Scientific Data Extraction from Oceanographic Papers

Scientific data collected in the oceanographic domain is invaluable to researchers when performing meta-analyses and examining changes over time in oceanic environments. However, many of the data samples and subsequent analyses published by researchers are not uploaded to a repository leaving the scientific paper as the only available source. Automated extraction of scientific data is, therefore, a valuable tool for such researchers. Specifically, much of the most valuable data in scientific papers are structured as tables, making these a prime target for information extraction research. Using the data relies on an additional step where the concepts mentioned in the tables, such as names of measures, units, and biological species, are identified within a domain ontology. Unfortunately, state-of-the-art table extraction leaves much to be desired and has not been attempted on a large scale on oceanographic papers. Furthermore, while entity linking in the context of a full paragraph of text has been heavily researched, it is still lacking in this harder task of linking single concepts. In this work, we present an annotated benchmark dataset of data tables from oceanographic papers. We further present the result of an evaluation on the extraction of these tables and the linking of the contained entities to the domain and general-purpose knowledge bases using the current state of the art. We highlight the challenges and quantify the performance of current tools for table extraction and table-concept linking.