Research On Matching Algorithms Of Question-answering Pairs In Interrogation Notes

In recent years,the development of artificial intelligence has brought many conveniences to people's lives,and it has also brought new possibilities to the development of all walks of life."Intelligent judiciary" is the focus of major companies competing for research in the near future.It can not only improve the work efficiency of judicial personnels,but also create a new model of judicial operation,better improve the credibility of the judiciary,and safeguard social fairness and justice.The content of this research is " matching algorithm for similar question and answer pairs of interrogation transcripts".We use "interrogation transcripts" as corpus to find similar question and answer pairs in the same case.The interrogation transcript is an important material in the judicial judgment process.It is used to record the confession of suspects and witnesses during the trial of the case.It contains a large number of key information such as testimony,used for case records and referencing judges.Through the machine learning method to quickly find similar transcripts Q & A,you can use the auxiliary sentencing,the confirmation of the evidence chain,to determine the authenticity of the testimony.The task of finding similar question and answer pairs can be regarded as a semantic matching task that calculates the matching degree between two texts.There are often feature-based statistical machine learning methods and neural network-based methods to do semantic textual matching.Feature-based statistical machine learning methods can achieve good results when the extracted features are suitable.However,it is quite timeconsuming to extract a large number of manual features.The neural network has powerful learning ability.Learning through vector representation often does not require too much manual feature design.In this paper,we first used feature-based methods to design a variety of features such as edit distance,keywords,and word-pair vectors.Then we used boosted tree and got an effective results in murder and fraud cases,better than a widely used CNN Matching model.However,this method requires designing a large number of features and is time consuming.Then,we greatly reduced model complexity and processing time through feature selection.We also implemented a variety of neural network models such as m LSTM,MV-LSTM,and interaction-based CNN models,which reduced the need for feature extraction.This paper performed experiments on six kinds of cases and got a general model with comparable results to boosted tree.This paper used a non-stationary approach in data processing to generate data sets by means of pairwise matching.Among the 6 cases we used,there were 5 cases that all reached a F1 value of 80 or more.This paper use the trained model to build a question and answer search system.When a transcript of a case is entered,the system can find the similar question and answer pairs in other transcripts by clicking on the question and answer pair of interest.