Sentiment-Aware Feature Recommendations for Maternal Mental-Health Apps via IndoBERT and BERTopic on Indonesian TikTok Data

Maternal mental-health concerns voiced by Indonesian pregnant women are widely shared on TikTok, yet systematic, data-driven insight into these expressions is still lacking. To bridge this gap, we propose a two-stage NLP pipeline that (i) fine-tunes IndoBERT for sentiment classification and (ii) applies BERTopic to uncover thematically coherent topics within each sentiment class. Using 10,000 TikTok posts collected between 2022 and 2024, we manually labeled 4,000 entries and employed class-weighted training to mitigate class imbalance; the resulting model achieved an F1score of 0.93 and accuracy of 0.94, outperforming a traditional SVM baseline (F1-score 0.85). Topic modeling on the sentimentlabeled corpus produced clusters with coherence scores of 0.4679 (positive) and 0.5294 (negative), indicating good interpretability for short, informal Indonesian text. Positive posts were dominated by social-support themes (about 98%), whereas negative posts centered on mental-health distress (about 51%), partner-related conflict (about 20%), and selfesteem concerns (about 14%). These findings offer actionable guidance for designing sentiment-aware digital interventions— such as peer-support forums, stress-management modules, and partner-engagement features—tailored to the emotional needs of pregnant women. The study is limited to a single platform and binary sentiment classes; future work will extend the framework to other social-media sources and employ finegrained emotion detection to enable real-time, cross-platform maternal-mental-health monitoring.