A graph neural network (GNN) refines representations by aggregating information from connected nodes. In retrieval systems, this lets related passages reinforce one another while reducing the influence of noisy or misleading evidence before generation. section 5.6.1
A GNN-based passage-refinement pipeline proceeds in three stages:
- Construct a graph whose nodes represent passages and whose edges connect closely related passages, such as passages sharing entities. section 5.6.1
- Aggregate neighboring passage information to strengthen relevant embeddings and dampen noise. section 5.6.1
- Rerank and prune the refined passages before sending a smaller, denser context to the FiD encoder and decoder. section 5.6.1
KG-FiD
KG-FiD applies a GNN as a pre-filter for multi-document question answering. By removing distracting passages before fusion, it reduces decoder workload and can improve answer quality when retrieval produces redundant or misleading context. section 5.6.1







