Abstract
Prior art retrieval refers to the process of identifying relevant prior arts for a given patent (or patent application). Prior art retrieval task is mainly used to support patent validity search or patentability search. Patent applicants often use peculiar or abstract terms to enlarge the legal monopoly scope of patents, which make the prior art retrieval a difficult task. However, existing techniques for prior art retrieval encounter some limitations. In response, we propose the citation-relatedness-based relevance feedback prior art retrieval (CRF-PAR) technique, which incorporates citation information of patents as knowledge source for performing relevance feedback. A hybrid similarity measure which combines text-based and citation-based similarities between patents is proposed to select top-ranked patents for expanding the original query patent. The expanded query patent is then applied to perform prior art retrieval. For empirical evaluation purpose, we collect 14,928 patents documents from the United States Patent and Trademark Office (USPTO) website and conduct a series of experiments using a traditional text-based prior art retrieval and a traditional relevance-feedback-based prior art retrieval as the performance benchmarks. Our evaluation results suggest that our proposed technique outperforms its benchmark techniques, measured by the top-m recall rate.