Gender Bias Analysis in Commonsense Knowledge Graph Embeddings

In this study, we explore the existence of gender bias in ConceptNet triples by employing knowledge graph embedding models, a crucial aspect that has gained prominence in the artificial intelligence domain. Commonsense knowledge graphs, such as ConceptNet, play a vital role in an array of natural language processing tasks; nonetheless, the potential gender bias embedded within these graphs may considerably influence the fairness of associated applications. To tackle this concern, we introduce a comprehensive gender bias analysis framework tailored for ConceptNet triples, with a focus on the link prediction task utilizing knowledge graph embedding models. By evaluating multiple state-of-the-art knowledge graph embedding models, we assess the degree of inherent gender bias across the models. Furthermore, we investigate mitigation strategies to attain a debiased representation of gender within commonsense knowledge graphs. Our results contribute to a broader comprehension of gender bias in knowledge graphs, and the proposed mitigation strategies lay the groundwork for developing more equitable AI systems that use ConceptNet.