Machine Learning for the Semantic Web Putting the user into the cycle
Luca Gilardoni, Christian Biasuzzi, Massimo Ferraro, Roberto Fonti, Piercarlo Slavazza · 2005
The main driver behind the assumed need for machine learning techniques for the semantic web is that the web is so huge that there is no hope to attach semantic descriptors without at least partial support from tools able to work autonomously, and these in turns cannot be developed with deterministic pre-programmed algorithms. While the assumption is surely correct I believe learning must be supported by an overall approach to annotation, training set building and cross validating overall results which should enable a better integration of end users within the cycle. Such an involvement is the only guarantee that machine derived results are really checked, machine learning tools are applied to real world training set, and training sets can seamlessly be built and grown to provide always better data sets and hence better results. Pushing towards such an integrated approach however pose strong constraints on how supporting tools must be designed and integrated, and on the supporting work methodologies, both for end users, which needs to be exposed to more complex albeit powerful tools, and for people working on ontologies and on learning algorithms and tools, which must provide tools much more integrated – or at least integrable – with end users working environments – and must develop methodologies to better support evolution.