Research On Construction And Applications Of Conditional Knowledge Graph

Knowledge has a profound impact on our production and life,and the acquisition and utilization of knowledge are the main manifestations of human intelligence.Knowledge graph(KG),as a research focus in the field of natural language processing and data mining,aims to provide the machine knowledge demand guarantees for the machine to achieve artificial intelligence,and to address issues such as intelligent question answering,reading comprehension,or automated text generation.KG takes entities and the relations between entities as the basic core elements,where the entities are presented in the form of nodes and the relations are presented in the form of edges to link entities.Finally,KG appears as a flat network graph structure.Relational entity facts are the core and only knowledge unit in the current constructed KG.However,this paper argues that conditions are essential for the verification and availability of facts,and we believe that KGs should consider conditions.On the one hand,the introduction of condition represents the factual knowledge established under conditions,and activates the verifiability and availability of facts in the KG.That will greatly promote the development of the KG research field.On the other hand,the conditions and facts are simultaneously structured and represented in the graph,so that lossless conversion of text to knowledge becomes possible.This paper first proposed Conditional Knowledge Graph,researching and designing the network representation structure of conditional knowledge graph in the background of open knowledge graph,exploring its automatic construction method,and studying its value in downstream tasks.We elaborated the research content from four aspects,with respect to the main four innovations:· Relation Inquiry Based Synchronous Joint Extractor for Entities and Relations The entity and its relationship extraction are the fundamental research in knowledge graphs,and the primary basis to explore the construction methods of conditional knowledge graphs.The existing entity and relation joint extraction methods mostly use an asynchronous framework,which tends to produce wordy intermediate redundant information,limited interaction between components,and exposure deviations from training to inference.The synchronous joint extraction framework learns entity and relationship models synchronously and outputs the extraction results synchronously,which suffers from the problem of overlapping tuples.This paper proposes a relation inquiry strategy to enhance the ability of the synchronization framework to extract overlapping triples and lead to sufficient interactive learning capabilities for entities and relational models.· Representation Design for Conditional Knowledge Graph and A Construction Approach Without considering the importance of factual conditions,the existing KGs express factual knowledge as a flat relational network of concepts,which not only loses the verifiability and conditionality integrity of KG,but reduces the availability of facts in KG.In this paper,we first propose the conditional knowledge graph and design a hierarchical network structure.Compared with the traditional flat network structure,it can simultaneously characterize facts and conditions with more flexible structures.Moreover,taking the existing research outcomes on natural language technology such as entity relationship extraction,we propose a sequence labeling model based on the multiple input facts-dual output conditions to automatically construct a condition-fact knowledge graph.· Dynamic Multi-Output based Conditional Knowledge Graph Construction Approach Our previous research has shown that problems such as overlapping tuple extractions and role assignment conflicts in fact or condition knowledge are the main challenging bottlenecks in the constructing of conditional KGs.We observe that 93.8% of sentences contain multi-fact or multi-condition tuples.It is necessary to extend the typical singleoutput sequence label to a dynamic multi-output design.It becomes a key research problem to dynamically determine the number of tuples of different roles.Considering the existing entity relation extraction results and entity relation pre-training encoder model,we further propose a dynamic multi-output sequence labeling model,which dynamically extracts multi-fact multi-condition tuples only based on the text information.Our proposed model significantly improve the construction accuracy of conditional knowledge graphs.· Conditional Knowledge Graph based Literature Search Application It is a low-loss structured representation that conditional KGs retain facts and condition information in text,which has extremely high value and significance for many natural language processing tasks.We consider the literature search as the application target of conditional KGs to explore the universal application technology of conditional KGs.Current literature search systems are mainly based on text to propose search algorithms,which is not conducive to match complex knowledge structures.We propose to adopt conditional KGs path matching and representation learning to address the literature search issues on the complex factconditional knowledge structure.In summary,this paper is proposed on the background of open KGs,exploring from the basic research on entity and relation joint extraction,to the automated accurate construction of conditional KGs.Furthermore,we focus on the application to present the real use value of the conditional KGs.We hoped that this research could provide a certain reference value for the researchers on knowledge graphs,natural language processing,and data mining.