基于深度学习的司法判决预测算法研究

Legal Judgement Prediction Algorithm Based on Deep Learning

司法判决预测任务指的是根据真实的案情描述文本来预测案件的判决结果,有助于司法专业人士高效的工作,在智能司法方面具有广阔的应用前景。在实践中,易混淆罪名和少样本罪名的判别问题是目前的两大难点,普通模型很容易在上述问题上出现误判。为使易混淆罪名得到更好的区分,结合BERT(bidirectional encoder representations from transformer)和双向门控循环单元(bidirectional gated recurrent unit,BiGRU)深度学习网络,提出BGAAT(BERT BiGRU attribute self-attention)网络模型。该模型引入具有区分性的罪名属性标签,通过注意力机制分别抽取属性特征与上下文特征,使用注意力分布对可解释性进行描述。为使少样本罪名更好地被识别,引入罪名分类权重,使模型在少样本罪名分类取得了更高的F1值。最后在真实的司法判决数据集上进行了实验,证明了所提出模型在该任务上有良好表现。

Legal judgment prediction task refers to predicting the judgment result of a case based on the real case description text,which is helpful for judicial professionals to work efficiently and has broad application prospects in smart justice.In practice,there are two major difficulties at present,which are the confusing charges and few-sample charges,and the common models are prone to make mistakes on the above tasks.In order to better distinguish confusing charges,the BERT BiGRU attribute self-attention(BGAAT)model was proposed by combining bidirectional encoder representations from transformer(BERT)and bidirectional gated recurrent unit(BiGRU)deep learning networks.The model introduced distinguishing crime attribute labels and attention mechanism was used to extracts attribute features and context features,in which attention distribution can describe interpretability.In order to better identify few-sample charges,the weight of charge classification was introduced so that the model can achieve a higher F1value in the classification of few-sample charges.Finally,it was conducted that the proposed model has good performance on this task by conducting experiments on the real legal judgment prediction dataset.