def test_transformed_feature():
# Phonological density is log-transformed.
drop_caches()
transformed_phonological_density = SubstitutionFeaturesMixin._transformed_feature("phonological_density")
assert transformed_phonological_density("time") == np.log(29)
assert np.isnan(transformed_phonological_density("wickiup"))
# Doc and name are transformed too.
assert (
transformed_phonological_density.__doc__
== "log(" + SubstitutionFeaturesMixin._phonological_density.__doc__ + ")"
)
assert transformed_phonological_density.__name__ == "_log_phonological_density"
# And the list of words is properly computed.
drop_caches()
with settings.file_override("CLEARPOND"):
with open(settings.CLEARPOND, "w") as f:
f.write("dog" + 5 * "\t" + "2" + 24 * "\t" + "3\n" "cat" + 5 * "\t" + "2" + 24 * "\t" + "3")
assert set(transformed_phonological_density()) == {"dog", "cat"}
# AoA is left untouched.
drop_caches()
transformed_aoa = SubstitutionFeaturesMixin._transformed_feature("aoa")
assert transformed_aoa("time") == 5.16
assert transformed_aoa("vocative") == 14.27
assert np.isnan(transformed_aoa("wickiup"))
# Doc and name are passed on.
assert transformed_aoa.__doc__ == SubstitutionFeaturesMixin._aoa.__doc__
assert transformed_aoa.__name__ == "_aoa"
# And the list of words is properly computed.
drop_caches()
with settings.file_override("AOA"):
with open(settings.AOA, "w") as f:
f.write("Word,Rating.Mean\nhave,2\ntell,3")
assert set(transformed_aoa()) == {"have", "tell"}
def test_orthographic_density_none():
drop_caches()
# Lemmas are all lowercase.
for word in SubstitutionFeaturesMixin._orthographic_density():
assert word.islower()
# And it's computed right.
drop_caches()
with settings.file_override("CLEARPOND"):
with open(settings.CLEARPOND, "w") as f:
f.write("dog" + 5 * "\t" + "2" + 24 * "\t" + "3\n" "cat" + 5 * "\t" + "2" + 24 * "\t" + "3")
assert set(SubstitutionFeaturesMixin._orthographic_density()) == {"dog", "cat"}
def test_aoa_none():
drop_caches()
# Lemmas are all lowercase.
for word in SubstitutionFeaturesMixin._aoa():
assert word.islower()
# And it's properly computed.
drop_caches()
with settings.file_override("AOA"):
with open(settings.AOA, "w") as f:
f.write("Word,Rating.Mean\nhave,2\ntell,3")
assert set(SubstitutionFeaturesMixin._aoa()) == {"have", "tell"}
def test_component():
drop_caches()
# Create a test PCA with features alternatively log-transformed and not,
# alternatively on tokens and lemmas.
features = ("letters_count", "aoa", "synonyms_count", "phonological_density")
pca = PCA(n_components=3)
# Trying this with a PCA fitted with the wrong shape fails.
pca.fit(np.array([[1, 1, 0], [0, 1, 0], [0, 1, 1]]))
with pytest.raises(AssertionError):
SubstitutionFeaturesMixin._component(0, pca, features)
with pytest.raises(AssertionError):
SubstitutionFeaturesMixin._component(1, pca, features)
# Trying this with unknown features fails.
with pytest.raises(ValueError) as excinfo:
SubstitutionFeaturesMixin._component(0, pca, ("letters_count", "unknown_feature", "aoa"))
assert "Unknown feature" in str(excinfo.value)
with pytest.raises(ValueError) as excinfo:
SubstitutionFeaturesMixin._component(1, pca, ("letters_count", "unknown_feature", "aoa"))
assert "Unknown feature" in str(excinfo.value)
