-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtraining_data_generator.py
226 lines (187 loc) · 9 KB
/
training_data_generator.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
import random
import pandas as pd
from tqdm import tqdm
from shared.utils import make_dirs
from shared.utils import load_from_json
import sys
class Training_Data_Generator(object):
""" Class for generating ground-truth dataset used for feature learning
:param random_seed: parameter used for reproducibility
:param num_samples: total number of negative samples
:param neg_type: negative samples type (simple or hard)
:param query_type: query type (faq or user_query)
:param loss_type: the loss type as method used for BERT Fine-tuning (softmax or triplet loss)
:param hard_filepath: the absolut path to hard negatives filepath
"""
def __init__(self, random_seed=5, num_samples=24, neg_type='simple', query_type='faq',
loss_type='triplet', hard_filepath=''):
self.random_seed = random_seed
self.num_samples = num_samples
self.hard_filepath = hard_filepath
self.neg_type = neg_type
self.query_type = query_type
self.loss_type = loss_type
self.pos_labels = []
self.neg_labels = []
self.num_pos_labels = 0
self.num_neg_labels = 0
self.id2qa = dict()
self.id2negids = dict()
self.df = pd.DataFrame()
self.seq_len_df = pd.DataFrame()
self.df_pos = pd.DataFrame()
self.df_neg = pd.DataFrame()
if self.query_type == 'faq':
self.hard_filepath = self.hard_filepath + "/hard_negatives_faq.json"
elif self.query_type == "user_query":
self.hard_filepath = self.hard_filepath + "/hard_negatives_user_query.json"
else:
raise ValueError('error, no query_type found for {}'.format(query_type))
def generate_pos_labels(self, query_answer_pairs):
""" Generate positive labels from qa pairs
:param qa_pairs: list of dicts
:return: list of positive labels
"""
qap_df = pd.DataFrame.from_records(query_answer_pairs)
qap_by_query_type = qap_df[qap_df['query_type'] == self.query_type]
pos_labels = []
for _, row in qap_by_query_type.iterrows():
id = row['id']
qa_pair = {
"id": id,
"label": 1,
"question": row['question'],
"answer": row['answer'],
"query_type": row['query_type']
}
pos_labels.append(qa_pair)
self.id2qa[id] = (qa_pair['question'], qa_pair['answer'], qa_pair['query_type'])
return pos_labels
def get_id2negids(self, id2qa):
""" Generate random negative sample ids for qa pairs
:param id2qa: dictionary (id: key, question-answer (tuple): value)
:return: dictionary (id: key, neg_ids: value)
"""
random.seed(self.random_seed)
id2negids = dict()
total_qa = len(id2qa)
ids = id2qa.keys()
for id, qa in id2qa.items():
neg_ids = random.sample([x for x in ids if x != id and x !=0], self.num_samples)
id2negids[id] = neg_ids
return id2negids
def generate_neg_labels(self, id2negids):
""" Generate negative labels from id2negids
:param id2negids: dictionary (id: key, neg_ids: value)
:return: list of negative labels as dictionaries
"""
neg_labels = []
for k, v in id2negids.items():
for id in v:
neg_label = dict()
neg_label['id'] = str(k)
neg_label['question'] = self.id2qa[k][0]
neg_answer = self.id2qa[id][1]
neg_label['answer'] = neg_answer
neg_label['label'] = 0
neg_label['query_type'] = self.id2qa[id][2]
neg_labels.append(neg_label)
return neg_labels
def get_seq_len_df(self, query_answer_pairs):
""" Get sequence length in dataframe
"""
seq_len = []
for qa in tqdm(query_answer_pairs):
qa['q_len'] = len(qa['question'])
qa['a_len'] = len(qa['answer'])
seq_len.append(qa)
seq_len_df = pd.DataFrame(seq_len)
return seq_len_df
def get_pos_neg_df(self, query_answer_pairs):
""" Generate positive, negative dataframes """
