基于机器阅读理解的案件要素识别方法

Method for case element recognition based on machine reading comprehension

传统的案件要素识别方法忽略了案件要素标签的语义信息。由于在裁判文书中案件要素存在严重数据不均衡问题,导致模型难以学习小样本案件要素中的语义特征。针对以上问题,提出基于机器阅读理解的案件要素识别方法。使用案件要素标签信息构建问题,指代所要识别的案件要素类型,将问题与文本拼接输入RoBERTa模型进行编码;通过双向注意力机制从两个方向出发为上下文和问题的交互提供补充信息;经过全连接网络预测案件要素是否存在;在词嵌入部分添加对抗扰动,提高模型的鲁棒性。该模型能够通过语言模型学习文本中的语义特征,有效避免了小样本案件要素中的语义特征稀疏问题。实验结果表明,该模型能有效提升数据集中小样本案件要素的识别性能。

Traditional case element recognition methods ignore the semantic information of the case element labels.Due to the serious data imbalance of case elements in judgment documents,it is difficult for the model to learn the semantic features of small sample case elements.In response to the above problems,a method for case element recognition based on machine reading comprehension was proposed.Case element label information was used to construct the question,referring to the type of case element to be identified,and the question and text were input into the RoBERTa model for encoding.The two-way attention mechanism was used to provide supplementary information for the interaction of context and question from two directions.The fully connected network predicted the existence of the element of the case.Anti-disturbance was added to the word embedding part to improve the robustness of the model.The model can learn the semantic features in the text through the language model,effectively avoiding the problem of sparse semantic features in the small sample case elements.Experimental results show that the model can effectively improve the recognition performance of small sample case elements in the data set.