A novel patentability detection model based on Siamese network

Patentability assessment is the most critical stage of the patent registration process. For an invention to be granted a patent, it must meet the patentability criteria of novelty, inventive step, and industrial applicability. Evaluating the novelty and inventive step requires a comparison of patent texts in terms of their technical scope. This comparison is conducted by experts from authorized patent offices and communicated to applicants through official reports. Under the EP and PCT conventions, patentability evaluations are expressed in search reports using the categories X, Y, and A. Due to the lengthy and complex structure of patent texts, as well as the specialized nature of novelty and inventive step comparisons, patentability assessment is a challenging and time-consuming process. Consequently, patent registration timelines are extended, and inventions are often submitted prematurely, before reaching a patentable level of development. The aim of this study is to address the patentability assessment problem from a unique perspective, distinguishing it from other semantic similarity tasks, by proposing a semantic similarity model specifically focused on patentability evaluation. The proposed model generates patentability text pairs with a focus on patent claims and determines the similarity between these pairs using the X, Y, and A categorization. The experiment of the proposed model shows that the integration of the Siamese network structure with CNN and LSTM deep learning architectures has proven to play a highly effective role in achieving accurate results. By utilizing the X, Y, and A codes identified by experts in published patent search reports for model training, this study marks a significant step forward in solving the patentability assessment problem.