Event Extraction Of Complicated Judicial Cases

In recent years,artificial intelligence has been rapidly applied in the judicial field of our country.Various case handling systems,which use artificial intelligence to assistant related task,have improved the case handling efficiency of judiciary.But from the actual situation,judicial artificial intelligence is still in the exploratory stage.The application of artificial intelligence in the judicial field still faces many challenges.In the judicial field,a case is formed by a number of basic events related to each other.One of the important tasks of judicial handling of cases is to analyze complex cases,identify basic cases,sort out the relationship between events,establish the context of cases,accurately depict the development of the case,and provide support for the understanding,analysis and adjudication of cases.Therefore,judicial case event extraction is the basic task of judicial intelligence.Event extraction is also one of the most basic and concerned tasks in natural language processing.The main task of event extraction is to detect the event from the text and obtain the relevant elements of the event,which are also called event arguments,such as the people involved,the place and time of the event and so on.The task of event extraction oriented to the judicial field has the following characteristics: 1)The judicial system of our country is sound,there are many accusations,and it is difficult to count and obtain the types of events involved.2)Judicial documents are highly narrative,complex judicial cases usually involve dozens or even hundreds of events,and the expressions of events are similar.3)Complex judicial cases involve a large number of case-related figures,who usually participate in multiple events.In addition,compared with the open field,the judicial-oriented event extraction task needs to meet the requirements of rigor and impartiality of the judicial system,and the extraction task must ensure good results.Therefore,in-depth study of the task method of event extraction in complex judicial cases has theoretical and practical significance.This paper focuses on solving the four problems of the event extraction task in the context of complex judicial cases: 1)How to obtain the event types in the judicial field and develop the event template to lay the foundation for the event extraction task.2)How to avoid the interference of similar events and improve the accuracy of multi-event detection in the event detection task of complex cases.3)how to identify the overlapping event arguments in the statement and assign them the correct argument role.4)How to reduce manual labeling and automatically construct large-scale event data sets in the judicial field.In order to solve the above four problems,this paper mainly studies the following four contents:1)A method for obtaining event types in judicial domain based on semantic clustering of trigger words.At present,there are no unified and standardized definitions of event types in judicial field.in order to solve this problem,this paper proposes a method to obtain event types in judicial field based on semantic clustering of trigger words.This paper notes that judicial documents are strongly narrative,so we use the optimized Text Rank-Info algorithm to obtain verbs and nouns that describe events in judicial documents,and construct a set of event triggering words.At the same time,the similarity of the expression context of the same event type is mined,and the clustering algorithm is used for the word vector of the trigger word set,which is expected to obtain the event type in the judicial field.In particular,incremental clustering algorithm is used to deal with the emergence of new event types in the future.A unified and reasonable definition of event types can clarify the task content of judicial event extraction and lay a good foundation for the task of event extraction.2)An Event detection method of complex judicial cases based on argument attention mechanism.The task of event extraction in the context of complex cases has the challenge of similar event types and multiple event instances,so this paper proposes a complex judicial case event detection method based on argument attention mechanism.In order to weaken the interference of similar event types to the event detection task and deal with the problem of sparse data,this paper refers to the criminal charge system in the judicial field,and completes the charge pre-classification of judicial documents based on the XLNet pre-training language model,which reduces the difficulty of the event detection task.On the basis of the above work,an end-to-end event detection model is built,and the key argument information is integrated into the model based on the attention mechanism to deal with the scene with multiple event instances in the sentence.improve the accuracy,recall and F value of event detection tasks,and provide good support for subsequent event argument detection tasks.3)A method of extracting overlapping arguments in complex judicial cases with syntactic information.Due to the strong narrative nature of judicial documents in complex cases,there will be multiple event instances sharing the same word as arguments in sentences.In order to deal with this situation,this paper proposes a method of extracting overlapping arguments in complex judicial cases with syntactic information.Making use of the semantic relationship between trigger words and event arguments,the information of trigger words is integrated into sentence semantic features to construct the sentence feature representation of different event instances.In particular,through the transformation of Transformer,syntactic information is added to the model as important reference information for identifying overlapping event arguments,and the potential syntactic dependency between triggers and event arguments is mined,which effectively solves the problem of sharing overlapping arguments among different event instances,and the result of argument extraction task is significantly improved.4)A construction method of large-scale judicial event data set based on remote supervision.In order to solve the dilemma of lack of large-scale event data set in judicial field,this paper proposes a construction method of large-scale event data set in judicial field based on remote supervision.This method makes use of the strong standardization of the description of specific judicial documents and makes rules based on syntactic analysis to obtain high-quality seed data sets.This paper makes an in-depth study of the structural characteristics of judicial events,defines the concept of key arguments and extends the trigger word set,and realizes the automatic tagging method of event data based on remote supervision.This method greatly reduces the labeling cost of event data,realizes the construction of large-scale event data set in judicial field,and strongly supports the training and learning of event extraction model.In this paper,the proposed method of event extraction from complex judicial cases is verified from two aspects of experiment and prototype system.All kinds of judicial documents disclosed by judicial organs are used as data sources.This paper makes an experimental evaluation on the acquisition method of event type in judicial field,the event detection model of complex judicial event,the event argument extraction model and the construction method of event data set.The experimental results show that the proposed method is better than the comparison method,and the event data set constructed has good quality.At the same time,in order to verify the feasibility of the method proposed in this paper,an event-based judicial case portrait system has been completed,which has been applied to the actual judicial project.