Research On Logical Relation Extraciton Technology In Knowledge Graph

As an important branch of artificial intelligence technology,knowledge graph,provides convenience for knowledge acquisition and intelligent information application in the Internet era.With the continuous promotion and application of artificial intelligence technology,the research of knowledge graph has important value and far-reaching significance.The process of knowledge graph construction includes information extraction,knowledge fusion,knowledge processing and other sub tasks.Relation extraction is an important task in the field of information extraction.It aims to extract the relation facts between entity pairs in unstructured texts and generate structured data.In this thesis,we study on the task of relation extraction and propose a method of relation extraction based on the combination of transfer learning and deep learning model,which solves the problem of relation extraction in the domains of few sample data.The main researchs are as follows:1.Aiming at the problem of relation extraction in the field of few samples,we present a relation extraction method based on piecewise convolutional neural network with transfer learning.The model combines the transfer learning method and deep learning model and uses the external knowledge of similar fields to help the few sample domain data to complete the task of relation extraction.Firstly,the text is represented by word vectors as the input of network structure.Considering the influence of adding different features on the final relation extraction(classification)results,we add word embeddings,position embeddings,part of speech features and syntactic features to the text representation.Then the source domain data is pre-trained in the model and the convolution layer parameters are retained.Finally,the convolution layer parameters are transfered and we use the target domain data fine-tune the network to improve the accuracy of the final relation classification results of the target data.Furthermore,we utilize multi granularity convolution filters to explore the influence of convolution filters granularity on relation classification results.2.The thesis proposes a new method of relation extraction based on fusion models and transfer learning.In this method,fusion models is proposed to solve the problem of limitation of convolutional neural network structure.Bi-directional long short-term memroy network structure has some advantages over convolutional neural network structure in feature extraction of long texts,which can extract temporal features of texts.In this thesis,we use bi-directional long short-term memroy architecture to pre-train the source domain data and contain the network parameters.In order to improve the effect of feature extraction in target domain data,we use the bi-directional long short-term memroy architecture with convolutional neural network model to extract the features and complete the final relation extraction task.The experimental results show that the deep learning model combined with the transfer learning method is effective for the task of relation extraction in the domains with few samples.The method proposed in this thesis can be applied to the tasks of relation discovery and automatic question answering.The task of relation extraction will bulid the foundation for further improving the knowledge graph.