Automatic Math Word Problem Solver With Deep And Reinforcement Learning

As one of the important standard tests for machine intelligence,the history of solving math word problems(MWP)automatically dates back to the 1960 s and has re-engaged the attention of a large number of researchers in recent years.The math word problem solver first maps human-readable sentences into machine-understandable logical forms,then infers the expressions,and finally calculates the answers.This task not only involves a deep understanding of the text,but also requires the solver to have strong reasoning ability,which is an important and difficult point in the study of natural language understanding and reasoning.In recent years,a large number of MWP solvers have been proposed,and the authors have claimed that their methods have achieved excellent results on their small-scale datasets.However,these methods can only support addition and subtraction,or require too much manual intervention,or as the difficulty of the data increases,the complexity of the solution increases dramatically and the performance drops dramatically.Therefore,in the past two years,large-scale MWP datasets have been released as more challenging test platforms for researchers,and the previous methods also have been tested on the datasets and their performance was poor.This paper focuses on the use of deep learning and reinforcement learning to construct an automatic solver for MWP.We first review the development of the two major MWP,i.e.,arithmetic and equations word problems,in recent years.Then the three solutions based on deep learning or reinforcement learning proposed in this paper are introduced: 1)An automatic solver for MWP based on deep-Q-network(DQN)is proposed,which reduces the exponential search space of constructing mathematical expression tree and can solve MWP more efficiently and accurately;2)An ensemble MWP solver based on three different types of SEQ2 SEQ and equation normalization are proposed.The solver reduces the manual cost on feature design and extraction,and equation normalization reduces the search space of the equation generation;3)A two-stage solution system for MWP based on recurrent neural network is proposed: we first apply a seq2 seq model to predict a tree-structure template,with inferred numbers as leaf nodes and unknown operators as inner nodes.Then,we design a recursive neural network to encode the quantity with Bi-LSTM and self attention,and infer the unknown operator nodes in a bottom-up manner.The method further reduces the search space of the template and improves the accuracy of solving MWP.At the end of the paper,based on the current research status,the paper discusses the points which are worthy of further research and exploration in MWP solving.