UNIVERSITY OF HERTFORDSHIRE

    COMPUTER SCIENCE RESEARCH COLLOQUIUM


                presents


   "Categorisation and the testing of categories:
            A neural computing approach"

                speaker:


          Prof. Khurshid Ahmad,
         Department of Computing,
        University of Surrey, U.K.



        27 November 2002 (Wednesday)

          Lecture Theatre E351
            Hatfield Campus
                4 - 5 pm



 Coffee/tea and biscuits will be available.


        Everyone is Welcome to Attend
            [Space Permitting]


Abstract:

The creation of categories interests computing scientists 
theoretically and experimentally.  Indeed, this vocation 
is much beloved of philosophers, biologists, linguists and 
kindred folk.  Once the categories are created then
various tests are administered to the output to ascertain 
the goodness of the hypothesis used to create the categories 
in the first place.  Neural computing systems, supervised and 
unsupervised, are used to create categories for a range of 
input data.  In this talk, I will discuss a new use of 
unsupervised neural networks for categorising news reports.  
We use the so-called self-organising feature maps for creating 
categories into which each of a set of the news stories can 
be categorised; the maps were created using the Surrey 
Artificial Neural Network System (SANNC).  The effectiveness 
of the categories is evaluated using a statistical test.  The
program created categories appear to overlap well with the 
categories of a human categoriser (sub-editor).  The category 
map is then used to test the effectiveness of a news 
summarisation program.  The summary of all news items in the 
training set is produced by one of our program SummariserPort,
and its representative vector categorised by the neural 
network trained on a set of full texts.  Each summary input 
is then allocated a category by SANNC: if the summary occupies 
the same space on the map as its source full text then one 
can argue that the summary was an accurate in so far as it 
was given the same category.  Typically, a human expert tests 
the output of a summariser; the test may sometimes have 
subjectivity that characterises almost all human evaluation.  
The use of a category map, produced by a neural computing 
program, in my opinion, reduces the bias.


--
Hertfordshire Computer Science Research Colloquium Abstracts
On-line: http://homepages.feis.herts.ac.uk/~nehaniv/colloq/