Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2020, Vol. 46 ›› Issue (1): 60-66,73. doi: 10.19678/j.issn.1000-3428.0053702

Previous Articles     Next Articles

Stream Document Structure Recognition Based on Bidirectional LSTM Network

ZHANG Zhena,b, LI Ninga,b, TIAN Ying'aia,b   

  1. a. Beijing Key Laboratory of Internet Culture and Digital Dissemination;b. Computer School, Beijing Information Science and Technology University, Beijing 100101, China
  • Received:2019-01-16 Revised:2019-03-05 Online:2020-01-15 Published:2019-03-14

基于双向LSTM网络的流式文档结构识别

张真a,b, 李宁a,b, 田英爱a,b   

  1. 北京信息科技大学 a. 网络文化与数字传播北京市重点实验室;b. 计算机学院, 北京 100101
  • 作者简介:张真(1994-),男,硕士研究生,主研方向为文档信息处理;李宁,教授、博士;田英爱,副教授、博士。
  • 基金资助:
    国家重点研发计划"私有云环境下服务化智能办公系统平台"(2018YFB1004100);国家自然科学基金"流式文档排版格式的智能化分析与优化方法"(61672105)。

Abstract: Stream document structure recognition is important to automatic typesetting optimization and information extraction.The existing rule-based structure recognition method has a poor performance,and the machine learning-based method has a low recognition accuracy rate as it does not consider the long distance dependency between document units.To address the problem,this paper proposes a stream document structure recognition method based on bidirectional Long Short-Term Memory(LSTM) network.The method extracts key features in terms of the format,content and semantics of document units.Then it reduces document structure recognition to sequence labeling,and uses bidirectional LSTM neural network to construct a recognition model to implement recognition of 18 logical labels.Experimental results show that the method can effectively recognize the document structure,and has a better recognition performance than Founder FX software.

Key words: document structure recognition, stream document, feature extraction, sequence labeling, Long Short-Term Memory(LSTM) network

摘要: 流式文档结构识别对于排版格式自动优化和信息提取等具有重要作用。基于规则的结构识别方法泛化能力较差,而基于机器学习的方法未考虑文档单元之间的长距离依赖关系,识别准确率较低。针对该问题,提出一种基于双向长短期时间记忆(LSTM)网络的流式文档结构识别方法。从文档单元的格式、内容与语义3个方面筛选关键特征,并将文档结构识别看作序列标注问题,使用双向LSTM神经网络构建识别模型,以实现对18种逻辑标签的识别。实验结果表明,该方法能够对文档结构进行有效识别,其识别效果优于方正飞翔软件。

关键词: 文档结构识别, 流式文档, 特征提取, 序列标注, 长短期时间记忆网络

CLC Number: