Research And Implementation Of Case-Based Reasoning Algorithm Based On Graph Neural Network And Pre-Training Model

Case-based reasoning is an important research direction in the field of natural language processing in artificial intelligence,aiming at inferring solutions to new problems from known cases,including four steps: case representation,case storage,case matching and case adaptation.Cases often exist in the form of long texts,which are much more difficult in feature extraction,information storage and calculation than short texts,and the probability of noise and errors is higher.Based on the pre-training model and graph neural network,thesis focuses on the improvement of case representation and case matching,and applies the improved algorithm to the legal field,and designs and implements a case retrieval system.The main contents and innovations of thesis are as follows:(1)A behavior extraction algorithm based on RoBERTa-BiLSTM-GCN is proposed.This algorithm transforms behavior extraction into semantic role labeling,and extracts behavior according to the structure of "agent+predicate+patient".The model includes an encoder and a decoder.The encoder consists of a pre-training model RoBERTa-wwmext-large,a self-attention layer,a bidirectional long-term memory network(BiLSTM)and a graph neural network(GNN),and the decoder consists of a conditional random field(CRF)layer.This algorithm is trained and tested on CoNLL-2012 data set.Compared with other methods,the optimal accuracy rate is 3.27 percentage points higher,the recall rate is 1.92 percentage points higher,and the F1 value is 2.60 percentage points higher.It is proved that this algorithm can effectively extract the pre-and-post dependencies in the sequence data in the text and the graph structure information based on syntax,and it has quite good behavior extraction ability.(2)A case matching algorithm based on graph model and SBERT-BiMPM is proposed.The key problem in finding similar cases is to calculate the similarity between the two cases.Firstly,based on the disorder of entities and the order of behaviors,an entity behavior graph model is constructed.Entities get word embedding through GloVe,and behaviors get sentence embedding through Sentiment-Bert.Then,the entity embedding representation sequence and behavior embedding representation sequence pass through the context representation layer,multi-view matching layer,aggregation layer and output layer in turn,and the similarity of the two cases is obtained.The multiview matching layer is the core layer,including three matching methods: maximum pooling matching,attention matching and maximum attention matching.Compared with other methods,the optimal accuracy of this model on CNSE and CNSS data sets is 2.72 and 1.92 percentage points higher,and the F1 value is 3.36 and 1.93 percentage points higher.This proves that this model can effectively obtain the key information in long texts,solve the problem that the information in long texts is not concentrated,and at the same time,it can make good use of the extracted key information to get the similarity between two long texts,which has quite good case matching ability.(3)The behavior extraction algorithm and case matching algorithm are applied to the legal field.Based on C/S(Client/Server)architecture,a case retrieval system is designed and implemented,and the new case is assisted in judgment through similar cases.Ten kinds of common causes were selected for the test,and the accuracy rate was above80% except one,which had certain practical value.