Rahmat Budianoor, Setyo Wahyu Saputro, Friska Abadi, Radityo Adi Nugroho, Andi Farmadi
Indonesian culinary comments on social media platforms such as Instagram are characterized by informal spelling, regional language mixing, slang expressions, and emojis, posing substantial challenges for automated sentiment classification. While IndoBERT has demonstrated strong performance across Indonesian natural language processing tasks, the contribution of individual preprocessing components to fine-tuning performance on informal text remains underexplored, particularly in the culinary domain. This study addresses this gap by conducting a systematic preprocessing ablation study on IndoBERT-Base fine-tuning for Indonesian culinary sentiment classification, accompanied by a comparative evaluation against Naive Bayes with TF-IDF, SVM with TF-IDF, and BiLSTM as representative baselines. A dataset of 3,500 manually labeled Instagram culinary comments across three sentiment classes was used, with a stratified 80/10/10 split. Six preprocessing variants were evaluated under identical experimental conditions to isolate the contribution of each component. The results show that slang normalization is the most impactful single preprocessing step, yielding a macro F1-score gain of +0.0609 over the no-preprocessing baseline, while the full pipeline achieves an accuracy of 0.8800 and a macro F1-score of 0.8465. IndoBERT-Base with the full pipeline outperforms all baselines across all evaluation metrics. Per-class analysis reveals that the negative class achieves the lowest F1-score of 0.7600, with sarcastic expressions and Banjar regional vocabulary identified as primary sources of misclassification. These findings indicate that preprocessing decisions have a measurable and non-uniform effect on IndoBERT fine-tuning performance. In this study, slang normalization provides the most substantial individual contribution in bridging the vocabulary gap between informal user-generated text and the model’s pre-training distribution.