Multi-label Patent Classification using Attention-Aware Deep Learning Model

Abstract

Patent classification is challenging and essential for any further patent analysis task. We tackle the classification task on lower level patent classification (subgroup level) by using AttentionXML. Recently, pretraining methods for Natural Language Processing (NLP), such as DistilBERT pre-trained model, have achieved state-of-the-art results on some NLP tasks such as text classification. In this work we focus on investigating the effect of applying DistilBERT pre-trained model and fine-tuning it for the important task of multi-label patent classification. Moreover, the large USPTO-3M dataset (3,050,625 patents) based on CPC subclass and subgroup level is used for the purpose of comparing previous deep learning related studies.