def test_read_from_file(self, lazy):
conll_reader = ConllCorefReader(max_span_width=self.span_width, lazy=lazy)
instances = ensure_list(conll_reader.read(str(AllenNlpTestCase.FIXTURES_ROOT /
'coref' / 'coref.gold_conll')))
assert len(instances) == 2
fields = instances[0].fields
text = [x.text for x in fields["text"].tokens]
assert text == ['In', 'the', 'summer', 'of', '2005', ',', 'a', 'picture', 'that',
'people', 'have', 'long', 'been', 'looking', 'forward', 'to',
'started', 'emerging', 'with', 'frequency', 'in', 'various', 'major',
'Hong', 'Kong', 'media', '.', 'With', 'their', 'unique', 'charm', ',',
'these', 'well', '-', 'known', 'cartoon', 'images', 'once', 'again',
'caused', 'Hong', 'Kong', 'to', 'be', 'a', 'focus', 'of', 'worldwide',
'attention', '.', 'The', 'world', "'s", 'fifth', 'Disney', 'park',
'will', 'soon', 'open', 'to', 'the', 'public', 'here', '.']
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(span_starts, span_ends, text)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [(candidate_mentions[i], x)
for i, x in gold_indices_with_ids]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
gold_mentions_with_ids.remove((["Hong", "Kong"], 0))
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
assert (["their"], 1) in gold_mentions_with_ids
# This is a span which exceeds our max_span_width, so it should not be considered.
assert not (["these", "well", "known", "cartoon", "images"], 1) in gold_mentions_with_ids
fields = instances[1].fields
text = [x.text for x in fields["text"].tokens]
assert text == ['The', 'area', 'of', 'Hong', 'Kong', 'is', 'only', 'one', 'thousand', '-', 'plus',
'square', 'kilometers', '.', 'The', 'population', 'is', 'dense', '.', 'Natural',
'resources', 'are', 'relatively', 'scarce', '.', 'However', ',', 'the', 'clever',
'Hong', 'Kong', 'people', 'will', 'utilize', 'all', 'resources', 'they', 'have',
'created', 'for', 'developing', 'the', 'Hong', 'Kong', 'tourism', 'industry', '.']
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(span_starts, span_ends, text)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [(candidate_mentions[i], x)
for i, x in gold_indices_with_ids]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
gold_mentions_with_ids.remove((["Hong", "Kong"], 0))
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
assert (["they"], 1) in gold_mentions_with_ids
assert (['the', 'clever', 'Hong', 'Kong', 'people'], 1) in gold_mentions_with_ids
def test_read_from_file(self, lazy):
conll_reader = ConllCorefReader(max_span_width=self.span_width, lazy=lazy)
instances = ensure_list(conll_reader.read('tests/fixtures/coref/coref.gold_conll'))
assert len(instances) == 2
fields = instances[0].fields
text = [x.text for x in fields["text"].tokens]
assert text == ['In', 'the', 'summer', 'of', '2005', ',', 'a', 'picture', 'that',
'people', 'have', 'long', 'been', 'looking', 'forward', 'to',
'started', 'emerging', 'with', 'frequency', 'in', 'various', 'major',
'Hong', 'Kong', 'media', '.', 'With', 'their', 'unique', 'charm', ',',
'these', 'well', '-', 'known', 'cartoon', 'images', 'once', 'again',
'caused', 'Hong', 'Kong', 'to', 'be', 'a', 'focus', 'of', 'worldwide',
'attention', '.', 'The', 'world', "'s", 'fifth', 'Disney', 'park',
'will', 'soon', 'open', 'to', 'the', 'public', 'here', '.']
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(span_starts, span_ends, text)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [(candidate_mentions[i], x)
for i, x in gold_indices_with_ids]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
gold_mentions_with_ids.remove((["Hong", "Kong"], 0))
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
assert (["their"], 1) in gold_mentions_with_ids
# This is a span which exceeds our max_span_width, so it should not be considered.
assert not (["these", "well", "known", "cartoon", "images"], 1) in gold_mentions_with_ids
fields = instances[1].fields
text = [x.text for x in fields["text"].tokens]
assert text == ['The', 'area', 'of', 'Hong', 'Kong', 'is', 'only', 'one', 'thousand', '-', 'plus',
'square', 'kilometers', '.', 'The', 'population', 'is', 'dense', '.', 'Natural',
'resources', 'are', 'relatively', 'scarce', '.', 'However', ',', 'the', 'clever',
'Hong', 'Kong', 'people', 'will', 'utilize', 'all', 'resources', 'they', 'have',
'created', 'for', 'developing', 'the', 'Hong', 'Kong', 'tourism', 'industry', '.']
