Argumentation in scholarly biomedical literature: Computational theory, implementation, and supporting deep learning software

点击翻译按钮获取中文摘要

Scientific research articles are structured arguments meant to persuade the reader to accept the authors' research claim. My research program is (1) to develop a computational theory of argumentation structures found specifically in published biomedical research papers, and (2) to translate these text-based argumentation structures to Abstract Dialectical Frameworks (ADFs), the most general computational model of argumentation in the Nonmonotonic Reasoning sub-field of Artificial Intelligence. There are a number of novel aspects of my research program: 1) the argumentation to be identified will include arguments in single papers and across multiple papers, 2) it incorporates deep learning techniques, and 3) it uses a Big Data problem to test the robustness of ADFs. Outcomes from our text analytics will have profound societal benefits in the fields of experimental and clinical medicine. The research program has started and will continue to develop a computational theory of rhetorical moves, a precursor of argumentation structure, for biochemistry articles. Following on our successful investigation of argumentation in the Discussion sections of biochemistry journal papers, we will further investigate argumentation in Introduction, Results, and Methods sections of these articles. The culmination of this is an argumentation structure for a single paper. Tracing citations will permit the formation of arguments across multiple papers. The argumentation is rooted in the author's claim and is supported by the accompanying experimental evidence. To construct the argumentation structure, these components must be identified in the text. This has been accomplished using standard supervised machine learning (ML) methodology which requires analysis of the text using CL techniques to extract text-based features to generate the model. This methodology will be enhanced with the use of deep learning. The text analysis will use the many widely used information resources (ontologies, databases, software tools) that have been developed by biomedical and CL researchers for this purpose. We are developing novel information extraction resources to distinguish some of these text features: an ontology of verb usage in scientific writing, phenotype name recognition, and method-mention recognition, among others. A final goal of this computational approach to argumentation structure in biomedical texts is to map texts into Abstract Dialectical Frameworks. Having computer-based representations of scientific arguments can lead to significant changes in how biomedical research is performed. Inconsistent experimental results can be detected, unrecognized biomedical relationships can be discovered, and the vast biomedical literature will be synthesized into a more accessible form for clinicians' use.