Patent Query Formulation by Synthesizing Multiple Sources of Relevance Evidence

Abstract

Patent prior art search is a task in patent retrieval with the goal of finding documents which describe prior art work related to a query patent. A query patent is a full patent application composed of hundreds of terms which does not represent a single focused information need. Fortunately, other relevance evidence sources (i.e., classification tags and bibliographical data) provide additional details about the underlying information need. In this article, we propose a unified framework that integrates multiple relevance evidence components for query formulation. We first build a query model from the textual fields of a query patent. To overcome the term mismatch, we expand this initial query model with the term distribution of documents in the citation graph, modeling old and recent domain terminology. We build an IPC lexicon and perform query expansion using this lexicon incorporating proximity information. We performed an empirical evaluation on two patent datasets. Our results show that employing the temporal features of documents has a precision enhancing effect, while query expansion using IPC lexicon improves the recall of the final rank list.