TL;DR

UniKGQA collapses KG subgraph retrieval and answer reasoning into a single UniModel with shared PLM parameters, outperforming pipeline baselines on WebQSP, CWQ, and MetaQA through joint end-to-end training.

Summary

Multi-hop knowledge graph question answering (KGQA) traditionally separates subgraph retrieval and answer reasoning into a pipeline, causing error propagation between stages. UniKGQA collapses both into a single UniModel built on a pretrained language model that shares parameters across retrieval (finding relevant KG paths) and reasoning (reading paths to produce answers). Joint end-to-end training allows retrieval signals to inform reasoning and vice versa. On WebQSP, CWQ, and MetaQA benchmarks, UniKGQA outperforms pipeline and retrieval-only baselines, confirming that joint optimization of retrieval and reasoning is superior to staged approaches.

Key contributions

  1. Proposes UniModel, a single PLM-based model that shares parameters across subgraph retrieval and answer reasoning for KGQA.
  2. Eliminates error propagation between retrieval and reasoning stages through joint end-to-end training.
  3. Frames subgraph retrieval as path scoring, enabling differentiable integration with the reasoning component.
  4. Outperforms pipeline baselines on WebQSP, CWQ, and MetaQA multi-hop benchmarks with a smaller overall model footprint.

When to cite

  1. When arguing that unified retrieve-and-reason models outperform modular pipelines for multi-hop KGQA.
  2. When discussing joint training as preferable to staged approaches when retrieval and reasoning have complementary supervision signals.
  3. When citing evidence that KG path scoring can be formulated as a sequence task compatible with PLM fine-tuning.
  4. When designing systems where multi-hop reasoning benefits from retrieving full reasoning chains rather than individual KG triples.

LLM KG Knowledge Graph RAG