一种求解数学应用题的多粒度图神经网络编码器

Solving Math Word Problems by Multi-grained Graph Neural Networks

近几年,数学应用题自动解答(Math Word Problems,MWP)的研究受到越来越多学者关注,大多数研究的重点是对编码器的改进。然而目前的研究在编码器的改进方面还存在以下问题:(1)输入文本的颗粒度一般是字级别,这会导致泛化能力不足;(2)大多数模型对文本信息的挖掘没有充分利用文本内实体、词性等信息,只是停留在时序信息层面。该文针对以上问题,在双向GRU(Gated Recurrent Unit)的基础上提出了一种新颖的基于多粒度分词和图卷积网络的编码器结构(Multi-grained Graph Neural Networks,MGNet)。多粒度分词是通过对文本的每个词进行不同颗粒度的分词,增加了样本容量,并且通过引入一些噪声样本,提高了模型的泛化能力。图卷积神经网络通过构建文本内实体、数字、日期之间的不同的属性图,对它们之间隐含的关系进行建模。在Math23K和Ape210K数据集的实验显示,该文提出的模型MGNet准确率分别达到77.73%和80.8%。

In recent years,the task of automatically solving Math Word Problems(MWP)has received more and more attention and most of researchers focused on improving the encoders.The issues in current encoders include:(1)The input granularity is character level and it will cause insufficient generalization ability;(2)Only the text sequence is modeled without capturing the entities,part of speech and other textual information.To alleviate the above issues,this paper proposes a novel encoder structure based on multi-grained word segmentation and graph convolution(MGNet)using the bidirectional GRU(Gated Recurrent Unit).The multi-grained word segmentation increases the sample capacity by segmenting the text with different granularities and improves the generalization ability of the model by introducing some noise samples.The graph convolutional neural networks can learn the implicit relationship among name entities,numbers and dates by constructing different attribute graphs among them.Experiments on two public benchmarks Math23K and Ape210K datasets show that our proposed MGNet can achieve the accuracy of 77.73% and 80.8% respectively.