University College London, UK, m.sadrzadeh@ucl.ac.uk, kin.lo.20@ucl.ac.uk
(joint work with Samson Abramsky)
Natural language ambiguities give rise to probability distributions that can be studied by the mathematics of Quantum contextuality. These ambiguities can be catalogued into two groups: syntactic and semantic. Previously, in CQMB 2020, we found sheaf theoretic and CbD contextual examples of instances of the semantic group. In this work, we focus on syntax and model the anaphoric ambiguities of discourse. Anaphoric discourse ambiguities arise from the possibility of a pronoun in the current discourse referring to more than one noun of a previous discourse. Examples of these ambiguities have given rise to the famous Winograd Schema Challenge in Natural Language Processing [1,2], which improves on the Turing test. We construct a schema for these ambiguities in such a way that the instances exhibit logical quantum-like contextuality and conduct two experiments. In experiment (1), we take advantage of the neural word embedding engine BERT to instantiate the schemas to natural language and extract probability distributions for the instances. In experiment (2), we collect probabilities for an instance of the schema using the Amazon Turk crowd sourcing engine. The probabilities are analyzed in the CbD framework and in a recently developed signalling fraction of the sheaf-theoretic framework. Plenty of examples exhibited contextuality. Our hope is that these experiments will pave the way to use the quantum advantage in natural language processing.
[1] Winograd, Terry (1972). “Understanding Natural Language", Cognitive Psychology. 3 (1): 1-191.
[2] Levesque, Hector; Davis, Ernest; Morgenstern, Leora (2012).“ The Winograd Schema Challenge", Proceedings of the Thirteenth International Conference on Principles of Knowledge Representation and Reasoning, 552-561.