Correcting Real-Word Spelling Errors: A New Hybrid Approach

TL;DR

This paper introduces a hybrid spelling correction method for real-word errors that combines statistical and syntactic knowledge, improving detection and correction accuracy over previous models.

Contribution

A novel hybrid approach utilizing Constraint Grammar and trigram probabilities to enhance real-word spelling error correction.

Findings

Outperforms previous models like WordNet-based methods

Effective on Wall Street Journal corpus

More practical than existing approaches

Abstract

Spelling correction is one of the main tasks in the field of Natural Language Processing. Contrary to common spelling errors, real-word errors cannot be detected by conventional spelling correction methods. The real-word correction model proposed by Mays, Damerau and Mercer showed a great performance in different evaluations. In this research, however, a new hybrid approach is proposed which relies on statistical and syntactic knowledge to detect and correct real-word errors. In this model, Constraint Grammar (CG) is used to discriminate among sets of correction candidates in the search space. Mays, Damerau and Mercer's trigram approach is manipulated to estimate the probability of syntactically well-formed correction candidates. The approach proposed here is tested on the Wall Street Journal corpus. The model can prove to be more practical than some other models, such as WordNet-based method of Hirst and Budanitsky and fixed windows size method of Wilcox-O'Hearn and Hirst.