Previous Ph.D. and M.Phil. Students

School of ECM
University of Surrey
Guildford, Surrey
GU2 5XH, UK

Tel: +44 (0)1483 259823
Fax: +44 (0)1483 876051

Previous Ph.D. and M.Phil. Students

Tracey Bale, B.Sc., Ph.D.

"Modular Connectionist Architectures and the Learning of Quantification Skills"
(Department of Computing, January 1999)

Abstract
Modular connectionist systems comprise autonomous, communicating modules, achieving a behaviour more complex than that of a single neural network. The component modules, possibly of different topologies, may operate under various learning algorithms. Some modular connectionist systems are constrained at the representational level, in that the connectivity of the modules is hard-wired by the modeller; others are constrained at an architectural level, in that the modeller explicitly allocates each module to a specific subtask. Our approach aims to minimise these constraints, thus reducing the bias possibly introduced by the modeller. This is achieved, in the first case, through the introduction of adaptive connection weights and, in the second, by the automatic allocation of modules to subtasks as part of the learning process. The efficacy of a minimally constrained system, with respect to representation and architecture, is demonstrated by a simulation of numerical development amongst children.

The modular connectionist system MASCOT (Modular Architecture for Subitising and Counting Over Time) is a dual-routed model simulating the quantification abilities of subitising and counting. A gating network learns to integrate the outputs of the two routes in determining the final output of the system. MASCOT simulates subitising through a numerosity detection system comprising modules with adaptive weights that self-organise over time. The effectiveness of MASCOT is demonstrated in that the distance effect and Fechner's law for numbers are seen to be consequences of this learning process. The automatic allocation of modules to subtasks is illustrated in a simulation of learning to count. Introducing feedback into one of two competing expert networks enables a mixture-of-experts model to perform decomposition of a task into static and temporal subtasks, and to allocate appropriate expert networks to those subtasks. MASCOT successfully performs decomposition of the counting task with a two-gated mixture-of-experts model and exhibits childlike counting errors.

Stephen Griffin, B.Sc., M.Phil.

"Exploiting Linguistic and Societal Metaphors for Knowledge Acquisition"
(Department of Mathematical and Computing Sciences, September 1996)

Abstract
Our inter-disciplinary research examines new approaches to knowledge acquisition through the exploitation of linguistic and societal metaphors. We argue that conventional knowledge acquisition relies too heavily on a psychological metaphor, and that this is insufficient in broad domains, where geographical and political issues make the expertise more socially situated, because it lacks input from the society in which the knowledge exists. We attempt to provide a methodology which captures this input by introducing a Domain Interface Group to support the knowledge engineer in his/her tasks. This presents a changing role for the knowledge engineer to primarily that of a group facilitator, and we suggest guidelines for brainstorming sessions to facilitate consensus decision making. We advocate the continued use of expert interviews, but suggest ways to improve their productivity. In particular, we attempt to alleviate reductive bias through the use and understanding of domain specific terminology and lexical semantics, during all domain communication and particularly during knowledge acquisition from text. We situate our work in the constructivist modelling paradigm and describe mediating representations which emphasise the importance of human comprehension of the model, for the knowledge engineer, the expert and the end user, above programming considerations. We have undertaken an evaluation of our methodology and an audit of a resulting paper knowledge base, and present the results in an attempt to prove the efficiency, effectiveness and accuracy of our approach.

Mohamed Benbrahim, Ingénieur d'Etat, Ph.D.

"Automatic Text Summarisation through Lexical Cohesion Analysis"
(Department of Mathematical and Computing Sciences, April 1996)

Abstract
A methodology for automatically summarising scientific texts is presented using the patterns of lexical cohesion found in such texts. Lexical cohesion is a type of cohesion whereby certain lexical features of the text connect sentences with each other in the text. An analysis of lexical cohesion in text, primarily by counting repetitions, synonyms and paraphrase, leads to the establishment of a network of sentences, some tightly bonded through lexical cohesion relations, some others having weak bonds or no bonds at all. The strength of connections in this cohesion network is used to identify key sentences in a text. Some sentences open key topics, some close topics, whilst others consolidate a given topic. Topic opening, closing, and consolidating, or central sentences, have different strengths and different connectivity patterns. A selection of these sentences can be construed as a summary of a given text. TELE-PATTAN (TExt and LExical cohesion PATTerns ANalysis), a system for summarising text automatically, extracts patterns of lexical cohesion in a text, categorises its sentences and subsequently produces summaries of the text on the basis of these patterns. Experiments were conducted with human subjects to evaluate the summaries. The results of this preliminary evaluation are encouraging.

John Wright, B.Sc., M.Sc., PhD.

"Connectionist Architectures for Language Disorder Simulation"
(Department of Mathematical and Computing Sciences, December 1995)

Abstract
Our interdisciplinary research focuses on the application of connectionist modelling techniques to the study of language disorders. In recent years, artificial neural network models of aphasia have enabled cognitive neuropsychologists to explore contemporary theories of language processing. Such work may, in the future, lead to the development of innovative strategies for the rehabilitation of brain-damaged patients. The aim of our work has been to analyse the modelling techniques employed in existing connectionist accounts of language disorders and, on the basis of our findings, to propose novel and computationally well-grounded architectures which may be used to explore cognitive neuropsychological theories.

