Incorporating Forward and Backward Instances in a Bi-LSTM-CNN Model for Relation Classification
Taoling Xu, Yajun Du, Chunlong Fu, Chuan Xie · 2018
Relation classification is a task that identifies entity's semantic relationships from unstructured text. Lately, the method based on convolution neural network (CNN) achieves competitive performance comparing with other complex-structured networks by only using a standard convolution layer, a pooling layer and a softmax layer. However, CNN difficultly extract effective features or even wrong features when processing samples of large-pitch entities, decreasing their classification accuracy. Besides the existing methods could obtain inconsistent relation classifications for the forward and backward instances of the same sample. In this paper, a Bi-LSTM encoding layer is employed to enhance the ability of CNN to captured the contextual information of entities. Then, combining with the forward and backward instances of samples, a classification framework is proposed for relation classification. We conduct experiments on the public dataset SemEval 2010 task 8 verify to the effectiveness of our method. The method we proposed has the significant performance, even without additional artificial features.