Spaces:
Runtime error
Runtime error
Commit
·
6308102
1
Parent(s):
cbda4b2
Create questiongenerator.py
Browse files- questiongenerator.py +345 -0
questiongenerator.py
ADDED
@@ -0,0 +1,345 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import math
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import spacy
|
7 |
+
import re
|
8 |
+
import random
|
9 |
+
import json
|
10 |
+
import en_core_web_sm
|
11 |
+
from string import punctuation
|
12 |
+
|
13 |
+
#from transformers import T5Tokenizer, T5ForConditionalGeneration, T5Config
|
14 |
+
#from transformers import BertTokenizer, BertForSequenceClassification
|
15 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
|
16 |
+
class QuestionGenerator():
|
17 |
+
|
18 |
+
def __init__(self, model_dir=None):
|
19 |
+
|
20 |
+
QG_PRETRAINED = 'iarfmoose/t5-base-question-generator'
|
21 |
+
self.ANSWER_TOKEN = '<answer>'
|
22 |
+
self.CONTEXT_TOKEN = '<context>'
|
23 |
+
self.SEQ_LENGTH = 512
|
24 |
+
|
25 |
+
self.device = torch.device('cpu')
|
26 |
+
# self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
27 |
+
|
28 |
+
self.qg_tokenizer = AutoTokenizer.from_pretrained(QG_PRETRAINED)
|
29 |
+
self.qg_model = AutoModelForSeq2SeqLM.from_pretrained(QG_PRETRAINED)
|
30 |
+
self.qg_model.to(self.device)
|
31 |
+
|
32 |
+
self.qa_evaluator = QAEvaluator(model_dir)
|
33 |
+
|
34 |
+
def generate(self, article, use_evaluator=True, num_questions=None, answer_style='all'):
|
35 |
+
|
36 |
+
print("Generating questions...\n")
|
37 |
+
|
38 |
+
qg_inputs, qg_answers = self.generate_qg_inputs(article, answer_style)
|
39 |
+
print("qg_inputs, qg_answers=>",qg_inputs, qg_answers)
|
40 |
+
generated_questions = self.generate_questions_from_inputs(qg_inputs,num_questions)
|
41 |
+
print("generated_questions(generate)=>",generated_questions)
|
42 |
+
return generated_questions
|
43 |
+
message = "{} questions doesn't match {} answers".format(
|
44 |
+
len(generated_questions),
|
45 |
+
len(qg_answers))
|
46 |
+
assert len(generated_questions) == len(qg_answers), message
|
47 |
+
|
48 |
+
if use_evaluator:
|
49 |
+
|
50 |
+
print("Evaluating QA pairs...\n")
|
51 |
+
|
52 |
+
encoded_qa_pairs = self.qa_evaluator.encode_qa_pairs(generated_questions, qg_answers)
|
53 |
+
scores = self.qa_evaluator.get_scores(encoded_qa_pairs)
|
54 |
+
if num_questions:
|
55 |
+
qa_list = self._get_ranked_qa_pairs(generated_questions, qg_answers, scores, num_questions)
|
56 |
+
else:
|
57 |
+
qa_list = self._get_ranked_qa_pairs(generated_questions, qg_answers, scores)
|
58 |
+
|
59 |
+
else:
|
60 |
+
print("Skipping evaluation step.\n")
|
61 |
+
qa_list = self._get_all_qa_pairs(generated_questions, qg_answers)
|
62 |
+
|
63 |
+
return qa_list
|
64 |
+
|
65 |
+
def generate_qg_inputs(self, text, answer_style):
|
66 |
+
|
67 |
+
VALID_ANSWER_STYLES = ['all', 'sentences', 'multiple_choice']
|
68 |
+
|
69 |
+
if answer_style not in VALID_ANSWER_STYLES:
|
70 |
+
raise ValueError(
|
71 |
+
"Invalid answer style {}. Please choose from {}".format(
|
72 |
+
answer_style,
|
73 |
+
VALID_ANSWER_STYLES
|
74 |
+
)
|
75 |
+
)
|
76 |
+
|
77 |
+
inputs = []
|
78 |
+
answers = []
|
79 |
+
|
80 |
+
if answer_style == 'sentences' or answer_style == 'all':
|
81 |
+
segments = self._split_into_segments(text)
|
82 |
+
for segment in segments:
|
83 |
+
sentences = self._split_text(segment)
|
84 |
+
prepped_inputs, prepped_answers = self._prepare_qg_inputs(sentences, segment)
|
85 |
+
inputs.extend(prepped_inputs)
|
86 |
+
answers.extend(prepped_answers)
|
87 |
+
|
88 |
+
if answer_style == 'multiple_choice' or answer_style == 'all':
|
89 |
+
sentences = self._split_text(text)
|
90 |
+
prepped_inputs, prepped_answers = self._prepare_qg_inputs_MC(sentences)
|
91 |
+
inputs.extend(prepped_inputs)
|
92 |
+
answers.extend(prepped_answers)
|
93 |
+
|
94 |
+
return inputs, answers
|
95 |
+
|
96 |
+
def generate_questions_from_inputs(self, qg_inputs,num_questions):
|
97 |
+
generated_questions = []
|
98 |
+
count = 0
|
99 |
+
print("num que => ", num_questions)
|
100 |
+
for qg_input in qg_inputs:
|
101 |
+
if count < int(num_questions):
|
102 |
+
question = self._generate_question(qg_input)
|
103 |
+
|
104 |
+
question = question.strip() #remove trailing spaces
|
105 |
+
question = question.strip(punctuation) #remove trailing questionmarks
|
106 |
+
question += "?" #add one ?
