Knowledge Graph Representation Learning Via Incorporating Entity Type Information

In recent years,the representation learning of knowledge has been widely concerned.Representation learning aims to project the entities and relations in the knowledge graph into dense and low-dimensional real-valued vectors,so as to efficiently calculate the semantic correlations between entities and relations.Besides,the learned knowledge representation can be applied to the downstream tasks such as knowledge graph completion,relation extraction and automatic question answering to improve the performance of corresponding models.With the development of knowledge representation learning,some researchers have shifted their focus from improving the model itself to other aspects,including multisource information fusion and negative sampling.This thesis focuses on these two aspects and proposes a novel and effective model.By adopting the attention mechanism to capture the information covered by the rules manually formulated in other methods,it solves the problem that some models need to introduce additional information when fusing entity type information.In this way,the model can be simplified and the universality of the model can be improved.In addition,to make full use of multi-source information and fill the gap that multi-source information is rarely used in negative sampling,we propose a method of generating high-quality negative samples by using entity types to improve model performance.The main results of this thesis are as follows:(1)This thesis proposes a knowledge representation learning model TEKRL,which merges entity type information.While introducing multi-source information,it solves the problem that other models need to introduce additional rules when using entity type information.The model constructs two entity representations based on structure and type,and captures the potential association between the entity type and the relation of triple by introducing attention mechanism,which can automatically learn the degree of correlations between different types of entities in a specific relation.In this way,the tedious process of manual rule-making is simplified,and knowledge representation learning is more efficient by using entity type information.The experimental results show that TEKRL model has made significant improvement in link prediction and triple classification tasks,especially in entity prediction tasks.Compared with other methods,the Hit@10 indicator has increased by about 7.2%,and the Mean Rank indicator has increased by about 23%,indicating that our model can effectively use type information for better knowledge representation.(2)We define the concept of entity type similarity and propose a negative sampling method based on it to improve the quality of negative samples.In this method,the type similarity is used as the basis for selecting the replacement entity,and the negative sample sequence is generated by sorting the type similarity.The high-quality sampling is realized by dividing the replacement entity sequence and setting different sampling probability for each partition area according to the type similarity.The experimental results show that the performance of the model is further improved on the basis of the original model after the addition of TENS negative sampling method.The Hit@10 indicator has been further improved by 1.2%,and the Mean Rank indicator has been further improved by 3.9%,indicating that the negative sample generation method we proposed can effectively improve the learning ability of the model.