A pytorch implementation of our ICASSP2023 paper
From Easy to Hard: Two-stage Selector and Reader for Multi-hop Question Answering
Xin-Yi Li, Wei-Jun Lei, Yu-Bin Yang Accepted by ICASSP 2023
This repo is still under construction. Please feel free to contact us if you have any questions.
Our result has been published on HotpotQA Leaderboard.
Run the script to download the data, including HotpotQA data and spacy packages.
sh scripts/download.sh
Preprocess the training and dev sets in the distractor setting:
sh scripts/run_preprocess.sh
The code for our selector is in directory "selector_src". To train our two-stage selector and infer on the dev set:
sh scripts/run_selector_large.sh
The code for our reader is in directory "reader_src". To train the reader with our two-stage strategy on ALBERT, and evaluate the model on the dev set:
sh scripts/run_reader_large.sh