摘要
基于循环神经网络模型在命名实体识别任务中限制了GPU并行计算效率,提出融合全局词频统计的膨胀卷积模型。对训练语料文本向量化后,使用文本向量训练膨胀卷积神经网络各节点权重,通过条件随机场有效避免预测结果产生不符合规则标签。在MSRA数据集上,F1值达到了92.12%,在简历数据集上,F1值达到了93.48%。模型的训练速度相比Bi-LSTM-CRF模型提高了3倍。条件随机场虽然能够学习到标签之间的潜在关系,但极大地降低了模型的运行速度。同时在序列建模中,卷积神经网络无法考虑词的有序关系。实验结果表明,本文的模型可在保持高精度的同时,具有更快的处理速度。
Recurrent neural networks have limited GPU parallelism computation efficiency in named entity recognition tasks.To solve this problem,we proposed a dilated convolution neural network model by introducing global word frequency statistics.Firstly,we vectorize our corpus,and then calculate the node weights of the dilated convolution neural network with the vectorized corpus.Finally,a conditional random field is used to effectively prevent irregular results from training datasets.A maximum F1 score of 92.12%is obtained on the MSRA benchmark datasets,and a maximum F1 score of 93.48%is obtained on the resume benchmark datasets.The model is around 3 times faster than Bi-LSTM-CRF model.Although the conditional random field can learn the potential relationship between the labels,it greatly reduces the computing speed.The convolutional neural network cannot consider the ordered relation of the tokens in sequence modelling.The experimental results show that our model provides very high training speed while maintaining high accuracy.
作者
张东
迟呈英
战学刚
ZHANG Dong;CHI Chengying;ZHAN Xuegang(School of Computer Science and Software Engineering,University of Science and Technology Liaoning,Anshan 114051,China)
出处
《辽宁科技大学学报》
CAS
2020年第3期231-236,共6页
Journal of University of Science and Technology Liaoning
基金
国家自然科学基金(61672138)。
关键词
命名实体识别
膨胀卷积
序列标注
named entity recognition
dilated convolution
sequence labelling
作者简介
张东(1996-),男,辽宁海城人;通讯作者:迟呈英(1963-),女,辽宁鞍山人,教授。
