Research On Relation Extraction Method Based On ResCNN

With the rapid development and popularization of the Internet,it has become very difficult to obtain target information from massive unstructured data.Information extraction requires the structured processing of text,and relation extraction is a key part of it.The purpose of relation extraction is to extract the semantic relationship of entity pairs in the same sentence.It is also a hot issue in the field of natural language processing(NLP).Relation extraction is of great significance to the construction of knowledge graphs,and is conducive to the development of intelligent recommendation and information retrieval.In relation extraction tasks,supervised methods require manual data annotation,which is time-consuming,laborious and costly,and also relies heavily on NLP tools,and it have become a series of factors hindering the development of this field.Although the distant supervision method can automatically label the corpus,there are a lot of wrong labels.With the success of deep learning in the image field,more and more neural network models have been applied to relation extraction to fully learn features.Although the convolutional neural network has a strong ability to extract local features,it is not easy to discover long-distance dependencies in data samples.However,the problem of network degradation often arises when constructing deep networks.To address the above issues,the main works are as follows:To address the weak feature extraction problem of the shallow network model,this paper uses residual learning to design convolutional blocks to build a residual convolutional neural network,which can effectively avoid the network degradation problem on the one hand and transfer the bottom features to the top layer well without adding additional parameters on the other hand.At the same time,the Squeeze-and-Excitation block(SE)is added to recalibrate the features to enhance the transmission of effective features,and further strengthen the network's characterization and generalization capabilities.To address the problem that convolutional neural networks do not extract long-distance dependent information easily,this paper proposes using a bidirectional long short-term memory network to obtain long-distance contextual information in the sample data,which is used as the output of the vector representation layer together with word vectors and position vectors to provide rich features for the sentence encoder.Piecewise max pooling strategy is used in the pooling stage to replace the max pooling approach commonly used by convolutional neural networks in processing NLP tasks in order to preserve the structural features between entity pairs.To address the problem of a large number of incorrect labels in distant supervised datasets,this paper introduces a sentence-level attention mechanism after the pooling stage of the network model,such that the sentences having correct relationships receive high weights and those with incorrect labels obtain low weights.As a result,it can reduce the interference of noisy data and improve the accuracy of the proposed model.To address the problem of imbalanced samples and difficult-to-easy samples in distant supervised datasets,this paper proposes a combination of soft label and focus loss to improve the loss function,and then weights samples according to the degree of balance of the sample and the difficulty of sample learning,thereby improving the model performance.