conversion.py

import os
import string
import math
import random
import xml.etree.ElementTree as et
import jsonlines
import uuid
import pandas as pd

# set random seed for shuffling
random.seed(1)

# column names of the reference answers file
FILE_NUMBER_COL = 'file_number'
REFERENCE_ANSWER_COL = 'reference_answer'

# column names of the files with the data
QUESTION_COL = 'Frage'
ANSWER_COL = 'Antwort'
SCORE_COL = 'Score'
ERROR_CLASS_COL = 'Fehlerklasse'
FEEDBACK_COL = 'Feedback'

# labels for verification_feedback
CORRECT_LABEL = 'Correct'
PARTIALLY_CORRECT_LABEL = 'Partially correct'
INCORRECT_LABEL = 'Incorrect'

def convert_xlsx_to_jsonl(
    path_to_dataset,
    path_to_reference_answers_file,
    dir,
    filename,
    train_split=None):
    """
    Utility function used for conversion of .xlsx files from the dataset into JSON lines
    
    Params:
        path_to_dataset (string): path to the folder containing the dataset (in .xlsx format)
        path_to_reference_answers_file (string): path to the folder containing the reference answers (in .xlsx format)
        dir (string): name of the directory where the JSON lines file will be stored
        filename (string): name of the JSON lines file that will store the dataset
        train_split (float or None): if not None, defines which percentage of the dataset to use for the train and validation splits
    
    Returns:
        None: the file is saved JSON lines format in the specified location
    """
    def return_verification_feedback(score):
        if math.isclose(score, 1.0):
            return CORRECT_LABEL
        elif math.isclose(score, 0.0):
            return INCORRECT_LABEL
        else:
            return PARTIALLY_CORRECT_LABEL
    
    data = []
    
    # get reference answers from file
    reference_answers_df = pd.read_excel(path_to_reference_answers_file)
    # the keys of the dictionary are the number of the files padded with zeroes
    # so that it has two digits, and the values are the reference answers themselves
    reference_answers = {
        f'{row[FILE_NUMBER_COL]:02}': row[REFERENCE_ANSWER_COL].strip() 
        for _, row in reference_answers_df.iterrows()}
    
    # loop through all files in directory
    for f in os.listdir(path_to_dataset):
        if f.endswith('.xlsx'):
            # read file
            file_df = pd.read_excel(os.path.join(path_to_dataset, f))
            # get question
            question = file_df[QUESTION_COL].iat[0].strip()
            # get reference answer based on file name
            ref_answer = reference_answers[f.split('.')[0]]

            # loop through all rows and store the appropriate fields in a list
            for _, row in file_df.iterrows():
                response = row[ANSWER_COL].strip()
                score = float(row[SCORE_COL])
                feedback = str(row[FEEDBACK_COL]).strip()
                verification_feedback = return_verification_feedback(score)
                error_class = row[ERROR_CLASS_COL].strip()
                
                # create dictionary with the appropriate fields
                data.append({
                    'id': uuid.uuid4().hex, # generate unique id in HEX format
                    'question': question,
                    'reference_answer': ref_answer,
                    'provided_answer': response,
                    'answer_feedback': feedback,
                    'verification_feedback': verification_feedback,
                    'error_class': error_class,
                    'score': score
                })
   
    if not os.path.exists(dir):
        print('Creating directory where JSON file will be stored\n')
        os.makedirs(dir)
    
    if train_split is None:
        with jsonlines.open(f'{os.path.join(dir, filename)}.jsonl', 'w') as writer:
            writer.write_all(data)
    else:
        # shuffle data and divide it into train and validation splits
        random.shuffle(data)
        train_data = data[: int(train_split * (len(data) - 1))]
        val_data = data[int(train_split * (len(data) - 1)) :]

        # write JSON lines file with train data
        with jsonlines.open(f'{os.path.join(dir, filename)}-train.jsonl', 'w') as writer:
            writer.write_all(train_data)
        
        # write JSON lines file with validation data
        with jsonlines.open(f'{os.path.join(dir, filename)}-validation.jsonl', 'w') as writer:
            writer.write_all(val_data)

if __name__ == '__main__':
    # convert legal domain dataset (german) to JSON lines
    convert_xlsx_to_jsonl(
        'data/training', 'data/reference_answers.xlsx',
        'data/json', 'saf-legal-domain-german',
        train_split=0.8)

    convert_xlsx_to_jsonl(
        'data/unseen_answers', 'data/reference_answers.xlsx',
        'data/json', 'saf-legal-domain-german-unseen-answers')

    convert_xlsx_to_jsonl(
        'data/unseen_questions', 'data/reference_answers.xlsx',
        'data/json', 'saf-legal-domain-german-unseen-questions')