А. В. Горбунов The authors describe the approach to increasing quality of automated patent search and emphasize the inefficiency of existing systems. The approach is based on automated formulation of query terminological vector out from the application text with its further extension with quasisynonyms from the distributional thesaurus built on the body of patent documents, and with further enrichment with bibliographic data – IPC (international Patent Classification) codes. Mathematical formalization of acquiring and extending query vector is provided; acquisition of patent distributional thesaurus is discussed. The authors propose original indicators of retrieval quality assessment with account to patent document specific character, i. e. the so-called “patent families”, which enables to evaluate the system’s capability to find the relevant inventions beyond the corresponding numbers. The experiments with Russian and English-language collections demonstrate the S@20 indicator increase by 10% as compared to standard search by keywords. The authors conclude on the effect of inclusion of patent families on the evaluation of search results efficiency. The independent expertise of search performance in the Russian language patent collection confirmed that the system delivered at least one relevant document in 96.25% cases. The developed algorithms have been implemented into the Rospatent search platform.
Научная статья 
машинное обучение