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(span_starts, span_ends, text)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [(candidate_mentions[i], x)
for i, x in gold_indices_with_ids]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
gold_mentions_with_ids.remove((["Hong", "Kong"], 0))
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
assert (["they"], 1) in gold_mentions_with_ids
assert (['the', 'clever', 'Hong', 'Kong', 'people'], 1) in gold_mentions_with_ids
def test_max_sentences(self):
conll_reader = ConllCorefReader(max_span_width=self.span_width)
instances = ensure_list(
conll_reader.read(
str(AllenNlpTestCase.FIXTURES_ROOT / "coref" /
"coref.gold_conll")))
limited_conll_reader = ConllCorefReader(max_span_width=self.span_width,
max_sentences=2)
limited_instances = ensure_list(
limited_conll_reader.read(
str(AllenNlpTestCase.FIXTURES_ROOT / "coref" /
"coref.gold_conll")))
assert len(limited_instances) == len(instances) == 4
tokens_of = lambda instance: instance.fields["text"].tokens
text_of = lambda tokens: [token.text for token in tokens]
docs = [tokens_of(instance) for instance in instances]
limited_docs = [tokens_of(instance) for instance in limited_instances]
# Short ones; not truncated
assert limited_docs[1] == docs[1]
assert limited_docs[3] == docs[3]
# Truncation happened
assert len(limited_docs[0]) < len(docs[0])
assert len(limited_docs[2]) < len(docs[2])
assert "Disney" in text_of(docs[0]) and "Disney" not in text_of(
limited_docs[0])
assert "tourism" in text_of(docs[2]) and "tourism" not in text_of(
limited_docs[2])
# Truncated tokens are the prefixes
assert limited_docs[0] == docs[0][:len(limited_docs[0])]
assert limited_docs[2] == docs[2][:len(limited_docs[2])]
def test_read_from_file(self):
conll_reader = ConllCorefReader(max_span_width=self.span_width)
dataset = conll_reader.read('tests/fixtures/conll_2012/')
assert len(dataset.instances) == 1
instances = dataset.instances
fields = instances[0].fields
text = [x.text for x in fields["text"].tokens]
assert text == [
'Mali', 'government', 'officials', 'say', 'the', 'woman', "'s",
'confession', 'was', 'forced', '.', 'The', 'prosecution', 'rested',
'its', 'case', 'last', 'month', 'after', 'four', 'months', 'of',
'hearings', '.', 'Denise', 'Dillon', 'Headline', 'News', '.'
]
span_starts = fields["span_starts"].field_list
span_ends = fields["span_ends"].field_list
candidate_mentions = self.check_candidate_mentions_are_well_defined(
span_starts, span_ends, text)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x)
for i, x in enumerate(gold_span_labels.labels)
if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [
(candidate_mentions[i], x) for i, x in gold_indices_with_ids
]
assert gold_mentions_with_ids == [
(['the', 'woman', "'s"], 0),
(['the', 'woman', "'s", 'confession'], 1),
(['The', 'prosecution'], 2), (['its'], 2), (['Denise',
'Dillon'], 2)
]
