-
Notifications
You must be signed in to change notification settings - Fork 20
/
rouge.py
355 lines (281 loc) · 10.3 KB
/
rouge.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
"""
Borrowed from https://github.com/tensorflow/nmt/blob/master/nmt/scripts/rouge.py
ROUGE metric implementation.
Copy from tf_seq2seq/seq2seq/metrics/rouge.py.
This is a modified and slightly extended verison of
https://github.com/miso-belica/sumy/blob/dev/sumy/evaluation/rouge.py.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import itertools
import numpy as np
#pylint: disable=C0103
def _get_ngrams(n, text):
"""Calcualtes n-grams.
Args:
n: which n-grams to calculate
text: An array of tokens
Returns:
A set of n-grams
"""
ngram_set = set()
text_length = len(text)
max_index_ngram_start = text_length - n
for i in range(max_index_ngram_start + 1):
ngram_set.add(tuple(text[i:i + n]))
return ngram_set
def _split_into_words(sentences):
"""Splits multiple sentences into words and flattens the result"""
return list(itertools.chain(*[_.split(" ") for _ in sentences]))
def _get_word_ngrams(n, sentences):
"""Calculates word n-grams for multiple sentences.
"""
assert len(sentences) > 0
assert n > 0
words = _split_into_words(sentences)
return _get_ngrams(n, words)
def _len_lcs(x, y):
"""
Returns the length of the Longest Common Subsequence between sequences x
and y.
Source: http://www.algorithmist.com/index.php/Longest_Common_Subsequence
Args:
x: sequence of words
y: sequence of words
Returns
integer: Length of LCS between x and y
"""
table = _lcs(x, y)
n, m = len(x), len(y)
return table[n, m]
def _lcs(x, y):
"""
Computes the length of the longest common subsequence (lcs) between two
strings. The implementation below uses a DP programming algorithm and runs
in O(nm) time where n = len(x) and m = len(y).
Source: http://www.algorithmist.com/index.php/Longest_Common_Subsequence
Args:
x: collection of words
y: collection of words
Returns:
Table of dictionary of coord and len lcs
"""
n, m = len(x), len(y)
table = dict()
for i in range(n + 1):
for j in range(m + 1):
if i == 0 or j == 0:
table[i, j] = 0
elif x[i - 1] == y[j - 1]:
table[i, j] = table[i - 1, j - 1] + 1
else:
table[i, j] = max(table[i - 1, j], table[i, j - 1])
return table
def _recon_lcs(x, y):
"""
Returns the Longest Subsequence between x and y.
Source: http://www.algorithmist.com/index.php/Longest_Common_Subsequence
Args:
x: sequence of words
y: sequence of words
Returns:
sequence: LCS of x and y
"""
i, j = len(x), len(y)
table = _lcs(x, y)
def _recon(i, j):
"""private recon calculation"""
if i == 0 or j == 0:
return []
elif x[i - 1] == y[j - 1]:
return _recon(i - 1, j - 1) + [(x[i - 1], i)]
elif table[i - 1, j] > table[i, j - 1]:
return _recon(i - 1, j)
else:
return _recon(i, j - 1)
recon_tuple = tuple(map(lambda x: x[0], _recon(i, j)))
return recon_tuple
def rouge_n(evaluated_sentences, reference_sentences, n=2):
"""
Computes ROUGE-N of two text collections of sentences.
Sourece: http://research.microsoft.com/en-us/um/people/cyl/download/
papers/rouge-working-note-v1.3.1.pdf
Args:
evaluated_sentences: The sentences that have been picked by the summarizer
reference_sentences: The sentences from the referene set
n: Size of ngram. Defaults to 2.
Returns:
A tuple (f1, precision, recall) for ROUGE-N
Raises:
ValueError: raises exception if a param has len <= 0
"""
if len(evaluated_sentences) <= 0 or len(reference_sentences) <= 0:
raise ValueError("Collections must contain at least 1 sentence.")
evaluated_ngrams = _get_word_ngrams(n, evaluated_sentences)
reference_ngrams = _get_word_ngrams(n, reference_sentences)
reference_count = len(reference_ngrams)
evaluated_count = len(evaluated_ngrams)
# Gets the overlapping ngrams between evaluated and reference
overlapping_ngrams = evaluated_ngrams.intersection(reference_ngrams)
overlapping_count = len(overlapping_ngrams)
# Handle edge case. This isn't mathematically correct, but it's good enough
if evaluated_count == 0:
precision = 0.0
else:
precision = overlapping_count / evaluated_count
if reference_count == 0:
recall = 0.0
else:
recall = overlapping_count / reference_count
f1_score = 2.0 * ((precision * recall) / (precision + recall + 1e-8))
# return overlapping_count / reference_count
return f1_score, precision, recall
def _f_p_r_lcs(llcs, m, n):
"""
Computes the LCS-based F-measure score
Source: http://research.microsoft.com/en-us/um/people/cyl/download/papers/
rouge-working-note-v1.3.1.pdf
Args:
llcs: Length of LCS
m: number of words in reference summary
n: number of words in candidate summary
Returns:
Float. LCS-based F-measure score
"""
r_lcs = llcs / m
p_lcs = llcs / n
beta = p_lcs / (r_lcs + 1e-12)
num = (1 + (beta**2)) * r_lcs * p_lcs
denom = r_lcs + ((beta**2) * p_lcs)
f_lcs = num / (denom + 1e-12)
return f_lcs, p_lcs, r_lcs
def rouge_l_sentence_level(evaluated_sentences, reference_sentences):
"""
Computes ROUGE-L (sentence level) of two text collections of sentences.
