QA-GNN: Reasoning with Language Models and Knowledge Graphs for Question Answering

• https://arxiv.org/abs/2104.06378
  • code: https://github.com/michiyasunaga/qagnn
  • experiments: https://worksheets.codalab.org/worksheets/0xf215deb05edf44a2ac353c711f52a25f
• Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, Jure Leskovec

Knowledge graph,

Proposed QA GNN method. This method retrieves sub knowledge graph and then create a joint graph of the retrieved graph and the question context $z$. Then do the “reasoning” through Graph neural network, particularly Graph attention networks.

The relevance scoring of knowledge graph nodes is done by concatenating the entity to the QA context and calculate the likelihood using the pretrained LLMs. See Nangi2023dense for improved retrieval.

Cited by

• Nangi2023dense