Muhammad Aamir Nashrullah, Pradini Puspitaningayu, Hapsari Peni Agustin Tjahyaningtijas, Yulia Fransisca, Muhamad Bagus Fikril Alan
This study develops a binary sentiment detection framework for Indonesian slang-rich X (Twitter) comments using BERT-based transfer learning. The informal nature of Indonesian digital discourse, characterized by context-dependent slang (e.g., "santuy"[chill] masking sarcasm or "jayus"[failed humor] implying negativity), challenges conventional sentiment analysis tools. The proposed method fine-tunes the pre-trained "indobert-base-uncased"model on a heterogeneous dataset of Indonesian X comments covering three topics: East Java governor performance, cellular service providers, and movie reviews. Preprocessing includes slang normalization via a custom dictionary, stopword removal (combining NLTK and Sastrawi), and Sastrawi-based stemming. The model achieved 92% test accuracy, with word clouds identifying key sentiment indicators like "mantap"(positive) and "lebay"(negative). Contextual analysis revealed BERT's capacity to disambiguate slang, such as distinguishing sarcastic versus genuine uses of "woles"(relaxed). The results underscore transfer learning's efficacy in adapting global NLP architectures to low-resource languages like Indonesian, where slang dominates digital communication. This work provides a scalable solution for real-time sentiment monitoring without manual lexicon updates, addressing the limitations of rule-based approaches. It highlights BERT's potential to navigate linguistic fluidity in socio-culturally dynamic environments, offering actionable insights for businesses and researchers aiming to decode public sentiment in Indonesia's evolving social media landscape. © 2025 IEEE.
Universitas Negeri Surabaya, Department of Electrical Engineering, Surabaya, Indonesia
