def bnExtractDocSimilarity(doc1, doc2, similarity):
"""Measure the semantic similarity between two documents using
Word Movers Distance. Uses Textacy API
textacy.similarity.word_movers(doc1, doc2, metric=u'cosine')
"""
from textacy import similarity
#if similarity == 'Word Movers':
if similarity == 'cosine':
# Metric can be cosine, euclidian, I1, I2, or manhattan
s = similarity.word_movers(doc1, doc2, metric=u'cosine')
print(" Cosine Similarity between docs {} and {} is: {}".format( \
bnGetDocName(doc1), bnGetDocName(doc2), s))
elif similarity == 'Euclidian':
s = similarity.word_movers(doc1, doc2, metric=u'euclidian')
print(" Euclidian Similarity between docs {} and {} is: {}".format( \
bnGetDocName(doc1), bnGetDocName(doc2), s))
elif similarity == 'Manhattan':
s = similarity.word_movers(doc1, doc2, metric=u'manhattan')
print(" Manhattan Similarity between docs {} and {} is: {}".format( \
bnGetDocName(doc1), bnGetDocName(doc2), s))
elif similarity == 'word2vec':
s = similarity.word2vec(doc1, doc2)
print(" Semantic Similarity between docs {} and {} is: {}".format( \
bnGetDocName(doc1), bnGetDocName(doc2), s))
else:
# Unsupported similarity method
s = 0
return round(s, 5)
def test_identity(self, doc_pairs):
for doc1, doc2 in doc_pairs[:2]: # HACK
print(doc1, doc2)
assert similarity.word_movers(doc1,
doc1) == pytest.approx(1.0, rel=1e-3)
assert similarity.word_movers(doc2,
doc2) == pytest.approx(1.0, rel=1e-3)
def similar_ementas(ref, list):
referencia = nlp(ref)
for reference in list:
yield (list.index(reference),
similarity.word_movers(referencia,
nlp(reference),
metric='cosine'))
def similar(arquivo_referencia, lista_pets_escritorio):
referencia = nlp(resumidor(arquivo_referencia))
for peticao_referencia in lista_pets_escritorio:
yield (lista_pets_escritorio.index(peticao_referencia),
similarity.word_movers(referencia,
nlp(resumidor(peticao_referencia)),
metric='canberra'))
def get_word_movers(text_1, text_2):
"""
Gets textacy word movers number from comparing two texts.
Number between 0.0 to 1.0 where 0 is no similarity and 1 full similarity.
:param text_1: First text
:param text_2: Second text
:return: Returns tuple with result of word movers and both processed texts
"""
doc_1, preprocess_text_1 = NLPService.get_textacy_doc(text_1)
doc_2, preprocess_text_2 = NLPService.get_textacy_doc(text_2)
word_mover = word_movers(doc_1, doc_2, metric="cosine")
return word_mover, f"Zpracovaný text 1:\n{preprocess_text_1}\n\nZpracovaný text 2:\n{preprocess_text_2}"
def test_word_movers_identity(doc1, doc2):
assert similarity.word_movers(doc1, doc1) == pytest.approx(1.0, rel=1e-3)
def test_word_movers_metrics(doc1, doc2):
metrics = ("cosine", "l1", "manhattan", "l2", "euclidean")
for metric in metrics:
assert 0.0 <= similarity.word_movers(doc1, doc2, metric=metric) <= 1.0
def similar_tags(resumo_ref, tag_list):
referencia = nlp(resumo_ref)
for ref in tag_list:
yield (tag_list.index(ref),
similarity.word_movers(referencia, nlp(ref), metric='canberra'))
def score_summary_2(summary_text):
"""Score a summarized piece of text
"""
# Want high similarity between paragraphs
inter_paragraph_similarities = []
avg_similarity = None
sentences = [i.text for i in NLP(summary_text).sents]
# readability measures close to ebert baseline
readability = abs(
text_stats.TextStats(NLP(summary_text)).automated_readability_index -
EBERT_READABILITY) / EBERT_READABILITY
# Coh Metrix Indices
anaphor_score = anaphor_overlap(summary_text)
person_score = person_overlap(summary_text)
# more subjective is better
total_subjectivity = 0
for i in sentences:
total_subjectivity += TextBlob(i).sentiment[1]
subjectivity = total_subjectivity / len(sentences)
# thesis sentence doesn't have "this", "here", "it"
if sentences[0] not in [' ', '', '\n']:
thesis_penalty = sum(i in sentences[0]
for i in [" this ", " This ", " here ", " Here"])
elif sentences[1] not in [' ', '', '\n']:
thesis_penalty = sum(i in sentences[1]
for i in [" this ", " This ", " here ", " Here"])
else:
thesis_penalty = 0
# Prefer expressions from the author
author_count = 0
for s in sentences:
if any(i in s for i in ["I ", "I'd", "My"]):
author_count += 1
# iterate through the paragraphs
# sentiment within a paragraph is similar
paragraphs = summary_text.split('\n')
for i in range(1, len(paragraphs)):
if paragraphs[i - 1] not in [' ', '', '\n'] and paragraphs[i] not in [
' ', '', '\n'
]:
inter_paragraph_similarities.append(
similarity.word_movers(NLP(paragraphs[i - 1]),
NLP(paragraphs[i])))
max_diff = 0
for p in paragraphs:
p_sent_min = None
p_sent_max = None
for s in p.split('.'):
sent = TextBlob(s).sentiment[0]
if p_sent_min is None:
p_sent_min = sent
if p_sent_max is None:
p_sent_max = sent
if sent < p_sent_min:
p_sent_min = sent
if sent > p_sent_max:
p_sent_max = sent
if max_diff < abs(p_sent_max - p_sent_min):
max_diff = abs(p_sent_max - p_sent_min)
max_diff = 1 - max_diff
avg_similarity = sum(inter_paragraph_similarities) / len(
inter_paragraph_similarities)
# Make score
score = (0.25 * avg_similarity) + \
(0.20 * person_score) + \
(0.15 * anaphor_score) + \
(0.1 * max_diff) + \
(0.05 * readability) + \
(0.25 * subjectivity)
# boost by person count
score = score * (1 + (0.1 * author_count))
score = score - (0.2 * thesis_penalty)
return score
def test_word_movers_metrics(doc1, doc2):
metrics = ('cosine', 'l1', 'manhattan', 'l2', 'euclidean')
for metric in metrics:
assert 0.0 <= similarity.word_movers(doc1, doc2, metric=metric) <= 1.0
