Sezibra/conflict-text-analysis

NLP analysis of UN Security Council speeches on the Tigray conflict using BERTopic and LLM-based zero-shot classification. Identifies how geopolitical blocs frame humanitarian, security, and sovereignty narratives across 88 text units from 4 UNSC meetings (2021-2022). # LLM-Powered Conflict Text Analysis ## Research Question How do UN Security Council member states frame the Tigray conflict, and how does framing vary across geopolitical blocs and over time (November 2020 to November 2022)? ## Navigation | Section | Description | |---------|-------------| | Motivation | Why text analysis matters for conflict research | | Key Findings | Main results across both methods | | Data | 88 text units from 18 UN documents | | Methods | BERTopic, LLM-as-Annotator, Human Validation | | Results in Detail | Figures and interpretation for each analytical step | | Limitations | Honest assessment of constraints | | Notebooks | Four-notebook analytical progression | | How to Reproduce | Setup and replication instructions | | References | Academic sources | ## Motivation The UN Security Council discussed the Tigray conflict at least 15 times between November 2020 and November 2022, but geopolitical divisions prevented formal action. China and Russia repeatedly blocked resolutions, while Western members pushed for humanitarian access and accountability. The early meetings were held in closed format, producing no public record. Only four open meetings generated verbatim transcripts. These transcripts contain the diplomatic language through which Council members framed the conflict. Different states emphasized different aspects: humanitarian crisis, military security, human rights accountability, state sovereignty, or political negotiation. Identifying these framing patterns systematically requires computational methods that go beyond manual reading. This project applies two complementary approaches from the "Text as Data" tradition (Grimmer and Stewart, 2013): unsupervised topic discovery with BERTopic and theory-driven zero-shot classification with an LLM. The combination produces both data-driven and researcher-driven perspectives on the same corpus, and the disagreements between them are as informative as the agreements. ## Key Findings …