Modeling ESL Word Choice Similarities By Representing Word Intensions and Extensions

Automatic error correction systems for English as a Second Language(ESL) speakers often rely on the use of a confusion set to limit the choices of possible correction candidates. Typically, the confusion sets are either manually constructed or extracted from a corpus of manually corrected ESL writings. Both options require the involvement of English teachers. This paper proposes a method to automatically construct confusion sets for commonly used prepositions from non-ESL corpus without manual intervention. The proposed method simulates how ESL learners learn both the intensions and extensions of English words from standard English text. Our experimental results suggest that the automatically constructed confusion sets based on the similarities between the learned words’ intensions is competitive with those directly learned from an ESL corpus containing about 150K preposition usages.