def test_evaluate_evidence4(self):
N = 4
expect = {
'TPs': [6, 5, 4, 3], 'FPs': [0, 0, 0, 0], 'FNs': [0, 0, 0, 0],
'precision': 1, 'recall': 1, 'F-measure': 1,
'BLEU': 1, 'ROUGE': 1,
'exact_match': 1, 'partial_match': 1
}
self.assertEqual(eval.evaluate_evidence(self.test_chunk, 3, N), expect)
def test_evaluate_evidence5(self):
# pred: [1, 1, 1, 0, 0, 1]
# and: [0, 1, 1, 1, 1, 1]
N = 4
expect = {
'TPs': [3, 1, 0, 0], 'FPs': [1, 1, 1, 0], 'FNs': [2, 3, 3, 2],
'precision': 3/4, 'recall': 3/5, 'F-measure': 2/3,
'BLEU': (1/8)**(1/4), 'ROUGE': (1/50)**(1/4),
'exact_match': 0, 'partial_match': 1
}
result = eval.evaluate_evidence(self.test_chunk, 4, N)
float_keys = ['precision', 'recall', 'F-measure', 'BLEU', 'ROUGE']
for key in float_keys:
self.assertAlmostEqual(result[key], expect[key])
expect.pop(key)
result.pop(key)
self.assertEqual(result, expect)
def test_evaluate_evidence7(self):
N = 4
self.assertIsNone(eval.evaluate_evidence(self.test_chunk, 6, N))