JohnnyBoy00 commited on
Commit
50c21de
1 Parent(s): 77a3a56

Upload conversion.py

Browse files
Files changed (1) hide show
  1. conversion.py +130 -0
conversion.py ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import string
3
+ import math
4
+ import random
5
+ import xml.etree.ElementTree as et
6
+ import jsonlines
7
+ import uuid
8
+ import pandas as pd
9
+
10
+ # set random seed for shuffling
11
+ random.seed(1)
12
+
13
+ # column names of the reference answers file
14
+ FILE_NUMBER_COL = 'file_number'
15
+ REFERENCE_ANSWER_COL = 'reference_answer'
16
+
17
+ # column names of the files with the data
18
+ QUESTION_COL = 'Frage'
19
+ ANSWER_COL = 'Antwort'
20
+ SCORE_COL = 'Score'
21
+ ERROR_CLASS_COL = 'Fehlerklasse'
22
+ FEEDBACK_COL = 'Feedback'
23
+
24
+ # labels for verification_feedback
25
+ CORRECT_LABEL = 'Correct'
26
+ PARTIALLY_CORRECT_LABEL = 'Partially correct'
27
+ INCORRECT_LABEL = 'Incorrect'
28
+
29
+ def convert_xlsx_to_jsonl(
30
+ path_to_dataset,
31
+ path_to_reference_answers_file,
32
+ dir,
33
+ filename,
34
+ train_split=None):
35
+ """
36
+ Utility function used for conversion of .xlsx files from the dataset into JSON lines
37
+
38
+ Params:
39
+ path_to_dataset (string): path to the folder containing the dataset (in .xlsx format)
40
+ path_to_reference_answers_file (string): path to the folder containing the reference answers (in .xlsx format)
41
+ dir (string): name of the directory where the JSON lines file will be stored
42
+ filename (string): name of the JSON lines file that will store the dataset
43
+ train_split (float or None): if not None, defines which percentage of the dataset to use for the train and validation splits
44
+
45
+ Returns:
46
+ None: the file is saved JSON lines format in the specified location
47
+ """
48
+ def return_verification_feedback(score):
49
+ if math.isclose(score, 1.0):
50
+ return CORRECT_LABEL
51
+ elif math.isclose(score, 0.0):
52
+ return INCORRECT_LABEL
53
+ else:
54
+ return PARTIALLY_CORRECT_LABEL
55
+
56
+ data = []
57
+
58
+ # get reference answers from file
59
+ reference_answers_df = pd.read_excel(path_to_reference_answers_file)
60
+ # the keys of the dictionary are the number of the files padded with zeroes
61
+ # so that it has two digits, and the values are the reference answers themselves
62
+ reference_answers = {
63
+ f'{row[FILE_NUMBER_COL]:02}': row[REFERENCE_ANSWER_COL].strip()
64
+ for _, row in reference_answers_df.iterrows()}
65
+
66
+ # loop through all files in directory
67
+ for f in os.listdir(path_to_dataset):
68
+ if f.endswith('.xlsx'):
69
+ # read file
70
+ file_df = pd.read_excel(os.path.join(path_to_dataset, f))
71
+ # get question
72
+ question = file_df[QUESTION_COL].iat[0].strip()
73
+ # get reference answer based on file name
74
+ ref_answer = reference_answers[f.split('.')[0]]
75
+
76
+ # loop through all rows and store the appropriate fields in a list
77
+ for _, row in file_df.iterrows():
78
+ response = row[ANSWER_COL].strip()
79
+ score = float(row[SCORE_COL])
80
+ feedback = str(row[FEEDBACK_COL]).strip()
81
+ verification_feedback = return_verification_feedback(score)
82
+ error_class = row[ERROR_CLASS_COL].strip()
83
+
84
+ # create dictionary with the appropriate fields
85
+ data.append({
86
+ 'id': uuid.uuid4().hex, # generate unique id in HEX format
87
+ 'question': question,
88
+ 'reference_answer': ref_answer,
89
+ 'provided_answer': response,
90
+ 'answer_feedback': feedback,
91
+ 'verification_feedback': verification_feedback,
92
+ 'error_class': error_class,
93
+ 'score': score
94
+ })
95
+
96
+ if not os.path.exists(dir):
97
+ print('Creating directory where JSON file will be stored\n')
98
+ os.makedirs(dir)
99
+
100
+ if train_split is None:
101
+ with jsonlines.open(f'{os.path.join(dir, filename)}.jsonl', 'w') as writer:
102
+ writer.write_all(data)
103
+ else:
104
+ # shuffle data and divide it into train and validation splits
105
+ random.shuffle(data)
106
+ train_data = data[: int(train_split * (len(data) - 1))]
107
+ val_data = data[int(train_split * (len(data) - 1)) :]
108
+
109
+ # write JSON lines file with train data
110
+ with jsonlines.open(f'{os.path.join(dir, filename)}-train.jsonl', 'w') as writer:
111
+ writer.write_all(train_data)
112
+
113
+ # write JSON lines file with validation data
114
+ with jsonlines.open(f'{os.path.join(dir, filename)}-validation.jsonl', 'w') as writer:
115
+ writer.write_all(val_data)
116
+
117
+ if __name__ == '__main__':
118
+ # convert legal domain dataset (german) to JSON lines
119
+ convert_xlsx_to_jsonl(
120
+ 'data/training', 'data/reference_answers.xlsx',
121
+ 'data/json', 'saf-legal-domain-german',
122
+ train_split=0.8)
123
+
124
+ convert_xlsx_to_jsonl(
125
+ 'data/unseen_answers', 'data/reference_answers.xlsx',
126
+ 'data/json', 'saf-legal-domain-german-unseen-answers')
127
+
128
+ convert_xlsx_to_jsonl(
129
+ 'data/unseen_questions', 'data/reference_answers.xlsx',
130
+ 'data/json', 'saf-legal-domain-german-unseen-questions')