Kirsty Kitto | Research · Language

I am currently seeking postgraduate students to work in this area!
See my Research Students page for more details.

Language & Memory in Context

Consider the word "boxer"... are you thinking of a dog? or of someone who punches people? Depending upon the context in which this word is used you will be biased towards assuming one meaning for it. For example, if I talked about a boxer at a boxing ring then you would be unlikely to assume that I was talking about a dog.. unless I was pointing to one at the time. Humans are incredibly good at interpreting such ambiguous words (although we often get this wrong, we can usually backtrack and determine the intended meaning). How do we do it?

This contextuality of language is very difficult to model. Does a person have a different cognitve state for every interpretation? Or is something more sophisticated going on? One possible way in which to model such behaviour is to treat the context of the cognitive state separately, using a basis state in a Hilbert space.

Indicative Papers

  1. P.D. Bruza, K. Kitto, B. Ramm, L. Sitbon. A probabilistic framework for analyzing the compositionality of conceptual combinations. Under Review, available at arXiv
  2. Nelson, D., Kitto, K., Galea, D., McEvoy, C., Bruza, P.D. (In Press). How Activation, Entanglement, and Searching a Semantic Network Contribute to Event Memory, Memory and Cognition.
  3. K. Kitto, P. Bruza. (2012) Tests and models of non-compositional concepts. In Miyake, N, Peebles, D., & Cooper, R.P. (Eds.) Proceedings of the 34th Annual Conference of the Cognitive Science Society, Cognitive Science Society, Sapporo, Japan, pp1792-1797.
  4. P.D. Bruza, K. Kitto, D. Nelson and C. McEvoy, (2009). Is there something quantum-like about the human mental lexicon?, Journal of Mathematical Psychology, 53:362-377, doi:10.1016/ (Preprint here.)

Quantum Models of the Human Mental Lexicon

a quantum word

In the figure above, the word |A> is modelled with respect to two different bases, or contexts. In these different contexts it has a different probability of recall, which means that the quantum formalism provides a very natural way in which to incorporate the natural ambiguity of language, and its dependence upon context. See Is there something quantum-like about the human mental lexicon? for the full model.

When is Language Noncompositional?

Given the above simple contextual model, we can naturally ask how the interpretation that we apply to one word might affect another. Thus, is language compositional, or is it somehow non-separable? We make use of primes to bias the interpretation given to a biambiguous but nonlexicalised conceptual combination (e.g. a "star charge"), asking subjects to invent a meaning for it in the light of this bias (e.g. "a battery that has been charged by a star", or "a rampaging herd of famous people"). We then make use of Bell-style inequalities, to determine if that conceptual combination should be considered separable, or not, within the context that was defined by the primes used.

You can get a pretty good idea about what these experiments entail by looking at an early online version of one, or the new paper A probabilistic framework for analyzing the compositionality of conceptual combinations has details about the most recent (and rigourous) experiment, as well as an analysis of the data that we gathered with that protocol.

Semantic Networks & Memory

Finally, it is reasonable to ask how the entire Mental Lexicon behaves, and even how it arises in a person as they grow from a baby to a mature adult. The Semantic Network of humans is complex, it exhibits small world behaviour, and has some very odd properties, that make them very difficult to model its behaviour using the standard Spreading Activation models. In particular, Doug Nelson & Cathy McEvoy collected decades of word association data, which can be used to map out the semantic network (at least for University of South Florida psychology students :) a more representative data set is now being collected by Leuven University). For example, in extra list cueing experiments, they found that even associates that did not return to a target word could still affect the probability that word would be recalled. This is in stark contrast to the spreading activation models, which neglect associates that do not connect to the target. They proposed an ad hoc Spooky Activation at a Distance model of word activation (based upon quantum phenomenon of nonlocality), and we have since helped to refine this with a more detailed model.

See How Activation, Entanglement, and Searching a Semantic Network Contribute to Event Memory for a comprehensive review of Doug Nelson & Cathy McEvoy's decades of data collection & analysis in this area, as well as our most recent attempts to fit a quantum model to this data.