Automated patent classification from the perspective of technology management: Deep learning methods and transformer models

Patents have become a widely used source of data for technology development, monitoring, and prediction due to their ability to disclose prior technologies and provide information about new technologies. Examining a patent application and assigning it to the appropriate class is a time-consuming and labor-intensive process for the patent expert. Automatic patent classification approaches can significantly help with this problem. The aim of this thesis is therefore to make a contribution to solving this problem on the basis of various leading studies. The thesis comprises four studies, which can be summarized as follows. The first study includes a bibliometric analysis of automatic patent classification. The second study evaluates patent texts using a hierarchical attention mechanism. The third study compares ensembled deep learning methods with the transformer encoder model. The last study investigates how the pre-trained Bert model addresses the challenge of automatic patent classification. It is worth mentioning that the studies conducted in this thesis aim to provide practical and useful solutions for automatic patent classification