Combining text and graph structure for knowledge graph predictions

Knowledge graph link prediction typically relies on either text descriptions alone or flattened entity neighborhoods, losing relational structure. GA-S2S integrates a T5-small encoder-decoder with a Relational Graph Attention Network (RGAT) to jointly encode both textual features and k-hop subgraph topology around query entities. The model captures multi-hop relational patterns that sequential linearization discards. On the CoDEx dataset, it outperforms competitive Seq2Seq baselines by up to 19% in link prediction accuracy.