dp2022/squad_transform.py

import json
from dotenv import load_dotenv
from tqdm import tqdm

from squad_utils import print_squad
from translate_utils import translate_text


def sort_qas_by_answer_index(squad):
    for article in squad['data']:
        for paragraph in article['paragraphs']:
            impossible_qas = list(filter(lambda qas: qas['is_impossible'] == True, paragraph['qas']))
            possible_qas = list(filter(lambda qas: qas['is_impossible'] == False, paragraph['qas']))
            sorted_qas = sorted(possible_qas, key=lambda qas: qas['answers'][0]['answer_start'])

            for qas in sorted_qas:
                a = qas['answers'][0]
                a['answer_end'] = a['answer_start'] + len(a['text'])

            paragraph['qas'] = sorted_qas + impossible_qas


def transform_squad(squad):
    for article in squad['data']:
        for paragraph in article['paragraphs']:
            add_special_chars_to_paragraph(paragraph)


def add_special_chars_to_paragraph(paragraph):
    for counter, qas in enumerate(paragraph['qas']):
        # Skip if impossible question
        if qas["is_impossible"] == True: continue

        special_char = f"[{counter}]"

        if len(qas['answers']) > 1 or len(qas['answers']) == 0: continue

        current = qas['answers'][0]

        # Get start index
        start = current['answer_start']
        # Calculate end index
        end = current['answer_end']
        # Add special chars to context
        context = paragraph['context']
        paragraph['context'] = f"{context[:start]}{special_char} {context[start:end]} {special_char}{context[end:]}"

        # Recalculate indexes
        for q in paragraph['qas'][counter + 1:]: # Skip all answers before and current one
            if q["is_impossible"] == True: continue

            other = q['answers'][0]

            if other['answer_start'] >= current['answer_start'] and other['answer_end'] <= current["answer_end"]: # Other is being enclosed by current
                other['answer_start'] += len(special_char) +1
                other['answer_end'] += 2*len(special_char) +2

            elif other['answer_start'] < current['answer_end']: # Other is enclosing the current one
                other['answer_start'] += len(special_char) +1
                other['answer_end'] += len(special_char) +1

            else: # Other is after current
                other['answer_start'] += 2*len(special_char) +2
                other['answer_end'] += 2*len(special_char) +2

        # Fix indexes in current answer
        other = paragraph['qas'][counter]['answers'][0]

        if other == current: # Other answer is the one im working on
            other['answer_start'] += len(special_char) +1
            other['answer_end'] += len(special_char) +1


def translate_paragraphs(squad):
    for article in tqdm(squad["data"]):
        for paragraph in article["paragraphs"]:
            translated = translate_text(paragraph["context"])
            paragraph['translatedContext'] = translated


if __name__ == "__main__":
    load_dotenv()

    with open("./data/squad-v2-dev-small.json", "r") as f:
        squad = json.load(f)

    sort_qas_by_answer_index(squad)
    transform_squad(squad)
    translate_paragraphs(squad)

    with open("./data/squad-v2-dev-small-translated.json", "w") as f:
        json.dump(squad, f, indent=2)