Patent expanded retrieval via word embedding under composite-domain perspectives

Abstract

Patent prior art search uses dispersed information to retrieve all the relevant documents with strong ambiguity from the massive patent database. This challenging task consists in patent reduction and patent expansion. Existing studies on patent reduction ignore the relevance between technical characteristics and technical domains, and result in ambiguous queries. Works on patent expansion expand terms from external resource by selecting words with similar distribution or similar semantics. However, this splits the relevance between the distribution and semantics of the terms. Besides, common repository hardly meets the requirement of patent expansion for uncommon semantics and unusual terms. In order to solve these problems, we first present a novel composite-domain perspective model which converts the technical characteristic of a query patent to a specific composite classified domain and generates aspect queries. We then implement patent expansion with double consistency by combining distribution and semantics simultaneously.We also propose to train semantic vector spaces via word embedding under the specific classified domains, so as to provide domain-aware expanded resource. Finally, multiple retrieval results of the same topic are merged based on perspective weight and rank in the results. Our experimental results on CLEP-IP 2010 demonstrate that our method is very effective. It reaches about 5.43% improvement in recall and nearly 12.38% improvement in PRES over the state-of-the-art. Our work also achieves the best performance balance in terms of recall, MAP and PRES.