Research And Application Of The Extraction Of Legal Elements In Divorce Dispute Cases

As the basic task of legal artificial intelligent,the extraction of legal elements has attracted much attention in recent years.It not only contributes to the structuring of judicial documents,but also has important significance for the downstream tasks of legal artificial intelligent.Regarding the extraction of legal elements,there is no unified framework that can be applied to all types of legal cases.The current practice is mainly to design different legal elements extraction schemes according to the characteristics of different types of legal cases.Considering the ever-increasing number of divorce dispute cases in our country,which has brought tremendous pressure to court judgment,we select divorce dispute cases as the research object.According to the different characteristics of the legal elements in the judicial documents of divorce dispute cases,we divide legal elements into basic legal elements and key legal elements,and then formulate the corresponding extraction scheme.Further,we apply legal elements extraction to the task of judicial case similarity calculation to verify the effectiveness of legal elements for downstream tasks.Specifically,the main research work of this thesis can be summarized as follows:(1)The data set of legal elements extraction and judicial case similarity calculation are constructed and implemented.Currently,there is no publicly available data set suitable for this thesis,and we need to construct it ourselves.In this thesis,legal elements are subdivided into basic legal elements and key legal elements.These two types of legal elements have different characteristics and cannot adopt a unified legal elements extraction scheme.Therefore,we construct the basic legal elements extraction data set and the key legal elements extraction data set respectively.Besides,in order to verify the effectiveness of legal elements extraction for downstream tasks,we also build a judicial case similarity calculation data set based on divorce dispute cases.(2)The basic legal elements extraction model and the key legal elements extraction model are designed and implemented.For the task of basic legal elements extraction,we extract them through the serial annotation method,and then take Bi LSTM-CRF as the baseline model.Aiming at the problem of the baseline model's inability to recognize polysemous words and lack of access to local features of the text,we propose an improved design that combines the pre-trained language model BERT and attention mechanism.For the key legal elements extraction task,we formalize it as a multi-label text classification task.Most of the existing multi-label text classification models tend to ignore the potential correlation between labels,while some existing models do not consider that different words in the text have different importance to label prediction.In this regard,we propose a key legal elements extraction model that combines the attention mechanism and the sequence generation form.Experimental results show that our models achieve significant improvements over baselines on these tasks.(3)The similarity calculation of judicial cases combined with legal elements extraction.We calculate the similarity of judicial cases based on neural network.In view of the lack of application of legal elements extraction,we design a joint model that integrates our legal elements extraction model.Experimental results show that the application of legal elements extraction can improve the performance of judicial case similarity calculation model.