Relation enhancement for noise resistance in open-world link prediction

Abstract

Open-world link prediction significantly expands the applicability of knowledge graphs by leveraging textual information to predict new entities. However, the transition from closed-world to open-world setting presents numerous new challenges, often resulting in a substantial decline in link prediction performance. It is important to investigate the causes of this decline. Through experiments assessing the impact of text noise on performance, it is found that text noise is the core factor leading to the degradation of prediction. Based on this finding, an effective anti-noise model enhanced by relation information is proposed. Firstly, a hierarchical gated fusion attention structure is designed to enhance the ability of the model to capture key semantic features by leveraging relational information in the knowledge graph, thereby significantly boosting its noise resistance. Secondly, this paper proposes a relation clustering algorithm for constructing relation specific mapping functions, which further enhances the utilization of relation information. Experimental results demonstrate that compared to similar mapping-based models, the proposed model exhibits a marked improvement in noise resistance. This method mitigates the negative impact of text noise on model performance and notably enhances the accuracy of open-world link prediction . These results show that improving noise resistance is highly consistent with improving link prediction performance in open-world scenarios. Finally, experiments show that the proposed model achieves the state-of-the-art among similar mapping-based models on two public datasets.