	distinct anonymous clients contributing (always >= 5)

Rows are directional: ន -> ណ and ណ -> ន are counted separately, and the asymmetry is informative.

Uses

Weighted edit distance for Khmer spell checking and fuzzy search, OCR and ASR confusion modelling, keyboard layout and input method design, and quantitative description of where Khmer orthography underdetermines spelling from sound.

Derived from anonymous search logs of khmerdict.com, collected December 2024 - August 2026 across roughly 3.2M searches by ~153,000 anonymous clients.

Privacy

This dataset contains no logs and no identifiers. It was produced by an aggregation pipeline that:

• requires every row to be attributable to at least 5 distinct clients, so no row can describe a single person's search;
• discards client identifiers, IP addresses, user agents and timestamps entirely — they are never read into the output;
• restricts all Khmer text to single words matching [\u1780-\u17DD]+, excluding digits, Latin characters, punctuation and whitespace, so no row can carry a phone number, email address, URL or freeform phrase;
• constrains words to dictionary headwords or strings within two edits of one.

At k=5, roughly 9% of distinct queries survive while retaining about 79% of total search volume: the discarded tail is where re-identification risk lives, not where the linguistic value is.

The source logs are not published and will not be.