Spaces:
Sleeping
Sleeping
Create question_generation.py
Browse files- question_generation.py +97 -0
question_generation.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import re
|
3 |
+
|
4 |
+
@torch.no_grad()
|
5 |
+
def question_generation_sampling(
|
6 |
+
g1_model,
|
7 |
+
g1_tokenizer,
|
8 |
+
g2_model,
|
9 |
+
g2_tokenizer,
|
10 |
+
context,
|
11 |
+
num_questions,
|
12 |
+
device,
|
13 |
+
):
|
14 |
+
qa_input_ids = prepare_qa_input(
|
15 |
+
g1_tokenizer,
|
16 |
+
context=context,
|
17 |
+
device=device,
|
18 |
+
)
|
19 |
+
max_repeated_sampling = int(num_questions * 1.5) # sometimes generated question+answer is invalid
|
20 |
+
num_valid_questions = 0
|
21 |
+
questions = []
|
22 |
+
for q_ in range(max_repeated_sampling):
|
23 |
+
# Stage G.1: question+answer generation
|
24 |
+
outputs = g1_model.generate(
|
25 |
+
qa_input_ids,
|
26 |
+
max_new_tokens=128,
|
27 |
+
do_sample=True,
|
28 |
+
)
|
29 |
+
question_answer = g1_tokenizer.decode(outputs[0], skip_special_tokens=False)
|
30 |
+
question_answer = question_answer.replace(g1_tokenizer.pad_token, "").replace(g1_tokenizer.eos_token, "")
|
31 |
+
question_answer_split = question_answer.split(g1_tokenizer.sep_token)
|
32 |
+
if len(question_answer_split) == 2:
|
33 |
+
# valid Question + Annswer output
|
34 |
+
num_valid_questions += 1
|
35 |
+
else:
|
36 |
+
continue
|
37 |
+
question = question_answer_split[0].strip()
|
38 |
+
answer = question_answer_split[1].strip()
|
39 |
+
|
40 |
+
# Stage G.2: Distractor Generation
|
41 |
+
distractor_input_ids = prepare_distractor_input(
|
42 |
+
g2_tokenizer,
|
43 |
+
context = context,
|
44 |
+
question = question,
|
45 |
+
answer = answer,
|
46 |
+
device = device,
|
47 |
+
separator = g2_tokenizer.sep_token,
|
48 |
+
)
|
49 |
+
outputs = g2_model.generate(
|
50 |
+
distractor_input_ids,
|
51 |
+
max_new_tokens=128,
|
52 |
+
do_sample=True,
|
53 |
+
)
|
54 |
+
distractors = g2_tokenizer.decode(outputs[0], skip_special_tokens=False)
|
55 |
+
distractors = distractors.replace(g2_tokenizer.pad_token, "").replace(g2_tokenizer.eos_token, "")
|
56 |
+
distractors = re.sub("<extra\S+>", g2_tokenizer.sep_token, distractors)
|
57 |
+
distractors = [y.strip() for y in distractors.split(g2_tokenizer.sep_token)]
|
58 |
+
options = [answer] + distractors
|
59 |
+
|
60 |
+
while len(options) < 4:
|
61 |
+
options.append(options[-1])
|
62 |
+
|
63 |
+
question_item = {
|
64 |
+
'question': question,
|
65 |
+
'options': options,
|
66 |
+
}
|
67 |
+
questions.append(question_item)
|
68 |
+
if num_valid_questions == num_questions:
|
69 |
+
break
|
70 |
+
return questions
|
71 |
+
|
72 |
+
|
73 |
+
def prepare_qa_input(t5_tokenizer, context, device):
|
74 |
+
"""
|
75 |
+
input: context
|
76 |
+
output: question <sep> answer
|
77 |
+
"""
|
78 |
+
encoding = t5_tokenizer(
|
79 |
+
[context],
|
80 |
+
return_tensors="pt",
|
81 |
+
)
|
82 |
+
input_ids = encoding.input_ids.to(device)
|
83 |
+
return input_ids
|
84 |
+
|
85 |
+
|
86 |
+
def prepare_distractor_input(t5_tokenizer, context, question, answer, device, separator='<sep>'):
|
87 |
+
"""
|
88 |
+
input: question <sep> answer <sep> article
|
89 |
+
output: distractor1 <sep> distractor2 <sep> distractor3
|
90 |
+
"""
|
91 |
+
input_text = question + ' ' + separator + ' ' + answer + ' ' + separator + ' ' + context
|
92 |
+
encoding = t5_tokenizer(
|
93 |
+
[input_text],
|
94 |
+
return_tensors="pt",
|
95 |
+
)
|
96 |
+
input_ids = encoding.input_ids.to(device)
|
97 |
+
return input_ids
|