Learning NLP Tasks with SARDSRN

James Hammerton, james.hammerton@ucd.ie

We present here a connectionist approach to learning NLP tasks which we hope will cope with large-scale tasks. The approach uses a network architecture known as SARDSRN:

http://www.cs.utexas.edu/users/nn/pages/publications/abstracts.html#mayberry.sardsrn.ps.gz

SARDSRN is based on the Simple Recurrent Network (SRN) and incorporates a self-organising map for sequences (known as SARDNET). SARDSRN has been shown to offer better performance than SRNs. To represent the output from the parser we intend to use a connectionist compositional representation of trees indicating the noun-phrases either through bracketing or annotation. In doing this we aim to see whether such techniques can scale-up to large tasks. We hope to present results of an exploratory experiment at the workshop.