Reading management tone with AI sentiment analysis
This week, XBRL International’s Revathy Ramanan published the third article in our series Using LLMs to Analyse Narrative Disclosures, shifting the focus from patterns in reported numbers to the tone of management commentary. Following earlier articles on liquidity risk and anomaly detection, this instalment explores how sentiment analysis can uncover signals embedded in narrative disclosures.
This week Revathy zeroed in on management reports, using a large language model to classify sentiment as positive, neutral or negative. What makes this powerful is the structured foundation. Because the narrative is precisely tagged in XBRL, the relevant text is already isolated, removing the need to extract or clean data from lengthy documents. The LLM can then be applied quickly and consistently across hundreds of reports.
Using a sample of more than 900 Ukrainian annual reports for 2022, the results were revealing. Most management reports showed a neutral tone, reflecting factual, compliance-driven disclosures. Where sentiment diverged, positive language tended to highlight stability or strategic progress, while negative sentiment often reflected external pressures such as conflict or financial strain. As with anomaly detection, these deviations provide useful starting points for deeper analysis.
Together with topic analysis and anomaly detection, sentiment analysis shows how narrative disclosures can be explored in ways that were previously impractical at scale. Structured XBRL tagging makes the text precise and comparable, while LLMs make it fast and flexible to interrogate. As these techniques are combined, narrative reporting moves from something that is merely read to something that can be (and is!) systematically analysed, compared and monitored across entire reporting populations.
Read the article here.
