Contextual Spelling Correction Using Latent Semantic Analysis

Contextual Spelling Correction Using Latent Semantic Analysis Michael P. Jones and James H. Martin Dept. of Computer Science and Institute of Cognitive Science University of Colorado Boulder, CO 80309-0430 {mj ones, mart in}@cs, colorado, edu Abstract Contextual spelling errors are defined as the use of an incorrect, though valid, word in a particular sentence or context. Tra- ditional spelling checkers flag misspelled words, but they do not typically attempt to identify words that are used incorrectly in a sentence. We explore the use of Latent Se- mantic Analysis for correcting these incor- rectly used words and the results are com- pared to earlier work based on a Bayesian classifier. 1 Introduction Spelling checkers are now available for all major word processing systems. However, these spelling checkers only catch errors that result in misspelled words. If an error results in a different, but incor- rect word, it will go undetected. For example, quite may easily be mistyped as quiet. Another type of er- ror occurs when a writer simply doesn't know which word of a set of homophones 1 (or near homophones) is the proper one for a particular context. For ex- ample, the usage of affect and effect is commonly confused. Though the cause is different for the two types of errors, we can treat them similarly by examining the contexts in which they appear. Consequently, no effort is made to distinguish between the two er- ror types and both are called contextual spelling er- rors. Kukich (1992a; 1992b) reports that 40% to 45% of observed spelling errors are contextual er- rors. Sets of words which are frequently misused or mistyped for one another are identified as confusion sets. Thus, from our earlier examples, { quiet, quite} and { affect, effect} are two separate confusion sets. In this paper, we introduce Latent Semantic Anal- ysis (LSA) as a method for correcting contextual spelling errors for a given collection of confusion sets. 1 Homophones are words that sound the same, but are spelled differently. LSA was originally developed as a model for infor- mation retrieval (Dumais et al., 1988; Deerwester et al., 1990), but it has proven useful in other tasks too. Some examples include an expert Expert lo- cator (Streeter and Lochbaum, 1988) and a confer- ence proceedings indexer (Foltz, 1995) which per- forms better than a simple keyword-based index. Recently, LSA has been proposed as a theory of se- mantic learning (Landauer and Dumais, (In press)). Our motivation in using LSA was to test its effec- tiveness at predicting words based on a given sen- tence and to compare it to a Bayesian classifier. LSA makes predictions by building a high-dimensional, "semantic" space which is used to compare the sim- ilarity of the words from a confusion set to a given context. The experimental results from LSA predic- tion are then compared to both a baseline predic- tor and a hybrid predictor based on trigrams and a Bayesian classifier. 2 Related Work Latent Semantic Analysis has been applied to …

Content truncated to 500 words