Pattern-based extraction of argumentation from the scientific literature

As the number of publications in the biomedical field continues its exponential increase, techniques for automatically summarizing information from this body of literature have become more diverse. In addition, the targets of summarization have become more subtle; initial work focused on extracting the factual assertions from full-text papers, while recent interest has shifted to recovering statements involving certainty, like speculations and agreements or disagreements with other research. Scientific writing is rife with such argumentation, and the premises, evidence, conjectures, objections and rebuttals that writers use to persuade the reader represent a rich vein of expert knowledge for summarization. However, recovering these presents substantial challenges as well: processing natural language leads to ambiguity; arguments are made implicitly instead of explicitly; and arguments are nested into complex structures. Agreement, disagreement, and conjecture are often expressed in highly scripted ways in scientific writing, and this feature makes these arguments recoverable by pattern-based search. Here, I present the PARROT software; it recognizes claims pertaining to scientific methods, cognition, discourse, negation, causation, and modality and uses discourse cues to combine these claims recursively into larger rhetorical structures to recover the shape of the arguments made in a scientific publication, which uses OpenDMAP patterns in combination with a Protégé ontology. PARROT outperforms an SVM classifier in identifying statements of support and conflict at the sentence level. Additionally, PARROT adapts to graphical representation of the arguments it finds, which makes it an valuable tool for summarizing the reasoning behind scientists' conclusions and identifying areas of consensus and contention.