# Now training with the right shape.
pca.fit(np.array([[1, 0, 0, 0], [-1, 0, 0, 0], [0, 1, 0, 0], [0, -1, 0, 0], [0, 0, 1, 0], [0, 0, -1, 0]]))
with settings.file_override("TOKENS"):
with open(settings.TOKENS, "wb") as f:
pickle.dump({"these", "are", "tokens"}, f)
c0 = SubstitutionFeaturesMixin._component(0, pca, features)
c1 = SubstitutionFeaturesMixin._component(1, pca, features)
c2 = SubstitutionFeaturesMixin._component(2, pca, features)
# Doc and name are properly set.
assert c0.__name__ == "_component_0"
assert c0.__doc__ == "component 0"
assert c1.__name__ == "_component_1"
assert c1.__doc__ == "component 1"
assert c2.__name__ == "_component_2"
assert c2.__doc__ == "component 2"
# We get the expected hand-computed values.
assert c0("time") == -5.16
assert c1("time") == 0.62860865942237421
assert c2("time") == -4
assert np.isnan(c0("makakiki"))
assert np.isnan(c1("makakiki"))
assert np.isnan(c2("makakiki"))
# And the list of words is properly computed. (These are not the true
# values since we overrode the tokens list.)
assert len(c0()) == 157863
assert len(c1()) == 157863
assert len(c2()) == 157863
def test_component_average():
drop_caches()
# Two subtitutions.
q1a = Quote(string="Chase it others is the dogs hound")
q1b = Quote(string="Others is the hound hound")
s1 = Substitution(source=q1a, destination=q1b, start=2, position=3)
q2a = Quote(string="Chase it others is the frisbee hound")
q2b = q1b
s2 = Substitution(source=q2a, destination=q2b, start=2, position=3)
# Create a test PCA that will use features we later override.
features = ("aoa", "phonological_density")
pca = PCA(n_components=2)
# Trying this with a PCA fitted with the wrong shape fails.
pca.fit(np.array([[1, 1, 0], [0, 1, 0], [0, 1, 1]]))
with pytest.raises(AssertionError):
s1.component_average(0, pca, features)
with pytest.raises(AssertionError):
s1.component_average(0, pca, features, source_synonyms=True)
with pytest.raises(AssertionError):
s1.component_average(0, pca, features, source_synonyms=False, sentence_relative="mean")
with pytest.raises(AssertionError):
s1.component_average(0, pca, features, source_synonyms=True, sentence_relative="mean")
# Trying this with unknown features fails.
with pytest.raises(ValueError) as excinfo:
s1.component_average(0, pca, ("letters_count", "unknown_feature", "aoa"))
assert "Unknown feature" in str(excinfo.value)
with pytest.raises(ValueError) as excinfo:
s1.component_average(0, pca, ("letters_count", "unknown_feature", "aoa"), source_synonyms=True)
assert "Unknown feature" in str(excinfo.value)
with pytest.raises(ValueError) as excinfo:
s1.component_average(
0, pca, ("letters_count", "unknown_feature", "aoa"), source_synonyms=False, sentence_relative="mean"
)
assert "Unknown feature" in str(excinfo.value)
with pytest.raises(ValueError) as excinfo:
s1.component_average(
0, pca, ("letters_count", "unknown_feature", "aoa"), source_synonyms=True, sentence_relative="mean"
)
assert "Unknown feature" in str(excinfo.value)
# Now with features we override to test manual values.
drop_caches()
pca.fit(np.array([[2, 1], [1, -2]]))
sign = np.sign(pca.components_[:, 0])
with settings.file_override("AOA", "CLEARPOND"):
with open(settings.AOA, "w") as f:
f.write("Word,Rating.Mean\n" "dog,2\nhound,3\nfrisbee,4\nchase,6\ncad,7\nother,8")
with open(settings.CLEARPOND, "w") as f:
f.write(
"dog" + 5 * "\t" + "0" + 24 * "\t" + "2\n"
"hound" + 5 * "\t" + "0" + 24 * "\t" + "3\n"
"frisbee" + 5 * "\t" + "0" + 24 * "\t" + "4\n"
"screen" + 5 * "\t" + "0" + 24 * "\t" + "5\n"
"chase" + 5 * "\t" + "0" + 24 * "\t" + "6\n"
"other" + 5 * "\t" + "0" + 24 * "\t" + "8\n"
"others" + 5 * "\t" + "0" + 24 * "\t" + "9"
)
