An anchor vector is a summary representation built from the most important sentences in retrieved passages. In multi-document question answering, it concentrates fine-grained evidence before the decoder processes the remaining context, improving grounding and reducing decoder memory reads. section 5.6.1
A multi-granularity fusion process forms and uses the vector in four steps:
- Re-rank the retrieved passages. section 5.6.1
- Select the most important sentence from each top-ranked passage. section 5.6.1
- Aggregate those sentences into an anchor vector. section 5.6.1
- Prune less important passages and give the decoder the vector together with the remaining passages. section 5.6.1
Evidence-focused fusion
Suppose several retrieved passages contain mixed-quality evidence. The system keeps one important sentence from each top-ranked passage, combines those sentences into the anchor vector, removes weaker passages, and lets the decoder read a more focused context. The selected evidence also helps show which passages supported the answer. section 5.6.1
