Improved Performance of Cyberbullying Detection in Indonesian Using Sentence-Level Semantic Expansion

Abstract

Social media platforms, such as Twitter, have afforded more opportunities for unfettered expression than were ever previously possible. But it has also been responsible for the rise of cyberbullies, particularly in Indonesia. The identification of cyberbullying in this sphere is particularly difficult due to the high frequency of brief, ambiguous statements and informal/deformed language that tends to render unreliable traditional classifier systems useless. To address these limitations, this paper presents a novel framework for improving text representation via sentence-level semantic expansion. The process enriches social media text data with semantically related terms from an external lexical resource (Kateglo is based on the Big Indonesian Dictionary, KBBI) to capture additional semantic contexts. The linguists verified and preprocessed the annotated dataset prior to semantic enrichment. Next, semantically augmented texts were fed directly to embedding algorithms (Word2Vec and BERT) to obtain vector representations. We then fed these embeddings into three deep learning classifiers: a Convolutional Neural Network (CNN), a Bidirectional Gated Recurrent Unit (BiGRU), and a Long Short-Term Memory (LSTM). Experimental results on three public datasets show that using the semantic extension with contextual embeddings yields higher classification accuracy. The CNN-BERT and BiGRU-BERT hybrid models achieved the highest accuracy at the high-level: 88.39% and 86.83%, respectively. The results showed that combining linguistic features with deep learning-based approaches is beneficial for cyberbullying detection in Indonesian-language content. However, there are challenges associated with context-specific words or phrases, slang, and multi-label situations. Future work may also include semantic-level inference at a higher level and the use of larger, more diverse datasets to improve detection performance under practical conditions.