Research On Some Issues Of Knowledge Reasoning For Knowledge Graphs

With the dramatic development of mobile internet,unstructured data including text,images,audio,and video ushered in explosive growth.Meanwhile,these data are massive,heterogeneous,diverse,and unorganized,which brings huge challenges for people to acquire and utilize the knowledge contained in them.Knowledge graph(KG),also known as a large-scale semantic network,has irreplaceable advantages in knowledge organization and efficient utilization.It lays the foundation for the structured storage and intelligent application of massive data,which has become the cornerstone of information services in the era of artificial intelligence.However,limited by the accuracy of existing information extraction and construction algorithms,KGs are usually incomplete or even incorrect,which impedes its large-scale implementation.Knowledge reasoning technology that can infer missing elements utilizing existing entities or relations in the KG and evaluate the credibility of triplets provides new ideas for resolving the above-mentioned issues.In recent years,deep learning-based reasoning methods that obtain embeddings of entities and relations in the KG with the help of knowledge representation learning models,and perform inference tasks through numerical operations between vectors have become the mainstream techniques,because these methods have the advantages of simplicity,efficiency,and computability,which better meets the reasoning requirements of large-scale KGs.Unfortunately,many limitations still exist with regard to reasoning due to the multi-source heterogeneity and dynamic openness of KGs.Therefore,it’s important to study knowledge reasoning techniques for KGs.Through careful analysis of the limitations and drawbacks of existing methods,this dissertation conducts an in-depth study of reasoning issues in scenarios such as multi-source information fusion,dynamic temporal knowledge,openworld assumptions,and multimodal KGs.The main contributions of this dissertation are as follows:1.A knowledge reasoning model combining entity description and hierarchical type information is proposed.KGs not only contain a large number of facts in the form of triplets but also come with entity description and hierarchical type information.These data,serving as an effective complement to the structural information of triplets,can help the model describe entities and relations more precisely,thus improving the accuracy rate of reasoning methods.However,existing models focus excessively on the structural information of triplets and neglect the utilization of this auxiliary information.To address this issue,this dissertation proposes a knowledge reasoning method that integrates entity description and hierarchical type information.The model first uses two different neural networks and hierarchical type encoders to construct description-based representations and type-based representations for entities,and then trains them together with structure-based representations to learn entities and relations embeddings,thereby providing prior knowledge for reasoning models.Moreover,to further improve the accuracy of the knowledge reasoning,the model introduces a type restriction strategy in both training and testing phases.Experimental results show that the proposed method can fully excavate the semantic information implied in entity descriptions and hierarchical types,which helps the model handle complex relations and improve the effectiveness of knowledge reasoning.2.A reasoning model with temporal consistency constraints for temporal knowledge graphs(TKGs)is proposed.TKGs such as YAGO and Wikidata contain a large number of time-sensitive facts,and these triplets hold only at specific time intervals or points.However,existing reasoning methods mostly assume that entities and relations in KGs are static,ignoring the impact of time on knowledge evolution.Although the improved models have taken time factors into account,they only model the chronological order of relations and do not encode temporal information directly.In response to this problem,this dissertation proposes a reasoning method over TKG that combines time consistency constraints,where the model uses a long short-term memory network to train the concatenate sequence of relation and timestamp for learning time-aware relation embedding.Further more,it introduces three kinds of time consistency constraints and transforms the knowledge reasoning task into an optimization problem with constraints.Then we use an integer linear programming method to solve this problem and filter out unreasonable candidate results.Experimental results indicate that the proposed model can effectively describe the dynamic evolution of TKG and improve the credibility of prediction results.3.A reasoning model based on neighborhood aggregation using graph attention mechanism for open-world KG is proposed.Large KGs evolve quickly with new facts being added quickly.In this process,some facts will be removed when invalid,and new entities will be connected with existing facts.In fact,these previously unseen entities in KGs should not be considered as wrong instances,which is the alleged open-world assumption.Unfortunately,existing reasoning models mostly operate under the closedworld assumption in which all entities need be observed in the training data.When new entities or relations are added to the KG,repeatedly re-training the whole KG is impractical for a large-scale knowledge base with billions of nodes.To tackle this issue,this dissertation proposes a reasoning method that aggregates neighborhood information using graph attention networks for open-world KG.The model utilizes graph neural networks to aggregate neighborhood information and takes relation features into account.It is worth mentioning that our model builds representations for out-of-KG entities without using external resources and re-training the entire KG.Experimental results indicate that the graph attention mechanism helps our model enable specifying different weights to different nodes,which improves the expressive ability of the proposed method.4.A reasoning method based on knowledge-aware pre-trained mechanism for multimodal KG is proposed.In recent years,several multi-modal KGs have been constructed in academia and industry.Multi-modal KGs create images for entities and add links between them based on traditional KGs.Therefore,how to fuse heterogeneous text and image information becomes the primary challenge for reasoning over multi-modal KGs.Existing methods simply view entity images as a complement to textual information and ignore the deep interaction of multimodal information.In addition,the process of modeling textual information is still troubled by the polysemy phenomenon.Concerning the issues outlined above,this dissertation proposes a method based on knowledge-aware pre-trained mechanism for reasoning over multi-modal KG.The model adopts a knowledge-enriched pre-training language model as the backbone and utilizes a multi-modal knowledge encoder to achieve deep fusion of multimodal information such as textual descriptions,knowledge embeddings,and entity images.Experimental results show that the proposed method can effectively extract features of different modalities and provide richer semantic knowledge for reasoning tasks.In summary,this dissertation conducts research on some issues of knowledge reasoning for KGs and proposes practical and feasible solutions,which is important for improving the trustworthiness of knowledge reasoning and promoting the further application of KGs.