generate.py

import json

import pandas as pd
from Bio import Entrez
from retry import retry
from tqdm import tqdm
import dask.dataframe as dd

# provided your NIH credentials
# read from .json file
with open("credentials.json") as f:
    credentials = json.load(f)
    Entrez.email = credentials["email"]
    Entrez.api_key = credentials["api_key"]


# change output file names here if necessary
RAW_EVALUATION_DATASET = "./raw_data/training11b.json"
PATH_TO_PASSAGE_DATASET = "./data/passages.parquet"
PATH_TO_EVALUATION_DATASET = "./data/test.parquet"

# only use questions that have at most MAX_PASSAGES passages to control the size of the dataset
# set to None to use all questions
MAX_PASSAGES = None


@retry()
def get_abstract(passage_id):
    with Entrez.efetch(
        db="pubmed", id=passage_id, rettype="abstract", retmode="text"
    ) as response:
        # get only the abstract - no metadata
        r = response.read()
        r = r.split("\n\n")
        abstract = max(r, key=len)
        return abstract


if __name__ == "__main__":
    # load the training data containing the questions, answers and the ids of relevant passages
    # but lacks the actual passages
    with open(RAW_EVALUATION_DATASET) as f:
        eval_data = json.load(f)["questions"]

    eval_df = pd.DataFrame(eval_data, columns=["body", "documents", "ideal_answer"])
    eval_df = eval_df.rename(
        columns={
            "body": "question",
            "documents": "relevant_passage_ids",
            "ideal_answer": "answer",
        }
    )
    eval_df.answer = eval_df.answer.apply(lambda x: x[0])
    # get abstract id from url
    eval_df.relevant_passage_ids = eval_df.relevant_passage_ids.apply(
        lambda x: [int(url.split("/")[-1]) for url in x]
    )
    if MAX_PASSAGES:
        eval_df["passage_count"] = eval_df.relevant_passage_ids.apply(lambda x: len(x))
        eval_df = eval_df.drop(columns=["passage_count"])

    # remove duplicate passage ids
    eval_df.relevant_passage_ids = eval_df.relevant_passage_ids.apply(lambda x: set(x))
    eval_df.relevant_passage_ids = eval_df.relevant_passage_ids.apply(lambda x: list(x))

    # get all passage ids that are relevant
    passage_ids = set().union(*eval_df.relevant_passage_ids)
    passage_ids = list(passage_ids)
    passages = pd.DataFrame(index=passage_ids)

    for i, passage_id in enumerate(tqdm(passages.index)):
        passages.loc[passage_id, "passage"] = get_abstract(passage_id)

        # intermediate save
        if i % 1000 == 0:
            passages.index.name = "id"
            dd.from_pandas(passages, npartitions=1).to_parquet(PATH_TO_PASSAGE_DATASET)


    # filter out the passages whos pmids (pubmed ids) where not available
    unavailable_passages = passages[passages["passage"] == "1. "]
    passages = passages[passages["passage"] != "1. "]
    passages.index.name = "id"
    dd.from_pandas(passages, npartitions=1).to_parquet(PATH_TO_PASSAGE_DATASET)

    # remove passages from evaluation dataset whose abstract could not be retrieved from pubmed website
    unavailable_ids = unavailable_passages.index.tolist()
    eval_df["relevant_passage_ids"] = eval_df["relevant_passage_ids"].apply(
        lambda x: [i for i in x if i not in unavailable_ids]
    )
    eval_df.index.name = "id"
    eval_df = eval_df[["question", "answer", "relevant_passage_ids"]]
    dd.from_pandas(eval_df, npartitions=1).to_parquet(PATH_TO_EVALUATION_DATASET)