Complex Multi-fold Relation Embedding And Text Knowledge Enhancement Based On Knowledge Graph

The rapid development of the Internet has given rise to massive amounts of network data.In order to effectively understand and utilize the key information in these diverse,semantically complex and heterogeneous data,people resort to knowledge graph(KG)to provide intelligent services via organizing and managing associated knowledge.As an important part of artificial intelligence,KG describes entities,concepts and their relations in real life in the form of graph structure.The knowledge is open interconnected and shows strong ability of semantic reasoning,which can be used as the external repository in scenarios like intelligent customer service,intelligent search,disease diagnosis and intelligent marking,producing important application values.However,there are problems in the existing construction and application of KG: First of all,knowledge is usually represented in triplet and lack of rational representation and embedding methods for complex multi-fold relations,which is difficult to completely express the complex multi-fold associated knowledge such as events and programs in the real world;Secondly,there is no effective method to utilize background knowledge in structured knowledge graph to assist natural language processing tasks,the symbolic representation of KG is inapppropriate for vector calculation.The method based on symbolic query and distant supervision of knowledge graph is limited for accurate semantic calculation and inference,which hinders the application and popularization of knowledge graph in natural language processing and other fields.Based on the above problems,indepth research is carried out from the following four aspects:The instance aggregation effect in multi-fold relational knowledge graph representation learning is studied.An instance group-constrained optimization strategy and a knowledge graph multi-fold relation embedding method based on this strategy were proposed to map the entities and fact nodes of multi-fold relation instances from entity space to relation space.The influence of different distance constraint methods on multi-fold relation mapping results was studied.Experimental results on multiple datasets show that compared with the existing multi-fold relation representation methods,this strategy can effectively model the clustering effect of multi-fold relation instances,achieve better performance in link prediction and instance classification tasks and has consistent enhancement effect on different multi-fold relation embedding models extended from binary ones.The characteristics of heterogeneous knowledge association in multi-fold relation knowledge graph representation learning are studied.This thesis puts forward a kind of knowledge graph multi-fold relation embedding method based on the facts associated modeling.The multiple instances of the entities,relations,facts are mapped to different low dimensional vector space.It also models the entity-relation,entity-fact and relation-fact associations respectively to form a combinatorial optimization problem.Finally,the stochastic gradient descent algorithm is exploited for problem solving.Experimental results on two datasets on multi-fold relations and four datasets on binary relations show that compared with the existing multi-fold relation embedding methods,the proposed method can effectively model the fact correlation characteristics within multi-fold relation instances,and the model is more robust and scalable than the existing methods.The problem of few-shot text event extraction with low resources based on knowledge graph transfer is studied.In the absence of annotated corpora,we model the structured event knowledge in the knowledge graph and the candidate event mention in the aligned text corpora,and use transfer learning method to deduce and distinguish the candidate event category in the free text based on the semantic structure information of the event in the knowledge graph.Empirical analysis results on ACE-2005 and MAVEN datasets show that this method can not only improve the classification accuracy of seen events,but also significantly improve the performance of unseen event extraction under the condition of few or even zero samples.The topic control problem of automatic text generation utilizing external knowledge graph is studied.A framework for automatic discourse topic text generation based on knowledge graph enhancement is proposed.In this framework,topic description statements are represented by joint embedding in combination with knowledge graph,and then writing topic planning is generated through semantic space search.Finally,an improved topic controllable text generation model is proposed to generate each paragraph text based on keywords in each cluster.The model proposes specific improvement strategies for the benchmark model of text generation from four aspects: text topic distribution,attention scoring,commonsense knowledge search and topic word coverage generation,and proves the effectiveness of the proposed method through experiments,which solves the problem of unsupervised topic controllable text generation.In summary,the proposed multi-fold relation representation learning method of and text knowledge enhancement method in this thesis make full use of the structural knowledge correlation characteristics of knowledge graph,provides diverse and heterogeneous knowledge modeling,knowledge representation and reasoning,text event extraction and topic controllable text generation under the condition of low resources,solving the difficulties in popularizing large-scale knowledge graph.