-
Notifications
You must be signed in to change notification settings - Fork 0
/
EASTER_en.py
191 lines (160 loc) · 7.73 KB
/
EASTER_en.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
import os
import random
import numpy as np
import pandas as pd
import tensorflow as tf
from abc import ABC
from typing import List
from keras.api._v2.keras import Model
from keras.api._v2.keras import layers
from keras.api._v2.keras import optimizers
from keras.api._v2.keras.utils import Sequence
from keras.api._v2.keras.callbacks import ModelCheckpoint
from transformers import RobertaTokenizerFast, TFRobertaForSequenceClassification
from sklearn.preprocessing import OneHotEncoder
from config import *
class DataGenerator(ABC, Sequence):
tokenizer = RobertaTokenizerFast.from_pretrained(ROBERTA_PATH)
def __init__(self, texts, labels):
max_len = 0
input_ids = []
attention_mask = []
self.encoder = OneHotEncoder().fit([[-1], [0], [1]])
for text in texts:
t = DataGenerator.tokenizer(text, return_tensors='tf')
input_ids.append(t['input_ids'].numpy())
attention_mask.append(t['attention_mask'].numpy())
max_len = max(len(input_ids[-1][0]), max_len)
self.max_len = max_len
self.input_ids = input_ids
self.attention_mask = attention_mask
self.labels = self.transform_label(labels)
def __len__(self):
return len(self.input_ids)
def __getitem__(self, batch_ids):
return {'input_ids': self.input_ids[batch_ids], 'attention_mask': self.attention_mask[batch_ids]}, \
np.array([self.labels[batch_ids]])
def _pad_sequences(seq: List[np.ndarray], padding=0, padding_method='same'):
padded_seq = []
for index in range(len(seq)):
t = seq[index]
pad_l = padding - t.shape[0]
if pad_l > 0:
pad_t = np.array([t[-1]])
if padding_method == 'one':
pad_t = np.ones(pad_t.shape, pad_t.dtype)
elif padding_method == 'zero':
pad_t = np.zeros(pad_t.shape, pad_t.dtype)
pad_t = np.concatenate([pad_t for _ in range(pad_l)])
t = np.concatenate((t, pad_t))
padded_seq.append(t)
return padded_seq
def pad_to(self, length):
self.input_ids = [i.reshape((1, -1)) for i in
self._pad_sequences([i[0] for i in self.input_ids], length, 'zero')]
self.attention_mask = [i.reshape((1, -1)) for i in
self._pad_sequences([i[0] for i in self.attention_mask], length, 'same')]
self.max_len = max(self.max_len, length)
return self
def transform_label(self, labels):
return self.encoder.transform(np.asarray(labels).reshape((-1, 1))).toarray()
def build_model(kernel_size, filters, strides, units, include_text_cnn=True):
# load roberta model
roberta = TFRobertaForSequenceClassification.from_pretrained(ROBERTA_PATH)
