一种自动求解数学应用题的双路文本编码器  

A Dual Channel Text Encoder for Solving Math Word Problems

作　　者：肖菁[1] 何岱俊 曹阳[1] XIAO Jing;HE Daijun;CAO Yang(School of Computer Science,South China Normal University,Guangzhou 510631,China)

机构地区：[1]华南师范大学计算机学院,广州510631

出　　处：《华南师范大学学报（自然科学版）》2023年第1期36-44,共9页Journal of South China Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(62177015);国防科技重点实验室稳定支持经费项目(WDZC20205250410)。

摘　　要：近年来,得益于人工智能技术(Artificial Intelligence,AI)的快速发展,关于自动求解数学应用题(Math Word Problem,MWP)的研究越来越趋向成熟。在自动求解数学应用题任务中,对问题文本进行建模至关重要。针对这一问题,文章提出了一个基于循环神经网络(Recursive Neural Network,RNN)和Transformer编码网络的双路文本编码器(Dual Channel Text Encoder,DCTE):首先,使用循环神经网络对文本进行初步的编码;然后,利用基于自注意力(Self-attention)机制的Transformer编码网络来获得词语的远距离上下文语义信息,以增强词语和文本的语义表征。结合DCTE和GTS(Goal-Driven Tree-structured MWP Solver)解码器,得到了数学应用题求解器(DCTE-GTS模型),并在Math23k数据集上,将该模型与Graph2Tree、HMS等模型进行了对比实验;同时,为探讨编码器配置方法对模型效果的影响,进行了消融实验。对比实验结果表明:DCTE-GTS模型均优于各基准模型,答案正确率达到77.6%。消融实验结果表明双路编码器的配置方法是最优的。In recent years,with the rapid development of artificial intelligence(AI)technology,researches on automatic solving of Math Word Problems(MWP)have been improved.In the task of automatically solving MWP,the modeling of the problem text is very important.For this issue,a Dual Channel Text Encoder(DCTE)based on Recursive Neural Network(RNN)and Transformer is proposed.DCTE firstly uses an RNN to initially encode the problem text,and then uses the Transformer based on the self-attention mechanism to obtain the long-range contextual information to enhance the representation of the word and problem text.Combining DCTE and GTS(Goal-Driven Tree-structured MWP Solver)decoder to obtain our math word problem solver DCTE-GTS.In this paper,DCTE-GTS is experimented on Math23k dataset and compared with Graph2Tree,HMS and other models.Meanwhile,the ablation experiments are also conducted to explore impact of encoder configuration on model’s perfor-mance.The experimental results show that the proposed DCTE-GTS is better than the baseline models,obtaining an answer accuracy of 77.6%.The ablation experiments show that the configuration of DCTE is the best.

关 键 词：人工智能 数学应用题 循环神经网络 自注意力 树形解码器 

分 类 号：TP319[自动化与计算机技术—计算机软件与理论]