http://research.microsoft.com/en-us/um/people/cyl/download/papers/
rouge-working-note-v1.3.1.pdf
Calculated according to:
R_lcs = LCS(X,Y)/m
P_lcs = LCS(X,Y)/n
F_lcs = ((1 + beta^2)*R_lcs*P_lcs) / (R_lcs + (beta^2) * P_lcs)
where:
X = reference summary
Y = Candidate summary
m = length of reference summary
n = length of candidate summary
Args:
evaluated_sentences: The sentences that have been picked by the summarizer
reference_sentences: The sentences from the referene set
Returns:
A float: F_lcs
Raises:
ValueError: raises exception if a param has len <= 0
"""
if len(evaluated_sentences) <= 0 or len(reference_sentences) <= 0:
raise ValueError("Collections must contain at least 1 sentence.")
reference_words = _split_into_words(reference_sentences)
evaluated_words = _split_into_words(evaluated_sentences)
m = len(reference_words)
n = len(evaluated_words)
lcs = _len_lcs(evaluated_words, reference_words)
return _f_p_r_lcs(lcs, m, n)
def _union_lcs(evaluated_sentences, reference_sentence):
"""
Returns LCS_u(r_i, C) which is the LCS score of the union longest common
subsequence between reference sentence ri and candidate summary C. For example
if r_i= w1 w2 w3 w4 w5, and C contains two sentences: c1 = w1 w2 w6 w7 w8 and
c2 = w1 w3 w8 w9 w5, then the longest common subsequence of r_i and c1 is
"w1 w2" and the longest common subsequence of r_i and c2 is "w1 w3 w5". The
union longest common subsequence of r_i, c1, and c2 is "w1 w2 w3 w5" and
LCS_u(r_i, C) = 4/5.
Args:
evaluated_sentences: The sentences that have been picked by the summarizer
reference_sentence: One of the sentences in the reference summaries
Returns:
float: LCS_u(r_i, C)
ValueError:
Raises exception if a param has len <= 0
"""
if len(evaluated_sentences) <= 0:
raise ValueError("Collections must contain at least 1 sentence.")
lcs_union = set()
reference_words = _split_into_words([reference_sentence])
combined_lcs_length = 0
for eval_s in evaluated_sentences:
evaluated_words = _split_into_words([eval_s])
lcs = set(_recon_lcs(reference_words, evaluated_words))
combined_lcs_length += len(lcs)
lcs_union = lcs_union.union(lcs)
union_lcs_count = len(lcs_union)
union_lcs_value = union_lcs_count / combined_lcs_length
return union_lcs_value
def rouge_l_summary_level(evaluated_sentences, reference_sentences):
"""
Computes ROUGE-L (summary level) of two text collections of sentences.
http://research.microsoft.com/en-us/um/people/cyl/download/papers/
rouge-working-note-v1.3.1.pdf
Calculated according to:
R_lcs = SUM(1, u)[LCS<union>(r_i,C)]/m
P_lcs = SUM(1, u)[LCS<union>(r_i,C)]/n
F_lcs = ((1 + beta^2)*R_lcs*P_lcs) / (R_lcs + (beta^2) * P_lcs)
where:
SUM(i,u) = SUM from i through u
u = number of sentences in reference summary
C = Candidate summary made up of v sentences
m = number of words in reference summary
n = number of words in candidate summary
Args:
evaluated_sentences: The sentences that have been picked by the summarizer
reference_sentence: One of the sentences in the reference summaries
Returns:
A float: F_lcs
Raises:
ValueError: raises exception if a param has len <= 0
"""
if len(evaluated_sentences) <= 0 or len(reference_sentences) <= 0:
raise ValueError("Collections must contain at least 1 sentence.")
# total number of words in reference sentences
m = len(_split_into_words(reference_sentences))
# total number of words in evaluated sentences
n = len(_split_into_words(evaluated_sentences))
union_lcs_sum_across_all_references = 0
for ref_s in reference_sentences:
union_lcs_sum_across_all_references += _union_lcs(evaluated_sentences,
ref_s)
return _f_p_r_lcs(union_lcs_sum_across_all_references, m, n)
def rouge(hypotheses, references):
"""Calculates average rouge scores for a list of hypotheses and
references"""
# Filter out hyps that are of 0 length
# hyps_and_refs = zip(hypotheses, references)
# hyps_and_refs = [_ for _ in hyps_and_refs if len(_[0]) > 0]
# hypotheses, references = zip(*hyps_and_refs)
# Calculate ROUGE-1 F1, precision, recall scores
rouge_1 = [
rouge_n([hyp], [ref], 1) for hyp, ref in zip(hypotheses, references)
]
rouge_1_f, rouge_1_p, rouge_1_r = map(np.mean, zip(*rouge_1))
# Calculate ROUGE-2 F1, precision, recall scores
rouge_2 = [
rouge_n([hyp], [ref], 2) for hyp, ref in zip(hypotheses, references)
]
rouge_2_f, rouge_2_p, rouge_2_r = map(np.mean, zip(*rouge_2))
# Calculate ROUGE-L F1, precision, recall scores
rouge_l = [
rouge_l_sentence_level([hyp], [ref])
for hyp, ref in zip(hypotheses, references)
]
rouge_l_f, rouge_l_p, rouge_l_r = map(np.mean, zip(*rouge_l))
return {
"rouge_1/f_score": rouge_1_f,
"rouge_1/r_score": rouge_1_r,
"rouge_1/p_score": rouge_1_p,
"rouge_2/f_score": rouge_2_f,
"rouge_2/r_score": rouge_2_r,
"rouge_2/p_score": rouge_2_p,
"rouge_l/f_score": rouge_l_f,
"rouge_l/r_score": rouge_l_r,
"rouge_l/p_score": rouge_l_p,
}