A multi-collection latent topic model for federated search

Collection selection is a crucial function, central to the effectiveness and
efficiency of a federated information retrieval system. A variety of solutions have been proposed for collection selection adapting proven techniques used in centralised retrieval. This paper defines a new approach to collection selection that models the topical distribution in each collection. We describe an extended version of latent Dirichlet allocation that uses a hierarchical hyperprior to enable the different topical distributions found in each collection to be modelled. Under the model, resources are ranked based on the topical relationship between query and collection. By modelling collections in a low dimensional topic space, we can implicitly smooth their term-based characterisation with appropriate
terms from topically related samples, thereby dealing with the problem of missing
vocabulary within the samples. An important advantage of adopting this hierarchical model over current approaches is that the model generalises well to unseen documents given small samples of each collection. The latent structure of each collection can therefore be estimated well despite imperfect information for each collection such as sampled documents obtained through query-based sampling. Experiments demonstrate that this new, fully integrated topical model is more robust than current state of the art collection selection algorithms.