An Improved Automatic Math Problem Solver Based On Temporal Convolutional Networks And Multi-head Attention

Math word problems are any numerical problems written in natural languages,based on any subject domain(mathematics,physics,chemistry,biology etc.)Due to the variety of types and complexity of math word problems,it is necessary to make appropriate reasoning for the expressions and answers to solve these problems described by natural language.Therefore,the design of automatic math word problem solvers has been a hot topic in applying artificial intelligence to education.In recent years,most top-level math word problem solvers have adopted deep learning methods.However,they generally ignore the importance of data preprocessing and lack the consideration of the time sequential problem of the text of the math word problems.Therefore,this paper aims to explore different ways to improve the performance of the current popular automatic math word problem solver.The core of designing automatic math word problem solver is to apply the advanced natural language processing technology and computer application technology to the whole model of automatic math word problem solver.It mainly includes two modules: data preprocessing and model training.First of all,this paper comprehensively analyzes the characteristics of the datasets of math word problems.Its application value in the automatic math word problem solver model is explored.Secondly,this paper analyzes and summarizes the related algorithms based on deep learning model.The performance advantages of different deep learning algorithms in building automatic math word problem solvers are explored.Finally,this paper proposes an automatic math word problem solver model TMASeq2 seq,which integrates the Seq2 seq model,Multi-Head Attention Mechanism and Temporal Convolution Network(TCN).Three improvement schemes are proposed for the establishment of automatic math word problem solver.Experiments are conducted on two common public datasets.All of them improve the accuracy of the automatic mathematics application problem solver.The main work of this paper is as follows:(1)TMASeq2seq adopts the better performance of word segmentation tool(Fool NLTK)and word embedding method(ELMO)to form the data preprocessing module.More attention is paid to the effect of data preprocessing on the performance of math word problem solver.(2)TMASeq2seq has added a Multi-Head Attention Mechanism.When calculating the context attention,it adopts the method of partial analysis first,and then whole reasoning.The original context attention is divided into several sub domains for calculation,and then combined reasoning.This method can not only achieve good parallelism,but also capture longer distance dependencies.(3)TMASeq2seq combines the traditional Seq2 seq model with the time convolutional network TCN for the first time to solve math word problems automatically.It enlarges the receptive field of the decoder and enhances the processing ability of sequential data to a certain extent.(4)On large-scale datasets,TMASeq2 seq achieves better performance than the current best integrated seq2 seq model(Ensemble model),with accuracy increasing from 68.4% to 70.5%.On a small dataset,TMASeq2 seq also achieved performance improvements,with an accuracy rate of 91.3%.