Validation-Based Hyperparameter Optimization of a Hybrid IndoBERT-BiGRU Model for Indonesian Sentiment Analysis

Public sentiment across digital platforms spans multiple domains, each with different linguistic characteristics. This study investigated the hyperparameter configurations selected for a hybrid IndoBERT-BiGRU model across three Indonesian-language datasets and an aggregated dataset. Sentiment pseudo-labels were generated using the Indonesian Sentiment Lexicon (InSet), while Optuna with a tree-structured Parzen estimator (TPE) optimized BiGRU hidden size, number of layers, and dropout rate for the four training settings. Validation Macro F1 was used as the optimization objective. The Global and MyPertamina models selected 384 hidden units, two BiGRU layers, and a dropout rate of 0.2465; the Shin Tae-yong model selected 256 hidden units, one layer, and a dropout rate of 0.2301; and the One For All model selected 128 hidden units, one layer, and a dropout rate of 0.3543. Their validation Macro F1-scores were 81.25%, 82.16%, 61.40%, and 65.71%. Agreement between InSet-generated pseudo-labels and Human_Sentiment labels across 14,526 records, measured using Cohen’s kappa, yielded a value of 0.2511. On the held-out test sets, evaluated against InSet-derived pseudo-labels, the models achieved Macro F1-scores of 49.19%, 46.86%, 22.71%, and 30.00% for Global, MyPertamina, Shin Tae-yong, and One For All, respectively. Per-class evaluation showed consistently low F1-scores for the Neutral class. These findings demonstrate differences in the validation-selected configurations but do not establish that domain-specific tuning is superior or that dataset size determines model complexity. The results should be interpreted in light of the limited agreement between InSet and Human_Sentiment labels and the very small One For All test set, which contains only eight records.