def load_swda_corpus_data(swda_directory):
print('Loading SwDA Corpus...')
corpus_reader = CorpusReader(swda_directory)
talks = []
talk_names = []
tags_seen = set()
tag_occurances = {}
for transcript in corpus_reader.iter_transcripts(False):
name = 'sw' + str(transcript.conversation_no)
talk_names.append(name)
conversation_content = []
conversation_tags = []
for utterance in transcript.utterances:
conversation_content.append( utterance.text_words(True) )
tag = utterance.damsl_act_tag()
conversation_tags.append( tag )
if tag not in tags_seen:
tags_seen.add(tag)
tag_occurances[tag] = 1
else:
tag_occurances[tag] += 1
talks.append( (conversation_content, conversation_tags) )
print('\nFound ' + str(len(tags_seen))+ ' different utterance tags.\n')
tag_indices = {tag:i for i, tag in enumerate(sorted(list(tags_seen)))}
for talk in talks:
talk_tags = talk[1]
for i, tag in enumerate(talk_tags):
talk_tags[i] = tag_indices[ tag ]
print('Loaded SwDA Corpus.')
return talks, talk_names, tag_indices, tag_occurances
def __init__(self, data_dir, tokenizer, word2id, task='', seed=42):
self.corpus = CorpusReader(data_dir)
self.data_dir = data_dir
self.tokenizer = tokenizer
self.word2id = word2id
self.task = task
# self.utt_num = 0
# for utt in self.corpus.iter_utterances():
# self.utt_num += 1
self.trans_num = 1155
self.deleted_utterances = 0
self.deleted_tokens = 0
self.in_tokens = 0
# for trans in self.corpus.iter_transcripts():
# self.trans_num += 1
self.da2num = switchboard_da_mapping()
self.stopwords = get_stopwords(data_dir, word2id)
# CAUTION: make sure that for each task the seed is the same s.t. the splits will be the same!
train_ixs, val_ixs = train_test_split(range(self.trans_num), shuffle=True, train_size=0.8, random_state=seed)
val_ixs, test_ixs = train_test_split(val_ixs, shuffle=True, train_size=0.5, random_state=seed)
self.train_ixs, self.val_ixs, self.test_ixs = train_ixs, val_ixs, test_ixs
self.utt_da_pairs = []
prev_da = "%"
for i, utt in enumerate(self.corpus.iter_utterances()):
sentence = self.clean_utt(utt.text)
if not sentence: continue
self.in_tokens += len(sentence)
sentence = self.word2id(sentence)
if len(sentence) == 0:
continue
if not self.utt_acceptable(sentence):
continue
act = utt.damsl_act_tag()
if act == None: act = "%"
if act == "+": act = prev_da
_, swda_name = os.path.split(utt.swda_filename)
swda_name = swda_name[:-4] if swda_name.endswith('.csv') else swda_name
ix = utt.utterance_index
self.utt_da_pairs.append((sentence, act, swda_name, ix))
def __init__(self, data_dir, tokenizer, word2id, task='', seed=42):
self.corpus = CorpusReader(data_dir)
self.data_dir = data_dir
self.tokenizer = tokenizer
self.word2id = word2id
self.task = task
self.utt_num = 0
for utt in self.corpus.iter_utterances():
self.utt_num += 1
self.trans_num = 0
for trans in self.corpus.iter_transcripts():
self.trans_num += 1
self.da2num = switchboard_da_mapping()
# CAUTION: make sure that for each task the seed is the same s.t. the splits will be the same!
train_ixs, val_ixs = train_test_split(range(self.trans_num),
shuffle=True,
train_size=0.8,
random_state=seed)
val_ixs, test_ixs = train_test_split(val_ixs,
shuffle=True,
train_size=0.5,
random_state=seed)
self.train_ixs, self.val_ixs, self.test_ixs = train_ixs, val_ixs, test_ixs
self.utt_da_pairs = []
prev_da = "%"
for i, utt in enumerate(self.corpus.iter_utterances()):
sentence = re.sub(r"([+/\}\[\]]|\{\w)", "", utt.text)
sentence = self.word2id(self.tokenizer(sentence))
act = utt.damsl_act_tag()
if act == None: act = "%"
if act == "+": act = prev_da
_, swda_name = os.path.split(utt.swda_filename)
swda_name = swda_name[:-4] if swda_name.endswith(
'.csv') else swda_name
ix = utt.utterance_index
self.utt_da_pairs.append((sentence, act, swda_name, ix))
def load_swda_corpus_data(swda_directory):
print('Loading SwDA Corpus...')
