-
Notifications
You must be signed in to change notification settings - Fork 0
/
sentiment_classification.py
530 lines (485 loc) · 18 KB
/
sentiment_classification.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
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
'''
Author: Julia Jeng, Shu Wang, Arman Anwar
Brief: AIT 726 Homework 2
Usage:
Put file 'sentiment_classification.py' and folder 'twitter' in the same folder.
Command to run:
python sentiment_classification.py
Description:
Build and train a feed forward neural network (FFNN) with 2 layers with hidden vector size 20.
Initalized weights: random weights.
Loss function: mean squared error.
Activation function: sigmoid.
Learning rate: 0.01.
Train/valid rate: 4:1
Emoticon tokenizer: TweetTokenizer
'''
import os
import re
import sys
import math
import random
import numpy as np
import nltk
from nltk.tokenize import TweetTokenizer
from nltk.corpus import stopwords
from nltk import word_tokenize
from nltk.stem import PorterStemmer
from itertools import chain
from sklearn.model_selection import train_test_split
import torch
import torch.nn as nn
import torch.optim as optim
import torch.utils.data as torchdata
nltk.download('stopwords')
nltk.download('punkt')
# global path
logPath = './sentiment_classification.txt'
datPath = './tweet/'
tmpPath = './tmp/'
# Logger: redirect the stream on screen and to file.
class Logger(object):
def __init__(self, filename = "log.txt"):
self.terminal = sys.stdout
self.log = open(filename, "a")
def write(self, message):
self.terminal.write(message)
self.log.write(message)
def flush(self):
pass
# The main function.
def main():
# initialize the log file.
sys.stdout = Logger(logPath)
print("-- AIT726 Homework 2 from Julia Jeng, Shu Wang, and Arman Anwar --")
# create the vocabulary.
if (os.path.exists(tmpPath + '/Train.npz') and
os.path.exists(tmpPath + '/Test.npz') and
os.path.exists(tmpPath + '/Vocab.npz')):
print('Successfully find data from \'' + tmpPath + '\'.')
else:
CreateVocabulary()
# run demo.
DemoFFNN('Stem')
DemoFFNN('noStem')
return
# a demo of neural network classifier.
def DemoFFNN(lStem = 'noStem'):
'''
a demo of neural network classifier.
:param lStem: stem setting - 'noStem', 'Stem'
:return: none
'''
# input validation.
if lStem not in ['noStem', 'Stem']:
print('Error: stem setting invalid!')
return
# extract training features with 'lStem' dataset.
if os.path.exists(tmpPath + '/featTrain_' + lStem + '.npy'):
featTrain = np.load(tmpPath + '/featTrain_' + lStem + '.npy')
print('Successfully load ' + tmpPath + '/featTrain_' + lStem + '.npy')
else:
featTrain = ExtractFeatures('Train', lStem)
np.save(tmpPath + '/featTrain_' + lStem + '.npy', featTrain)
# get the model parameters.
model = TrainFFNN(featTrain)
# extract testing features with 'lStem' dataset.
if os.path.exists(tmpPath + '/featTest_' + lStem + '.npy'):
featTest = np.load(tmpPath + '/featTest_' + lStem + '.npy')
print('Successfully load ' + tmpPath + '/featTest_' + lStem + '.npy')
else:
featTest = ExtractFeatures('Test', lStem)
np.save(tmpPath + '/featTest_' + lStem + '.npy', featTest)
# get testing predictions using model parameters.
accuracy, confusion = TestFFNN(model, featTest)
# output the results on screen and to files.
OutputFFNN(accuracy, confusion, lStem)
# debug
return
# Read train/test sets and create vocabulary.
def CreateVocabulary():
'''
read train and test sets and create vocabulary.
