trspell-10

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 TS Corpus.
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.