From Dictionaries to Transformer Models: Financial Sentiment Analysis for Romanian News

This paper creates a financial sentiment indicator from a Romanian news archive and analyzes its relevance for financial stability purposes. The ongoing literature on text-based indicators highlights the informational value of financial news, but existing methods often rely either on manual labels or on purely dictionary-based approaches. We address the limitations that can affect central banks or low-resource languages, because labeled datasets are scarce, and interpretability is essential.

We build on the author’s previous work, where we introduced a Natural Language Processing toolbox for the National Bank of Romania. We expand this base by proposing a weakly supervised learning methodology where the Financial-Stability dictionary is used to compute sentiment at the chunk level. These pseudo-labels we built, serve as a training sample for a Transformer model. Also, we apply unsupervised pre-training on our financial news archive, so the model learns domain-specific language before the fine-tuning step.

It is clear that continual learning improved the stability and coherence of sentiment estimates and enhance the detection of growing financial stress periods. The aggregated daily sentiment indicator shows correlations with market variables, indicating its use for early warning monitoring. Overall, this research shows that the mix of expert dictionary, continual learning, and Transformer models creates a robust and interpretable method for financial sentiment analysis, giving both methodological and practical value to the growing literature on text-based economic indicators.