Generation And Application Of Entity Descriptions Based On Large-scale Knowledge Base

Knowledge graphs gain its importance to the industry applications and is meanwhile a significant infrastructure in the field of artificial intelligence.There are many facts in the knowledge graphs or knowledge bases,which mainly consist of an enormous number of entities and the relations among them.Besides,knowledge graphs will provide an additional description sentence briefly introducing each entity.These entity descriptions could directly present the basic information of entities to the end uses and are thus widely applied in the industry.This paper proposed a method to generate entity description automatically based on the existing knowledge bases,which conducted a knowledge completion for the graph,generated the text using an end-to-end neural network model,and finally presented a method to incorporate entity descriptions in the question answering system.Knowledge graph completion involves the representation learning methods,which learns the low-dimensional representations of entities and relations.Existing work had been mostly about designing discriminative models,while this paper proposed a new method to combining discriminative and generative models under the newly-proposed adversarial training framework.The performance was further improved upon existing discriminative models.After the completion of knowledge bases,this paper proposed an end-to-end neural network model by adopting the encoder-decoder framework.Both the encoder and decoder are neural networks here,and the decoder further adopted attention mechanism to model the correlation between each word of the entity description and structural data in knowledge graph.Using this way did the model jointly learn two important steps of entity description generation,the content selection and surface realization.Furthermore,this paper discussed the importance of multi-hop facts on the knowledge bases.The model encoded multi-hop facts in knowledge bases and improved the fluency and the evaluation results.Eventually,the question answering system is one of the typical applications of knowledge bases.To show the effect of entity descriptions,this paper presented a method to use entity descriptions in the question answering system.First this paper constructed a question answering system targeting real situations by two main means,the template matching and semantic parsing,to adapt to different user questions.Entity descriptions can be embedded in the existing framework as a ranking feature.This paper designed a neural matching network,which is implemented as a micro-service to model the text similarity and give a score for each entity against a single question.