Answering Complex Open-Domain Questions with Multi-Hop Dense Retrieval

Authors: Wenhan Xiong, Xiang Li, Srini Iyer, Jingfei Du, Patrick Lewis, William Yang Wang, Yashar Mehdad, Scott Yih, Sebastian Riedel, Douwe Kiela, Barlas Oguz

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We propose a simple and efficient multi-hop dense retrieval approach for answering complex open-domain questions, which achieves state-of-the-art performance on two multi-hop datasets, Hotpot QA and multi-evidence FEVER. Our experiments focus on two datasets: Hotpot QA and Multi-evidence FEVER.
Researcher Affiliation Collaboration 1University of California, Santa Barbara 2University of Massachusetts Amherst Facebook AI University College London {xwhan, william}@cs.ucsb.edu, xiangl@cs.umass.edu, {sviyer, jingfeidu, plewis, mehdad, scottyih, sriedel, dkiela, barlaso}@fb.com
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes 1https://github.com/facebookresearch/multihop_dense_retrieval.
Open Datasets Yes Our experiments focus on two datasets: Hotpot QA and Multi-evidence FEVER. Hotpot QA (Yang et al., 2018) includes 113k multi-hop questions... Multi-evidence FEVER includes 20k claims from the FEVER (Thorne et al., 2018) fact verification dataset...
Dataset Splits Yes Our experiments focus on two datasets: Hotpot QA and Multi-evidence FEVER. Hotpot QA (Yang et al., 2018) includes 113k multi-hop questions... We conduct further analysis on Hotpot QA dev.
Hardware Specification Yes All the experiments are conducted on a machine with 8 32GB V100 GPUs.
Software Dependencies No The paper mentions software like 'Huggingface Transformers' and 'FAISS' but does not specify their version numbers, which is required for reproducibility.
Experiment Setup Yes Table 11: Hyperparameters of Retriever [lists learning rate 2e-5, batch size 150, etc.] Table 12: Hyperparameters of Extractive Reader (ELECTRA) [lists learning rate 5e-5, batch size 128, etc.]