Word and Insertion-Gap Error Detection in English Learner Writing: A Cross-Corpus Study of Pretrained and Syntactic Features

Abstract

Grammatical error detection must identify both problematic source words and locations where material is missing. This study evaluates these decisions separately and examines whether explicit dependency features improve classifiers built on frozen pretrained representations. An audited W&I+LOCNESS partition supplies 31,214 training sentences, 2,083 tuning sentences, and 2,286 internal development sentences. External evaluation uses 2,514 FCE test sentences after excluding 181 exact normalized source overlaps with the W&I partitions. Local and contextual lexical controls are compared with a frozen BART encoder followed by linear or multilayer classifiers. Dependency-relation and finite-clause-depth features are tested in parameter-matched ablations across three training seeds. The syntax-free BART multilayer model obtains mean external word and gap detection F1 of 48.84% and 40.29%, respectively; adding both syntactic features yields 48.89% and 40.03%. The syntactic additions do not provide a consistent improvement across both detection task. Uncertainty is estimated by paired resampling of the 97 external writer groups, and parser-defined structure and clean-sentence false alarms are reported separately. The contribution is a reproducible, overlap-controlled comparison of word and pure-insertion-gap detection. These span-derived detection scores do not measure correction quality or establish state-of-the-art performance.