NEWS · RESEARCH · #132
FedV-KGQA in Practice: Design Lessons and an Interactive Prototype
This arXiv poster reports empirical results for FedV-KGQA, a federated approach to multi-hop question answering over vertically partitioned knowledge graphs where organizations share entity identifiers but keep disjoint relation types. Each silo trains local KG embeddings and a server concatenates silo-specific entity views, anchors the question at a topic entity, and ranks candidates without exchanging raw triples or relation embeddings; experiments show federated fusion recovers most centralized accuracy, anchoring and enrichment matter more than embedding choice, and the cheapest encoder depends on target accuracy; the paper also provides four design lessons and an interactive prototype with released checkpoints.
KEY POINTS
- This arXiv poster reports empirical results for FedV-KGQA, a federated approach to multi-hop question answering over vertically partitioned knowledge graphs where organizations share entity identifiers but keep disjoint relation types.
- Each silo trains local KG embeddings and a server concatenates silo-specific entity views, anchors the question at a topic entity, and ranks candidates without exchanging raw triples or relation embeddings; experiments show federated fusion recovers most centralized accuracy, anchoring and enrichment matter more than embedding choice, and the cheapest encoder depends on target accuracy; the paper also provides four design lessons and an interactive prototype with released checkpoints.
- This work provides practical empirical evidence and design guidance for privacy-preserving, multi-party KG question answering, showing federated fusion can approach centralized performance without sharing raw triples.
WHY IT MATTERS
This work provides practical empirical evidence and design guidance for privacy-preserving, multi-party KG question answering, showing federated fusion can approach centralized performance without sharing raw triples.