def create_alphabets(alphabet_directory,
train_path,
data_paths=None,
max_vocabulary_size=100000,
embedd_dict=None,
min_occurrence=0,
normalize_digits=False):
def expand_vocab():
vocab_set = set(vocab_list)
for data_path in data_paths:
# logger.info("Processing data: %s" % data_path)
sentences = parse(open(data_path, 'r').read())
for sentence in sentences:
for word in sentence:
form = word['form']
pos = word['upostag']
type = word['deprel']
real_word = form.split('_BERT_')[0]
for char in real_word:
char_alphabet.add(char)
form = DIGIT_RE.sub("0",
form) if normalize_digits else form
pos_alphabet.add(pos)
type_alphabet.add(type)
if form not in vocab_set and (form in embedd_dict or
form.lower() in embedd_dict):
vocab_set.add(form)
vocab_list.append(form)
logger = get_logger("Create Alphabets")
word_alphabet = Alphabet('word', defualt_value=True, singleton=False)
char_alphabet = Alphabet('character', defualt_value=True)
pos_alphabet = Alphabet('pos')
type_alphabet = Alphabet('type')
if not os.path.isdir(alphabet_directory):
logger.info("Creating Alphabets: %s" % alphabet_directory)
char_alphabet.add(PAD_CHAR)
pos_alphabet.add(PAD_POS)
type_alphabet.add(PAD_TYPE)
char_alphabet.add(ROOT_CHAR)
pos_alphabet.add(ROOT_POS)
type_alphabet.add(ROOT_TYPE)
char_alphabet.add(END_CHAR)
pos_alphabet.add(END_POS)
type_alphabet.add(END_TYPE)
vocab = defaultdict(int)
sentences = parse(open(train_path, 'r').read())
for sentence in sentences:
for word in sentence:
form = word['form']
pos = word['upostag']
type = word['deprel']
real_word = form.split('_BERT_')[0]
for char in real_word:
char_alphabet.add(char)
form = DIGIT_RE.sub("0", form) if normalize_digits else form
vocab[form] += 1
pos_alphabet.add(pos)
type_alphabet.add(type)
# collect singletons
singletons = set(
[word for word, count in vocab.items() if count <= min_occurrence])
# if a singleton is in pretrained embedding dict, set the count to min_occur + c
if embedd_dict is not None:
assert isinstance(embedd_dict, OrderedDict)
for word in vocab.keys():
if word in embedd_dict or word.lower() in embedd_dict:
vocab[word] += min_occurrence
vocab_list = _START_VOCAB + sorted(vocab, key=vocab.get, reverse=True)
logger.info("Total Vocabulary Size: %d" % len(vocab_list))
logger.info("Total Singleton Size: %d" % len(singletons))
vocab_list = [
word for word in vocab_list
if word in _START_VOCAB or vocab[word] > min_occurrence
]
logger.info("Total Vocabulary Size (w.o rare words): %d" %
len(vocab_list))
if len(vocab_list) > max_vocabulary_size:
vocab_list = vocab_list[:max_vocabulary_size]
if data_paths is not None and embedd_dict is not None:
expand_vocab()
for word in vocab_list:
word_alphabet.add(word)
if word in singletons:
word_alphabet.add_singleton(word_alphabet.get_index(word))
word_alphabet.save(alphabet_directory)
char_alphabet.save(alphabet_directory)
pos_alphabet.save(alphabet_directory)
type_alphabet.save(alphabet_directory)
else:
word_alphabet.load(alphabet_directory)
char_alphabet.load(alphabet_directory)
pos_alphabet.load(alphabet_directory)
type_alphabet.load(alphabet_directory)
word_alphabet.close()
char_alphabet.close()
pos_alphabet.close()
type_alphabet.close()
logger.info("Word Alphabet Size (Singleton): %d (%d)" %
(word_alphabet.size(), word_alphabet.singleton_size()))
logger.info("Character Alphabet Size: %d" % char_alphabet.size())
logger.info("POS Alphabet Size: %d" % pos_alphabet.size())
