University College London, UK, d.wang@cs.ucl.ac.uk, m.sadrzadeh@ucl.ac.uk
(joint work with Samson Abramsky)
Natural language ambiguities give rise to probability distributions and these can be studied in the mathematical frameworks of Quantum contextuality. Such ambiguities are manifold, but can be catalogued into two broad groups: syntactic and semantic. In [1], sheaf theoretic and CbD contextual examples of instances of the semantic group were discussed. In this work, we focus on syntax and model parsing ambiguities via the so-called garden-path effect. Garden-path sentences first appeared in [2] and sparked a research program on the biological and cognitive bases of human interaction patterns. The 80’s and 90’s, saw a wave of Psycholinguistic experiments with the aim of classifying these ambiguities and measuring their cognitive dissonance levels. Recently, Natural Language Processing researchers have found correlations between the statistics learnt by deep neural network language models and human reading times [3]. They failed, however, when it came to predicting degrees of difficulty and reanalysis. We show that garden-path sentences can be modelled and analysed by the signaling and causal fractions from the sheaf theoretic framework for contextuality, and by the degree of direct influences of the Contextuality-by-Default (CbD) model. We consider incrementally increasing subphrases as contexts, and study their local grammatical structures. The consistency of these local structures from one context to the other is calculated using the signaling and causal fractions, their degree of direct influences on one another is computed via CbD’s ∆. Our goal is to use neural language models statistics and relate the predictions of the contextuality models to human reading times, as well as to the levels of difficulty of garden path sentences.
[1] Wang D., Sadrzadeh M., Abramsky S., Cervantes V. (2021). “Analysing Ambiguous Nouns and Verbs with Quantum Contextuality Tools”. Journal of Cognitive Science, 22 (3) 391-420.
[2] Bever, T. (1970). “The Cognitive Basis for Linguistic Structures”.
[3] Van Schijndel M., Linzen, T. (2021). “Single-stage prediction models do not explain the magnitude of syntactic disambiguation difficulty”. Cognitive science, 45(6), e12988.