ILPS at TREC 2017 Common Core Track

Abstract: The TREC 2017 Common Core Track aimed at gathering a diverse set of participating runs and building a new test collection using advanced pooling methods. In this paper, we describe the participation of the IlpsUvA team at the TREC 2017 Common Core Track. We submitted runs created using two methods to the track: (1) BOIR uses Bayesian optimization to automatically optimize retrieval model hyperparameters. (2) NVSM is a latent vector space model where representations of documents and query terms are learned from scratch in an unsupervised manner. We find that BOIR is able to optimize hyperparameters as to find a system that performs competitively amongst track participants. NVSM provides rankings that are diverse, as it was amongst the top automated unsupervised runs that provided the most unique relevant documents.