Knowledge Representation Learning And Application Based On Multi-source Information

With the fast development of Internet technology and AI technology,knowledge graph bas been widely concerned by major enterprises and research institutions as an important basic technology of AI.However,with the scale of knowledge graph growing rapidly,traditional symbolic knowledge representation which regard nodes as entities and edges as relations is suffering from low computational efficiency and poor scalability.To solve these problems,knowledge representation learning which embeds the entities and relations into the low dimensional semantic vector space emerges.Represented by Trans E model,knowledge representation learning not only shows its high efficiency in knowledge graph inference and semantic similarity computing,which significantly increases the performance of knowledge graph merging,complement and inference.But also easy to be utilized in other areas.However,traditional knowledge graph learning focuses on entities and relations' structure information,and thus it is difficult for these methods to solve complex relational modeling problem and data sparsity problem very well.In fact,external texts beyond knowledge graph structure contain lots of information and thus are able to extend knowledge graph structure and alleviate data sparsity problem effectively.How to utilize these texts to improve the performance of knowledge representation learning has become a popular research.In another respect,the success of knowledge representation learning in QA system and recommender system shows the knowledge graph structure information can serve as prior knowledge in other areas,including constraints and expansions.But it is difficult to incorporate knowledge graph into traditional machine learning tasks such as Click Through Rate Prediction Task.It has become a popular research issues recently how to efficiently solve this problem.This paper makes an in-depth study on the problems mentioned above,and the main works and contributions are as follows:1.We investigate recent knowledge representation learning models which integrate external text information.Aiming at solving the problems of these models,this paper proposes the ETRL model(Entity Topic based Representation Learning).ETRL model studies the entity information from entity descriptions of knowledge graph and integrates with the structure information by topic model.Experiments show that,ETRL model can effectively alleviate the complex relations modeling problem and expand the semantic of entities and relations by merging the entity topic information.Moreover,ETRL model performs better than other models in knowledge graph completion task.2.This paper tries to utilize knowledge representation learning in Click Through Rate prediction task which is suffering from data sparsity and high-level interaction features problem.This paper proposes KSDeep FM model(Knowledge Structure based Deep Factorization Machines)which integrates the knowledge graph structure information modeling by Trans R in Deep FM model by distance supervising.Experiments show that,the knowledge structure information can alleviate data sparsity and feature learning problems in CTR prediction task and improve the performance of the model of CTR prediction.3.This paper designs and implements a knowledge graph management system KGPlus(Knowledge Graph Plus System),which can store knowledge graphs and implements the knowledge graph cleaning,completion,entity extending query,and so on.This system can reduce the cost of manual cleaning and has a strong application value.