# Now check that we don't collect spans greater than the max width.
conll_reader = ConllCorefReader(max_span_width=2)
dataset = conll_reader.read('tests/fixtures/conll_2012/')
instances = dataset.instances
fields = instances[0].fields
text = [x.text for x in fields["text"].tokens]
span_starts = fields["span_starts"].field_list
span_ends = fields["span_ends"].field_list
candidate_mentions = self.check_candidate_mentions_are_well_defined(
span_starts, span_ends, text)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x)
for i, x in enumerate(gold_span_labels.labels)
if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [
(candidate_mentions[i], x) for i, x in gold_indices_with_ids
]
assert gold_mentions_with_ids == [(['The', 'prosecution'], 2),
(['its'], 2),
(['Denise', 'Dillon'], 2)]
def test_wordpiece_modeling(self):
tokenizer = PretrainedTransformerTokenizer("bert-base-cased")
conll_reader = ConllCorefReader(
max_span_width=self.span_width, wordpiece_modeling_tokenizer=tokenizer
)
instances = ensure_list(
conll_reader.read(str(AllenNlpTestCase.FIXTURES_ROOT / "coref" / "coref.gold_conll"))
)
assert len(instances) == 4
fields = instances[3].fields
text = [x.text for x in fields["text"].tokens]
assert text == [
"[CLS]",
"Hong",
"Kong",
"Wet",
"##land",
"Park",
",",
"which",
"is",
"currently",
"under",
"construction",
",",
"is",
"also",
"one",
"of",
"the",
"designated",
"new",
"projects",
"of",
"the",
"Hong",
"Kong",
"SA",
"##R",
"government",
"for",
"advancing",
"the",
"Hong",
"Kong",
"tourism",
"industry",
".",
"[SEP]",
]
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(
span_starts, span_ends, text
)
# Asserts special tokens aren't included in the spans
assert all(span_start > 0 for span_start in span_starts)
assert all(span_end < len(text) - 1 for span_end in span_ends)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [
(candidate_mentions[i], x) for i, x in gold_indices_with_ids
]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
# Within span_width before wordpiece splitting but exceeds afterwards
assert (["the", "Hong", "Kong", "SA", "##R", "government"], 0) not in gold_mentions_with_ids
fields = instances[1].fields
text = [x.text for x in fields["text"].tokens]
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(
span_starts, span_ends, text
)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [
(candidate_mentions[i], x) for i, x in gold_indices_with_ids
]
# Prior to wordpiece tokenization, 's was one token; wordpiece tokenization splits it into 2
assert (["the", "city", "'", "s"], 0) in gold_mentions_with_ids
def test_read_from_file(self, lazy):
conll_reader = ConllCorefReader(max_span_width=self.span_width, lazy=lazy)
instances = ensure_list(
conll_reader.read(str(AllenNlpTestCase.FIXTURES_ROOT / "coref" / "coref.gold_conll"))
)
assert len(instances) == 4
fields = instances[0].fields
text = [x.text for x in fields["text"].tokens]
assert text == [
"In",
"the",
"summer",
"of",
"2005",
",",
"a",
"picture",
"that",
"people",
"have",
"long",
"been",
"looking",
"forward",
"to",
"started",
"emerging",
"with",
"frequency",
"in",
"various",
"major",
"Hong",
"Kong",
"media",
".",
"With",
"their",
"unique",
"charm",
",",
"these",
"well",
"-",
"known",
"cartoon",
"images",
"once",
"again",
"caused",
"Hong",
"Kong",
"to",
"be",
"a",
"focus",
"of",
"worldwide",
"attention",
".",
"The",
"world",
"'s",
"fifth",
"Disney",
"park",
"will",
"soon",
"open",
"to",
"the",
"public",
"here",
".",
]
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(
span_starts, span_ends, text
)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [
(candidate_mentions[i], x) for i, x in gold_indices_with_ids
]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
gold_mentions_with_ids.remove((["Hong", "Kong"], 0))
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
assert (["their"], 1) in gold_mentions_with_ids
# This is a span which exceeds our max_span_width, so it should not be considered.
assert (
["these", "well", "-", "known", "cartoon", "images"],
1,
) not in gold_mentions_with_ids
fields = instances[2].fields
text = [x.text for x in fields["text"].tokens]
assert text == [
"The",
"area",
"of",
"Hong",
"Kong",
"is",
"only",
"one",
"thousand",
"-",
"plus",
"square",
"kilometers",
".",
"The",
"population",
"is",
"dense",
".",
"Natural",
"resources",
"are",
"relatively",
"scarce",
".",
"However",
",",
"the",
"clever",
"Hong",
"Kong",
"people",
"will",
"utilize",
"all",
"resources",
"they",
"have",
"created",
"for",
"developing",
"the",
"Hong",
"Kong",
"tourism",
"industry",
".",
]
spans = fields["spans"].field_list
span_starts, span_ends = zip(*[(field.span_start, field.span_end) for field in spans])
candidate_mentions = self.check_candidate_mentions_are_well_defined(
span_starts, span_ends, text
)
gold_span_labels = fields["span_labels"]
gold_indices_with_ids = [(i, x) for i, x in enumerate(gold_span_labels.labels) if x != -1]
gold_mentions_with_ids: List[Tuple[List[str], int]] = [
(candidate_mentions[i], x) for i, x in gold_indices_with_ids
]
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
gold_mentions_with_ids.remove((["Hong", "Kong"], 0))
assert (["Hong", "Kong"], 0) in gold_mentions_with_ids
assert (["they"], 1) in gold_mentions_with_ids
assert (["the", "clever", "Hong", "Kong", "people"], 1) in gold_mentions_with_ids