Pseudo-Relevance Feedback for CLEF-CHiC Adhoc

In this paper we will briey describe the approaches taken by the Cheshire (Berkeley) Group for the CLEF CHiC Adhoc tasks (Mono- lingual, Bilingual and Multilingual retrieval for English, French and Ger- man). We used multiple translations of the topics for searching each of the CHiC Europeana English, French and German subcollections, employing Google Translate as our translation system. In addition we combined the original topics for various multilingual runs. Once again this year our approach was to use probabilistic text retrieval based on logistic regression and incorporating pseudo relevance feedback for all of the runs. The results overall, when viewed using the multilingual qrels based on the entire set of languages for the CHiC collection, were not good, while the individual monolingual runs using only the collection-specic qrels, appear to have performed reasonably well. There is some question about the qrels for the the entire multilingual collection, since there appear to be no relevant documents at all from the English collection.