Paul McKevitt

Integration of Natural Language and Vision Processing

The need for a formal language in which to express and reason about spatial concepts is of crucial importance in many areas of AI and visual systems. For the last five years, spatial reasoning research by the Qualitative Spatial Reasoning Group, University of Leeds, has centred on the development and application of such a language the RCC spatial logic. Below, we briefly describe the work of the group in this area.

From Chinese rooms to Irish rooms: new words on visions for language

Natural languages like English have been constrained in expressing perceptions like vision, sound and touch for years despite the efforts of Joyce (1922, 1939) and others. In situ, lexicons have been limited in their form and content. They have typically been structured in the form of sequences of natural language words with their content defined using flat symbolic descriptions in natural languages. In particular, we believe that todays dictionaries in general, and with respect to Artificial Intelligence (AI) systems in particular, are unnatural in the sense that they do not encode pictures for words just like we do in our heads. There is now a move towards integrated systems in Artificial Intelligence (AI) (see Dennett 1991; Mc Kevitt 1994a, 1994b) and that will cause a need for dire actions on lexical research in the form of integrated lexicons. We believe that lexicons must move towards a situation where natural language words are also defined in terms of spatial and visual structures. These spatial and visual structures will solve what have been two of the most prominent problems in the field of Natural Language Processing (NLP) for years: (1) Where are symbolic semantic primitive meanings in computer programs grounded? and (2) how come some words, typically in the defining vocabulary, in dictionaries have circular definitions so that words end up defining each other? We believe integrated lexicons will cause these two problems to go away and hence help solve Searles Chinese Room Problem and move more towards Irish Rooms of people like James Joyce.

Integrating vision processing and natural language processing with a clinical application

It is an interesting observation that humans are able to combine information from ostensibly disparate sources in order to better comprehend the world around them. For example, the sense information from the eyes and the ears are combined at every instant-tasks such as road crossing would be far more difficult without such integration. Such combinations are nonlinear in nature, so examining each sense in isolation will not shed light on how they combine within the cognitive frame. We are interested in the way humans combine their visual sense with natural language. This is an important area that appears to be little researched to date, despite the large amount of literature in both vision processing and natural language processing. In response to the nonlinear nature of the combination we proceed by synthesising the phenomenon rather than analysing it. To advance our work we focus upon a particular application that not only makes such a combination feasible but has a practical use in the field of medicine.

Review of "Proceedings of the 1988 Connectionist Models Summer School by David Touretzky, Geoffrey Hinton and Terrence Sejnowski", Morgan Kaufmann Publishers, 1989.

Here we have a set of papers by a number of researchers (students and faculty) at the time of the revival of connectionism, and some of the students have since made their names in the field (e.g., Bookman, Miikkulainen, and Regier). The proceedings is from the second Connectionist Models Summer School held at Carnegie Mellon University in 1988 and organized by Dave Touretzky with Geoffrey Hinton and Terrence Sejnowski as advisors.Here we have a set of papers by a number of researchers (students and faculty) at the time of the revival of connectionism, and some of the students have since made their names in the field (e.g., Bookman, Miikkulainen, and Regier). The proceedings is from the second Connectionist Models Summer School held at Carnegie Mellon University in 1988 and organized by Dave Touretzky with Geoffrey Hinton and Terrence Sejnowski as advisors.

Automatic reconstruction of vasculature

Two of the most difficult problems in Artificial Intelligence are processing visual scenes and processing natural languages. There has been a large amount of research in each of these fields but little on their integration. This is surprising given the potential importance of integrated systems, not only for understanding human cognition but also for the range of practical applications that will be enabled. We review previous work and provide an overview of our own work. We focus upon the medical application of reconstructing complicated cerebral blood vessel structures and associated pathologies from images and medical reports. This gives our work a clear and significant practical aim. We show how the ostensibly disparate technologies can be married using a single knowledge representation. Previous attempts at reconstruction have used images alone and no satisfactory solution exists. We believe that the synergy provided by integrating vision and natural language processing provides an information-rich environment that will enable progress toward an efficient and robust solution. Such an integration will have not only have important practical uses but also implications for Artificial Intelligence, Cognitive science, Philosophy, and Psychology.

Securing electronic data interchange

Abstract One of the growing areas of computer technology is electronic data interchange (EDI). Electronic data interchange is the electronic transfer of data from one agent to another. One of the most important problems with computer programs today is that they are not reliable. The reason for this is quite simple. Computer programs are built to solve particular problems and these problems can be quite large. By large we mean having lots of constraints, cases and rules as to what constitutes the problem. Real‐world problems are so complex that it is hard to be sure that when a program has been built it will never fail for an input that has not been taken into account. Hence, reliability of computer programs for EDI has serious implications because if a program fails to work then this can lead to problems with security. If the field of computer science concentrates on better techniques for increasing computer reliability then we will be able to build more reliable and secure programs for EDI.