:return: none
'''
# pre-process the data.
def Preprocess(data):
# remove url
pattern = r'http[s]?://(?:[a-zA-Z]|[0-9]|[$-_@.&+]|[!*\(\),]|(?:%[0-9a-fA-F][0-9a-fA-F]))+'
data = re.sub(pattern, '', data)
# remove html special characters.
pattern = r'&[(amp)(gt)(lt)]+;'
data = re.sub(pattern, '', data)
# remove independent numbers.
pattern = r' \d+ '
data = re.sub(pattern, ' ', data)
# lower case capitalized words.
pattern = r'([A-Z][a-z]+)'
def LowerFunc(matched):
return matched.group(1).lower()
data = re.sub(pattern, LowerFunc, data)
# remove hashtags.
pattern = r'[@#]([A-Za-z]+)'
data = re.sub(pattern, '', data)
return data
# remove stop words.
def RemoveStop(data):
dataList = data.split()
for item in dataList:
if item.lower() in stopwords.words('english'):
dataList.remove(item)
dataNew = " ".join(dataList)
return dataNew
# get tokens.
def GetTokens(data):
# use tweet tokenizer.
tknzr = TweetTokenizer()
tokens = tknzr.tokenize(data)
tokensNew = []
# tokenize at each punctuation.
pattern = r'[A-Za-z]+\'[A-Za-z]+'
for tk in tokens:
if re.match(pattern, tk):
subtokens = word_tokenize(tk)
tokensNew = tokensNew + subtokens
else:
tokensNew.append(tk)
return tokensNew
# process tokens with stemming.
def WithStem(tokens):
porter = PorterStemmer()
tokensStem = []
for tk in tokens:
tokensStem.append(porter.stem(tk))
return tokensStem
# if there is no 'tmp' folder, create one.
if not os.path.exists(tmpPath):
os.mkdir(tmpPath)
if not os.path.exists(datPath):
print('[ERROR]: Cannot find the data path \'' + datPath + '\'.')
# read the training data.
labelTrain = []
dataTrain = []
dataTrainStem = []
for root, ds, fs in os.walk(datPath + '/train/'):
for file in fs:
fullname = os.path.join(root, file)
# get the training label.
if "positive" in fullname:
label = 1
else: # "negative" in fullname
label = 0
labelTrain.append(label)
# get the training data.
data = open(fullname, encoding="utf8").read()
# print(data)
# preprocess the data.
data = Preprocess(data)
# print(data)
# remove stop words.
data = RemoveStop(data)
# print(data)
# get the tokens for the data.
tokens = GetTokens(data)
dataTrain.append(tokens)
# print(tokens)
# get the stemmed tokens for the data.
tokensStem = WithStem(tokens)
dataTrainStem.append(tokensStem)
# print(tokensStem)
print('Load TrainSet: %d/%d positive/negative samples.' % (sum(labelTrain), len(labelTrain)-sum(labelTrain)))
np.savez(tmpPath + '/Train.npz', labelTrain = labelTrain, dataTrain = dataTrain, dataTrainStem = dataTrainStem)
# build the vocabulary from training set.
vocab = list(set(list(chain.from_iterable(dataTrain))))
vocabStem = list(set(list(chain.from_iterable(dataTrainStem))))
print('Vocabulary: %d items.' % len(vocab))
print('Vocabulary (stem): %d items.' % len(vocabStem))
np.savez(tmpPath + '/Vocab.npz', vocab = vocab, vocabStem = vocabStem)
# read the testing data.
labelTest = []
dataTest = []
dataTestStem = []
for root, ds, fs in os.walk(datPath + '/test/'):
for file in fs:
fullname = os.path.join(root, file)
# get the testing label.
if "positive" in fullname:
label = 1
else: # "negative" in fullname
label = 0
labelTest.append(label)
# get the testing data.
data = open(fullname, encoding="utf8").read()
# print(data)
# preprocess the data.
data = Preprocess(data)
# print(data)
# remove stop words.
data = RemoveStop(data)
# print(data)
# get the tokens for the data.
tokens = GetTokens(data)
dataTest.append(tokens)
# print(tokens)
# get the stemmed tokens for the data.
tokensStem = WithStem(tokens)
dataTestStem.append(tokensStem)
# print(tokensStem)
print('Load TestSet: %d/%d positive/negative samples.' % (sum(labelTest), len(labelTest)-sum(labelTest)))
np.savez(tmpPath + '/Test.npz', labelTest = labelTest, dataTest = dataTest, dataTestStem = dataTestStem)
return
# extract tfidf features for a 'dataset' with or without 'stem'
def ExtractFeatures(dataset = 'Train', lStem = 'noStem'):
'''
extract features for a 'dataset' with or without 'stem'
:param dataset: dataset type - 'Train', 'Test'
:param lStem: stem setting - 'noStem', 'Stem'
:return: tfidf feature - D * V
'''
# input validation.
if dataset not in ['Train', 'Test']:
print('Error: dataset input invalid!')
