A Deep Graphical Model for Spelling Correction

We propose a deep graphical model for the correction of isolated spelling errors in text. The model is a deep autoencoder consisting of a stack of Restricted Boltzmann Machines, and learns to associate errorfree strings represented as bags of character n-grams with bit strings. These bit strings can be used to find nearest neighbor matches of spelling errors with correct words. This is a novel application of a deep learning semantic hashing technique originally proposed for document retrieval. We demonstrate the effectiveness of our approach for two corpora of spelling errors, and propose a scalable correction procedure based on small sublexicons