Research On Knowledge Representation Learning Based On Multi-Class Information Fusion

Knowledge graphs can provide effective structured information and become the foundation of several intelligent applications including web search,question answering and semantic analysis.For different knowledge graphs,people need to design special algorithms to store and utilize them.With the continuous increase of knowledge,there have been numerous problems in the traditional network-based knowledge representation,such as computation inefficiency problem and data sparsity problem,etc.Thus,the knowledge graph embedding came into being.Knowledge graph embedding aims to represent the semantic information of knowledge as dense low-dimensional vectors to provide a reasonable numerical computation framework for knowledge graphs,it can greatly promote the acquisition and reasoning of knowledge,and play an important role in the tasks of similarity calculation,knowledge graph completion and automatic answering.Most existing models of knowledge graph embedding solely concentrate on fact triples of the knowledge graph,but ignore the rich information contained in the text description and images of entities.In order to make full use of the text information and the image information to build a more accurate knowledge representation,and to filter out some of the noise information,first of all,we need to fully extract the features of the text information and image information.And for this,we propose an image encoder based on attention mechanism and a text encoder based on knowledge triples.The image encoder uses Alexnet model to be the main form of the encoder,based on this,an attention mechanism is added to the encoder,which enables the image encoder to automatically select high-quality entity images and filter out the noise data in the image information.On the basis of the Text-CNN model,the text encoder based on knowledge triples adds the influence of knowledge triples,so that the text encoder can automatically extract high-quality entity related information in text description and reduce the interference of useless information in text description.After fully extracting features from the text information and the image information of entities,it is more important to combine the structured information and the unstructured information to build a more reasonable knowledge representation.For this,we propose a knowledge representation learning model based on hyperplane projection which is called EHP(entity hyperplane projection).EHP model not only pays attention to the fact triples of the knowledge graph,but also establishes the strong correlation between structured information and unstructured information including the image information and the text information through hyperplane projection.EHP model secondly restrains the knowledge representation to get more accurate results by combining these three kinds of information.In summary,we propose an image encoder based on attention mechanism and a text encoder based on knowledge triples to make the two encoders more suitable for the task of the knowledge graph embedding.Based on these,we construct a knowledge representation learning model called EHP which integrates multi category information.EHP model not only pays attention to the fact triples of the knowledge graph,but also makes full use of the rich knowledge contained in text descriptions and image information of entities,thus,the performance of knowledge representation is greatly improved.