def _build_vocab_from_db(self, corpus):
try:
f = io.open(corpus, 'r')
for line in f:
elements = line.rstrip().split('\t')[1:]
for ele in elements:
tokens = to_tokens(ele)
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = '_'.join(tokens[i-t:i+1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
except UnicodeDecodeError:
f = open(corpus, 'r')
for line in f:
elements = line.rstrip().split('\t')[1:]
for ele in elements:
tokens = to_tokens(ele)
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = '_'.join(tokens[i-t:i+1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
def _build_vocab_from_corpus(self, corpus):
if not os.path.isfile(corpus): return
try:
f = io.open(corpus, 'r')
for line in f:
tokens = to_tokens(line.rstrip())
for i in range(len(tokens)):
for t in range(self.N):
if i - t < 0: continue
ngram = '_'.join(tokens[i - t:i + 1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
except UnicodeDecodeError:
f = open(corpus, 'r')
for line in f:
tokens = to_tokens(line.rstrip())
for i in range(len(tokens)):
for t in range(self.N):
if i - t < 0: continue
ngram = '_'.join(tokens[i - t:i + 1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
def _build_vocab_from_corpus(self, corpus):
if not os.path.isfile(corpus): return
try:
f = io.open(corpus, 'r')
for line in f:
tokens = to_tokens(line.rstrip())
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = '_'.join(tokens[i-t:i+1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
except UnicodeDecodeError:
f = open(corpus, 'r')
for line in f:
tokens = to_tokens(line.rstrip())
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = '_'.join(tokens[i-t:i+1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
def _build_vocab_from_db(self, corpus):
try:
f = io.open(corpus, 'r')
for line in f:
elements = line.rstrip().split('\t')[1:]
for ele in elements:
tokens = to_tokens(ele)
for i in range(len(tokens)):
for t in range(self.N):
if i - t < 0: continue
ngram = '_'.join(tokens[i - t:i + 1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
except UnicodeDecodeError:
f = open(corpus, 'r')
for line in f:
elements = line.rstrip().split('\t')[1:]
for ele in elements:
tokens = to_tokens(ele)
for i in range(len(tokens)):
for t in range(self.N):
if i - t < 0: continue
ngram = '_'.join(tokens[i - t:i + 1])
if ngram not in self.grams:
self.grams[ngram] = self.n
self.n += 1
f.close()
def _corrupt_value(self, val):
def _is_int(s):
try:
int(s)
return True
except ValueError:
return False
def _is_float(s):
try:
float(s)
return True
except ValueError:
return False
# 中文不用NLTK
# tokens = nltk.word_tokenize(val)
tokens = to_tokens(val)
if len(tokens)>1: tokens.pop(random.randrange(len(tokens)))
out = []
for t in tokens:
if _is_int(t):
out.append(str(int(random.gauss(int(t),0.5))))
elif _is_float(t):
out.append('%.1f' %random.gauss(float(t),0.5))
else:
out.append(t)
return ' '.join([o for o in out])
def _update_state(self, user_utterance, upd=UPD, verbose=False):
prev_act, prev_slot = self.state['prevact'].split('@')
s_t = to_tokens(user_utterance)
slot_match = self._search_slots(s_t) # search slots
val_match = self._search_values(s_t) # search values
for slot, values in val_match.iteritems():
requested = (prev_act=='request') and (prev_slot==slot)
matched = (slot in slot_match)
if not values:
if requested: # asked for value but did not get it
self.state['database'].delete_slot(slot)
self.state['num_requests'][slot] = 1000
self.state['dont_care'].add(slot)
else:
for y, match in values.iteritems():
#y = self.movie_dict.dict[slot].index(val)
if verbose:
print 'Detected %s' %self.movie_dict.dict[slot][y], ' update = ', match
if matched and requested:
alpha = upd*(match + 1. + slot_match[slot])
elif matched and not requested:
alpha = upd*(match + slot_match[slot])
elif not matched and requested:
alpha = upd*(match + 1.)
else:
alpha = upd*match
self.state['inform_slots'][slot][y] += alpha
self.state['slot_tracker'].add(slot)
def _search_values(self, s_t):
'''
在用户输入中查找出现的slot value，这里已经转为了token操作
:param s_t: 用户输入的token化结果
:return: 匹配统计结果
'''
print('-' * 100 + "\nsearching values: ")
for v in s_t:
print(
v.encode("utf8")
if v is not None and type(v) == unicode else v)
matches = {}
for slot in self.state['database'].slots:
matches[slot] = defaultdict(float)
for ss in s_t:
if ss in self.movie_dict.tokens[slot]:
for vi in self.movie_dict.tokens[slot][ss]:
matches[slot][vi] += 1.
for vi, f in matches[slot].iteritems():
val = self.movie_dict.dict[slot][vi]
# 中文版删掉了nltk的东西
# print(nltk.word_tokenize(val))
# matches[slot][vi] = f/len(nltk.word_tokenize(val))
matches[slot][vi] = f / len(to_tokens(val))
print('-' * 100)
return matches
def _build_token_index(self):
self.tokens = {}
for slot,vals in self.dict.iteritems():
self.tokens[slot] = defaultdict(list)
for vi,vv in enumerate(vals):
w_v = to_tokens(vv)
for w in w_v: self.tokens[slot][w].append(vi)
def featurize(self, text):
'''
基于N-Gram的方式构造文本text中的特征
:param text: 自然语言文本
:return: 长度为|Grams|的向量，向量中的每个值表示该Gram的数量
'''
vec = np.zeros((len(self.grams), )).astype('float32')
embeddings = []
# 转到中文之后，N-Gram不再必要了，全部使用Embedding即可
for ngram in to_tokens(text):
if ngram in self.grams:
vec[self.grams[ngram]] += 1.
