NTCIR-4 Patent Retrieval Experiments at RICOH

Focusing on the document structure of patent, two search databases AC and WH were built and compared in retrieval performance, where AC includes only abstract and claim sections in patent and WH includes the whole patent texts. Moreover, we attempted to combine search results for the two databases to improve retrieval performance. Another point of our experiments is cross-lingual patent retrieval using a large and high-quality parallel corpus. The query submitted against the parallel database was partially translated and expanded using the same mechanism for pseudo-relevance feedback.