PATENT CLASSIFICATION MODEL BASED ON DENSE VECTOR REPRESENTATION, SORTING TECHNIQUES, AND KNOWLEDGE EXPLICATION

Resumo

This study addresses the automatic classification of patents, aiming to assist examiners in efficiently categorizing documents. The objective is to propose a model that utilizes unstructured text data, taking into account the ordering of subclasses and the explication of knowledge. An integrative literature review was conducted to identify appropriate methods. The model is evaluated in two scenarios using data from the USPTO. In the general scenario, employing transformer-based neural network architectures, an accuracy of approximately 80% is achieved in recommending the top 5 subclasses, considering 50 retrieved documents. In the specific scenario, compared to traditional neural networks, an accuracy of 90% is obtained. Additionally, the feasibility of a knowledge graph is explored to support the ordering of patent subclasses, with the aim of facilitating the explication and visualization of results. The findings indicate that the proposed model can optimize the patent classification process, making it easier for examiners to select appropriate subclasses.