The Algorithm And Technology Research Of Knowledge Graph Based On Deep Learning

The information age produces huge data,and the forward technique of AI improves the user's application requirements of data.Knowledge Graph(KG)can give artificial intelligence the ability of in-depth understanding,which makes the intelligence of the network closer to human's cognitive thinking.The language element of knowledge graph is entity that can convey information,so the techniques used to create entities can impact the quality of the knowledge graph.Entities in the KG can be extracted using named entity recognition.Disambiguation technology can make the meaning of the entity more accurately.The accuracy of entity can improve the accuracy of KG.At present,there are two main challenges facing Chinese knowledge graph.One is that entity boundary is difficult to identify in Chinese plain text information,which leads to high difficulty in entity extraction.Another is that in short text,the entity in the input text cannot be accurately related to the entity with the correct meaning of knowledge base due to insufficient feature information in context.In view of these two challenges,this paper determined the research direction,which can enhance the information features extracted,increasing the model's accuracy and recall rate.This thesis proposes an entity recognition model based on multi-granularity vector.Moreover,the model can solve the problem of difficult extraction of entities from pure text information.In addition,the feature enhancement method was used to improve the feature information of the entity in the entity disambiguation model.The accuracy rate of the established entity disambiguation model in the entity disambiguation of short texts reached 91.14%,and the recall rate was over 0.88%.The entity recognition model studied in this paper takes BI-LSTM and CRF as the network framework and combines the self-attention mechanism to complete entity recognition.On the basis of this research,Multi-granularity vector model is put forward to further enrich the semantic features of the text.The model uses a multi-granular vector method to combine word vectors with word vectors as input.Using BERT-wwm for pre-training,feature vectors acquired from pre-training input to the Bi-LSTM,which can fully consider context information.At the same time,introducing a self-attention mechanism,focusing on local features,CRF obtains the labeling result of entity recognition.Experiments show that the recognition effect of this model gets better.Building an entity disambiguation model uses contextual features based on deep learning.When input is the text to be disambiguated,the sequence features of the text to be disambiguated are extracted using a bidirectional LSTM network.Extracting the first word vector,last word vector,maximum pooling,and Self-Attention of the entity sequence in the text to be disambiguated,then weighted summation after splicing the obtained features to acquire the feature vector in the text to be disambiguated.Retrieve the set of candidate entities using the dictionary and enter the description text for the candidate entities,using BI-LSTM network to obtain the input text sequence.Then,attention mechanism can further extract semantic information,finally obtaining the feature vector describing the text.By calculating the similarity of entity features,the output of entity disambiguation is obtained.The laboratory findings show that the optimized entity disambiguation model has a high accuracy 91.14%in disambiguation of Chinese short text data with limited context features.