KG-FiD: Infusing Knowledge Graph in Fusion-in-Decoder for Open-Domain Question Answering

• DOI: 10.48550/arxiv.2110.04330
• Publication Date: 2021-01-01
• Authors: Donghan Yu; Chenguang Zhu; Yuwei Fang; Wenhao Yu; Shuohang Wang; Xu Yi‐chong; Xiang Ren; Yiming Yang; Michael Zeng
• Keywords: Benchmark (surveying)

Current Open-Domain Question Answering (ODQA) model paradigm often contains a retrieving module and reading module. Given an input question, the predicts answer from relevant passages which are retrieved by retriever. The recent proposed Fusion-in-Decoder (FiD), is built on top of pretrained generative T5, achieves state-of-the-art performance in Although being effective, it remains constrained inefficient attention all contain lot noise. In this work, we propose novel method KG-FiD, filters noisy leveraging structural relationship among with knowledge graph. We initiate passage node embedding FiD encoder then use graph neural network (GNN) to update representation for reranking. To improve efficiency, build GNN intermediate layer output only pass few reranked into higher layers decoder generation. also apply based reranking enhance retrieval results Extensive experiments common ODQA benchmark datasets (Natural TriviaQA) demonstrate that KG-FiD can vanilla up 1.5% exact match score achieve comparable 40% computation cost.