def preserve_case(token):
"""
Returns True if `token` is a proper noun or acronym, False otherwise.
Args:
token (``spacy.Token``): parent document must have POS information
Returns:
bool
"""
if token.doc.is_tagged is False:
raise ValueError('token is not POS-tagged')
return token.pos == PROPN or is_acronym(token.text)
def preserve_case(token):
"""
Returns True if `token` is a proper noun or acronym, False otherwise.
Args:
token (``spacy.Token``): parent document must have POS information
Returns:
bool
"""
if token.doc.is_tagged is False:
raise ValueError('token is not POS-tagged')
return token.pos == PROPN or is_acronym(token.text)
def preserve_case(token):
"""
Returns True if `token` is a proper noun or acronym, False otherwise.
Args:
token (``spacy.Token``): parent document must have POS information
Returns:
bool
TODO: use universal pos PROPN instead of english-specific tags as soon as
Honnibal decides to include them in his model...
"""
if token.doc.is_tagged is False:
raise ValueError('token is not POS-tagged')
return token.tag_ in {'NNP', 'NNPS'} or is_acronym(token.text)
def _get_acronym_definition(acronym, window, threshold=0.8):
"""
Identify most likely definition for an acronym given a list of tokens.
Args:
acronym (str): acronym for which definition is sought
window (``spacy.Span``): a span of tokens from which definition
extraction will be attempted
threshold (float, optional): minimum "confidence" in definition required
for acceptance; valid values in [0.0, 1.0]; higher value => stricter threshold
Returns:
Tuple[str, float]: most likely definition for given acronym ('' if none found),
along with the confidence assigned to it
References:
Taghva, Kazem, and Jeff Gilbreth. "Recognizing acronyms and their definitions."
International Journal on Document Analysis and Recognition 1.4 (1999): 191-198.
"""
def build_lcs_matrix(X, Y):
m = len(X)
n = len(Y)
b = zeros((m, n), dtype=int)
c = zeros((m, n), dtype=int)
for i in range(0, m):
for j in range(0, n):
if X[i] == Y[j]:
c[i, j] = c[i - 1, j - 1] + 1
b[i, j] = 1
elif c[i - 1, j] >= c[i, j - 1]:
c[i, j] = c[i - 1, j]
else:
c[i, j] = c[i, j - 1]
return c, b
def parse_lcs_matrix(b, start_i, start_j, lcs_length, stack, vectors):
m = b.shape[0]
n = b.shape[1]
for i in range(start_i, m):
for j in range(start_j, n):
if b[i, j] == 1:
s = (i, j)
stack.append(s)
if lcs_length == 1:
vec = [NaN] * n
for k, l in stack:
vec[l] = k
vectors.append(vec)
else:
parse_lcs_matrix(b, i + 1, j + 1, lcs_length - 1, stack, vectors)
stack = []
return vectors
def vector_values(v, types):
vv = {}
first = v.index(int(nanmin(v)))
last = v.index(int(nanmax(v)))
vv['size'] = (last - first) + 1
vv['distance'] = len(v) - last
vv['stop_count'] = 0
vv['misses'] = 0
for i in range(first, last + 1):
if v[i] >= 0 and types[i] == 's':
vv['stop_count'] += 1
elif v[i] is None and types[i] not in ['s', 'h']:
vv['misses'] += 1
return vv
def compare_vectors(A, B, types):
vv_A = vector_values(A, types)
vv_B = vector_values(B, types)
# no one-letter matches, sorryboutit
if vv_A['size'] == 1:
return B
elif vv_B['size'] == 1:
return A
if vv_A['misses'] > vv_B['misses']:
return B
elif vv_A['misses'] < vv_B['misses']:
return A
if vv_A['stop_count'] > vv_B['stop_count']:
return B
if vv_A['stop_count'] < vv_B['stop_count']:
return A
if vv_A['distance'] > vv_B['distance']:
return B
elif vv_A['distance'] < vv_B['distance']:
return A
if vv_A['size'] > vv_B['size']:
return B
elif vv_A['size'] < vv_B['size']:
return A
return A
# get definition window's leading characters and word types
def_leads = []
def_types = []
for tok in window:
tok_text = tok.text
if tok.is_stop:
def_leads.append(tok_text[0])
def_types.append('s')
elif text_utils.is_acronym(tok_text):
def_leads.append(tok_text[0])
def_types.append('a')
elif '-' in tok_text and not tok_text.startswith('-'):
tok_split = [t[0] for t in tok_text.split('-') if t]
def_leads.extend(tok_split)
def_types.extend('H' if i == 0 else 'h' for i in range(len(tok_split)))
else:
def_leads.append(tok_text[0])
