MUSER

Source code and dataset for CIKM 2023 paper "MUSER: A Multi-View Similar Case Retrieval Dataset". [paper]

This paper won CIKM2023 Best Resource Paper Honorable Mention.

Dataset Structure

/MUSER/data is the root directory of all MUSER data. The data structure is shown below:

data
- cases
  - cands_by_querys.json  # queries and corresponding candidates IDs
  - train_test.json  # train set and test set for deep nerual model training
- labels
  - golden_labels.json  # queries and relevant candidates IDs
  - sim_scores.json  # relevance scores between queies and candidates
  - top30_dict.json  # top 30 relevent candidates
- legal_items  # legal element label schema
  - aqbq.json  # legal fact
  - ft.json  # law statutory
  - zyjd.json  # dispute focus
- predictions  # retrieval results of baseline models
  - bm25_top100.json
  - labels_top100.json
  - lfm_top100.json
  - lmir_top100.json
  - tfidf_top100.json
- utils
  - stopword.txt

For total case documents cases_pool.json, see this link.

For jieba tokenized corpus corpus.json, see this link.

Legal Element Label Schema

For our hierarchical legal element label schema and their English meanings, see this link.

Keys of the Case Document

case_pool.json contains all cases in MUSER. The key definition of this file is as below:

{
  "uid": "unique id of the case",
  "caseID": "case id assigned by the court",
  "content": {
    "本院查明": "The court's findings of fact",
    "本院认为": "The court's opinion",
    "法条": "cited articles"
  },
  "labels": {
    "案情标签": "legal fact labels annotated at sentences",
    "争议焦点": "disputed focus labels annotated at sentences"
  },
  "split_labels": {
    "aqbq": "legal fact labels splited by level",
    "zyjd": "disputed focus labels splited by level"
  }
}

Experiment

Legal Element Prediction

Run /src/experiments/label_pred/data_preprocess.py to preprocess the sentence-level training data.

The dataset we proposed on this repo has already been preprocessed. You can directly use it to train the model.

For legal fact label prediction, run /src/experiments/label_pred/train.py. For dispute focus prediction, run /src/experiments/label_pred/train_zhengyi.py. There are two training models you can select: BERT and Lawformer.

Similar Case Retrieval

For traditional bag-of-word retrieval models (BM25, TF-IDF, and LMIR) and our proposed legal-element-based model, run /src/experiments/solve_{model_name}.py to get retrieval results at /data/predictions/{model_name}_top100.json.

For deep neural model based on Lawformer, run /src/experiments/train.sh to fine-tune it, run /src/experiments/test.py to get the retrieval result lfm_top100.json.

For evaluation, run /src/experiments/metrics.py, results will be printed on the console.

For our deep neural model train set, see this link; test set see this link.

For our fine-tuned checkpoint, see this link.

Citation

If you use this dataset, please cite our paper.

@inproceedings{li2023muser,
  title={MUSER: A Multi-View Similar Case Retrieval Dataset},
  author={Li, Qingquan and Hu, Yiran and Yao, Feng and Xiao, Chaojun and Liu, Zhiyuan and Sun, Maosong and Shen, Weixing},
  booktitle={Proceedings of the 32nd ACM International Conference on Information and Knowledge Management},
  pages={5336--5340},
  year={2023}
}