Skip to content
Switch branches/tags
Go to file

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time
May 20, 2018
May 11, 2018
Jan 12, 2018
May 20, 2018
May 12, 2018
May 12, 2018
May 10, 2018
May 11, 2018
May 12, 2018


This is source code of TopJudge which is used to solve the Legal Judgment Prediction(LJP). The details of this model can be found from


For training, you can run like:

python --config CONFIG_FILE --gpu GPU_ID

The gpu config is optional.

For testing, you can run like:

python --config CONFIG_FILE --gpu GPU_ID --model MODEL_PATH

If you want to use our data setting, you need to promise that there are the frequency count result in the data folder. You can run python to generate the count result. This version of has not been written in OOP form yet.

If your data format is different from this project's setting, you need to modify net/ You need to implement two functions, check and parse. The function check has two parameters: data and config, and it will return whether this data is suitable for the config. The function parse has three parameters: data, config and transformer. The transformer is the Word2Vec object, you can view the function net/ for usage. The parse function should return three things. The first is the documents embedding with size $sentence_num\times sentence_len\times vec_size$. The second is the length of every sentence with total size $sentence_num$. The third is the label whose size should be fixed for your tasks.

Config File

To run our project, a config file is needed. If your config file is config, we will first query the config setting in config, and then if failed, query the config setting in default_config. You can find default_config in the directory config. If you want to learn the details of the config file, you can read the following list:

  • Field net
    • name: The type of model. There are some possible models listing following:
      • CNN: The model using CNN as encoder.
      • LSTM: The model using LSTM as encoder.
      • MultiLSTM: The model using hierarchical LSTM as encoder.
      • ArtFact: The model in Luo, Bingfeng, et al. "Learning to Predict Charges for Criminal Cases with Legal Basis." (2017).
      • ArtFactSeq: Combine the model metioned before with SeqJudge.
      • CNNSeq: Combine CNN encoder with SeqJudge.
      • MultiLSTMSeq: Combine hierarchical LSTM encoder with SeqJudge.
      • Article: Use artilce information as known knowledge to improve the results of SeqJudge.
    • max_gram: The max gram of CNN encoder.
    • min_gram: The min gram of CNN encoder.
    • fc1_feature: The num of features of fc layer.
    • filters: The num of filters of CNN.
    • more_fc: Whether to add a more fc for every tasks or not.
    • hidden_size: The hidden size of LSTM cell.
    • attention: Whether to use attention or not.
    • num_layers: The number of layers in LSTM.
    • method: The method of LSTM to generate embedding. The possible options are:
      • MAX: Using max-pooling on the LSTM output.
      • LAST: Using the last output.
  • Field data
    • data_path: The path of data. Under the data path.
    • dataset: The dataset under the path of data. There should be two files named as crit.txt and law.txt, containing the frequency information.
    • train_data: The list of train data filenames.
    • test_data: The list of test data filenames.
    • type_of_label: The list of tasks.
    • type_of_loss: The list of loss of every task. The posssble loss function are:
      • single_classification: Single class classifcation loss, with cross entropy.
      • multi_classification: Multi classes classifcation loss, with cross entropy.
      • log_regression: $\log^2\left(\left|y_{label}-y_{predict}\right|+1\right)$, unimplemented yet.
    • graph: The DAG structure of SeqJudge.
    • batch_size: The batch size.
    • shuffle: Whether to shuffle the data or not.
    • vec_size: The vector size of every word embedding.
    • sentence_num: The maximum number of sentences in documents.
    • sentence_len: The maximum length of a sentence.
    • min_frequency: The min frequency of a label in the data.
    • word2vec : The word2vec path.
    • top_k: The top-k relevant articles, see the paper we mentioned.
  • Field train
    • epoch: The maximum training epoches.
    • learning_rate: The learning rate.
    • momentum: The momentum of sgd.
    • weight_decay: The weight decay.
    • optimizer: The optimizer, now you can use adam or sgd.
    • dropout: The dropout ratio.
    • train_num_process: The num of processes to read training data.
    • test_num_process: The num of processes to read testing data.
  • Field debug
    • output_time: The num of batches to output debug information.
    • model_path: The basic path to save models.
    • test_path: The basic path to save test results,
    • model_name: The name of this config model. We will store all the things in the directory with name basic path add model name.
    • test_time: The num of epoches to run a test.



  Title                    = {Legal Judgment Prediction via Topological Learning},
  Author                   = {Zhong, Haoxi and Guo, Zhipeng and Tu, Cunchao and Xiao, Chaojun and Liu, Zhiyuan and Sun, Maosong},
  Booktitle                = {Proceedings of EMNLP},
  Year                     = {2018}


No description or website provided.





No releases published


No packages published