Design And Implementation Of Guided Retrieval System For Criminal Similar Cases

Since 2010,the Supreme People’s Court has proposed the case-guidance system,as a result,the retrieval mechanism of similar case is playing an important role in the judicial auxiliary decision system.The significance of the retrieval mechanism of similar cases lies in that it can assist judges to refer to similar cases,achieve "the same case and the same judgment",and let judges,parties and prosecutors participate in the trial to the greatest extent.However,the existing retrieval system for similar cases has the following problems: First of all,in terms of user interaction,the retrieval process for similar case is relatively complex,and the needs of different users are not taken into account.For example,some senior judges are not good at using internet technology,so it is difficult for them to skillfully enter the appropriate retrieval elements;In addition,for some parties involved in the case,due to the lack of legal professional knowledge,they cannot accurately understand the professional terms in the retrieval process.Secondly,the user satisfaction of retrieval mechanism for similar cases is not high,which is mainly reflected in the long average response time of retrieval and the low accuracy of retrieval results.To solve the above problems,this thesis takes criminal cases as an example,designed a retrieval system for similar cases.The work completed in this thesis mainly includes the following aspects:(1)In order to take into account the needs of different users,a collection process of retrieval elements for similar cases based on multiple rounds of dialogue is designed.Through the dialogue robot to guide the user to communicate with the robot,and the system automatically extracts the key elements needed for case retrieval from the user’s reply.In the implementation,the task-oriented dialogue module is introduced to realize multiple rounds of dialogue guidance.Task-oriented dialogue module includes two tasks: natural language understanding and dialogue management.The natural language understanding task uses the BILSTM-CRF model for named entity recognition,and extracts the elements required for similar case retrieval from the user’s replies,in addition,an intention classification component based on Bert is designed to classify the user’s replies.System will display the extracted retrieval elements to the user at the end of the dialogue.(2)In order to improve the user satisfaction of retrieval mechanism for similar cases,an improved vertical search scheme is designed.Firstly,in order to improve the accuracy of retrieval results,this thesis improved the existing retrieval function for similar cases.By establishing the adjustment mechanism of element weight,users can adjust the order of retrieval element list according to their attention to different elements,system quantifies the weight of each element according to the list order and dynamically adjusts the relevance score of the document according to the weight of the retrieval element.In addition,search association and aggregation analysis functions are added to assist users to complete retrieval.At the same time,the document scoring mechanism in the retrieval process under the distributed environment is optimized.Secondly,in order to reduce the average response time of the system,a distributed storage and retrieval mechanism is introduced.Eliminate the restriction of data transmission through multiple data nodes,and add data replica to improve data access ability.The system tests show that the accuracy of the optimized task-oriented dialogue module in the field of named entity identification and intent classification tasks has been significantly improved,and the dialogue management based on slot filling strategy makes the guidance of multiple rounds of dialogue more reasonable.For the vertical search module,the improvement of the retrieval mechanism for similar cases and the addition of the auxiliary retrieval function enhance the accuracy of the retrieval results.The response time is reduced by introducing the distributed vertical search mechanism.The system achieves the desired effect basically.