Automatic sentiment and topic annotation of Italian economic documents: a comparison of different LLM-based approaches

Sentiment and topic analysis of news articles and central bank communications is highly valuable for understanding macroeconomics trends. The sentiment captured from these texts can supplement traditional economic indicators, as central bank communications influence consumer expectations while news capture market sentiment. Traditional sentiment analysis approaches for economic documents rely mainly on lexicon-based methods, which suffer from context-blindness, limited flexibility and language dependency. Large Language Models (LLMs) represent a promising alternative, combining deeper contextual understanding and the ability to learn from few examples without requiring expensive fine-tuning. This thesis explores the use of LLMs to automatically annotate sentiment and topic of Italian news articles and European Central Bank (ECB) bulletins. Different prompting strategies are compared to assess their ability to replicate human annotations, using manually labeled data as ground truth. The study provides a comparative analysis of LLM-based methods for annotating Italian economic texts, highlighting both the capabilities and limitations of automatic approaches in this specific domain. Results show that these models approximate human annotators in topic identification, while sentiment analysis proves to be challenging for both models and humans. Nevertheless, the incorporation of few-shot approaches yields consistent and measurable improvements across tasks.