Implementation Of Automatic Problem Solving And Generation System Based On Mathematical Semantic Understanding

Practicing math word problems is an essential part of the teaching process that can help students improve their ability to apply mathematical knowledge to solve practical issues.However,manually solving or creating math word problems is a tedious and repetitive task that wastes a significant amount of teachers’ resources and time.With the advancement of artificial intelligence technology,utilizing natural languageunderstanding and natural language generation techniques to enhance the quality and quantity of math word problems resources has become a trend.This thesis addresses the problem of deep learning models’ inadequate understanding of mathematical semantic information due to the unique characteristics of math word problem’s text.To achieve the goal of understanding and representing mathematical semantic information,this thesis investigates related techniques for text feature representation and structural feature representation.Based on these techniques,an algorithmic model for automatic problem-solving and problem-generating is designed and implemented,which relies on mathematical semantic comprehension.The main contributions and accomplishments of this thesis are outlined as follows:(1)In response to the issue that the BERT model trained on a general corpus cannot fully represent semantic information with domain-specific features,this thesis designs three types of pre-training tasks based on the text features of math word problems and employs corresponding mask strategies for different character types to obtain the MBERT model that possesses mathematical semantic information.Experimental results demonstrate that compared with the BERT model,the MBERT model has stronger mathematical semantic representation ability.(2)In response to the issue of inadequate utilization of semantic information in pre-training models by existing automated problem-solving models,this thesis proposes to use the intermediate layer semantic information of the MBERT model for automatic mathematical problemsolving tasks.Based on the construction of an entity attribute graph that describes the structure information of the text of the problem,the node feature representation is enhanced using the intermediate layer of the MBERT model.The corresponding attention weight matrix of the intermediate layer is used to weight the edges in the entity attribute graph and construct a complementary dependency graph,thereby constructing an encoder model with intermediate layer semantic information.Experimental results show that compared to the Graph2Tree model,the proposed automated problem-solving model has higher expression prediction accuracy and answer value prediction accuracy.(3)A framework for an automatic question generation model was proposed to enhance semantic information.The consistency between the generated question text and the input topic words and expressions was achieved from two aspects.In terms of encoder and decoder design,a directed graph was used to enhance expression structure information and entity structure information expression,and to obtain enhanced input information representation.The generated question text was then inputted into the automatic problem-solving model described in this article.Based on the corresponding expression of the generated question text and the input expression,a reward score was calculated,and the model parameters were jointly adjusted with the model’s own loss function to further improve the quality of the generated question text.(4)An automatic problem-solving and problem-generating system for mathematical application questions is developed,which is implemented using a browser/server architecture to provide basic functions such as exercise and user profiling.The system is equipped with the algorithm models for automatic problem-solving and problem-generating described in this paper,which enabled intelligent functions for automatic problemsolving and problem-generating.The system aims to improve both the quality and quantity of mathematical application question resources and to provide users with a good user experience.