Eureka: Neural Insight Learning for Knowledge Graph Reasoning

Proceedings of the 29th International Conference on Computational Linguistics, pages 4517–4527 October 12–17, 2022. 4517 Eureka: Neural Insight Learning for Knowledge Graph Reasoning Alex X. Zhang1, Xun Liang1∗, Bo Wu1, Xiangping Zheng1, Sensen Zhang1, Yuhui Guo1, Jun Wang2, Xinyao Liu3 1Renmin University of China 2Swinburne University of Technology 3Hong Kong University of Science and Technology {zhangxuanalex, xliang, wubochn, xpzheng, sensen0126, yhguo}@ruc.edu.cn junwang@swin.edu.au liuxinyao575599@163.com Abstract The human recognition system has presented the remarkable ability to effortlessly learn novel knowledge from only a few trigger events based on prior knowledge, which is called insight learning. Mimicking such behavior on Knowledge Graph Reasoning (KGR) is an interesting and challenging re- search problem with many practical applica- tions. Simultaneously, existing works, such as knowledge embedding and few-shot learning models, have been limited to conducting KGR in either “seen-to-seen” or “unseen-to-unseen” scenarios. To this end, we propose a neural insight learning framework named Eureka to bridge the “seen” to “unseen” gap. Eureka is empowered to learn the seen relations with suf- ficient training triples while providing the flex- ibility of learning unseen relations given only one trigger without sacrificing its performance on seen relations. Eureka meets our expecta- tion of the model to acquire seen and unseen re- lations at no extra cost, and eliminate the need to retrain when encountering emerging unseen relations. Experimental results on two real- world datasets demonstrate that the proposed framework also outperforms various state-of- the-art baselines on datasets of both seen and unseen relations. 1 Introduction Human knowledge provides a formal understand- ing of the world. Knowledge graphs (KGs) that represent structural relations between entities in the form of (head entity, relation, tail entity) have become an increasingly popular research direction towards cognition and human-level intelligence. These triples of KGs are abbreviated using (h, r, t) in this paper. Over the last few years, the works (Bordes et al., 2013; Sun et al., 2019) on knowl- edge embedding have achieved impressive results in the knowledge graph reasoning (KGR) task. To ∗Corresponding author successfully learn a set of relations, these methods usually require a large number of training triples and cannot infer missing facts of unseen relations due to the sparse interactions, which are essentially transductive learning processes in terms of the rela- tions. Thus, we categorize this line of knowledge embedding research as “seen-to-seen” methods; i.e., reasoning from seen relations to seen rela- tions. Moreover, the representations of the rela- tions in KGs produced by knowledge embedding models always remain fixed after training. They may be sub-optimal since the real-world large-scale KGs dynamically evolve quickly with new relations emerging every day, rather than staying static. Suppose that we would like to expand the set of relations that the knowledge embedding models can recognize. We need to collect training triples for the emerging (unseen) relations; i.e., those not in the initial training set, and then restart the afore- mentioned computationally costly training proce- dure on the enhanced training set. Not to mention the fact that the model may not perform …

Content truncated to 500 words