Please note: This PhD defence will take place in DC 2584 and online.
Ankit Vadehra, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Pascal Poupart, Olga Vechtomova
Automated text generation models have been utilized for a variety of tasks, like conversational systems (chatbots), translating sentences, captioning, summarization systems, creative text generation, etc. In this thesis, I focus on Text Quality Improvement models for the task of Grammatical Error Correction (GEC) and Automatic Speech Recognition (ASR) Transcript Error Correction via Post-Editing (PE). I also investigate and introduce Post-Editing Effort Estimation in Time (PEET) for GEC Tool evaluation.
For ASR Transcript Error Correction, I focus on a Low-Resource Setting, utilizing different semi-supervised synthetic data augmentation approaches. By performing analysis on ASR error correction, I propose a final Post-Editing Model for ASR transcription error correction. Building on the impact of different correction edit types, I propose stronger GEC baseline models optimized using edit-type (substitution, insertion, and deletion) ensemble techniques. Finally, I address the limitations of the traditional GEC evaluation metric by considering PE Temporal Effort dependent on the number of correction edits and their type. I propose a new evaluation metric called the PEET Scorer to estimate PE time for GEC Tool evaluation.
In this talk, I present the challenges of ASR transcript correction and grammatical error correction tasks, results from the models trained to improve text quality, and the work on evaluating GEC tools by temporal post-editing effort.
To attend this PhD defence in person, please go to DC 2584. You can also attend virtually on MS Teams.