Modified self-attentive bi-directional long-short term memory for detecting clickbait in Indonesian news headline

One of the sources of income for online news media is the advertisements displayed on their web pages. Of course, someone is more likely to advertise on the certain web if the web is widely accessed by the public. Various methods can be used to attract the attention of readers to access or click the news, one of them is clickbait. We are interested to knowing whether the news headline contains clickbait or not because it indirectly affects the quality and credibility of the news. Research related to click-bait has been carried out quite a lot, but it is still very limited in Indonesia. Fakhruzzaman has done this in 2021 using M-BERT, this research will be carried out using Modified SA-BiDLSTM. A recurrent neural network with attentional mechanisms to process the information contained in news headlines, followed by concatenation with other relevant structural variables, and overlaid by a sigmoid classifier was trained to detect clickbait. Using a total of 10,000 training data, this model is able to produce an accuracy value of 0.817 and an F1-Score of 0.78, which indicates the model has succeeded in classifying 81.7% of the data correctly, in other words, the model can correctly determine a news headline contain a clickbait or not with percentage 81.7. The average correct classification for each label is 78%.