def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
article_headline = get_tokenized_lemmas(s.articleHeadline)
claim_headline = get_tokenized_lemmas(s.claimHeadline)
mat[i, 0] = len(set(article_headline).intersection(claim_headline)) / \
float(len(set(article_headline).union(claim_headline)))
return mat
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
article_headline = get_tokenized_lemmas(s.articleHeadline)
claim_headline = get_tokenized_lemmas(s.claimHeadline)
mat[i, 0] = len(set(article_headline).intersection(claim_headline)) / \
float(len(set(article_headline).union(claim_headline)))
return mat
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
idx = get_aligned_data().get((s.claimId, s.articleId))
if idx:
try:
claim_tok = get_tokenized_lemmas(s.claimHeadline)
article_tok = get_tokenized_lemmas(s.articleHeadline)
mat[i, 0] = self._sts(claim_tok, article_tok, idx)
except:
pass
return mat
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
idx = get_aligned_data().get((s.claimId, s.articleId))
if idx:
try:
claim_tok = get_tokenized_lemmas(s.claimHeadline)
article_tok = get_tokenized_lemmas(s.articleHeadline)
mat[i, 0] = self._sts(claim_tok, article_tok, idx)
except:
pass
return mat
def transform(self, X):
mat = np.zeros((len(X), len(_refuting_words)))
for i, (_, s) in enumerate(X.iterrows()):
# article_headline = [get_stem(w) for w in get_tokenized_lemmas(s.articleHeadline)]
article_headline = get_tokenized_lemmas(s.articleHeadline)
# mat[i, :] = np.array([1 if get_stem(w) in article_headline else 0 for w in _refuting_words])
mat[i, :] = np.array([1 if w in article_headline else 0 for w in _refuting_words])
return mat
def transform(self, X):
bc_data, bc_data_idx = get_brown_cluster_data(self.cluster_size)
mat = dok_matrix((len(X), len(bc_data_idx.values())), dtype=np.float32)
for i, (_, s) in enumerate(X.iterrows()):
claim_headline = get_tokenized_lemmas(s.claimHeadline)
article_headline = get_tokenized_lemmas(s.articleHeadline)
word_pairs = it.product(article_headline, claim_headline)
for v, w in word_pairs:
v_cluster = bc_data.get(v)
w_cluster = bc_data.get(w)
if v_cluster is None or w_cluster is None:
continue
idx = bc_data_idx[(v_cluster, w_cluster)]
mat[i, idx] = 1
return mat
def transform(self, X):
bc_data, bc_data_idx = get_brown_cluster_data(self.cluster_size)
mat = dok_matrix((len(X), len(bc_data_idx.values())), dtype=np.float32)
for i, (_, s) in enumerate(X.iterrows()):
claim_headline = get_tokenized_lemmas(s.claimHeadline)
article_headline = get_tokenized_lemmas(s.articleHeadline)
word_pairs = it.product(article_headline, claim_headline)
for v, w in word_pairs:
v_cluster = bc_data.get(v)
w_cluster = bc_data.get(w)
if v_cluster is None or w_cluster is None:
continue
idx = bc_data_idx[(v_cluster, w_cluster)]
mat[i, idx] = 1
return mat
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
idx = get_aligned_data().get((s.claimId, s.articleId))
f = 0
if idx:
claim_tok = get_tokenized_lemmas(s.claimHeadline)
article_tok = get_tokenized_lemmas(s.articleHeadline)
for x, y in idx:
if x > 0 and y == 0:
f = self._match(claim_tok[x-1])
elif x == 0 and y > 0:
f = self._match(article_tok[y-1])
elif [x-1, y-1] not in idx:
f = self._match(claim_tok[x-1]) or self._match(article_tok[y-1])
mat[i, 0] = f
return mat
def transform(self, X):
mat = np.zeros((len(X), len(_refuting_words)))
for i, (_, s) in enumerate(X.iterrows()):
# article_headline = [get_stem(w) for w in get_tokenized_lemmas(s.articleHeadline)]
article_headline = get_tokenized_lemmas(s.articleHeadline)
# mat[i, :] = np.array([1 if get_stem(w) in article_headline else 0 for w in _refuting_words])
mat[i, :] = np.array(
[1 if w in article_headline else 0 for w in _refuting_words])
return mat
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
idx = get_aligned_data().get((s.claimId, s.articleId))
f = 0
if idx:
claim_tok = get_tokenized_lemmas(s.claimHeadline)
article_tok = get_tokenized_lemmas(s.articleHeadline)
for x, y in idx:
if x > 0 and y == 0:
f = self._match(claim_tok[x - 1])
elif x == 0 and y > 0:
f = self._match(article_tok[y - 1])
elif [x - 1, y - 1] not in idx:
f = self._match(claim_tok[x - 1]) or self._match(
article_tok[y - 1])
mat[i, 0] = f
return mat
def calc_hungarian_alignment_score(s, t):
"""Calculate the alignment score between the two texts s and t
using the implementation of the Hungarian alignment algorithm
provided in https://pypi.python.org/pypi/munkres/."""
