LLM-Based Analysis Outperforms Traditional NLP in Political News Framing

A recent investigation available on arXiv (2608.05155) contrasts conventional sentiment analysis (SA) with a multi-dimensional framing analysis utilizing large language models (LLMs) for assessing political news. Analyzed were 50 political news articles from 17 global media sources, revealing a significant flaw in traditional SA known as 'neutral collapse.' The widely used SA model RoBERTa categorized 70% of the articles as neutral, oversimplifying complex political narratives into a vague category. Alarmingly, 23% of these neutral articles had negative probability scores exceeding 0.30, suggesting underlying bias. Conversely, the LLM-based method effectively identified the direction and intensity of political bias, sensationalism, emotional resonance, and framing aspects, crucial for social sciences and humanities research. This study emphasizes the limitations of polarity-based sentiment analysis in capturing intricate political discussions and highlights the advantages of LLMs for deeper analytical insights. The findings are vital for scholars in political communication, computational social science, and media studies, providing a more advanced tool for political news analysis.