Research On Sentence Matching Implementation Method For Legal Case Simulation System Evaluation

The progress of artificial intelligence technology and the publicity of judicial data have a far-reaching impact on the legal industry.With the help of natural language processing technology,establishing knowledge base of professional domain,assisting teachers in the automatic evaluation of students is one of the trends of intelligent teaching.The legal case simulation system contains a large number of legal cases,fully relies on Internet technology,simulates the real litigation process,and realizes intelligent teaching,and provides some inspiration for exploring the way out of legal education.Through the demand analysis of the legal case simulation system,the task of evaluating the lawsuit requests of the simulation system is transformed into matching problem of litigation request and judgment items.At present,there are few related literature reports in this aspect of research work.Therefore,this thesis conducts a tentative study on the method of sentence matching implementation for legal case simulation system evaluation.The main work has the following aspects:(1)Develop a legal case lawsuit request and judgment matter sentence matching data set for the evaluation of the legal case simulation system.The work content of this part mainly includes:analyzing the characteristics of the judgment,the crawling of the judgment data,the extraction of litigation requests and judgments,and the design of appropriate labeling specifications.(2)Study unsupervised litigation request and judgment matter matching method.A matching method based on simple weighted word vector averaging is used as a baseline system for unsupervised matching method,we implemented a matching method based on smooth inverse frequency weighted word vector averaging.This method uses pre-training word vectors and integrates the word frequency information of large-scale corpus.The corpus frequency information is used to weight the sum of the word vectors corresponding to all the words in the sentence,and then uses the principal component analysis to reduce the noise,and finally calculate the cosine similarity of the sentence vector.The experimental results show that the machting method based on smooth inverse frequency weighted word vector averaging achieves relatively good performance compared to the matching method based on simple weighted word vector averaging.(3)Study supervised litigation requests and judgments matching methods.The bilateral multi-perspective matching model is used as the baseline system for supervised matching method.The method matches the vector representation of a certain time of a sentence with the representation of all time vectors of another sentence vector in two directions.Match the strategy and finally aggregate and classifiy.Study a matching method based on multiple attention mechanism model.The method enhances the word-level interaction in sentence modeling by combining multiple attention mechanisms,and uses the "gate" mechanism to extract key information,and finally classifiy by multi-layer perceptron.The experimental results show that the multi-attention mechanism model achieves relatively good performance compared to the bilateral multi-perspective matching model.(4)Design and construction of submodule of litigation request evaluation of legal case simulation system.Simulate the real case litigation process,and use the multi-attention mechanism model sentence matching method to realize the intelligent evaluation of the litigation request written by the user.