def load_raw(self):
"""Load for sklearn training"""
labels = [] # string labels
examples = [] # examples as strings
no_labels = [] # docs with no labels
# document id -> label mapping
doc2label = i2b2.parse_standoff(self.annot_xml, self.disease,
self.judgement)
for f in os.listdir(self.corpus_path):
doc_id = f.split('.')[0]
file_path = os.path.join(self.corpus_path, f)
file_feat_list = utils.read_cuis(file_path)
# no labels for some documents for some reason
if doc_id in doc2label:
string_label = doc2label[doc_id]
int_label = LABEL2INT[string_label]
labels.append(int_label)
examples.append(' '.join(file_feat_list))
else:
no_labels.append(doc_id)
print '%d documents with no labels for %s/%s in %s' \
% (len(no_labels), self.disease,
self.judgement, self.annot_xml.split('/')[-1])
return examples, labels
def make_token_alphabet(self):
"""Map tokens (CUIs) to integers"""
# count tokens in the entire corpus
token_counts = collections.Counter()
for f in os.listdir(self.corpus_path):
file_path = os.path.join(self.corpus_path, f)
file_feat_list = None
if self.use_cuis:
file_feat_list = utils.read_cuis(file_path)
else:
file_feat_list = utils.read_tokens(file_path)
token_counts.update(file_feat_list)
# now make alphabet (high freq tokens first)
index = 1
self.token2int['oov_word'] = 0
outfile = open(ALPHABET_FILE, 'w')
for token, count in token_counts.most_common():
if count > self.min_token_freq:
outfile.write('%s|%s\n' % (token, count))
self.token2int[token] = index
index = index + 1
# pickle alphabet
pickle_file = open(self.alphabet_pickle, 'wb')
pickle.dump(self.token2int, pickle_file)
def __init__(self, corpus_path, max_tokens_in_file):
"""Load documents as strings"""
self.samples = []
for f in os.listdir(corpus_path):
file_path = os.path.join(corpus_path, f)
file_feat_list = utils.read_cuis(file_path)
if len(file_feat_list) < max_tokens_in_file:
self.samples.append(' '.join(file_feat_list))
def load(self, maxlen=float('inf'), tokens_as_set=True):
"""Convert examples into lists of indices for keras"""
labels = [] # int labels
examples = [] # examples as int sequences
no_labels = [] # docs with no labels
# document id -> label mapping
doc2label = i2b2.parse_standoff(self.annot_xml, self.disease,
self.judgement)
# load examples and labels
for f in os.listdir(self.corpus_path):
doc_id = f.split('.')[0]
file_path = os.path.join(self.corpus_path, f)
file_ngram_list = None
if self.use_cuis == True:
file_feat_list = utils.read_cuis(file_path)
else:
file_feat_list = utils.read_tokens(file_path)
example = []
if tokens_as_set:
file_feat_list = set(file_feat_list)
for token in file_feat_list:
if token in self.token2int:
example.append(self.token2int[token])
else:
example.append(self.token2int['oov_word'])
if len(example) > maxlen:
example = example[0:maxlen]
# no labels for some documents for some reason
if doc_id in doc2label:
string_label = doc2label[doc_id]
int_label = LABEL2INT[string_label]
labels.append(int_label)
examples.append(example)
else:
no_labels.append(doc_id)
print('%d documents with no labels for %s/%s in %s' \
% (len(no_labels), self.disease,
self.judgement, self.annot_xml.split('/')[-1]))
return examples, labels
def load_vectorized(self, exclude, maxlen=float('inf')):
"""Same as above but labels are vectors"""
labels = [] # int labels
examples = [] # examples as int sequences
no_labels = [] # docs with no labels
# document id -> vector of labels
doc2labels = i2b2.parse_standoff_vectorized(self.annot_xml,
self.judgement, exclude)
# load examples and labels
for f in os.listdir(self.corpus_path):
doc_id = f.split('.')[0]
file_path = os.path.join(self.corpus_path, f)
file_feat_list = utils.read_cuis(file_path)
example = []
# TODO: use unique tokens or not?
for token in set(file_feat_list):
if token in self.token2int:
example.append(self.token2int[token])
else:
example.append(self.token2int['oov_word'])
if len(example) > maxlen:
example = example[0:maxlen]
# no labels for some documents for some reason
if doc_id in doc2labels:
label_vector = doc2labels[doc_id]
labels.append(label_vector)
examples.append(example)
else:
no_labels.append(doc_id)
print '%d documents with no labels for %s/%s in %s' \
% (len(no_labels), self.disease,
self.judgement, self.annot_xml.split('/')[-1])
return examples, labels
