process.py

""" Script to process raw SQuADshift file for Question Generation format
cd data/processed
gsplit -l 1500 -d --additional-suffix=.jsonl new_wiki.test.jsonl new_wiki.test
gsplit -l 1500 -d --additional-suffix=.jsonl nyt.test.jsonl nyt.test
gsplit -l 1500 -d --additional-suffix=.jsonl reddit.test.jsonl reddit.test
gsplit -l 1500 -d --additional-suffix=.jsonl amazon.test.jsonl amazon.test

gsplit -l 1500 -d --additional-suffix=.jsonl new_wiki.train.jsonl new_wiki.train
gsplit -l 1500 -d --additional-suffix=.jsonl nyt.train.jsonl nyt.train
gsplit -l 1500 -d --additional-suffix=.jsonl reddit.train.jsonl reddit.train
gsplit -l 1500 -d --additional-suffix=.jsonl amazon.train.jsonl amazon.train

gsplit -l 1500 -d --additional-suffix=.jsonl new_wiki.validation.jsonl new_wiki.validation
gsplit -l 1500 -d --additional-suffix=.jsonl nyt.validation.jsonl nyt.validation
gsplit -l 1500 -d --additional-suffix=.jsonl reddit.validation.jsonl reddit.validation
gsplit -l 1500 -d --additional-suffix=.jsonl amazon.validation.jsonl amazon.validation

rm -rf new_wiki.test.jsonl
rm -rf nyt.test.jsonl
rm -rf reddit.test.jsonl
rm -rf amazon.test.jsonl

rm -rf new_wiki.train.jsonl
rm -rf nyt.train.jsonl
rm -rf reddit.train.jsonl
rm -rf amazon.train.jsonl

rm -rf new_wiki.validation.jsonl
rm -rf nyt.validation.jsonl
rm -rf reddit.validation.jsonl
rm -rf amazon.validation.jsonl

"""
import json
import os
import re
from random import shuffle, seed
from tqdm import tqdm

import spacy
from datasets import load_dataset

DATASET_NAME = "squadshifts"
DATASET_TYPES = ['new_wiki', 'nyt', 'reddit', 'amazon']
HIGHLIGHT_TOKEN = '<hl>'
GENERATE_TEST_SPLIT = True
SPLITTER = spacy.load('en_core_web_sm')


def get_sentence(document: str): return [str(sent) for sent in SPLITTER(document).sents]


def process_single_data(question: str, paragraph: str, answer: str):
    """ Convert single raw json data into QG format """
    example = {'question': question, 'paragraph': paragraph, 'answer': answer}
    start = example['paragraph'].find(example['answer'])
    end = start + len(answer)
    assert paragraph[start:end] == answer
    # get sentence
    before_tmp = get_sentence(example['paragraph'][:start])
    if len(before_tmp) == 0:
        before = ''
        before_sentence = ''
    else:
        if before_tmp[-1].endswith('.'):
            before = ' '.join(before_tmp)
            before_sentence = ''
        else:
            before = ' '.join(before_tmp[:-1])
            before_sentence = before_tmp[-1]
            before_sentence = before_sentence if before_sentence.endswith(' ') else f'{before_sentence} '
    after_tmp = get_sentence(example['paragraph'][start + len(example['answer']):])
    if len(after_tmp) == 0:
        after = ''
        after_sentence = ''
    else:
        after = ' '.join(after_tmp[1:])
        after_sentence = after_tmp[0]
        after_sentence = after_sentence if after_sentence.startswith(' ') else f' {after_sentence}'
    example['sentence'] = f"{before_sentence}{example['answer']}{after_sentence}"

    # get paragraph_sentence
    before = '' if before == '' else f'{before} '
    after = '' if after == '' else f' {after}'
    source_text = '{0}{1} {2} {1}{3}'.format(before, HIGHLIGHT_TOKEN, example['sentence'], after)
    example['paragraph_sentence'] = re.sub(r'\s+', ' ', source_text)

    # get paragraph_answer
    source_text = '{0}{1} {2} {1}{3}'.format(
        example['paragraph'][:start], HIGHLIGHT_TOKEN, example['answer'],
        example['paragraph'][start + len(example['answer']):])
    example['paragraph_answer'] = re.sub(r'\s+', ' ', source_text)

    # get sentence_answer
    if len(before_tmp) == 0 or before_tmp[-1].endswith('.'):
        before = ''
    else:
        before = before_tmp[-1] if before_tmp[-1].endswith(' ') else f'{before_tmp[-1]} '
    if len(after_tmp) == 0:
        after = ''
    else:
        after = after_tmp[0] if after_tmp[0].startswith(' ') else f' {after_tmp[0]}'
    source_text = '{0}{1} {2} {1}{3}'.format(before, HIGHLIGHT_TOKEN, example['answer'], after)
    example['sentence_answer'] = re.sub(r'\s+', ' ', source_text)

    return example


if __name__ == '__main__':
    output = './data/processed'
    os.makedirs(output, exist_ok=True)
    for data_type in DATASET_TYPES:
        dataset = load_dataset(DATASET_NAME, data_type)
        _split = 'test'
        tmp_dataset = dataset[_split]
        full_data = []
        for single_data in tqdm(tmp_dataset):
            question_str = single_data['question']  #.replace("\n", ".").replace('"', "'")
            paragraph_str = single_data['context']  #.replace("\n", ".").replace('"', "'")
            answer_str = single_data['answers']['text']
            if type(answer_str) == list:
                answer_str = answer_str[0]
            assert type(answer_str) is str, answer_str
            assert type(question_str) is str, question_str
            assert type(paragraph_str) is str, paragraph_str
            tmp_data = process_single_data(question=question_str, paragraph=paragraph_str, answer=answer_str)
            tmp_data['paragraph_id'] = single_data['id']
            full_data.append(tmp_data)

        # split test into train/valid/test
        test_size = int(len(full_data)/2)
        train_size = int((len(full_data) - test_size) * 2/3)
        # train_size = 2500
        valid_size = len(full_data) - train_size - test_size
        assert train_size + test_size + valid_size == len(full_data), f"{train_size}, {test_size}, {valid_size}"
        paragraph_ids = list(set([i['paragraph_id'] for i in full_data]))
        data_dict = {p: [i for i in full_data if i['paragraph_id'] == p] for p in paragraph_ids}
        seed(0)
        shuffle(paragraph_ids)
        lines_train = []
        lines_test = []
        lines_valid = []

        for i in paragraph_ids:
            if len(lines_test) < test_size:
                lines_test += data_dict[i]
            elif len(lines_train) < train_size:
                lines_train += data_dict[i]
            else:
                lines_valid += data_dict[i]
        print(f'STATS(train/valid/test): {data_type}| {len(lines_train)}/{len(lines_valid)}/{len(lines_test)}')
        with open(f'{output}/{data_type}.test.jsonl', 'w') as f:
            f.write('\n'.join([json.dumps(i) for i in lines_test]))
        with open(f'{output}/{data_type}.train.jsonl', 'w') as f:
            f.write('\n'.join([json.dumps(i) for i in lines_train]))
        with open(f'{output}/{data_type}.validation.jsonl', 'w') as f:
            f.write('\n'.join([json.dumps(i) for i in lines_valid]))