def_types.append('w')
def_leads = ''.join(def_leads).lower()
def_types = ''.join(def_types)
# extract alphanumeric characters from acronym
acr_leads = ''.join(c for c in acronym if c.isalnum())
# handle special cases of '&' and trailing 's'
acr_leads = acr_leads.replace('&', 'a')
if acr_leads.endswith('s'):
# bail out if it's only a 2-letter acronym to start with, e.g. 'Is'
if len(acr_leads) == 2:
return ('', 0)
acr_leads = acr_leads[:-1]
acr_leads = acr_leads.lower()
c, b = build_lcs_matrix(acr_leads, def_leads)
# 4.4.1
lcs_length = c[c.shape[0] - 1, c.shape[1] - 1]
confidence = lcs_length / len(acronym)
if confidence < threshold:
return ('', confidence)
vecs = parse_lcs_matrix(b, 0, 0, lcs_length, [], [])
# first letter of acronym must be present
vecs = [vec for vec in vecs if 0 in vec]
if not vecs:
return ('', confidence)
best_vec = vecs[0]
for vec in vecs[1:]:
best_vec = compare_vectors(best_vec, vec, def_types)
first = best_vec.index(int(nanmin(best_vec)))
last = best_vec.index(int(nanmax(best_vec)))
definition = window[first: last + 1].text
if len(definition.split()) == 1:
return ('', confidence)
return (definition, confidence)
def acronyms_and_definitions(doc, known_acro_defs=None):
"""
Extract a collection of acronyms and their most likely definitions, if available,
from a spacy-parsed doc. If multiple definitions are found for a given acronym,
only the most frequently occurring definition is returned.
Args:
doc (``textacy.Doc`` or ``spacy.Doc`` or ``spacy.Span``)
known_acro_defs (dict, optional): if certain acronym/definition pairs
are known, pass them in as {acronym (str): definition (str)};
algorithm will not attempt to find new definitions
Returns:
dict: unique acronyms (keys) with matched definitions (values)
References:
Taghva, Kazem, and Jeff Gilbreth. "Recognizing acronyms and their definitions."
International Journal on Document Analysis and Recognition 1.4 (1999): 191-198.
"""
# process function arguments
acro_defs = defaultdict(list)
if not known_acro_defs:
known_acronyms = set()
else:
for acro, defs in known_acro_defs.items():
if not isinstance(defs, list):
acro_defs[acro] = [defs]
known_acronyms = set(acro_defs.keys())
if isinstance(doc, SpacySpan):
sents = [doc]
else: # textacy.Doc or spacy.Doc
sents = doc.sents
# iterate over sentences and their tokens
for sent in sents:
max_ind = len(sent) - 1
for i, token in enumerate(sent):
token_ = token.text
if token_ in known_acronyms or text_utils.is_acronym(token_) is False:
continue
# define definition search window(s)
window_size = min(2 * len(token_), len(token_) + 5)
windows = [sent[max(i - window_size, 0): i],
sent[min(i + 1, max_ind): min(i + window_size + 1, max_ind)]]
# if candidate inside (X) or -X-, only look in pre-window
if 0 < i < max_ind:
adjacent_tokens = sent[i - 1].text + sent[i + 1].text
if adjacent_tokens in {'()', '--', '––'}:
windows.pop()
# iterate over possible windows
# filtering for valid definition strings
for window in windows:
window_ = window.text
# window text can't be all uppercase
if window_.isupper():
continue
# window can't contain separating punctuation
if '!' in window_ or '?' in window_ or ':' in window_ or ';' in window_:
continue
# acronym definition can't contain itself: no ouroboros!
if token_ in window_:
continue
# window must contain at least one character used in acronym
if not any(char in window_ for char in token_):
continue
definition, confidence = _get_acronym_definition(
token_, window, threshold=0.8)
if definition:
acro_defs[token_].append((definition, confidence))
if not acro_defs.get(token_):
acro_defs[token_].append(('', 0.0))
# vote by confidence score in the case of multiple definitions
for acro, defs in acro_defs.items():
if len(defs) == 1:
acro_defs[acro] = defs[0][0]
else:
acro_defs[acro] = sorted(defs, key=itemgetter(1), reverse=True)[0][0]
return dict(acro_defs)
def _get_acronym_definition(acronym, window, threshold=0.8):
"""
Identify most likely definition for an acronym given a list of tokens.
