Automatic detection and extraction of key resources from tables in biomedical papers.

Abstract

Background: Tables are useful information artifacts that allow easy detection of missing data 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. STAR*Methods key resource tables have increased the "findability" of these key resources, improving transparency of the paper by warning authors (before publication) about any problems, such as key resources that cannot be uniquely identified or those that are known to be problematic, but they have not been commonly available outside of the Cell Press journal family. We believe that processing preprints and adding these 'resource table candidates' automatically will improve the availability of structured and linked information about research resources in a broader swath of the scientific literature. However, if the authors have already added a key resource table, that table must be detected, and each entity must be correctly identified and faithfully restructured into a standard format.

Methods: We introduce four end-to-end table extraction pipelines to extract and faithfully reconstruct key resource tables from biomedical papers in PDF format. The pipelines employ machine learning approaches for key resource table page identification, "Table Transformer" models for table detection, and table structure recognition. We also introduce a character-level generative pre-trained transformer (GPT) language model for scientific tables pre-trained on over 11 million scientific tables. We fine-tuned our table-specific language model with synthetic training data generated with a novel approach to alleviate row over-segmentation significantly improving key resource extraction performance.

Results: The extraction of key resource tables in PDF files by the popular GROBID tool resulted in a Grid Table Similarity (GriTS) score of 0.12. All of our pipelines have outperformed GROBID by a large margin. Our best pipeline with table-specific language model-based row merger achieved a GriTS score of 0.90.

Conclusions: Our pipelines allow the detection and extraction of key resources from tables with much higher accuracy, enabling the deployment of automated research resource extraction tools on BioRxiv to help authors correct unidentifiable key resources detected in their articles and improve the reproducibility of their findings. The code, table-specific language model, annotated training and evaluation data are publicly available.