/
bow_baseline.py
288 lines (241 loc) · 14.1 KB
/
bow_baseline.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
#!/usr/bin/env python
__author__ = 'Isabelle Augenstein'
import numpy as np
from collections import Counter
import tokenize_tweets
from twokenize_wrapper import tokenize
from tokenize_tweets import readTweetsOfficial
from training_eval import *
from affect import getAffect
from autoencoder_eval import extractFeaturesCrossTweetTarget, extractFeaturesAutoencoder
from emoticons import analyze_tweet
from word2vec_integration import extractW2VHashFeatures
from gensim.models import word2vec, Phrases
from tokenize_tweets import filterStopwords
# select features, compile feature vocab
def extractFeatureVocab(tweets, keyword="all", usephrasemodel=True, phrasemodel="phrase.model", anon_targets=False):
tokencounts = Counter()
features_final = []
bigram = Phrases(phrasemodel)
#tokens_topic = []
#if keyword == "all":
# for top in tokenize_tweets.TOPICS:
# if top != 'clinton':
# for tok in tokenize(tokenize_tweets.TOPICS_LONG[top]):
# tokens_topic.append(tok)
#else:
# tokens_topic = tokenize(tokenize_tweets.TOPICS_LONG[keyword])
for tweet in tweets:
if usephrasemodel == False:
tokenised_tweet = tokenize(tweet)
for token in tokenised_tweet: #unigram features
tokencounts[token] += 1
#for toktopic in tokens_topic:
# tokencounts[toktopic + '|' + token] += 1
for l in zip(*[tokenised_tweet[i:] for i in range(2)]): #bigram features
tokencounts["_".join(l)] += 1
#for ltop in zip(*[tokens_topic[i:] for i in range(2)]):
# tokencounts["_".join(ltop) + '|' + "_".join(l)] += 1
else:
# this includes unigrams and frequent bigrams
tokens = filterStopwords(tokenize(tweet.lower())) #For Trump it's [1]
phrasetoks = bigram[tokens]
target_keywords = []
if anon_targets==True:
for top in tokenize_tweets.TOPICS:
if top == "climate": # hack, this is the only non-list value
target_keywords.append("climate")
else:
#for keyw in tokenize_tweets.KEYWORDS[top]:
target_keywords.extend(tokenize_tweets.KEYWORDS[top])
phrasetoks_new = []
for token in phrasetoks:
for keyw in target_keywords:
if keyw in token:
token = token.replace(keyw, "TARGET")
phrasetoks_new.append(token)
phrasetoks = phrasetoks_new
for token in phrasetoks:
tokencounts[token] += 1
for l in zip(*[phrasetoks[i:] for i in range(2)]):
tokencounts["_".join(l)] += 1
for token, count in tokencounts.most_common():
if count > 1:
features_final.append(token)
#print token, count
return features_final
# extract BOW n-gram features, returns matrix of features
def extractFeaturesBOW(tweets, targets, features_final, anon_targets=False, usephrasemodel=False, phrasemodel="phrase.model"):
bigram = Phrases(phrasemodel)
matrix = [] # np.zeros((len(features_final), len(tweets)))
for i, tweet in enumerate(tweets):
vect = np.zeros((len(features_final)))
if usephrasemodel == False:
tokenised_tweet = tokenize(tweet)
for token in tokenised_tweet:
insertIntoVect(features_final, vect, token)
#for toktopic in tokens_topic:
# insertIntoVect(features_final, vect, toktopic + '|' + token)
for l in zip(*[tokenised_tweet[i:] for i in range(2)]):
insertIntoVect(features_final, vect, "_".join(l))
#for ltop in zip(*[tokens_topic[i:] for i in range(2)]):
# insertIntoVect(features_final, vect, "_".join(ltop) + '|' + "_".join(l))
else:
inv_topics = {v: k for k, v in tokenize_tweets.TOPICS_LONG.items()}
target_keywords = tokenize_tweets.KEYWORDS.get(inv_topics.get(targets[i]))
tokens = filterStopwords(tokenize(tweet.lower())) #For Trump it's [1]
phrasetoks = bigram[tokens]
if anon_targets==True:
phrasetoks_new = []
for token in phrasetoks:
if target_keywords == "climate":
if target_keywords in token:
token = token.replace(keyw, "TARGET")
else:
for keyw in target_keywords:
if keyw in token:
token = token.replace(keyw, "TARGET")
phrasetoks_new.append(token)
phrasetoks = phrasetoks_new
for token in phrasetoks:
insertIntoVect(features_final, vect, token)
for l in zip(*[phrasetoks[i:] for i in range(2)]):
insertIntoVect(features_final, vect, "_".join(l))
matrix.append(vect)
#print " ".join(str(v) for v in vect), "\n"
