Abstract
Whenever a protest takes place, a politician visits a foreign country, or two countries negotiate a deal; it takes almost no time for these events to be reported in the television, newspapers and more recently, the vast array of online news sources. For decades, International Relations researchers have investigated political interactions among states, societies and organisations leading to studies of war, peace, cooperation and conflicts. Such studies help to underpin why countries act the way they do and the political motivations behind the events around the world. However, there has been little work on mining norms in international relations. Moreover, there has been little work on empirically evaluating event data sets to construct norms. This thesis explores this gap by employing norm identification concepts from multi-agent systems.
Norm identification has been an increasingly popular topic of research in multi-agent systems technology that models software entities as agents. The core essence of an agent-based system lies in how software agents in a complex, rule-governed environment act in order to achieve their goals while conforming to social roles and expectations. Norms play a significant role to help agents conform to these social roles or expectations. We apply these concepts to extract norms from international relations. We select a one year dataset from the Global Database of Event, Language and Tone (GDELT) that is the most advanced event dataset of political events, comprising of over 300 million records. We use the Bayesian norm identification technique, a recently proposed approach to identify normative behaviour in terms of Prohibition and Obligation norms. To do this, the dataset was analysed to extract event dialogues between countries using a cluster computing tool. Root event categories such as ‘Protest’, ‘Consult’ from the CAMEO taxonomy were used to encode these dialogues from the dataset. The Sequence Memoizer, a language modelling tool was trained to learn baseline probabilities of event sequences from these event dialogues.
We extracted norms by learning the odds of each norm in a set of norm hypotheses with respect to the null hypothesis that there are no norms. Our norm mining model extracted seven most likely norms from the one-year dataset. The findings showed that a model combining the baseline probabilistic model with normative reasoning based on the seven top norms explained the data significantly better than the baseline model alone.