Research On Temporal Knowledge Graph Completion And Question Answering Methods

Knowledge Graph(KG),as a semantic network-based knowledge representation model,aims to integrate a large amount of data on the Internet into a massive knowledge base,establish connections between data,and create a structured representation of data.By integrating diverse data,a more comprehensive,accurate,and in-depth knowledge representation model can be constructed,improving the efficiency of information retrieval and usage.Traditional KGs ignore the temporal information of facts and do not consider that some facts may change over time.Therefore,Temporal Knowledge Graph(TKG)extends the traditional KG by attaching a timestamp to each fact,which can be a point or a period,to represent the occurrence time or validity period of the fact.Like the traditional KG,TKG is also incomplete,and Temporal Knowledge Graph Completion aims to predict missing facts in TKG.Most existing methods treat TKG as a set of independent facts,ignoring the implicit correlations between them.In fact,as a dynamic heterogeneous graph,TKG’s graph structure can reflect rich information.Therefore,this paper proposes a Temporal Knowledge Graph Completion method based on Comp GCN,which incorporates timestamp into entity and relation representations,divides TKG into time intervals,learns graph structure evolution features of different time intervals through GRU components,and obtains a vectorized representation of TKG.Finally,a decoder based on convolutional neural networks is used to achieve the link prediction task.Experimental results show that the proposed TKG completion method has competitive performance on both temporal inference and temporal extrapolation tasks.Question Answering over Temporal Knowledge Graph(TKGQA),as a downstream task of TKG completion,has also received much attention from researchers in recent years.Current research assumes that the question is fully annotated before being input to the system and regards the question answering as a link prediction task.In addition,since the reasoning process is carried out in the vector space,the answer prediction process is not interpretable.Therefore,this paper proposes a method called AE-TQ based on semantic parsing,which uses Abstract Meaning Representation(AMR)to understand complex temporal questions and generates problem-oriented semantic information for explicit and effective temporal reasoning.We evaluate our method on the largest TKGQA dataset Cron Questions,and experimental results show that AE-TQ outperforms several competitive methods under various settings.