|
107 |
+
if question not in generated_questions:
|
108 |
+
generated_questions.append(question)
|
109 |
+
print("question ===> ",question)
|
110 |
+
count += 1
|
111 |
+
else:
|
112 |
+
return generated_questions
|
113 |
+
return generated_questions #
|
114 |
+
def _split_text(self, text):
|
115 |
+
MAX_SENTENCE_LEN = 128
|
116 |
+
|
117 |
+
sentences = re.findall('.*?[.!\?]', text)
|
118 |
+
|
119 |
+
cut_sentences = []
|
120 |
+
for sentence in sentences:
|
121 |
+
if len(sentence) > MAX_SENTENCE_LEN:
|
122 |
+
cut_sentences.extend(re.split('[,;:)]', sentence))
|
123 |
+
# temporary solution to remove useless post-quote sentence fragments
|
124 |
+
cut_sentences = [s for s in sentences if len(s.split(" ")) > 5]
|
125 |
+
sentences = sentences + cut_sentences
|
126 |
+
|
127 |
+
return list(set([s.strip(" ") for s in sentences]))
|
128 |
+
|
129 |
+
def _split_into_segments(self, text):
|
130 |
+
MAX_TOKENS = 490
|
131 |
+
|
132 |
+
paragraphs = text.split('\n')
|
133 |
+
tokenized_paragraphs = [self.qg_tokenizer(p)['input_ids'] for p in paragraphs if len(p) > 0]
|
134 |
+
|
135 |
+
segments = []
|
136 |
+
while len(tokenized_paragraphs) > 0:
|
137 |
+
segment = []
|
138 |
+
while len(segment) < MAX_TOKENS and len(tokenized_paragraphs) > 0:
|
139 |
+
paragraph = tokenized_paragraphs.pop(0)
|
140 |
+
segment.extend(paragraph)
|
141 |
+
segments.append(segment)
|
142 |
+
return [self.qg_tokenizer.decode(s) for s in segments]
|
143 |
+
|
144 |
+
def _prepare_qg_inputs(self, sentences, text):
|
145 |
+
inputs = []
|
146 |
+
answers = []
|
147 |
+
|
148 |
+
for sentence in sentences:
|
149 |
+
qg_input = '{} {} {} {}'.format(
|
150 |
+
self.ANSWER_TOKEN,
|
151 |
+
sentence,
|
152 |
+
self.CONTEXT_TOKEN,
|
153 |
+
text
|
154 |
+
)
|
155 |
+
inputs.append(qg_input)
|
156 |
+
answers.append(sentence)
|
157 |
+
|
158 |
+
return inputs, answers
|
159 |
+
|
160 |
+
def _prepare_qg_inputs_MC(self, sentences):
|
161 |
+
|
162 |
+
spacy_nlp = en_core_web_sm.load()
|
163 |
+
docs = list(spacy_nlp.pipe(sentences, disable=['parser']))
|
164 |
+
inputs_from_text = []
|
165 |
+
answers_from_text = []
|
166 |
+
|
167 |
+
for i in range(len(sentences)):
|
168 |
+
entities = docs[i].ents
|
169 |
+
if entities:
|
170 |
+
for entity in entities:
|
171 |
+
qg_input = '{} {} {} {}'.format(
|
172 |
+
self.ANSWER_TOKEN,
|
173 |
+
entity,
|
174 |
+
self.CONTEXT_TOKEN,
|
175 |
+
sentences[i]
|
176 |
+
)
|
177 |
+
answers = self._get_MC_answers(entity, docs)
|
178 |
+
inputs_from_text.append(qg_input)
|
179 |
+
answers_from_text.append(answers)
|
180 |
+
|
181 |
+
return inputs_from_text, answers_from_text
|
182 |
+
|
183 |
+
def _get_MC_answers(self, correct_answer, docs):
|
184 |
+
|
185 |
+
entities = []
|
186 |
+
for doc in docs:
|
187 |
+
entities.extend([{'text': e.text, 'label_': e.label_} for e in doc.ents])
|
188 |
+
|
189 |
+
# remove duplicate elements
|
190 |
+
