-
Notifications
You must be signed in to change notification settings - Fork 8
/
topic_coherence.py
125 lines (109 loc) · 3.92 KB
/
topic_coherence.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
import pandas as pd
import numpy as np
from cluster import nmf_articles
from itertools import permutations
from progressbar import ProgressBar, Percentage
import matplotlib.pyplot as plt
import seaborn as sns
def get_avg_coherence(df, n_topics):
print '{} Topics Processing...'.format(n_topics)
nmf, X, W, W_percent, labels, topic_words, feature_names, reverse_lookup = nmf_articles(df, n_topics=n_topics, n_features=10000, random_state=1, max_df=0.8, min_df=5)
print 'Factorizing Done...'
pbar = ProgressBar()
coherence = []
for words in pbar(topic_words):
coherence.append(topic_coherence(X, reverse_lookup, words))
print '\n'
return np.mean(coherence)
def topic_coherence(X, reverse_lookup, topic_words, e=0.01, num_top_words=10):
'''
Using the UMass topic coherence as definied in http://anthology.aclweb.org/D/D12/D12-1087.pdf
'''
if not num_top_words:
num_top_words = len(topic_words)
result = 0.0
perm = permutations(topic_words[:num_top_words], 2)
while True:
try:
word1, word2 = perm.next()
result += coherence_score(X, reverse_lookup, word1, word2, e)
except:
return result
def coherence_score(X, reverse_lookup, word1, word2, e):
return np.log((document_word_count(X, reverse_lookup, word1, word2) + e) / document_word_count(X, reverse_lookup, word2))
def document_word_count(X, reverse_lookup, word1, word2=None):
'''
INPUTS: X - tfidf matrix
reverse_lookup - Dictionary relating word to the cooresponding column in the tfidf matrix
word1 - First word to search for
word2 - Second word to search for.
OUTPUT: Int - Number of documents in the corpus that have either word1 or word1 AND word2
'''
if word2:
return len(np.where((X[:, reverse_lookup[word1]] > 0).toarray() & (X[:, reverse_lookup[word2]] > 0).toarray())[0])
else:
return len(np.where((X[:, reverse_lookup[word1]] > 0).toarray())[0])
def make_coherence_plot(n_topics, coherence, show=False):
fig = plt.figure(figsize=(12, 8))
plt.plot(n_topics, coherence)
plt.xlabel('Number of Topics')
plt.ylabel('Avg Coherence Score')
plt.suptitle('Average Coherence Score Among Topics', fontsize=18)
plt.subplots_adjust(left=0.08, bottom=0.09, right=0.95, top=0.91, hspace=0.16)
if show:
plt.show()
else:
plt.savefig('./plots/nmf_coherence.png', dpi=350)
if __name__=='__main__':
df = pd.read_pickle('election_data.pkl')
num_topics = range(2, 31, 2) + range(35, 101, 5) + range(110, 201, 10) + range(225, 401, 25)
avg_coherence = [get_avg_coherence(df, n_topic) for n_topic in num_topics]
make_coherence_plot(num_topics, avg_coherence)
# topics = range(2, 31, 2) + range(35, 101, 5) + range(110, 201, 10) + range(225, 401, 25)
# avg_coherence = [-74.949694804000032,
# -79.499050543893304,
# -83.20766537930065,
# -88.350138020815962,
# -90.806811290179454,
# -88.966148457213947,
# -92.259849347336086,
# -93.693694511355105,
# -96.229715459278808,
# -101.02106266796712,
# -105.26140489793111,
# -106.44992830555778,
# -111.19900256168015,
# -108.3888383643815,
# -110.42210903014306,
# -113.06946716707321,
# -114.56167197849281,
# -116.51518854832943,
# -120.75116955758651,
# -122.07414946308113,
# -122.73389201912455,
# -123.20835369178577,
# -124.4706818825221,
# -125.2930064826773,
# -126.67173047873919,
# -126.11364157215765,
# -126.91972720705586,
# -127.58764173111905,
# -130.76500367061993,
# -133.18690997621755,
# -144.88052698246312,
# -156.09331805752294,
# -192.39619162591814,
# -193.04223270498383,
# -228.03546894957145,
# -254.53502103391168,
# -286.54564970701426,
# -294.97466219245791,
# -320.66635603102822,
# -365.14886227907732,
# -397.78072283424865,
# -421.15376018755478,
# -440.271437412821,
# -465.41217065103001,
# -472.07237254507726,
# -490.61559381230097,
# -499.6807662116052]