def main(sentences, language, outfile, confidence, processes):
""" Perform entity linking over a set of input sentences.
The service is Dandelion Entity Extraction API:
https://dandelion.eu/docs/api/datatxt/nex/v1/ .
Links having confidence score below the given
threshold are discarded.
"""
def worker(row):
sentence = json.loads(row)
text = sentence.get('text')
if text:
sentence['linked_entities'] = link(text, confidence, language)
return json.dumps(sentence)
count = 0
for each in parallel.map(worker, sentences, processes):
outfile.write(each)
outfile.write('\n')
count += 1
if count % 1000 == 0:
logger.info('Linked %d sentences', count)
if count > 0:
logger.info("Dumped linked sentences to '%s'" % outfile.name)
logger.info('Done, linked %d sentences', count)
def main(sentences, model, language, outfile, processes, gazetteer):
gazetteer = reverse_gazetteer(json.load(gazetteer)) if gazetteer else {}
logger.info("Loading model from '%s' ...", model)
model, extractor = joblib.load(model)
classifier = SentenceClassifier(model, extractor, language, gazetteer)
def worker(batch):
data = (json.loads(s) for s in batch)
for classified in classifier.classify_sentences(data):
yield json.dumps(classified)
count = 0
for each in parallel.map(worker, sentences, batch_size=1000,
flatten=True, processes=processes):
outfile.write(each)
outfile.write('\n')
count += 1
if count % 1000 == 0:
logger.info('Classified %d sentences', count)
logger.info('Done, classified %d sentences', count)
if count > 0:
logger.info("Dumped classified sentences to '%s'", outfile.name)
def main(corpus, verbs, processes, outfile, sub_sentences):
""" Compute the LU distribution in the corpus, i.e. how many LUs per sentence
"""
global splitter, tagger, parser, all_verbs
splitter = PunktSentenceSplitter('en')
tagger = TTPosTagger('en')
parser = StanfordParser(
path_to_jar='dev/stanford-corenlp-3.6.0.jar',
path_to_models_jar='dev/stanford-corenlp-3.6.0-models.jar',
java_options=' -mx1G -Djava.ext.dirs=dev/'
) # no way to make classpath work
all_verbs = reduce(lambda x, y: x.union(y),
imap(set,
json.load(verbs).values()), set())
all_verbs.discard('be')
all_verbs.discard('have')
args = load_corpus(corpus, 'bio', text_only=True)
worker = worker_with_sub_sentences if sub_sentences else worker_with_sentences
counter = defaultdict(int)
for i, counts in enumerate(parallel.map(worker, args, processes)):
for k, v in counts.iteritems():
counter[k] += v
if (i + 1) % 10000 == 0:
logger.info('Processed %d documents', i + 1)
counter = OrderedDict(sorted(counter.items(), key=lambda (k, v): k))
for k, v in counter.iteritems():
print k, v
json.dump(counter, outfile, indent=2)
def process_corpus(self, items, output_file, dump_unresolved_file=None, genealogics=None, processes=0):
count = skipped = 0
genealogics_url_to_id = {}
for success, item in parallel.map(self.serialize_item, items, processes, flatten=True):
if success:
subj, prop, val, url = item
statement = wikidata.finalize_statement(
subj, prop, val, self.language, url,
resolve_property=False, resolve_value=False
)
if not statement:
continue
output_file.write(statement.encode('utf8'))
output_file.write('\n')
if genealogics and url.startswith('http://www.genealogics.org/'):
genealogics_url_to_id[url] = subj
count += 1
if count % 10000 == 0:
logger.info('Produced %d statements so far, skipped %d names', count, skipped)
else:
skipped += 1
if dump_unresolved_file:
dump_unresolved_file.write(json.dumps(item))
dump_unresolved_file.write('\n')
logger.info('Produced %d statements so far, skipped %d names', count, skipped)
return genealogics_url_to_id, count, skipped
def main(sentences, language, outfile, confidence, processes):
""" Perform entity linking over a set of input sentences.
The service is Dandelion Entity Extraction API:
https://dandelion.eu/docs/api/datatxt/nex/v1/ .
Links having confidence score below the given
threshold are discarded.
