Using first and second language models to correct preposition errors in second language authoring

In this paper, we investigate a novel approach to correcting grammatical and lexical errors in texts written by second language authors. Contrary to previous approaches which tend to use unilingual models of the user's second language (L2), this new approach uses a simple roundtrip Machine Translation method which leverages information about both the author's first (L1) and second languages. We compare the repair rate of this roundtrip translation approach to that of an existing approach based on a unilingual L2 model with shallow syntactic pruning, on a series of preposition choice errors. We find no statistically significant difference between the two approaches, but find that a hybrid combination of both does perform significantly better than either one in isolation. Finally, we illustrate how the translation approach has the potential of repairing very complex errors which would be hard to treat without leveraging knowledge of the author's L1.