Combining pattern matching with word embeddings for the extraction of experimental variables from scientific literature

Scientists frequently use experiments published in other articles or reports by governing entities (e.g. NIH) as templates for reporting on their own experiments. Those templates occasionally change to reflect new discoveries. For creating retrospective studies and meta-analyses, finding the template parameters associated with scientific results can be critical. To aid in the extraction of experimental parameters (e.g. animal housing temperature) in a corpus of ∼8M scientific reports, we used a combination of pattern matching, part of speech tagging, units and measures extraction, and machine learning. We describe a use case where the housing temperature used for experiments involving mice was shown to impact their response to tumor reduction drugs. We show that 1) combining deep learning and pattern matching is a good model to address the problem described and 2) that researcher's behavior and experimental template usage takes a while to change after the publication of an important discovery.