FabianWillner commited on
Commit
4ee625a
·
1 Parent(s): 98e4725

Upload triviaQARC.py

Browse files
Files changed (1) hide show
  1. triviaQARC.py +109 -0
triviaQARC.py ADDED
@@ -0,0 +1,109 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from collections import Counter
2
+ import datasets
3
+
4
+ class triviaQARC(datasets.Metric):
5
+ def _info(self):
6
+ return datasets.MetricInfo(
7
+ description= "Idk",
8
+ citation= "idk",
9
+ features=datasets.Features(
10
+ {
11
+ "predictions": {"id": datasets.Value("string"), "prediction_text": datasets.Value("string")},
12
+ "references": {
13
+ "id": datasets.Value("string"),
14
+ "answers": datasets.features.Sequence(
15
+ {
16
+ "text": datasets.Value("string"),
17
+ "answer_start": datasets.Value("int32"),
18
+ }
19
+ ),
20
+ },
21
+ }
22
+ ),
23
+ )
24
+
25
+ def _compute(self, predictions, references):
26
+ pred_dict = {prediction["id"]: prediction["prediction_text"] for prediction in predictions}
27
+ dataset = [
28
+ {
29
+ "paragraphs": [
30
+ {
31
+ "qas": [
32
+ {
33
+ "answers": [{"text": answer_text} for answer_text in ref["answers"]["text"]],
34
+ "id": ref["id"],
35
+ }
36
+ for ref in references
37
+ ]
38
+ }
39
+ ]
40
+ }
41
+ ]
42
+ score = evaluate(dataset=dataset, predictions=pred_dict)
43
+ return score
44
+
45
+ def evaluate(dataset, predictions):
46
+ f1 = exact_match = total = recall = precision= 0
47
+ for article in dataset:
48
+ for paragraph in article["paragraphs"]:
49
+ for qa in paragraph["qas"]:
50
+ total += 1
51
+ if qa["id"] not in predictions:
52
+ message = "Unanswered question " + qa["id"] + " will receive score 0."
53
+ print(message, file=sys.stderr)
54
+ continue
55
+ ground_truths = list(map(lambda x: x["text"], qa["answers"]))
56
+ prediction = predictions[qa["id"]]
57
+ exact_match += metric_max_over_ground_truths(exact_match_score, prediction, ground_truths)
58
+ temp_f1, temp_precision, temp_recall = metric_max_over_ground_truths(f1_score, prediction, ground_truths)
59
+ f1 += temp_f1
60
+ precision += temp_precision
61
+ recall += temp_recall
62
+
63
+ exact_match = 100.0 * exact_match / total
64
+ f1 = 100.0 * f1 / total
65
+
66
+ return {"exact_match": exact_match, "f1": f1, "recall": recall, "precision": precision}
67
+
68
+ def normalize_answer(s):
69
+ """Lower text and remove punctuation, articles and extra whitespace."""
70
+
71
+ def remove_articles(text):
72
+ return re.sub(r"\b(a|an|the)\b", " ", text)
73
+
74
+ def white_space_fix(text):
75
+ return " ".join(text.split())
76
+
77
+ def remove_punc(text):
78
+ exclude = set(string.punctuation)
79
+ return "".join(ch for ch in text if ch not in exclude)
80
+
81
+ def lower(text):
82
+ return text.lower()
83
+
84
+ return white_space_fix(remove_articles(remove_punc(lower(s))))
85
+
86
+
87
+ def f1_score(prediction, ground_truth):
88
+ prediction_tokens = normalize_answer(prediction).split()
89
+ ground_truth_tokens = normalize_answer(ground_truth).split()
90
+ common = Counter(prediction_tokens) & Counter(ground_truth_tokens)
91
+ num_same = sum(common.values())
92
+ if num_same == 0:
93
+ return 0
94
+ precision = 1.0 * num_same / len(prediction_tokens)
95
+ recall = 1.0 * num_same / len(ground_truth_tokens)
96
+ f1 = (2 * precision * recall) / (precision + recall)
97
+ return f1, precision, recall
98
+
99
+
100
+ def exact_match_score(prediction, ground_truth):
101
+ return normalize_answer(prediction) == normalize_answer(ground_truth)
102
+
103
+
104
+ def metric_max_over_ground_truths(metric_fn, prediction, ground_truths):
105
+ scores_for_ground_truths = []
106
+ for ground_truth in ground_truths:
107
+ score = metric_fn(prediction, ground_truth)
108
+ scores_for_ground_truths.append(score)
109
+ return max(scores_for_ground_truths)