Research On Information Extraction Techniques For Knowledge Graph Construction In Marine Industry

As a subversive information technology,knowledge graph has strong semantic processing ability and information interconnection ability.Therefore,the construction of marine industry knowledge mapping not only can provide the core driving force for the development of marine economy,but also contribute to the sustainable development of marine economy.Since the most important problem in the construction of knowledge graph of Marine industry is the acquisition of data,and the unstructured text contains a large amount of knowledge,information extraction from the unstructured text is an important step in the construction of knowledge graph.The key technologies of information extraction are named entity recognition and relationship extraction,but there are still many problems to be solved.The first is how to avoid different types of features from interfering with each other in the neural network in the task of named entity recognition.Then it is how to extract the entity knowledge in the text of the marine industry field.Finally,how to avoid adding external language tools to build graph structures when using GCN in relation extraction tasks.Therefore,for these three problems,we propose a new data set and improve the existing algorithm.The research content mainly includes the following two points.(1)In order to solve the problem that different types of features interfere with each other in neural network,we propose the multi-channel named entity recognition algorithm(MCNER).The algorithm uses multi-channel embedding to capture different types of features of the input,and proposes a multi-channel Bi GRU network to avoid mutual interference between channels.In order to integrate features,we also propose intra-channel and inter-channel attention.It can assign different weights to the features in each channel according to their importance degree,and dynamically integrate the features between each channel also according to their importance degree.The integrated features are input to the CRF layer to classify the entity category of each character.(2)Because the graph structure of GCN encoding usually needs to be constructed by adding external language tools,it will lead to problems such as large amount of calculation,unable to carry out end-to-end training and not suitable for knowledge extraction in professional fields.After the algorithm performs feature extraction on the sequence,self-attention is used in each layer of DGCN to infer the graph structure without the need to use external language tools.After inferring the graph structure,GCN is then used for encoding.Finally,all the layers of GCN are used to represent the input fully connected layer for feature dimensionality reduction,so as to realize the relation classification between entitiesIn this paper,named entity recognition and relation extraction experiments were carried out on the Chinese and English data sets to verify the generalization performance of the model.For named entity recognition,experiments have proved that the multi-channel named entity recognition algorithm can effectively avoid the mutual interference between different types of features,and has good classification performance in the marine industry data set and the Co NLL-2003 data set.For relation extraction,experiments have proved that the dynamic graph convolutional neural network relation extraction algorithm does not need to add external language tools to construct the graph structure,and it has good classification performance in the San Wen dataset and the Sem Eval-2010 Task 8 dataset.However,the information extraction framework of this article is a pipeline model,which will also lead to the transmission of errors.In the future,we need to further study the model of entity and relation jointly extraction.