Lexical Error Correction Using Contextual Linguistic Expectations

Abstract

Natural Language Understanding systems should be able to handle incorrect inputs user-friendly by correcting errors autonomously. Lexical error correction methods based solely on the use of morphologic information cannot cope with these requirements. To achieve a more intelligent lexical error correction, it is necessary to consider not only the isolated word but also its intrasentential context. Context information can be used to obtain expectations on the linguistic attributes of an erroneous word. Therefore, it is reasonable to integrate morphologic, syntactic and semantic knowledge into the correction process.

We present an error classification scheme where errors are categorized according to the level of linguistic knowledge, which is needed to correct and understand the input sentence. The presented error correction approach is based on a semantically oriented parsing strategy using both syntactic and semantic expectations on incorrect words. Probabilistically rated syntactic expectations on word categories are computed by applying Bayes’ theorem, thus taking into account sentence fragments already parsed. To integrate semantic knowledge the domain of discourse is represented explicitly.

Examining the intrasentential context of a incorrect word makes it feasible to correct masked mistakes, which are only resolvable by the use of syntactic and semantic information.