Freedom House'i demokraatia indeksi ühtsuse analüüs BERT keelemudeliga

Abstract

This research examines Freedom House’s democracy index, Freedom in the World (FITW). Academic literature has shown, that there are a lot of problems with academic literature. They have found to be biased and to have poor aggregation rules. The study investigates FITW using a machine learning-trained BERT language model, exploring how machine learning has been utilized in social sciences and its potential for further applications. BERT is used to create embeddings, that a vectorial representation of texts. Embeddings allow for a qualitative examination of the any texts. Since FITW index uses descriptive texts, these descriptive texts can be examined. Hence, this research examined, how FITW descriptions are tied to their scores. A hypothesis was created, that descriptive texts, that are most similar to one another, are also the ones with the highest scores. The empirical findings demonstrate, that texts and score are indeed linked. This was done through creating a score difference index for all questions. Next, the score difference between the top 1% most similar texts was done. Through this it was shown that the score and texts are most strongly linked, when the score are high, and linked the weakest, when the scores are low. T-SNE method was used to show embeddings in 2-d projections. Through this the score were also proven to be linked visually. Visualizations were also used to show, that texts from the same regions of the world cluster together. Further geographical analysis revealed, that countries generally cluster according to their geographical locations, especially in the Americas and Europe. Overall, the study concludes that FITW descriptions strongly correlate with high scores but weaken as scores decrease. Geographically, FITW descriptions are most consistent in the Americas and Europe, while the index is less uniform in Africa.