High-Order Sequence Modeling for Language Learner Error Detection

We address the problem of detecting English language learner errors by using a discriminative high-order sequence model. Unlike most work in error-detection, this method is agnostic as to specific error types, thus potentially allowing for higher recall across different error types. The approach integrates features from many sources into the error-detection model, ranging from language model-based features to linguistic analysis features. Evaluation results on a large annotated corpus of learner writing indicate the feasibility of our approach on a realistic, noisy and inherently skewed set of data. High-order models consistently outperform low-order models in our experiments. Error analysis on the output shows that the calculation of precision on the test set represents a lower bound on the real system performance.