corpus_reader = CorpusReader(swda_directory)
talks = []
talk_names = []
tags_seen = {}
tag_occurances = {}
num_tags_seen = 0
X=[]
Y=[]
for transcript in corpus_reader.iter_transcripts(False):
name = 'sw' + str(transcript.conversation_no)
talk_names.append(name)
conversation_content = []
conversation_tags = []
for utterance in transcript.utterances:
conversation_content.append( utterance.text_words(True) )
tag = utterance.damsl_act_tag()
conversation_tags.append( tag )
if tag not in tags_seen:
tags_seen[tag] = num_tags_seen
num_tags_seen += 1
tag_occurances[tag] = 1
else:
tag_occurances[tag] += 1
talks.append( (conversation_content, conversation_tags) )
X.append(conversation_content)
Y.append(conversation_tags)
print('\nFound ' + str(len(tags_seen))+ ' different utterance tags.\n')
for talk in talks:
conversation_tags = talk[1]
for i in range(len(conversation_tags)):
conversation_tags[i] = tags_seen[ conversation_tags[i] ]
print('Loaded SwDA Corpus.')
return X,Y,talks, talk_names, tags_seen, tag_occurances
class SwitchboardConverter:
def __init__(self, data_dir, tokenizer, word2id, task='', seed=42):
self.corpus = CorpusReader(data_dir)
self.data_dir = data_dir
self.tokenizer = tokenizer
self.word2id = word2id
self.task = task
# self.utt_num = 0
# for utt in self.corpus.iter_utterances():
# self.utt_num += 1
self.trans_num = 1155
self.deleted_utterances = 0
self.deleted_tokens = 0
self.in_tokens = 0
# for trans in self.corpus.iter_transcripts():
# self.trans_num += 1
self.da2num = switchboard_da_mapping()
self.stopwords = get_stopwords(data_dir, word2id)
# CAUTION: make sure that for each task the seed is the same s.t. the splits will be the same!
train_ixs, val_ixs = train_test_split(range(self.trans_num), shuffle=True, train_size=0.8, random_state=seed)
val_ixs, test_ixs = train_test_split(val_ixs, shuffle=True, train_size=0.5, random_state=seed)
self.train_ixs, self.val_ixs, self.test_ixs = train_ixs, val_ixs, test_ixs
self.utt_da_pairs = []
prev_da = "%"
for i, utt in enumerate(self.corpus.iter_utterances()):
sentence = self.clean_utt(utt.text)
if not sentence: continue
self.in_tokens += len(sentence)
sentence = self.word2id(sentence)
if len(sentence) == 0:
continue
if not self.utt_acceptable(sentence):
continue
act = utt.damsl_act_tag()
if act == None: act = "%"
if act == "+": act = prev_da
_, swda_name = os.path.split(utt.swda_filename)
swda_name = swda_name[:-4] if swda_name.endswith('.csv') else swda_name
ix = utt.utterance_index
self.utt_da_pairs.append((sentence, act, swda_name, ix))
def clean_utt(self, utterance):
utterance, cnt = re.subn(r"([+/\}\[\],\-\(\)#]|\{\w)", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r"<+.*>+", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r"\*\w+", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r">[\s\w'?]+$", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r"\*.+$", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r"\^\w+$", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r"^uh+$", "", utterance) ; self.deleted_tokens += cnt
utterance, cnt = re.subn(r"(uh+)", "", utterance) ; self.deleted_tokens += cnt
ml = re.search("(^\s*\.\s*$)|>", utterance)
if ml:
self.deleted_utterances += 1
return None
utterance = [w.lower() for w in utterance.split(" ") if len(w) > 0 and not re.search("[Uu][Hh]+", w)]
return utterance
def utt_acceptable(self, utt):
