基于ALBERT-CAW模型的时政新闻命名实体识别方法

Miscnews named entity recognition method based on ALBERT-CAW

针对时政新闻领域词语的含义复杂、新词更新速度快等问题,提出了一种基于ALBERTCAW的时政新闻命名实体识别模型。使用预训练语言模型ALBERT获取文本的动态字词向量,在CAW层中利用多层CNN提取词语的局部特征,使用LSTM获得每个词的上下文语意,将两者结果融合,输入BiLSTM获取深层特征,通过条件随机场(CRF)获取最有可能的标签作为识别结果。在自建的人民日报新闻数据集上取得了87.3%的F1值,优于对比模型。实验结果表明,该模型能较好地应用于时政新闻命名实体识别任务。

Aiming at solving the problem that words in the field of miscnews have complex meanings and update fast,a new named entity recognition model based on ALBERT-CAW is proposed.The model uses the pre⁃trained language model ALBERT to obtain the dynamic word vector and of the text.In the CAW layer,the model uses the multi⁃layer CNN layer to extract the local features of the word and uses LSTM to extract contextual semantics.The result of the fusion of the two is input into BiLSTM to obtain deep features.The most likely label is obtained as the recognition result through the conditional random field(CRF).On the self⁃built People’s daily news data set,the model obtaines a F1 value of 87.3%which is better than compared models.The experimental results show that the model is better applied to the task of miscnews named entity recognition.