基于字符嵌入与BiGRU的命名实体识别

Named Entity Recognition Based on Character Embedding and BiGRU

在命名实体识别任务中,一般使用词级嵌入向量化句子,忽略了字符级特征所包含的信息。为此,提出一种将字符级特征与词级特征将结合的模型,首先使用BiLSTM捕捉字符级信息,结合输出的字符向量与Glove模型获取词向量输入BiGRU网络,然后引入自注意力层机制重新分配特征权重,最后使用CRF层输出具有全局最优节的标注序列。实验表明,该方法相较于BiGRU-CRF模型在命名实体识别任务中的性能更好,在CoNLL-2003、OntoNotes5.0数据集上的F1值分别达到91.69%、88.97%,对提高命名实体识别任务的性能有所帮助。

In named entity recognition tasks,word-level embeddings are generally used to vectorize sentences,while ignoring the information contained in character-level features.Therefore,propose a model that combines character-level features and word-level features,uses BiL⁃STM to capture character-level information,and then combines the output character vector with the word vector obtained by the Glove model and outputs it to the BiGRU network.The attention layer mechanism redistributes the weights of the features,and finally outputs the label se⁃quence with the globally optimal section by using the CRF layer.The experimental results show that the proposed method achieves better per⁃formance in the named entity recognition task than the BiGRU-CRF model.The F1 values on the CoNLL-2003 and OntoNotes5.0 datasets reach 91.69%and 88.97%,respectively,which proved the propose method helps improve the performance of named entity recognition tasks.