entities_json = [json.dumps(kv) for kv in entities]
|
191 |
+
pool = set(entities_json)
|
192 |
+
num_choices = min(4, len(pool)) - 1 # -1 because we already have the correct answer
|
193 |
+
|
194 |
+
# add the correct answer
|
195 |
+
final_choices = []
|
196 |
+
correct_label = correct_answer.label_
|
197 |
+
final_choices.append({'answer': correct_answer.text, 'correct': True})
|
198 |
+
pool.remove(json.dumps({'text': correct_answer.text, 'label_': correct_answer.label_}))
|
199 |
+
|
200 |
+
# find answers with the same NER label
|
201 |
+
matches = [e for e in pool if correct_label in e]
|
202 |
+
|
203 |
+
# if we don't have enough then add some other random answers
|
204 |
+
if len(matches) < num_choices:
|
205 |
+
choices = matches
|
206 |
+
pool = pool.difference(set(choices))
|
207 |
+
choices.extend(random.sample(pool, num_choices - len(choices)))
|
208 |
+
else:
|
209 |
+
choices = random.sample(matches, num_choices)
|
210 |
+
|
211 |
+
choices = [json.loads(s) for s in choices]
|
212 |
+
for choice in choices:
|
213 |
+
final_choices.append({'answer': choice['text'], 'correct': False})
|
214 |
+
random.shuffle(final_choices)
|
215 |
+
return final_choices
|
216 |
+
|
217 |
+
def _generate_question(self, qg_input):
|
218 |
+
self.qg_model.eval()
|
219 |
+
encoded_input = self._encode_qg_input(qg_input)
|
220 |
+
with torch.no_grad():
|
221 |
+
output = self.qg_model.generate(input_ids=encoded_input['input_ids'])
|
222 |
+
return self.qg_tokenizer.decode(output[0])
|
223 |
+
|
224 |
+
def _encode_qg_input(self, qg_input):
|
225 |
+
return self.qg_tokenizer(
|
226 |
+
qg_input,
|
227 |
+
pad_to_max_length=True,
|
228 |
+
max_length=self.SEQ_LENGTH,
|
229 |
+
truncation=True,
|
230 |
+
return_tensors="pt"
|
231 |
+
).to(self.device)
|
232 |
+
|
233 |
+
def _get_ranked_qa_pairs(self, generated_questions, qg_answers, scores, num_questions=10):
|
234 |
+
if num_questions > len(scores):
|
235 |
+
num_questions = len(scores)
|
236 |
+
print("\nWas only able to generate {} questions. For more questions, please input a longer text.".format(num_questions))
|
237 |
+
|
238 |
+
qa_list = []
|
239 |
+
for i in range(num_questions):
|
240 |
+
index = scores[i]
|
241 |
+
qa = self._make_dict(
|
242 |
+
generated_questions[index].split('?')[0] + '?',
|
243 |
+
qg_answers[index])
|
244 |
+
qa_list.append(qa)
|
245 |
+
return qa_list
|
246 |
+
|
247 |
+
def _get_all_qa_pairs(self, generated_questions, qg_answers):
|
248 |
+
qa_list = []
|
249 |
+
for i in range(len(generated_questions)):
|
250 |
+
qa = self._make_dict(
|
251 |
+
generated_questions[i].split('?')[0] + '?',
|
252 |
+
qg_answers[i])
|
253 |
+
qa_list.append(qa)
|
254 |
+
return qa_list
|
255 |
+
|
256 |
+
def _make_dict(self, question, answer):
|
257 |
+
qa = {}
|
258 |
+
qa['question'] = question
|
259 |
+
qa['answer'] = answer
|
260 |
+
return qa
|
261 |
+
|
262 |
+
|
263 |
+
class QAEvaluator():
|
264 |
+
def __init__(self, model_dir=None):
