Comparative Study of BERT Based Architectures for Multi Task News Classification and Threat Detection

We present a comparative analysis of single-task and multi-task BERT-based models for Indonesian news classification across two objectives: category prediction (five classes) and threat detection (binary). Using 8,951 annotated news titles, single-task baselines achieved weighted F1 scores of 0.84±0.01 (category) and 0.87±0.00 (threat). Multi-task hybrids integrating CNN, LSTM, and Bi-LSTM layers performed comparably overall and improved minority class threat detection, with the BERT CNN variant attaining the highest threat F1 (0.88±0.01). Per-class results confirmed that the Ideology category, represented by only 279 samples, remained the most challenging. Efficiency benchmarks on an NVIDIA L4 demonstrated practical feasibility, with batch size 32 throughput of approximately 450-470 items per second (equivalent to 2.1 2.2 ms per title) and single-item latency of around 67-69 ms. Training times ranged from 138±12 to 193±22 s across seeds. These findings indicate that multi-task BERT hybrids can improve threat detection while sustaining near real-time throughput, supporting their applicability in large scale monitoring of Indonesian news streams.