<![CDATA[This study aims to classify cyberbullying text on Indonesian social media using an enhanced Bidirectional Long Short-Term Memory (BiLSTM) model integrated with 300-dimensional FastText pre-trained embeddings. The key challenges addressed include the high prevalence of informal language, slang, abbreviations, and writing variations that complicate automatic text analysis, as well as the semantic limitations of random embeddings. The proposed method comprises a preprocessing pipeline including text cleaning, slang normalization with toxic word preservation, tokenization, and padding, followed by a BiLSTM model built using the Functional API with dual pooling (GlobalMaxPooling + GlobalAveragePooling), L2 regularization, SpatialDropout1D, BatchNormalization, and class weight balancing. The dataset consists of 2,109 Indonesian comments collected from TikTok and Instagram, with a nearly balanced class distribution (1,058 Non-CB and 1,051 CB). Experimental results demonstrate that the proposed model achieves improved performance over the baseline, with a significant reduction in the overfitting gap between training and testing accuracy. This study demonstrates that combining FastText pre-trained embeddings with an optimized BiLSTM architecture is an effective and resource-efficient approach for Indonesian cyberbullying text classification.]]>
Copyrights © 2026