"""
def worker(row):
sentence = json.loads(row)
text = sentence.get('text')
if text:
sentence['linked_entities'] = link(text, confidence, language)
return json.dumps(sentence)
count = 0
for each in parallel.map(worker, sentences, processes):
outfile.write(each)
outfile.write('\n')
count += 1
if count % 1000 == 0:
logger.info('Linked %d sentences', count)
if count > 0:
logger.info("Dumped linked sentences to '%s'" % outfile.name)
logger.info('Done, linked %d sentences', count)
def label_sentences(self, sentences, normalize_numerical, score_type, core_weight,
processes=0, input_encoded=False, output_encoded=False):
""" Process all the given sentences with the rule-based classifier,
optionally giving a confidence score
:param sentences: List of sentence data
:param normalize_numerical: Whether to automatically
normalize numerical expressions
:param score_type: Which type of score (if any) to use to
compute the classification confidence
:param core_weight: Weight of the core FEs (used in the scoring)
:param processes: how many processes to use to concurrently label sentences
:param input_encoded: whether the corpus is an iterable of dictionaries or an
iterable of JSON-encoded documents. JSON-encoded documents are preferable
over large size dictionaries for performance reasons
:param output_encoded: whether to return a generator of dictionaries or a generator
of JSON-encoded documents. Prefer encoded output for performance reasons
:return: Generator of labeled sentences
"""
def worker(item):
if input_encoded:
item = json.loads(item)
labeled = self.label_sentence(item, normalize_numerical,
score_type, core_weight)
if labeled:
return json.dumps(labeled) if output_encoded else labeled
for each in parallel.map(worker, sentences, processes):
yield each
def lu_count(sentences, processes=0, input_encoded=False):
""" Count how many sentences per LU there are for each source
:param iterable sentences: Corpus with the POS-tagged sentences
:param int processes: how many processes to use for parallel execution
:param bool input_encoded: whether the corpus is an iterable of dictionaries
or an iterable of JSON-input_encoded documents. JSON-input_encoded
documents are preferable over large size dictionaries for performance reasons
:return: A dictionary source -> frequencies, where frequencies is
another dictionary lemma -> count
:type: bool
"""
def worker(batch):
freqs = defaultdict(lambda: 0)
for row in batch:
sentence = json.loads(row) if input_encoded else row
parsed = urlparse(sentence['url'])
if not parsed.netloc:
logger.warn('cannot parse URL: %s', sentence['url'])
return
lu = sentence['lu']
freqs[(parsed.netloc, lu)] += 1
return freqs.items()
frequencies = defaultdict(lambda: defaultdict(lambda: 0))
for (source, lemma), count in parallel.map(worker,
sentences,
processes,
batch_size=100,
flatten=True):
frequencies[source][lemma] += count
return frequencies
def main(corpus, verbs, processes, outfile, sub_sentences):
""" Compute the LU distribution in the corpus, i.e. how many LUs per sentence
"""
global splitter, tagger, parser, all_verbs
splitter = PunktSentenceSplitter("en")
tagger = TTPosTagger("en")
parser = StanfordParser(
path_to_jar="dev/stanford-corenlp-3.6.0.jar",
path_to_models_jar="dev/stanford-corenlp-3.6.0-models.jar",
java_options=" -mx1G -Djava.ext.dirs=dev/",
) # no way to make classpath work
all_verbs = reduce(lambda x, y: x.union(y), imap(set, json.load(verbs).values()), set())
all_verbs.discard("be")
all_verbs.discard("have")
args = load_corpus(corpus, "bio", text_only=True)
worker = worker_with_sub_sentences if sub_sentences else worker_with_sentences
counter = defaultdict(int)
for i, counts in enumerate(parallel.map(worker, args, processes)):
for k, v in counts.iteritems():
counter[k] += v
if (i + 1) % 10000 == 0:
logger.info("Processed %d documents", i + 1)
counter = OrderedDict(sorted(counter.items(), key=lambda (k, v): k))
for k, v in counter.iteritems():
print k, v
json.dump(counter, outfile, indent=2)
def test_batches(self):
def consumer(bulk):
self.assertEqual(len(bulk), batch_size)
return True
for batch_size in range(1, 10, 2):
data = range(batch_size * 5)
self.assertTrue(all(parallel.map(consumer, data, processes=5, batch_size=batch_size)))
def about_sources(corpus, processes, with_bio):
""" Items' sources
"""
def worker(items):
sources = defaultdict(int)
for doc in items:
url = doc.get('url')
if not url:
logger.warn('found an item without URL, name: %s, bio: %s',
doc.get('name'),
doc.get('bio', '')[:100] + ' ...')
