Throughout the Kim Jong Un era, North Korea has engaged in an unprecedented level of missile, nuclear weapons, and nuclear delivery systems development and testing. In particular, North Korea has made significant strides in advancing its intercontinen...
Throughout the Kim Jong Un era, North Korea has engaged in an unprecedented level of missile, nuclear weapons, and nuclear delivery systems development and testing. In particular, North Korea has made significant strides in advancing its intercontinental ballistic missile (ICBM) program and has now successfully tested four missiles which are estimated to be capable of reaching the United States: Hwasong-14, -15, -17, and -18. It has also made advancements in the intermediate (IRBM), short range (SRBM), submarine launched (SLBM), and hypersonic missile (HGV) categories. North Korea’s significant strides to expand and improve its missile capabilities has been a major destabilizing variable in the regional security of Northeast Asia, not only in terms of the arms-racing conditions they provoke but also in terms of the credible threat that it enables North Korea to impose on its enemies in South Korea, Japan, and the United States and its territories.
North Korean missile launches are considered by neighboring states to be major provocations that destabilize regional security. In light of this, this study proposes a framework for identifying and interpreting textual signals that foreshadow North Korean missile provocations. This is the first study to use the internal North Korean state newspaper Rodong Sinmun as a datasource to analyze and detect missile provocations with machine learning. It demonstrates that rhetorical patterns in domestic propaganda (not only outward-facing English-language media) contain quantifiable and potentially stronger signals of strategic behavior.
Three ML models were central to this analysis: two trained using the CatBoost classifier on a temporally balanced dataset from 2012–2022 and on a dataset limited to 2016 onward; and one was trained using a Support Vector Machine (SVM) on the 2016–2022 dataset. Each model was trained to classify whether a given edition of the Rodong Sinmun fell within the seven-day period preceding selected North Korean missile launches, based solely on its textual features. When evaluated on unseen and unlabeled test data, the CatBoost model trained on the temporally balanced 2012–2022 dataset achieved 81.9% precision and 73.8% recall for an F1 score of 0.776. The CatBoost model trained on the 2016–2022 subset attained 86.8% precision and 69.7% recall for an F1 score of 0.773. The SVM model trained on the 2016–2022 data achieved the highest recall (77.3%) and F1 score (0.797), making it the best overall performer in terms of early warning effectiveness, despite a slightly lower precision (82.3%) than the CatBoost model trained on the same data (86.8%).
SHAP (SHapley Additive exPlanations) analysis of the models revealed the most influential predictive concepts driving threat identification. In the 2012–2022 dataset, the CatBoost model ranked keywords such as 만리마 (Mallima speed campaign), 리졸브 (Key Resolve), 태양절 (Day of the Sun), 싸드 (THAAD), and 남조선 (South Korea, pejorative) as the strongest rhetorical indicators. In contrast, when trained on post-2016 data, the model emphasized terms such as 만리마, 미제 (American imperialist), 선군 (military-first), 북침 (invasion of the North), and 미국 (United States), along with continued prominence of 백두산 (Mount Paektu), 괴뢰 (puppet regime), and 남조선. These different terms becoming more influential when pre-2016 data was excluded reflects evolving priorities in state rhetoric, suggesting a shift from internally focused ideological mobilization to, in the later Kim Jong Un
period, more explicit confrontational signaling about specific external enemies in the lead-up to missile activity.
A closer investigation into the broader rhetorical environment and original texts of the Rodong Sinmun reveals how these textual indicators often coincide with certain types of messaging themes in the lead-up to provocations. Compared to prior studies analyzing conventional provocations via KCNA (Whang et al., 2016, 2017), this study demonstrates comparable or superior recall and accuracy in detecting provocations through Rodong Sinmun, suggesting that North Korea’s domestic-facing media carries reliable indicators of imminent provocative behavior. Further research based on this methodology may be well situated to examine the signal power of Rodong Sinmun textual elements for other types of provocations or events.