Deterrent and Spiral Beliefs about Force and War
Authors: Sohini Timbadia & Sergi Margalef
Date: 10 May 2024
The World Wars in the 20th century are two of the most tragic events in human history, and unfortunately, conflict and war still happen today. When faced with a conflict or an aggressor, people can have different beliefs and attitudes towards it. One way of acting could be with appeasement, due to the fear of escalating the conflict, as in the early stages of World War I (1914-1918). Others may believe that when facing an aggressor, deterrent action is needed; arguing that stronger opposition to Nazi Germany might have prevented World War II (1939-1945).
These perspectives align with two prominent theories in international relations: the spiral model and the deterrent model. The spiral model says that conflicts escalate when one side tries to punish the other, and on the other hand, the deterrent model posits that to avoid an escalation, deterrent actions are necessary. Although people can hold both these views, they tend to lean towards one or the other.
Our research aims to develop a classifier model that identifies these spiral and deterrent beliefs from human text responses. Another objective is to uncover their relationship with other attitudes, specifically militant assertiveness and isolationism. Militant assertiveness refers to the tendency to use aggressive or confrontational measures in the international arena, whereas isolationism relates to the tendency to distance oneself from international affairs. We hypothesise that people with high militant assertiveness prefer deterrent actions, and isolationists may demonstrate more spiral ones.
We sourced our data from Joshua Alley's 2024 survey for the working paper ‘Provocation, Intimidation, and Nuclear Threats’. It involved 3,600 participants from the U.S. and Germany who were asked multiple-choice questions on how they would respond to a hypothetical scenario wherein a nation faces aggression from another, with their answers helping to assess their inclinations towards militant assertiveness or isolationism. Participants also had the option to provide additional insights in a text box, which formed the basis for training the classifier models. After excluding 18 responses without additional text, 3,582 responses remained, with 80% used for training the models and 20% for testing.
Two coders manually labelled the text responses on a five-point scale: 'Spiral', 'Maybe Spiral', 'Neither', 'Maybe Deterrent', or 'Deterrent'. When both coders agreed, that label was final. For slight disagreements (one step apart), the lesser intense label was chosen, reducing overconfidence in label assignment. For larger discrepancies (more than one step apart), a middle value was selected, effectively neutralising major differences and ensuring a balanced approach. This allowed for nuanced and consensus-driven labelling. We achieved an intercoder reliability of 69.5%. There were marginal differences in labelling 21.8% of the time, and stark differences occurred in 8.7% of the cases.
Given the unique challenges of classifying the specific type of text in our study, we explored several machine learning models known for their effectiveness with textual data to find the most suitable one for our dataset. The models we evaluated include Bag of Words, Multinomial Naive Bayes, Complement Naive Bayes, Linear Support Vector Classification (SVC), Long Short-Term Memory (LSTM), and Convolutional Neural Network (CNN).
We anticipate that the Bag of Words model may not achieve high accuracy since it ignores the order and structure of words, which are crucial for accurate classification. Nevertheless, it is our baseline for comparison with more sophisticated models. The Multinomial Naive Bayes model, while efficient, performs poorly on imbalanced datasets, prompting the use of Complement Naive Bayes. We also implemented LSTM and CNN models to determine if deep learning techniques could outperform traditional methods like Linear SVC. Notably, LSTMs and CNNs learn features from raw text, whereas Linear SVC relies on precise feature engineering but may still be competitive, especially with smaller datasets.
In our initial attempts to train classifiers using the five-point labelling system, the Linear SVC model performed the best, achieving an accuracy of 59%. However, this is still relatively low, indicating that none of the models performed optimally. The low accuracies observed across our classifiers can be attributed to the unbalanced distribution of labels, stemming from the initial use of a five-point scale. To address this, we converged the labels into a three-point scale, assigning texts labelled as ‘Spiral’ and ‘Maybe Spiral’ as ‘Spiral’ and ‘Maybe Deterrent’ and ‘Deterrent’ as ‘Deterrent’.
We further refined our dataset by only selecting entries with unanimous coder agreement from the original five-point labels. Then, we converged those labels to the three-point scale, improving the consistency and integrity of our dataset, ensuring more reliable results. This boosted the performance of our classifiers. Here, the Multinomial Naive Bayes model excelled with a 77% accuracy, followed closely by the CNN at 76%, as shown below.
Table 1: Model Accuracies for 3 Point Label with Coder Agreements
Despite these improvements, the Bag of Words model continued to underperform. However, all other models surpassed 71% accuracy. The shift to a three-point scale contributed to these positive changes by creating a more balanced distribution and simplifying the classification process with fewer labels.
To confirm the relationship between spiral and deterrent beliefs with militant assertiveness and isolationism, we conducted an analysis of variance test. The results showed statistically significant differences in mean scores. We visualised these results:
Figure 1: Mean of Militant Assertiveness & Isolationism for 3-Point Scale with Coder Agreements
These plots confirm our assumption that people with spiral beliefs tend to lean towards isolationism, and those with deterrent beliefs lean towards militant assertiveness.
In international relations, the spiral and deterrent models explain how conflicts escalate. The spiral model views aggressive responses as escalating conflicts, while the deterrent model suggests that leniency can provoke further demands. Our study uses text analysis of survey responses to identify these beliefs, achieving a 77% accuracy. There's potential for further improvements, but care must be taken to avoid overfitting. Future research could extend to analysing structured texts like political speeches and employing advanced techniques like BERT models. Our findings also associate spiral beliefs with isolationism and deterrent beliefs with militant assertiveness, suggesting avenues for future studies to explore these beliefs as indicators of broader international stances, which could enhance our understanding of international policy preferences.