Grammatical Error Correction in Low Error Density Domains: A New Benchmark and Analyses

Evaluation of grammatical error correction (GEC) systems has primarily\nfocused on essays written by non-native learners of English, which however is\nonly part of the full spectrum of GEC applications. We aim to broaden the\ntarget domain of GEC and release CWEB, a new benchmark for GEC consisting of\nwebsite text generated by English speakers of varying levels of proficiency.\nWebsite data is a common and important domain that contains far fewer\ngrammatical errors than learner essays, which we show presents a challenge to\nstate-of-the-art GEC systems. We demonstrate that a factor behind this is the\ninability of systems to rely on a strong internal language model in low error\ndensity domains. We hope this work shall facilitate the development of\nopen-domain GEC models that generalize to different topics and genres.\n