Leveraging LLM for Sentiment Detection in News Headlines

Abstract

In this study, a longitudinal dataset of more than 23 million news headlines from 47 U.S.-based media outlets is used to investigate the use of large language models (LLMs) for sentiment detection. Recent developments in LLMs offer potential gains in contextual understanding, adaptability, and generalization, even though traditional models have been used extensively for emotion classification. Using zero-shot and few-shot learning paradigms, which enable the model to generalize to new contexts with little task-specific data, we apply LLM-based emotion detection and assess its performance. The usefulness of sophisticated NLP models for media sentiment research is demonstrated by our analysis, which focuses on how well LLMs capture subtle emotional cues across time periods and news sources. The findings establish a new benchmark for emotion analysis in extensive, real-world text datasets by showing that LLMs can perform competitively and frequently better even with little supervision.