TüReuth Legal at SemEval 2023 Task 6 Code

This repository contains the code to replicate experiments for our participation at SemEval Task 6: LegalAI. Note that we only participate in subtask (A), predicting Rhetorical Roles.

Usage

1. Create Folder data
2. Place 3 json files called train.json, dev.json, and test.json in folder data. These files are provided by the shared task organisers, but have to be renamed accordingly
3. Run python main.py. This will train and save all models and output test predictions in file test_predictions.pickle.
4. Run python make_test_predictions.py. This will create a file called RR_TEST_DATA_FS.json which contains test set predictions in the right format for submission to the shared task.

Note that training MLPs and fine-tuning LMs requires GPU and internet access to download models from huggingface model hub.

Requirements

• torch (with GPU support)
• transformers
• datasets
• nltk
• numpy
• scipy
• pandas
• tqdm