Abstract
This dissertation presents a computational framework for analyzing media framing through event-based methods, addressing limitations of traditional topic-level classification approaches that fail to capture the nuanced mechanisms by which news sources construct divergent narratives. The research has been organized around four phases, each contributing novel datasets, models, and analytical tools. In the first phase, we tackled Cross-Document Event Coreference (CDEC) by constructing the Richer EventCorefBank (RECB) dataset through sentence-level decontextualization, significantly improving the scalability and robustness of cross-document event coreference systems. In the second phase, we developed an event-based framing pipeline for media attitude detection that leverages event coreference, contextual event omission and inclusion, and causal relationships to classify news article stances on contentious geopolitical topics, demonstrating the effectiveness of event-centric features over traditional baselines. In the third phase, we introduced the Framing-divergent Event Coreference (FRECO) task and corpus, the first contrastive framing dataset centered on identifying equivalent events described with divergent frames across news sources, and evaluated various large language model (LLM) fine-tuning strategies including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and bootstrapped mining for capturing subtle framing distinctions. In the fourth phase, we extended the framework to Framing-Aware Event Causality Identification (FRECI), examining how causal attributions between events are constructed differently across sources and languages. Collectively, these contributions have established a unified, interpretable, and scalable pipeline for computational framing analysis, advancing both Natural Language Processing (NLP) methodology and practical applications in media transparency and narrative analysis.