# check whether an utterance is acceptable for perturbation
if not utt: return False
stop_cnt = 0
for w in utt:
if w in self.stopwords:
stop_cnt += 1
if (len(utt)-stop_cnt >= 4): # (float(stop_cnt) / float(len(utt))) < 0.999 and
return True
return False
def draw_rand_sent(self):
r = random.randint(0, len(self.utt_da_pairs)-1)
return self.utt_da_pairs[r]
def create_vocab(self):
print("Creating Vocab file for Switchboard")
cnt = Counter()
for utt in self.corpus.iter_utterances():
sentence = re.sub(r"([+/\}\[\]]|\{\w)", "",
utt.text)
sentence = self.tokenizer(sentence)
for w in sentence:
cnt[w] += 1
itos_file = os.path.join(self.data_dir, "itos.txt")
itosf = open(itos_file, "w")
for (word, _) in cnt.most_common(25000):
itosf.write("{}\n".format(word))
#getKeysByValue
def swda_permute(self, sents, amount, speaker_ixs):
if amount == 0:
return []
permutations = [list(range(len(sents)))]
segment_permutations = []
amount = min(amount, factorial(len(sents))-1)
segm_ixs = self.speaker_segment_ixs(speaker_ixs)
segments = list(set(segm_ixs.values()))
for i in range(amount):
while True:
permutation = []
segm_perm = np.random.permutation(len(segments))
segment_permutations.append(segm_perm)
for segm_ix in segm_perm:
utt_ixs = sorted(getKeysByValue(segm_ixs, segm_ix))
permutation = permutation + utt_ixs
if permutation not in permutations:
break
permutations.append(permutation)
return permutations[1:] , segment_permutations #the first one is the original, which was included s.t. won't be generated
def speaker_segment_ixs(self, speaker_ixs):
i = 0
segment_indices = dict()
prev_speaker = speaker_ixs[0]
for j,speaker in enumerate(speaker_ixs):
if speaker != prev_speaker:
prev_speaker = speaker
i += 1
segment_indices[j] = i
return segment_indices
def swda_half_perturb(self, amount, speaker_ixs):
segm_ixs = self.speaker_segment_ixs(speaker_ixs)
segments = list(set(segm_ixs.values()))
segment_permutations = []
permutations = [list(segm_ixs.keys())]
for _ in range(amount):
speaker = random.randint(0,1) # choose one of the speakers
speaker_to_perm = list(filter(lambda x: (x-speaker) % 2 == 0, segments))
speaker_orig = list(filter(lambda x: (x-speaker) % 2 != 0, segments))
#TODO: rename either speaker_ix or speaker_ixs, they are something different, but the names are too close
if len(speaker_to_perm) < 2:
return []
while True:
permuted_speaker_ix = np.random.permutation(speaker_to_perm).tolist()
new_segments = [None]*(len(speaker_orig)+len(permuted_speaker_ix))
if speaker == 0 :
new_segments[::2] = permuted_speaker_ix
new_segments[1::2] = speaker_orig
else:
new_segments[1::2] = permuted_speaker_ix
new_segments[::2] = speaker_orig
segment_permutations.append(new_segments)
permutation = []
for segm_ix in new_segments:
utt_ixs = sorted(getKeysByValue(segm_ixs, segm_ix))
permutation = permutation + utt_ixs
if not permutation in permutations:
permutations.append(permutation)
break
return permutations, segment_permutations
def swda_utterance_insertion(self, speaker_ixs, amounts):
segment_ixs = self.speaker_segment_ixs(speaker_ixs)
segments = list(set(segment_ixs.values()))
segment_permutations = []
permutations = []
i = 0
for _ in range(amounts):
while True: # actually: do ... while permutation not in permutations
i_from = random.randint(0, len(segments)-1)
i_to = random.randint(0, len(segments)-2)
segm_perm = deepcopy(segments)
rem_elem = segments[i_from]
segm_perm = segm_perm[0:i_from] + segm_perm[i_from+1:]
