Automatic Dialogue System For Short Text Questions Of Marriage Law Based On Decision Tree

With the development of the Internet,there are more and more online legal consultations,but at present most of them are undertaken by lawyers.In order to liberate the work of lawyers,various consulting systems have developed.The current legal consulting system also faces problems such as weak logical reasoning,poor sentence representation,etc.This paper mainly focuses on the ability of sentence representation,the ability to extract attribute values,and the design of dialogue processes.This article mainly studies the following:(1)Less structural information in the existing sentence semantic similarity calculation leads to low precision,a sentence semantic similarity calculation method of Tree-LSTM based on the multi-head attention mechanism(MA-Tree-LSTM)is proposed according to the dependency tree.First,for achieving a fusion of multi-head attention mechanism and Tree-LSTM,MA-Tree-LSTM takes external instructive features as input,and then the input combined with the multi-head attention mechanism work on all child nodes of the Tree-LSTM tree node,which assigns different weight to each child node.Second,three-layers MA-Tree-LSTM is used to sentence semantic similarity calculation and realizes the mutual guidance of sentence pairs.Then,the multi-layers representation of the semantic features in sentences can be obtained.Finally,the semantic feature of the multi-layer is chosen to establish the semantic similarity calculation model of sentence,which makes full use of the semantic structure features in the sentence pairs.The proposed method is robust,interpretable,and insensitive to the order of sentence words,and does not require feature engineering.According to the experimental results on the SICK and STS datasets,the accuracy of sentence semantic similarity calculation based on MA-Tree-LSTM is better than Tree-LSTM method with the non-attention mechanism and the BiLSTM method with multi-head attention mechanism.(2)For the existing BERT attribute value extraction method,it is unable to capture long-distance features and weak generalization ability caused low accuracy.This paper proposes a attribute value extraction method based on multi-model integration.This method implements fine-tuning of the model by adding BERT_CLS,BERT_AVG,BERT_BiLSTM,and BERT_CNN hidden layers to the BERT output.Then,20 models are obtained by training with a 5-fold cross-validation method.Finally,the results of 20 models are integrated through the stacking ensemble learning method.The method proposed in this paper is easy to parallel,has strong generalization ability,does not require feature engineering.The experimental results on the RACE dataset show that the BERT model based on multi-model integration is superior to the single-model BERT method in accuracy.(3)A task-oriented automatic dialogue system based on decision tree was designed and implemented for real-time legal consultation.The system first discretizes the conclusions of legal consultation into category attributes.The information of the users involved in the conclusion is discretized into attributes,thereby translating the legal consultation problem into a classified prediction problem.Secondly,by collecting actual cases as training samples,a legal consultation classification prediction model based on C4.5 decision tree algorithm was established.Finally,when there is a new user comes to consult,with a simple decision tree of one round of question and answer,the method of calculating the similarity of sentences is used to return the result.For a decision tree of multiple rounds of question and answer,starting from the decision attribute of the root node of the established decision tree,the problem corresponding to the decision attribute is thrown to the user to ask questions.For the answer of the user,the attribute value corresponding to the decision attribute is obtained by the SVM algorithm.The attribute value determines the decision tree branch,the next attribute,and the question.The consultation process does not end until the leaf node is reached,and classification category corresponding to the leaf node is returned to the user.In the case of the divorce problem,the designed automatic dialogue system is highly explanatory,high precision,less problematic,high-time,and greatly reduces the amount of manual work.How to extract basic attributes and how to prevent overfitting of decision tree through post-pruning are further research directions.In the subsequent work,we will consider the type of problem and implement an interpretable model.We will further study how to explain the reasoning of the model on the RACE dataset.