Research On CNN-based Entity Relation Extraction Method

Entity Relation Extraction is one of the important basic technologies in Natural Language Processing,and it is also a key sub-task of Information Extraction.It has an important research value and a wide range of application scenarios.In recent years,deep learning based methods have been widely used in Entity Relation Extraction.Different from rule-based and statistics-based methods,deep learning can automatically acquire text features through the network and integrate feature learning into the process of model building,which reduces the incompleteness of artificial design features.Entity Relation Extraction mainly uses two basic models:Recurrent Neural Networks(RNN)and Convolutional Neural Networks(CNN).RNN-based method will lose part of the local features when modeling sentences and CNN-based method can integrate the local information of sentences into global information well.Therefore,CNN is used as the model basis in our work.After investigation,we find that CNN-based Entity Relation Extraction method faces the following challenges.One the one hand,deep learning method requires a large amount of annotation data.For relation extraction,the lack of labeled data is a major problem at present.On the other hand,the use of attention mechanism has further improved the performance of the neural network model.However,the existing attention mechanisms mostly focus on lexical-level features and lack attention to high-level semantic features such as sentence global information.In response to the above challenges,this paper proposes corresponding solutions.The main innovations and contributions of this paper are as follows:(1)The high-level semantic attention-based piecewise convolutional neural networks(PCNN_HSATT)model is proposed.After the piecewise max pooling layer of the CNN,an attention mechanism layer is added to pay attention to the sentence global information that contributes greatly to relation prediction.Since different convolution kernels can extract different sentence global information,the attention mechanism layer can be used to reasonably assign weights to obtain more effective features.(2)The hypernym information of the external semantic resource HowNet is merged in the vector representation layer of the network,which enriches the information of the vector representation layer and improves the F1-value.(3)For the lack of Chinese entity relation labeled data problem,the paper proposes a data augmentation method that integrates the Tongyici Cilin and adds dependency syntactic constraints.This method effectively increased the COAE2016 training set from 988 to 11 328 sentences.Compared with PCNN and other existing models,the experimental results showed that the proposed PCNN_HSATT model has more advantages in relation extraction task.We used the data augmentation method to obtain a large amount of valid data and verified its validity in experiments.