Effective spell checking by learning user behavior

Abstract

This paper describes a spell checking system that learns user behavior. Based on that insight, the system with high likelihood suggests correct replacements for incorrect words and declares unknown, but correct words to be correct. The system relies on three dictionaries, a so-called user history file, and two logic modules to carry out the learning and spell checking. Tests have proved that the system is very fast and highly reliable. Specifically, the top ranked replacement word for an incorrect word was the correct word 96% of the time. Words that were not in the large dictionary but that nevertheless were correct, for example, persons' names, compound words, and control commands, were declared to be correct 82% of the time. It was never observed that an incorrect word was accepted as correct.