| Entity relationship extraction aims to identify semantic relationships between entities in text,which is not only a key subtask in natural language processing,but also an important support for tasks such as question and answer systems,intelligent search,and opinion analysis.Recently,the mainstream relationship extraction models use deep neural networks to extract higher-order abstract semantic features from the input corpus.However,when deeper network structures are used,the introduction of irrelevant semantics and the generation of gradient disappearance can lead to a weaker representation of the underlying semantic features and generate the problem of semantic disappearance.To solve this problem,this thesis first designs a stacked percolation network structure that preserves the significant semantic information in the original utterance.Then,on the basis of this model structure,a two-channel percolation network model is designed and implemented,which can fuse the semantic information of multiple granularities in the shallow convolutional network and the deep convolutional network to enable the neural network to learn a more comprehensive semantic representation of the text and improve the performance of acquiring inter-entity relationships.The main work accomplished is as follows:(1)Proposes a relationship extraction method based on a stacked layer percolation network.A stacked layer Deep Penetration Network(DPN)combining convolution and upsampling methods is proposed for the problem of semantic disappearance when a deeper network structure is used in relationship extraction tasks.The model adds a text upsampling layer that enhances semantic features before convolution;after convolution,k-max pooling is used to extract significant semantic features.The method finally makes the significant semantic features percolate through the convolutional neural network,and builds semantic dependencies among the semantic features.(2)A two-channel percolation network-based relationship extraction method is proposed.Based on the characteristics of text corpus with multiple granularities,a deep network model supporting entity relationship extraction is designed and implemented on the basis of stacked percolation network(Two Channel Deep Penetration Network,TC-DPN).The model has channels that can learn semantic features of different granularities and fuse semantic features of different depths in the network to achieve higher performance than the classical mainstream models on several relationship extraction datasets: 88.7%,95.5%,84.9%,84.9%,84.9%,84.9%,84.9%,84.9%,84.9%,84.9%,84.9% and 91.6% on five biological relationship extraction datasets AImed,Bio Infer,IEPA,HPRD50,and LLL,respectively,88.0%,and 91.6% of F1 value scores,and 81.07% and 81.25% of F1 value scores were achieved on ACE2005 and CLTC relationship extraction datasets,respectively.The experimental results show that the accuracy and stability of the method have been further improved compared with the classical deep learning method. |