segm_perm = segm_perm[0:i_to] + [rem_elem] + segm_perm[i_to:]
permutation = []
for segm_ix in segm_perm:
utt_ixs = sorted(getKeysByValue(segment_ixs, segm_ix))
permutation = permutation + utt_ixs
if permutation not in permutations:
permutations.append(permutation)
segment_permutations.append(segm_perm)
break
return permutations, segment_permutations
def swda_utterance_sampling(self, utterances, acts, speaker_ixs, amount):
segm_ixs = self.speaker_segment_ixs(speaker_ixs)
segments = list(set(segm_ixs.values()))
permutations = []
for i in range(amount):
(sentence, act, swda_name, ix) = self.draw_rand_sent()
# insert_ix = random.choice(range(len(utterances)))
while(True):
insert_ix = np.random.choice(range(len(utterances)))
utt = utterances[insert_ix]
act_orig = acts[insert_ix]
if self.utt_acceptable(utt) and act != act_orig:
break
permutations.append((sentence, act, swda_name, ix, insert_ix))
return permutations
def convert_dset(self, amounts):
# create distinct train/validation/test files. they'll correspond to the created
# splits from the constructor
train_output_file = os.path.join(self.data_dir, 'train', 'coherency_dset_{}.txt'.format(self.task))
val_output_file = os.path.join(self.data_dir, 'validation', 'coherency_dset_{}.txt'.format(self.task))
test_output_file = os.path.join(self.data_dir, 'test', 'coherency_dset_{}.txt'.format(self.task))
if not os.path.exists(os.path.join(self.data_dir, 'train')):
os.makedirs(os.path.join(self.data_dir, 'train'))
if not os.path.exists(os.path.join(self.data_dir, 'validation')):
os.makedirs(os.path.join(self.data_dir, 'validation'))
if not os.path.exists(os.path.join(self.data_dir, 'test')):
os.makedirs(os.path.join(self.data_dir, 'test'))
trainfile = open(train_output_file, 'w')
valfile = open(val_output_file, 'w')
testfile = open(test_output_file, 'w')
shuffled_path = os.path.join(self.data_dir, "shuffled_{}".format(self.task))
if not os.path.isdir(shuffled_path):
os.mkdir(shuffled_path)
for i,trans in enumerate(tqdm(self.corpus.iter_transcripts(display_progress=False), total=1155)):
utterances = []
acts = []
speaker_ixs = []
prev_act = "%"
for utt in trans.utterances:
sentence = self.clean_utt(utt.text)
sentence = self.word2id(sentence)
# print(sentence, " ## DAs: ", utt.act_tag)
utterances.append(sentence)
act = utt.damsl_act_tag()
if act == None: act = "%"
if act == "+": act = prev_act
acts.append(self.da2num[act])
prev_act = act
if "A" in utt.caller:
speaker_ixs.append(0)
else:
speaker_ixs.append(1)
if self.task == 'up':
permuted_ixs , segment_perms = self.swda_permute(utterances, amounts, speaker_ixs)
elif self.task == 'us':
permuted_ixs = self.swda_utterance_sampling(utterances, acts, speaker_ixs, amounts)
elif self.task == 'hup':
permuted_ixs , segment_perms = self.swda_half_perturb(amounts, speaker_ixs)
elif self.task == 'ui':
permuted_ixs, segment_perms = self.swda_utterance_insertion(speaker_ixs, amounts)
swda_fname = os.path.split(trans.swda_filename)[1]
shuffle_file = os.path.join(shuffled_path, swda_fname) # [:-4]
with open(shuffle_file, "w") as f:
#TODO: analogous to DD, write switchboard name into the file
csv_writer = csv.writer(f)
if self.task == 'us':
for perm in permuted_ixs:
(utt, da, name, ix, insert_ix) = perm
row = [name, ix,insert_ix]
csv_writer.writerow(row)
else:
for perm in segment_perms:
csv_writer.writerow(perm)
if self.task == 'us':
