I had the honour to officiate as an opponent at the public defense of Johan Eklund’s doctoral dissertation With or without context: automatic text categorization using semantic kernels at the Swedish School of Library and Information Science in Borås. Johan used topological models to work out the scope of categorisation approaches, and then performed a number of experiments for text categorisation using variants of distributional approaches.
There were a number of methodological and formal questions to discuss.
- How are the topological models applicable in practice?
- Is a hierarchical knowledge model a reasonable model of human knowledge organisation?
- Johan used only single-term lexical features, aggressively filtered for efficiency. What might happen if we used non-lexical features?
- Johan used standard data sets. The data set quite obviously influenced the result and should be viewed as a parameter for this type of experimentation. How does the palette of categories influence the result? (This was not tested.)
- Is the F1-score as useless a metric as I myself consider it to be. (Johan did not convince me of its usefulness, but neither did I convince him of its uselessness).
Anyone interested in the mathematics of categorisation will be inspired by the models delineated in Johan’s dissertation!