File size: 7,375 Bytes
280d87f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import json
import random
from collections import defaultdict

import torch
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data import (DataLoader, RandomSampler, SequentialSampler,
                              TensorDataset)

from .config_train import DEFAULT_TEXT_ANNOTATION_FILE, batch_size, tokenizer

with open(DEFAULT_TEXT_ANNOTATION_FILE, 'r', encoding='utf-8') as file:
    data = json.load(file)
    
# Prepare sentences and labels
sentences = [item[0] for item in data["annotations"]]
"""
List[str]: A list of sentences extracted from the dataset.
Each sentence corresponds to an annotation in the dataset.
"""

labels = [item[1]['entities'] for item in data["annotations"]]
"""
List[List[Tuple[str, str]]]: A list of entity labels for each sentence.
Each label is a tuple containing the entity text and its corresponding tag.
"""

# Define tags
tags = data["classes"]
"""
List[str]: A list of all possible entity tags (classes) in the dataset.
These tags will be used to label the tokens in each sentence.
"""

# Convert tags to indices
tag2idx = {tag: 0 for idx, tag in enumerate(tags)}
for label in labels:
    for entity in label:
        tag2idx[entity[1]] = tag2idx[entity[1]] + 1
# Sort the dictionary by values
sorted_tags = dict(sorted(tag2idx.items(), key=lambda item: item[1],reverse=True))
sorted_tags = {key: value for key, value in sorted_tags.items() if value != 0}

new_tag = {'<pad>': 0}

sorted_tags = {**new_tag, **sorted_tags}

# Convert tags to indices
tag2idx = {tag: idx for idx, tag in enumerate(list(sorted_tags.keys()))}
"""
Dict[str, int]: A dictionary mapping each tag to a unique index.
This is used to convert tag labels into numerical format for model training.
"""

# Count the occurrences of each tag
tag_counts = defaultdict(int)
for item in data["annotations"]:
    for entity in item[1]["entities"]:
        tag_counts[entity[1]] += 1

# Create a list of annotations for each tag
tag_annotations = defaultdict(list)
for item in data["annotations"]:
    for entity in item[1]["entities"]:
        tag_annotations[entity[1]].append(item)

# Split the annotations for each tag into training and testing sets
train_annotations = []
test_annotations = []
for tag, annotations in tag_annotations.items():
    random.shuffle(annotations)
    split_point = len(annotations) // 2
    train_annotations.extend(annotations[:split_point])
    test_annotations.extend(annotations[split_point:])

# Prepare training and testing datasets
sentences_train = [item[0] for item in train_annotations]
sentences_test = [item[0] for item in test_annotations]
labels_train = [item[1]['entities'] for item in train_annotations]
labels_test = [item[1]['entities'] for item in test_annotations]

# Encode the sentences and labels into input IDs and tag IDs
input_ids_train = []
tag_ids_train = []
for index in range(len(sentences_train)):
    input_ids = tokenizer.encode(sentences_train[index])  # Tokenize the sentence
    input_ids_train.append(input_ids)

    tag_ids = [0 for ids in input_ids]  # Initialize tag IDs with padding
    for labels in labels_train[index]:
        key_ids = tokenizer.encode(labels[0])  # Encode entity text
        for ids in key_ids[1:len(key_ids)-1]:
            ids = input_ids.index(ids)  # Find the index of the token
            tag_ids[ids] = tag2idx[labels[1]]  # Assign tag ID
    tag_ids_train.append(tag_ids)

input_ids_test = []
tag_ids_test = []
for index in range(len(sentences_test)):
    input_ids = tokenizer.encode(sentences_test[index])  # Tokenize the sentence
    input_ids_test.append(input_ids)

    tag_ids = [0 for ids in input_ids]  # Initialize tag IDs with padding
    for labels in labels_test[index]:
        key_ids = tokenizer.encode(labels[0])  # Encode entity text
        for ids in key_ids[1:len(key_ids)-1]:
            ids = input_ids.index(ids)  # Find the index of the token
            tag_ids[ids] = tag2idx[labels[1]]  # Assign tag ID
    tag_ids_test.append(tag_ids)


