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
|
import numpy as np
from sklearn.metrics import euclidean_distances
from sklearn.neighbors import KNeighborsClassifier
from sklearn.preprocessing import normalize
from sklearn.utils import check_array
import ot
from lapjv import lapjv
from mosestokenizer import MosesTokenizer
from pathos.multiprocessing import ProcessingPool as Pool
class Wasserstein_Matcher(KNeighborsClassifier):
"""
Implements a nearest neighbors classifier for input distributions using the Wasserstein distance as metric.
Source and target distributions are l_1 normalized before computing the Wasserstein distance.
Wasserstein is parametrized by the distances between the individual points of the distributions.
"""
def __init__(self,
W_embed,
n_neighbors=1,
n_jobs=1,
verbose=False,
sinkhorn=False,
sinkhorn_reg=0.1):
"""
Initialization of the class.
Arguments
---------
W_embed: embeddings of the words, np.array
verbose: True/False
"""
self.sinkhorn = sinkhorn
self.sinkhorn_reg = sinkhorn_reg
self.W_embed = W_embed
self.verbose = verbose
super(Wasserstein_Matcher, self).__init__(n_neighbors=n_neighbors,
n_jobs=n_jobs,
metric='precomputed',
algorithm='brute')
def _wmd(self, i, row, X_train):
union_idx = np.union1d(X_train[i].indices, row.indices)
W_minimal = self.W_embed[union_idx]
W_dist = euclidean_distances(W_minimal)
bow_i = X_train[i, union_idx].A.ravel()
bow_j = row[:, union_idx].A.ravel()
if self.sinkhorn:
return ot.sinkhorn2(
bow_i,
bow_j,
W_dist,
self.sinkhorn_reg,
numItermax=50,
method='sinkhorn_stabilized',
)[0]
else:
return ot.emd2(bow_i, bow_j, W_dist)
def _wmd_row(self, row):
X_train = self._fit_X
n_samples_train = X_train.shape[0]
return [self._wmd(i, row, X_train) for i in range(n_samples_train)]
def _pairwise_wmd(self, X_test, X_train=None):
n_samples_test = X_test.shape[0]
if X_train is None:
X_train = self._fit_X
pool = Pool(nodes=self.n_jobs
) # Parallelization of the calculation of the distances
dist = pool.map(self._wmd_row, X_test)
return np.array(dist)
def fit(self, X, y): # X_train_idf
X = check_array(X, accept_sparse='csr',
copy=True) # check if array is sparse
X = normalize(X, norm='l1', copy=False)
return super(Wasserstein_Matcher, self).fit(
X, y) # X_train_idf, np_ones(document collection size)
def predict(self, X):
X = check_array(X, accept_sparse='csr', copy=True)
X = normalize(X, norm='l1', copy=False)
dist = self._pairwise_wmd(X)
dist = dist * 1000 # for lapjv, small floating point numbers are evil
return super(Wasserstein_Matcher, self).predict(dist)
def kneighbors(self, X, n_neighbors=1): # X : X_train_idf
X = check_array(X, accept_sparse='csr', copy=True)
X = normalize(X, norm='l1', copy=False)
dist = self._pairwise_wmd(X)
dist = dist * 1000 # for lapjv, small floating point numbers are evil
return lapjv(dist) # and here is the matching part
class Wasserstein_Retriever(KNeighborsClassifier):
"""
Implements a nearest neighbors classifier for input distributions using the Wasserstein distance as metric.
Source and target distributions are l_1 normalized before computing the Wasserstein distance.
Wasserstein is parametrized by the distances between the individual points of the distributions.
"""
def __init__(self,
W_embed,
n_neighbors=1,
n_jobs=1,
verbose=False,
sinkhorn=False,
sinkhorn_reg=0.1):
"""
Initialization of the class.
