University of Amsterdam at TREC 2014

This paper presents the University of Amsterdam’s participation in TREC 2014. For the Contextual Suggestion Track, we experimented with the use of anchor text representations in the language modeling framework, and base our runs either on full ClueWeb12 or the subset of touristic aggregators (e.g., tripadvisor) provided by the organizers of the track. We also look at the effectiveness of priors (in particular, PageRank) and ways of formulating the query based on the context. Our main finding is that the anchor text representation is effective for retrieving candidate attractions, and performs better than a standard document text index. A linear combination of both anchor and document text leads to further improvement. For the Web Track, we continued our experiment with the fusion of anchor text relative to the text-based baseline run. Our main finding is, again, that the combination of anchor and document text leads to improvement, and we demonstrate how the fusion weight can be used as a handle to tune the amount of risk acceptable for the risk sensitive evaluation.