Typo Correction Example

sinlib’s TypoDetector uses a three-stage hybrid pipeline to detect and correct Sinhala spelling errors:

1. Phonological Trie — retrieves phonologically close dictionary candidates
2. BiGRU Seq2Seq — generates correction candidates from a neural encoder-decoder
3. Stupid Backoff LM — reranks candidates using a 151k-article news corpus bigram model

---

Basic sentence correction

from sinlib import TypoDetector

detector = TypoDetector.from_pretrained()

result = detector("ගුරුවරයා අපට උගන්වය්")
print(result)
# 'ගුරුවරයා අපට උගන්වයි'

Diacritic errors

result = detector("සිංහල බාෂාව ලස්සනයි")
print(result)
# 'සිංහල භාෂාව ලස්සනයි'

School-related correction

result = detector("මම පසලට ගියෙමි")
print(result)
# 'මම පාසලට ගියෙමි'

Context-aware suggestions

Passing prev_word and next_word enables bigram reranking to prefer contextually appropriate corrections over phonologically close ones.

# Without context — returns closest phonological match
detector.suggest_correction("ලසන")
# ['වසන', 'ලේඛන', ...]

# With context — promotes the semantically correct candidate
detector.suggest_correction(
    "ලසන",
    prev_word="ලංකාව",
    next_word="රටක්"
)
# ['ලස්සන', ...]

Checking valid words

# A correctly spelled word passes through unchanged
result = detector("මගේ ගෙදර ලස්සනයි")
print(result)
# 'මගේ ගෙදර ලස්සනයි'

Detection without correction

detector.is_word_suspicious("සිංහල")   # False — valid
detector.is_word_suspicious("සින්හල")  # True  — likely typo

Scoring words manually

prob = detector.word_ngram_probability("සිංහල")
print(prob)
# ~3.2e-05  — plausible word

prob = detector.word_ngram_probability("xzqabc")
print(prob)
# ~1e-27  — implausible, would be corrected

Tuning threshold

# Stricter: flag more words as suspicious
strict = TypoDetector(threshold=1e-6)

# Lenient: only flag obvious typos
lenient = TypoDetector(threshold=1e-12)