return matrix
def insertIntoVect(feats, vect, expr):
try:
ind = feats.index(expr)
vect[ind] = 1
except (ValueError, IndexError):
pass
return vect
# extract emoticon features
def extractEmoticons(tweets):
vects = [] # BOTH_HS, HAPPY, SAD, NA
vocab = ["BOTH_HS", "HAPPY", "SAD", "NA"]
for i, tweet in enumerate(tweets):
vect = np.zeros(4)
emo = analyze_tweet(tweet)
if emo == "NA":
vect[0] = 1
elif emo == "HAPPY":
vect[1] = 1
elif emo == "SAD":
vect[2] = 1
elif emo == "BOTH_HS":
vect[3] = 1
vects.append(vect)
return vects, vocab
# extract features autoencoder plus n-gram bow
def extractFeaturesMulti(features=["auto_false", "bow", "targetInTweet", "emoticons", "affect", "w2v", "bow_phrase"]
, automodel="model.ckpt", w2vmodel="skip_nostop_multi_300features_10minwords_10context", phrasemodel="phrase.model",
useDev=True):
if useDev==False:
tweets_train, targets_train, labels_train = readTweetsOfficial(tokenize_tweets.FILETRAIN, 'windows-1252', 2)
tweets_dev, targets_dev, labels_dev = readTweetsOfficial(tokenize_tweets.FILEDEV, 'windows-1252', 2)
else:
tweets_train, targets_train, labels_train = readTweetsOfficial(tokenize_tweets.FILETRAIN, 'windows-1252', 2)
tweets_origdev, targets_origdev, labels_origdev = readTweetsOfficial(tokenize_tweets.FILEDEV, 'windows-1252', 2)
tweets_train.extend(tweets_origdev)
targets_train.extend(targets_origdev)
labels_train.extend(labels_origdev)
tweets_dev, targets_dev, labels_dev = readTweetsOfficial(tokenize_tweets.FILETEST, 'windows-1252', 2)
features_final = []
if features.__contains__("bow"):
features_final = extractFeatureVocab(tweets_train)
features_train = extractFeaturesBOW(tweets_train, targets_train, features_final)
features_dev = extractFeaturesBOW(tweets_dev, targets_dev, features_final)
elif features.__contains__("targetInTweet"):
features_train = extractFeaturesCrossTweetTarget(tweets_train, targets_train)
features_dev = extractFeaturesCrossTweetTarget(tweets_dev, targets_dev)
features_final.append("targetInTweet")
if features.__contains__("bow_phrase") or features.__contains__("bow_phrase_anon"):
if features.__contains__("bow_phrase"):
features_vocab = extractFeatureVocab(tweets_train, usephrasemodel=True)
features_train_phrbow = extractFeaturesBOW(tweets_train, targets_train, features_vocab, usephrasemodel=True)
features_dev_phrbow = extractFeaturesBOW(tweets_dev, targets_dev, features_vocab, usephrasemodel=True)
elif features.__contains__("bow_phrase_anon"):
features_vocab = extractFeatureVocab(tweets_train, usephrasemodel=True, anon_targets=True)
features_train_phrbow = extractFeaturesBOW(tweets_train, targets_train, features_vocab, usephrasemodel=True, anon_targets=True)
features_dev_phrbow = extractFeaturesBOW(tweets_dev, targets_dev, features_vocab, usephrasemodel=True, anon_targets=True)
features_final.extend(features_vocab)
if features.__contains__("auto_added"):
useph=False
if "phrase" in automodel:
useph=True
features_train_auto, labels_train, features_dev_auto, labels_dev = extractFeaturesAutoencoder(automodel, tweets_train, targets_train, labels_train, tweets_dev, targets_dev, labels_dev, "added", usephrasemodel=useph)
elif features.__contains__("auto_true"):
useph=False
if "phrase" in automodel:
useph=True
features_train_auto, labels_train, features_dev_auto, labels_dev = extractFeaturesAutoencoder(automodel, tweets_train, targets_train, labels_train, tweets_dev, targets_dev, labels_dev, "true", usephrasemodel=useph)
elif features.__contains__("auto_false"):
useph=False
if "phrase" in automodel:
useph=True
features_train_auto, labels_train, features_dev_auto, labels_dev = extractFeaturesAutoencoder(automodel, tweets_train, targets_train, labels_train, tweets_dev, targets_dev, labels_dev, "false", usephrasemodel=useph)
targetInTweetTrain = []
targetInTweetDev = []
if features.__contains__("targetInTweet") and features.__contains__("bow"):
targetInTweetTrain = extractFeaturesCrossTweetTarget(tweets_train, targets_train)
targetInTweetDev = extractFeaturesCrossTweetTarget(tweets_dev, targets_dev)
features_final.append("targetInTweet")
