Spelling Correction of OCR-Generated Hindi Text Using Word Embedding and Levenshtein Distance

Optical Character Recognition (OCR) systems for Indian languages including Hindi often suffer from poor accuracy due to the wide character variety, compound characters, vowel signs, and other issues. Spelling errors play the key role in the low accuracy. Output level post processing can be an effective technique to detect and correct the spelling errors. In this paper we propose a system for spelling error correction from OCR-generated text in Hindi. The proposed technique is based on neural word embedding and Levenshtein distance. The Continuous Bag-of-Word (CBOW) model is used to learn the word embeddings using large corpora. Dictionary, named entity recognition and context information are used to detect the spelling errors. Once the errors are detected, then the word embeddings are used to find the top most likely words in that context that can replace the erroneous word. These words form the candidate list of words. Then Levenshtein distance is computed between the wrong word and these candidates list of words. The word having the least distance is chosen as the corrected word. The system is tested on ‘The Gita’ where we achieve a reasonable accuracy.