This repository includes the evaluation results in "Enhancing Legal Argument Mining with Domain Pre-training and Neural Networks".
Six embedding models are used as basic modules for different combinations: ELMo, GloVe, Legal-BERT-base, Legal-BERT-echr, Legal-BERT-harv, Legal-BERT-harv (custom). The six folders each contains different model combinations with classic neural networks: embedding, embedding+cnn, embedding+bilstm, embedding+resnet.
Each .XLSX file includes four tables: task1, task1-law, task2, task3-c, task3-p, which respectively include evaluation results for the three argument mining tasks (as well as sub-tasks). Specifically, task1 for argument clause recognition (document level), task1-law for argument clause recognition (section level), task2 for argument relation mining,task3-c for conclusion classification sub-task in argument component classification, and task3-p for premise classification sub-task in argument component classification.