ConNR: A continual N-ary knowledge reasoner for growing N-ary knowledge graphs

N-ary Knowledge Graphs (NKGs), containing facts with more than two entities, are prevalent in the real world. Link Prediction in NKGs (LPN) is to predict missing elements in facts therein, which helps populate and enrich the NKGs. Existing LPN methods implicitly assume that NKGs remain static. However, real-world NKGs usually grow, frequently introducing new facts over time. When new facts emerge, existing LPN methods need to be fine-tuned with new data or retrained on the entire dataset, often resulting in knowledge forgetting or requiring a lot of training time. Consequently, this paper introduces a new task, i.e., Continual Link Prediction in growing NKGs (CLPN), aiming to predict missing elements in facts in growing NKGs. To conduct CLPN, we propose a Continual N-ary knowledge Reasoner, called ConNR. ConNR includes an embedding generation module to initialize embeddings of the unseen elements, a embedding update module to globally update embeddings of all elements, and a fact decoder that captures the semantics of new facts and predicts missing elements. To evaluate CLPN, we have carefully constructed five datasets with different NKG growth scenarios. Experimental results on all these datasets demonstrate the superior performance of ConNR over existing representative models.