for p in permuted_ixs:
a = " ".join([str(x) for x in acts])
u = str(utterances)
# (sentence, act, swda_name, ix, insert_ix)
insert_sent, insert_da, name, ix, insert_ix = p
insert_da = self.da2num[insert_da]
p_a = deepcopy(acts)
p_a[insert_ix] = insert_da
pa = " ".join([str(x) for x in p_a])
p_u = deepcopy(utterances)
p_u[insert_ix] = insert_sent
if i in self.train_ixs:
trainfile.write("{}|{}|{}|{}|{}\n".format("0",a,u,pa,p_u))
trainfile.write("{}|{}|{}|{}|{}\n".format("1",pa,p_u,a,u))
if i in self.val_ixs:
valfile.write("{}|{}|{}|{}|{}\n".format("0",a,u,pa,p_u))
valfile.write("{}|{}|{}|{}|{}\n".format("1",pa,p_u,a,u))
if i in self.test_ixs:
testfile.write("{}|{}|{}|{}|{}\n".format("0",a,u,pa,p_u))
testfile.write("{}|{}|{}|{}|{}\n".format("1",pa,p_u,a,u))
else:
for p in permuted_ixs:
a = " ".join([str(x) for x in acts])
u = str(utterances)
pa = [acts[i] for i in p]
p_a = " ".join([str(x) for x in pa])
pu = [utterances[i] for i in p]
p_u = str(pu)
if i in self.train_ixs:
trainfile.write("{}|{}|{}|{}|{}\n".format("0",a,u,p_a,p_u))
trainfile.write("{}|{}|{}|{}|{}\n".format("1",p_a,p_u,a,u))
if i in self.val_ixs:
valfile.write("{}|{}|{}|{}|{}\n".format("0",a,u,p_a,p_u))
valfile.write("{}|{}|{}|{}|{}\n".format("1",p_a,p_u,a,u))
if i in self.test_ixs:
testfile.write("{}|{}|{}|{}|{}\n".format("0",a,u,p_a,p_u))
testfile.write("{}|{}|{}|{}|{}\n".format("1",p_a,p_u,a,u))
# similarity_processor = BertSimilarity()
QUESTIONS = ["qy", "qw", "qo", "qr"]
def get_similarity(sentence1, sentence2):
print(sentence1, sentence2)
return 1
# similarity_processor.get_similarity(sentence1, sentence2)
def is_question(utterance):
return utterance.damsl_act_tag() in QUESTIONS
if __name__ == '__main__':
data_dir = "swda/swda"
scan_range = 5
cr = CorpusReader("swda1/swda")
for dialog in cr.iter_transcripts(display_progress=True):
for index, utterance in enumerate(dialog.utterances):
if index < scan_range or index >= len(
dialog.utterances) - scan_range:
continue
if is_question(utterance):
for i in range(-5, 5):
print(
get_similarity(utterance.text,
dialog.utterances[index + i].text))
# pyplot.bar(range(scan_range*2), aggregate)
'ng', 'ny', 'qw^d', 'bd', 'qy^d', 'bf', 'ft', 'ba', 'bh', 'bk', 'fa', 'fc',
'br', 'qh', 'oo', 'b', 'qw', 'qy', 'h', 't3', 'o', 't1', '^h', 'aap', '^q',
'x', 'sd', '^2', 'qo', '^g'
]
convert = {
'+': 'sd',
'fo_o_fw_"_by_bc': 'sd',
'oo_co_cc': 'sd',
'arp_nd': 'no',
'aap_am': 'sd'
}
# assuming SWDA corpus installed in path-to-project/swda
# url of repo is https://github.com/cgpotts/swda
# proprocessor script for this model is in https://github.com/miyamotost/swda
corpus = CorpusReader('swda/swda')
with open('dataset/swda_datset_training.txt',
mode='a') as f1, open('dataset/swda_datset_test.txt',
mode='a') as f2:
for i, trans in enumerate(corpus.iter_transcripts(display_progress=False)):
speakerids = [pad, pad, pad, pad]
utts = [pad, pad, pad, pad]
labels = [pad, pad, pad, pad]
print('iter: {}'.format(i + 1))
#
# speakerid : utt.caller_no
# main_topics : trans.topic_description しばらく"PAD"で対応（無視する）
# pos : utt.act_tag しばらく"PAD"で対応（無視する）
# utt : utt.text
pos = pos_list[k][1]
add_token(language, sentence, word, lemma, pos, i, j, k + 1)
tree = etree.ElementTree(text)
tree.write(out_file,
pretty_print=True,
xml_declaration=True,
encoding='utf-8')
def add_token(language, sentence, token, lemma, pos, i, j, k):
"""
Converts a CONLL-U token to a OPUS-xml 'w'-tag.
