Context Influence on Sequence to Sequence Turkish Spelling Correction

In this paper, we implement a Turkish spelling correction system using sequence to sequence (seq2seq) deep neural networks (DNN). In the baseline system, the letters of the misspelled and reference words are fed into the seq2seq model as input and outputs, respectively. In order to improve the performance, we utilize right and left context of the misspelled words. The context words are coded into three letter sequences. The first three consonants of context words are taken when the length of the word is longer than three. The seq2seq models are trained using a large text corpus collected from internet. We randomly add substitution, deletion and insertion spelling errors to the words in the training set. In the experiments, we observe that, the context-dependent model improves the correction performance by 9.2% absolute when compared to the baseline seq2seq model in which no context is utilized.