return
if lStem not in ['noStem', 'Stem']:
print('Error: stem setting invalid!')
return
# sparse the corresponding dataset.
dset = np.load(tmpPath + dataset + '.npz', allow_pickle = True)
if 'Stem' == lStem:
data = dset['data' + dataset + lStem]
else:
data = dset['data' + dataset]
D = len(data)
# sparse the corresponding vocabulary.
vset = np.load(tmpPath + '/Vocab.npz', allow_pickle = True)
if 'Stem' == lStem:
vocab = vset['vocab' + lStem]
else:
vocab = vset['vocab']
V = len(vocab)
vocabDict = dict(zip(vocab, range(V)))
# get the feature matrix (tfidf):
# get freq and bin features.
termFreq = np.zeros((D, V))
termBin = np.zeros((D, V))
for ind, doc in enumerate(data):
for item in doc:
if item in vocabDict:
termFreq[ind][vocabDict[item]] += 1
termBin[ind][vocabDict[item]] = 1
# get tf (1+log10)
tf = np.zeros((D, V))
for ind in range(D):
for i in range(V):
if termFreq[ind][i] > 0:
tf[ind][i] = 1 + math.log(termFreq[ind][i], 10)
del termFreq
# find idf
if 'Train' == dataset:
# get df
df = np.zeros((V, 1))
for ind in range(D):
for i in range(V):
df[i] += termBin[ind][i]
# get idf (log10(D/df))
idf = np.zeros((V, 1))
for i in range(V):
if df[i] > 0:
idf[i] = math.log(D, 10) - math.log(df[i], 10)
del df
np.save(tmpPath + '/idf_' + lStem + '.npy', idf)
else:
# if 'Test' == dataset, get idf from arguments.
idf = np.load(tmpPath + '/idf_' + lStem + '.npy')
del termBin
# get tfidf
tfidf = np.zeros((D, V))
for ind in range(D):
for i in range(V):
tfidf[ind][i] = tf[ind][i] * idf[i]
return tfidf
# class defination: feed forward neural network.
class FeedForwardNeuralNetwork(nn.Module):
def __init__(self, dims):
super(FeedForwardNeuralNetwork, self).__init__()
self.L1 = nn.Linear(dims, 20)
self.L2 = nn.Linear(20, 1)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
a1 = self.sigmoid(self.L1(x))
a2 = self.sigmoid(self.L2(a1))
return a2
# train the feed forward neural network.
def TrainFFNN(featTrain):
'''
train a feed forward neural network using train features.
:param featTrain: train features - D * V
:return: model - a FeedForwardNeuralNetwork object
'''
# initialize network weights with uniform distribution.
def weight_init(m):
if isinstance(m, nn.Linear):
nn.init.uniform_(m.weight)
nn.init.uniform_(m.bias)
# calculate the train accuracy.
def trainAccuracy(y, yhat):
total = len(y)
cnt = 0
for i in range(total):
err = y[i] - yhat[i]
if abs(err) < 0.5:
cnt += 1
return cnt / total
# get V and D.
V = len(featTrain[0])
D = len(featTrain)
# sparse the corresponding label.
dset = np.load(tmpPath + '/Train.npz', allow_pickle = True)
labelTrain = dset['labelTrain']
# shuffle the data and label.
index = [i for i in range(D)]
random.shuffle(index)
featTrain = featTrain[index]
labelTrain = labelTrain[index]
# split the train and valid set.
xTrain, xValid, yTrain, yValid = train_test_split(featTrain, labelTrain, test_size=0.2)
# convert data (x,y) into tensor.
xTrain = torch.Tensor(xTrain).cuda()
yTrain = torch.LongTensor(yTrain).cuda()
yTrain = yTrain.reshape(len(yTrain), 1)
xValid = torch.Tensor(xValid).cuda()
yValid = torch.LongTensor(yValid).cuda()
yValid = yValid.reshape(len(yValid), 1)
# convert to mini-batch form.
batchsize = 256
train = torchdata.TensorDataset(xTrain, yTrain)
numTrain = len(train)
trainloader = torchdata.DataLoader(train, batch_size=batchsize, shuffle=False)
valid = torchdata.TensorDataset(xValid, yValid)
numValid = len(valid)
validloader = torchdata.DataLoader(valid, batch_size=batchsize, shuffle=False)
# build the model of feed forward neural network.
model = FeedForwardNeuralNetwork(V)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model.apply(weight_init)
model.to(device)
# optimizing with stochastic gradient descent.
optimizer = optim.SGD(model.parameters(), lr = 0.5)
# seting loss function as mean squared error.
criterion = nn.MSELoss()
# run on each epoch.
accList = [0]
for epoch in range(10000):
# training phase.
model.train()
lossTrain = 0
accTrain = 0
for iter, (data, label) in enumerate(trainloader):
data = data.to(device)
label = label.to(device)
optimizer.zero_grad() # set the gradients to zero.
