Malay Tweet Sentiment (BERT)

mesolitica/bert-base-standard-bahasa-cased fine-tuned to classify Malaysian tweets as Negative, Neutral or Positive.

• Test macro-F1 0.627, accuracy 0.66 on 300 held-out tweets (TF-IDF baseline: 0.576).
• Trained on 1,397 human-labelled tweets. Casual Twitter Malay; formal text may be less accurate.
• Clean input like training did: links replaced with URL, usernames with @USER.

Code, error analysis and full results: https://github.com/szal31/Malay-Tweet-Sentiment-Analysis-Using-BERT