Indonesian News Stance Classification Based on Hybrid Bidirectional LSTM and Transformer Based Embedding

Stance classification is used to understand the relationship between sentences so that the model can recognize the attitude of a response to a topic, where the attitudes are classified into three, namely supporting (for), neutral (observing), and opposing (against). Furthermore, stance classification could aid the automatic fake news detection. This research is specially made for Indonesian news titles. The proposed model used to recognize these news attitudes is Bidirectional Long Short-Term Memory (Bi-LSTM). Thus, to obtain the word representation vector, the pre-trained Bidirectional Encoder Representations from Transformers (BERT) embedding model from indoBERT is used to process words in Indonesian. In Bi-LSTM, each word representation will be processed twice in a forward and backward direction sequentially, so to get a vector representation of the sentence from the input, the output is taken from the results of the representation process of the last word in the forward direction process and the representation process results of the first word in the backward direction. Then the results of the two outputs are combined to serve as a sentence representation. Based on the test results on the Indonesian news dataset, the model for stance classification task was able to achieve an F1 score with an average of 78.30%, with an F1 score label for (supportive) of 73.10%, label observing (neutral) of 89.57%, and label against (against) by 72.23%. The performance is on par with the results of experiments with several Large Language Models currently available.