Forecasting the Level and Types of North Korea’s Provocations with Text Mining

Abstract

Research into the feasibility of predicting specific events using Text Mining techniques has been actively pursued in conjunction with the advancement of Machine Learning. Consequently, the potential for predicting North Korea’s provocations utilizing Text Mining methods has emerged. However, the field lags behind other domains due to challenges in acquiring high-quality training data and the complexity associated with event classification. This study addresses these limitations by leveraging a Pre-trained BERT model to establish a comprehensive classification framework for North Korea’s provocative behavior, moving beyond binary classifications (provocation or peace) used in previous research. Original data from the Korean Central News Agency (KCNA) and domestic media sources were gathered and analyzed as training data. Notably, the findings demonstrated that employing original data from the KCNA increased prediction accuracy compared to utilizing data from domestic media. This study offers a way to enhance the informational value of North Korea’s provocations through scientific predictions, ultimately bolstering the reliability of qualitative expert judgments.