def compute_stanford_ner(source_file,
service_url=STANFORD_CORE_NLP_URL,
excludes=None):
excludes = excludes or ('O', )
assert os.path.isfile(source_file), 'File missing!'
tagger = corenlp.CoreNLPParser(url=service_url,
encoding='utf8',
tagtype='ner')
reader = text_corpus.CompressedFileReader(source_file)
document_index = domain_logic.compile_documents_by_filename(
reader.filenames)
stream = domain_logic.get_document_stream(reader,
'en',
document_index=document_index)
i = 0
ner_data = []
for filename, text, metadata in stream:
print(filename)
document_id = document_index.loc[document_index.filename == filename,
'document_id'].values[0]
ner = recognize_named_entities(tagger, document_id, text, excludes)
ner_data.extend(ner)
i += 1
if i % 10 == 0:
logger.info('Processed {} files...'.format(i))
break
return ner_data
def get_document_stream(source, lang, document_index=None, id_extractor=None):
assert document_index is not None
if 'document_id' not in document_index.columns:
document_index['document_id'] = document_index.index
id_extractor = lambda filename: filename.split('_')[0]
lang_pattern = re.compile("^(\w*)\_" + lang + "([\_\-]corr)?\.txt$")
item_filter = lambda x: lang_pattern.match(
x) # and id_extractor(x) in document_index.index
if isinstance(source, str):
print('Opening archive: {}'.format(source))
reader = text_corpus.CompressedFileReader(source,
pattern=lang_pattern,
itemfilter=item_filter)
else:
reader = source
id_map = {
filename: id_extractor(filename)
for filename in reader.filenames if item_filter(filename)
}
if len(set(document_index.index) - set(id_map.values())) > 0:
logger.warning(
'Treaties not found in archive: ' +
', '.join(list(set(document_index.index) - set(id_map.values()))))
columns = ['signed_year', 'party1', 'party2']
df = document_index[columns]
for filename, text in reader:
document_id = id_map.get(filename, None)
if document_id not in df.index:
continue
metadata = df.loc[document_id].to_dict()
metadata['filename'] = filename
metadata['document_id'] = document_id
metadata['treaty_id'] = document_id
yield filename, document_id, text, metadata
def get_document_stream(source, lang, **kwargs):
id_map = {}
if isinstance(source, str):
# FIXME Use "smart_open" or "open_sesame" library instead
reader = text_corpus.CompressedFileReader(source)
else:
reader = source
lookup = compile_documents_by_filename(
reader.filenames).set_index('filename')
lookup['filename'] = lookup.index
row_id = 0
for filename, text in reader:
metadata = lookup.loc[filename].to_dict()
yield filename, text, metadata
row_id += 1
def get_document_stream(source, lang, **kwargs):
reader = text_corpus.CompressedFileReader(source) if isinstance(
source, str) else source
df_corpus_index = compile_unesco_corpus_index(reader)
#reader.filenames = sorted(list(df_corpus_index[(df_corpus_index.year//10).isin([194, 195, 196, 197, 198])].filename.values))
logger.info('Note! Filter is applied: Only first file for each year.')
reader.filenames = list(
df_corpus_index[df_corpus_index.filename.str.contains('en')].groupby(
'year')['filename'].min().values)
df_corpus_index = df_corpus_index.loc[df_corpus_index.filename.isin(
reader.filenames)].sort_values('filename')
assert len(reader.filenames) == len(df_corpus_index)
n_words_threshold = kwargs.get('n_words_threshold', 10)
n_bytes_trunc_size = kwargs.get('n_bytes_trunc_size', None)
logger.info('INFO: Files having less than {} words will be skipped'.format(
n_words_threshold))
logger.info('INFO: Files greater than {} bytes will be truncated'.format(
n_bytes_trunc_size))
processed_count = 0
empty_count = 0
truncated_count = 0
yielded_count = 0
for filename, text in reader:
processed_count += 1
local_number, lang = split_name(filename)
local_number = int(local_number)
metadata = df_corpus_index.loc[local_number].to_dict()
if metadata['n_words'] < n_words_threshold:
#logger.info('WARNING: Skipping empty file {} '.format(filename))
empty_count += 1
continue
if metadata['lang'] != 'en':
logger.info('WARNING: Skipping file (unknown language) {} '.format(
filename))
continue
if n_bytes_trunc_size is not None and metadata[
'n_bytes'] > n_bytes_trunc_size:
text = text[:n_bytes_trunc_size]
truncated_count += 1
# i = 0
# for segment in chunk_text(text, step=50000):
# basename = str(local_number).zfill(10) + '_' + str(i).zfill(3)
# yield filename, text[:100000], metadata
# i += 1
yield filename, text, metadata
yielded_count += 1
logger.info(
'Corpus read done: {} processed files, {} empty files, {} truncated files, {} files yielded'
.format(processed_count, empty_count, truncated_count, yielded_count))
def generate_textacy_corpus(data_folder,
wti_index,
container,
source_path,
language,
merge_entities,
overwrite=False,
period_group='years_1935-1972',
treaty_filter='',
parties=None,
disabled_pipes=None,
tick=utility.noop,
treaty_sources=None):
for key in container.__dict__:
container.__dict__[key] = None
nlp_args = {'disable': disabled_pipes or []}
container.source_path = source_path
container.language = language
container.textacy_corpus = None
container.prepped_source_path = utility.path_add_suffix(
source_path, '_preprocessed')
if overwrite or not os.path.isfile(container.prepped_source_path):
textacy_utility.preprocess_text(container.source_path,
container.prepped_source_path,
tick=tick)
container.textacy_corpus_path = textacy_utility.generate_corpus_filename(
container.prepped_source_path,
container.language,
nlp_args=nlp_args,
compression=None,
period_group=period_group)
container.nlp = textacy_utility.setup_nlp_language_model(
container.language, **nlp_args)
if overwrite or not os.path.isfile(container.textacy_corpus_path):
logger.info('Working: Computing new corpus ' +
container.textacy_corpus_path + '...')
