Deep Learning For English Relation Extraction

Relation Extraction is of great significance in the field of information extraction.This paper is mainly talked about the combination of using ResNet,RNN and Attention mechanism to do Relation Extraction task.A common practice in many effective experiments is to use CNN as the encoder only.After a multi-layer convolution operation,the result of pooling layers is classified by Softmax.When using RNN,the final result is weighted by Attention mechanism.In this task,very few people combine the two models to do the task of Relation Extraction.In this paper,the convolutional result is processed by using the RNN plus Attention mechanism instead of using the maximum pooling of the convolutional neural network,based on the use of the residual network as the encoder.It improves the performance of deep CNNs on the task of RE.On the fully supervised dataset,some papers have pointed using RNN after convolution operation is effective.However,the effect of the combined model is not satisfactory because of the large amount of noise on the weekly supervised data.In the experiment,combined with the characteristics of the ResNet residual block,the residual block and the RNN and the Attention mechanism are simultaneously used for the Weakly-Supervised Relation Extraction.The main contributions are as follows:Consider using a complex model in the Weakly-Supervised relation extraction,and combine ResNet and RNN to process the noise in the data.Using the complex model achieved better experimental performance than using a single model.The final result was almost same with PCNN+ATT.Combining ResNet,RNN and the attention mechanism can be easily changed and then applied to other NLP tasks.