R3E: Reading, Reasoning and Reporting Explanations
Abstract
SIFTs R3 (Reading, Reasoning, Remembering) project in DARPAs Big Mechanism program was focused on the problems of (1) developing an automated reading system for scientific articles capable of extracting semantic descriptions of relevant details about mechanisms, and (2) interpreting those descriptions in the context of an evolving background model of the larger mechanistic system being studied a process we call localization against a model. That is, we sought to develop methods to identify how descriptions of mechanisms and the entities involved in those mechanisms in texts relate to parts of a previously known but incomplete model, so that the model could be extended or revised with the new information described in those texts. Our reading system,R3/SPARSER was developed largely during the first two years of the project based on a previously existing system, SPARSER, and emphasized multi-sentential reading and co-reference. Since SPARSER was designed for semantic information extraction, it is extremely fast, indeed much faster than TRIPS/DRUM, and even slightly faster than REACH, the other two readers used by the CURE consortium to assemble its mechanism descriptions. SPARSER can process most articles it reads in well under 1 second on a MacBook Pro, and frequently under second. By June of 2017, we were processing and providing HMS with native INDRA/CURE formatted output for more than 17,000 open text articles. By the end of the project Sparser had read close to 1 million articles.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 20, 2018
- Accession Number
- AD1070821
Entities
People
- Mark H. Burstein
Organizations
- Smart Information Flow Technologies