rmtariq/multilingual-emotion-classifier

Fine-tuned emotion classifier (7-class) for Malaysian higher-education feedback. Trained on MYUniDialectSentiment840 (840 samples, 14 dialects, 20 topics, 15 learning contexts), a hand-curated balanced corpus covering Standard Malay, 13 regional dialects, and Manglish code-switching.

Labels

• happy
• love
• surprise
• no_clear_emotion
• sadness
• anger
• fear

Held-out test metrics (n=147, stratified)

split	accuracy	f1_macro
validation (n=105)	0.9524	0.9440
test (n=147)	0.9796	0.9753

Intended use

Sentiment / emotion monitoring of student feedback for Malaysian higher-education institutions. Designed to handle code-switched, dialect-heavy and informal academic discourse.

Training details

• Base: previous revision of rmtariq/multilingual-emotion-classifier
• Optimizer: AdamW (lr=2e-5, weight_decay=0.01, warmup_ratio=0.1)
• Epochs: 5 with early stopping on validation macro-F1
• Batch size: 16 (train) / 32 (eval), max_length=128
• Hardware: Apple Silicon MPS
• Class-weighted cross-entropy (for emotion only)

Dataset

MYUniDialectSentiment840 — 840 samples, balanced on sentiment, stratified 70/12.5/17.5 train/val/test by sentiment-x-dialect.

Citation / authors

• Raja Mohd Tariqi B. Raja Lope Ahmad — Ts., Fiscal Digest Sdn. Bhd.
• Raja Qatrun Nada Bin Raja Mohd Tariqi — Master of Education, UKM