Deparadoxifying Strategic Decisions: An Integrated Approach Utilizing Machine Learning and Natural Language Processing: Analyzing Strategic Decision Making within the UN Security Council by Applying the BERT Model and Inferential Statistics

Abstract

The purpose of this paper is to answer the call for empirical research on deparadoxification by demonstrating a new approach which utilize the rapid technological developments that has taken place since Luhmann introduced the concept of deparadoxificaiton. In doing so, the paper seeks to combine two academic fields which not yet have been connected, namely deparadoxification and machine learning. The paper demonstrates a feature-based approach with BERT and random forest, to classify the paragraphs in the United Nations Security (UNSC) meeting minutes into deparadoxification strategies. The contextual model will be compared with a non-contextual model (using TF-IDF) to investigate whether deparadoxification is context dependent or not. Due to the lack of existing labeled data, the authors constructs their own training dataset through iterative manual labeling using active learning with least confidence sampling. The model will be used to uncover the distribution of deparadoxification strategies and provide classified data for the regression analysis to investigate whether the strategies affect resolution voting outcome. The model reached 0.53 accuracy and 0.44 F1 macro after five iterations of labeling, followed by hyperparameter tuning. The non-contextual model reached 0.47 accuracy and 0.35 F1 macro. Both models outperformed ZeroR and the uniform dummy classifier (UDC) by a large margin. Both the labeled and predicted distribution suggested that the strategies does not follow a uniform distribution, but are rather imbalanced. The regression analysis suggested that the strategies (only based on occurrences) does not explain any of the variation in the voting outcome. We argue that this is due to the majority of resolution votes being unanimous. We believe, however, that as this merging of fields gets more attention, larger datasets, sufficient in size to train complex models, will be made available, which might lead to different results. While the regression analysis did not show significant results, the fact that both models outperformed ZeroR and UDC proves that they were in fact able to pick up on a pattern, demonstrating that it is possible to detect and measure deparadoxification through machine learning