sources['_skipped_'] += 1
continue
elif with_bio and len(doc.get('bio') or '') < 5:
continue
parsed = urlparse(url)
if parsed.netloc:
sources[parsed.netloc] += 1
else:
logger.warn('cannot parse URL: %s', url)
sources['_skipped_'] += 1
return sources
aggregated_sources = defaultdict(int)
corpus = parallel.make_batches(load_scraped_items(corpus), 1000)
for sources in parallel.map(worker, corpus, processes):
for k, v in sources.iteritems():
aggregated_sources[k] += v
aggregated_sources = sorted(aggregated_sources.items(),
key=lambda (_, v): v,
reverse=True)
for source, count in aggregated_sources:
print source, count
try:
import matplotlib.pyplot as plt
except ImportError:
logger.warn('Cannot import matplotlib, skipping chart')
return
count = sum(c for s, c in aggregated_sources)
display_sources = filter(lambda (s, v): float(v) / count >= 0.01,
aggregated_sources)
sources, values = map(list, zip(*display_sources))
sources.append('Rest')
values.append(count - sum(values))
plt.pie(values, labels=sources)
plt.axis('equal')
plt.show()
def extract_sentences(sentences,
probabilities,
processes=0,
input_encoded=False,
output_encoded=False):
""" Extracts some sentences from the corpus following the given probabilities
:param iterable sentences: Extracted sentences
:param dict probabilities: Conditional probabilities of extracting a sentence containing
a specific LU given the source of the sentence. It is therefore a mapping
source -> probabilities, where probabilities is itself a mapping LU -> probability
:param int processes: how many processes to use for parallel execution
:param bool input_encoded: whether the corpus is an iterable of dictionaries or an
iterable of JSON-encoded documents. JSON-encoded documents are preferable
over large size dictionaries for performance reasons
:param bool output_encoded: whether to return a generator of dictionaries or a generator
of JSON-encoded documents. Prefer encoded output for performance reasons
:return: Generator of sentences
"""
def worker(batch):
for row in batch:
sentence = json.loads(row) if input_encoded else row
parsed = urlparse(sentence['url'])
if not parsed.netloc:
logger.warn('cannot parse URL: %s', sentence['url'])
return
lu = sentence['lu']
p = probabilities[(parsed.netloc, lu)]
if random.random() < p:
yield parsed.netloc, lu, json.dumps(
sentence) if output_encoded else sentence
counts = defaultdict(lambda: 0)
for source, lu, sentence in parallel.map(worker,
sentences,
processes,
batch_size=100,
flatten=True):
counts[(source, lu)] += 1
yield sentence
aggs_lu = defaultdict(lambda: 0)
aggs_source = defaultdict(lambda: 0)
for (source, lu), n in counts.iteritems():
aggs_lu[lu] += n
aggs_source[source] += n
logger.debug('aggregated statistics per LU: %s', aggs_lu)
logger.debug('aggregated statistics per source: %s', aggs_source)
def find_ranking(self, processes=0):
""" Ranks the verbs
:param int processes: How many processes to use for parallel ranking
:return: tuple with average tf-idf and average standard deviation ordered rankings
:rtype: tuple of (OrderedDict, OrderedDict)
"""
tfidf_ranking = {}
stdev_ranking = {}
for lemma, tfidf, stdev in parallel.map(self.score_lemma, self.verbs, processes):
tfidf_ranking[lemma] = tfidf
stdev_ranking[lemma] = stdev
return (OrderedDict(sorted(tfidf_ranking.items(), key=lambda x: x[1], reverse=True)),
OrderedDict(sorted(stdev_ranking.items(), key=lambda x: x[1], reverse=True)))
def main(corpus, document_key, language_code, outfile, processes):
""" Split an input corpus into sentences """
corpus = load_corpus(corpus, document_key, text_only=True)
s = PunktSentenceSplitter(language_code)
logger.info("Starting sentence splitting of the input corpus ...")
def worker((i, text)):
sentences = list(s.split(text))
return json.dumps({i: sentences}) if sentences else None
for sentences in parallel.map(worker, enumerate(corpus), processes):
outfile.write(sentences)
outfile.write('\n')
return 0
def main(corpus, document_key, language_code, outfile, processes):
""" Split an input corpus into sentences """
corpus = load_corpus(corpus, document_key, text_only=True)
s = PunktSentenceSplitter(language_code)
logger.info("Starting sentence splitting of the input corpus ...")
