Original title: Measuring Sentiment News with Transformer-Based Language Models
Maria Saveria Mavillonio, Stefano Borgioli, Caterina Giannetti, Chiara Ongari +1 · July 2026
Financial experts often track daily news mood to see how people feel about the economy. Older tools count positive and negative words in an article, but simple word counts miss context, negation, and complex sentence structure. To fix this, researchers built a new system using a transformer-based language model, a type of artificial intelligence that reads sentences while tracking how words connect to each other. The tool grades each sentence as positive, neutral, or negative, and then combines those scores into daily mood scores. To test how well it works, the researchers asked human volunteers to read articles and score them independently. The model scores matched human opinions far better than old word-counting methods did. It also did a much better job separating positive, neutral, and negative news, showing that understanding context matters.
You get a text from a friend: 'Great, just what I needed, another flat tire.' If you only counted the words 'great' and 'needed', you would think they were happy. You need the whole sentence and the situation to hear the sarcasm. In this paper, the old method counts words and misses that tone, while the new model reads the full sentence and lands closer to what a human actually thinks.
Financial news moves markets, but only if we can tell good news from bad. Old word-count methods miss context, so 'not a loss' can look like bad news. This paper shows transformer models capture that context and align much more closely with what human readers think.
The abstract doesn't say whether the new indices predict market movements, returns, or economic outcomes — it only reports agreement with human judgments. It also doesn't say whether the approach generalizes to other languages, domains, or time periods.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Measuring Sentiment News with Transformer-Based Language Models — Maria Saveria Mavillonio, Stefano Borgioli, Caterina Giannetti et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.