Techniques for automatically correcting words in text

Reviewer: Graeme J. Hirst

It is often easy to tell when a poor speller or poor typist has used a spelling checker on a document: each word is correctly spelled, but not all are the words that the author intended. And optical character recognition of documents, with its occasional misrecognitions, has given the world a whole new source of spelling errors. Although spelling checkers (sometimes called “spell checkers” by people who need syntax checkers) have been available for many years now, there is still much room for improvement. In this paper, Kukich presents a careful and exhaustive survey of the techniques—many of them fascinating and ingenious—that have been developed for efficiently finding and correcting errors in spelling; she summarizes each method and its strengths and weaknesses. The problem divides into two parts: detecting an error, which might be a non-word or a wrong real word; and correcting such errors, either in isolation or in context. Non-word detection is the easiest form of the problem, and so the simplest spelling checkers are those that merely draw the user's attention to suspect words. The main techniques used are n -gram probabilities (for example, the trigram fkh has zero chance of occurring in an English word) and lexicons of correctly spelled words (which must be neither too big nor too small). Kukich finds the former better for detecting OCR errors, the latter better for human typing. To correct the possible error, once it is found, a set of candidate corrections must be generated and ranked. These may be presented to the user for the final judgment, or the substitution may be automatic. Kukich reviews a wide variety of methods—including minimum edit distance, similarity keys, the Viterbi algorithm, and neural nets—but finds none wholly satisfactory; in particular, neural nets, which might have been thought to be ideally suited to a problem of this kind, require a prohibitive amount of training. The hardest form of the problem is the detection and correction of erroneous real words, which generally requires some linguistic knowledge (and, in the worst case, a complete understanding of the meaning of the text). For example, a parser can determine when a real-word error causes a syntax error in the sentence; this technique is the basis for many grammar-based writer's aids, such as CRITIQUE [1]. Word bigram or trigram probabilities, derived from large text corpora, can improve other techniques. Because it admits so many different kinds of approaches, spelling checking is a problem that attracts an audience from many different subfields of computing. Despite much effort and many clever ideas, it remains far from solved. Kukich's review will become the definitive reference for work done up to this point; any computer scientist will enjoy reading it.