Triple confidence-aware encoder–decoder model for commonsense knowledge graph completion
Hongzhi Chen,Fu Zhang,Qinghui Li,Xiang Li,Yifan Ding,Daqing Zhang,Jingwei Cheng,Xing Wang
DOI: https://doi.org/10.1007/s13042-024-02378-y
2024-09-14
International Journal of Machine Learning and Cybernetics
Abstract:Commonsense knowledge is essential for performing inference and retrieval in many artificial intelligence applications, including those in natural language processing and expert system. However, a large amount of valuable commonsense knowledge exists implicitly or is missing in commonsense knowledge graphs (KGs). In this case, commonsense knowledge graph completion (CKGC) is proposed to solve this incomplete problem by inferring missing parts of commonsense triples, e.g., (? , HasPrerequisite, turn computer on ) or ( get onto web, HasPrerequisite, ?). Some existing methods attempt to learn as much entity semantic information as possible by exploiting the structural and semantic context of entities for improving the performance of CKGC. However, we found that the existing models only pay attention to entities and relations of the commonsense triples and ignore the important confidence ( weight ) information related to the commonsense triples. In this paper we innovatively introduce commonsense triple confidence into CKGC and propose a confidence-aware encoder–decoder CKGC model. In the encoding stage, we propose a method to incorporate the commonsense triple confidence into RGCN (relational graph convolutional network), so that the encoder can learn a more accurate semantic representation of a triple by considering the triple confidence constraints. Moreover, the commonsense KGs are usually sparse, because there are a large number of entities with an in-degree of 1 in the commonsense triples. Therefore, we propose to add a new relation (called similar edge) between two similar entities for compensating the sparsity of commonsense KGs. In the decoding stage, considering that entities in the commonsense triples are sentence-level entities (e.g., the tail entity turn computer on mentioned above), we propose a joint decoding model by fusing effectively the existing InteractE and ConvTransE models. Experiments show that our new model achieves better performance compared to the previous competitive models. In particular, the incorporating of the confidence of triples actually brings significant improvements to CKGC.
computer science, artificial intelligence