Research And Implementation Of Knowledge-based Automatic Math Word Problem Solving Method

With the development of the information industry and the breakthrough of computing power bottlenecks,various artificial intelligence technologies represented by deep learning neural networks have made amazing progress.How to test the intelligence level of these artificial intelligences is a challenging task,and as one of the test tasks,automatic solving of mathematical word problems has attracted more and more researchers.The automatic math word problem solving requires the solution system to understand,model,and reason the input math word problem text before obtain the corresponding solution expression and calculate the final answer.The hard part of this task is that the model can not obtain the answers through current document question answering system,it needs to map natural language into a logical form that the machine can understand and perform reasoning to solve it.Considered to be a good scheme for evaluating artificial intelligence in natural language understanding,the emergence of a high-accuracy math word problem solver will be regarded as an important milestone in the development of artificial intelligence.The study of automatic math word problem solving has a long history.Since the 1960s,it has experienced three stages of development:symbolic rules,probability and statistics,and deep learning.Although the methods in the first two stages have good interpretability,they suffered such as relying on a large number of manual features or being unable to work on large-scale datasets.In recent years,the development of deep learning neural networks and the proposal of large-scale datasets have brought new research directions for the automatic math word problem solving,and more and more automatic math word problem solvers based on deep learning neural networks have emerged.However,most of these works are based on the information provided by the current input math word problem text,but ignore the external knowledge information such as common sense values and entity semantic relation,which are required for solving math word problems.How to introduce the external knowledge information into the automatic math word problem solving model so that the model can"think like a human being" has become a new challenge.This dissertation focuses on using various knowledge information to improve the automatic math word problem solving model.The main contributions of this dissertation are as follows:(1)An external knowledgeaware module is designed to associate related entity information.(2)Prompt template is designed to import external common sense values and adapt the tasks to the form of pre-training tasks.A large-scale externalvalue candidate dataset containing 246793 math word problems was constructed,and the search space was narrowed by searching similar questions combined with the Prompt template to solve external common sense values problems.(3)To address the problem of factual errors,we design a fact consistency verification module.And in the process of model generation,we propose a token-granularity generation verification scheme.(4)A large-scale untitled English math word problem dataset called CS355K,containing 355023 questions,was constructed to enable GPT-3 to perform few-shot learning based on a wider range of untitled math word problem data.(5)A knowledge-based automatic math word problem solving system was Implemented to provide users with automatic math word problem solving services.