Automatic Detection and Extraction of Key Resources from Tables in Biomedical Papers

Abstract Tables are useful information artifacts that allow easy detection of data “missingness” by humans and have been deployed by several publishers to improve the amount of information present for key resources and reagents such as antibodies, cell lines, and other tools that constitute the inputs to a study. The STAR*Methods tables, specifically, have increased the “findability” of these key resources, but they have not been commonly available outside of the Cell Press journal family. To improve the availability of these tables in the broader biomedical literature, we have attempted to automatically process BioRxiv preprints to create tables from text or to recognize tables already created by authors and structure them for later use by publishers and search systems, to improve “findability” of resources in a larger amount of the scientific literature. The extraction of key resource tables in PDF files by the best in class tools resulted in Grid Table Similarity (GriTS) score of 0.12, so we have created several multimodal pipelines employing machine learning approaches for key resource table page identification, Table Transformer models for table detection and table structure recognition and a new table-specific language model for row over-segmentation to improve the extraction of text in tables created by biomedical authors and published on BioRxiv to around GriTS score of 0.90 enabling the deployment of automated research resource extraction tools onto BioRxiv. Author summary Tables are useful information artifacts that allow for easy detection of data “missingness” by humans and have been implemented by several publishers to improve the amount of information present for key resources and reagents such as antibodies, cell lines, and other tools that constitute the inputs to a study. To improve the availability of these tables in the broader biomedical literature, we introduced four pipelines for key resource table extraction from biomedical documents in PDF format. Our approach reconstructs key resource tables using image level table detection and structure detection generated table boundary, column (and row) bounding box information together with PDF text alignment. To remedy row over-segmentation resulting from overflowing table cell contents, we introduced a language modeling (LM) based row merging solution where a character-level generative pre-trained transformer (GPT) model was pre-trained on more than 11 million scientific table contents from PubMed Central Open Access Subset (PMC OAS). All introduced pipelines significantly outperformed GROBID baseline while our Table LM based row merging based pipeline, significantly outperformed all other pipelines including our OCR based pipeline.