A machine-compiled macroevolutionary history of Phanerozoic life

Many aspects of macroevolutionary theory and our knowledge of biotic responses to global environmental change derive from literature-based compilations of paleontological data. Although major features in the macroevolutionary history of life, notably long-term patterns of biodiversity, are similar across compilations, critical assessments of synthetic databases have been limited to the enumeration of taxonomic and geochronological errors in single lineages. Existing databases also leverage a small fraction of relevant published knowledge and are difficult to extend with new data types. Here, we develop a statistical machine reading and learning system, PaleoDeepDive, to automatically find and extract data from the text, tables, and figures of publications. We show that PaleoDeepDive requires comparatively little training data to perform comparably to humans in many complex data extraction tasks, and then deploy the system to extend the human-constructed Paleobiology Database to include nearly ten times more journal articles. Large-scale Phanerozoic taxonomic diversity and genus-level extinction and origination patterns are robust, even when derived from different bodies of literature. Unlike traditional databases, PaleoDeepDive produces a probabilistic database that improves as new information is added and that is extendable to include data not previously accessible on large scales, including morphological data in biological illustrations. Although literature-based compilations will always be subject to errors caused by inconsistent and erroneous data reporting, our high quality machine-reading approach to data synthesis and integration brings within reach questions that are now underdetermined and does so in ways that may stimulate new modes of inquiry.