Hoax Detection in Informal Indonesian Language through a Hybrid Approach: Integrating Lexicon and Deep Learning

Informal Indonesian is highly variable: slang, abbreviations, misspellings, and other nonstandard forms often express the same meaning in different surface forms. This variability can complicate automatic hoax detection because the classifier must interpret both the underlying message and its unconventional wording. This study examines a hybrid pipeline in which lexicon-based normalization is applied before IndoBERT fine-tuning. Here, hybrid refers to a sequential combination of lexical preprocessing and deep learning classification; it does not imply a new neural architecture or training algorithm. The dataset initially contained 3,628 labeled Indonesian texts collected in 2026. Nine records with invalid split information were removed, leaving 3,258 samples for development and 361 samples for a fixed hold-out test. Two otherwise identical configurations were evaluated: one used the original text, while the other used lexicon-normalized text. Using 10-fold StratifiedGroupKFold grouped by topic, the original-text configuration obtained a Macro-F1 of 75.29 +/- 3.10%, compared with 75.21 +/- 3.66% for the normalized-text configuration. Neither the Wilcoxon signed-rank test (W = 24.0; p = 0.7695) nor the exact McNemar test (41 versus 40 discordant cases; p = 1.00) indicated a significant difference. On the fixed hold-out set, the normalized-text configuration achieved 80.89% accuracy and 71.39% Macro-F1 (95% bootstrap CI: 65.4-76.9). Because a majority-class predictor already reaches 80.61% accuracy, the accuracy figure should be read cautiously. Overall, explicit normalization did not improve global performance, although exploratory subset results indicate that its influence may depend on how many tokens are actually normalized.