Thai MBTI Psychographic Segmentation: Complement NaïVe Bayes vs. Thai Transformer on Short Social Media Posts

We present a transparent Thai MBTI-style text classifier for short
social posts that outperforms a Thai specialist transformer while
remaining easy to audit and deploy. Using surface-preserving
normalization and MBTI-string masking, we train a calibrated TF-IDF +
Complement Naïve Bayes (CNB) model and compare it to WangchanBERTa under
author disjoint splits. On 1,338 held-out posts, CNB achieves Accuracy
0.84 / Macro-F1 0.83 versus 0.57 / 0.55 for the transformer. Calibrated
probabilities, an abstain option, and user-level aggregation turn
post-level predictions into reliable, privacy aware communication-style
signals for segmentation, targeting, and creative positioning. A web
demo mirrors the pipeline and includes smoke tests (public figures and
volunteer posts) to validate parity, latency, and leakage controls.