# We find the hand-computed values alright.
assert abs(-sign[0] * s1.component_average(0, pca, features) - (-2.7921497899976822)) < 1e-14
assert abs(-sign[0] * s2.component_average(0, pca, features) - (-2.7921497899976822)) < 1e-14
assert abs(-sign[1] * s1.component_average(1, pca, features) - (-2.3369703188414315)) < 1e-14
assert abs(-sign[1] * s2.component_average(1, pca, features) - (-2.3369703188414315)) < 1e-14
# Same with synonyms. Computed on synonyms of 'dog' (lemma of
# 'dogs'). 'frisbee' has no synonyms, hence the NaN for s2.
assert (
abs(-sign[0] * s1.component_average(0, pca, features, source_synonyms=True) - (-2.7940486530122683)) < 1e-14
)
assert np.isnan(s2.component_average(0, pca, features, source_synonyms=True))
assert (
abs(-sign[1] * s1.component_average(1, pca, features, source_synonyms=True) - (-2.2309281091642896)) < 1e-14
)
assert np.isnan(s2.component_average(1, pca, features, source_synonyms=True))
# Same without synonyms but with sentence_relative.
# Each feature uses either lemmas or tokens (whereas above it was
# all lemmas).
assert (
abs(
-sign[0] * s1.component_average(0, pca, features, source_synonyms=False, sentence_relative="mean")
- 0.34030374468910285
)
< 1e-14
)
assert (
abs(
-sign[0] * s2.component_average(0, pca, features, source_synonyms=False, sentence_relative="mean")
- 0.34030374468910285
)
< 1e-14
)
assert (
abs(
-sign[1] * s1.component_average(1, pca, features, source_synonyms=False, sentence_relative="mean")
- 0.51902095047064112
)
< 1e-14
)
assert (
abs(
-sign[1] * s2.component_average(1, pca, features, source_synonyms=False, sentence_relative="mean")
- 0.51902095047064112
)
< 1e-14
)
# Same with synonyms and sentence_relative.
assert (
abs(
-sign[0] * s1.component_average(0, pca, features, source_synonyms=True, sentence_relative="mean")
- 0.3390378360127122
)
< 1e-14
)
assert np.isnan(s2.component_average(0, pca, features, source_synonyms=True, sentence_relative="median"))
assert (
abs(
-sign[1] * s1.component_average(1, pca, features, source_synonyms=True, sentence_relative="mean")
- 0.58971575692206901
)
< 1e-14
)
assert np.isnan(s2.component_average(1, pca, features, source_synonyms=True, sentence_relative="mean"))
def test_feature_average():
# Two subtitutions.
q1a = Quote(string="Chase it others is the dogs hound")
q1b = Quote(string="Others is the hound hound")
s1 = Substitution(source=q1a, destination=q1b, start=2, position=3)
q2a = Quote(string="Chase it others is the frisbee hound")
q2b = q1b
s2 = Substitution(source=q2a, destination=q2b, start=2, position=3)
# Test a non-transformed feature (AoA), computed on lemmas.
drop_caches()
with settings.file_override("AOA"):
with open(settings.AOA, "w") as f:
f.write("Word,Rating.Mean\n" "dog,2\nhound,3\nfrisbee,4\nchase,6\ncad,7\nother,8")
assert s1.feature_average("aoa") == 30 / 6
assert s2.feature_average("aoa") == 30 / 6
assert s1.feature_average("aoa", source_synonyms=True) == np.mean([3, 6, 7])
# 'frisbee' has no synonyms.
assert np.isnan(s2.feature_average("aoa", source_synonyms=True))
assert s1.feature_average("aoa", source_synonyms=False, sentence_relative="mean") == (-0.33333333333333304)
assert s2.feature_average("aoa", source_synonyms=False, sentence_relative="mean") == (-0.33333333333333304)
assert s1.feature_average("aoa", source_synonyms=True, sentence_relative="mean") == (-0.11111111111111072)
