Semantic Segmentation of Legal Documents via Rhetorical Roles
- Created a new corpus of 100 legal cases, annotated with rhetorical role labels. This is the largest RR corpus.
- We propose new multi-task learning model MTL-BiLSTM-CRF(BERT-SC) which uses label shift as an auxillary task.
- Our model has an F1 score of 0.71 which is better than various other baselines and sequence classification models.
- We also show an application of RR for the judgement prediction task.