/
chrF_pp.py
272 lines (210 loc) · 10.2 KB
/
chrF_pp.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# **Note: This code is adopted from https://github.com/m-popovic/chrF
# Copyright 2017 Maja Popovic
# The program is distributed under the terms
# of the GNU General Public Licence (GPL)
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU General Public License for more details.
# You should have received a copy of the GNU General Public License
# along with this program. If not, see <http://www.gnu.org/licenses/>.
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
# Publications of results obtained through the use of original or
# modified versions of the software have to cite the authors by refering
# to the following publication:
# Maja Popović (2015).
# "chrF: character n-gram F-score for automatic MT evaluation".
# In Proceedings of the Tenth Workshop on Statistical Machine Translation (WMT15), pages 392–395
# Lisbon, Portugal, September 2015.
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
import sys
import math
import unicodedata
import argparse
from collections import defaultdict
import time
import string
def separate_characters(line):
return list(line.strip().replace(" ", ""))
def separate_punctuation(line):
words = line.strip().split()
tokenized = []
for w in words:
if len(w) == 1:
tokenized.append(w)
else:
lastChar = w[-1]
firstChar = w[0]
if lastChar in string.punctuation:
tokenized += [w[:-1], lastChar]
elif firstChar in string.punctuation:
tokenized += [firstChar, w[1:]]
else:
tokenized.append(w)
return tokenized
def ngram_counts(wordList, order):
counts = defaultdict(lambda: defaultdict(float))
nWords = len(wordList)
for i in range(nWords):
for j in range(1, order+1):
if i+j <= nWords:
ngram = tuple(wordList[i:i+j])
counts[j-1][ngram]+=1
return counts
def ngram_matches(ref_ngrams, hyp_ngrams):
matchingNgramCount = defaultdict(float)
totalRefNgramCount = defaultdict(float)
totalHypNgramCount = defaultdict(float)
for order in ref_ngrams:
for ngram in hyp_ngrams[order]:
totalHypNgramCount[order] += hyp_ngrams[order][ngram]
for ngram in ref_ngrams[order]:
totalRefNgramCount[order] += ref_ngrams[order][ngram]
if ngram in hyp_ngrams[order]:
matchingNgramCount[order] += min(ref_ngrams[order][ngram], hyp_ngrams[order][ngram])
return matchingNgramCount, totalRefNgramCount, totalHypNgramCount
def ngram_precrecf(matching, reflen, hyplen, beta):
ngramPrec = defaultdict(float)
ngramRec = defaultdict(float)
ngramF = defaultdict(float)
factor = beta**2
for order in matching:
if hyplen[order] > 0:
ngramPrec[order] = matching[order]/hyplen[order]
else:
ngramPrec[order] = 1e-16
if reflen[order] > 0:
ngramRec[order] = matching[order]/reflen[order]
else:
ngramRec[order] = 1e-16
denom = factor*ngramPrec[order] + ngramRec[order]
if denom > 0:
ngramF[order] = (1+factor)*ngramPrec[order]*ngramRec[order] / denom
else:
ngramF[order] = 1e-16
return ngramF, ngramRec, ngramPrec
def computeChrF(fpRef, fpHyp, nworder, ncorder, beta, sentence_level_scores = None):
norder = float(nworder + ncorder)
# initialisation of document level scores
totalMatchingCount = defaultdict(float)
totalRefCount = defaultdict(float)
totalHypCount = defaultdict(float)
totalChrMatchingCount = defaultdict(float)
totalChrRefCount = defaultdict(float)
totalChrHypCount = defaultdict(float)
averageTotalF = 0.0
nsent = 0
for hline, rline in zip(fpHyp, fpRef):
nsent += 1
# preparation for multiple references
maxF = 0.0
bestWordMatchingCount = None
bestCharMatchingCount = None
hypNgramCounts = ngram_counts(separate_punctuation(hline), nworder)
hypChrNgramCounts = ngram_counts(separate_characters(hline), ncorder)
# going through multiple references
refs = rline.split("*#")
for ref in refs:
refNgramCounts = ngram_counts(separate_punctuation(ref), nworder)
refChrNgramCounts = ngram_counts(separate_characters(ref), ncorder)
# number of overlapping n-grams, total number of ref n-grams, total number of hyp n-grams
matchingNgramCounts, totalRefNgramCount, totalHypNgramCount = ngram_matches(refNgramCounts, hypNgramCounts)
matchingChrNgramCounts, totalChrRefNgramCount, totalChrHypNgramCount = ngram_matches(refChrNgramCounts, hypChrNgramCounts)
# n-gram f-scores, recalls and precisions
ngramF, ngramRec, ngramPrec = ngram_precrecf(matchingNgramCounts, totalRefNgramCount, totalHypNgramCount, beta)
