/
joint_intent_slot_dataset.py
258 lines (211 loc) · 8.63 KB
/
joint_intent_slot_dataset.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
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
# Copyright 2018 The Google AI Language Team Authors and
# The HuggingFace Inc. team.
# Copyright (c) 2019, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Utility functions for Token Classification NLP tasks
Some parts of this code were adapted from the HuggingFace library at
https://github.com/huggingface/pytorch-pretrained-BERT
"""
import random
import numpy as np
from torch.utils.data import Dataset
from nemo import logging
from nemo.collections.nlp.data.datasets.datasets_utils.preprocessing import get_stats
__all__ = ['BertJointIntentSlotDataset', 'BertJointIntentSlotInferDataset']
def get_features(
queries,
max_seq_length,
tokenizer,
pad_label=128,
raw_slots=None,
ignore_extra_tokens=False,
ignore_start_end=False,
):
all_subtokens = []
all_loss_mask = []
all_subtokens_mask = []
all_segment_ids = []
all_input_ids = []
all_input_mask = []
sent_lengths = []
all_slots = []
with_label = False
if raw_slots is not None:
with_label = True
for i, query in enumerate(queries):
words = query.strip().split()
subtokens = ['[CLS]']
loss_mask = [1 - ignore_start_end]
subtokens_mask = [0]
if with_label:
slots = [pad_label]
for j, word in enumerate(words):
word_tokens = tokenizer.tokenize(word)
subtokens.extend(word_tokens)
loss_mask.append(1)
loss_mask.extend([not ignore_extra_tokens] * (len(word_tokens) - 1))
subtokens_mask.append(1)
subtokens_mask.extend([0] * (len(word_tokens) - 1))
if with_label:
slots.extend([raw_slots[i][j]] * len(word_tokens))
subtokens.append('[SEP]')
loss_mask.append(not ignore_start_end)
subtokens_mask.append(0)
sent_lengths.append(len(subtokens))
all_subtokens.append(subtokens)
all_loss_mask.append(loss_mask)
all_subtokens_mask.append(subtokens_mask)
all_input_mask.append([1] * len(subtokens))
if with_label:
slots.append(pad_label)
all_slots.append(slots)
max_seq_length = min(max_seq_length, max(sent_lengths))
logging.info(f'Max length: {max_seq_length}')
get_stats(sent_lengths)
too_long_count = 0
for i, subtokens in enumerate(all_subtokens):
if len(subtokens) > max_seq_length:
subtokens = ['[CLS]'] + subtokens[-max_seq_length + 1 :]
all_input_mask[i] = [1] + all_input_mask[i][-max_seq_length + 1 :]
all_loss_mask[i] = [1 - ignore_start_end] + all_loss_mask[i][-max_seq_length + 1 :]
all_subtokens_mask[i] = [0] + all_subtokens_mask[i][-max_seq_length + 1 :]
if with_label:
all_slots[i] = [pad_label] + all_slots[i][-max_seq_length + 1 :]
too_long_count += 1
all_input_ids.append([tokenizer._convert_token_to_id(t) for t in subtokens])
if len(subtokens) < max_seq_length:
extra = max_seq_length - len(subtokens)
all_input_ids[i] = all_input_ids[i] + [0] * extra
all_loss_mask[i] = all_loss_mask[i] + [0] * extra
all_subtokens_mask[i] = all_subtokens_mask[i] + [0] * extra
all_input_mask[i] = all_input_mask[i] + [0] * extra
if with_label:
all_slots[i] = all_slots[i] + [pad_label] * extra
all_segment_ids.append([0] * max_seq_length)
logging.info(f'{too_long_count} are longer than {max_seq_length}')
return (all_input_ids, all_segment_ids, all_input_mask, all_loss_mask, all_subtokens_mask, all_slots)
class BertJointIntentSlotDataset(Dataset):
"""
Creates dataset to use for the task of joint intent
and slot classification with pretrained model.
