Research On Reasoning Methods Based On Knowledge Graph Represent Learning

Knowledge reasoning aims to find results or sufficient paths with a certain path or an entity pair given,which has been widely studied recently.And it has giant importance to searching in closed field,knowledge graph completion,questioning and answering in opened field and indexing problems.Though there were lots of work about logical reasoning and one-hop rea-soning which were proved effective and widely used,it is still difficult for models in multi-hop reasoning with long paths and complex relation.This thesis proposes a knowledge graph repre-sentation learning model based on hyperbolic space to learn hierarchical structure,which is able to embed entities and relations in knowledge graph into low-dimension space.Then,hyperbolic embeddings and additional action space are used to ensure the sufficiency in reasoning and help agents in reinforcement learning framework to search for paths efficiently.Lastly,hierarchical information are engaged to promote accuracy and sufficiency of agents and models.The main contributions of this thesis includes:1)A knowledge graph representation learning method based on hyperbolic space,HBE,is proposed:HBE embeds entities of knowledge graph in tangent space into hyperbolic space to learn hierarchical structure in Euclidean space,which contributes to embeddings.Meanwhile,a method similar to polar coordinate transformation is proposed to reduce the problem of insufficient precision of floating-point numbers in calculation.At the same time,the non-Euclidean hypersphere and constraint term are used to deal with the bound-ary condition and to learn symmetry and other properties in relations.In addition,the problem of inner product and addition in hyperbolic space is dealt with angle transforma-tion approximation to optimize parameters in hyperbolic space which is similar to ones in Euclidean space.2)A multi-hop reasoning model based on reinforcement learning,PAAR,is proposed:PAAR aims to learn multi-hop reasoning in knowledge graph with reinforcement learning.And hyperbolic embeddings are and additional action space are used to restrict agents' actions into narrow limits,which guarantees the sufficiency of reasoning.Compared with other models with reinforcement learning,PAAR outperforms the others and utilizes sufficient constraint and hierarchical information from hyperbolic embeddings better.