Machine knowledge about entities and their relationships has been a
long-standing goal for AI researchers. Over the last 15 years, thousands
of public knowledge graphs have been au-tomatically constructed from
various web sources. They are crucial for use cases such as search
engines. Yet, existing web-scale knowledge graphs focus on collecting
positive state-ments, and store very little to no negatives. Due to
their incompleteness, the truth of absent information remains unknown,
which compromises the usability of the knowledge graph. In this
dissertation: First, I make the case for selective materialization of
salient negative statements in open-world knowledge graphs. Second, I
present our methods to automatically infer them from encyclopedic and
commonsense knowledge graphs, by locally inferring closed-world topics
from reference comparable entities. I then discuss our evaluation
fin-dings on metrics such as correctness and salience. Finally, I
conclude with open challenges and future opportunities.
Contact
Petra Schaaf
+49 681 9325 5000
--email hidden
passcode not visible
Petra Schaaf, 10/18/2023 10:05 -- Created document.