chrNgramF, chrNgramRec, chrNgramPrec = ngram_precrecf(matchingChrNgramCounts, totalChrRefNgramCount, totalChrHypNgramCount, beta)
sentRec = (sum(chrNgramRec.values()) + sum(ngramRec.values())) / norder
sentPrec = (sum(chrNgramPrec.values()) + sum(ngramPrec.values())) / norder
sentF = (sum(chrNgramF.values()) + sum(ngramF.values())) / norder
if sentF > maxF:
maxF = sentF
bestMatchingCount = matchingNgramCounts
bestRefCount = totalRefNgramCount
bestHypCount = totalHypNgramCount
bestChrMatchingCount = matchingChrNgramCounts
bestChrRefCount = totalChrRefNgramCount
bestChrHypCount = totalChrHypNgramCount
# all the references are done
# write sentence level scores
if sentence_level_scores:
sentence_level_scores.write("%i::c%i+w%i-F%i\t%.4f\n" % (nsent, ncorder, nworder, beta, 100*maxF))
# collect document level ngram counts
for order in range(nworder):
totalMatchingCount[order] += bestMatchingCount[order]
totalRefCount[order] += bestRefCount[order]
totalHypCount[order] += bestHypCount[order]
for order in range(ncorder):
totalChrMatchingCount[order] += bestChrMatchingCount[order]
totalChrRefCount[order] += bestChrRefCount[order]
totalChrHypCount[order] += bestChrHypCount[order]
averageTotalF += maxF
# all sentences are done
# total precision, recall and F (aritmetic mean of all ngrams)
totalNgramF, totalNgramRec, totalNgramPrec = ngram_precrecf(totalMatchingCount, totalRefCount, totalHypCount, beta)
totalChrNgramF, totalChrNgramRec, totalChrNgramPrec = ngram_precrecf(totalChrMatchingCount, totalChrRefCount, totalChrHypCount, beta)
totalF = (sum(totalChrNgramF.values()) + sum(totalNgramF.values())) / norder
averageTotalF = averageTotalF / nsent
totalRec = (sum(totalChrNgramRec.values()) + sum(totalNgramRec.values())) / norder
totalPrec = (sum(totalChrNgramPrec.values()) + sum(totalNgramPrec.values())) / norder
return totalF, averageTotalF, totalPrec, totalRec
def corpus_chrf(ref, hyp, ncorder=6, nworder=2, beta=2, sent = None):
"""
About:
Get chrF++ score (modified chrF) of a hypothesis translation given a reference translation.
Args:
@ref: (List), List of reference sentences.
@hyp: (List), List of target/hypothesis sentences.
@ncorder: (int), character n-gram order (default value is 6)
@nworder: (int), word n-gram order (default value is 2)
@beta: (float), beta parameter (default=2.0)
@sent: (bool), if True then print sentences label scores.
Outputs:
@totalF: (float), overall document/corpus level F-score (scaled to 100)
@averageTotalF: (float), overal macro-averaged document level F-score (arithmetic average of the sentence level scores)
Use:
import chrF_pp
ref = ["this is a test.", "I want a glass of water"]
hyp = ["this is a test.", "He is crossing the road."]
chrF_pp.corpus_chrf(ref,hyp,sent=True) #Sentence-level and corpus-level
chrF_pp.corpus_chrf(ref,hyp) #corpus-level
"""
sentence_level_scores = None
if sent:
sentence_level_scores = sys.stdout # Or stderr?
totalF, averageTotalF, totalPrec, totalRec = computeChrF(ref, hyp, nworder, ncorder, beta, sentence_level_scores)
totalF = totalF * 100
averageTotalF = averageTotalF * 100
totalPrec = totalPrec * 100
totalRec = totalRec * 100
scores = {"totalF":totalF, "averageTotalF":averageTotalF, "totalPrec":totalPrec, "totalRec":totalRec}
return scores
def main():
argParser = argparse.ArgumentParser()
argParser.add_argument("-R", "--reference", help="reference translation", required=True)
argParser.add_argument("-H", "--hypothesis", help="hypothesis translation", required=True)
args = argParser.parse_args()
rtxt = open(args.reference, 'r')
htxt = open(args.hypothesis, 'r')
ref = []
hyp = []
for hline, rline in zip(rtxt, htxt):
ref.append(hline)
hyp.append(rline)
htxt.close()
rtxt.close()
print("ref list: {}\n".format(ref))
print("hyp list: {}\n".format(hyp))
# Default params:
ncorder = 6; nworder = 2; beta=2; sent = None
# Get the scores
scores = corpus_chrf(ref, hyp, ncorder, nworder, beta, sent)
sys.stdout.write("start_time:\t%i\n" % (time.time()))
sys.stdout.write("c%i+w%i-F%i\t%.4f\n" % (ncorder, nworder, beta, scores["totalF"]))
sys.stdout.write("c%i+w%i-avgF%i\t%.4f\n" % (ncorder, nworder, beta, scores["averageTotalF"]))
#sys.stdout.write("c%i+w%i-Prec\t%.4f\n" % (args.ncorder, args.nworder, 100*totalPrec))
#sys.stdout.write("c%i+w%i-Rec\t%.4f\n" % (args.ncorder, args.nworder, 100*totalRec))
sys.stdout.write("end_time:\t%i\n" % (time.time()))
if __name__ == "__main__":
main()