1.3 Some Methodological Considerations

To end this introductory chapter, I would like to point out a few methodological stumbling blocks which I have encountered while developing CLASPnet.

• Uncertainty: Neither the field of theoretical linguistics nor that of neural network research can be considered to be very stable. The controversies dividing linguists are well known, as are their many competing proposals for the explanation of most phenomena. Choosing which ones to accept for my own research has been a minor challenge in itself. Similarly, much about the behavior of neural network models is still unknown. While there are rules of thumb to guide one's choice of network architecture or to help find an appropriate learning algorithm, there are no lookup tables yet which provide ready-made solutions for different classes of problems.

• Originality: To the best of my knowledge, no one has ever described a connectionist model with a task very similar to the one which is presented in this dissertation. (Chapter 2 will summarize some related work, though.) Much of the field of connectionist linguistics is still terra incognita, of course, but there have been moments when results from independently performed studies would have been very welcome -- in a field without established foundations, it is seldom self-evident how to proceed from one stage to the next: there is an overabundance of factors to be investigated, but only finite time.

• Ambiguity: As has been mentioned above, CLASPnet tries to bring connectionist models closer to theoretical linguistics. Both fields, however, have their own methodological assumptions, and their own standards by which to judge research. Having mixed elements from both sides, I am in the uncomfortable position of knowing perfectly well that I have not fully lived up to the standards of either field. In theory, this shouldn't be a problem as interdisciplinary explorations is what Cognitive Science is all about (Mandler 1984).

• Scarcity of Utilities: unsurprisingly, not that many software tools were readily available for the task I had in mind for CLASPnet. The one which stands out in this respect is the X Window System package Stuttgart Neural Network Simulator (SNNS), of which I have used versions 3.3 and 4.1 for running my own network simulations. (Note 2) The other tool I have made extensive use of is the programming language Perl -- it has proved invaluable both for easily generating new training and test corpora, and for analyzing the results of the network simulations. (Note 3) While writing the programs for CLASPnet, I have tried to make the code as reusable as possible. Hence, I hope that parts may prove useful for other models. (The programs are listed in the Appendices and can also be downloaded. (Note 4))

  
  

  Note 2
  The SNNS software package runs on most UNIX architectures and is available via the Internet from ftp://ftp.informatik.uni-stuttgart.de/pub/SNNS/.

  Note 3
  Perl is available from many sites on the Internet, but http://www.perl.com/ provides a starting point.

  Note 4
  The URL for this dissertation is http://snow.ucsd.edu/~ezra/msc/.