INTEGRASI INDOBERT DAN LONG SHORT-TERM MEMORY UNTUK KLASIFIKASI SENTIMEN PASAR BITCOIN BERBASIS MEDIA SOSIAL X

Abstract

Bitcoin's price is notoriously volatile, and much of that volatility traces back to how investors feel about the market at any given moment, which makes an automatic system for reading Indonesian-language sentiment genuinely useful. Word-embedding approaches that treat each word in isolation tend to fall short here because they cannot hold onto sentence-level context or the way meaning unfolds across a sequence of informal, slang-heavy words. This study builds a three-class sentiment classifier (positive, neutral, negative) for the Bitcoin market by pairing IndoBERT, used purely as a contextual feature extractor, with a Long Short-Term Memory (LSTM) layer that learns the sequential pattern on top of it. The dataset came from crawling X (formerly Twitter) for Indonesian-language tweets between January and December 2024, yielding 3,579 raw posts that were cleaned and manually labeled down to 3,542 usable tweets with a mild class imbalance. Preprocessing covered cleaning, case folding, normalization of crypto slang and emoji, and a selective stopword removal that deliberately kept negation words intact. All twelve IndoBERT encoder layers were frozen, and only a 128-unit LSTM with 0.4 dropout sat on top, trained with Focal Loss and class weighting to keep the imbalance from skewing predictions, then evaluated on a split of 2,856 training, 339 validation, and 347 test tweets. The resulting hybrid model reached 90.20% accuracy, 91.13% macro precision, 90.33% macro recall, and a 90.58% macro F1-score, performing best on the negative class (F1-score 92.19%) while most of its remaining confusion involved positive tweets being read as neutral, largely because of ambiguous crypto slang. Overall, combining IndoBERT with LSTM produces a Bitcoin sentiment classifier accurate enough to serve as one input among several for investors weighing their next move.