Research On Relation Extraction Based On Adaptive Chosen Convolution Kernel Network

The target of relation extraction is to identify the relationship between entities from natural language text,which is one of the main tasks of information extraction.At present,the method of relation extraction based on convolutional neural networks have been widely used in the task of relation extraction.There are various convolutional networks for feature extraction,and the performance of networks affects the quality of relation extraction tasks.However,traditional convolutional networks mostly use fixed and single convolutional kernels,which makes it difficult to dynamically select the features generated by different sizes of convolutional kernels,resulting in weak network representation ability.It results in poor network presentation capabilities.In addition,many methods of relational extraction ignore the overall semantic information of the text,which leads to inadequate extraction of text features.Facing the above problems,this thesis combines with the Selective Kernel Networks to carry out the research on relation extraction,and improves the accuracy of relation extraction by improving the representation ability of network.Main work of this thesis is:1.We propose a relation extraction model named selective kernel and multi granularity network(SK-MGNet).And combined with the characteristics of SK-MGNet and relation extraction,the multi pooling is proposed.Firstly,the text is represented as vector as the input of the network.Then SK-MGNet is used to make the network adaptively select the features generated by different sizes of convolutional kernels according to specific task,which can give more weight to the appropriate convolution.And combined with the idea of multi granularity,strong features are highlighted.The multi pooling method proposed in this thesis takes into account the strong features and global information that appear many times.Finally,these features are passed a fully connected softmax layer to predict the final relationships.In order to confirm the validity of the model,experiments were performed on the NYT-Free Base dataset,and compared with the current popular baseline methods.The experimental results show that SK-MGNet can improve the quality of the task of relational extraction.2.Aiming at the problem that the whole semantic information of text is ignored,this thesis proposes a method based on semantic feature fusion,which can further improve the ability of network representation.Firstly,sentences are preprocessed according to the characteristics of relational extraction.Then,we use the idea of transfer learning to train on a source domain dataset.At the same time,in order to improve the performance of the model,adversarial training is added to get the weighting parameters.Then,the weight parameters are transferred to the target domain dataset to get the semantic features representing the whole information of the text.Finally,the semantic features are fused with the extracted features of SK-MGNet to provide richer information for the whole model.The experimental results show that the method proposed in this thesis can obtain more accurate features,provide more information for the network,and improve the precision of relation extraction.