Datasets:
HVU_QA
Browse files- HVU_QA/generate_question.py +139 -139
HVU_QA/generate_question.py
CHANGED
@@ -1,139 +1,139 @@
|
|
1 |
-
import json
|
2 |
-
from difflib import SequenceMatcher
|
3 |
-
from transformers import T5Tokenizer, T5ForConditionalGeneration
|
4 |
-
from transformers.utils import logging as hf_logging
|
5 |
-
|
6 |
-
hf_logging.set_verbosity_error()
|
7 |
-
|
8 |
-
MODEL_DIR = "t5-viet-qg-finetuned"
|
9 |
-
DATA_PATH = "30ktrain.json"
|
10 |
-
|
11 |
-
tokenizer = T5Tokenizer.from_pretrained(MODEL_DIR)
|
12 |
-
model = T5ForConditionalGeneration.from_pretrained(MODEL_DIR)
|
13 |
-
|
14 |
-
def find_best_match_from_context(user_context, squad_data):
|
15 |
-
best_score, best_entry = 0.0, None
|
16 |
-
ui = user_context.lower()
|
17 |
-
|
18 |
-
for article in squad_data.get("data", []):
|
19 |
-
context_title = article.get("title", "")
|
20 |
-
score_title = SequenceMatcher(None, ui, context_title.lower()).ratio()
|
21 |
-
|
22 |
-
for paragraph in article.get("paragraphs", []):
|
23 |
-
context = paragraph.get("context", "")
|
24 |
-
for qa in paragraph.get("qas", []):
|
25 |
-
answers = qa.get("answers", [])
|
26 |
-
if not answers:
|
27 |
-
continue
|
28 |
-
answer_text = answers[0].get("text", "").strip()
|
29 |
-
question_text = qa.get("question", "").strip()
|
30 |
-
|
31 |
-
score = score_title
|
32 |
-
if score > best_score:
|
33 |
-
best_score = score
|
34 |
-
best_entry = (context, answer_text, question_text)
|
35 |
-
|
36 |
-
return best_entry
|
37 |
-
|
38 |
-
def _near_duplicate(q, seen, thr=0.90):
|
39 |
-
for s in seen:
|
40 |
-
if SequenceMatcher(None, q, s).ratio() >= thr:
|
41 |
-
return True
|
42 |
-
return False
|
43 |
-
|
44 |
-
def generate_questions(user_context,
|
45 |
-
total_questions=20,
|
46 |
-
batch_size=10,
|
47 |
-
top_k=60,
|
48 |
-
top_p=0.95,
|
49 |
-
temperature=0.9,
|
50 |
-
max_input_len=512,
|
51 |
-
max_new_tokens=64):
|
52 |
-
with open(DATA_PATH, "r", encoding="utf-8") as f:
|
53 |
-
squad_data = json.load(f)
|
54 |
-
|
55 |
-
best_entry = find_best_match_from_context(user_context, squad_data)
|
56 |
-
if best_entry is None:
|
57 |
-
print("Không tìm thấy dữ liệu phù hợp trong file JSON.")
|
58 |
-
return
|
59 |
-
|
60 |
-
context, answer, _ = best_entry
|
61 |
-
|
62 |
-
|
63 |
-
input_text = f"answer: {answer}
|
64 |
-
inputs = tokenizer(
|
65 |
-
input_text,
|
66 |
-
return_tensors="pt",
|
67 |
-
truncation=True,
|
68 |
-
max_length=max_input_len
|
69 |
-
)
|
70 |
-
|
71 |
-
unique_questions = []
|
72 |
-
remaining = total_questions
|
73 |
-
|
74 |
-
while remaining > 0:
|
75 |
-
n = min(batch_size, remaining)
|
76 |
-
outputs = model.generate(
|
77 |
-
**inputs,
|
78 |
-
do_sample=True,
|
79 |
-
top_k=top_k,
|
80 |
-
top_p=top_p,
|
81 |
-
temperature=temperature,
|
82 |
-
max_new_tokens=max_new_tokens,
|
83 |
-
num_return_sequences=n,
|
84 |
-
no_repeat_ngram_size=3,
|
85 |
-
repetition_penalty=1.12
|
86 |
-
)
|
87 |
-
|
88 |
-
for out in outputs:
|
89 |
-
q = tokenizer.decode(out, skip_special_tokens=True).strip()
|
90 |
-
if len(q) < 5:
|
91 |
-
continue
|
92 |
-
if not _near_duplicate(q, unique_questions, thr=0.90):
|
93 |
-
unique_questions.append(q)
|
94 |
-
|
95 |
-
remaining = total_questions - len(unique_questions)
|
96 |
-
if remaining <= 0:
|
97 |
-
break
|
98 |
-
|
99 |
-
unique_questions = unique_questions[:total_questions]
|
100 |
-
|
101 |
-
|
102 |
-
print("Các câu hỏi mới được sinh ra:")
|
103 |
-
for i, q in enumerate(unique_questions, 1):
|
104 |
-
if not q.endswith("?"):
|
105 |
-
q += "?"
