Formulating Simple Structured Queries using Temporal and Distributional Cues in Patents
Patent prior art retrieval aims to find related publications, especially patents, which may invalidate the patent. The task exhibits its own characteristic because of the possible use of a whole patent as a query. This work focuses on the use of date fields and content fields of the query patent to formulate effective structured queries. Retrieval is performed on the collection of patents which also share the same structure as the query patent, mainly priority dates, application date, publication date and content fields. Unsurprisingly, results show that filtering using date information improves retrieval significantly. However, results also show that a careful choice of the date filter is important, given the multiple date fields existent in a patent. The actual ranking query is constructed based on word distributions of title, claims and content fields of the query patent. The overall MAP of this citation finding task is still in the lower 0.1 range. An error analysis focusing on the lower performing topics finds that the citation finding task (given publication recommend citations, which is a very similar setup as this year’s prior art evaluation) can be very different from the prior art task (finding patents that invalidates the query patent). It raises the concern that just the citations included in query patents can be a biased and incomplete set of relevance judgements for the prior art task.