Improving Named Entity Recognition Using Multi-Domain Knowledge Transfer

Neural networks have been widely used for named entity recognition (NER) and have shown stateof-the-art performance. However, in real applications, we often need to consider new types of entities in new domains where we do not have sufficient annotated data, which inevitably leads to a degradation in NER model performance. To this end, we propose an end-to-end domain adversarial named entity recognition (DANER) model. Unlike previous rule-based methods, DANER explicitly addresses NER-specific token-level context shifts by deeply integrating a gradient reversal layer with a sequence tagging architecture. Extensive experiments on our newly constructed FIND-2026, to our knowledge, the first large-scale Chinese NER corpus explicitly enriched with fine-grained domain annotations, demonstrate the superiority of our proposed model over six state-of-the-art methods. Moreover, on two public datasets (MSRA and Weibo), our approach achieves remarkable zero-shot knowledge transfer capabilities, yielding an absolute F-score gain of more than 22 percentage points compared with non-domain baselines without utilizing any target-domain labels.