/attribute_charge

The source code of our COLING'18 paper "Few-Shot Charge Prediction with Discriminative Legal Attributes".

Primary LanguagePython

Few-Shot Charge Prediction with Discriminative Legal Attributes

Source code and datasets of COLING2018 paper: "Few-Shot Charge Prediction with Discriminative Legal Attributes". (pdf)

Dataset

Please download the dataset here, unzip it and you will get three folders: "data", "data_20w", "data_38w". Then put the folder "data" under this directory. It contains following files:

  • words.vec: Pre-trained word embeddings, each line contains a word and its embedding.
  • attributes: The legal attributes for each charge.
  • train: data for training from small dataset.
  • test: data for test from small dataset.
  • valid: data for validation from small dataset.

If you want to train and test on middle dataset, please copy the files in "data_20w" folder to "data" folder. If you want to train and test on large dataset, please copy the files in "data_38w" folder to "data" folder.

Run

Run the following command for training our model:

cd code/
python train.py

Dependencies

  • Tensorflow == 0.12
  • Scipy == 0.18.1
  • Numpy == 1.11.2
  • Python == 2.7

Log

After start training, a new folder "log" will be created.There are 4 directories in it:

  • /evaluation_charge_log/: stores model's performance of charge prediction on test data during training.
  • /evaluation_attr_log/: stores model's performance of attribute prediction on test data during training.
  • /validation_charge_log/: stores model's performance of charge prediction on validation data during training.
  • /validation_attr_log/: stores model's performance of attribute prediction on validation data during training.

Cite

If you use the code, please cite this paper:

Zikun Hu, Xiang Li, Cunchao Tu, Zhiyuan Liu, Maosong Sun. Few-Shot Charge Prediction with Discriminative Legal Attributes. The 27th Iinternational Conference on Computational Liguisitics (COLING 2018).

For more related works, please refer to my homepage.