|
106 |
-
print(f"{i}. {q}")
|
107 |
-
|
108 |
-
if __name__ == "__main__":
|
109 |
-
user_context = input("\nNhập đoạn văn bản:\n ").strip()
|
110 |
-
|
111 |
-
raw_n = input("\nNhập vào số lượng câu hỏi bạn cần:").strip()
|
112 |
-
if raw_n == "":
|
113 |
-
total_questions = 20
|
114 |
-
else:
|
115 |
-
try:
|
116 |
-
total_questions = int(raw_n)
|
117 |
-
except ValueError:
|
118 |
-
print("Giá trị không hợp lệ. Dùng mặc định 20.")
|
119 |
-
total_questions = 20
|
120 |
-
|
121 |
-
if total_questions < 1:
|
122 |
-
total_questions = 1
|
123 |
-
if total_questions > 200:
|
124 |
-
total_questions = 200
|
125 |
-
|
126 |
-
batch_size = 20 if total_questions >= 30 else min(20, total_questions)
|
127 |
-
|
128 |
-
print("\nĐang phân tích dữ liệu...\n")
|
129 |
-
|
130 |
-
generate_questions(
|
131 |
-
user_context=user_context,
|
132 |
-
total_questions=total_questions,
|
133 |
-
batch_size=batch_size,
|
134 |
-
top_k=60,
|
135 |
-
top_p=0.95,
|
136 |
-
temperature=0.9,
|
137 |
-
max_input_len=512,
|
138 |
-
max_new_tokens=64
|
139 |
-
)
|
|
|
1 |
+
import json
|
2 |
+
from difflib import SequenceMatcher
|
3 |
+
from transformers import T5Tokenizer, T5ForConditionalGeneration
|
4 |
+
from transformers.utils import logging as hf_logging
|
5 |
+
|
6 |
+
hf_logging.set_verbosity_error()
|
7 |
+
|
8 |
+
MODEL_DIR = "t5-viet-qg-finetuned"
|
9 |
+
DATA_PATH = "30ktrain.json"
|
10 |
+
|
11 |
+
tokenizer = T5Tokenizer.from_pretrained(MODEL_DIR)
|
12 |
+
model = T5ForConditionalGeneration.from_pretrained(MODEL_DIR)
|
13 |
+
|
14 |
+
def find_best_match_from_context(user_context, squad_data):
|
15 |
+
best_score, best_entry = 0.0, None
|
16 |
+
ui = user_context.lower()
|
17 |
+
|
18 |
+
for article in squad_data.get("data", []):
|
19 |
+
context_title = article.get("title", "")
|
20 |
+
score_title = SequenceMatcher(None, ui, context_title.lower()).ratio()
|
21 |
+
|
22 |
+
for paragraph in article.get("paragraphs", []):
|
23 |
+
context = paragraph.get("context", "")
|
24 |
+
for qa in paragraph.get("qas", []):
|
25 |
+
answers = qa.get("answers", [])
|
26 |
+
if not answers:
|
27 |
+
continue
|
28 |
+
answer_text = answers[0].get("text", "").strip()
|
29 |
+
question_text = qa.get("question", "").strip()
|
30 |
+
|
31 |
+
score = score_title
|
32 |
+
if score > best_score:
|
33 |
+
best_score = score
|
34 |
+
best_entry = (context, answer_text, question_text)
|
35 |
+
|
36 |
+
return best_entry
|
37 |
+
|
38 |
+
def _near_duplicate(q, seen, thr=0.90):
|
39 |
+
for s in seen:
|
40 |
+
if SequenceMatcher(None, q, s).ratio() >= thr:
|
41 |
+
return True
|
42 |
+
return False
|
43 |
+
|
44 |
+
def generate_questions(user_context,
|
45 |
+
total_questions=20,
|
46 |
+
batch_size=10,
|
47 |
+
top_k=60,
|
48 |
+
top_p=0.95,
|
49 |
+
temperature=0.9,
|
50 |
+
max_input_len=512,
|
51 |
+
max_new_tokens=64):
|
52 |
+
with open(DATA_PATH, "r", encoding="utf-8") as f:
|
53 |
+
squad_data = json.load(f)
|
54 |
+
|
55 |
+
best_entry = find_best_match_from_context(user_context, squad_data)
|
56 |
+
if best_entry is None:
|
57 |
+
print("Không tìm thấy dữ liệu phù hợp trong file JSON.")
