Deriving General World Knowledge from Texts by Abstraction of Logical Forms

University Of Rochester, Rochester NY

Investigators

Abstract

The goal of this project is to break through the knowledge acquisition bottleneck in the effort to endow artificial intelligence (AI) systems with common sense. The novel idea is to abstract general propositions about what is possible in the world from the specific assertions made in miscellaneous texts (including fiction). This will be accomplished by compositional interpretation of parse trees, with heuristics to decide on verb argument structure and with concomitant simplification and abstraction of logical forms. The research will employ state-of-the-art parsers to derive general knowledge from unannotated texts. Semantic classifications of words will strengthen and disambiguate the knowledge derived from text. There will be an emphasis on the extraction of causal knowledge, since this is so central to commonsense understanding of the world. This will be made possible by the use of event variables in logical forms and fuller interpretation of adverbial modification. Research will also investigate the representation of general (and often uncertain or even inconsistent) knowledge and methods of using such knowledge for inference. One result will be a demonstration system that is able to answer simple questions eliciting general knowledge. The broader impact of this work includes enhanced learning and active research participation by doctoral candidates as well as undergraduates. By helping to break through the knowledge acquisition bottleneck, this research will pave the way for building more user-friendly AI systems for a broad range of potential applications, such as personal agents that mediate between a user and more specialized software, medical advice systems, tutoring systems, and computer games. It could also serve to bootstrap natural language understanding systems towards more human-like understanding.