ctl commited on
Commit
8cc63c8
1 Parent(s): 9b7e9c3

Update cer.py

Browse files
Files changed (1) hide show
  1. cer.py +39 -25
cer.py CHANGED
@@ -1,4 +1,4 @@
1
- # coding=utf-8
2
  # Copyright 2021 The HuggingFace Datasets Authors.
3
  #
4
  # Licensed under the Apache License, Version 2.0 (the "License");
@@ -12,12 +12,35 @@
12
  # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
  # See the License for the specific language governing permissions and
14
  # limitations under the License.
15
- """ Word Error Ratio (WER) metric. """
16
 
17
- from jiwer import wer
 
 
18
 
19
  import datasets
20
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
 
22
  _CITATION = """\
23
  @inproceedings{inproceedings,
@@ -30,15 +53,13 @@ _CITATION = """\
30
  """
31
 
32
  _DESCRIPTION = """\
33
- Word error rate (WER) is a common metric of the performance of an automatic speech recognition system.
34
-
35
- The general difficulty of measuring performance lies in the fact that the recognized word sequence can have a different length from the reference word sequence (supposedly the correct one). The WER is derived from the Levenshtein distance, working at the word level instead of the phoneme level. The WER is a valuable tool for comparing different systems as well as for evaluating improvements within one system. This kind of measurement, however, provides no details on the nature of translation errors and further work is therefore required to identify the main source(s) of error and to focus any research effort.
36
 
37
- This problem is solved by first aligning the recognized word sequence with the reference (spoken) word sequence using dynamic string alignment. Examination of this issue is seen through a theory called the power law that states the correlation between perplexity and word error rate.
38
 
39
- Word error rate can then be computed as:
40
 
41
- WER = (S + D + I) / N = (S + D + I) / (S + D + C)
42
 
43
  where
44
 
@@ -48,25 +69,25 @@ I is the number of insertions,
48
  C is the number of correct words,
49
  N is the number of words in the reference (N=S+D+C).
50
 
51
- WER's output is always a number between 0 and 1. This value indicates the percentage of words that were incorrectly predicted. The lower the value, the better the
52
- performance of the ASR system with a WER of 0 being a perfect score.
53
  """
54
 
55
  _KWARGS_DESCRIPTION = """
56
- Computes WER score of transcribed segments against references.
57
  Args:
58
  references: list of references for each speech input.
59
  predictions: list of transcribtions to score.
60
  Returns:
61
- (float): the word error rate
62
 
63
  Examples:
64
 
65
  >>> predictions = ["this is the prediction", "there is an other sample"]
66
  >>> references = ["this is the reference", "there is another one"]
67
- >>> wer = datasets.load_metric("wer")
68
- >>> wer_score = wer.compute(predictions=predictions, references=references)
69
- >>> print(wer_score)
70
  0.5
71
  """
72
 
@@ -87,16 +108,9 @@ class CER(datasets.Metric):
87
  codebase_urls=["https://github.com/jitsi/jiwer/"],
88
  reference_urls=[
89
  "https://en.wikipedia.org/wiki/Word_error_rate",
 
90
  ],
91
  )
92
 
93
  def _compute(self, predictions, references):
94
- preds = [char for seq in predictions for char in list(seq)]
95
- refs = [char for seq in references for char in list(seq)]
96
- return wer(refs, preds)
97
- """
98
- wers = []
99
- for pred, ref in zip(predictions, references):
100
- wers.append(wer(list(ref), list(pred)))
101
- return sum(wers) / len(wers)
102
- """
 
1
+ # coding=utf-8
2
  # Copyright 2021 The HuggingFace Datasets Authors.
3
  #
4
  # Licensed under the Apache License, Version 2.0 (the "License");
 
12
  # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
  # See the License for the specific language governing permissions and
14
  # limitations under the License.
15
+ """ Character Error Ratio (CER) metric. """
16
 
17
+ import jiwer
18
+ import jiwer.transforms as tr
19
+ from typing import List
20
 
21
  import datasets
22
 
23
+ class SentencesToListOfCharacters(tr.AbstractTransform):
24
+ def process_string(self, s: str):
25
+ return list(s)
26
+
27
+ def process_list(self, inp: List[str]):
28
+ chars = []
29
+
30
+ for sentence in inp:
31
+ chars.extend(self.process_string(sentence))
32
+
33
+ return chars
34
+
35
+
36
+ cer_transform = tr.Compose(
37
+ [
38
+ tr.RemoveMultipleSpaces(),
39
+ tr.Strip(),
40
+ SentencesToListOfCharacters(),
41
+ ]
42
+ )
43
+
44
 
45
  _CITATION = """\
46
  @inproceedings{inproceedings,
 
53
  """
54
 
55
  _DESCRIPTION = """\
56
+ Character error rate (CER) is a common metric of the performance of an automatic speech recognition system.
 
 
57
 
58
+ CER is similar to Word Error Rate (WER), but operate on character insted of word. Please refer to docs of WER for further information.
59
 
60
+ Character error rate can be computed as:
61
 
62
+ CER = (S + D + I) / N = (S + D + I) / (S + D + C)
63
 
64
  where
65
 
 
69
  C is the number of correct words,
70
  N is the number of words in the reference (N=S+D+C).
71
 
72
+ CER's output is always a number between 0 and 1. This value indicates the percentage of characters that were incorrectly predicted. The lower the value, the better the
73
+ performance of the ASR system with a CER of 0 being a perfect score.
74
  """
75
 
76
  _KWARGS_DESCRIPTION = """
77
+ Computes CER score of transcribed segments against references.
78
  Args:
79
  references: list of references for each speech input.
80
  predictions: list of transcribtions to score.
81
  Returns:
82
+ (float): the character error rate
83
 
84
  Examples:
85
 
86
  >>> predictions = ["this is the prediction", "there is an other sample"]
87
  >>> references = ["this is the reference", "there is another one"]
88
+ >>> cer = datasets.load_metric("cer")
89
+ >>> cer_score = cer.compute(predictions=predictions, references=references)
90
+ >>> print(cer_score)
91
  0.5
92
  """
93
 
 
108
  codebase_urls=["https://github.com/jitsi/jiwer/"],
109
  reference_urls=[
110
  "https://en.wikipedia.org/wiki/Word_error_rate",
111
+ "https://sites.google.com/site/textdigitisation/qualitymeasures/computingerrorrates#whitespace",
112
  ],
113
  )
114
 
115
  def _compute(self, predictions, references):
116
+ return jiwer.wer(references, predictions, truth_transform=cer_transform, hypothesis_transform=cer_transform)