Mathematical Word Problem Solver Based On Deep Adaptive Graph Convolutional Networ

The study of mathematical application problem solvers is one of the important tasks in the field of natural language processing.Mathematical application problem solvers reason out the information between quantities by recognizing the information in the problem text to derive the correct expressions and answers.In order to make the mathematical application problem solver accurately infer the position of the quantities in the expression and the order of operations,this paper proposes a new model of the mathematical application problem solver.The paper proposes a new model of mathematical application problem solver based on Graph2 Tree model,and names the new model of mathematical application problem solver as GDR model.The GDR model is a graph convolutional network that improves the graph encoding part of Graph2 Tree model,while constructing the relationship graph from the input to the graph encoding part,and applies the R-Drop algorithm on the training data.In order to verify the performance of the GDR model,two datasets are selected for relevant experiments in this paper,the two datasets are Math23 K,a large dataset,and MAWPS,a small dataset,and compared with other excellent models in mathematical application problem solvers,and the experimental results show that the GDR model outperforms the models of other mathematical supposed problem solvers.The main work of this paper is as follows.(1)In this paper,four relationship graphs between quantities and between quantities and word nodes are constructed as input for graph coding : the relationship graph of adjacent word nodes,the relationship graph of quantity nodes and adjacent word nodes,the relationship graph of quantity node sizes,and the relationship graph of the same word nodes adjacent to quantities.These four relationship diagrams can better represent the relationship between quantity nodes and word nodes,and capture the relationship between quantities through four relationship diagrams to enrich the representation of quantities.(2)In this paper,the graph convolutional network is improved based on the depth-adaptive graph neural network,and the improved network is a depth-adaptive graph convolutional network.The depth adaptive graph convolutional network is used as the graph coding part of the model to alleviate the occurrence of over-smoothing problem,which can make the graph coding stacked deeper and get higher level representation.(3)In this paper,the R-Drop algorithm is applied in the process of training the dataset to minimize the KL scatter between two distributions by randomly Dropout twice,so that the output is consistent twice,thus alleviating the overfitting phenomenon,which can improve the performance of the mathematical application problem solver model.(4)In this paper,we conduct experiments on two datasets,MAWPS and Math23 K,respectively.Math23 K is a large Chinese dataset with 23162 mathematical application problems,while MAWPS is a small dataset with 2373 mathematical application problems.The accuracy of the GDR model of the Math23 K dataset is 78.1%,and the accuracy of the MAWPS dataset is 85.5%,and the experimental results show that the GDR model outperforms other models for solving mathematical application problems.