基于深度学习的非结构化表格文档数据抽取方法  

Data Extraction Method of Unstructured form Documents Based on Deep Learning

作　　者：奚建飞 王志英[2] 邹文景 甘莹 XI Jianfei;WANG Zhiying;ZOU Wenjing;GAN Ying(Digital Grid Research Institute, CSG, Guangzhou 511455, China;China Southern Power Grid Co. Ltd., Guangzhou 510663, China)

机构地区：[1]南方电网数字电网研究院有限公司,广东广州511455 [2]中国南方电网有限责任公司,广东广州510663

出　　处：《微型电脑应用》2022年第2期102-105,共4页Microcomputer Applications

摘　　要：非结构化表格文档结构性较低,模式多样且数据冗杂,但此类文档里潜藏大量有价值数据,数据高精度抽取对分析数据价值存在增值作用,为此提出基于深度学习的非结构化表格文档数据抽取方法。在数据抽取前,采用基于循环和卷积神经网络的文本分类方法,对非结构化表格文档实施分类,获取所需表格文档,由此缩小后续数据抽取范围,提高抽取效率与精度;在分类后的表格文档中使用基于深度学习的数据自动抽取模型,通过双向循环神经网络编码获取中间语义向量,然后通过注意力模型和单项循环神经网络解密中间语义向量,获取非结构化表格文档数据。实验结果表明,采用所提方法抽取10种非结构化表格文档数据时,均获取显著的抽取效果。Documents with unstructured form have low structure,diverse patterns and redundant data.There are a lot of valuable data in this kind of documents,and high-precision data extraction has a catalytic effect on the analysis of data.Before data extraction,the text classification method based on cyclic and convolutional neural network is used to classify unstructured form documents and obtain the required form documents,so as to reduce the scope of subsequent data extraction and improve the efficiency and accuracy of extraction.After classification,the automatic data extraction model based on deep learning is used in the documents,and the two-way cyclic neural network coding is used to obtain the data.The intermediate semantic vector is then decrypted by attention model and single loop neural network to obtain unstructured form document data.The experimental results show that when the proposed method is used to extract 10 kinds of unstructured form document data,significant extraction results are obtained.

关 键 词：深度学习 非结构化表格 文档数据 抽取方法 卷积神经网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]