基于深度学习的多通道多任务学习判决预测模型

Multi-channel Multi-task Learning Judgment Prediction Model Based on Deep Learning

针对目前深度学习在判决领域多聚焦于独立处理单一任务,无法捕捉到法条和罪名预测任务间的关联性,使得罪名预测效果存在准确率瓶颈并缺少法条依据和支持的问题。本文将法条和罪名预测相结合,提出一种基于深度学习的多通道多任务学习判决预测模型。针对词嵌入模型Word2Vec无法解决一词多义及动态优化的问题,引入基于Transformer的双向编码表征(Bidirectional Encoder Representation from Transformers,BERT)词嵌入方法,实现词向量的多任务输送,并通过搭建不同的特征抽取层,提出了基于注意力的双向编码表征法条预测模型(BERT-BiGRU-Attention,BERT-BA)和基于注意力和卷积神经网络的双向编码表征双通道罪名预测模型(BERT-BA-BERT-Convolutional Neural Network,BERT-BABC)。模型从多个视角捕捉多个维度的特征向量,获取更丰富的语义要素,以提升模型的预测效果和泛化能力。实验结果表明,在法条预测中模型准确率达到了87.24%,在罪名预测中模型准确率达到了88.21%,准确率均高于基准模型。

Application of deep learning to the judgment field is mainly focused on processing single task independently and unable to capture the correlation between the article and the crime prediction tasks,which brings the bottleneck of crime prediction accuracy and the issue of lacking the legal basis and support.In this paper,a multi-channel multi-task learning judgment prediction model based on deep learning is proposed by combining the article prediction and the crime prediction.To address the issue that the word embedding model Word2Vec is incapable of solving the problems of polysemy and dynamic optimization,a Bidirectional Encoder Representation from Transformers(BERT)word embedding method based on Transformer is introduced to realize the multi-task transmission of word vectors.By establishing diverse feature extraction layers,an attention-based BERT article prediction model(named BERT-BiGRU-Attention,BERT-BA)and a dual-channel BERT crime prediction model(named BERT-BA-BERT-Convolutional Neural Network,BERT-BABC)are proposed.In the proposed models,feature vectors are captured from multiple perspectives to obtain abundant semantic elements to improve the models’prediction effectiveness and generalization capability.Experimental results show that the proposed models achieve accuracy rate of 87.24%in article prediction and 88.21%in crime prediction,both of which are higher than those of the benchmark model.