"""
word = etree.SubElement(sentence, 'w')
word.text = token
word.set('id', 'w{}.{}.{}'.format(i, j, k))
word.set('lem', lemma)
word.set('tree', pos)
if __name__ == '__main__':
corpus = CorpusReader('swda/swda')
for transcript in corpus.iter_transcripts():
out_file = os.path.splitext(os.path.basename(
transcript.swda_filename))[0]
out_path = 'swda-opus'
os.makedirs(out_path, exist_ok=True)
process_single('en', transcript,
os.path.join(out_path, out_file + '.xml'))
import logging
from swda.swda import CorpusReader
corpus = CorpusReader('swda/swda')
train_set_idx = [
'sw2005', 'sw2006', 'sw2008', 'sw2010', 'sw2012', 'sw2015', 'sw2018',
'sw2019', 'sw2020', 'sw2022', 'sw2024', 'sw2025', 'sw2027', 'sw2028',
'sw2032', 'sw2035', 'sw2038', 'sw2039', 'sw2040', 'sw2041', 'sw2051',
'sw2060', 'sw2061', 'sw2062', 'sw2064', 'sw2065', 'sw2073', 'sw2078',
'sw2079', 'sw2085', 'sw2086', 'sw2090', 'sw2092', 'sw2093', 'sw2094',
'sw2095', 'sw2101', 'sw2102', 'sw2104', 'sw2105', 'sw2107', 'sw2109',
'sw2110', 'sw2111', 'sw2113', 'sw2120', 'sw2122', 'sw2124', 'sw2125',
'sw2130', 'sw2137', 'sw2139', 'sw2145', 'sw2149', 'sw2154', 'sw2155',
'sw2157', 'sw2168', 'sw2171', 'sw2177', 'sw2178', 'sw2180', 'sw2181',
'sw2184', 'sw2185', 'sw2187', 'sw2190', 'sw2191', 'sw2197', 'sw2205',
'sw2220', 'sw2221', 'sw2226', 'sw2227', 'sw2228', 'sw2231', 'sw2232',
'sw2234', 'sw2235', 'sw2237', 'sw2241', 'sw2244', 'sw2247', 'sw2248',
'sw2249', 'sw2252', 'sw2259', 'sw2260', 'sw2262', 'sw2263', 'sw2264',
'sw2265', 'sw2266', 'sw2268', 'sw2275', 'sw2278', 'sw2279', 'sw2283',
'sw2285', 'sw2287', 'sw2290', 'sw2292', 'sw2293', 'sw2295', 'sw2296',
'sw2300', 'sw2301', 'sw2302', 'sw2303', 'sw2304', 'sw2305', 'sw2308',
'sw2309', 'sw2313', 'sw2314', 'sw2316', 'sw2323', 'sw2324', 'sw2325',
'sw2330', 'sw2331', 'sw2334', 'sw2336', 'sw2339', 'sw2342', 'sw2344',
'sw2349', 'sw2353', 'sw2354', 'sw2355', 'sw2362', 'sw2365', 'sw2366',
'sw2368', 'sw2370', 'sw2372', 'sw2376', 'sw2379', 'sw2380', 'sw2382',
'sw2383', 'sw2386', 'sw2387', 'sw2389', 'sw2393', 'sw2397', 'sw2405',
'sw2406', 'sw2407', 'sw2413', 'sw2418', 'sw2421', 'sw2423', 'sw2424',
'sw2426', 'sw2427', 'sw2429', 'sw2431', 'sw2432', 'sw2433', 'sw2435',
'sw2436', 'sw2437', 'sw2439', 'sw2442', 'sw2445', 'sw2446', 'sw2448',
'sw2450', 'sw2451', 'sw2452', 'sw2457', 'sw2460', 'sw2465', 'sw2466',
'sw2467', 'sw2469', 'sw2471', 'sw2472', 'sw2476', 'sw2477', 'sw2478',
def prep_swda():
"""
Put the conversations into a json format that torchtext can read easily.
