Researchers at the University of California, Davis have developed an artificial intelligence system that utilizes a deep learning neural network and an autoencoder to identify prior patent applications relevant to any proposed or filed new patent application.
Citations included in a typical patent application may not reveal all relevant, prior filings. Therefore, patent examiners and others performing prior art searches are usually required to manually identify additional relevant patents. Currently, search algorithms that can assist patent examiners rely mainly on keywords, and match those keywords with ones used in existing patents. This method generally leads to low accuracy in finding patents related to the patent application being considered/reviewed – primarily due to a high variance in word usage.
Researchers at the University of California, Davis have developed an artificial intelligence system that utilizes a deep learning neural network to identify prior-patent applications/filings that are relevant to any new patent filing under consideration/being reviewed. The neural network uses patent citation graph and content-level features - including both the plain text of the application and any image attached - to create a holistic, learnt representation. An autoencoder then compares this representation to all patents in a database and calculates how related or correlated the prior patent filings are to the inquiry. Experiments comparing this new technology to a leading text-mining search engine showed a detection rate of approximately 63% and a false-alarm rate of only 0.5%, compared to the text-mining engine's detection rate of 0.33% and a much higher false-alarm rate of 54.7%. Thus, this new technology finds a higher volume of relevant results while maintaining high accuracy and minimizing "false positives."
Patent prior art search, correlation detection, similarity detection, deep learning, autoencoder, matrix completion