treaties = wti_index.get_treaties(language=container.language,
period_group=period_group,
treaty_filter=treaty_filter,
parties=parties,
treaty_sources=treaty_sources)
reader = text_corpus.CompressedFileReader(
container.prepped_source_path)
stream = domain_logic.get_document_stream(reader,
container.language,
document_index=treaties)
logger.info('Working: Stream created...')
tick(0, len(treaties))
container.textacy_corpus = textacy_utility.create_textacy_corpus(
stream, container.nlp, tick)
container.textacy_corpus.save(container.textacy_corpus_path)
tick(0)
else:
logger.info('Working: Loading corpus ' +
container.textacy_corpus_path + '...')
tick(1, 2)
container.textacy_corpus = textacy.Corpus.load(
container.nlp, container.textacy_corpus_path)
logger.info('Loaded corpus with {} documents.'.format(
len(container.textacy_corpus)))
tick(0)
if merge_entities:
logger.info('Working: Merging named entities...')
for doc in container.textacy_corpus:
named_entities = textacy.extract.named_entities(doc)
merge_spans(named_entities, doc.spacy_doc)
else:
logger.info('Named entities not merged')
logger.info('Done!')
def generate_textacy_corpus(
domain_logic,
data_folder,
container,
document_index, # data_frame or lambda corpus: corpus_index
source_path,
language,
merge_entities,
overwrite=False,
binary_format=True,
use_compression=True,
disabled_pipes=None,
tick=utility.noop):
for key in container.__dict__:
container.__dict__[key] = None
nlp_args = {'disable': disabled_pipes or []}
store_format = 'binary' if binary_format else 'pickle'
store_extension = 'bin' if binary_format else 'pkl'
store_compression = 'bz2' if use_compression else ''
container.source_path = source_path
container.language = language
container.textacy_corpus = None
container.prepped_source_path = utility.path_add_suffix(
source_path, '_preprocessed')
if not os.path.isfile(container.prepped_source_path):
textacy_utility.preprocess_text(container.source_path,
container.prepped_source_path,
tick=tick)
container.textacy_corpus_path = textacy_utility.generate_corpus_filename(
container.prepped_source_path,
container.language,
nlp_args=nlp_args,
extension=store_extension,
compression=store_compression)
container.nlp = textacy_utility.setup_nlp_language_model(
container.language, **nlp_args)
if overwrite or not os.path.isfile(container.textacy_corpus_path):
logger.info('Working: Computing new corpus ' +
container.textacy_corpus_path + '...')
#FIXME VARYING ASPECTS:
reader = text_corpus.CompressedFileReader(
container.prepped_source_path)
stream = domain_logic.get_document_stream(
reader, container.language, document_index=document_index)
logger.info('Working: Stream created...')
if False:
tick(0, len(reader.filenames))
container.textacy_corpus = textacy_utility.create_textacy_corpus(
stream, container.nlp, tick)
logger.info('storing corpus (this might take some time)...')
textacy_utility.save_corpus(container.textacy_corpus,
container.textacy_corpus_path,
format=store_format)
tick(0)
else:
textacy_utility.create_textacy_corpus_streamed(
stream,
container.nlp,
container.textacy_corpus_path,
format='binary',
tick=utility.noop)
container.textacy_corpus = textacy_utility.load_corpus(
container.textacy_corpus_path, container.nlp, format='binary')
else:
logger.info('Working: Loading corpus ' +
container.textacy_corpus_path + '...')
tick(1, 2)
logger.info('...reading corpus (this might take several minutes)...')
container.textacy_corpus = textacy_utility.load_corpus(
container.textacy_corpus_path, container.nlp, format=store_format)
tick(0)
if merge_entities:
logger.info('Working: Merging named entities...')
try:
for doc in container.textacy_corpus:
named_entities = textacy.extract.named_entities(doc)
textacy.spacier.utils.merge_spans(named_entities,
doc.spacy_doc)
except Exception as ex:
logger.error(ex)
logger.info('NER merge failed')
else:
logger.info('Named entities not merged')
logger.info('Done!')
def create_source_stream(source_path, lang, document_index=None):
reader = text_corpus.CompressedFileReader(source_path)
stream = domain_logic.get_document_stream(reader, lang, document_index=document_index)
return stream
