Since the existing machine solving methods of math word problems cannot adaptively understand the text of the problem with changing semantics,and have a limit in the improvement of solving accuracy,this paper proposed a machine sol-ving method based on semantic understanding enhancement.Firstly,this method designed a semantically enhanced pre-train-ing language model SeBERT to accurately understand the topic through a multi-granularity knowledge modeling strategy and con-tinuous semantic integration strategy.Secondly,this method constructed the solution model SeBERT-PT,which adopted the solution structure of language model-pool-tree to effectively improve the semantic understanding deviation of word problems and the accuracy of understanding problems.Finally,it introduced a confidence-based judgment mechanism to directly deter-mine the failure of solving untrustworthy predictions,ensure the accuracy of the solution,and improve the training efficiency of solving models.The experimental results show that the accuracy results on Chinese and English datasets are 85.7％and 77.9％respectively,which is superior to other baseline methods,especially on problems involving complex semantic under-standing and logical reasoning.It has proved the effectiveness of the method in improving the accuracy of solving math word problems and demonstrates its wide applicability in cross-language environments.
关键词
数学应用题求解/预训练语言模型/语义增强/池化/置信度
Key words
math word problem solution/pre-trained language model/semantic enhancement/pooling/confidence
维普
万方数据