Neural Spelling Correction for Azerbaijani Language

Abstract

Spelling correction is one of the most needed steps in many natural language processing tasks. It is the operation of correcting orthographical mistakes. For languages such as English, spelling correction is usually performed by dictionary lookup for words with minimum edit distance to the given word which is not feasible for agglutinative languages such as Azerbaijani (ISO 639-3:aze). Thanks to their rich morphological structure, languages such as Azerbaijani can have quite complex words with several derivational and inflectional affixes. Indeed, agglutination of the affixes in the new word and meaning formation while adhering to the certain linguistic rules such as vowel harmony is what makes the neural models suitable for the correction of the certain classes of the errors in such languages, helping understand the patterns that the agglutination brings with it. In this paper, the possibility of building such a model for the mentioned purpose is explored and the results are reported.