if features.__contains__("emoticons"):
emoticons_train, emoticons_vocab = extractEmoticons(tweets_train)
emoticons_dev, emoticons_vocab = extractEmoticons(tweets_dev)
for emo in emoticons_vocab:
features_final.append("Emoticon_" + emo)
if features.__contains__("affect"):
affect_train, affect_vocab = getAffect(tweets_train)
affect_dev, affect_vocab = getAffect(tweets_dev)
for aff in affect_vocab:
features_final.append("WNaffect_" + aff)
if features.__contains__("hash"):
phmodel = Phrases.load(phrasemodel)
w2vmodel = word2vec.Word2Vec.load(w2vmodel)
features_train_w2v, features_w2v_vocab = extractW2VHashFeatures(w2vmodel, phmodel, "hash", tweets_train, targets_train, labels_train)
features_dev_w2v, features_w2v_vocab = extractW2VHashFeatures(w2vmodel, phmodel, "hash", tweets_dev, targets_dev, labels_dev)
elif features.__contains__("w2v_hash"): # this contains hash
phmodel = Phrases.load(phrasemodel)
w2vmodel = word2vec.Word2Vec.load(w2vmodel)
features_train_w2v, features_w2v_vocab = extractW2VHashFeatures(w2vmodel, phmodel, "w2v_hash", tweets_train, targets_train, labels_train)
features_dev_w2v, features_w2v_vocab = extractW2VHashFeatures(w2vmodel, phmodel, "w2v_hash", tweets_dev, targets_dev, labels_dev)
# combine features
for i, featvec in enumerate(features_train):#features_train_auto)
if features.__contains__("auto_added") or features.__contains__("auto_true") or features.__contains__("auto_false"):
features_train[i] = np.append(features_train[i], features_train_auto[i]) # numpy append works as extend works for python lists
if features.__contains__("targetInTweet") and features.__contains__("bow"):
features_train[i] = np.append(features_train[i], targetInTweetTrain[i])
if features.__contains__("bow_phrase") or features.__contains__("bow_phrase_anon"):
features_train[i] = np.append(features_train[i], features_train_phrbow[i])
if features.__contains__("emoticons"):
features_train[i] = np.append(features_train[i], emoticons_train[i])
if features.__contains__("affect"):
features_train[i] = np.append(features_train[i], affect_train[i])
if features.__contains__("w2v_hash") or features.__contains__("hash"):
features_train[i] = np.append(features_train[i], features_train_w2v[i])
for i, featvec in enumerate(features_dev):#features_dev_auto):
if features.__contains__("auto_added") or features.__contains__("auto_true") or features.__contains__("auto_false"):
features_dev[i] = np.append(features_dev[i], features_dev_auto[i])
if features.__contains__("targetInTweet") and features.__contains__("bow"):
features_dev[i] = np.append(features_dev[i], targetInTweetDev[i])
if features.__contains__("bow_phrase") or features.__contains__("bow_phrase_anon"):
features_dev[i] = np.append(features_dev[i], features_dev_phrbow[i])
if features.__contains__("emoticons"):
features_dev[i] = np.append(features_dev[i], emoticons_dev[i])
if features.__contains__("affect"):
features_dev[i] = np.append(features_dev[i], affect_dev[i])
if features.__contains__("w2v_hash") or features.__contains__("hash"):
features_dev[i] = np.append(features_dev[i], features_dev_w2v[i])
return features_train, labels_train, features_dev, labels_dev, features_final
if __name__ == '__main__':
# Options: "auto_false", "bow", "targetInTweet", "emoticons", "affect", "w2v", "hash", "bow_phrase"
features_train, labels_train, features_dev, labels_dev, feature_vocab = extractFeaturesMulti(["auto_false", "targetInTweet"],
"model_phrase_100_samp500_it2000.ckpt", useDev=False)# "model_trump_phrase_100_samp500_it2600.ckpt")#"model_100_samp500.ckpt")
#train_classifiers_TopicVOpinion(features_train, labels_train, features_dev, labels_dev, "out.txt")
# train_classifier_3waySGD is another option, for testing elastic net regularisation, doesn't work as well as just l2 though
train_classifier_3way(features_train, labels_train, features_dev, labels_dev, "out_hillary_bow_targetInTweet.txt", feature_vocab, "false", "false", useDev=False)
#train_classifiers_PosVNeg(features_train, labels_train, features_dev, labels_dev, "out.txt")
eval(tokenize_tweets.FILEDEV, "out_hillary_bow_targetInTweet.txt")