Leveraging Graph Structure in Seq2Seq Models for Knowledge Graph Link Prediction
Researchers have developed a new framework called Graph-Augmented Sequence-to-Sequence (GA-S2S) that enhances knowledge graph link prediction. This model combines a T5-small encoder-decoder with a Relational Graph Attention Network (RGAT) to incorporate both textual entity descriptions and the underlying graph structure. By processing multi-hop relational patterns and textual information together, GA-S2S achieved a significant improvement in link prediction accuracy, showing up to a 19% relative gain on the CoDEx dataset compared to existing methods. AI
IMPACT This new framework could improve the accuracy of knowledge graph completion and reasoning tasks.