Legal Charge Prediction Via Capsule Neural Networks

In recent years,the artificial intelligence technology represented by deep learning and natural language processing has made substantial breakthroughs,and has begun to emerge in the field of legal intelligence.Legal intelligence empowers machines to understand legal texts,analyze cases,and predict defendants’ charges based on the descriptions of cases.Automatic charge prediction plays a critical role in assisting judges and lawyers to improve the efficiency of legal decisions,and thus has received much attention.However,most of existing automatic charge prediction researches focus on high-frequency charge prediction,but few researches propose effective methods for low-frequency charge prediction.Therefore,in this paper we propose a Sequence Enhanced Capsule model,dubbed as SECaps model,to relieve this problem.Specifically,following the work of capsule neural networks,we propose the seq-caps layer,which considers sequence representation and spatial representation of legal texts simultaneously.Then we design a attention mechanism-based residual unit,which provides auxiliary information for charge prediction.In addition,in order to make the model focus more on hard-to-classify samples during training,SECaps model introduces focal loss,which relieves the problem of imbalanced charges.Comparing the state-of-the-art methods,SECaps model obtains 4.5%,2.5% and 6.4% absolutely considerable improvements under F1 in Criminal-S,Criminal-M and Criminal-L respectively.In addition,SECaps model achieves 4.1% improvements than the state-of-the-art baseline on few-shot charge scenarios.The experimental results consistently demonstrate the superiorities and competitiveness of SECaps model on few-shot charge scenarios.In addition,there are some cases which contain multiple charges in real-world,the existing works of charge prediction often ignore the multiple charges because of its too complex to solve if these cases contained.Based on this,in this paper we propose Multiple Charge Prediction via Label Capsule neural network,dubbed as LCaps model,to relieve this problem.Specifically,we first employ Long short-term memory(LSTM)to capture sequence representation of legal case’s descriptions.Then,we propose label capsule to represent each charge,and judge the charges of legal cases according to the probability representation of label capsules.We argue that relevant law articles play an important role in this task,and therefore propose a method to jointly model the charge prediction task and the relevant article extraction task in unified framework.To the best of our knowledge,this is the first work which focuses on the issue of multiple charge prediction.Comparing the state-of-the-art methods,LCaps model obtains 9.4% absolutely considerable improvements under F1 in CAIL2018-Small dataset.The experimental results consistently demonstrate the superiorities and competitiveness of SECaps model on multiple charge scenarios.