Title: The Gap between Deep Learning and Law: Predicting Employment Notice

Abstract: We explore the effect of domain adaptation of transformer-based models by evaluation on the task of Determination of Reasonable Notice through the usage of the Bardal Factors. We categorize free-text summaries by months of reasonable notice as the ground truth, using a window to account for subjectivity. We observe reduced performance of our models after pre-training with full cases. Slightly better performance is observed in comparison, when a large corpus of legal data is introduced, reinforcing conclusions drawn from previous work.

Bio of the main author: Jason Lam is a recent MSc graduate who wrote his thesis on applying deep learning to legal analytics. He continues to research with the Conflict Analytics Lab at Queen's University which aims to provide the general public an access to justice.