The Effects of Learner Errors on the Development of a Collocation Detection Tool

摘要

Texts produced by language learners often contain a fair amount of noise, such as misspellings, grammar errors and word-choice errors. These pose a challenge to designing an automated tool to process these texts, partly because most existing tools used in text processing preliminary to linguistic analysis, such as POS-tagging and syntactic parsing, are trained on native-speaker data from which errors have been edited out, and are not designed to deal with atypical errors produced by language learners. In designing and implementing an NLP or a computer-assisted language learning (CALL) tool, determining which "non-pertinent" errors (i.e. errors not specifically targeted by the tool) to deal with and how exactly to deal with them can have a measurable impact on the tool performance. This paper discusses how dealing with some of the "non-pertinent" learner errors in the development of an automated tool to detect miscollocations in learner texts significantly reduces potential tool errors.