Inferential Commonsense Knowledge from Text

To enable human-level artificial intelligence, machines must have access to the same kind of commonsense knowledge about the world that people have. The best source of such knowledge is text -- learning by reading. Implicit in linguistic discourse is information about what people assume to be possible or expect to happen. From these references, I obtain an extensive collection of semantically underspecified 'factoids' -- simple predications and conditional rules. Using lexical-semantic resources and corpus frequencies, these factoids are generalized and partially disambiguated to form a collection of reasonable commonsense knowledge. Together with lexical axioms from the interpretation of WordNet, these probabilistic logical inference rules allow a reasoner to draw conclusions about everyday situations as might be encountered while reading a story or conversing with a person.