Upload Custom WER.py
Browse files- Custom WER.py +171 -0
Custom WER.py
ADDED
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from sklearn.feature_extraction.text import CountVectorizer
|
2 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
3 |
+
import numpy as np
|
4 |
+
from typing import List, Optional
|
5 |
+
import pandas as pd
|
6 |
+
|
7 |
+
def cosine_sim_wer(references: List[str], predictions: List[str]) -> float:
|
8 |
+
"""
|
9 |
+
Calculate a WER-like metric based on cosine similarity between reference and prediction texts.
|
10 |
+
|
11 |
+
This function computes character-level n-gram similarities between each reference-prediction pair
|
12 |
+
and returns an error rate (100% - average similarity). Handles empty inputs and provides
|
13 |
+
detailed similarity statistics.
|
14 |
+
|
15 |
+
Args:
|
16 |
+
references: List of reference transcript strings
|
17 |
+
predictions: List of model prediction strings
|
18 |
+
|
19 |
+
Returns:
|
20 |
+
float: Error rate based on cosine similarity (100% - average similarity)
|
21 |
+
|
22 |
+
Example:
|
23 |
+
>>> references = ["hello world", "good morning"]
|
24 |
+
>>> predictions = ["hello world", "good evening"]
|
25 |
+
>>> error_rate = cosine_sim_wer(references, predictions)
|
26 |
+
"""
|
27 |
+
# Validate and clean inputs
|
28 |
+
valid_refs, valid_preds = [], []
|
29 |
+
|
30 |
+
for ref, pred in zip(references, predictions):
|
31 |
+
if not ref.strip() or not pred.strip():
|
32 |
+
continue # Skip empty strings
|
33 |
+
valid_refs.append(ref.strip())
|
34 |
+
valid_preds.append(pred.strip())
|
35 |
+
|
36 |
+
# Handle case with no valid pairs
|
37 |
+
if not valid_refs:
|
38 |
+
print("Warning: No valid reference-prediction pairs found")
|
39 |
+
return 100.0 # Maximum error if no valid data
|
40 |
+
|
41 |
+
# Calculate pairwise similarities
|
42 |
+
similarities = []
|
43 |
+
for ref, pred in zip(valid_refs, valid_preds):
|
44 |
+
try:
|
45 |
+
# Use character-level n-grams (2-3 chars) for robust comparison
|
46 |
+
vectorizer = CountVectorizer(
|
47 |
+
analyzer='char_wb', # Word-boundary aware character n-grams
|
48 |
+
ngram_range=(2, 3) # Bigrams and trigrams
|
49 |
+
)
|
50 |
+
|
51 |
+
# Create document-term matrices
|
52 |
+
vectors = vectorizer.fit_transform([ref, pred])
|
53 |
+
|
54 |
+
# Compute cosine similarity
|
55 |
+
similarity = cosine_similarity(vectors[0:1], vectors[1:2])[0][0]
|
56 |
+
similarities.append(similarity * 100) # Convert to percentage
|
57 |
+
|
58 |
+
except Exception as e:
|
59 |
+
print(f"Error calculating similarity: {e}")
|
60 |
+
similarities.append(0.0) # Default to 0% similarity on error
|
61 |
+
|
62 |
+
# Compute statistics
|
63 |
+
avg_similarity = np.mean(similarities)
|
64 |
+
min_similarity = np.min(similarities)
|
65 |
+
max_similarity = np.max(similarities)
|
66 |
+
error_rate = 100.0 - avg_similarity
|
67 |
+
|
68 |
+
# Print diagnostics
|
69 |
+
print(f"Similarity Statistics:")
|
70 |
+
print(f" - Average: {avg_similarity:.2f}%")
|
71 |
+
print(f" - Range: {min_similarity:.2f}% to {max_similarity:.2f}%")
|
72 |
+
print(f" - Valid samples: {len(similarities)}/{len(references)}")
|
73 |
+
|
74 |
+
return error_rate
|
75 |
+
|
76 |
+
|
77 |
+
def create_wer_analysis_dataframe(
|
78 |
+
references: List[str],
|
79 |
+
predictions: List[str],
|
80 |
+
normalized_references: Optional[List[str]] = None,
|
81 |
+
normalized_predictions: Optional[List[str]] = None,
|
82 |
+
output_csv: str = "wer_analysis.csv"
|
83 |
+
) -> pd.DataFrame:
|
84 |
+
"""
|
85 |
+
Create a comprehensive DataFrame comparing reference and prediction texts with multiple metrics.
