| Extracting structured information from paper documents is a hot research topic.The current solutions often rely more on manually written rules,which are inefficient.Paper documents will still coexist with digital documents for a long time,and there is an urgent need for automated information extraction of these contents.In this subject,a neural network-based scheme is designed to extract structured information and reduce the pressure of enterprise personnel registration.Specifically,the main achievements of this study include:1.Design and implement a multi-modal processing model for obtaining reading order based on spatial position(ROBS).The model uses spatial position to replace the original reading order,integrates text semantic information,text structure information and image information.The experimental results show that the model proposed in this paper effectively improves the performance of document information extraction2.Extend the encoder model-based architecture and propose a generative modelbased structure,which can directly generate answers,improve the end-to-end performance of the model and have error correction capabilities.Compared with the traditional model,the model in this paper has very superior performance in complex layout analysis.3.A series of engineering optimizations were performed on the model.Proposed solutions for the shortcomings of traditional natural language processing models,such as difficulty in processing long texts,and difficulty in multi-model feature fusion.By optimizing the generative model generation logic and proposing a shared encoder structure,the model inference speed is greatly improved.For the problems of document text error and text recognition errors encountered in practical applications,the error data generation method is designed and a dataset is made,which verifies the robustness of the model scheme in this paper to complex layout and text recognition errors;4.Abstracted the training process and design a document-structured information extraction platform that is easy for enterprise personnel to use.Users do not need to master professional knowledge,they only need to collect the layouts that need to be processed and mark the fields they are interested in,and then they can customize their own models.The platform supports operations such as dataset management,model training,and model export,and is highly user-friendly. |
