The Deep Learning Method for Information Extracting about the Components of Patented Devices and their Relationships

The study presents an approach to constructing semantic trees based on extracted information about the components of patented devices and their relationships. The object of analysis for data extraction is the abstract of Russian-language patents. Information about the device components can be used in many tasks of analyzing a patent array, such as determining the novelty of an invention, generating a patent application text, but the search for effective approaches to extracting such information is currently ongoing. This article proposes an approach to automating the extraction of morphological features of an invention from patent texts using machine learning methods. The developed system of 3 entities and 5 relations for representing the components of a patented device is shown. The process of forming a training sample consisting of 21,543 entities {SYSTEM - 1961, COMPONENT - 15,028, ATTRIBUTE - 4554} is given. This training sample and the models trained on its basis (ruBert, ruRoberta) were tested on the tasks of healing device components and their relationships.