# define model inputs
inputs = {'input_ids': layers.Input((None,), dtype=tf.int32),
'attention_mask': layers.Input((None,), dtype=tf.int32)}
# Using the first layer of the Roberta model to process input data.
roberta_main = roberta.layers[0](inputs)[0]
# one road
if include_text_cnn:
conv_outputs = []
for size, filter_, stride in zip(kernel_size, filters, strides):
x = layers.Conv1D(filter_, size, stride, activation='relu')(roberta_main)
x = layers.GlobalMaxPool1D()(x)
conv_outputs.append(x)
x = layers.concatenate(conv_outputs)
x = layers.Dropout(0.5)(x)
x = layers.Dense(units, activation='relu')(x)
x = layers.Dropout(0.5)(x)
output_0 = layers.Dense(3, activation='relu')(x)
# another road
output_1 = roberta.layers[1](roberta_main)
if include_text_cnn:
x = layers.concatenate([output_0, output_1])
else:
x = output_1
outputs = layers.Dense(3, activation='softmax')(x)
model_ = Model(inputs, outputs)
model_.compile(optimizers.Adam(5e-6), 'categorical_crossentropy', ['accuracy'])
return model_
def train_model(target: str, model_name: str, model: Model, epochs):
train_data = pd.read_csv(f'{BASE_DIR}/data/sentiment_acos/{target}_train.csv')
train_data = DataGenerator(train_data['text'], train_data['sentiment'])
valid_data = pd.read_csv(f'{BASE_DIR}/data/sentiment_acos/{target}_dev.csv')
valid_data = DataGenerator(valid_data['text'], valid_data['sentiment'])
callbacks = [
ModelCheckpoint(f'{BASE_DIR}/data/pretrained/{model_name}.h5', 'val_accuracy', 1, True, True)
]
model.fit(train_data, epochs=epochs, validation_data=valid_data, callbacks=callbacks, verbose=1)
return model
def load_model(model_name: str, kernel_size, filters, strides, units, include_text_cnn):
model = build_model(kernel_size, filters, strides, units, include_text_cnn)
model.load_weights(f'{BASE_DIR}/data/pretrained/{model_name}.h5')
return model
def evaluate_model(target: str, model: Model):
test_data = pd.read_csv(f'{BASE_DIR}/data/sentiment_acos/{target}_test.csv')
real_data = list(test_data['sentiment'])
test_data = DataGenerator(test_data['text'], test_data['sentiment'])
pred_data = model.predict(test_data)
pred_data = list(map(lambda x: int(x) - 1, np.where(pred_data == np.max(pred_data, axis=1).reshape(-1, 1))[1]))
result = dict()
for sentiment in (-1, 0, 1):
pred = set(map(lambda y: y[0], filter(lambda x: x[1] == sentiment, enumerate(pred_data))))
real = set(map(lambda y: y[0], filter(lambda x: x[1] == sentiment, enumerate(real_data))))
TP = len(list(filter(lambda x: x in pred, real)))
FP = len(pred) - TP
FN = len(real) - TP
result[sentiment] = (TP, FP, FN)
tag = {-1: 'negative', 0: 'neutral', 1: 'positive', 2: 'overall'}
result[2] = tuple(map(lambda y: sum(map(lambda x: x[y], result.values())), range(3)))
print('P,R,F')
for k in (-1, 0, 1, 2):
print(tag[k], end=':')
try:
P, R = result[k][0] / (result[k][0] + result[k][1]), result[k][0] / (result[k][0] + result[k][2])
F = (2 * P * R) / (P + R)
except ZeroDivisionError:
P,R,F = None,None,None
print(f'{P},{R},{F}')
return result
model_params = {
'laptop': ((2, 2, 3), (300, 200, 200), (2, 1, 1), 150),
'rest16': ((2, 2, 3), (200, 200, 150), (1, 1, 1), 150),
}
epoch = 30
def analysis_sentiment(target_name,model_name,test_sentiment_fname):
# Train the model and evaluate its performance on the sentiment test data：
# Default include_text_cnn = True
include_text_cnn = True
model = build_model(*model_params[target_name], include_text_cnn)
train_model(f'{target_name}', model_name, model, epoch)
best_model = load_model(model_name, *model_params[target_name], include_text_cnn)
evaluate_model(f'{target_name}', best_model)
# Use the best model to predict all the AOS test data
test_text_fname = f'{BASE_DIR}/data/acos/{target_name}_quad_test.tsv'
with open(test_text_fname, 'r', encoding='utf-8') as f:
text_list = list(map(lambda x: x.split('\t')[0], f.read().splitlines()))
sentiment_list = best_model.predict(DataGenerator(text_list, [0 for _ in text_list]))
sentiment_list = list( map(lambda x: int(x) - 1, np.where(sentiment_list == np.max(sentiment_list, axis=1).reshape(-1, 1))[1]))
sentiment_df = pd.DataFrame({ 'text': text_list,'sentiment': sentiment_list })
sentiment_df.to_csv(test_sentiment_fname, index=False)
return best_model, text_list, sentiment_list
if __name__ == "__main__":
"""
first step work，train model，predict sentiment
"""
target_name = "laptop"
model_name = f'{target_name}_model'
test_sentiment_fname = f'{BASE_DIR}/data/pred_senti/{target_name}_test_sentiment.csv'
analysis_sentiment(target_name,model_name,test_sentiment_fname);