import os import string import math import random import xml.etree.ElementTree as et import jsonlines import uuid # set random seed for shuffling random.seed(1) def convert_xml_to_jsonl(path_to_dataset, dir, filename, train_split=None): """ Utility function used for conversion of XML files from the dataset into JSON lines Params: path_to_dataset (string): path to the folder containing the dataset (in XML 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 in JSON lines format in the specified location """ data = [] # loop through all files in directory for f in os.listdir(path_to_dataset): if f.endswith('.xml'): root = et.parse(os.path.join(path_to_dataset, f)).getroot() # get question question = root.find('questionText').text.replace('\n', ' ') # get reference and student answers ref_answers = [x for x in root.find('referenceAnswers')] student_answers = [x for x in root.find('studentAnswers')] if len(ref_answers) == 1: # get reference answer and clear all spaces ref_answer = ref_answers[0].text.strip() # loop through all student answers and store the appropriate fields in a list for answer in student_answers: response = answer.find('response').text.strip() score = float(answer.find('score').text) feedback = answer.find('response_feedback').text.strip() verification_feedback = answer.find('verification_feedback').text.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, '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 micro job dataset (german) to JSON lines convert_xml_to_jsonl( 'data/training/german', 'data/json', 'saf-micro-job-german', train_split=0.8) convert_xml_to_jsonl( 'data/unseen_answers/german', 'data/json', 'saf-micro-job-german-unseen-answers') convert_xml_to_jsonl( 'data/unseen_questions/german', 'data/json', 'saf-micro-job-german-unseen-questions')