Leveraging Graph Structure in Seq2Seq Models for Knowledge Graph Link Prediction

In three linesGA-S2S combines a T5-small encoder-decoder with a Relational Graph Attention Network for knowledge graph link prediction. The model jointly encodes textual features and full k-hop subgraph topology around the query entity, capturing multi-hop relational patterns. On CoDEx, GA-S2S outperforms Seq2Seq baselines with up to 19% relative accuracy gain.

Summary generated by Claude — human-verified