Machine Reading and Reasoning Technology
Abstract
(U) The Machine Reading and Reasoning Technology program (previously funded in PE 0602304E, Project COG-02) will develop enabling technologies to acquire, integrate, and use high performance reasoning strategies in knowledge-rich domains. Such technologies will provide DoD decision makers with rapid, relevant knowledge from a broad spectrum of sources that may be dynamic and/or inconsistent. To address the significant challenges of context, temporal information, complex belief structures, and uncertainty, new capabilities are needed to extract key information and metadata, and to exploit these via context-capable search and inference. Cognitive inference has traditionally emphasized deduction via theorem-proving and induction via statistical techniques, but abduction —also known as “inference to the best explanation”— is also likely to play a large role. DoD systems sense, capture, and store information in the form of text, audio, imagery, and video, and so advanced machine reasoning capabilities must extract knowledge from, and reason about, all types of multimedia data. New visual faculties will enable cognitive systems to learn from visual experience, to reason about action in the real world, and to apply that knowledge in a broad range of domains to solve problems in tactical and security contexts. (U) Machine Reading addresses the prohibitive cost of handcrafting information by replacing the expert, and associated knowledge engineer, with un-supervised or self-supervised learning systems, systems that “read” natural text and insert it into AI knowledge bases, i.e. data stores especially encoded to support subsequent machine reasoning. Machine Reading requires the integration of multiple technologies: natural language processing must be used to transform the text into candidate internal representations, and knowledge representation and reasoning techniques must be used to test this new information to determine how it is to be integrated into the system’s evolving models so that it can be used for effective problem solving.
Document Details
- Document Type
- Accomplishment
- Publication Date
- Oct 01, 2011
- Source ID
- a8c26f7d47b0645a1044454071d3ec6c
Related Documents
- Root: MACHINE INTELLIGENCE