Mitigating False-Negative Contexts in Multi-Document Question Answering with Retrieval Marginalization

Authors: Ansong Ni, Matt Gardner and Pradeep Dasigi

Full Model Outputs for IIRC

We include the full model outputs on the validation and test set for IIRC in joint_retrieval_results. If you wish to use another QA model to improve over our performance, feel free to use the predicted links or retrieved contexts by our model. The jsonl file is organized as follows:

{
question: # orginal question in IIRC, 
original_paragraph: # introductory paragraph in IIRC, 
link_prediction: {
        predicted_links: # a list of the predicted links to other articles,
        gold_question_links: # a list of the gold links to other articles,
},
context_retrieval: {
        gold_link_name_sent_list: # gold text snippets from different documents, 
        predicted_link_name_set_list: # predicted text snippets from different documents,
},
qa: {
        predicted_answer_ability: # we use NumNet+, which predicts the question type first,
        predicted_answer: # the final predicted answer,
        gold_answer_type: # the ground truth answer type
        gold_answer: # a dictionary of the gold answer that NumNet+ uses
        em: # the exact match score, either 0 or 1
        f1: # the f1 score, between 0 and 1
}
}

Citation

If you use our code or model outputs, please cite:

@inproceedings{ni-etal-2021-mitigating,
    title = "Mitigating False-Negative Contexts in Multi-document Question Answering with Retrieval Marginalization",
    author = "Ni, Ansong  and
      Gardner, Matt  and
      Dasigi, Pradeep",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.497",
    doi = "10.18653/v1/2021.emnlp-main.497",
    pages = "6149--6161",
}