s_toks = get_tokenized_lemmas(s)
t_toks = get_tokenized_lemmas(t)
df = pd.DataFrame(index=s_toks, columns=t_toks, data=0.)
for c in s_toks:
for a in t_toks:
df.ix[c, a] = compute_paraphrase_score(c, a)
matrix = df.values
cost_matrix = make_cost_matrix(matrix, lambda cost: _max_ppdb_score - cost)
indexes = _munk.compute(cost_matrix)
total = 0.0
for row, column in indexes:
value = matrix[row][column]
total += value
return indexes, total / float(np.min(matrix.shape))
def calc_hungarian_alignment_score(s, t):
"""Calculate the alignment score between the two texts s and t
using the implementation of the Hungarian alignment algorithm
provided in https://pypi.python.org/pypi/munkres/."""
s_toks = get_tokenized_lemmas(s)
t_toks = get_tokenized_lemmas(t)
df = pd.DataFrame(index=s_toks, columns=t_toks, data=0.)
for c in s_toks:
for a in t_toks:
df.ix[c, a] = compute_paraphrase_score(c, a)
matrix = df.values
cost_matrix = make_cost_matrix(matrix, lambda cost: _max_ppdb_score - cost)
indexes = _munk.compute(cost_matrix)
total = 0.0
for row, column in indexes:
value = matrix[row][column]
total += value
return indexes, total / float(np.min(matrix.shape))
def _compute_overlap(row):
claim_lemmas = get_tokenized_lemmas(row.claimHeadline)
article_lemmas = get_tokenized_lemmas(row.articleHeadline)
intersect = set(claim_lemmas).intersection(article_lemmas)
union = set(claim_lemmas).union(article_lemmas)
return float(len(intersect)) / len(union)
def _calc_polarity(s):
tokens = get_tokenized_lemmas(s)
return sum([t in _refuting_words for t in tokens]) % 2
def _get_bigram_clusters(s, bc_data):
clusters = filter(None, [bc_data.get(l) for l in get_tokenized_lemmas(s)])
return ngrams(clusters, 2)
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
if '?' in get_tokenized_lemmas(s.articleHeadline):
mat[i, 0] = 1
return mat
def _get_bigram_clusters(s, bc_data):
clusters = filter(None, [bc_data.get(l) for l in get_tokenized_lemmas(s)])
return ngrams(clusters, 2)
def transform(self, X):
mat = np.zeros((len(X), 1))
for i, (_, s) in enumerate(X.iterrows()):
if '?' in get_tokenized_lemmas(s.articleHeadline):
mat[i, 0] = 1
return mat
def _calc_polarity(s):
tokens = get_tokenized_lemmas(s)
return sum([t in _refuting_words for t in tokens]) % 2
import pickle
from aligner import align
from model.utils import get_dataset, get_tokenized_lemmas
def _get_unaligned_tokens(tokens, alignment):
aligned = [a-1 for (a, _) in alignment]
unaligned = [i for i in range(len(tokens)) if i not in aligned]
return [tokens[i] for i in unaligned]
if __name__ == "__main__":
df = get_dataset()
data = {}
for id, row in df.iterrows():
article_hl_tok = get_tokenized_lemmas(row.articleHeadline)
claim_hl_tok = get_tokenized_lemmas(row.claimHeadline)
try:
alignment = align(claim_hl_tok, article_hl_tok)
data[(row.claimId, row.articleId)] = [(s-1, t-1) for (s, t) in alignment[0]]
except:
print 'Unable to align', article_hl_tok, 'and', claim_hl_tok
print row.articleId, row.claimId
with open(os.path.join('..', 'data', 'pickled', 'aligned-data.pickle'), 'wb') as f:
pickle.dump(data, f, pickle.HIGHEST_PROTOCOL)
from aligner import align
from model.utils import get_dataset, get_tokenized_lemmas
def _get_unaligned_tokens(tokens, alignment):
aligned = [a - 1 for (a, _) in alignment]
unaligned = [i for i in range(len(tokens)) if i not in aligned]
return [tokens[i] for i in unaligned]
if __name__ == "__main__":
df = get_dataset()
data = {}
for id, row in df.iterrows():
article_hl_tok = get_tokenized_lemmas(row.articleHeadline)
claim_hl_tok = get_tokenized_lemmas(row.claimHeadline)
try:
alignment = align(claim_hl_tok, article_hl_tok)
data[(row.claimId, row.articleId)] = [(s - 1, t - 1)
for (s, t) in alignment[0]]
except:
print 'Unable to align', article_hl_tok, 'and', claim_hl_tok
print row.articleId, row.claimId
with open(os.path.join('..', 'data', 'pickled', 'aligned-data.pickle'),
'wb') as f:
pickle.dump(data, f, pickle.HIGHEST_PROTOCOL)
def _compute_overlap(row):
claim_lemmas = get_tokenized_lemmas(row.claimHeadline)
article_lemmas = get_tokenized_lemmas(row.articleHeadline)
intersect = set(claim_lemmas).intersection(article_lemmas)
union = set(claim_lemmas).union(article_lemmas)
return float(len(intersect)) / len(union)
