Abstract
This paper describes R3 (Reading, Reasoning, and Reporting), our system for deep language understanding and extension of mechanistic models. The overall purpose of R3 is to read about biochemical signaling pathways from PubMed Central journal articles and integrate information into its model. Its initial background model of these biochemical pathways is derived from an imported Reactome model of biological pathways, events, complexes, and proteins. We describe some significant issues for semantic parsing in this domain and how R3 uses pre- and post-analysis reasoning to bridge the differences between the semantic information that can be derived from a text and the codified mechanistic information in the curated biomedical database. We also present extensions to relational structure mapping to detect corroboration between the semantic parse and the model and extend the model via analogical inference from the parse. We close with a description of empirical results with R3, including semantic parsing, model extension, grounding entity and event references, and modeling entity behavior using knowledge learned by reading.