|
86 |
+
|
87 |
+
For each sample, calculates:
|
88 |
+
- Word Error Rate (WER) for original and normalized texts
|
89 |
+
- Cosine similarity for original and normalized texts
|
90 |
+
- Length statistics and differences
|
91 |
+
|
92 |
+
Args:
|
93 |
+
references: List of original reference texts
|
94 |
+
predictions: List of original prediction texts
|
95 |
+
normalized_references: Optional list of normalized reference texts
|
96 |
+
normalized_predictions: Optional list of normalized prediction texts
|
97 |
+
output_csv: Path to save results (None to skip saving)
|
98 |
+
|
99 |
+
Returns:
|
100 |
+
pd.DataFrame: Analysis results with one row per sample
|
101 |
+
|
102 |
+
Example:
|
103 |
+
>>> df = create_wer_analysis_dataframe(
|
104 |
+
... references=["hello world"],
|
105 |
+
... predictions=["hello there"],
|
106 |
+
... output_csv="analysis.csv"
|
107 |
+
... )
|
108 |
+
"""
|
109 |
+
from jiwer import wer # Import here to avoid dependency if not using WER
|
110 |
+
|
111 |
+
records = []
|
112 |
+
|
113 |
+
for i, (ref, pred) in enumerate(zip(references, predictions)):
|
114 |
+
# Skip empty samples
|
115 |
+
if not ref.strip() or not pred.strip():
|
116 |
+
continue
|
117 |
+
|
118 |
+
# Get normalized versions if provided
|
119 |
+
norm_ref = normalized_references[i] if normalized_references else ref
|
120 |
+
norm_pred = normalized_predictions[i] if normalized_predictions else pred
|
121 |
+
|
122 |
+
# Calculate metrics with error handling
|
123 |
+
metrics = {
|
124 |
+
'index': i,
|
125 |
+
'reference': ref,
|
126 |
+
'prediction': pred,
|
127 |
+
'normalized_reference': norm_ref,
|
128 |
+
'normalized_prediction': norm_pred,
|
129 |
+
'ref_length': len(ref.split()),
|
130 |
+
'pred_length': len(pred.split()),
|
131 |
+
'length_difference': len(pred.split()) - len(ref.split())
|
132 |
+
}
|
133 |
+
|
134 |
+
# Calculate WER metrics
|
135 |
+
try:
|
136 |
+
metrics['wer'] = wer(ref, pred) * 100
|
137 |
+
metrics['normalized_wer'] = wer(norm_ref, norm_pred) * 100
|
138 |
+
except Exception as e:
|
139 |
+
print(f"WER calculation failed for sample {i}: {e}")
|
140 |
+
metrics.update({'wer': np.nan, 'normalized_wer': np.nan})
|
141 |
+
|
142 |
+
# Calculate cosine similarities
|
143 |
+
for prefix, text1, text2 in [
|
144 |
+
('', ref, pred),
|
145 |
+
('normalized_', norm_ref, norm_pred)
|
146 |
+
]:
|
147 |
+
try:
|
148 |
+
vectorizer = CountVectorizer(
|
149 |
+
analyzer='char_wb',
|
150 |
+
ngram_range=(2, 3)
|
151 |
+
vectors = vectorizer.fit_transform([text1, text2])
|
152 |
+
similarity = cosine_similarity(vectors[0:1], vectors[1:2])[0][0] * 100
|
153 |
+
metrics[f'{prefix}similarity'] = similarity
|
154 |
+
except Exception as e:
|
155 |
+
print(f"Similarity calculation failed for sample {i}: {e}")
|
156 |
+
metrics[f'{prefix}similarity'] = np.nan
|
157 |
+
|
158 |
+
records.append(metrics)
|
159 |
+
|
160 |
+
# Create DataFrame
|
161 |
+
df = pd.DataFrame(records)
|
162 |
+
|
163 |
+
# Save to CSV if requested
|
164 |
+
if output_csv:
|
165 |
+
try:
|
166 |
+
df.to_csv(output_csv, index=False)
|
167 |
+
print(f"Analysis saved to {output_csv}")
|
168 |
+
except Exception as e:
|
169 |
+
print(f"Failed to save CSV: {e}")
|
170 |
+
|
171 |
+
return df
|