pos_df = None
neg_df = None
pos_labels = []
neg_labels = []
id2negids = dict()
if self.loss_type == "triplet" or self.loss_type == "softmax":
if self.neg_type == "simple":
pos_labels = self.generate_pos_labels(query_answer_pairs)
id2negids = self.get_id2negids(self.id2qa)
neg_labels = self.generate_neg_labels(id2negids)
pos_df = pd.DataFrame(pos_labels)
neg_df = pd.DataFrame(neg_labels)
elif self.neg_type == "hard":
neg_labels = load_from_json(self.hard_filepath)
pos_labels = query_answer_pairs
neg_df = pd.DataFrame.from_records(neg_labels)
neg_df = neg_df[neg_df['rank'] <= self.num_samples]
pos_df = pd.DataFrame.from_records(pos_labels)
else:
raise ValueError("error, no neg_type found for".format(self.neg_type))
else:
raise ValueError("error, no loss_type found for".format(self.loss_type))
self.id2negids = id2negids
self.pos_labels = pos_labels
self.neg_labels = neg_labels
self.num_pos_labels = len(pos_labels)
self.num_neg_labels = len(neg_labels)
return pos_df, neg_df
def generate_triplet_dataset(self, query_answer_pairs, output_path):
""" Generate ground-truth dataset for feature learning
:param qa_pairs: question-answer pair list
:param output_path: output path name
"""
# create directory structure
output_path = output_path + "/dataset/" + self.loss_type + "/" + self.query_type
make_dirs(output_path)
df = None
pos_df = None
neg_df = None
if self.loss_type == "triplet":
if self.neg_type == "simple":
# generate pos, neg dataframes
pos_df, neg_df = self.get_pos_neg_df(query_answer_pairs)
# rename colnames
pos_df['positive'] = pos_df['answer']
neg_df['negative'] = neg_df['answer']
# drop columns answer, label, query_type
pos_df.drop(['answer', 'label', 'query_type'], axis=1, inplace=True)
neg_df.drop(['answer', 'label', 'query_type'], axis=1, inplace=True)
# convert colname to data types
pos_df['id'] = pos_df['id'].astype(int)
pos_df['question'] = pos_df['question'].astype(str)
neg_df['id'] = neg_df['id'].astype(int)
neg_df['question'] = neg_df['question'].astype(str)
df = pd.merge(pos_df, neg_df, on=['question', 'id'])
df.drop(['id'], axis=1, inplace=True)
elif self.neg_type == "hard":
# generate pos, neg dataframes
pos_df, neg_df = self.get_pos_neg_df(query_answer_pairs)
if ('label' in pos_df.columns) and ('label' in neg_df.columns):
pos_df.drop(['label'], axis=1, inplace=True)
neg_df.drop(['label'], axis=1, inplace=True)
pos_df.rename(columns={'answer': 'pos_answer', 'question': 'query_string'}, inplace=True)
df = pd.merge(pos_df, neg_df, on=['query_string'])
df.drop_duplicates(inplace=True)
df.drop(['id', 'query_type', 'question', 'question_answer', 'rank', 'score'], axis=1, inplace=True)
df.rename(columns={'query_string': 'question', 'pos_answer': 'positive', 'neg_answer': 'negative'}, inplace=True)
elif self.loss_type == "softmax":
if self.neg_type == "simple":
# generate pos, neg dataframes
pos_df, neg_df = self.get_pos_neg_df(query_answer_pairs)
df = pd.concat([pos_df, neg_df])
df.drop(['id', 'query_type'], axis=1, inplace=True)
elif self.neg_type == "hard":
# generate pos, neg dataframes
pos_df, neg_df = self.get_pos_neg_df(query_answer_pairs)
pos_df.drop(['id', 'query_type'], axis=1, inplace=True)
neg_df.drop(['question_answer', 'rank', 'score', 'question'], axis=1, inplace=True)
neg_df.rename(columns={'query_string': 'question', 'neg_answer': 'answer'}, inplace=True)
df = pd.concat([pos_df, neg_df])
else:
raise ValueError("error, no loss_type found for {}".format(self.loss_type))
df.to_csv(output_path + "/" + self.neg_type + "_" + self.query_type + "_dataset.csv", index=False)
self.pos_df = pos_df
self.neg_df = neg_df
self.df = df