|
58 |
+
return
|
59 |
+
|
60 |
+
context, answer, _ = best_entry
|
61 |
+
|
62 |
+
|
63 |
+
input_text = f"answer: {answer}\ncontext: {context}\nquestion:"
|
64 |
+
inputs = tokenizer(
|
65 |
+
input_text,
|
66 |
+
return_tensors="pt",
|
67 |
+
truncation=True,
|
68 |
+
max_length=max_input_len
|
69 |
+
)
|
70 |
+
|
71 |
+
unique_questions = []
|
72 |
+
remaining = total_questions
|
73 |
+
|
74 |
+
while remaining > 0:
|
75 |
+
n = min(batch_size, remaining)
|
76 |
+
outputs = model.generate(
|
77 |
+
**inputs,
|
78 |
+
do_sample=True,
|
79 |
+
top_k=top_k,
|
80 |
+
top_p=top_p,
|
81 |
+
temperature=temperature,
|
82 |
+
max_new_tokens=max_new_tokens,
|
83 |
+
num_return_sequences=n,
|
84 |
+
no_repeat_ngram_size=3,
|
85 |
+
repetition_penalty=1.12
|
86 |
+
)
|
87 |
+
|
88 |
+
for out in outputs:
|
89 |
+
q = tokenizer.decode(out, skip_special_tokens=True).strip()
|
90 |
+
if len(q) < 5:
|
91 |
+
continue
|
92 |
+
if not _near_duplicate(q, unique_questions, thr=0.90):
|
93 |
+
unique_questions.append(q)
|
94 |
+
|
95 |
+
remaining = total_questions - len(unique_questions)
|
96 |
+
if remaining <= 0:
|
97 |
+
break
|
98 |
+
|
99 |
+
unique_questions = unique_questions[:total_questions]
|
100 |
+
|
101 |
+
|
102 |
+
print("Các câu hỏi mới được sinh ra:")
|
103 |
+
for i, q in enumerate(unique_questions, 1):
|
104 |
+
if not q.endswith("?"):
|
105 |
+
q += "?"
|
106 |
+
print(f"{i}. {q}")
|
107 |
+
|
108 |
+
if __name__ == "__main__":
|
109 |
+
user_context = input("\nNhập đoạn văn bản:\n ").strip()
|
110 |
+
|
111 |
+
raw_n = input("\nNhập vào số lượng câu hỏi bạn cần:").strip()
|
112 |
+
if raw_n == "":
|
113 |
+
total_questions = 20
|
114 |
+
else:
|
115 |
+
try:
|
116 |
+
total_questions = int(raw_n)
|
117 |
+
except ValueError:
|
118 |
+
print("Giá trị không hợp lệ. Dùng mặc định 20.")
|
119 |
+
total_questions = 20
|
120 |
+
|
121 |
+
if total_questions < 1:
|
122 |
+
total_questions = 1
|
123 |
+
if total_questions > 200:
|
124 |
+
total_questions = 200
|
125 |
+
|
126 |
+
batch_size = 20 if total_questions >= 30 else min(20, total_questions)
|
127 |
+
|
128 |
+
print("\nĐang phân tích dữ liệu...\n")
|
129 |
+
|
130 |
+
generate_questions(
|
131 |
+
user_context=user_context,
|
132 |
+
total_questions=total_questions,
|
133 |
+
batch_size=batch_size,
|
134 |
+
top_k=60,
|
135 |
+
top_p=0.95,
|
136 |
+
temperature=0.9,
|
137 |
+
max_input_len=512,
|
138 |
+
max_new_tokens=64
|
139 |
+
)
|