Label Unification for Cross-Dataset Generalization in Cybersecurity NER
The study tried to combine four cybersecurity datasets by unifying their labels into a single set. They evaluated this unified approach using BiLSTM models across different datasets to see if it improved generalization. The results showed that simple label unification introduced errors, and alternative architectures like multi-head and graph-based transfer models did not provide consistent performance gains over single-dataset baselines.