An automatic selective PDF table-extraction method for collecting materials data from literature

Abstract

Table data in scientific literature is an important and economic data source for constructing materials database. The existing PDF table-extraction method is mainly designed for the common table type, which has no difference in various disciplines and does not have the ability to automatically filter the tabular data and extract non-full-framed tables with high precision. In view of this, we propose herein the use of unique coordinates for each object in a PDF and a method of automated table extraction from scientific literature based on text-state characteristics including six stages. In this method, we analyze the special presentation of table content and decode the PDF content stream to detect tables by key words of the table caption, especially use data ontology to filter irrelevant table data, and restore the data structure of tables according to the certainty and uniqueness of character coordinates. The proposed method automatically and accurately extracts table data from scientific literature without relying on table grid lines, thereby overcoming the drawbacks of existing technology for extracting data from three-line tables. The validity and advantages of the proposed method are verified by applying it to squeeze casting literature. Experiments show that the recall rate and precision of the proposed method reach 0.891 and 0.861. The comprehensive performance outperforms the main tools in the market for scientific literature table extraction.