The Automated Sentiment Intelligence framework converted Thai student comments into four topic domains and three sentiment classes. From 4,145 records, 2,873 remained after preprocessing. XLM-R outperformed the BiLSTM implementation, reaching 0.922 topic accuracy and 0.80 sentiment accuracy, with clearer gains on neutral and negative feedback.
Key findings
- XLM-R produced higher topic accuracy than the BiLSTM baseline. • The transformer showed notable gains for difficult neutral and negative sentiment classes. • The framework connects data cleaning, annotation, dual-task modelling, and error analysis for institutional use.
Why this matters globally
The framework could reduce manual review workload and help institutions monitor teaching, curriculum, and facility concerns systematically. It also demonstrates how multilingual transformers can be adapted to languages without consistent word boundaries.
Thai researcher contribution
Researchers from Phuket Rajabhat University developed the framework around Thai-language student feedback, with supporting data reported as openly available through GitHub.
Limitations to consider
The dataset came from one vocational information system and was reduced to 2,873 records after preprocessing. Performance may shift across institutions, dialects, topics, and class distributions, and accuracy alone does not justify fully automated decisions.