XML-IR: coverage as a part of relevance

Relevance is a multidimensional concept, not only consisting of linguistic-only properties but also enriched by various other relevance dimensions that are largely orthogonal to the topicality (i.e. content-based relevance) of a document. The question is how to capture such dimensions of relevance effectively in a retrieval model. In this paper we propose a model where we regard additional relevance dimensions independent (given a document instantiation). The independence assumption is made because it is very difficult to predict influence of relevance dimensions a-priori. The model also reflects our belief that modeling of additional knowledge with prior probabilities (in a probabilistic setting) is a counter-intuitive approach because of 1) the orthogonality of additional relevance dimensions and 2) the difficulty to reliably (re-)estimate dimension models, due to possible `noise' introduced by nondimension related priors. Also, relevance feedback needs to be able to handle multiple dimensions of relevance effectively. Feedback in the model is done with dimension-specific feedback sets. We can only report informally on the results of our model; based on the experimental scenarios performed, the model is appearing to perform very well, although quantitative assessments using an assessed collection are necessary to confirm this and draw further conclusions.