UMass Machine Learning and Friends Lunch | Main / Building A Semantic Representation Of Personal Information

A 'semantic desktop' has been suggested as a way to organize personal information for improved accessibility. However, existing systems have focused on building a complex ontology and require the user to make nontrivial amounts of annotation, which has been a critical barrier for their adoption. In this talk, I suggest a simple model of semantic representation for personal information that is composed of documents, concepts and the weighted links between them. The proposed system is easy to understand, and requires minimal user intervention in that it can learn to predict links between concepts and documents using click feedback from a user. Our experiment shows that it is possible to make effective predictions using small amounts of feedback.