Arguments
---------
W_embed: embeddings of the words, np.array
verbose: True/False
"""
self.sinkhorn = sinkhorn
self.sinkhorn_reg = sinkhorn_reg
self.W_embed = W_embed
self.verbose = verbose
super(Wasserstein_Retriever, self).__init__(n_neighbors=n_neighbors,
n_jobs=n_jobs,
metric='precomputed',
algorithm='brute')
def _wmd(self, i, row, X_train):
union_idx = np.union1d(X_train[i].indices, row.indices)
W_minimal = self.W_embed[union_idx]
W_dist = euclidean_distances(W_minimal)
bow_i = X_train[i, union_idx].A.ravel()
bow_j = row[:, union_idx].A.ravel()
if self.sinkhorn:
return ot.sinkhorn2(
bow_i,
bow_j,
W_dist,
self.sinkhorn_reg,
numItermax=50,
method='sinkhorn_stabilized',
)[0]
else:
return ot.emd2(bow_i, bow_j, W_dist)
def _wmd_row(self, row):
X_train = self._fit_X
n_samples_train = X_train.shape[0]
return [self._wmd(i, row, X_train) for i in range(n_samples_train)]
def _pairwise_wmd(self, X_test, X_train=None):
n_samples_test = X_test.shape[0]
if X_train is None:
X_train = self._fit_X
pool = Pool(nodes=self.n_jobs)
dist = pool.map(self._wmd_row, X_test)
return np.array(dist)
def fit(self, X, y):
X = check_array(X, accept_sparse='csr', copy=True)
X = normalize(X, norm='l1', copy=False)
return super(Wasserstein_Retriever, self).fit(X, y)
def predict(self, X):
X = check_array(X, accept_sparse='csr', copy=True)
X = normalize(X, norm='l1', copy=False)
dist = self._pairwise_wmd(X)
return super(Wasserstein_Retriever, self).predict(dist)
def kneighbors(self, X, n_neighbors=1):
X = check_array(X, accept_sparse='csr', copy=True)
X = normalize(X, norm='l1', copy=False)
dist = self._pairwise_wmd(X)
return super(Wasserstein_Retriever, self).kneighbors(dist, n_neighbors)
def align(self, X, n_neighbors=1):
"""
Wrapper function over kneighbors to return
precision at one and percentage values
"""
_, preds = self.kneighbors(X, n_neighbors)
_, p_at_one = mrr_precision_at_k(list(range(len(preds))), preds)
percentage = p_at_one * 100
return (p_at_one, percentage)
def load_embeddings(path, dimension=300):
"""
Loads the embeddings from a word2vec formatted file.
word2vec format is one line per word and it's associated embedding
(dimension x floating numbers) separated by spaces
The first line may or may not include the word count and dimension
"""
vectors = {}
with open(path, mode='r', encoding='utf8') as fp:
first_line = fp.readline().rstrip('\n')
if first_line.count(' ') == 1:
# includes the "word_count dimension" information
(word_count, dimension) = map(int, first_line.split())
else:
# assume the file only contains vectors
fp.seek(0)
for line in fp:
elems = line.split()
vectors[" ".join(elems[:-dimension])] = " ".join(
elems[-dimension:])
return vectors
def clean_corpus_using_embeddings_vocabulary(
embeddings_dictionary,
corpus,
vectors,
language,
):
'''
Cleans corpus using the dictionary of embeddings.
Any word without an associated embedding in the dictionary is ignored.
Adds '__target-language' and '__source-language' at the end of the words according to their language.
'''
clean_corpus, clean_vectors, keys = [], {}, []
words_we_want = set(embeddings_dictionary)
tokenize = MosesTokenizer(language)
for key, doc in enumerate(corpus):
clean_doc = []
words = tokenize(doc)
for word in words:
if word in words_we_want:
clean_doc.append(word + '__%s' % language)
clean_vectors[word + '__%s' % language] = np.array(
vectors[word].split()).astype(np.float)
if len(clean_doc) > 3 and len(clean_doc) < 25:
keys.append(key)
clean_corpus.append(' '.join(clean_doc))
tokenize.close()
return np.array(clean_corpus), clean_vectors, keys
def mrr_precision_at_k(golden, preds, k_list=[1,]):
"""
Calculates Mean Reciprocal Error and Hits@1 == Precision@1
"""
my_score = 0
precision_at = np.zeros(len(k_list))
for key, elem in enumerate(golden):
if elem in preds[key]:
location = np.where(preds[key] == elem)[0][0]
my_score += 1 / (1 + location)
for k_index, k_value in enumerate(k_list):
if location < k_value:
precision_at[k_index] += 1
return my_score / len(golden), (precision_at / len(golden))[0]
|
