maximedb commited on
Commit
343756c
·
1 Parent(s): 977eb10
Files changed (1) hide show
  1. train_valid_split.py +38 -19
train_valid_split.py CHANGED
@@ -27,21 +27,39 @@ def filter_valid(questions):
27
  return new_questions
28
 
29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
30
  def format_to_valid(questions):
 
 
 
 
 
 
 
 
 
 
31
  answers_txt = [e["answer"] for e in questions]
32
- questions_txt = [e["question"] for e in questions]
33
- vectorizer = TfidfVectorizer()
34
- vectorizer.fit(answers_txt + questions_txt)
35
- answer_vectors = vectorizer.transform(answers_txt)
36
- for i, question in enumerate(questions):
37
- similarities = linear_kernel(answer_vectors[[i]], answer_vectors).flatten()
38
- answer_scores = [(j, sim) for j, sim in enumerate(similarities) if sim != 1]
39
- answer_scores = sorted(answer_scores, key=lambda x: x[1], reverse=True)
40
- sorted_answers = [questions[j]["answer"] for j, _ in answer_scores if questions[j]["answer"] != question["answer"]]
41
- negative_answer = sorted_answers[0]
42
- assert question["answer"] not in sorted_answers
43
- question["candidates"] = [question["answer"]] + sorted_answers
44
- question["negative_example"] = negative_answer
45
  return questions
46
 
47
 
@@ -57,18 +75,19 @@ def valid_train_split(filename, mapping=None):
57
  line = json.loads(line_txt.strip())
58
  domain = line["domain"]
59
  if domain != previous_domain and previous_domain != "":
 
60
  if len(mapping[previous_domain]) > 1:
61
- train.extend(domain_data["questions"])
62
  elif len(valid) > 2000:
63
- train.extend(domain_data["questions"])
64
  elif len(domain_data["pages"]) > 1:
65
- train.extend(domain_data["questions"])
66
  elif len(domain_data["questions"]) < 15:
67
- train.extend(domain_data["questions"])
68
  else:
69
  questions = filter_valid(domain_data["questions"])
70
  if len(questions) < 15:
71
- train.extend(questions)
72
  else:
73
  questions = format_to_valid(questions)
74
  valid.extend(questions)
@@ -76,7 +95,7 @@ def valid_train_split(filename, mapping=None):
76
  domain_data["questions"].append(line)
77
  domain_data["pages"].add(line["domain_index"])
78
  previous_domain = domain
79
- train.extend(domain_data["questions"])
80
  return train, valid, filename
81
 
82
 
 
27
  return new_questions
28
 
29
 
30
+ # def format_to_valid(questions):
31
+ # answers_txt = [e["answer"] for e in questions]
32
+ # questions_txt = [e["question"] for e in questions]
33
+ # vectorizer = TfidfVectorizer()
34
+ # vectorizer.fit(answers_txt + questions_txt)
35
+ # answer_vectors = vectorizer.transform(answers_txt)
36
+ # for i, question in enumerate(questions):
37
+ # similarities = linear_kernel(answer_vectors[[i]], answer_vectors).flatten()
38
+ # answer_scores = [(j, sim) for j, sim in enumerate(similarities) if sim != 1]
39
+ # answer_scores = sorted(answer_scores, key=lambda x: x[1], reverse=True)
40
+ # sorted_answers = [questions[j]["answer"] for j, _ in answer_scores if questions[j]["answer"] != question["answer"]]
41
+ # negative_answer = sorted_answers[len(sorted_answers) // 2]
42
+ # assert question["answer"] not in sorted_answers
43
+ # question["candidates"] = [question["answer"]] + sorted_answers
44
+ # question["negative_example"] = negative_answer
45
+ # return questions
46
+
47
+
48
  def format_to_valid(questions):
49
+ answers = [e["answer"] for e in questions]
50
+ for question in questions:
51
+ answer = question["answer"]
52
+ candidates = [e for e in answers if e != answer]
53
+ candidates = [answer] + candidates
54
+ question["candidates"] = candidates
55
+ return questions
56
+
57
+
58
+ def format_to_train(questions):
59
  answers_txt = [e["answer"] for e in questions]
60
+ answers_shifted = answers_txt[1:] + [answers_txt[0]]
61
+ for question, answer in zip(questions, answers_shifted):
62
+ question["negative"] = answer
 
 
 
 
 
 
 
 
 
 
63
  return questions
64
 
65
 
 
75
  line = json.loads(line_txt.strip())
76
  domain = line["domain"]
77
  if domain != previous_domain and previous_domain != "":
78
+ form_questions = format_to_train(domain_data["questions"])
79
  if len(mapping[previous_domain]) > 1:
80
+ train.extend(form_questions)
81
  elif len(valid) > 2000:
82
+ train.extend(form_questions)
83
  elif len(domain_data["pages"]) > 1:
84
+ train.extend(form_questions)
85
  elif len(domain_data["questions"]) < 15:
86
+ train.extend(form_questions)
87
  else:
88
  questions = filter_valid(domain_data["questions"])
89
  if len(questions) < 15:
90
+ train.extend(form_questions)
91
  else:
92
  questions = format_to_valid(questions)
93
  valid.extend(questions)
 
95
  domain_data["questions"].append(line)
96
  domain_data["pages"].add(line["domain_index"])
97
  previous_domain = domain
98
+ # train.extend(form_questions)
99
  return train, valid, filename
100
 
101