logger.info("Type Alphabet Size: %d" % type_alphabet.size())
return word_alphabet, char_alphabet, pos_alphabet, type_alphabet
def read_bucketed_data(source_path: str,
word_alphabet: Alphabet,
char_alphabet: Alphabet,
pos_alphabet: Alphabet,
type_alphabet: Alphabet,
max_size=None,
normalize_digits=True,
symbolic_root=False,
symbolic_end=False):
data = [[] for _ in _buckets]
max_char_length = [0 for _ in _buckets]
print('Reading data from %s' % source_path)
counter = 0
reader = CoNLLXReader(source_path, word_alphabet, char_alphabet,
pos_alphabet, type_alphabet)
inst = reader.getNext(normalize_digits=normalize_digits,
symbolic_root=symbolic_root,
symbolic_end=symbolic_end)
while inst is not None and (not max_size or counter < max_size):
counter += 1
if counter % 10000 == 0:
print("reading data: %d" % counter)
inst_size = inst.length()
sent = inst.sentence
for bucket_id, bucket_size in enumerate(_buckets):
if inst_size < bucket_size:
data[bucket_id].append([
sent.word_ids, sent.char_id_seqs, inst.pos_ids, inst.heads,
inst.type_ids
])
max_len = max([len(char_seq) for char_seq in sent.char_seqs])
if max_char_length[bucket_id] < max_len:
max_char_length[bucket_id] = max_len
break
inst = reader.getNext(normalize_digits=normalize_digits,
symbolic_root=symbolic_root,
symbolic_end=symbolic_end)
reader.close()
print("Total number of data: %d" % counter)
bucket_sizes = [len(data[b]) for b in range(len(_buckets))]
data_tensors = []
for bucket_id in range(len(_buckets)):
bucket_size = bucket_sizes[bucket_id]
if bucket_size == 0:
data_tensors.append((1, 1))
continue
bucket_length = _buckets[bucket_id]
char_length = min(MAX_CHAR_LENGTH, max_char_length[bucket_id])
wid_inputs = np.empty([bucket_size, bucket_length], dtype=np.int64)
cid_inputs = np.empty([bucket_size, bucket_length, char_length],
dtype=np.int64)
pid_inputs = np.empty([bucket_size, bucket_length], dtype=np.int64)
hid_inputs = np.empty([bucket_size, bucket_length], dtype=np.int64)
tid_inputs = np.empty([bucket_size, bucket_length], dtype=np.int64)
masks = np.zeros([bucket_size, bucket_length], dtype=np.float32)
single = np.zeros([bucket_size, bucket_length], dtype=np.int64)
lengths = np.empty(bucket_size, dtype=np.int64)
for i, inst in enumerate(data[bucket_id]):
wids, cid_seqs, pids, hids, tids = inst
inst_size = len(wids)
lengths[i] = inst_size
# word ids
wid_inputs[i, :inst_size] = wids
wid_inputs[i, inst_size:] = PAD_ID_WORD
for c, cids in enumerate(cid_seqs):
cid_inputs[i, c, :len(cids)] = cids
cid_inputs[i, c, len(cids):] = PAD_ID_CHAR
cid_inputs[i, inst_size:, :] = PAD_ID_CHAR
# pos ids
pid_inputs[i, :inst_size] = pids
pid_inputs[i, inst_size:] = PAD_ID_TAG
# type ids
tid_inputs[i, :inst_size] = tids
tid_inputs[i, inst_size:] = PAD_ID_TAG
# heads
hid_inputs[i, :inst_size] = hids
hid_inputs[i, inst_size:] = PAD_ID_TAG
# masks
masks[i, :inst_size] = 1.0
for j, wid in enumerate(wids):
if word_alphabet.is_singleton(wid):
single[i, j] = 1
words = torch.from_numpy(wid_inputs)
chars = torch.from_numpy(cid_inputs)
pos = torch.from_numpy(pid_inputs)
heads = torch.from_numpy(hid_inputs)
types = torch.from_numpy(tid_inputs)
masks = torch.from_numpy(masks)
single = torch.from_numpy(single)
lengths = torch.from_numpy(lengths)
data_tensor = {
'WORD': words,
'CHAR': chars,
'POS': pos,
'HEAD': heads,
'TYPE': types,
'MASK': masks,
'SINGLE': single,
'LENGTH': lengths
}
