Page 4 - i1052-5173-32-11
P. 4
Answering Geosciences Research Questions
at a Global Scale via a Hybrid Machine-
Human Learning Approach: A Case Study of
the Link between Climate and Volcanism
Seongjin Park*, Barbara Carrapa , University of Arizona, Tucson, Arizona 85721, USA; Mihai N. Ducea, University of Arizona, Tucson,
#
Arizona 85721, USA, and Faculty of Geology and Geophysics, University of Bucharest, 010041, Bucharest, Romania; Mihai Surdeanu,
Robert Hayes, Dan Collins, University of Arizona, Tucson, Arizona 85721, USA
ABSTRACT experiment conclude that volcanism influ- particular area can still be tracked by the
A common challenge in science is the ences climate change in deep time, but there interested geologist (the number of papers is
human capability to evaluate the real impact is no complete consensus on this question. still within reach of human processing), the
of an observation and a data set. This is a Similarly, any question of potential global importance of so many global-scale multi-
complex task due to having only partial significance, such as the impact of human disciplinary interpretations is difficult to
information and/or to the complexity of the activities on climate change, can be posed as evaluate. For example, did erosion of Earth’s
problem, requiring different fields to be an interrogating technique for our vast and surface increase globally since the Pliocene
combined. In order to overcome these fast-growing literature in the field of geosci- as the result of increased climate variability
important limitations, we need to be able to ences. Such an approach has the potential to (Zhang et al., 2001; Herman et al., 2013)?
review all the available data and interpreta- be applied to a variety of complex problems, Was tectonics the cause of CO drawdown
2
tions. This would allow us to evaluate the hence addressing some of the major limita- and global cooling in the Cenozoic (e.g.,
global distribution of a specific process or tions with cross-disciplinary research. Raymo and Ruddiman, 1992; Gernon et al.,
phenomenon of interest. The increasing 2021)? Did Earth’s surface topography
number of scientific publications prevents INTRODUCTION affect biodiversity through time (Badgley et
scientists from being able to keep up with all One of the cornerstone theories in natural al., 2017)? These are just a couple of exam-
the available literature especially when sci- sciences, Darwin’s evolutionism, states that ples of far-reaching but hard-to-evaluate
entific papers cross disciplines. These chal- the evolution of flora and fauna in the geo- research questions in a science that increas-
lenges prevent us from evaluating the global logic past goes through temporally deter- ingly requires ingestion of too much infor-
impact of a certain process and are particu- mined and irreversible extinctions corrobo- mation at a global scale and that commonly
larly relevant today given the impact of our rated with the development of new species. needs to be placed into a complex deep
scientific assessment on one of the most That theory has been vetted by innumerable time–space framework.
pressing issues of our time, which is climate observations and stands today because of To address these issues, we built a hybrid
change and its impact on society. We present that. However, most potentially ground- machine-human approach for the systematic
here an application of artificial intelligence breaking research questions in natural sci- analysis of scientific discoveries in geosci-
to geosciences: We conduct a systematic ences have a difficult time being resolved at ences. The proposed approach employs
analysis of geoscience literature through a global scales because of the complexity of machine reading to ingest publications at
hybrid machine-human approach. Such observations. In order to answer such ques- scale and aggregate scientific discoveries.
applications are more common in other tions at a global scale, we need to have a These models allow scientists to attempt a
fields such as biomedicine and are in their global review of the scientific literature. This wider understanding of science, which facili-
infancy in the geosciences because of vari- task has turned into a near impossible chal- tates the identification of (apparent) contra-
ous difficulties the machines encounter in lenge in recent years due to the vast amount dictions in scientific findings, as well as
parsing geologic literature. We describe of scientific data that have been published, “blank spaces” in the research landscape.
here some of these limitations and how we which exceeds human capacity for process- Note that approaches that summarize sci-
overcame them. We then use the following ing and interpretation. This is particularly entific work already exist, such as SCITE
case study as an example to test our problematic in fields like geosciences that (https://scite.ai) and SCITLDR (https://
approach: We ask whether climate is influ- require the interpretation of data and research scitldr.apps.allenai.org), both of which are
enced by volcanism in the geologic past. questions on a global scale and over large trained on previously published papers.
Our case study results show, as expected, time intervals. Whereas data pertaining to a However, their goals are different from what
that most analyzed literature in this specific field (e.g., regional geology) of a we aimed to achieve in our study. SCITE
GSA Today, v. 32, https://doi.org/10.1130/GSATG528A.1. BY-CC-NC.
*This author is currently affiliated with Educational Testing Service.
# Corresponding author: bcarrapa@arizona.edu
4 GSA TODAY | November 2022