yhat = model.forward(data) # get output
loss = criterion(label.float(), yhat)
loss.backward()
optimizer.step()
# statistic
lossTrain += loss.item()
preds = (yhat > 0.5).long()
accTrain += torch.sum(torch.eq(preds, label).long()).item()
lossTrain /= (iter + 1)
accTrain *= 100 / numTrain
# validation phase.
model.eval()
accValid = 0
with torch.no_grad():
for iter, (data, label) in enumerate(validloader):
data = data.to(device)
label = label.to(device)
yhat = model.forward(data) # get output
# statistic
preds = (yhat > 0.5).long()
accValid += torch.sum(torch.eq(preds, label).long()).item()
accValid *= 100 / numValid
accList.append(accValid)
# output information.
if 0 == (epoch + 1) % 100:
print('[Epoch %03d] loss: %.3f, train acc: %.3f%%, valid acc: %.3f%%' % (
epoch + 1, lossTrain, accTrain, accValid))
# save the best model.
if accList[-1] > max(accList[0:-1]):
torch.save(model.state_dict(), tmpPath + '/model.pth')
# stop judgement.
if (epoch + 1) >= 100 and accList[-1] < min(accList[-100:-1]):
break
# load best model.
model.load_state_dict(torch.load(tmpPath + '/model.pth'))
return model
# test the feed forward neural network.
def TestFFNN(model, featTest):
'''
run test data using the feed forward neural network
:param model: a FeedForwordNeuralNetwork object.
:param featTest: test features - D' * V
:return: accuracy - 0~1
:return: confusion - confusion matrix 2 * 2
'''
# get predictions for testing samples with model parameters.
def GetPredictions(model, featTest):
D = len(featTest)
x = torch.Tensor(featTest).cuda()
yhat = model.forward(x)
predictions = np.zeros(D)
for ind in range(D):
if yhat[ind] > 0.5:
predictions[ind] = 1
return predictions
# evaluate the predictions with gold labels, and get accuracy and confusion matrix.
def Evaluation(predictions):
# sparse the corresponding label.
dset = np.load(tmpPath + '/Test.npz', allow_pickle = True)
labelTest = dset['labelTest']
D = len(labelTest)
cls = 2
# get confusion matrix.
confusion = np.zeros((cls, cls))
for ind in range(D):
nRow = int(predictions[ind])
nCol = int(labelTest[ind])
confusion[nRow][nCol] += 1
# get accuracy.
accuracy = 0
for ind in range(cls):
accuracy += confusion[ind][ind]
accuracy /= D
return accuracy, confusion
# get predictions for testing samples.
predictions = GetPredictions(model, featTest)
# get accuracy and confusion matrix.
accuracy, confusion = Evaluation(predictions)
return accuracy, confusion
# output the results.
def OutputFFNN(accuracy, confusion, lStem):
'''
output the results.
:param accuracy: test accuracy 0~1
:param confusion: confusion matrix 2 * 2
:param lStem: stem setting - 'noStem', 'Stem'
:return: none
'''
# input validation.
if lStem not in ['noStem', 'Stem']:
print('Error: stem setting invalid!')
return
# output on screen and to file.
print('-------------------------------------------')
print('Feed Forward Neural Network | ' + lStem )
print('test accuracy : %.2f%%' % (accuracy * 100))
print('confusion matrix : (actual)')
print(' Neg Pos')
print('(predicted) Neg %-4d(TN) %-4d(FN)' % (confusion[0][0], confusion[0][1]))
print(' Pos %-4d(FP) %-4d(TP)' % (confusion[1][0], confusion[1][1]))
print('-------------------------------------------')
return
# The program entrance.
if __name__ == "__main__":
main()