融合多头注意力的中文命名实体识别方法OA

To address the limited feature extraction capacity and the difficulty in modeling long-range dependencies in existing Chinese named entity recognition(NER)models,this study proposed a method that integrates multi-head attention(MA)with a bidirectional gated recurrent unit(BiGRU).Specifically,it first used the ALBERT(a lite BERT)pre-trained language model to generate contextualized representations,which are then fed into a BiGRU to extract global semantic features.A multi-head attention mechanism is subsequently applied to capture long-range dependencies and further enhance the semantic representation then the multi-head attention mechanism is used to capture long-distance dependent information to enhance semantic representations.Finally,a conditional random field(CRF)layer decodes the optimal label sequence.Experimental results show that the proposed method achieves F1 scored above 95%on both the People's Daily and MSRA datasets,outperforming competing models.In addition,compared with the BERT-BiLSTM-CRF model,this approach reduces training time by approximately 14.5%,demonstrating its effectiveness and generalizability.