A patent search strategy based on machine learning for the emerging field of service robotics

Emerging technologies are often not part of any official industry, patent or trademark classification systems. Thus, delineating boundaries to measure their early development stage is a nontrivial task. This paper is aimed to present a methodology to automatically classify patents concerning service robots. We introduce a synergy of a traditional technology identification process, namely keyword extraction and verification by an expert community, with a machine learning algorithm. The result is a novel possibility to allocate patents which (1) reduces expert bias regarding vested interests on lexical query methods, (2) avoids problems with citation approaches, and (3) facilitates evolutionary changes. Based upon a small core set of worldwide service robotics patent applications, we derive apt n-gram frequency vectors and train a support vector machine, relying only on titles, abstracts, and IPC categorization of each document. Altering the utilized Kernel functions and respective parameters, we reach a recall level of 83% and precision level of 85%.