Effect of Word Embedding on Indonesian Social Media Hate Speech Classification Using Hybrid CNN-SimpleRNN

Mas Muhammad Rizqi Adiguna (Telkom University, Bandung)
Yuliant Sibaroni (Telkom University, Bandung)

Building of Informatics, Technology and Science
Vol 8 No 1 (2026): June 2026

The rapid rise in the number of people using social media platforms, specifically platform X, poses great difficulties for spotting any possible harm from hate speech. In particular, due to the high informality of discourse of Indonesian internet users, expressed through slang, abbreviation, and distorted spelling, automatic moderation becomes even more complicated. The current study attempts at classifying hate speech on social media platform X, proposing a hybrid model of CNN combined with a Bidirectional SimpleRNN architecture, alongside a comparative study on word embedding approaches. The combination between CNN and SimpleRNN is chosen because it allows for exploiting both CNN's capability to extract local spatial features by finding toxic n-grams and the strength of Bidirectional SimpleRNN for capturing long contextual dependency in the sequence of text data. Given that the problem of OOV is highly significant, TF-IDF, FastText, and Word2Vec have been rigorously tested not only separately but also combined in different ways. Compared to the baseline configuration using standalone TF-IDF (which achieved 84.56% accuracy), the results show that the use of the hybrid TF-IDF + FastText provided the best performance with average accuracy of 86.49%, average precision of 86.32%, recall of 86.80% and F1 score of 86.55%. Conversely, the combination of multiple dense semantic vectors (Word2Vec and FastText) led to semantic drift and feature ambiguity; this created feature overlap and computational noise that obscured classification decision boundaries, resulting in redundancy and poor performance. It shows that the combination of lexico-statistical significance and semantic subword context greatly contributes to achieving better results in the Indonesian language setting and is very resistant to slang and OOV words found in digital settings. This study contributes to the field of natural language processing by providing a lightweight, highly accurate, and computationally efficient lexico-semantic framework tailored for moderating highly unstructured Indonesian social media text.