KG-FiD addresses a common Fusion-in-Decoder failure: noisy, redundant, or misleading passages distract the decoder and leave relevant evidence unextracted. It refines passages before standard FiD encoding and decoding. section 5.6.1
Its pipeline uses three stages:
- Graph construction connects passage nodes with edges when passages are closely related, such as when they share entities. section 5.6.1
- GNN refinement updates each passage embedding by aggregating information from neighboring passages, strengthening relevant evidence and dampening noise. section 5.6.1
- Reranking and pruning use the refined embeddings to select a smaller, curated passage set for the standard FiD encoder and decoder. section 5.6.1
Entity-linked passage refinement
When retrieved passages share entities, graph edges let neighboring evidence influence passage embeddings before reranking. The decoder then receives a denser context with fewer distracting passages. section 5.6.1
