Machine Reading for Scientific Discovery

Abstract

The aim of this work is to accelerate scientific discovery by advancing machine reading approaches designed to extract claims and assertions made in the literature, assemble these statements into cohesive models, and generate novel hypotheses that synthesize findings from isolated research communities. Over 1 million new publications are added to the biomedical literature each year. This poses a serious challenge to researchers needing to understand the state of the field. It is effectively impossible for an individual to summarize the larger body of work or even remain abreast of research findings directly relevant to a subtopic. As the boundaries between disciplines continue to blur, the question of what to read grows more complicated. Researchers must inevitably turn to machine reading techniques to summarize findings, detect contradictions, and illuminate the inner workings of complex systems. Machine reading is a research program in artificial intelligence centered on teaching computers to read and comprehend natural language text. Through large-scale machine reading of the scientific literature, we can greatly advance our understanding of the natural world. Despite remarkable progress (Gunning et al., 2010; Berant et al., 2014; Cohen, 2015a), current machine reading systems face two major obstacles which impede wider adoption: <i>Assembly</i> The majority of machine reading systems extract disconnected findings from the literature (Berant et al., 2014). In areas of study such as biology, which involve large mechanistic systems with many interdependent components, it is essential that the insights scattered across the literature be contextualized and carefully integrated. The single greatest challenge facing machine reading is in learning to piece together this intricate puzzle to form coherent models and mitigate information overload. In this work, I will demonstrate how disparate biomolecular statements mined from text can be causally ordered into chains of reactions (Hahn-Powell et al., 2016b) that extend our understanding of mechanistic biology. Then, moving beyond a single domain, we will see how machine-read fragments (influence relations) drawn from a multitude of disciplines can be assembled into models of children’s heath. <i>Hypothesis generation and “undiscovered public knowledge”</i> (Swanson, 1986a) Without a notion of research communities and their interaction, machine reading systems struggle to identify knowledge gaps and key ideas capable of bridging disciplines and fostering the kind of collaboration that accelerates scientific progress. With this aim in mind, I introduce a procedure for detecting research communities using a large citation network and derive semantic representations that encode a measure of the flow of information between these groups. Finally, I leverage these representations to uncover influence relation pathways which connect otherwise isolated communities.