|
265 |
+
|
266 |
+
QAE_PRETRAINED = 'iarfmoose/bert-base-cased-qa-evaluator'
|
267 |
+
self.SEQ_LENGTH = 512
|
268 |
+
|
269 |
+
self.device = torch.device('cpu')
|
270 |
+
# self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
271 |
+
|
272 |
+
self.qae_tokenizer = AutoTokenizer.from_pretrained(QAE_PRETRAINED)
|
273 |
+
self.qae_model = AutoModelForSequenceClassification.from_pretrained(QAE_PRETRAINED)
|
274 |
+
self.qae_model.to(self.device)
|
275 |
+
|
276 |
+
|
277 |
+
def encode_qa_pairs(self, questions, answers):
|
278 |
+
encoded_pairs = []
|
279 |
+
for i in range(len(questions)):
|
280 |
+
encoded_qa = self._encode_qa(questions[i], answers[i])
|
281 |
+
encoded_pairs.append(encoded_qa.to(self.device))
|
282 |
+
return encoded_pairs
|
283 |
+
|
284 |
+
def get_scores(self, encoded_qa_pairs):
|
285 |
+
scores = {}
|
286 |
+
self.qae_model.eval()
|
287 |
+
with torch.no_grad():
|
288 |
+
for i in range(len(encoded_qa_pairs)):
|
289 |
+
scores[i] = self._evaluate_qa(encoded_qa_pairs[i])
|
290 |
+
|
291 |
+
return [k for k, v in sorted(scores.items(), key=lambda item: item[1], reverse=True)]
|
292 |
+
|
293 |
+
def _encode_qa(self, question, answer):
|
294 |
+
if type(answer) is list:
|
295 |
+
for a in answer:
|
296 |
+
if a['correct']:
|
297 |
+
correct_answer = a['answer']
|
298 |
+
else:
|
299 |
+
correct_answer = answer
|
300 |
+
return self.qae_tokenizer(
|
301 |
+
text=question,
|
302 |
+
text_pair=correct_answer,
|
303 |
+
pad_to_max_length=True,
|
304 |
+
max_length=self.SEQ_LENGTH,
|
305 |
+
truncation=True,
|
306 |
+
return_tensors="pt"
|
307 |
+
)
|
308 |
+
|
309 |
+
def _evaluate_qa(self, encoded_qa_pair):
|
310 |
+
output = self.qae_model(**encoded_qa_pair)
|
311 |
+
return output[0][0][1]
|
312 |
+
|
313 |
+
|
314 |
+
def print_qa(qa_list, show_answers=True):
|
315 |
+
for i in range(len(qa_list)):
|
316 |
+
space = ' ' * int(np.where(i < 9, 3, 4)) # wider space for 2 digit q nums
|
317 |
+
|
318 |
+
print('{}) Q: {}'.format(i + 1, qa_list[i]['question']))
|
319 |
+
|
320 |
+
answer = qa_list[i]['answer']
|
321 |
+
|
322 |
+
# print a list of multiple choice answers
|
323 |
+
if type(answer) is list:
|
324 |
+
|
325 |
+
if show_answers:
|
326 |
+
print('{}A: 1.'.format(space),
|
327 |
+
answer[0]['answer'],
|
328 |
+
np.where(answer[0]['correct'], '(correct)', ''))
|
329 |
+
for j in range(1, len(answer)):
|
330 |
+
print('{}{}.'.format(space + ' ', j + 1),
|
331 |
+
answer[j]['answer'],
|
332 |
+
np.where(answer[j]['correct'] == True, '(correct)', ''))
|
333 |
+
|
334 |
+
else:
|
335 |
+
print('{}A: 1.'.format(space),
|
336 |
+
answer[0]['answer'])
|
337 |
+
for j in range(1, len(answer)):
|
338 |
+
print('{}{}.'.format(space + ' ', j + 1),
|
339 |
+
answer[j]['answer'])
|
340 |
+
print('')
|
341 |
+
|
342 |
+
# print full sentence answers
|
343 |
+
else:
|
344 |
+
if show_answers:
|
345 |
+
print('{}A:'.format(space), answer, '\n')
|