def from_nersuite(cls, text):
"""Return Document given NERsuite format file."""
sentences = []
lines = []
offset = 0
for line in split_keep_separator(text):
if not line:
pass
elif not line.isspace():
lines.append(line)
else:
sentences.append(Sentence.from_nersuite(lines, offset))
if sentences[-1].tokens:
offset = sentences[-1].tokens[-1].end + 1 # guess
lines = []
if lines:
sentences.append(Sentence.from_nersuite(lines, offset))
# Add spaces for gaps implied by token positions but not
# explitly recorded in NERsuite format
for s, next_s in pairwise(sentences):
if s.tokens and next_s.tokens:
gap = next_s.tokens[0].start - s.tokens[-1].end
s.text = s.text + ' ' * gap
# Assure document-final newline (text file)
if sentences and not sentences[-1].text.endswith('\n'):
sentences[-1].text = sentences[-1].text + '\n'
text = ''.join(s.text for s in sentences)
return cls(text, sentences)
def retag_document(document, tagset):
if tagset == IO_TAGSET:
mapper = BIO_to_IO
elif tagset == IOBES_TAGSET:
mapper = BIO_to_IOBES
else:
raise ValueError('tagset {}'.format(tagset))
for sentence in document.sentences:
for t, next_t in pairwise(sentence.tokens, include_last=True):
next_tag = next_t.tag if next_t is not None else None
t.tag = mapper(t.tag, next_tag)
def retag_document(document, tagset):
if tagset == IO_TAGSET:
mapper = BIO_to_IO
elif tagset == IOBES_TAGSET:
mapper = BIO_to_IOBES
else:
raise ValueError('tagset {}'.format(tagset))
for sentence in document.sentences:
for t, next_t in pairwise(sentence.tokens, include_last=True):
next_tag = next_t.tag if next_t is not None else None
t.tag = mapper(t.tag, next_tag)
def document_to_ascii(document):
# traverse sentences, store sentence head (text before the
# first token), then traverse tokens in pairs, storing the
# tail of each token (space between the token and the next),
# convert token text to ASCII using convert_u2a(), redo token
# offsets, then redo the sentence text using the head, token
# texts and tails, and finally redo the document text using
# the sentences.
read_u2a_data()
document.unicode_text = document.text
offset = 0
for sentence in document.sentences:
sentence.unicode_text = sentence.text
sent_end = sentence.base_offset + len(sentence.text)
ascii_base_offset = offset
if sentence.tokens:
t_start = sentence.tokens[0].start
sent_head = document.text[sentence.base_offset:t_start]
else:
sent_head = sentence.text
assert not sent_head or sent_head.isspace()
offset += len(sent_head)
for t, next_t in pairwise(sentence.tokens, include_last=True):
t.unicode_text = t.text
tail_end = next_t.start if next_t is not None else sent_end
t.tail = document.text[t.end:tail_end]
assert not t.tail or t.tail.isspace()
t.text = convert_u2a(t.text)
t.start, t.end = offset, offset + len(t.text)
offset += len(t.text) + len(t.tail)
sentence.base_offset = ascii_base_offset
sentence.text = sent_head + ''.join([t.text + t.tail
for t in sentence.tokens])
assert sentence.is_valid()
document.text = ''.join(s.text for s in document.sentences)
assert document.is_valid()
def document_to_ascii(document):
# traverse sentences, store sentence head (text before the
# first token), then traverse tokens in pairs, storing the
# tail of each token (space between the token and the next),
# convert token text to ASCII using convert_u2a(), redo token
# offsets, then redo the sentence text using the head, token
# texts and tails, and finally redo the document text using
# the sentences.
read_u2a_data()
document.unicode_text = document.text
offset = 0
for sentence in document.sentences:
sentence.unicode_text = sentence.text
sent_end = sentence.base_offset + len(sentence.text)
ascii_base_offset = offset
if sentence.tokens:
t_start = sentence.tokens[0].start
sent_head = document.text[sentence.base_offset:t_start]
else:
sent_head = sentence.text
assert not sent_head or sent_head.isspace()
offset += len(sent_head)
for t, next_t in pairwise(sentence.tokens, include_last=True):
t.unicode_text = t.text
tail_end = next_t.start if next_t is not None else sent_end
t.tail = document.text[t.end:tail_end]
assert not t.tail or t.tail.isspace()
t.text = convert_u2a(t.text)
t.start, t.end = offset, offset + len(t.text)
offset += len(t.text) + len(t.tail)
sentence.base_offset = ascii_base_offset
sentence.text = sent_head + ''.join(
[t.text + t.tail for t in sentence.tokens])
assert sentence.is_valid()
document.text = ''.join(s.text for s in document.sentences)
assert document.is_valid()
def get_tagged(self, relative_offsets=False):
"""Return list of (type, start, end) based on Token tags.
If relative_offsets is True, start and end offsets are
relative to sentence beginning; otherwise, they are absolute
offsets into the document text.
"""
tagged = []
first = None
for t, next_t in pairwise(self.tokens, include_last=True):
if is_start_tag(t.tag):
first = t
if first and not (next_t and is_continue_tag(next_t.tag)):
tagged.append((first.tagged_type(), first.start, t.end))
first = None
if relative_offsets:
tagged = [(t[0], t[1] - self.base_offset, t[2] - self.base_offset)
for t in tagged]
return tagged
def tagged_spans(tags):
"""Given a sequence of tags, return corresponding list TypedSpan,
i.e. (type, start, end) triples where start and end are token
indices.
The end index is that of the first token occurring after a tagged
span, i.e. the spanned tokens are tokens[start:end].
"""
# TODO: eliminate redundancy with Sentence.get_tagged() (DRY!)
spans = []
first = None
index = 0
for t, next_t in pairwise(tags, include_last=True):
if is_start_tag(t):
first = index
if first is not None and not (next_t and is_continue_tag(next_t)):
_, type_ = parse_tag(t)
spans.append(TypedSpan(type_, first, index + 1))
first = None
index += 1
return spans
