Next: IR issues Up: Adaptive information retrieval Previous: Adaptive information retrieval

Programmed and learned knowledge

As suggested above, a defining tension in my research is generated by my desire to enjoy the advantages of semantically well-understood representational schemes of traditional AI, while simultaneously demanding the ability to modify the representation in the face of experience. Much of my research on the relative strengths and weaknesses of programmed and learned knowledge has been in the context of the IR problem, but I have also analyzed the problem with respect to classifier systems and more generally [10][25].

Sometimes this tension takes the form of constraints imposed on the learning system. For example the special features of the IR problem make novel demands on connectionist learning that have lead me to consider unusual network topologies, dynamics and learning rules [18]. Conversely, a central contribution of the SCALIR system is a mechanism allowing the logical behavior we desire of programmed knowledge (e.g., an indexing thesaurus) encoded in a semantic network to be reconciled with the probabilistic behavior we rely upon in an overlayed connectionist network [47].

The relative merits of traditional symbolic approaches versus those of ``sub-symbolic'' systems for AI systems and cognitive models is an extremely active debate [20]. Because there are often advantages to combinations of the two approaches, I have assisted a previous student, A. Wilson, with the development of CONNCERT, a programming language for connectionist networks [51]. We have also argued that hybrids of symbolic and sub-symbolic technologies are not simply efficacious engineering solutions but demanded by a science of cognition [44].

A second important example of the way in which these two types of programmed and learned knowledge can profitably interact is the use of both manually constructed and automatically derived indexing structures for free-text searching. Carefully constructed symbolic representations, such as the West Keynumber classification of the legal literature, or Encyclopædia Britannica's Propædia classification, become excellent benchmarks against which statistically-derived sub-symbolic representations may be compared [49]. Conversely, these automatic techniques can provide an excellent source of ``quality assurance'' checking on the manual procedures. The number of corpora which have benefitted from such careful, manual attention are few, however. Consequently the ultimate goal of this research is to extrapolate our methods from their use on these special collections to the vast majority of others which can not afford such attention.

Next: IR issues Up: Adaptive information retrieval Previous: Adaptive information retrieval