Research On Police Question Answering Technology Based On Machine Reading Comprehension

In the process of law enforcement and case handling,the public security police will form a large number of unstructured electronic data represented by the files.The police need to understand the details of the person,place,object and so on in the massive files to assist law enforcement and case handling.With the advancement of the construction of the rule of law in China,the requirements of relevant laws,regulations and policy documents for law enforcement and social governance are increasingly improved.The police should fully understand and master the corresponding laws and regulations in the handling of specific matters,and carry out their work in strict accordance with the prescribed procedures and requirements.The police question and answer system can meet the needs of police officers to obtain relevant information in their work through the interactive way of "question-answer".Machine reading comprehension,as one of the technologies of intelligent question answering,is characterized by its ability to find answers from unstructured text paragraphs according to questions.In the application of police question and answer,it is found that the current machine reading comprehension technology is unable to model the lengthy text efficiently and misidentify the sentence components.Based on this,this thesis carries out research,and the specific work is as follows:Firstly,aiming at the problem that the length of text sequence in public security business is far beyond that which can be processed by most of the existing pre-training language models,an efficient long text machine reading comprehension model FastMRC based on the improvement of additive attention is proposed.FastMRC uses this attention mechanism for global context modeling,which greatly improves the interaction speed between the generated global vector and each character in the text sequence,reduces the computational complexity of the model,and thus realizes efficient long text modeling with linear complexity.In this experiment,the CJRC judgment document data set was used for model training.The results showed that FastMRC improved EM by 5.1 percentage points and F1 by 4.1 percentage points compared to the baseline model.Furthermore,it is found in the experiment that both FastMRC and the pre-trained language model are difficult to accurately identify key sentence components in the processing of lengthy texts,which affects the accuracy of machine reading to understand the answers.Therefore,FastSemMRC is proposed based on the architecture of FastMRC,which integrates semantic dependency technology.FastSemMRC adds semantic tags to the coding layer and applies the mask strategy based on semantic dependency relationship to the semantic interaction layer,so as to realize explicit constraints on the coding sequence based on semantic information in the model,thus improving the accuracy of the model question answering.The results showed that FastSemMRC increased EM value by 6.5 percentage points and F1 value by 5.9 percentage points compared with the baseline model.Finally,this thesis takes FastSemMRC as the core model to design the police question and answer system,which includes machine reading comprehension module,data storage module and front-end page display module.The prototype police question and answer system supported by relevant policy documents of the Ministry of Public Security is realized,and the validity of the model in this thesis is verified.