TypoDetector

Sinhala spell checker using a character-level n-gram language model combined with edit-distance candidate generation.

Import

from sinlib import TypoDetector

Quick Reference

Method / Property	Returns	Description
TypoDetector.from_pretrained(repo)	TypoDetector	Load from HF Hub
detector(text)	str	Correct a sentence
detector.suggest_correction(word)	list[str]	Closest dictionary matches
detector.word_ngram_probability(word)	float	N-gram likelihood score
detector.get_dictionary()	set[str]	Full word list
detector.get_ngram_probs()	dict	Full n-gram table
detector.dictionary	str	Human-readable summary
detector.ngram_probs	str	Human-readable summary

Usage

Correct a sentence

from sinlib import TypoDetector

detector = TypoDetector.from_pretrained("Ransaka/sinlib")

detector("අපකරියට ගිය")
# 'අපකීර්තියට ගිය'

Get correction suggestions

detector.suggest_correction("අඩිරාජ")
# ['අධිරාජ']

detector.suggest_correction("xyz")
# ['No suggestion']

Score a word

prob = detector.word_ngram_probability("සිංහල")
# 0.000032  (higher = more likely to be a real word)

Inspect the dictionary

print(detector.dictionary)
# Dictionary containing 45231 words.

words = detector.get_dictionary()
"ගෙදර" in words  # True

Behaviour Details

For each word in the input sentence the detector:

1. Checks if the word is in the known dictionary — if yes, passes through unchanged.
2. Estimates the word’s character-level bigram probability.
   • If prob < threshold (default 1e-8): replaces with the top suggest_correction result.
   • If threshold <= prob < 1.0: emits a UserWarning but keeps the word.
3. On any processing error, emits a UserWarning and keeps the original word.