def evaluate_cosine_similarity(evaluated_sentences, reference_sentences):
evaluated_words = tuple(chain(*(s.words for s in evaluated_sentences)))
reference_words = tuple(chain(*(s.words for s in reference_sentences)))
evaluated_model = TfDocumentModel(evaluated_words)
reference_model = TfDocumentModel(reference_words)
return cosine_similarity(evaluated_model, reference_model)
def analyze_summaries(book_id, args):
"""
Analyze the summary.
Compares the created summary with the ground truth summary.
Expects the ground truth summary to be in the data/summaries directory.
Parameters:
book_id: (str) the book identifier
args: the command line arguments, used to determine the filename for the created summary.
Outputs:
Saves the analysis to a csv file in the results/analysis directory
"""
if not exists('../results/analysis'):
makedirs('../results/analysis')
analysis_data = []
summary_doc, summary_model = load_summary(
get_data_filename(book_id, 'summaries'))
new_summary_doc, new_summary_model = load_summary(
get_results_filename(book_id, args))
if summary_doc != '':
if new_summary_doc != '':
analysis_data.append([
'word embeddings similarity',
summary_doc.similarity(new_summary_doc)
])
if summary_model != '':
if new_summary_model != '':
analysis_data.append([
'cosine similarity',
cosine_similarity(summary_model, new_summary_model)
])
with open(get_analysis_filename(book_id, args), 'w') as csvFile:
writer = csv.writer(csvFile)
writer.writerows(analysis_data)
def test_cosine_half_match(self):
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Veta aká sa tu len veľmi ťažko hľadá",
tokenizer)
model2 = TfDocumentModel("Teta ktorá sa tu iba veľmi zle hľadá",
tokenizer)
self.assertAlmostEqual(cosine_similarity(model1, model2), 0.5)
def test_cosine_no_match(self):
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Toto je moja veta. To sa nedá poprieť!",
tokenizer)
model2 = TfDocumentModel("Hento bolo jeho slovo, ale možno klame.",
tokenizer)
self.assertAlmostEqual(cosine_similarity(model1, model2), 0.0)
def test_cosine_half_match(self):
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Veta aká sa tu len veľmi ťažko hľadá",
tokenizer)
model2 = TfDocumentModel("Teta ktorá sa tu iba veľmi zle hľadá",
tokenizer)
self.assertAlmostEqual(cosine_similarity(model1, model2), 0.5)
def test_cosine_no_match(self):
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Toto je moja veta. To sa nedá poprieť!",
tokenizer)
model2 = TfDocumentModel("Hento bolo jeho slovo, ale možno klame.",
tokenizer)
self.assertAlmostEqual(cosine_similarity(model1, model2), 0.0)
def test_cosine_no_match():
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Toto je moja veta. To sa nedá poprieť!",
tokenizer)
model2 = TfDocumentModel("Hento bolo jeho slovo, ale možno klame.",
tokenizer)
assert cosine_similarity(model1, model2) == approx(0.0)
def test_wrong_arguments():
text = "Toto je moja veta, to sa nedá poprieť."
model = TfDocumentModel(text, Tokenizer("czech"))
with pytest.raises(ValueError):
cosine_similarity(text, text)
with pytest.raises(ValueError):
cosine_similarity(text, model)
with pytest.raises(ValueError):
cosine_similarity(model, text)
def test_empty_model():
text = "Toto je moja veta, to sa nedá poprieť."
model = TfDocumentModel(text, Tokenizer("czech"))
empty_model = TfDocumentModel([])
with pytest.raises(ValueError):
cosine_similarity(empty_model, empty_model)
with pytest.raises(ValueError):
cosine_similarity(empty_model, model)
with pytest.raises(ValueError):
cosine_similarity(model, empty_model)
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_cosine_exact_match(self):
text = "Toto je moja veta, to sa nedá poprieť."
model = TfDocumentModel(text, Tokenizer("czech"))
self.assertAlmostEqual(cosine_similarity(model, model), 1.0)
def test_cosine_half_match():
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Veta aká sa tu len veľmi ťažko hľadá", tokenizer)
model2 = TfDocumentModel("Teta ktorá sa tu iba veľmi zle hľadá", tokenizer)
assert cosine_similarity(model1, model2) == approx(0.5)
def test_cosine_no_match():
tokenizer = Tokenizer("czech")
model1 = TfDocumentModel("Toto je moja veta. To sa nedá poprieť!", tokenizer)
model2 = TfDocumentModel("Hento bolo jeho slovo, ale možno klame.", tokenizer)
assert cosine_similarity(model1, model2) == approx(0.0)
def test_cosine_exact_match():
text = "Toto je moja veta, to sa nedá poprieť."
model = TfDocumentModel(text, Tokenizer("czech"))
assert cosine_similarity(model, model) == approx(1.0)
def test_cosine_exact_match(self):
text = "Toto je moja veta, to sa nedá poprieť."
model = TfDocumentModel(text, Tokenizer("czech"))
self.assertAlmostEqual(cosine_similarity(model, model), 1.0)
