A Method for Unsupervised Broad-Coverage Lexical Error Detection and Correction

We describe and motivate an unsupervised lexical error detection and correction algorithm and its application in a tool called Lexbar appearing as a query box on the Web browser toolbar or as a search engine interface. Lexbar accepts as user input candidate strings of English to be checked for acceptability and, where errors are detected, offers corrections. We introduce the notion of hybrid n-gram and extract these from BNC as the knowledgebase against which to compare user input. An extended notion of edit distance is used to identify most likely candidates for correcting detected errors. Results are illustrated with four types of errors.