Args:
acronym (str): acronym for which definition is sought
window (``spacy.Span``): a span of tokens from which definition
extraction will be attempted
threshold (float, optional): minimum "confidence" in definition required
for acceptance; valid values in [0.0, 1.0]; higher value => stricter threshold
Returns:
(str, float): most likely definition for given acronym ('' if none found),
along with the confidence assigned to it
References:
Taghva, Kazem, and Jeff Gilbreth. "Recognizing acronyms and their definitions."
International Journal on Document Analysis and Recognition 1.4 (1999): 191-198.
"""
def build_lcs_matrix(X, Y):
m = len(X)
n = len(Y)
b = zeros((m, n), dtype=int)
c = zeros((m, n), dtype=int)
for i in range(0, m):
for j in range(0, n):
if X[i] == Y[j]:
c[i, j] = c[i - 1, j - 1] + 1
b[i, j] = 1
elif c[i - 1, j] >= c[i, j - 1]:
c[i, j] = c[i - 1, j]
else:
c[i, j] = c[i, j - 1]
return c, b
def parse_lcs_matrix(b, start_i, start_j, lcs_length, stack, vectors):
m = b.shape[0]
n = b.shape[1]
for i in range(start_i, m):
for j in range(start_j, n):
if b[i, j] == 1:
s = (i, j)
stack.append(s)
if lcs_length == 1:
vec = [NaN] * n
for k, l in stack:
vec[l] = k
vectors.append(vec)
else:
parse_lcs_matrix(b, i + 1, j + 1, lcs_length - 1, stack, vectors)
stack = []
return vectors
def vector_values(v, types):
vv = {}
first = v.index(int(nanmin(v)))
last = v.index(int(nanmax(v)))
vv['size'] = (last - first) + 1
vv['distance'] = len(v) - last
vv['stop_count'] = 0
vv['misses'] = 0
for i in range(first, last + 1):
if v[i] >= 0 and types[i] == 's':
vv['stop_count'] += 1
elif v[i] is None and types[i] not in ['s', 'h']:
vv['misses'] += 1
return vv
def compare_vectors(A, B, types):
vv_A = vector_values(A, types)
vv_B = vector_values(B, types)
# no one-letter matches, sorryboutit
if vv_A['size'] == 1:
return B
elif vv_B['size'] == 1:
return A
if vv_A['misses'] > vv_B['misses']:
return B
elif vv_A['misses'] < vv_B['misses']:
return A
if vv_A['stop_count'] > vv_B['stop_count']:
return B
if vv_A['stop_count'] < vv_B['stop_count']:
return A
if vv_A['distance'] > vv_B['distance']:
return B
elif vv_A['distance'] < vv_B['distance']:
return A
if vv_A['size'] > vv_B['size']:
return B
elif vv_A['size'] < vv_B['size']:
return A
return A
# get definition window's leading characters and word types
def_leads = []
def_types = []
for tok in window:
tok_text = tok.text
if tok.is_stop:
def_leads.append(tok_text[0])
def_types.append('s')
elif text_utils.is_acronym(tok_text):
def_leads.append(tok_text[0])
def_types.append('a')
elif '-' in tok_text and not tok_text.startswith('-'):
tok_split = [t[0] for t in tok_text.split('-') if t]
def_leads.extend(tok_split)
def_types.extend('H' if i == 0 else 'h' for i in range(len(tok_split)))
else:
def_leads.append(tok_text[0])
def_types.append('w')
def_leads = ''.join(def_leads).lower()
def_types = ''.join(def_types)
# extract alphanumeric characters from acronym
acr_leads = ''.join(c for c in acronym if c.isalnum())
# handle special cases of '&' and trailing 's'
acr_leads = acr_leads.replace('&', 'a')
if acr_leads.endswith('s'):
# bail out if it's only a 2-letter acronym to start with, e.g. 'Is'
if len(acr_leads) == 2:
return ('', 0)
acr_leads = acr_leads[:-1]
acr_leads = acr_leads.lower()
c, b = build_lcs_matrix(acr_leads, def_leads)
# 4.4.1
lcs_length = c[c.shape[0] - 1, c.shape[1] - 1]
confidence = lcs_length / len(acronym)
if confidence < threshold:
return ('', confidence)
vecs = parse_lcs_matrix(b, 0, 0, lcs_length, [], [])
# first letter of acronym must be present
vecs = [vec for vec in vecs if 0 in vec]
if not vecs:
return ('', confidence)
best_vec = vecs[0]
for vec in vecs[1:]:
best_vec = compare_vectors(best_vec, vec, def_types)
first = best_vec.index(int(nanmin(best_vec)))
last = best_vec.index(int(nanmax(best_vec)))
definition = window[first: last + 1].text
if len(definition.split()) == 1:
return ('', confidence)
return (definition, confidence)
def acronyms_and_definitions(doc, known_acro_defs=None):
"""
Extract a collection of acronyms and their most likely definitions, if available,
from a spacy-parsed doc. If multiple definitions are found for a given acronym,
only the most frequently occurring definition is returned.
