This investigation delves into political polarization in Georgia, examining social media discourses and influential factors using machine learning methodologies grounded in language models. In line with the European Union's recommendations to the Georgian government, the study aims to monitor and assess the country's efforts to decrease polarization as part of its EU integration process. We amassed a dataset of 120,000 posts from 154 social media accounts linked to diverse political parties, originating from four leading platforms. By dissecting ten primary polarization discourses, we scrutinized opposing narratives fueling polarization. We applied 42,000 annotation points from 5,000 distinctive posts to develop and validate text classification models. The study employed three pre-trained models—BERT-Multilingual, XLM-RoBERTa, and Electra-Ka—for accurate classification and analysis of Georgia's political polarization discourse and narratives.

Introduction

Since its independence, Georgia has consistently expressed its European aspirations, which are enshrined in the Georgian Constitution and supported by the country's citizens and political spectrum. The Georgian government's commitment to progress on its European path is exemplified by its initial intention to submit an EU membership application in 2024. In response to the European Union's recommendations, this study aims to monitor Georgia's efforts to reduce political polarization, a critical aspect of the country's integration process. By analyzing social media discourses and key factors contributing to polarization, we hope to provide valuable insights into Georgia's progress toward meeting the EU's expectations and fostering a more inclusive and cooperative political landscape.

This research offers valuable insights into Georgia's political polarization, delving into discourses and influential factors through social media-based analysis. By leveraging state-of-the-art language models and machine learning methodologies, we were able to examine a comprehensive dataset spanning multiple platforms and political parties. Our findings underscore the importance of continued efforts in reducing political polarization in Georgia, aligned with the EU's recommendations, to facilitate the country's integration process and promote a more inclusive political climate. This study provides a solid foundation for future research on Georgia's political landscape and serves as a useful model for analyzing polarization in other contexts.

Methodology and Data Collection

The methodology employed in this study involved a thorough examination of data collected from diverse social media accounts associated with various political parties and their members. In total, 154 public accounts were carefully monitored using a rigorous selection process, which facilitated the analysis of political polarization in Georgia. The dataset encompassed posts published between 2021 and March 2023, amounting to 120,000 posts extracted from four major social media platforms: Facebook, YouTube, Telegram, and Vkontakte. This comprehensive approach to data collection allowed for a wide-ranging perspective on Georgia's political landscape, ensuring a balanced representation of political opinions and ideologies.

Political parties covered in the study:

• ● Georgian Dream
• ● United National Movement
• ● For Georgia
• ● Girchi
• ● Girchi More Freedom
• ● For People
• ● Strategy Agmashenebeli
• ● LELO
• ● DROA
• ● European Georgia
• ● Mokalakeebi
• ● Conservative Movement
• ● People's Power
• ● Alliance of Patriots
• ● Labour Party
• ● ERI

Polarisation discourses

In our analysis, we identified ten key discourses of polarization within Georgia's political landscape, leveraging topic labeling and entity extraction techniques. For each discourse, we selected two opposing narrative groups to comprehensively examine the diverse perspectives driving polarization. The selected discourses include:

• ● The 2008 August war between Russia and Georgia
• ● The conflict between Russia and Ukraine
• ● Georgia's role in the conflict between Russia and Ukraine
• ● The custody and treatment of Mikheil Saakashvili
• ● The confrontation between liberal and conservative values
• ● Pro-Western versus Anti-Western foreign policy
• ● Pro-Russian versus Anti-Russian foreign policy
• ● The role of NGOs in the country's political processes
• ● Media independence and the role of media in political processes
• ● The influence of oligarchs in political processes

By examining these discourses, we aim to provide a comprehensive understanding of the complex dynamics shaping political polarization in Georgia and the underlying factors driving these divisive narratives.

Two additional discourses Law against foreign agents, and Sanctions against Georgian Court members, will be added in the next phase of the study.

Data Annotation

To facilitate a robust data annotation process, 26 master's students from Tbilisi State University and ILIA University, with a focus on Political and Social studies, were chosen to partake in the text annotation exercise. This exercise produced a total of 42,000 annotation points, addressing 5,000 unique posts from the collected dataset. To guarantee the quality and reliability of the annotated text examples, each post was independently annotated by four distinct annotators. This strategy minimized the risk of bias or misinterpretation, leading to a more objective analysis. The annotated examples were then used to train the text classification models, as well as to validate the accuracy of the discourse and narrative prediction models, thus reinforcing the rigor and robustness of our research framework.

Classification Models

We used several pre-trained models for our text and entity classification tasks, including BERT-Multilingual, XLM-RoBERTa, and Electra-Ka. We fine-tuned these models with our dataset to best suit our requirements. Additionally, we employed the foundational models of GPT3.5 and GPT4 in our final analysis, which provided superior accuracy. The F1 scores for each model are presented in the table below:

Model	F1 Score for original posts in Georgian	F1 Score for post translated in English
BERT-Multilingual	0.67	0.73
XLM-RoBERTa	0.68	0.71
Electra-Ka	0.70	-
GPT3.5	0.68	0.88
GPT4	0.71	0.91

As the results indicate, GPT3.5 and GPT4 achieved superior performance in our analysis, guiding our final decision.

Discussion and Implications

The findings of this study provide a comprehensive understanding of the complex dynamics shaping political polarization in Georgia, including the key factors driving these divisive narratives. By examining the identified discourses and opposing narrative groups, we have illuminated areas where cooperation and collaboration between political parties can be encouraged, as recommended by the EU.

Limitations

While this study provides valuable insights into political polarization in Georgia through the lens of social media discourses, there are certain limitations to consider. Firstly, the data was sourced exclusively from official channels. These channels often exercise caution in their postings, potentially omitting the most polarized statements. As a result, our dataset might not capture the full spectrum of polarized sentiments prevalent in the broader social media landscape. Secondly, to gain a more comprehensive understanding of the situation, it would be beneficial to incorporate media monitoring. Such an approach would likely unearth additional nuances and insights that official channels might overlook or deliberately exclude. This broader perspective could offer a more holistic view of the political polarization in the region.

Conclusion

Our analysis reveals that Georgia has made some strides in addressing political polarization, with tangible progress evident in specific discourses. However, further efforts are needed to ensure robust cooperation and mutual understanding across the political landscape. The country's ongoing commitment to EU integration requires addressing the root causes of polarization, fostering a more inclusive political environment, and strengthening democratic institutions.

The study was conducted in two distinct stages, with each stage focusing on different discourses and timeframes. During the first stage, we concentrated on four polarization discourses, analyzing social media data from January 1st, 2023, to April 1st, 2023. This phase laid the groundwork for understanding the emerging trends and patterns of polarization in the Georgian political landscape.

In the second stage, which is currently planned, we aim to expand the scope of our analysis by including all 10 polarization discourses. Furthermore, we intend to incorporate transcriptions from video content and broaden the dataset to cover a more extensive timeframe, spanning from January 1st, 2021, to June 1st, 2023. This comprehensive approach will enable a deeper exploration of the intricacies of political polarization in Georgia, providing valuable insights into the underlying factors that contribute to its persistence. By analyzing an extended period, the second stage will offer a better view of polarization dynamics over time, illuminating how these discourses have evolved and their potential impact on Georgia's political climate.