Classificação de documentos de patentes usando o doc2vec

Abstract

Patents are considered extremely useful sources for activities related to the search and analysis of information and the generation of new knowledge. In this article, we use a paragraph vector algorithm doc2vec, an extension of word2vec, which learns representations of sentences in a document, in a supervised deep learning scheme for automatic patent classification. The classification was carried out in documents with summaries of patents in english, in a hierarchical process comprising sections, classes, subclasses according to the International Patent Classification (IPC). The tests were developed in four steps, required, due to the large number of classes and subclasses, with the objective of identifying primary or secondary IPC codes, if it is associated with a set of classifications related to other aspects expressed in the patent. The tests presented very promising results in the classification of patents. The next steps will be to produce qualitative assessments and compare them with other models of machine learning present in the literature.