Report on the CLEF-IP 2012 Experiments: Search of Topically Organized Patents

This technical report presents the work which has been carried out using Distributed Information Retrieval methods for federated search of patent documents for the passage retrieval starting from claims (patentability or novelty search) task. Patent documents produced worldwide have manuallyassigned classification codes which in our work are used to cluster, distribute and index patents through hundreds or thousands of sub-collections. We tested different combinations of source selection (CORI, BordaFuse, Reciprocal Rank) and results merging algorithms (SSL, CORI). We also tested different combinations of the number of collections requested and documents retrieved from each collection. One of the aims of the experiments was to test older DIR methods that characterize different collections using collection statistics like term frequencies and how they perform in patent search. Also to experiment with newer DIR methods which focus on explicitly estimating the number of relevant documents in each collection and usually attain improvements in precision over previous approaches, but their recall is usually lower. However, the most important aim was to examine how DIR methods will perform if patents are topically organized using their IPC and if DIR methods can approximate the performance of a centralized index approach. We submitted 8 runs. According to PRES @100 our best DIR approach ranked 7 across 31 submitted results, however our best DIR (not submitted) run outperforms all submitted runs.