This is the repository for our paper "End-to-End Beam Retrieval for Multi-Hop Question Answering".

Our repository is under construction, feel free to contact us if you have any questions.

Cheers! Our results have been published on MuSiQue-Ans , 2WikiMultihopQA and HotpotQA. Many thanks to the dataset authors.

We use three original datasets MuSiQue-Ans, HotpotQA and 2WikiMultihopQA for our main experiments and three paritial datasets sampled by IRCoT.

We use DeBERTa as our backbone model.

The code for our Beam Retrieval is in directory retrieval. To train our Beam Retrieval, choose the script from run_train_retr_musique.sh, run_train_beam_retr.sh, run_train_2wiki.sh, which aim at MuSiQue-Ans, HotpotQA and 2WikiMultihopQA respectively. Note that you should edit your actual url of data and model in the script.

For open domain retrieval setting, we use the data produced by MDR, and we format them in directory fullwiki, then train our Beam Retrieval using script run_train_fullwiki_reranker.

The code for the supervised downstream reader is in directory qa, while the code for LLMs is llm_exp_long.py.

All the results of retrieval and downstream reader are in directory results.

You can also obtain the scores through running test_model_tmp.py after training.