def test_f_score_2():
sentences = (("s1", "s3", "s6"), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=0.5)
p = 2/3
r = 2/5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (1.25 * p * r) / (0.25 * p + r)
assert result == approx(expected)
def test_f_score_2(self):
sentences = (("s1", "s3", "s6"), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=0.5)
p = 2/3
r = 2/5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (1.25 * p * r) / (0.25 * p + r)
self.assertAlmostEqual(result, expected)
def test_f_score_1(self):
sentences = (("s1",), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=2.0)
p = 1/1
r = 1/5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (5 * p * r) / (4 * p + r)
self.assertAlmostEqual(result, expected)
def test_f_score_1(self):
sentences = (("s1", ), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=2.0)
p = 1 / 1
r = 1 / 5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (5 * p * r) / (4 * p + r)
self.assertAlmostEqual(result, expected)
def test_f_score_2(self):
sentences = (("s1", "s3", "s6"), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=0.5)
p = 2 / 3
r = 2 / 5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (1.25 * p * r) / (0.25 * p + r)
self.assertAlmostEqual(result, expected)
def test_f_score_1():
sentences = (("s1", ), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=2.0)
p = 1 / 1
r = 1 / 5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (5 * p * r) / (4 * p + r)
assert result == approx(expected)
def test_f_score_1():
sentences = (("s1",), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=2.0)
p = 1/1
r = 1/5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (5 * p * r) / (4 * p + r)
assert result == approx(expected)
def test_f_score_2():
sentences = (("s1", "s3", "s6"), ("s1", "s2", "s3", "s4", "s5"))
result = f_score(*sentences, weight=0.5)
p = 2 / 3
r = 2 / 5
# ( (W^2 + 1) * P * R ) / ( W^2 * P + R )
expected = (1.25 * p * r) / (0.25 * p + r)
assert result == approx(expected)
def evaluate(summary, sumref, debug=False):
sumstring = sumtostr(summary)
sumtuple = sumtotup(summary)
refstring = sumtostr(sumref)
reftuple = sumtotup(sumref)
summodel = TfDocumentModel(sumstring, Tokenizer("english"))
refmodel = TfDocumentModel(refstring, Tokenizer("english"))
if debug:
print(reftuple)
print(sumtuple)
cos_val = cosine_similarity(summodel, refmodel)
unit_val = unit_overlap(summodel, refmodel)
precision_val = precision(sumtuple, reftuple)
recall_val = recall(sumtuple, reftuple)
f_val = f_score(sumtuple, reftuple)
return cos_val, unit_val, precision_val, recall_val, f_val
def test_basic_f_score_empty_reference():
with pytest.raises(ValueError):
f_score(("s1", "s2", "s3", "s4", "s5"), ())
def test_basic_f_score_equals():
sentences = ("s1", "s2", "s3", "s4", "s5")
result = f_score(sentences, sentences)
assert result == approx(1.0)
def test_basic_f_score_no_match(self):
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s6", "s7", "s8"))
self.assertEqual(result, 0.0)
def test_basic_f_score_equals(self):
sentences = ("s1", "s2", "s3", "s4", "s5")
result = f_score(sentences, sentences)
self.assertAlmostEqual(result, 1.0)
def test_basic_f_score_evaluated_smaller():
result = f_score(("s1", ), ("s1", "s2", "s3", "s4", "s5"))
assert result == approx(1 / 3)
def test_basic_f_score_reference_smaller(self):
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s1", ))
self.assertAlmostEqual(result, 1 / 3)
def test_basic_f_score_evaluated_smaller(self):
result = f_score(("s1", ), ("s1", "s2", "s3", "s4", "s5"))
self.assertAlmostEqual(result, 1 / 3)
def test_basic_f_score_reference_smaller(self):
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s1",))
self.assertAlmostEqual(result, 1/3)
def test_basic_f_score_no_match(self):
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s6", "s7", "s8"))
self.assertEqual(result, 0.0)
def test_basic_f_score_equals(self):
sentences = ("s1", "s2", "s3", "s4", "s5")
result = f_score(sentences, sentences)
self.assertAlmostEqual(result, 1.0)
def test_basic_f_score_evaluated_smaller(self):
result = f_score(("s1",), ("s1", "s2", "s3", "s4", "s5"))
self.assertAlmostEqual(result, 1/3)
def test_basic_f_score_no_match():
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s6", "s7", "s8"))
assert result == 0.0
def test_basic_f_score_empty_evaluated():
with pytest.raises(ValueError):
f_score((), ("s1", "s2", "s3", "s4", "s5"))
def test_basic_f_score_empty_evaluated():
with pytest.raises(ValueError):
f_score((), ("s1", "s2", "s3", "s4", "s5"))
def test_basic_f_score_evaluated_smaller():
result = f_score(("s1",), ("s1", "s2", "s3", "s4", "s5"))
assert result == approx(1/3)
def test_basic_f_score_equals():
sentences = ("s1", "s2", "s3", "s4", "s5")
result = f_score(sentences, sentences)
assert result == approx(1.0)
def test_basic_f_score_empty_reference():
with pytest.raises(ValueError):
f_score(("s1", "s2", "s3", "s4", "s5"), ())
def test_basic_f_score_no_match():
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s6", "s7", "s8"))
assert result == 0.0
def test_basic_f_score_reference_smaller():
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s1", ))
assert result == approx(1 / 3)
def test_basic_f_score_reference_smaller():
result = f_score(("s1", "s2", "s3", "s4", "s5"), ("s1",))
assert result == approx(1/3)