The majority of connectionist language disorder models reported in the literature may be categorised as network-level models, consisting of a single homogeneous structure built from identical processing elements. We believe that in order to simulate more fully the complexity of human language processing, it may be necessary to move away from this approach, in favour of nervous system-level models, in which a number of network-level models are interconnected to form a modular connectionist architecture. The suitability of these architectures for language disorder simulation has been assessed through the construction of LISA: a Language Impairment Simulation Architecture. LISA comprises a number of linked connectionist networks which have been collectively trained to simulate object naming and word repetition. By lesioning one or more components of our modular system, it is possible to simulate the impaired language production of an aphasic patient. We present our attempts to simulate an acquired disorder of repetition, deep dysphasia and a progressive disorder, semantic dementia, using LISA. The results of our experiments are encouraging, and lead us to conclude that the cognitive neuropsychology community may indeed benefit from the use of modular connectionist architectures in the simulation of both progressive and acquired language disorders.

Paul Holmes-Higgin, B.Sc., PhD.

"Text Knowledge: The Quirk Experiments"
(Department of Mathematical and Computing Sciences, March 1995)

Abstract
Our research examines text knowledge: the knowledge encoded in text and the knowledge about a text. We approach text knowledge from different perspectives, describing the theories and techniques that have been applied to extracting, representing and deploying this knowledge, and propose some novel techniques that may enhance the understanding of text knowledge. These techniques include the concept of virtual corpus hierarchies, hybrid symbolic and connectionist representation and reasoning, text analysis and self-organising corpora. We present these techniques in a framework that embraces the different facets of text knowledge as a whole, be it corpus organisation and text identification, text analysis or knowledge representation and reasoning. This framework comprises three phases, that of organisation, analysis and evaluation of text, where a single text might be a complete work, a technical term, or even a single letter. The techniques proposed are demonstrated by implementations of computer systems and some experiments based on these implementations: the Quirk Experiments. Through these experiments we show how the highly interconnected nature of text knowledge can be reduced or abstracted for specific purposes, from a range of techniques based on explicit symbolic representations and self-organisation connectionist schemes.

Syed Sibte Raza Abidi, B.Eng., M.S., Ph.D.

"A Connectionist Simulation: Towards a Model of Child Language Development"
(Department of Mathematical and Computing Sciences, September 1994).

Abstract
Our research focuses on the connectionist simulation of child language development within the age group 9û24 months. We present a hybrid connectionist modelùACCLAIM (A Connectionist Child Language development and Imitation Model), comprising 'supervised' and 'unsupervised' learning connectionist networks that take into account the diverse nature of inputs to and outputs from a child learning his or her first language. The model is used to simulate the child's development of concepts, acquisition of words, ostensive naming of concepts, understanding of conceptual and semantic relations and the learning of word-order. The simulation produces child-like one-word and two-word sentences. The simulation of aspects of child language development are 'language informed', in that the data used in the simulation were taken from extant child language corpora. Theoretical underpinnings of our simulation were based on Jean Piaget's notions of cognitive development. The efficacy of hybrid connectionist models is demonstrated through the operationalisation of real child language data. The simulations indicate that connectionist networks can simulate developmental behaviour, and both connectionist and developmental psychology communities can benefit from such a contribution.

Abu Turab Alam, B.Sc., M.Sc., Ph.D.

"The Elicitation of Software Requirements: The Role of Natural Language Processing"
(Department of Mathematical and Computing Sciences, February 1991.)

Abstract
The engineering of a software system depends crucially upon the requirements specification of the system. The specification of requirements is a complex and interactive process involving an analyst and a client in a requirements definition activity. The principal medium for this activity is natural language, and we observe that special terms or jargon are used to abbreviate the communication between an analyst and the client. The information available to an analyst during this communication is inherently ambiguous and incomplete and often defined by the client without context.

We emphasise the all-persuasive use of natural language during the requirements definition activity. Natural language is used from the very start of a project and used throughout requirements acquisition, expression and analysis for software specification. Furthermore, a substantial amount of relevant information about the client's system is also available in natural language.

An analyst performs various tasks to elicit and understand software requirements. We identify a number of techniques to expedite these tasks for an analyst. These techniques have their origins in three different fields: knowledge engineering (for underlying the user's domain directly from its text); and natural language studies (schema for formalising the user's domain knowledge).

The main advantage of our framework is that it does not constrain (in the form of arbitrary method constructs) the thinking processes of an analyst. Instead, our framework emphasises the functional behaviour of natural language in a specific domain and allows the analyst to elicit and understand the requirements themselves in natural language.

Ian Gerald Wells, B.Sc., M.Sc., Ph.D.

"Knowledge Representation in Clinical Biochemistry"
(Department of Mathematical and Computing Sciences/Department of Biochemistry, March 1990.)

Abstract
Clinical decision-making is related to direct medical observation and treatment of patients, and includes both complex decisions, often made with inadequate information, together with routine judgements which require little rigorous analysis. An examination of the clinical decision-making process suggests that for complex problems the tendency is to reason backwards from a small number of hypotheses, but reasoning forwards from the available evidence can be more effective in other cases. Clinical errors are generally caused by defective knowledge rather than poor reasoning, and focus on the incorrect interpretation of a small number of significant clinical cues.

A computer-based development environment called PROSE has been designed and implemented in order to demonstrate a methodical approach to the analysis of a domain and to provide a tool for constructing knowledge-based systems which emulate clinical decision-making. PROSE has been developed within a computational framework where the domain primitives have been encoded as objects and relationships, and reasoning is effected by novel control features which allow alternative solutions to be explored. PROSE has been implemented in the logic programming language Prolog, and its performance evaluated using five selected problems in clinical biochemistry. PROSE has also been found to be of greater relevance through its successful application to a range of practical tasks in engineering and computer science.

Direct comments or questions to: c.jones

Last Modified by Gemma Stevens 15 July 1999.