Args:
doc (``spacy.Doc`` or ``spacy.Span``)
known_acro_defs (dict, optional): if certain acronym/definition pairs
are known, pass them in as {acronym (str): definition (str)};
algorithm will not attempt to find new definitions
Returns:
dict: unique acronyms (keys) with matched definitions (values)
References:
Taghva, Kazem, and Jeff Gilbreth. "Recognizing acronyms and their definitions."
International Journal on Document Analysis and Recognition 1.4 (1999): 191-198.
"""
# process function arguments
acro_defs = defaultdict(list)
if not known_acro_defs:
known_acronyms = set()
else:
for acro, defs in known_acro_defs.items():
if not isinstance(defs, list):
acro_defs[acro] = [defs]
known_acronyms = set(acro_defs.keys())
try:
sents = doc.sents
except AttributeError:
sents = [doc]
# iterate over sentences and their tokens
for sent in sents:
max_ind = len(sent) - 1
for i, token in enumerate(sent):
token_ = token.text
if token_ in known_acronyms or text_utils.is_acronym(token_) is False:
continue
# define definition search window(s)
window_size = min(2 * len(token_), len(token_) + 5)
windows = [sent[max(i - window_size, 0): i],
sent[min(i + 1, max_ind): min(i + window_size + 1, max_ind)]]
# if candidate inside (X) or -X-, only look in pre-window
if 0 < i < max_ind:
adjacent_tokens = sent[i - 1].text + sent[i + 1].text
if adjacent_tokens in {'()', '--', '––'}:
_ = windows.pop()
# iterate over possible windows
# filtering for valid definition strings
for window in windows:
window_ = window.text
# window text can't be all uppercase
if window_.isupper():
continue
# window can't contain separating punctuation
if '!' in window_ or '?' in window_ or ':' in window_ or ';' in window_:
continue
# acronym definition can't contain itself: no ouroboros!
if token_ in window_:
continue
# window must contain at least one character used in acronym
if not any(char in window_ for char in token_):
continue
definition, confidence = _get_acronym_definition(
token_, window, threshold=0.8)
if definition:
acro_defs[token_].append((definition, confidence))
if not acro_defs.get(token_):
acro_defs[token_].append(('', 0.0))
# vote by confidence score in the case of multiple definitions
for acro, defs in acro_defs.items():
if len(defs) == 1:
acro_defs[acro] = defs[0][0]
else:
acro_defs[acro] = sorted(defs, key=itemgetter(1), reverse=True)[0][0]
return dict(acro_defs)
def test_is_acronym_exclude(self):
self.assertFalse(text_utils.is_acronym('NASA', exclude={'NASA'}))
def test_is_acronym_bad(self):
for item in BAD_ACRONYMS:
self.assertFalse(text_utils.is_acronym(item))
def test_is_acronym_good(self):
for item in GOOD_ACRONYMS:
self.assertTrue(text_utils.is_acronym(item))
def test_is_acronym_bad(token):
assert not text_utils.is_acronym(token)
def test_is_acronym_exclude():
assert not text_utils.is_acronym("NASA", exclude={"NASA"})
def test_is_acronym_bad():
for item in BAD_ACRONYMS:
assert not text_utils.is_acronym(item)
def test_is_acronym_good():
for item in GOOD_ACRONYMS:
assert text_utils.is_acronym(item)
def test_is_acronym_exclude():
assert not text_utils.is_acronym('NASA', exclude={'NASA'})
def test_is_acronym_exclude(self):
self.assertFalse(text_utils.is_acronym('NASA', exclude={'NASA'}))
def test_is_acronym_bad(self):
for item in BAD_ACRONYMS:
self.assertFalse(text_utils.is_acronym(item))
def test_is_acronym_good(self):
for item in GOOD_ACRONYMS:
self.assertTrue(text_utils.is_acronym(item))
def test_is_acronym_good(token):
assert text_utils.is_acronym(token)
def test_is_acronym_exclude(token, exclude, expected):
assert text_utils.is_acronym(token, exclude=exclude) == expected
