2015 GSA Annual Meeting in Baltimore, Maryland, USA (1-4 November 2015)

Paper No. 135-28
Presentation Time: 9:00 AM-6:30 PM

STROMATOLITE DISTRIBUTION IN SPACE AND TIME: A MACHINE-READING ASSISTED QUANTITATIVE ANALYSIS


WILCOTS, Julia1, HUSSON, Jon2 and PETERS, Shanan E.2, (1)Geosciences, Princeton University, Princeton, NJ 08544, (2)Department of Geoscience, University of Wisconsin–Madison, 1215 W Dayton St, Madison, WI 53706, jwilcots@princeton.edu

Stromatolites are products of microbial interactions between carbonate sediments and water, and are among the oldest evidence for macro-scale life on Earth. Because stromatolite growth is known to be inhibited by the presence of grazing metazoans, it is generally recognized that stromatolites are widespread and frequent in occurrence within Precambrian-aged sedimentary deposits and that Phanerozoic stromatolite occurrences are confined to saline or otherwise restricted aqueous environments. Despite the fact that stromatolites are routinely described in published accounts of stratigraphic successions, few studies have quantified their occurrence on large spatial and temporal scales. Here we attempt to provide such a measure of stromatolite distribution and abundance by: (1) manually searching the literature for descriptions of stromatolites in carbontate-rich units, and (2) utilizing machine reading and learning approaches to discover and extract information on stromatolite occurrences from publications. For the former, we utilized the Macrostrat database in North America; for the latter we used the entity recognition, distant supervision and probabilistic inference capabilities of DeepDive, a machine learning and knowledge base construction system, which we customized for the purpose of stromatolite occurrence recognition. Our DeepDive application was then run using thousands of published documents currently accessible via the UW–Madison GeoDeepDive infrastructure. We manually trained the DeepDive system with known instances of stromatolites and by annotating DeepDive extractions, thereby generating a larger number of increasingly accurate results. In combination with a manual search for stromatolite-bearing carbonate units in the Macrostrat database, we provide a new estimate of the spatial and temporal frequency of stromatolitic units that is measured against the total number of potentially stromatolite-bearing carbonate units in North America. Quantifying the stratigraphic distribution of stromatolitic carbonate units over the last three billion years provides a new perspective on the environmental signal captured by stromatolite growth.