BERT-based Contextual Semantic analysis for English Preposition Error Correction

Abstract

In the writing of English as a Second Language (ESL) learners, preposition related errors account for nearly 30 percent in the all grammatical errors (Bitchener et al. 2005)[1]. Unlike verbs and nouns, isolated preposition does not mean anything, but in different contexts, prepositions usually have different meanings so that ESL students hardly master how to use all prepositions correctly. In this publication, we present a preposition error correction (PEC) model based BERT (Devlin et al., 2018)[2] which contains rich contextual semantic information, fine-tuned BERT models to intensify sentence semantic in output, and calculate the preposition matching scores according to the formula.