Skip to content

Latest commit

 

History

History
41 lines (28 loc) · 2.64 KB

TDD-C-202109-CC-003.md

File metadata and controls

41 lines (28 loc) · 2.64 KB

This dataset consists of 10K unique corrupted words, which simulates a typo with the manually corrected annotations. The dataset is presented as one test split.

All the words in the generated dataset are extracted from the following data sources:

Dataset Curation

To create this dataset, we randomly sample 8500 Turkish words from the TS Corpus Word List (Sezer, 2013, 2017). We develop artificial misspellings by applying random insertions, deletions, and substitutions on 65% of the words, where we use at most two operations on the same word. The remaining 35% of the words are unchanged. Moreover, we added 1K random foreign words, and 500 randomly generated word-like character sequences.

A list of unique words is extracted from each dataset with their frequency information. Intersections of the unique words of each document are computed; then, the words in the presented dataset are chosen randomly with a weighted sampling according to their frequency.

For randomly selected 70% of the words, the following corruptions, with the following probabilities: [1,2,3] p = [0.7, 0.25, 0.05], are applied to simulate possible typos:

  • For a probability of 1/2, the word is asciified.
  • For a probability of 1/6, a random character in the word is substituted by another character sampled from a distribution simulating the placement of keys in standard Turkish-Qwerty keyboards.
  • For a probability of 1/6, a random character is inserted into the word sampled from a distribution simulating the placement of keys in standard Turkish-Qwerty keyboards.
  • For a probability of 1/6, a random character is deleted from a word sampled from a distribution simulating the placement of keys in standard Turkish-Qwerty keyboards.

Dataset format

The dataset is provided in a csv file, where it has two words per line separated by a comma. The first word is the gold (right spelling) word, and the second is the input of the spellchecker to evaluate.

gold input
partililerinden partililerşinden
bulaştıklarına ulaştıklarına
suçlanmadı suçlanmaı
ilgilendirmiyor ilgilendirmiyor
düşüştedir düşüştevdir
sindireceğini sidirecegini
zorlanıyoruz zorlanıyogruz
umduğunuzu umduğunuzu
nitekim nktekim
yapamıyorum yapamiyorum

Dataset Quality

As a quality check of these artificial misspellings, given a list of corrupted words, we ask our annotators to provide us a list of suggestions up to 10 suggestions per word. Their suggestion lists had the gold output 91% of the time.

Contact

Uploaded and documented by Ali Safaya: alisafaya gmail com