# 'frisbee' has no synonyms.
assert np.isnan(s2.feature_average("aoa", source_synonyms=True, sentence_relative="mean"))
# Test a log-transformed feature (phonological density), computed on
# tokens.
drop_caches()
with settings.file_override("CLEARPOND"):
with open(settings.CLEARPOND, "w") as f:
f.write(
"dog" + 5 * "\t" + "0" + 24 * "\t" + "2\n"
"hound" + 5 * "\t" + "0" + 24 * "\t" + "3\n"
"frisbee" + 5 * "\t" + "0" + 24 * "\t" + "4\n"
"chase" + 5 * "\t" + "0" + 24 * "\t" + "6\n"
"cad" + 5 * "\t" + "0" + 24 * "\t" + "7\n"
"other" + 5 * "\t" + "0" + 24 * "\t" + "8"
)
assert s1.feature_average("phonological_density") == np.log([2, 3, 4, 6, 7, 8]).mean()
assert s2.feature_average("phonological_density") == np.log([2, 3, 4, 6, 7, 8]).mean()
# Even though phonological density is computed on tokens, the synonyms
# come from the lemmas.
assert s1.feature_average("phonological_density", source_synonyms=True) == np.log([3, 6, 7]).mean()
# 'frisbee' has no synonyms.
assert np.isnan(s2.feature_average("phonological_density", source_synonyms=True))
# Features for the 'sentence_relative' part are still taken from the
# tokens, which leads us to drop 'others'.
assert (
s1.feature_average("phonological_density", source_synonyms=False, sentence_relative="mean")
== 0.20029093819187427
)
assert (
s2.feature_average("phonological_density", source_synonyms=False, sentence_relative="mean")
== 0.20029093819187427
)
assert (
s1.feature_average("phonological_density", source_synonyms=True, sentence_relative="mean")
== 0.25674084015785814
)
# 'frisbee' has no synonyms.
assert np.isnan(s2.feature_average("phonological_density", source_synonyms=True, sentence_relative="median"))
# _synonyms_count(word=None) returns a list of words, some of which have
# a _synonyms_count(word) == np.nan (because 0 synonyms is returned as
# np.nan). So check that synonyms_count feature average is not np.nan.
assert np.isfinite(s1.feature_average("synonyms_count"))
def test_frequency_none():
drop_caches()
with settings.file_override("FREQUENCY"):
with open(settings.FREQUENCY, "wb") as f:
pickle.dump({"dog": 2, "cat": 3}, f)
assert set(SubstitutionFeaturesMixin._frequency()) == {"dog", "cat"}
def test_clustering_none():
drop_caches()
with settings.file_override("CLUSTERING"):
with open(settings.CLUSTERING, "wb") as f:
pickle.dump({"dog": 2, "cat": 3}, f)
assert set(SubstitutionFeaturesMixin._clustering()) == {"dog", "cat"}
def test_betweenness_none():
drop_caches()
with settings.file_override("BETWEENNESS"):
with open(settings.BETWEENNESS, "wb") as f:
pickle.dump({"dog": 2, "cat": 3}, f)
assert set(SubstitutionFeaturesMixin._betweenness()) == {"dog", "cat"}
def test_pagerank_none():
drop_caches()
with settings.file_override("PAGERANK"):
with open(settings.PAGERANK, "wb") as f:
pickle.dump({"dog": 2, "cat": 3}, f)
assert set(SubstitutionFeaturesMixin._pagerank()) == {"dog", "cat"}
def test_degree_none():
drop_caches()
with settings.file_override("DEGREE"):
with open(settings.DEGREE, "wb") as f:
pickle.dump({"dog": 2, "cat": 3}, f)
assert set(SubstitutionFeaturesMixin._degree()) == {"dog", "cat"}
def test_letters_count_none():
drop_caches()
with settings.file_override("TOKENS"):
with open(settings.TOKENS, "wb") as f:
pickle.dump({"these", "are", "tokens"}, f)
assert SubstitutionFeaturesMixin._letters_count() == {"these", "are", "tokens"}