data_tensors.append(data_tensor)
return data_tensors, bucket_sizes
def read_data(source_path: str, word_alphabet: Alphabet, char_alphabet: Alphabet,
pos_alphabet: Alphabet, chunk_alphabet: Alphabet, ner_alphabet: Alphabet,
max_size=None, normalize_digits=True):
data = []
max_length = 0
max_char_length = 0
print('Reading data from %s' % source_path)
counter = 0
reader = CoNLL03Reader(source_path, word_alphabet, char_alphabet, pos_alphabet, chunk_alphabet, ner_alphabet)
inst = reader.getNext(normalize_digits)
while inst is not None and (not max_size or counter < max_size):
counter += 1
if counter % 10000 == 0:
print("reading data: %d" % counter)
sent = inst.sentence
data.append([sent.word_ids, sent.char_id_seqs, inst.pos_ids, inst.chunk_ids, inst.ner_ids])
max_len = max([len(char_seq) for char_seq in sent.char_seqs])
if max_char_length < max_len:
max_char_length = max_len
if max_length < inst.length():
max_length = inst.length()
inst = reader.getNext(normalize_digits)
reader.close()
print("Total number of data: %d" % counter)
data_size = len(data)
char_length = min(MAX_CHAR_LENGTH, max_char_length)
wid_inputs = np.empty([data_size, max_length], dtype=np.int64)
cid_inputs = np.empty([data_size, max_length, char_length], dtype=np.int64)
pid_inputs = np.empty([data_size, max_length], dtype=np.int64)
chid_inputs = np.empty([data_size, max_length], dtype=np.int64)
nid_inputs = np.empty([data_size, max_length], dtype=np.int64)
masks = np.zeros([data_size, max_length], dtype=np.float32)
single = np.zeros([data_size, max_length], dtype=np.int64)
lengths = np.empty(data_size, dtype=np.int64)
for i, inst in enumerate(data):
wids, cid_seqs, pids, chids, nids = inst
inst_size = len(wids)
lengths[i] = inst_size
# word ids
wid_inputs[i, :inst_size] = wids
wid_inputs[i, inst_size:] = PAD_ID_WORD
for c, cids in enumerate(cid_seqs):
cid_inputs[i, c, :len(cids)] = cids
cid_inputs[i, c, len(cids):] = PAD_ID_CHAR
cid_inputs[i, inst_size:, :] = PAD_ID_CHAR
# pos ids
pid_inputs[i, :inst_size] = pids
pid_inputs[i, inst_size:] = PAD_ID_TAG
# chunk ids
chid_inputs[i, :inst_size] = chids
chid_inputs[i, inst_size:] = PAD_ID_TAG
# ner ids
nid_inputs[i, :inst_size] = nids
nid_inputs[i, inst_size:] = PAD_ID_TAG
# masks
masks[i, :inst_size] = 1.0
for j, wid in enumerate(wids):
if word_alphabet.is_singleton(wid):
single[i, j] = 1
words = torch.from_numpy(wid_inputs)
chars = torch.from_numpy(cid_inputs)
pos = torch.from_numpy(pid_inputs)
chunks = torch.from_numpy(chid_inputs)
ners = torch.from_numpy(nid_inputs)
masks = torch.from_numpy(masks)
single = torch.from_numpy(single)
lengths = torch.from_numpy(lengths)
data_tensor = {'WORD': words, 'CHAR': chars, 'POS': pos, 'CHUNK': chunks,
'NER': ners, 'MASK': masks, 'SINGLE': single, 'LENGTH': lengths}
return data_tensor, data_size
def create_alphabets(alphabet_directory,
train_path,
data_paths=None,
max_vocabulary_size=100000,
embedd_dict=None,
min_occurrence=1,
normalize_digits=True):
def expand_vocab():
vocab_set = set(vocab_list)
for data_path in data_paths:
# logger.info("Processing data: %s" % data_path)
with open(data_path, 'r') as file:
for line in file:
line = line.strip()
if len(line) == 0 or line.startswith(
'#'): # conllu format. Attardi
continue
tokens = line.split('\t')
if '-' in tokens[0] or '.' in tokens[0]: # conllu. Attardi
continue
for char in tokens[1]:
char_alphabet.add(char)
word = DIGIT_RE.sub(
