面向低资源命名实体识别的BiLSTM-Att-BCRF模型

在低资源场景下,由于受训练数据量少的限制,现有模型的参数不能拟合到预期效果,所以导致模型识别实体的性能不佳.该文提出一种融入伯努利分布(Bernoulli distribution)的新型损失函数,使模型能较好拟合数据.此外,该文在BiLSTM-CRF模型基础上融合多层字符特征信息和自注意力机制,并结合基于伯努利分布的新型损失函数,构建了BiLSTM-Att-BCRF模型.BiLSTM-Att-BCRF模型在20%的CONLL2003和20%的BC5CDR的数据集上,F₁值在BiLSTM-CRF模型基础上分别提升了7.00%和4.08%,能较好地适应低资源命名实体识别任务.

The BiLSTM-Att-BCRF Model for Low Resource Named Entity Recognition

In low-resource scenarios,the existing models are limited by the small amount of training data,and the parameters are not fitted to the expected effect,resulting in poor performance of the model in recognizing entities.In this paper,a new loss function incorporating Bernoulli distribution is proposed to allow the model to fit the data better.In addition,a BiLSTM-Att-BCRF model based on the BiLSTM-CRF model is constructed by fusing multi-layer character feature information and self-attention mechanism,combined with a novel loss function based on Bernoulli distribution.The BiLSTM-Att-BCRF model proposed in this paper improves the F₁ values by 7.00% and 4.08% based on the BiLSTM-CRF model on the datasets of 20% CONLL2003 and 20% BC5CDR,respectively.The model is better adapted to low resource named entity recognition tasks.