naturalqs.py

"""
Natural Questions: a Benchmark for Question Answering Research
https://storage.googleapis.com/pub-tools-public-publication-data/pdf/1f7b46b5378d757553d3e92ead36bda2e4254244.pdf

The Natural Questions (NQ) corpus is a question-answering dataset that contains
questions from real users and requires QA systems to read and comprehend an entire
Wikipedia article that may or may not contain the answer to the question. The
inclusion of real user questions, and the requirement that solutions should read
an entire page to find the answer, cause NQ to be a more realistic and challenging
task than prior QA datasets.

TODO: NaturalQS has a *really* large train set that huggingface just automatically
downloads even if you dont use it. we should try and only download the val set and
not even bother with the train set. 

Homepage: https://ai.google.com/research/NaturalQuestions
"""
import random
from . common import HFTask
from itertools import islice


_CITATION = """
@article{47761,
    title={Natural Questions: a Benchmark for Question Answering Research},
    author={Tom Kwiatkowski and Jennimaria Palomaki and Olivia Redfield and Michael Collins and Ankur Parikh and Chris Alberti and Danielle Epstein and Illia Polosukhin and Matthew Kelcey and Jacob Devlin and Kenton Lee and Kristina N. Toutanova and Llion Jones and Ming-Wei Chang and Andrew Dai and Jakob Uszkoreit and Quoc Le and Slav Petrov},
    year={2019},
    journal={Transactions of the Association of Computational Linguistics}
}
"""


class NaturalQs(HFTask):
    VERSION = 0
    DATASET_PATH = "natural_questions"
    DATASET_NAME = None

    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
        return False

    def training_docs(self):
        # Cache training for faster few-shot.
        # Data is too large to fit in memory.
        return self.data["train"]

    def fewshot_examples(self, k, rnd):
        # Data is too large to fit in memory. We just sample from the first bit.
        if self._training_docs is None:
            self._training_docs = list(islice(self.training_docs(), 0, 100000))

        return rnd.sample(self._training_docs, k)

    def doc_to_text(self, doc):
        return 'Q: ' + doc['question']['text'] + '\n\n' + 'A:'

    def doc_to_target(self, doc):
        # There's a short answer and a long answer. Based on the paper, I'm using the long answer.
        short_answer = doc['annotations']['short_answers'][0]['text']
        long_answer_start = doc['annotations']['long_answer'][0]['start_token']
        long_answer_end = doc['annotations']['long_answer'][0]['end_token']
        long_answer_span = doc['document']['tokens']['token'][long_answer_start:long_answer_end]
        long_answer_is_html = doc['document']['tokens']['is_html'][long_answer_start:long_answer_end]
        long_answer_chars = [tok for (tok, is_html) in zip(long_answer_span, long_answer_is_html) if not is_html]
        long_answer = " ".join(long_answer_chars)
        return long_answer # Replace with short_answer[0] for short answer

    def construct_requests(self, doc, ctx):
        """ Uses RequestFactory to construct Requests and returns an iterable of 
        Requests which will be sent to the LM.

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
            The context string, generated by fewshot_context. This includes the natural 
            language description, as well as the few shot examples, and the question
            part of the document for `doc`. 
        """
        # TODO: implement evaluation.
        raise NotImplementedError('Evaluation not implemented')
    
    def process_results(self, doc, results):
        """Take a single document and the LM results and evaluates, returning a 
        dict where keys are the names of submetrics and values are the values of 
        the metric for that one document

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
        """
        # TODO: implement evaluation.
        raise NotImplementedError('Evaluation not implemented')

    def aggregation(self):
        """
        :returns: {str: [float] -> float}
            A dictionary where keys are the names of submetrics and values are 
            functions that aggregate a list of metrics
        """
        # TODO: implement evaluation.
        raise NotImplementedError('Evaluation not implemented')

    def higher_is_better(self):
        """
        :returns: {str: bool}
            A dictionary where keys are the names of submetrics and values are 
            whether a higher value of the submetric is better
        """
        # TODO: implement evaluation.
        raise NotImplementedError('Evaluation not implemented')