Spell Checking and Correction

Todo list:

• Dataset creation and cleaning
  • TDK Güncel Türkçe Sözlük
  • Using data from big corpus (OSCAR)
  • ZemberekNLP data
• Data Augmentation script
• Prepare test set
• Evaluate on existing spell checkers
• Research on creating better .aff/.dic training

Notes:

• No reply for the developers of 10fastfingers.com ?

Turkish Spell Checker Benchmark

We artificially created two benchmarks for Turkish spell checking and correction task.

The following table shows the performance of Turkish spell checkers on official_test.csv dataset.

Spell Checker	Error detection Precision	Error detection Recall	Error detection F1-Score	Correction accuracy
TurkishSpellChecker	94.69	39.09	55.34	60.91
zemberek-nlp	99.06	94.05	96.49	99.24
zemberek-python	91.07	94.42	92.71	91.61
velhasil	96.77	94.19	95.46	93.92
Hunspell-based
hunspell-tr (vdemir)	81.23	97.16	88.48	80.11
hunspell-tr (hrzafer)	92.73	96.51	94.58	79.68
tr-spell	87.42	96.96	91.94	87.37
TDD hunspell-tr (ours)	97.23	94.38	95.78	97.15

The following table shows the performance of Turkish spell checkers on official_test_v2.csv dataset.

Spell Checker	Error detection Precision	Error detection Recall	Error detection F1-Score	Correction accuracy
TurkishSpellChecker	96.54	63.91	76.91	36.09
zemberek-nlp	97.83	87.75	92.52	63.52
zemberek-python	96.62	96.05	96.33	53.84
velhasil	96.74	94.73	95.73	52.99
Hunspell-based
hunspell-tr (vdemir)	90.96	97.47	94.10	48.41
hunspell-tr (hrzafer)	95.38	97.27	96.32	48.93
tr-spell	93.66	97.24	95.42	54.65
TDD hunspell-tr (ours)	97.85	95.25	96.53	56.38