<![CDATA[The effect of hyperparameters on part-of-speech tagging (POS Tagging) performance for the Indonesian language in deep learning models is rarely examined in isolation, because earlier studies usually add complex components (pretrained embeddings, CNN, CRF, or BERT), so the effect of hyperparameters on a plain Bidirectional Long Short-Term Memory (BiLSTM) cannot be isolated. The novelty of this study is a stepwise analysis of five hyperparameters (embedding dimension, hidden dimension, learning rate, train-test split ratio, and number of epochs) on a three-layer BiLSTM built from scratch in NumPy, without pretrained embeddings or CRF, on basic-level Indonesian sentences, that is, sentences commensurate with the comprehension level of elementary school (SD) students, represented by first-grade SD/MI textbooks. Secondary data comprised 10,747 sentences with 23 labels, combining the IDN Tagged Corpus and 747 textbook sentences. The effect of hyperparameters proved highly uneven: learning rate and hidden-layer capacity determined performance, whereas embedding dimension, split ratio, and number of epochs shifted accuracy by less than 0.5 percentage points. The best configuration reached 95.92% accuracy and 0.9966 weighted ROC-AUC, with labels X, UH, and FW being the hardest to predict. A five-seed test of the final configuration showed stable results (accuracy 95.82 ± 0.08% at 50 epochs and 95.76 ± 0.17% at 98 epochs; the difference between the two was not significant, p = 0.26), but small differences between hyperparameter settings have not been tested statistically. This study provides a simple and transparent baseline and a priority order for hyperparameter tuning as a starting point for NLP practitioners.]]>
