Reasoning about Plausibility for the Winograd Schema Challenge. Golovin, D. Master's thesis, Department of Computer Science, RWTH Aachen University, April, 2017.
abstract   bibtex   
The Winograd Schema Challenge (WSC) has been proposed as an alternative to the Turing Test for measuring a machine's intelligence by letting it solve pronoun resolution problems that cannot be tackled by statistical analysis alone, but require commonsense, everyday background knowledge and some form of deeper "understanding" of the question. WSCs are thus hard to solve for machines, but easy for humans. Many solutions so far are based on machine learning and natural language processing, and achieve results that are hardly better than guessing. Moreover, most knowledge-based approaches to the WSC have been purely theoretical. The goal of this thesis was to develop and implement a knowledge-based WSC solver. In particular, a logic of conditional beliefs called BO is employed that is capable of dealing with incomplete or even inconsistent information (which commonsense knowledge often is). It does so by formalising the observation that humans often reason by picturing different contingencies of what the world could be like, and then choose to believe what is thought to be most plausible. Relevant commonsense background information furthermore is obtained from the ConceptNet semantic network and translated into BO, and processed by the Limbo reasoner.
@thesis{Golovin2017,
  author      = {Denis Golovin},
  title       = {Reasoning about Plausibility for the {Winograd}
                  Schema Challenge},
  school      = {Department of Computer Science, RWTH Aachen
                  University},
  year        = {2017},
  month       = apr,
  advisor     = {Cla{\ss}en, Jens},
  abstract    = {The Winograd Schema Challenge (WSC) has been proposed
                  as an alternative to the Turing Test for measuring a
                  machine's intelligence by letting it solve pronoun
                  resolution problems that cannot be tackled by
                  statistical analysis alone, but require commonsense,
                  everyday background knowledge and some form of
                  deeper "understanding" of the question. WSCs are
                  thus hard to solve for machines, but easy for
                  humans.  Many solutions so far are based on machine
                  learning and natural language processing, and
                  achieve results that are hardly better than
                  guessing. Moreover, most knowledge-based approaches
                  to the WSC have been purely theoretical. The goal of
                  this thesis was to develop and implement a
                  knowledge-based WSC solver. In particular, a logic
                  of conditional beliefs called BO is employed that is
                  capable of dealing with incomplete or even
                  inconsistent information (which commonsense
                  knowledge often is). It does so by formalising the
                  observation that humans often reason by picturing
                  different contingencies of what the world could be
                  like, and then choose to believe what is thought to
                  be most plausible. Relevant commonsense background
                  information furthermore is obtained from the
                  ConceptNet semantic network and translated into BO,
                  and processed by the Limbo reasoner.},
  type        = {mathesis}
}

Downloads: 0