面向用户生成内容的旅游领域命名实体识别方法OA

In allusion to the challenges of named entity recognition(NER)in tourism-oriented user-generated content(UGC),including a large amount of noisy data,ambiguous nested entity boundaries,and lengthy text corpora under web-based contexts,a novel NER model integrating multi-head self-attention and adversarial training is proposed.The ERNIE 2.0 is used to encode tourism-specific corpora,generating semantically enriched dynamic word embedding.Adversarial training such as fast gradient method(FGM)and projected gradient descent(PGD)are introduced at the embedding layer by injecting minimal perturbations into word vectors to generate adversarial examples,so as to simulate noise characteristics in tourism UGC.A hybrid feature extraction layer combining bidirectional long short-term memory(BiLSTM)and multi-head self-attention mechanisms is constructed to capture the dependency relationship between entity boundary information and long-distance text,so as to dynamically adjust the feature weight distributions.A conditional random field(CRF)is used to decode the global optimal label sequence.The experiments were conducted on the self-built tourism dataset and the open-source news dataset CLUENER2020.The results show that the accuracy,recall,and F1-score of the proposed model on both datasets are all improved compared with the baseline model.It indicates that the model can still maintain high recognition accuracy on datasets from different fields,verifying its good generalization and robustness.