PUTTING MDS AND LSI TO WORK

Creative Approaches

Much of the work in applied MDS has come from the fields of advertising and cognitive psychology (where it is also known as perceptual mapping). Researchers in both fields use the technique to transform questionnaires about relative preferences and similarities into a visual representation using the scaling techniques we have outlined. These techniques do not appear to have been applied to linguistic data until relatively recently.

This illustrates a common theme in latent semantic research - combining familiar techniques from different disciplines in a novel way to tackle problems in data retrieval. This kind of creative juxtaposition is one of the things that makes LSI interesting to work on, and levels the playing field between major research institutions and liberal arts colleges. One does not need an enormous supercomputer or advanced mathematical knowledge to do interesting work with these techniques. In fact, because LSI research draws on pure and applied mathematics, linguistics, computer science, psychology, information retrieval, and the social sciences, what really matters is breadth of knowledge. There are likely to be connections further afield that remain to be discovered.

With this eclectic background in mind, here are some potential applications of semantic indexing coupled with MDS data visualization:

1. Archive Management Tools:
   

   We already mentioned the potential use of LSI as an archivist's assistant, using LSI to highlight content patterns in a data collection, and more traditional taxonomies to formalize and heighten those patterns. One intuitive method for creating such tools is to display data visually using MDS, and allow for human feedback. An interactive program using multi-dimensional scaling would allow an archivist to graphically manipulate data, draw boundaries between clusters, examine content relationships and add classifiers using an intuitive, click-and-drag type interface. What's more, different expert users would be able to use MDS to generate their own personal view of a data set, and then reconcile or combine those views.
   

2. Concept Maps:
   

   Concept maps take this notion of interactivity and classification further, letting users manipulate and edit LSI-generated views of a data collection to produce a spatial map of topics and concepts. By drawing connections between items and moving them around, users can create their own view of a data collection. These views can be 'untangled' using mathematical techniques to create clear, visually direct concept maps. These maps can be shared, combined, and compared with others, making a unique pedagogical or research tool.

3. Bioinformatics:
   

   The same LSI techniques we use to find similarities in language have enormous potential in the field of bioinformatics. Both DNA and protein molecules consist of long strings of biochemical 'words'. Finding and understanding patterns in those words is one of the major research problems in modern biology. Using the tools we describe would make it possible to detect and visualize such patterns, and conduct important basic research in this nascent field.