def preprocess_raw_data(file, tokenize, token2id_dct, **kwargs):
"""
# 处理自有数据函数模板
# file文件数据格式: 句子1\t句子2
# [filter] 过滤
# [segment] 分词
# [build vocab] 构造词典
# [split] train-dev-test
"""
seg_file = file.rsplit('.', 1)[0] + '_seg.txt'
if not os.path.exists(seg_file):
items = utils.file2items(file)
# 过滤
# filter here
print('过滤后数据量', len(items))
# 分词
for i, item in enumerate(items):
item[0] = ' '.join(tokenize(item[0]))
item[1] = ' '.join(tokenize(item[1]))
utils.list2file(seg_file, items)
print('保存分词后数据成功', '数据量', len(items), seg_file)
else:
# 读取分词好的数据
items = utils.file2items(seg_file)
# 划分 不分测试集
train_items, dev_items = utils.split_file(items,
ratio='19:1',
shuffle=True,
seed=1234)
# 构造词典(option)
need_to_rebuild = []
for token2id_name in token2id_dct:
if not token2id_dct[token2id_name]:
print(f'字典{token2id_name} 载入不成功, 将生成并保存')
need_to_rebuild.append(token2id_name)
if need_to_rebuild:
print(f'生成缺失词表文件...{need_to_rebuild}')
for items in [train_items, dev_items]: # 字典只统计train和dev
for item in items:
if 'word2id' in need_to_rebuild:
token2id_dct['word2id'].to_count(item[0].split(' '))
token2id_dct['word2id'].to_count(item[1].split(' '))
if 'word2id' in need_to_rebuild:
token2id_dct['word2id'].rebuild_by_counter(
restrict=['<pad>', '<unk>', '<eos>'],
min_freq=5,
max_vocab_size=30000)
token2id_dct['word2id'].save(f'{curr_dir}/../data/s2s_word2id.dct')
else:
print('使用已有词表文件...')
return train_items, dev_items, None
def preprocess_common_dataset_Toutiao(file, tokenize, token2id_dct, **kwargs):
train_file = f'{curr_dir}/../data/train.toutiao.cls.txt'
dev_file = f'{curr_dir}/../data/valid.toutiao.cls.txt'
test_file = f'{curr_dir}/../data/test.toutiao.cls.txt'
items_lst = []
for file in [train_file, dev_file, test_file]:
seg_file = file.rsplit('.', 1)[0] + '_seg.txt' # 原始文本分词并保存为_seg.txt后缀文件
if not os.path.exists(seg_file):
items = utils.file2items(file, deli='\t')
# 过滤
# filter here
print('过滤后数据量', len(items))
# 分词
for i, item in enumerate(items):
item[0] = ' '.join(tokenize(item[0]))
utils.list2file(seg_file, items)
print('保存分词后数据成功', '数据量', len(items), seg_file)
items_lst.append(items)
else:
# 读取分词好的数据
items_lst.append(utils.file2items(seg_file))
train_items, dev_items, test_items = items_lst
# 构造词典(option)
need_to_rebuild = []
for token2id_name in token2id_dct:
if not token2id_dct[token2id_name]:
print(f'字典{token2id_name} 载入不成功, 将生成并保存')
need_to_rebuild.append(token2id_name)
if need_to_rebuild:
print(f'生成缺失词表文件...{need_to_rebuild}')
for items in [train_items, dev_items]: # 字典只统计train和dev
for item in items:
if 'word2id' in need_to_rebuild:
token2id_dct['word2id'].to_count(item[0].split(' '))
if 'label2id' in need_to_rebuild:
token2id_dct['label2id'].to_count([item[1]])
if 'word2id' in need_to_rebuild:
token2id_dct['word2id'].rebuild_by_counter(restrict=['<pad>', '<unk>'], min_freq=1, max_vocab_size=20000)
token2id_dct['word2id'].save(f'{curr_dir}/../data/toutiao_cls_word2id.dct')
if 'label2id' in need_to_rebuild:
token2id_dct['label2id'].rebuild_by_counter(restrict=['<unk>'])
token2id_dct['label2id'].save(f'{curr_dir}/../data/toutiao_cls_label2id.dct')
else:
print('使用已有词表文件...')
