Iteratively Questioning and Answering for Interpretable Legal Judgment Prediction

Authors: Haoxi Zhong, Yuzhong Wang, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, Maosong Sun1250-1257

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments on several realworld datasets. Experimental results show that QAjudge can provide interpretable judgments while maintaining comparable performance with other state-of-the-art LJP models.
Researcher Affiliation Collaboration Haoxi Zhong, 1 Yuzhong Wang, 1 Cunchao Tu,1 Tianyang Zhang,2 Zhiyuan Liu, 1 Maosong Sun1 1Department of Computer Science and Technology Institute for Artificial Intelligence, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology, China 2Beijing Powerlaw Intelligent Technology Co., Ltd., China
Pseudocode Yes Algorithm 1 Training the Question Net 1: D {}. 2: while Training do 3: Run QAjudge with input x while the prediction is y. 4: for t 1 to K do 5: D D {(s(t), qt, rt, s(t + 1), d)}. 6: end for 7: Sample a mini-batch from D, and update WQ with loss L(WQ) in Eq. 8 using the mini-batch. 8: end while
Open Source Code Yes The codes can be found from https://github.com/thunlp/QAjudge.
Open Datasets Yes Following the task settings of Zhong et al. (2018), we select three different LJP datasets, namely CJO, PKU, and CAIL for experiments. CJO contains the legal documents published by the Chinese government on China Judgment Online1, while PKU contains the legal documents collected from Peking University Law Online2. Moreover, CAIL is a LJP competition constructed by Xiao et al. (2018) with hundreds of participants. ... 1http://wenshu.court.gov.cn/ 2http://www.pkulaw.com/
Dataset Splits No We randomly select 20% of the data as testing set.
Hardware Specification No No specific hardware details (like GPU models, CPU types, or memory) used for running experiments are mentioned in the paper.
Software Dependencies No The paper mentions software like Bert, Adam, and Bert Adam, but does not provide specific version numbers for these or any other software dependencies.
Experiment Setup Yes The learning rate is 10 5 for Bert and 10 3 for all other models. ... The size of the mini-batch is 4, 096, and hyper-parameters are tuned for different experiments.