Abstract: 【Purpose/significance】PDF documents can faithfully preserve the content and appearance of the original document, but it
brings great difficulties to the parsing of the document. In order to mine multilingual scientific and technological texts more compre?
hensively and automatically, the methods of effective parsing and knowledge extraction for PDF documents of multilingual scientific
papers are explored.【Method/process】This paper proposes a Rule-Faster-RCNN-based method for the structural frame elements of
multilingual scientific papers in PDF file format. The structural frame elements of papers' full text are divided into text elements and
graph-table elements, and the rules are combined with deep learning methods to extract them respectively. The rule method uses the
characteristics of the paper layout to identify the text frame elements and graph-table elements, the deep learning method regards
graph-table recognition as target detection and constructs a Faster-RCNN network to supplement the deficiencies of the rule method.
【Result/conclusion】Experiments show that the PDF parsing method proposed in this paper is superior to the baseline method, and the full-text knowledge of scientific papers is successfully obtained effectively.【Innovation/limitation】The rules and deep learning meth? ods are used to extract the structural framework elements of the full-text documents of multilingual scientific papers more precisely, and the effectiveness of the method is verified by experiments; however, due to the complexity of PDF documents, the table elements are only extracted as pictures, and the text information inside the table cannot be mined.