Compositional Distributional Cognition

Lead Research Organisation: University of Oxford
Department Name: Computer Science


This project falls within the EPSRC "Information and Communication Technologies" research area.

Our project investigates cognitive processes from a computational standpoint covering aspects of cognition such as neural architecture, knowledge representation, and computation. We propose that the
interaction of the quantuminspired compositional distributional (DisCo) model of meaning [Clark & Pulman, 2007], [Clark et al., 2008], [Coecke et al., 2011] and the integrated cognitive architectures of
[Gärdenfors, 2004], [Smolensky & Legendre, 2005], [Eliasmith, 2013] is a fruitful area of research for cognitive science. Building on existing DisCo models of meaning, we aim to present a robust and
expressive novel unified cognitive architecture. The higherand lowerlevel substrates of our architecture shall interface via an instruction set of basic functions to manipulate cognitive representations. Paramount to our research is an examination of the inbuilt quantum properties of our model, as well as an understanding of how to evaluate 'meaningfulness'. Looking ahead, a notion of 'meaningfulness' shall allow us to consider potential applications with a focus on mending nonoptimal semantic structures. Once we have set up the theoretical foundations of our compositional distributional model of cognition (DisCog), a practical implementation of some sort shall complete our research.


10 25 50
publication icon
Al-Mehairi Y (2017) Quantum Interaction

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/N509711/1 01/10/2016 30/09/2021
1732662 Studentship EP/N509711/1 07/10/2016 31/10/2020 Khalid Yaared Al-Mehairi
Description We have improved on Smolensky's model of Connectionist and Symbolic Cognitive architecture with a focus on Natural Language Processing. Using Coecke and Pullman's Distributional Models of Meaning, we have made steps to integrate the two seemingly opposed models to produce a way of composing words so as not to witness an explosion in the size of their meaning space. In essence, the grammatical structure of an instance of language - i.e. composition rules a la Lambek - are functorially mapped to algorithms within the distributional meaning space of words which then govern the "filler/role binding procedure" from Smolensky. The result is that grammar and meaning work hand-in-hand as they should. This allows for two grammatically valid strings of words with different structures to exist within the same meaning space and therefore be compared etc.

The next step is to abstract further and consider nouns as first-class citizens of a language. We draw inspiration from Dynamic Epistemic Logic and model nouns - just as agents or notions of reality - as processes encoding information in an event-update (Kripke-like) model where events can be of the epistemic (i.e. belief) or ontic (i.e. factual) class. With this approach, like Benthem, we take the broad view that logic is "the science of information processing in evolving contexts". With respect to language, we can think of nouns as the main actors within a literary universe which are then updated or characterized via static description-like sentences. In essence, we are dealing with threads of knowledge that evolve and exchange information. So to give a general account of the dynamics of knowledge and belief we are working on a framework that provides (i) a recipe to represent a state of knowledge (or belief) of multiple agents and (ii) a procedure to model perception, information exchange and action. What will then be most interesting is the study of the relationships between different flavours of dynamic logics and how they interact and compose - for example language and vision. At present, we are finding the category of relations to be fruitful for the modelling of vanilla Dynamic Epistemic Logic whereby we model ambiguity via disjunction and have natural interpretations for epistemic phenomena such as distributed knowledge, group knowledge and common knowledge. Of course, the next immediate step is to then progress to dynamic logics with richer representations of information.
Exploitation Route Work on distributional models of meaning and dynamic logics should be most evidently fruitful for linguists and logicians as outlined above.
Sectors Digital/Communication/Information Technologies (including Software)