1. Background
case life-cycle information
难点:
- 不同角色的词汇空间可能都不一样,传统NLP算法很难整(consume)这个。
- 当事人陈述与最后判决识别的事实之间的gap。
2. 模型MSJudge
多任务
MSJudge:同时从court debate中识别legal facts和预测每个claim的judgment result
(这里面的抽取的事实部分我是有点好奇,可以用最后判决书里的事实描述文本来做teacher forcing吗?)
可视化components( “debate and fact”, “fact and claim” and “across claims”)之间的互相影响
Multi-Stage Context Encoding:模仿法官理解court debate和pre-trial claims
Debate Utterance Encoder:word embedding + role embedding(随机初始化,联合训练)→Bi-LSTM+attention→utterance embedding
Debate Dialogue encoder:Bi-LSTM,建模得到utterance的全局表征
Pre-trial Claim Encoder:Bi-LSTM+attention(debate和claim共享词嵌入矩阵)
Multi-Stage Content Interaction:建模debates和claims、facts和claims、claims之间的关联,加强claim表征
Debate-to-Claim
Debate-to-Fact
Fact-to-Claim
Fusion
Across-Claim
Fact Recognition and Judgment Prediction
3. 实验
- word embeddings (skip-gram) and role embeddings维度:300
Bi-LSTM隐藏层维度:256
Adam 学习率0.001
batch size 16
dropout rate 0.8
- grid search tuning method and cross-validation
- 把每个claim加上所有debate然后做预测
其他略,待补。