return train_items, dev_items, test_items
def Doubanchange2items(file):
# 转为[multi_src, tgt]格式 按字分
exm_lst = []
sess_lst = utils.file2items(file)
for sess in sess_lst:
sess = [' '.join(s) for s in sess] # 按字分
for i in range(1, len(sess)):
multi_src = '$$$'.join(sess[:i])
tgt = sess[i]
exm_lst.append([multi_src, tgt])
return exm_lst
def Doubanchange2items(file):
# 转为[multi_src, tgt]格式
# 分词
seg_file = file.rsplit('.', 1)[0] + '_seg.txt'
if not os.path.exists(seg_file):
items = utils.file2items(file)
# 分词
for i, item in enumerate(items):
for j in range(len(item)):
items[i][j] = ' '.join(tokenize(items[i][j]))
utils.list2file(seg_file, items)
print('保存分词后数据成功', '数据量', len(items), seg_file)
else:
items = utils.file2items(seg_file)
exm_lst = []
sess_lst = items
for sess in sess_lst:
for i in range(1, len(sess)):
multi_src = '$$$'.join(sess[:i])
tgt = sess[i]
exm_lst.append([multi_src, tgt])
return exm_lst
def change2line(file):
exm_lst = []
items = utils.file2items(file, deli=' ')
curr_sent = []
curr_bmeo = []
for item in items:
if len(item) == 1: # 分隔标志 ['']
if curr_sent and curr_bmeo:
exm_lst.append([' '.join(curr_sent), ' '.join(curr_bmeo)])
curr_sent, curr_bmeo = [], []
continue
curr_sent.append(item[0])
curr_bmeo.append(item[1])
if curr_sent and curr_bmeo:
exm_lst.append([' '.join(curr_sent), ' '.join(curr_bmeo)])
return exm_lst
def preprocess_raw_data(file, tokenize, token2id_dct, **kwargs):
"""
# 处理自有数据函数模板
# file文件数据格式: 句子(以空格分好)\t标签(以空格分好)
# [filter] 过滤
# [segment] 分词 ner一般仅分字，用空格隔开，不需分词步骤
# [build vocab] 构造词典
# [split] train-dev-test
"""
items = utils.file2items(file)
# 过滤
# filter here
print('过滤后数据量', len(items))
# 划分
train_items, dev_items, test_items = utils.split_file(items, ratio='18:1:1', shuffle=True, seed=1234)
# 构造词典(option)
need_to_rebuild = []
for token2id_name in token2id_dct:
if not token2id_dct[token2id_name]:
print(f'字典{token2id_name} 载入不成功, 将生成并保存')
need_to_rebuild.append(token2id_name)
if need_to_rebuild:
print(f'生成缺失词表文件...{need_to_rebuild}')
for items in [train_items, dev_items]: # 字典只统计train和dev
for item in items:
if 'char2id' in need_to_rebuild:
token2id_dct['char2id'].to_count(item[0].split(' '))
if 'bmeo2id' in need_to_rebuild:
token2id_dct['bmeo2id'].to_count(item[1].split(' '))
if 'char2id' in need_to_rebuild:
token2id_dct['char2id'].rebuild_by_counter(restrict=['<pad>', '<unk>'], min_freq=1, max_vocab_size=5000)
token2id_dct['char2id'].save(f'{curr_dir}/../data/s2l_char2id.dct')
if 'bmeo2id' in need_to_rebuild:
token2id_dct['bmeo2id'].rebuild_by_counter(restrict=['<pad>', '<unk>'])
token2id_dct['bmeo2id'].save(f'{curr_dir}/../data/s2l_bmeo2id.dct')
else:
print('使用已有词表文件...')
return train_items, dev_items, test_items
def preprocess_raw_data(file, tokenize, token2id_dct, **kwargs):
"""
# 处理自有数据函数模板
# file文件数据格式: 多轮对话句子1\t多轮对话句子2\t...\t多轮对话句子n
# [filter] 过滤
# [segment] 分词
# [build vocab] 构造词典
# [split] train-dev-test
"""
seg_file = file.rsplit('.', 1)[0] + '_seg.txt'
if not os.path.exists(seg_file):
sess_lst = utils.file2items(file)
# 过滤
# filter here
print('过滤后数据量', len(sess_lst))
# 分词
for i, sess in enumerate(sess_lst):
# sess_lst[i] = [' '.join(s) for s in sess] # 按字分
sess_lst[i] = [' '.join(tokenize(s)) for s in sess] # 按词分
utils.list2file(seg_file, sess_lst)
print('保存分词后数据成功', '数据量', len(sess_lst), seg_file)
else:
# 读取分词好的数据
sess_lst = utils.file2items(seg_file)
# 转为多轮格式 multi-turn之间用$$$分隔
items = []
for sess in sess_lst:
for i in range(1, len(sess)):
multi_src = '$$$'.join(sess[:i])
tgt = sess[i]
items.append([multi_src, tgt])
# items: [['w w w$$$w w', 'w w w'],...]
# 划分 不分测试集
train_items, dev_items = utils.split_file(items,
ratio='19:1',
shuffle=True,
seed=1234)
# 构造词典(option) 字词联合
need_to_rebuild = []
for token2id_name in token2id_dct:
if not token2id_dct[token2id_name]:
print(f'字典{token2id_name} 载入不成功, 将生成并保存')
need_to_rebuild.append(token2id_name)
if need_to_rebuild:
print(f'生成缺失词表文件...{need_to_rebuild}')
for items in [train_items, dev_items]: # 字典只统计train和dev
for item in items:
if 'word2id' in need_to_rebuild:
for sent in item[0].split('$$$'):
token2id_dct['word2id'].to_count(sent.split(' '))
token2id_dct['word2id'].to_count(item[1].split(' '))
if 'char2id' in need_to_rebuild:
for sent in item[0].split('$$$'):
token2id_dct['char2id'].to_count(
list(sent.replace(' ', '')))
token2id_dct['char2id'].to_count(
list(item[1].replace(' ', '')))
if 'word2id' in need_to_rebuild:
token2id_dct['word2id'].rebuild_by_counter(
restrict=['<pad>', '<unk>', '<eos>'],
min_freq=1,
max_vocab_size=30000)
token2id_dct['word2id'].save(
f'{curr_dir}/../data/mmch_word2id.dct')
if 'char2id' in need_to_rebuild:
token2id_dct['char2id'].rebuild_by_counter(
restrict=['<pad>', '<unk>', '<eos>'],
min_freq=1,
max_vocab_size=4000)
token2id_dct['char2id'].save(
f'{curr_dir}/../data/mmch_char2id.dct')
else:
print('使用已有词表文件...')
# 负采样
train_items = train_helper.gen_pos_neg_sample(train_items,
sample_idx=1,
num_neg_exm=4)
dev_items = train_helper.gen_pos_neg_sample(dev_items,
sample_idx=1,
num_neg_exm=4)
return train_items, dev_items, None
