基于Multi-BiLSTM-Attention的关系抽取

Relation Extraction Based on Multiple Bidirectional Long Short-Term Memory-Attention

在关系抽取任务中,单一的长短期记忆模型只能学习到某个特定维度的特征,而卷积神经网络可以利用多个卷积核学习不同维度特征。基于上述两个特点,该文提出了一种多层双向长短期记忆-注意力模型,该方法通过给长短期记忆模型设置不同大小的隐藏层,使其能自动从原始输入中抽取不同维度的、带依赖信息的抽象特征,并利用注意力机制捕获全局信息。实验显示,该方法能显著提高中文关系抽取效果,在ACE RDC 2005中文数据集上取得71.6%的F值。

In relation extraction,single long-short term memory network can only learn a certain dimensional feature. While convolutional neural network can use multiple convolution kernels to learn different dimensional features. Based on the above two points,this paper proposes a multiple bidirectional long-short term memory-attention model. Through setting different hidden layers,it can automatically extract different dimensional and abstract features with dependent information from the raw inputs,and use attention mechanism to capture global information. Experiments show that this method can significantly improve the result of relation extraction. Testing on the ACE RDC 2005 Chinese dataset,a 71.6% score on F1-score is gotten.