Each "example" is a conversation comprised of a list of utterances
and a list of dialogue act tags (each the same length)
"""
log.info("Loading SWDA corpus.")
if not os.path.isfile(SWDA_CORPUS_DIR):
with zipfile.ZipFile("swda/swda.zip") as zip_ref:
zip_ref.extractall('data')
corpus = CorpusReader(SWDA_CORPUS_DIR)
corpus = {t.conversation_no: t for t in corpus.iter_transcripts()}
bert_vocab_file = BERT_VOCAB_FILE.format(BERT_MODEL)
if not os.path.isfile(bert_vocab_file):
log.info("Customizing BERT vocab.")
customize_bert_vocab()
log.info("Loading BERT vocab/tokenizer.")
bert_tokenizer = BertTokenizer.from_pretrained(bert_vocab_file,
never_split = BERT_RESERVED_TOKENS + BERT_CUSTOM_TOKENS)
log.info("Getting splits.")
splits_file = SWDA_SPLITS.format('splits')
if os.path.isfile(splits_file): # use existing SWDA splits (for reproducibility purposes)
with open(splits_file) as f:
splits = json.load(f)
else: # save the splits file
splits = gen_splits(list(corpus.keys()))
with open(splits_file, 'w') as f:
json.dump(splits, f)
def words_to_ints(ws):
maxvalue = max(vocab.values())
for w in ws:
if w not in vocab:
maxvalue += 1
vocab[w] = maxvalue
xs = [vocab[x] for x in ws]
return xs
def tag_to_int(tag):
maxvalue = max(tag_vocab.values()) if tag_vocab else -1
if tag not in tag_vocab:
maxvalue += 1
tag_vocab[tag] = maxvalue
return tag_vocab[tag]
def extract_example(transcript):
""" Gets the parts we need from the SWDA utterance object """
tags, tags_ints, utts, utts_ints, utts_ints_bert , utts_ints_nl, utts_ints_bert_nl = [], [], [], [], [], [], []
for utt in transcript.utterances:
# Regex tokenization
words = "[SPKR_{}] ".format(utt.caller) + tokenize(utt.text.lower())
words_nl = remove_laughters(remove_disfluencies(words))
utts.append(words)
utts_ints.append(words_to_ints(words.split()))
utts_ints_nl.append(words_to_ints(words_nl.split()))
# BERT wordpiece tokenization
bert_text = "[CLS] [SPKR_{}] ".format(utt.caller) + utt.text
bert_tokens = bert_tokenizer.tokenize(bert_text) # list of strings
utts_ints_bert.append(bert_tokenizer.convert_tokens_to_ids(bert_tokens))
bert_text_nl = remove_laughters(remove_disfluencies(bert_text))
bert_tokens_nl = bert_tokenizer.tokenize(bert_text_nl)
utts_ints_bert_nl.append(bert_tokenizer.convert_tokens_to_ids(bert_tokens_nl))
# dialogue act tags
tag = damsl_tag_cluster(utt.act_tag)
tags.append(tag)
tags_ints.append(tag_to_int(tag))
return {'id': transcript.conversation_no, 'utts': utts, 'utts_ints': utts_ints,
'utts_ints_bert': utts_ints_bert, 'tags': tags, 'tags_ints': tags_ints,
'utts_ints_bert_nl': utts_ints_bert_nl, 'utts_ints_nl': utts_ints_nl}
log.info("Extracting data and saving splits.")
for split in splits:
data = []
for ex_id in tqdm(splits[split], desc=split):
data.append(extract_example(corpus[ex_id]))
with open(SWDA_SPLITS.format(split), 'w') as f:
json.dump(data, f)
log.info("Vocab size: {}". format(len(vocab)))
with open(SWDA_SPLITS.format("vocab"), 'w') as f:
json.dump(vocab, f)
log.info("Tag vocab size: {}". format(len(tag_vocab)))
with open(SWDA_SPLITS.format("tag_vocab"), 'w') as f:
json.dump(tag_vocab, f)