Mind the Phrase: Annotation and Identification of Multiword Expressions in Grammatical Error Correction

Multiword Expressions (MWEs) are an important part of language competence which are both commonplace across languages and challenging for language learners to master. Since Grammatical Error Correction (GEC) systems tend to provide suggestions at the token level, they may fail to handle the semantic unity of MWEs. We present annotations of MWEs in learner English essays while identifying their intended spans. This not only allows us to analyse learner MWE errors in more detail for the first time, but also enables us to build GEC systems that provide more meaningful feedback to learners. We have annotated 1,850 essays that were previously tagged for grammatical errors with various types of MWEs. We extend an error annotation toolkit with an MWE error-type and present a method that leverages these annotations to make GEC evaluation more explainable, enabling fine-grained analysis of MWE errors. Experiments with current GEC systems show that an edit-based model fine-tuned on our dataset achieves higher F0.5 scores for the correction of common MWE error types, demonstrating its effectiveness for handling MWEs.