"0", tokens[1]) if normalize_digits else tokens[1]
pos = tokens[4]
type = tokens[7]
pos_alphabet.add(pos)
type_alphabet.add(type)
if word not in vocab_set and (word in embedd_dict or
word.lower() in embedd_dict):
vocab_set.add(word)
vocab_list.append(word)
logger = get_logger("Create Alphabets")
word_alphabet = Alphabet('word', defualt_value=True, singleton=True)
char_alphabet = Alphabet('character', defualt_value=True)
pos_alphabet = Alphabet('pos')
type_alphabet = Alphabet('type')
if not os.path.isdir(alphabet_directory):
logger.info("Creating Alphabets: %s" % alphabet_directory)
char_alphabet.add(PAD_CHAR)
pos_alphabet.add(PAD_POS)
type_alphabet.add(PAD_TYPE)
char_alphabet.add(ROOT_CHAR)
pos_alphabet.add(ROOT_POS)
type_alphabet.add(ROOT_TYPE)
char_alphabet.add(END_CHAR)
pos_alphabet.add(END_POS)
type_alphabet.add(END_TYPE)
vocab = defaultdict(int) # Attardi
with open(train_path, 'r') as file:
for line in file:
line = line.strip()
if len(line) == 0 or line.startswith('#'): # conllu. Attardi
continue
tokens = line.split('\t')
if '-' in tokens[0] or '.' in tokens[0]: # conllu. Attardi
continue
for char in tokens[1]:
char_alphabet.add(char)
word = DIGIT_RE.sub(
"0", tokens[1]) if normalize_digits else tokens[1]
vocab[word] += 1
pos = tokens[4]
pos_alphabet.add(pos)
type = tokens[7]
type_alphabet.add(type)
# collect singletons
singletons = set(
[word for word, count in vocab.items() if count <= min_occurrence])
# if a singleton is in pretrained embedding dict, set the count to min_occur + c
if embedd_dict is not None:
assert isinstance(embedd_dict, OrderedDict)
for word in vocab.keys():
if word in embedd_dict or word.lower() in embedd_dict:
vocab[word] += min_occurrence
vocab_list = _START_VOCAB + sorted(vocab, key=vocab.get, reverse=True)
logger.info("Total Vocabulary Size: %d" % len(vocab_list))
logger.info("Total Singleton Size: %d" % len(singletons))
vocab_list = [
word for word in vocab_list
if word in _START_VOCAB or vocab[word] > min_occurrence
]
logger.info("Total Vocabulary Size (w.o rare words): %d" %
len(vocab_list))
if len(vocab_list) > max_vocabulary_size:
vocab_list = vocab_list[:max_vocabulary_size]
if data_paths is not None and embedd_dict is not None:
expand_vocab()
for word in vocab_list:
word_alphabet.add(word)
if word in singletons:
word_alphabet.add_singleton(word_alphabet.get_index(word))
word_alphabet.save(alphabet_directory)
char_alphabet.save(alphabet_directory)
pos_alphabet.save(alphabet_directory)
type_alphabet.save(alphabet_directory)
else:
word_alphabet.load(alphabet_directory)
char_alphabet.load(alphabet_directory)
pos_alphabet.load(alphabet_directory)
type_alphabet.load(alphabet_directory)
word_alphabet.close()
char_alphabet.close()
pos_alphabet.close()
type_alphabet.close()
logger.info("Word Alphabet Size (Singleton): %d (%d)" %
(word_alphabet.size(), word_alphabet.singleton_size()))
logger.info("Character Alphabet Size: %d" % char_alphabet.size())
logger.info("POS Alphabet Size: %d" % pos_alphabet.size())
logger.info("Type Alphabet Size: %d" % type_alphabet.size())
return word_alphabet, char_alphabet, pos_alphabet, type_alphabet
def read_data(source_path: str, word_alphabet: Alphabet, char_alphabet: Alphabet, pos_alphabet: Alphabet, type_alphabet: Alphabet,
max_size=None, normalize_digits=True, symbolic_root=False, symbolic_end=False):
data = []
max_length = 0
max_char_length = 0
