Unmasking Public Sentiment: A Sample Efficient Approach to Analyzing Twitter Opinion on US Aid to Ukraine

Abstract

The U.S. aid to Ukraine is a bipartisan topic of extreme socio-political importance. While several organizations have conducted surveys to understand the public stance on this topic, there is no research to date that analyses public opinion on social media, possibly due to the lack of annotated data. Therefore, this research compares several sample-efficient methods (including in-context learning) to analyze tweet sentiments with minimal training data. First, we collect 11,289 tweets about the U.S. aid to Ukraine and mapped them to U.S. states. Next, we explore three different approaches to sentiment analysis: tool-based, embedding-based, and prompt-based. Our results indicate that GPT-4 Few Shot improves accuracy by 121.8% and 77.5% over TextBlob and Vader, respectively. Our geospatial analysis shows that Indiana has the most negative normalized net sentiment (NNS) of -0.83, while Vermont has the most positive NNS of +0.33. Finally, we perform a detailed thematic analysis to identify the common arguments that support or oppose the aid. We highlight that our results do not correlate with media surveys, possibly due to the presence of echo chambers and algorithmic biases.