-
Notifications
You must be signed in to change notification settings - Fork 1
/
metrics.py
executable file
·200 lines (149 loc) · 6.45 KB
/
metrics.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
import numpy as np
from collections import Counter
from scipy.optimize import linear_sum_assignment
from consts import CATEGORIES
def f1(p_num, p_den, r_num, r_den, beta=1):
p = 0 if p_den == 0 else p_num / float(p_den)
r = 0 if r_den == 0 else r_num / float(r_den)
return 0 if p + r == 0 else (1 + beta * beta) * p * r / (beta * beta * p + r)
class CorefCategories:
def __init__(self):
self.categories = {cat_name: {'tp': 0, 'fp': 0, 'fn': 0, 'tn': 0} for cat_name in CATEGORIES}
def update(self, coref_logits, categories_labels, clusters_labels):
for cat_name, cat_id in CATEGORIES.items():
cat_mask = categories_labels == cat_id
labels = clusters_labels[cat_mask]
logits = coref_logits[:, :, :-1][cat_mask] # without null cluster logit
self.categories[cat_name]['tn'] += np.logical_and(labels == 0., logits < 0).sum()
self.categories[cat_name]['fn'] += np.logical_and(labels == 1., logits < 0).sum()
self.categories[cat_name]['fp'] += np.logical_and(labels == 0., logits > 0).sum()
self.categories[cat_name]['tp'] += np.logical_and(labels == 1., logits > 0).sum()
def get_stats(self):
stats = {}
for cat_name in self.categories:
tp, fp = self.categories[cat_name]['tp'], self.categories[cat_name]['fp']
fn, tn = self.categories[cat_name]['fn'], self.categories[cat_name]['tn']
stats[cat_name] = {'true_pairs': tp + fn, 'false_pairs': tn + fp, 'precision': 0, 'recall': 0, 'f1': 0}
if tp + fp:
stats[cat_name]['precision'] = round((tp / (tp + fp)) * 100, 1)
if tp + fn:
stats[cat_name]['recall'] = round((tp / (tp + fn)) * 100, 1)
if tp + 0.5 * (fp + fn):
stats[cat_name]['f1'] = round((tp / (tp + 0.5 * (fp + fn))) * 100, 1)
return stats
class MentionEvaluator:
def __init__(self):
self.tp, self.fp, self.fn = 0, 0, 0
def update(self, predicted_mentions, gold_mentions):
predicted_mentions = set(predicted_mentions)
gold_mentions = set(gold_mentions)
self.tp += len(predicted_mentions & gold_mentions)
self.fp += len(predicted_mentions - gold_mentions)
self.fn += len(gold_mentions - predicted_mentions)
def get_f1(self):
pr = self.get_precision()
rec = self.get_recall()
return 2 * pr * rec / (pr + rec) if pr + rec > 0 else 0.0
def get_recall(self):
return self.tp / (self.tp + self.fn) if (self.tp + self.fn) > 0 else 0.0
def get_precision(self):
return self.tp / (self.tp + self.fp) if (self.tp + self.fp) > 0 else 0.0
def get_prf(self):
return self.get_precision(), self.get_recall(), self.get_f1()
class CorefEvaluator(object):
def __init__(self):
self.evaluators = [Evaluator(m) for m in (muc, b_cubed, ceafe)]
def update(self, predicted, gold, mention_to_predicted, mention_to_gold):
for e in self.evaluators:
e.update(predicted, gold, mention_to_predicted, mention_to_gold)
def get_f1(self):
return sum(e.get_f1() for e in self.evaluators) / len(self.evaluators)
def get_recall(self):
return sum(e.get_recall() for e in self.evaluators) / len(self.evaluators)
def get_precision(self):
return sum(e.get_precision() for e in self.evaluators) / len(self.evaluators)
def get_prf(self):
return self.get_precision(), self.get_recall(), self.get_f1()
class Evaluator(object):
def __init__(self, metric, beta=1):
self.p_num = 0
self.p_den = 0
self.r_num = 0
self.r_den = 0
self.metric = metric
self.beta = beta
def update(self, predicted, gold, mention_to_predicted, mention_to_gold):
if self.metric == ceafe:
pn, pd, rn, rd = self.metric(predicted, gold)
else:
pn, pd = self.metric(predicted, mention_to_gold)
rn, rd = self.metric(gold, mention_to_predicted)
self.p_num += pn
self.p_den += pd
self.r_num += rn
self.r_den += rd
def get_f1(self):
return f1(self.p_num, self.p_den, self.r_num, self.r_den, beta=self.beta)
def get_recall(self):
return 0 if self.r_num == 0 else self.r_num / float(self.r_den)
def get_precision(self):
return 0 if self.p_num == 0 else self.p_num / float(self.p_den)
def get_prf(self):
return self.get_precision(), self.get_recall(), self.get_f1()
def get_counts(self):
return self.p_num, self.p_den, self.r_num, self.r_den
def b_cubed(clusters, mention_to_gold):
num, dem = 0, 0
for c in clusters:
if len(c) == 1:
continue
gold_counts = Counter()
correct = 0
for m in c:
if m in mention_to_gold:
gold_counts[tuple(mention_to_gold[m])] += 1
for c2, count in gold_counts.items():
if len(c2) != 1:
correct += count * count
num += correct / float(len(c))
dem += len(c)
return num, dem
def muc(clusters, mention_to_gold):
tp, p = 0, 0
for c in clusters:
p += len(c) - 1
tp += len(c)
linked = set()
for m in c:
if m in mention_to_gold:
linked.add(mention_to_gold[m])
else:
tp -= 1
tp -= len(linked)
return tp, p
def phi4(c1, c2):
return 2 * len([m for m in c1 if m in c2]) / float(len(c1) + len(c2))
def ceafe(clusters, gold_clusters):
clusters = [c for c in clusters if len(c) != 1]
scores = np.zeros((len(gold_clusters), len(clusters)))
for i in range(len(gold_clusters)):
for j in range(len(clusters)):
scores[i, j] = phi4(gold_clusters[i], clusters[j])
row_ind, col_ind = linear_sum_assignment(-scores)
similarity = sum(scores[row_ind, col_ind])
return similarity, len(clusters), similarity, len(gold_clusters)
def lea(clusters, mention_to_gold):
num, dem = 0, 0
for c in clusters:
if len(c) == 1:
continue
common_links = 0
all_links = len(c) * (len(c) - 1) / 2.0
for i, m in enumerate(c):
if m in mention_to_gold:
for m2 in c[i + 1:]:
if m2 in mention_to_gold and mention_to_gold[m] == mention_to_gold[m2]:
common_links += 1
num += len(c) * common_links / float(all_links)
dem += len(c)
return num, dem