Machine Common Sense (MCS) Multi-modal Open World Grounded Learning and Inference (MOWGLI)

DARPA is the research and development office for the U.S. Department of Defense.  It maintains the technological superiority of the U.S. military and prevents technological surprise from harming our national security. TWC has been funded by DARPA for numerous projects, including this one. TWC collaborates to the project in two work streams:

• The DARPA MCS MOWGLI Extract-Transform-Load sub-project consists of a series of extract-transform-load (ETL) pipelines for adding common sense knowledge triples to the MOWGLI Common Sense Knowledge Graph (CSKG). The CSKG is used by downstream applications such as question answering systems and knowledge graph browsers. ConceptNet serves as the core of the CSKG, and other sources such as Wikidata are linked to it.
• Despite the incredible advancements by language model-based implementations in achieving close to human performance in many commonsense tasks, challenges remain in assessing in thoroughness their reasoning capabilities. TWC is tackling this problem by proposing novel approaches for the creation of benchmarks that are theoretically grounded in formal theories of common sense. Evaluations with such benchmarks support the assessment of common-sense reasoning completeness and soundness.