Temporal Knowledge Graph Reasoning with Dual Guiding Constraints

Temporal knowledge graphs (TKGs) can store and model dynamic relational information. Recently, research on generative models has addressed the limitations of embedding methods in terms of generality and scalability. However, generative models still face two challenges in the context of TKGs: (1) Inability to effectively handle a large volume of contextual information;(2) Inability to constrain the rationality of generated results. To this end, we propose a new temporal graph reasoning framework, called the Dual Guiding constrained generative model (DuGen), which primarily consists of a context summarizer and a generative reasoner. The former generates crucial contextual information using question-guided prompts, while the latter employs a type decoder to infer the credibility of candidate entities. DuGen maintains the integrity and flexibility of individual model structures through external nesting. Experimental results demonstrate that DuGen outperforms state-of-the-art baseline techniques.