Investing Strategies for Clause Recommendation

Source code for the paper, "Investigating Strategies for Clause Recommendation" published at JURIX 2022
You may find the publication here.

Reproducing the results

You'll find the scripts to run inside the ./scripts folder, and code for different components in other folders. Follow this order to prepare the workable subset from the LEDGAR corpus (download the LEDGAR dataset from here.) We'll be making use of the cleaned corpus.

Preparing the subset with contract mapping

bash ./scripts/run_prepare_clauserec_dataset.sh

Further pretrain a BERT or BERT-based model on the dataset

bash ./scripts/run_bert_pretrain.sh

Preparing the embedding-serialized datasets

bash ./scripts/run_create_clauserec_emb_dataset.sh
bash ./scripts/run_create_label_embeddings_dataset.sh

Preparing for training: create id files with train/dev/test splits, train a tokenizer

bash ./scripts/run_create_ids_file.sh
bash ./scripts/run_train_tokenizer.sh

Train for clause recommendation corresponding to a strategy

bash ./scripts/run_train_clause_decoder.sh

Evaluate the best training checkpoint using metrics

bash ./scripts/run_metrics_calculator.sh

Citation

If you're using this, please cite the work as: Joshi, Sagar, et al. "Investigating Strategies for Clause Recommendation." Legal Knowledge and Information Systems. IOS Press, 2022. 73-82.