A Time Series-Based Dynamic Topic Analysis and Future Keyword Prediction for Artificial Intelligence Patents

This paper aims to conduct a future keyword prediction study by applying dynamic topic modeling and the VAR model through time series analysis using artificial intelligence-related patent data from 2012 to 2022. Patent data was collected through the KIPRIS platform, and major keywords such as ‘information’, ‘data’, and ‘user’ were constantly dealt with in the annual topic trend graph derived through dynamic topic modeling, and the importance of topics such as ‘image’ and ‘video’ has been increasing in recent years. These results suggest that research trends in various application fields are changing with the development of artificial intelligence technology. Next, 10, 20, and 30 top keywords were extracted respectively from the yearly patent data collected for future keyword prediction using BoW, TF-IDF, and Word2Vec, which are three embedding methods. It was confirmed that ‘information’, ‘data’, ‘user’, ‘device’, and ‘system’ were the same important keywords in the annual top keyword analysis by year. The ranking of different patterns was subsequently shown for each embedding method. After that, an experiment was conducted to compare the keywords predicted through the VAR model with the actual 2023 keywords, and it was confirmed that the Word2Vec embedding method showed the highest prediction performance, and the overall performance improved as the number of keywords increased. This study provided important basic data for predicting technology trends and future trends through artificial intelligence patent data, and it is expected that it can be used as a basis for strategic use directions and investment decisions for various stakeholders such as companies, research institutes, and the government.