def testMatchingSentences(self):
"""If input=output=reference, the score should be 1."""
input_sentence = [3, 1, 4, 1, 5, 9, 2, 6, 5]
output = input_sentence
references = [input_sentence]
score, _, _, _ = sari_hook.get_sari_score(input_sentence, output,
references)
self.assertEqual(1, score)
def testMatchingOutputAndReference(self):
"""If output=reference, the score should be 1."""
input_sentence = [3, 1, 4, 1, 5, 9, 2, 6, 5]
output = [3, 1, 4, 1, 80, 70]
references = [output]
score, _, _, _ = sari_hook.get_sari_score(input_sentence, output,
references)
self.assertEqual(1, score)
def testIdsWithZeros(self):
"""Zeros should be ignored."""
input_sentence = [3, 1, 4, 0, 0, 0]
output = [3, 1, 4]
references = [[3, 1, 4, 0, 0, 0, 0, 0]]
score, _, _, _ = sari_hook.get_sari_score(input_sentence, output,
references)
self.assertEqual(1, score)
def testMatchingSentencesWithRepetitions(self):
"""Token frequencies should not matter if we only consider unigrams."""
input_sentence = [3, 1, 4]
output = [3, 3, 1, 1, 1, 4]
references = [[3, 3, 3, 1, 1, 4, 4]]
score, _, _, _ = sari_hook.get_sari_score(input_sentence, output,
references, max_gram_size=1)
self.assertEqual(1, score)
def testSariSent1(self):
"""Test case 1 from SARI-paper.
The score is slightly different from what is reported in the paper (0.2683)
since the authors' code seems to contain a bug in the keep recall score
computation.
"""
output = "About 95 you now get in .".split()
score, _, _, _ = sari_hook.get_sari_score(self.input_sentence, output,
self.references)
self.assertAlmostEqual(0.2695360, score)
def testSariSent3(self):
"""Test case 3 from SARI-paper."""
output = "About 95 species are currently agreed .".split()
score, _, _, _ = sari_hook.get_sari_score(self.input_sentence, output,
self.references)
self.assertAlmostEqual(0.5088682, score)
def testSariSent2(self):
"""Test case 2 from SARI-paper."""
output = "About 95 species are now agreed .".split()
score, _, _, _ = sari_hook.get_sari_score(self.input_sentence, output,
self.references)
self.assertAlmostEqual(0.6170966, score)
