transform.py

from datasets import load_dataset
import pandas as pd
import re
import json

# Load SQuAD dataset from Hugging Face
dataset = load_dataset('squad')

# Define a function to find the previous sentence-ending punctuation before a given index
def find_previous_punctuation(text, start_index):
    # Regex pattern for sentence-ending punctuation
    punctuation_pattern = r"[.?!;]"
    # Find all matching punctuation positions
    matches = list(re.finditer(punctuation_pattern, text[:start_index]))
    if matches:
        # Return the position just after the last match before start_index
        return matches[-1].end()
    return 0  # Return the start of the string if no punctuation is found

# Define a function to find the next sentence-ending punctuation after a given index
def find_next_punctuation(text, start_index):
    # Regex pattern for sentence-ending punctuation
    punctuation_pattern = r"[.?!;]"
    # Search for the next punctuation after start_index
    match = re.search(punctuation_pattern, text[start_index:])
    if match:
        # Return the index relative to the original string
        return start_index + match.end()
    return len(text)  # Return end of the string if no punctuation is found

# Extract the specific row's context from the dataset
def get_row_context(row):
    # Get the starting index of the answer
    answer_idx = row.get('answers').get('answer_start')[0]  # Assuming the first answer
    # Find the previous sentence-ending punctuation before the answer
    start_idx = find_previous_punctuation(row['context'], answer_idx)
    # Find the next sentence-ending punctuation after the answer
    end_idx = find_next_punctuation(row['context'], answer_idx + len(row.get('answers').get('text')[0]))
    # Return the substring of context containing the answer and its surrounding context
    return row['context'][start_idx:end_idx].strip()

# Function to join and process the dataset into a Pandas DataFrame
def join_squad_dataset(dataset):
    # Convert the dataset into a Pandas DataFrame
    df = pd.DataFrame(dataset)  # Use 'train' split
    # Apply `get_row_context` to generate the context sentence
    df['context'] = df.apply(get_row_context, axis=1)
    df['answer'] = df['answers'].apply(lambda x: x['text'][0])
    # Return the processed DataFrame with required columns
    return df[['context', 'question', 'answer']]

# Process the dataset
train_df = join_squad_dataset(dataset['train'])
test_df = join_squad_dataset(dataset['validation'])

with open("train.json", "w", encoding="utf-8") as f:
        json.dump(train_df.to_dict(orient='records'), f, indent=4, ensure_ascii=False)

with open("validation.json", "w", encoding="utf-8") as f:
        json.dump(test_df.to_dict(orient='records'), f, indent=4, ensure_ascii=False)