Research On The Solution Of Elementary Arithmetic Problems Based On Deep Learning

It is very challenging for machines to automatically solve elementary mathematics and arithmetic problems.The main reason is that there is a large semantic gap between the thinking of humans to understand mathematical problems and the logical representations understandable by machines.Although the research on the machine to solve arithmetic problems began in the 1960s,there is still a big gap between the research results and people’s expectations.The previous research methods require too many human rules and manual interventions,and the application scenarios are single,which often can only solve arithmetic problems in the scenarios set by the researchers in advance.In recent years,the research on the machine solution of mathematics and arithmetic problems based on deep learning has regained great development,which makes people see the hope of problem solving in this field.This article adopts the current mainstream encoder/decoder(Sequence to Sequence,Seq2Seq)neural network structure to construct a mixed model for solving elementary mathematics and arithmetic problems,that is,first extract semantic information from the math problem text and then generate the corresponding equations of the problem text.So as to answer.This article focuses on the automatic solution of mathematics and arithmetic problems.The main work done is as follows:(1)A comprehensive and detailed combing and summary of the existing research methods for machine solutions to elementary mathematics and arithmetic problems.The history,background and technology of domestic and foreign research in this direction are introduced in detail,and the answering technology is explained in the three directions of symbolic semantic analysis,structure prediction,and sequence-to-sequence learning.(2)Format the collected elementary mathematics and arithmetic problems,and then label the corresponding equations and answer values.An expanded version of the math23K data set is proposed and used in the model experiment of this article.The entire data set has the same format.Through the expanded data set,the model can fully learn the rich semantic features and quantitative related information in the text.(3)This paper designs an encoder decoder(Seq2Seq)network model containing a graph encoder(GraphEncoder)and a tree decoder(TreeDecoder).This architecture combines the advantages of graph-based encoders and tree-based decoders.To generate better mathematical equations for arithmetic problems.Our graph encoder framework includes two graphs:the quantity unit graph and the quantity comparison graph.They are designed to graphically represent text information by effectively representing the relationship between the quantities in mathematical arithmetic problems and the order of quantity information.Through GraphEncoder,the model can fully learn the original semantic features and quantitative information in the encoding stage of the original text.In the decoding stage,the equation space can be normalized through the tree representation of the equation,so that the model can be well applied to arithmetic problems.Machine answers.(4)Carry out experiments based on the data sets and models in this article,and compare the experimental results of the Seq2Seq models with different structures designed in this article.And the final model with a complete structure(GraphEncoder+TreeDecoder)was analyzed experimentally,including module ablation experiment,graph convolution number experiment,operator number experiment,and case analysis experiment.The final answer accuracy rate of the model reached 78.8%,which was verified The effectiveness of this article in the data set and model design work.