DOI: 10.11118/ecos2026_020

FINE-TUNING ROBERTA FOR SENTIMENT ANALYSIS OF SUPPORT CONVERSATIONS WITH PRIVACY-PRESERVING ANONYMISATION

Vojtech Mahdal¹, Oldrich Trenz¹

¹ Department of Informatics, Faculty of Business and Economics, Mendel University in Brno, Zemědělská 1, 613 00 Brno, Czech Republic

This article describes a prototype system for processing support-conversation text. The system accepts a text input, classifies sentiment into three classes (positive, neutral, negative), masks selected sensitive items such as names, e-mails, phone numbers, URLs, usernames and long numeric strings, and stores an anonymised record. The sentiment classifier is based on RoBERTa and was fine-tuned in three stages on public chat and dialogue datasets. Evaluation showed that the second stage reached the best score on its internal test split, but its accuracy on a separate combined test set fell from 83.41% to 70.06%. After a final fine-tuning stage on a smaller dataset with a higher share of neutral examples, the model reached 92.82% accuracy on 1,573 independent samples. The anonymisation layer is rule-based and is most suitable for structured patterns; names, especially sentence-initial capitalised words, remain the main limitation. The result is a working demonstration consisting of a FastAPI service, a JavaScript dashboard and an SQLite log, containerised with Docker and deployed on Hugging Face Spaces.

Keywords: sentiment analysis, RoBERTa, fine-tuning, anonymisation, GDPR, customer support

pages: 204-210, online: 2026