print('Reading data from %s' % source_path)
counter = 0
data_tensors = []
reader = CoNLLXReader(source_path, word_alphabet, char_alphabet, pos_alphabet, type_alphabet)
inst = reader.getNext(normalize_digits=normalize_digits, symbolic_root=symbolic_root, symbolic_end=symbolic_end)
while inst is not None and (not max_size or counter < max_size):
counter += 1
if counter % 10000 == 0:
print("reading data: %d" % counter)
sent = inst.sentence
data.append([sent.bert_ids, sent.sub_idx, sent.word_ids, sent.char_id_seqs, inst.pos_ids, inst.heads, inst.type_ids])
max_len = max([len(char_seq) for char_seq in sent.char_seqs])
if max_char_length < max_len:
max_char_length = max_len
if max_length < inst.length():
max_length = inst.length()
inst = reader.getNext(normalize_digits=normalize_digits, symbolic_root=symbolic_root, symbolic_end=symbolic_end)
reader.close()
print("Total number of data: %d" % counter)
data_size = len(data)
char_length = min(MAX_CHAR_LENGTH, max_char_length)
bert_wid_inputs = np.empty([data_size, max_length], dtype=np.int64)
subword_word_indicator = np.empty([data_size, max_length], dtype=np.int64)
wid_inputs = np.empty([data_size, max_length], dtype=np.int64)
cid_inputs = np.empty([data_size, max_length, char_length], dtype=np.int64)
pid_inputs = np.empty([data_size, max_length], dtype=np.int64)
hid_inputs = np.empty([data_size, max_length], dtype=np.int64)
tid_inputs = np.empty([data_size, max_length], dtype=np.int64)
masks = np.zeros([data_size, max_length], dtype=np.float32)
single = np.zeros([data_size, max_length], dtype=np.int64)
lengths = np.empty(data_size, dtype=np.int64)
bert_lengths = np.empty(data_size, dtype=np.int64)
for i, inst in enumerate(data):
bert_wids, subword_ids, wids, cid_seqs, pids, hids, tids = inst
inst_size = len(wids)
bpe_st_size = len(bert_wids)
lengths[i] = inst_size
bert_lengths[i] = bpe_st_size
# word ids
wid_inputs[i, :inst_size] = wids
wid_inputs[i, inst_size:] = PAD_ID_WORD
# bert ids
bert_wid_inputs[i, :bpe_st_size] = bert_wids
bert_wid_inputs[i, bpe_st_size:] = 0
# subword
subword_word_indicator[i, :len(subword_ids)] = subword_ids
subword_word_indicator[i, len(subword_ids):] = 0
for c, cids in enumerate(cid_seqs):
cid_inputs[i, c, :len(cids)] = cids
cid_inputs[i, c, len(cids):] = PAD_ID_CHAR
cid_inputs[i, inst_size:, :] = PAD_ID_CHAR
# pos ids
pid_inputs[i, :inst_size] = pids
pid_inputs[i, inst_size:] = PAD_ID_TAG
# type ids
tid_inputs[i, :inst_size] = tids
tid_inputs[i, inst_size:] = PAD_ID_TAG
# heads
hid_inputs[i, :inst_size] = hids
hid_inputs[i, inst_size:] = PAD_ID_TAG
# masks
masks[i, :inst_size] = 1.0
for j, wid in enumerate(wids):
if word_alphabet.is_singleton(wid):
single[i, j] = 1
bert_wid_inputs = torch.from_numpy(bert_wid_inputs)
words = torch.from_numpy(wid_inputs)
chars = torch.from_numpy(cid_inputs)
pos = torch.from_numpy(pid_inputs)
heads = torch.from_numpy(hid_inputs)
types = torch.from_numpy(tid_inputs)
masks = torch.from_numpy(masks)
single = torch.from_numpy(single)
lengths = torch.from_numpy(lengths)
bert_lengths = torch.from_numpy(bert_lengths)
subword_word_indicator = torch.from_numpy(subword_word_indicator)
data_tensor = {'WORD': words, 'CHAR': chars, 'POS': pos, 'HEAD': heads, 'TYPE': types,
'MASK': masks, 'SINGLE': single, 'LENGTH': lengths, 'BERT_LENGTH': bert_lengths,
'BERT_WORD': bert_wid_inputs, 'SUB_IDX': subword_word_indicator}
return data_tensor, data_size
