Please note: This PhD seminar will take place online.
Ankit Vadehra, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Pascal Poupart, Olga Vechtomova
The task of Grammar Error Correction (GEC) entails designing a system that is capable of performing text improvement and correcting semantic/syntax inconsistencies in a text span (sentence), while grammar error detection (GED) is used to classify whether a sentence is correct or not.
In this work, I evaluate and explore the utilization of a GED module to assist the task of GEC. The state-of-the-art GEC models are designed using a translation (sequence generation) or edit (classification) based objective. The sequence generation models rely on the use of large amounts of pseudo-data for training while the edit-based models rely on specific transformation labels. I evaluate the performance of GED assisted GEC approach on both edit-based and sequence generation models. To aid our task of comparing performance of different GEC models, I also introduce a GEC specific evaluation metric called “usefulness” that can act as a good estimator of how helpful a GEC model is for the end-user.