def extract_entity_frequencies(self, tokens):
pos_tagged_tokens = [nltk.pos_tag(t) for t in tokens]
loc = LocationChunker()
trees = batch_ne_chunk(pos_tagged_tokens)
entity_types = ['PERSON', 'ORGANIZATION', 'GPE', 'LOCATION', 'FACILITY']
for entity_type in entity_types:
entity_freq_dict = {}
chunks = [sub_leaves(t, entity_type) for t in trees]
for sent in chunks:
for c in sent:
entity = ' '.join([w[0] for w in c])
entity_freq_dict[entity] = entity_freq_dict.get(entity, 0) + 1
# A secondary attempt at extracting locations based on reference
# to lists of place names
if entity_type == 'LOCATION':
for sent in pos_tagged_tokens:
t = loc.parse(sent)
chunks = sub_leaves(t, 'LOCATION')
for c in chunks:
entity = ' '.join([w[0] for w in c])
entity_freq_dict[entity] = entity_freq_dict.get(entity, 0) + 1
entity_freq_list = [(entity_freq_dict[e], e) for e in entity_freq_dict.keys()]
entity_freq_list.sort(reverse=True)
for e in entity_freq_list:
self.entity_frequencies.append(
entity_type=entity_type.lower(),
entity=e[1],
frequency=e[0])
def entities(docs, strategy='stanford'):
"""
Named entity recognition on
a text document or documents.
Requires that a Stanford NER server is
running on localhost:8080.
Args:
| docs (list) -- the documents to process.
| doc (str) -- the document to process.
| strategy (str) -- the strategy to use, default is stanford.
Returns:
| list -- list of all entity mentions
"""
if type(docs) is str:
docs = [docs]
entities = []
if strategy == 'stanford':
tagger = ner.SocketNER(host='localhost', port=8080)
for doc in docs:
ents = tagger.get_entities(doc)
# We're only interested in the entity names,
# not their tags.
names = [ents[key] for key in ents]
# Flatten the list of lists.
names = [name for sublist in names for name in sublist]
entities += names
# TEMPORARILY REMOVED, THIS PART IS HANDLED EXTERNALLY BY A VECTORIZER.
# Calculate (rough, naive) normalized weights for the entities.
# Will likely want to find ways to recognize congruent entities which
# may not necessarily be consistently mentioned, i.e. "Bill Clinton" and "Clinton" (not yet implemented).
#counts = Counter(entities)
#if len(counts):
#top_count = counts.most_common(1)[0][1]
#results = []
#for entity, count in counts.items():
#results.append((entity, count/top_count))
#return results
return entities
elif strategy == 'nltk':
from nltk.tag import pos_tag
from nltk.chunk import batch_ne_chunk
for doc in docs:
sentences = sent_tokenize(doc)
tokenized_sentences = [word_tokenize(sent) for sent in sentences]
tagged = [pos_tag(sent) for sent in tokenized_sentences]
chunked = batch_ne_chunk(
tagged, binary=True
) # binary=False will tag entities as ORGANIZATION, etc.
for tree in chunked:
entities.extend(_extract_entities(tree))
return entities
else:
raise Exception('Unknown strategy specified.')
def entities(docs, strategy='stanford'):
"""
Named entity recognition on
a text document or documents.
Requires that a Stanford NER server is
running on localhost:8080.
Args:
| docs (list) -- the documents to process.
| doc (str) -- the document to process.
| strategy (str) -- the strategy to use, default is stanford.
Returns:
| list -- list of all entity mentions
"""
if type(docs) is str:
docs = [docs]
entities = []
if strategy == 'stanford':
tagger = ner.SocketNER(host='localhost', port=8080)
for doc in docs:
ents = tagger.get_entities(doc)
# We're only interested in the entity names,
# not their tags.
names = [ents[key] for key in ents]
# Flatten the list of lists.
names = [name for sublist in names for name in sublist]
entities += names
# TEMPORARILY REMOVED, THIS PART IS HANDLED EXTERNALLY BY A VECTORIZER.
# Calculate (rough, naive) normalized weights for the entities.
# Will likely want to find ways to recognize congruent entities which
# may not necessarily be consistently mentioned, i.e. "Bill Clinton" and "Clinton" (not yet implemented).
#counts = Counter(entities)
#if len(counts):
#top_count = counts.most_common(1)[0][1]
#results = []
#for entity, count in counts.items():
#results.append((entity, count/top_count))
#return results
return entities
elif strategy == 'nltk':
from nltk.tag import pos_tag
from nltk.chunk import batch_ne_chunk
for doc in docs:
sentences = sent_tokenize(doc)
tokenized_sentences = [word_tokenize(sent) for sent in sentences]
tagged = [pos_tag(sent) for sent in tokenized_sentences]
chunked = batch_ne_chunk(tagged, binary=True) # binary=False will tag entities as ORGANIZATION, etc.
for tree in chunked:
entities.extend(_extract_entities(tree))
return entities
else:
raise Exception('Unknown strategy specified.')
