Automatic Prior Art Searching and Patent Encoding at CLEF-IP '10.
Douglas Teodoro, Julien Gobeill, Emilie Pasche, Dina Vishnyakova, Patrick Ruch, Christian Lovis · 2010
In the intellectual property field two tasks are of high relevance: prior art searching and patent classification. Prior art search is fundamental for many strategic issues such as patent granting, freedom to operate and opposition. Accurate classification of patent documents according to the IPC code system is vital for the interoperability between different patent offices and for the prior art search task involved in a patent application procedure. In this paper, we report our experiments with prior art searching and patent classification in the context of CLEF-IP ’10 evaluation track. In the Prior Art Candidates search task, we strongly improved our last year’s model based on our experiments on training data (MAP 0.22), but official results, alas, were far from the expected ones (MAP 0.14). Regarding multilingual issues, our simple Google translator strategy achieved a 10% improvement. Nevertheless we think that the multilingual aspects in CLEF-IP’10 were less clear than for CLEF-IP’09. Finally, exploiting applicant’s citations led to a 30% improvement, but their visibility depends on who (the applicant or the examiner) performs the prior art search in the simulated task. This issue needs clarification by the organizers for the forthcoming campaigns. In the Classification task, we apply the k-NN algorithm in the categorisation process and explore different retrieval models, ranking combinations and languages features in order to enhance our results. Using multi-collection in the classification process improved the results by 2%. Both the prior art search and classification systems are in the top three rank among the participants.