Closes #427

[nomisto]

Closes #427

Dataset contains 8 different subset_id's (different dataset settings), each with a bigbio and source schema.

Furthermore there is an subset called mediqa_ans_all which includes all data (articles, sections, URLs of documents, all four different kinds of summaries, ...). I did not implement a bigbio schema for the all view as I think this does not make sense here. Since the bigbio schema is missing for all tests fail for subset mediqa_ans_all.

Tests:

python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_all
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_page2answer_multi_abstractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_page2answer_multi_extractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_page2answer_single_abstractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_page2answer_single_extractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_section2answer_multi_abstractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_section2answer_multi_extractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_section2answer_single_abstractive
python -m tests.test_bigbio biodatasets/mediqa_ans/mediqa_ans.py --subset_id mediqa_ans_section2answer_single_extractive

[nomisto]

This is the transformation of the source data to fit the t2t schema.
Basically the summarization works like: question + answer -> summarized_answer so for t2t schema I concatenated all interesting values with "\n" for the value of text_1.

An of example page2answer_single_abstractive:

"1_Answer4": {
        "summary": "Abetalipoproteimemia, also known as Bassen-Kornzweig syndrome, ... ",
        "articles": " Bassen-Kornzweig syndrome Abetalipoproteinemia Acanthocytosis Apolipoprotein B deficiency...",
        "question": "abetalipoproteimemia hi, I would like to know if there is any support for those suffering with abetalipoproteinemia? ...",
        "question_id": "1",
        "rating": "3-Incomplete"
}

where "1_Answer4" is answer_id above and "articles" corresponds to article["text"]

[sunnnymskang]

@nomisto In the description part, can you add information about subset_id (and mediqa_ans_all implements only source)? Confirmed that all other 8 subset id pass unit tests

[nomisto]

Hi @sunnnymskang , Sure, I've added a description to the value of _DESCRIPTION and the docstring.

[hakunanatasha]

@nomisto Can you remind me why this fits the t2t schema better than question answering? We want to merge this PR asap; it looks mostly ok.

[nomisto]

Hi @hakunanatasha , the name of this dataset is a little misleading: It is a summarization task, more specifically an answer summarization task. So the input is question + answer and the task is to generate a summarization of that answer.

[hakunanatasha]

@nomisto got it; I'll merge this later today. Sorry for the hold up. I assume since it's a summarization, the text-1/2-name are also blank as there is nothing to update here.