Flagging clickbait in Indonesian online news websites using fine-tuned transformers

Click counts are related to the amount of money that online advertisers paid news sites. Such business models forced some sites employ a dirty trick click-baiting, i.e., using hyperbolic and interesting words, sometimes unfinished sentences in headline purposefully tease readers. Some Indonesian also joined party clickbait, which indirectly degrade other established sites' credibility. A neural network with pre-trained language model multilingual bidirectional encoder representations from transformers (BERT) acted as an embedding layer is then combined 100 node-hidden topped sigmoid classifier was trained detect clickbait headlines. With total 6,632 headlines training dataset, performed remarkably well. Evaluated 5-fold cross-validation, it has accuracy score 0.914, F1-score precision 0.916, receiver operating characteristic-area under curve (ROC-AUC) 0.92. The usage BERT text classification task tested possible be enhanced further. Future possibilities, societal impact, limitations detection discussed.