"""
padded torch.Tensor: A padded tensor of tag IDs for the training set.
All tag sequences are padded to the same length as their corresponding token sequences.
"""
# Pad sequences
padded_input_ids_train = pad_sequence([torch.tensor(ids) for ids in input_ids_train], batch_first=True, padding_value=0)
# Convert to tensor
input_ids_train= padded_input_ids_train.clone().detach()

# Pad sequences
padded_input_ids_test = pad_sequence([torch.tensor(ids) for ids in input_ids_test], batch_first=True, padding_value=0)
# Convert to tensor
input_ids_test= padded_input_ids_test.clone().detach()

# Pad sequences
padded_tag_ids_train = pad_sequence([torch.tensor(ids) for ids in tag_ids_train], batch_first=True, padding_value=0)

# Convert to tensor
tag_ids_train= padded_tag_ids_train.clone().detach()

# Pad sequences
padded_tag_ids_test = pad_sequence([torch.tensor(ids) for ids in tag_ids_test], batch_first=True, padding_value=0)

# Convert to tensor
tag_ids_test= padded_tag_ids_test.clone().detach()

# Find the maximum sequence length
max_length = max(padded_input_ids_train.size(1), padded_input_ids_train.size(1), padded_tag_ids_train.size(1), padded_tag_ids_test.size(1))
# Pad input_ids and tag_ids to have the same length
padded_input_ids_train = torch.nn.functional.pad(padded_input_ids_train, (0, max_length - padded_input_ids_train.size(1)), value=tokenizer.pad_token_id)
padded_input_ids_test = torch.nn.functional.pad(padded_input_ids_test, (0, max_length - padded_input_ids_test.size(1)), value=tokenizer.pad_token_id)
padded_tag_ids_train = torch.nn.functional.pad(padded_tag_ids_train, (0, max_length - padded_tag_ids_train.size(1)), value=0)
padded_tag_ids_test = torch.nn.functional.pad(padded_tag_ids_test, (0, max_length - padded_tag_ids_test.size(1)), value=0)

# Convert to tensor
input_ids_train = padded_input_ids_train.clone().detach()
tag_ids_train = padded_tag_ids_train.clone().detach()

input_ids_test = padded_input_ids_test.clone().detach()
tag_ids_test = padded_tag_ids_test.clone().detach()

"""
Mask torch.Tensor: A tensor containing the attention masks.
This tensor indicates which tokens are real and which are padding.
"""
attention_masks_train = torch.tensor([[1.0 if i != tokenizer.pad_token_id and i!= tokenizer.cls_token_id and i!= tokenizer.sep_token_id else 0 for i in ii] for ii in input_ids_train])
attention_masks_test = torch.tensor([[1.0 if i != tokenizer.pad_token_id and i!= tokenizer.cls_token_id and i!= tokenizer.sep_token_id else 0 for i in ii] for ii in input_ids_test])

# Convert data to PyTorch tensors
train_inputs = input_ids_train.clone().detach()
validation_inputs = input_ids_test.clone().detach()
train_labels = tag_ids_train.clone().detach()
validation_labels = tag_ids_test.clone().detach()
train_masks = attention_masks_train.clone().detach()
validation_masks = attention_masks_test.clone().detach()

# Create DataLoader for training data
train_data = TensorDataset(train_inputs, train_masks, train_labels)
train_sampler = RandomSampler(train_data)
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=batch_size)

# Create DataLoader for validation data
validation_data = TensorDataset(validation_inputs, validation_masks, validation_labels)
validation_sampler = SequentialSampler(validation_data)
validation_dataloader = DataLoader(validation_data, sampler=validation_sampler, batch_size=batch_size)