Converts from raw data to an instance that can be used by
NMDataLayer.
For dataset to use during inference without labels, see
BertJointIntentSlotInferDataset.
Args:
input_file (str): file to sequence + label.
the first line is header (sentence [tab] label)
each line should be [sentence][tab][label]
slot_file (str): file to slot labels, each line corresponding to
slot labels for a sentence in input_file. No header.
max_seq_length (int): max sequence length minus 2 for [CLS] and [SEP]
tokenizer (Tokenizer): such as BertTokenizer
num_samples (int): number of samples you want to use for the dataset.
If -1, use all dataset. Useful for testing.
shuffle (bool): whether to shuffle your data.
pad_label (int): pad value use for slot labels.
by default, it's the neutral label.
"""
def __init__(
self,
input_file,
slot_file,
max_seq_length,
tokenizer,
num_samples=-1,
shuffle=True,
pad_label=128,
ignore_extra_tokens=False,
ignore_start_end=False,
):
if num_samples == 0:
raise ValueError("num_samples has to be positive", num_samples)
with open(slot_file, 'r') as f:
slot_lines = f.readlines()
with open(input_file, 'r') as f:
input_lines = f.readlines()[1:]
assert len(slot_lines) == len(input_lines)
dataset = list(zip(slot_lines, input_lines))
if shuffle or num_samples > 0:
random.shuffle(dataset)
if num_samples > 0:
dataset = dataset[:num_samples]
raw_slots, queries, raw_intents = [], [], []
for slot_line, input_line in dataset:
raw_slots.append([int(slot) for slot in slot_line.strip().split()])
parts = input_line.strip().split()
raw_intents.append(int(parts[-1]))
queries.append(' '.join(parts[:-1]))
features = get_features(
queries,
max_seq_length,
tokenizer,
pad_label=pad_label,
raw_slots=raw_slots,
ignore_extra_tokens=ignore_extra_tokens,
ignore_start_end=ignore_start_end,
)
self.all_input_ids = features[0]
self.all_segment_ids = features[1]
self.all_input_mask = features[2]
self.all_loss_mask = features[3]
self.all_subtokens_mask = features[4]
self.all_slots = features[5]
self.all_intents = raw_intents
def __len__(self):
return len(self.all_input_ids)
def __getitem__(self, idx):
return (
np.array(self.all_input_ids[idx]),
np.array(self.all_segment_ids[idx]),
np.array(self.all_input_mask[idx], dtype=np.long),
np.array(self.all_loss_mask[idx]),
np.array(self.all_subtokens_mask[idx]),
self.all_intents[idx],
np.array(self.all_slots[idx]),
)
class BertJointIntentSlotInferDataset(Dataset):
"""
Creates dataset to use for the task of joint intent
and slot classification with pretrained model.
Converts from raw data to an instance that can be used by
NMDataLayer.
This is to be used during inference only.
For dataset to use during training with labels, see
BertJointIntentSlotDataset.
Args:
queries (list): list of queries to run inference on
max_seq_length (int): max sequence length minus 2 for [CLS] and [SEP]
tokenizer (Tokenizer): such as BertTokenizer
pad_label (int): pad value use for slot labels.
by default, it's the neutral label.
"""
def __init__(self, queries, max_seq_length, tokenizer):
features = get_features(queries, max_seq_length, tokenizer)
self.all_input_ids = features[0]
self.all_segment_ids = features[1]
self.all_input_mask = features[2]
self.all_loss_mask = features[3]
self.all_subtokens_mask = features[4]
def __len__(self):
return len(self.all_input_ids)
def __getitem__(self, idx):
return (
np.array(self.all_input_ids[idx]),
np.array(self.all_segment_ids[idx]),
np.array(self.all_input_mask[idx], dtype=np.long),
np.array(self.all_loss_mask[idx]),
np.array(self.all_subtokens_mask[idx]),
)