def worker((i, text)):
sentences = list(s.split(text))
return json.dumps({i: sentences}) if sentences else None
for sentences in parallel.map(worker, enumerate(corpus), processes):
outfile.write(sentences)
outfile.write('\n')
return 0
def find_ranking(self, processes=0, bulk_size=10000, normalize=True):
ranking = defaultdict(int)
for score in parallel.map(self.score_from_tokens,
self._bulkenize(self.tags, bulk_size),
processes):
for k, v in score.iteritems():
ranking[k] += v
ranking = OrderedDict(sorted(ranking.items(), key=lambda x: x[1], reverse=True))
if normalize:
max_score = float(ranking[next(iter(ranking))])
for lemma, score in ranking.iteritems():
ranking[lemma] = score / max_score
return ranking
def main(classified, lexical_db, outfile, language, semistructured, processes,
dump_unresolved):
""" Serialize classification results into quickstatements
"""
if semistructured:
url_to_wid = map_url_to_wid(semistructured)
logger.info(
'Used semi-structured dataset to infer %d Wikidata Item IDs',
len(url_to_wid))
else:
url_to_wid = {}
logger.info('TIP: using the semi-structured dataset could help in '
'resolving the Wikidata Item ID of more subjects')
lexical_db = json.load(lexical_db)
count = skipped = 0
serializer = ClassificationSerializer(language, lexical_db, url_to_wid)
for success, item in parallel.map(serializer.to_statements,
classified,
processes=processes,
flatten=True):
if success:
outfile.write(item.encode('utf8'))
outfile.write('\n')
count += 1
else:
skipped += 1
if dump_unresolved:
dump_unresolved.write(json.dumps(item))
dump_unresolved.write('\n')
if count % 1000 == 0 and count > 0:
logger.info('Produced %d statements so far, skipped %d names',
count, skipped)
logger.info('Done, produced %d statements, skipped %d names', count,
skipped)
logger.info("Dataset serialized to '%s'" % outfile.name)
if dump_unresolved:
logger.info("Unresolved entities dumped to '%s'" %
dump_unresolved.name)
def main(classified, lexical_db, outfile, language,
semistructured, processes, dump_unresolved):
""" Serialize classification results into quickstatements
"""
if semistructured:
url_to_wid = map_url_to_wid(semistructured)
logger.info('Used semi-structured dataset to infer %d Wikidata Item IDs',
len(url_to_wid))
else:
url_to_wid = {}
logger.info('TIP: using the semi-structured dataset could help in '
'resolving the Wikidata Item ID of more subjects')
lexical_db = json.load(lexical_db)
count = skipped = 0
serializer = ClassificationSerializer(language, lexical_db, url_to_wid)
for success, item in parallel.map(serializer.to_statements, classified,
processes=processes, flatten=True):
if success:
outfile.write(item.encode('utf8'))
outfile.write('\n')
count += 1
else:
skipped += 1
if dump_unresolved:
dump_unresolved.write(json.dumps(item))
dump_unresolved.write('\n')
if count % 1000 == 0 and count > 0:
logger.info('Produced %d statements so far, skipped %d names', count, skipped)
logger.info('Done, produced %d statements, skipped %d names', count, skipped)
logger.info("Dataset serialized to '%s'" % outfile.name)
if dump_unresolved:
logger.info("Unresolved entities dumped to '%s'" % dump_unresolved.name)
def extract(self, processes=0):
""" Processes the corpus extracting sentences from each item
and storing them in the item itself.
:param int processes: how many processes to use for parallel tagging
:return: the extracted sentences
:type: generator of dicts
"""
self.setup_extractor()
try:
count = 0
for i, (item, extracted) in enumerate(
parallel.map(self.extract_from_item, self.corpus,
processes)):
if not item.get('name') or not item.get('url'):
logger.warn('Skipping item without name or URL')
continue
# assign an unique incremental ID to each sentence
# and store information about the originating document
for each in extracted:
each['id'] = count
each['url'] = item['url']
each['name'] = item['name']
count += 1
yield each
if (i + 1) % 10000 == 0:
logger.info('Processed %d items, extracted %d sentences',
i + 1, count)
logger.info('Done, total sentences extracted: %d', count)
finally:
self.teardown_extractor()
def extract(self, processes=0):
""" Processes the corpus extracting sentences from each item
and storing them in the item itself.
:param int processes: how many processes to use for parallel tagging
:return: the extracted sentences
:type: generator of dicts
"""
self.setup_extractor()
try:
count = 0
for i, (item, extracted) in enumerate(parallel.map(self.extract_from_item,
self.corpus, processes)):
if not item.get('name') or not item.get('url'):
logger.warn('Skipping item without name or URL')
continue
# assign an unique incremental ID to each sentence
# and store information about the originating document
for each in extracted:
each['id'] = count
each['url'] = item['url']
each['name'] = item['name']
count += 1
yield each
if (i + 1) % 10000 == 0:
logger.info('Processed %d items, extracted %d sentences',
i + 1, count)
logger.info('Done, total sentences extracted: %d', count)
finally:
self.teardown_extractor()
def test_exception_single(self):
self.assertRaises(ValueError, self.consume,
parallel.map(self.exc_function, self.list_in, processes=1,
raise_exc=True))
def test_flatten_multi_process(self):
list_out = parallel.map(self.multi_function, self.list_in, processes=2,
flatten=True)
self.assertEqual(Counter(list_out), Counter(self.correct_multi))
def test_more_workers(self):
list_out = set(parallel.map(self.function, self.list_in, processes=20))
self.assertEqual(list_out, self.correct)
def test_with_nones_multi_process(self):
list_out = set(parallel.map(self.function, self.list_in_nones, processes=1))
self.assertEqual(list_out, self.correct_nones)
def test_multi_process(self):
list_out = set(parallel.map(self.function, self.list_in, processes=2))
self.assertEqual(list_out, self.correct)
