Attention based deep learning for contextual classification of political conflict and mediation strategies in Indonesian digital news narratives

Political conflicts among parties in economic and social domains have increasingly been investigated using data-driven deep learning (DL) approaches. However, existing data-driven DL models often fail to identify discriminative patterns within political narratives. These challenges are primarily attributed to the scarcity of annotated political datasets, the linguistic diversity of regional languages, and the inherent ambiguity of political discourse. To address these limitations, an attention-based DL framework is proposed for contextual classification in Indonesian political narratives using Bidirectional Encoder Representations from Transformers (BERT) variants. Political data were systematically collected and annotated through a human-in-the-loop process, using expert-defined labels for political conflict and mediation strategies. The proposed framework consists of two stages: an extractive summarization module followed by a contextual classification module. Multi-task DL models with customized shared-backbone architectures based on BERT, DistilBERT, and RoBERTa were developed and trained on the compiled dataset. Performance was evaluated using five-fold cross-validation and benchmarked against conventional machine learning and state-of-the-art DL approaches. Experimental results demonstrated that the customized DistilBERT model achieved the best performance on the Indonesian political corpus, attaining F1 scores of 0.972 for political conflict classification and 0.909 for mediation strategy classification. These findings demonstrate the effectiveness of the proposed framework and highlight its potential as a human-centric analytical tool for supporting political analysts, policymakers, and researchers in interpreting complex political discourse.