SBIR Phase I: From Search to Research with Fast Patent-document Correlations

This SBIR Phase I project attempts to address a fundamental question for innovators: is my idea already patented? This do-it-yourself-initially service will empower small-business enterprises with synthesized research of over 10 million patent documents within 10 minutes based on the concepts contained within the idea. Such a comprehensive, real-time, low-cost offering is currently unavailable for small-business entrepreneurs. Current search tools do not synthesize the results into an executive summary, do not allow an entire document to be entered as the search input, and do not perform real-time concept/correlation computations. This proposed innovation will enable the inventor to submit an entire document (the idea) as the search query; then, utilizing network mathematics and artificial-intelligence algorithms, this service will synthesize search results in real-time summarizing what patent documents are most related to the idea based on natural-language processing. Such a service for small-business entrepreneurs would enable them to initially ascertain the novelty of their idea and give them an on-the-go education about the natural language used in patent documents in comparison to their idea. This meta-innovation would objectively ascertain the intellectual-property merit of any proposed technology, enable small-business innovators, and foster the acceleration of innovation development in the United States. The development of latent semantic analysis (LSA) has enabled algorithm development to extract latent (or hidden) semantic structure from documents addressing two important word-sense ambiguity issues that text-matching search cannot: polysemy (single term with multiple meanings; i.e., strike as to hit [verb], to start up [verb], or to cease working [noun]) and synonymy (multiple terms with single meaning; i.e., car and automobile). Albeit robust, this concept-search approach for large document collections is not tractable due to the high-complexity computational requirements for performing matrix singular value decomposition (SVD) necessary for LSA. Approximation techniques that use subset approaches necessarily introduce some amount of systematic error. To ascertain the most relevant documents in a large collection for a given focal document, this proposed innovation (search-subset LSA) will subset using proprietary search methodologies without any systematic error, reducing both the number of documents to compare and the number of terms to analyze thereby making real-time document correlations possible. The aims of this research are: to identify the optimal subset approach for comprehensive nomological capture of top-correlation candidates, to ascertain optimal input parameters for the focal query document, and to develop a statistical test to confirm that no systematic bias is present in this approach.