한국어 특허 문장 기반 CPC 자동분류 연구 ―인공지능 언어모델 KorPatBERT를 활용한 딥러닝 기법 접근―

With the advent of various new technologies in the 4th industrial revolution, securing intellectual property rights has become increasingly important to countries or companies for maintaining technological competitiveness and building growth engines. In particular, a patent is a technical document that contains the core technology and is widely used for measuring corporate value and analyzing competitive technologies. To make this support, the CPC that including latest and detailed technical fields has been developed and more than 62 million documents worldwide have been classified as CPC. And five advanced patent offices which account for more than 80% of the world’s patent applications invest big budget for CPC of new patent applications every year. In this study, we had generated the KorPatBERT that was pre-trained and outperformed in patent field using the BERT language model which understands the meaning of sentences beyond the limits of keywords. And we proposed the methods and constructed the dataset that relieved imbalanced distribution for each CPC code. And finally, we had generated the AI CPC model that can classify into main group level and verified through reliable evaluation indicators. Through this, we want to contribute the sustainable development of Korean patent based classification and NLP field.