Understanding the Structure of Streaming Documents based on Neural Network
Yutong Jiang, Ning Li, Yingai Tian · 2019
Document structure understanding can obtain the structural information of long-form documents, which plays a key role in the automatic layout of document formats. In the main document format, the structure of streaming documents is difficult to understand, and the current recognition effect is not ideal. According to the characteristics of streaming documents, neural network has obvious advantages for sequence labeling problems, and then a neural network based streaming document structure identification method is proposed. First, the format features, content features and semantic features are extracted, and deep and cross network model is introduced to apply the feature intersections in an automated manner. Secondly, using Long- and short-term memory(LSTM) neural network to construct the recognition model can identify eighteen types of logical labels and document structures more accurately. Finally, using the migration learning method, the structure identification of reports and other types of documents is initially realized. Experiments show that the recognition model proposed in this paper has better document structure recognition ability than other machine learning-based models or methods, and the effect is better than the current best products.