if ngram in self.embedding_vocab_t2n:
embeddings.append(
self.embedding_vectors[self.embedding_vocab_t2n[ngram]])
else:
embeddings.append(self.UNK_EMBEDDING)
seq_size = len(embeddings)
while len(embeddings) < self.seq_max_len:
embeddings.append(self.BAK_EMBEDDING)
embeddings = embeddings[:self.seq_max_len]
cat_embedding = torch.cat([x.view(1,
x.size()[0]) for x in embeddings], 0)
# TODO: embedding直接改为二维列表,numpy的二维array
return cat_embedding.numpy(), seq_size
def _corrupt_value(self, val):
def _is_int(s):
try:
int(s)
return True
except ValueError:
return False
def _is_float(s):
try:
float(s)
return True
except ValueError:
return False
# 中文不用nltk
# tokens = nltk.word_tokenize(val)
tokens = to_tokens(val)
if len(tokens)>1:
tokens.pop(random.randrange(len(tokens)))
out = set([' '.join(tokens)])
else:
t = tokens[0]
out = set()
if _is_int(t):
pert = round(random.gauss(0,0.5))
if pert>0: out.add('%d' %(int(t)+pert))
out.add(t)
elif _is_float(t):
pert = random.gauss(0,0.5)
if pert>0.05: out.add('%.1f' %(float(t)+pert))
out.add(t)
else:
out.add(t)
return out
def _vocab_search(self, text):
tokens = to_tokens(text)
for i in range(len(tokens)):
for t in range(self.N):
if i - t < 0: continue
ngram = '_'.join(tokens[i - t:i + 1])
if ngram in self.grams:
return True
return False
def _vocab_search(self, text):
tokens = to_tokens(text)
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = '_'.join(tokens[i-t:i+1])
if ngram in self.grams:
return True
return False
def featurize(self, text):
vec = np.zeros((len(self.grams),)).astype('float32')
tokens = to_tokens(text)
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = '_'.join(tokens[i-t:i+1])
if ngram in self.grams:
vec[self.grams[ngram]] += 1.
return vec
def _build_token_index(self):
self.tokens = {}
# tokens的结构:{slot_name:{word_token: [slot value IDs]}}
for slot, vals in self.dict.iteritems():
# print "db slot: {}\nslot values: {}".format(slot.encode("utf-8"), vals)
self.tokens[slot] = defaultdict(list)
for vi, vv in enumerate(vals):
w_v = to_tokens(vv)
for w in w_v:
self.tokens[slot][w].append(vi)
def featurize(self, text):
vec = np.zeros((len(self.grams), )).astype('float32')
tokens = to_tokens(text)
for i in range(len(tokens)):
for t in range(self.N):
if i - t < 0: continue
ngram = '_'.join(tokens[i - t:i + 1])
if ngram in self.grams:
vec[self.grams[ngram]] += 1.
return vec
def _vocab_search(self, text):
'''
按照2-Gram的方式查看是否用户输入中是否存在关键字
:param text: 用户输入的自然语言句子
:return: 是否包含关键字，或者说用户输入的问题是否有意义
'''
tokens = to_tokens(text)
for i in range(len(tokens)):
for t in range(self.N):
if i-t<0: continue
ngram = u''.join(tokens[i-t:i+1])
if ngram in self.grams:
return True
return False
def _build_vocab_from_corpus(self, corpus):
'''
根据对话文本构造全局vocabulary存到grams中，N-Gram实际上包括了全部[1,N]-Grams
:param corpus: 对话文本文件path
:return: None
'''
#转到中文之后，N-Gram不再必要了
if not os.path.isfile(corpus): return
with open(corpus, 'r') as f:
for line in f:
for ngram in to_tokens(line.strip()):
if ngram.strip() != "" and ngram not in self.grams:
self.grams[ngram] = self.n
# print(ngram.encode("utf8") if ngram is not None and type(ngram) == unicode else ngram)
self.n += 1
def _update_state(self, user_utterance, upd=UPD, verbose=False):
'''
根据用户输入，暴力查找用户输入中的关键字(token)，就地改变BT的状态
:param user_utterance: 用户输入
:param upd: 当前训练次数(模型更新次数)
:param verbose: 是否启用唠叨模式
:return: None
'''
prev_act, prev_slot = self.state['prevact'].split('@')
s_t = to_tokens(user_utterance)
slot_match = self._search_slots(s_t) # search slots
val_match = self._search_values(s_t) # search values
for slot, values in val_match.iteritems():
requested = (prev_act == 'request') and (prev_slot == slot)
matched = (slot in slot_match)
if not values:
if requested: # asked for value but did not get it，就不再关心这个slot了！
self.state['database'].delete_slot(slot)
self.state['num_requests'][slot] = 1000
self.state['dont_care'].add(slot)
else:
for y, match in values.iteritems():
# match记录的是支持度这样一个类似的概念
#y = self.movie_dict.dict[slot].index(val)
if verbose:
v = self.movie_dict.dict[slot][y]
print "检测到：{}, 更新置信度：{}".format(
v.encode("utf8") if v is not None
and type(v) == unicode else v, match)
# print 'Detected %s' %v.encode("utf8") if v is not None and type(v) == unicode else v, ' update = ', match
if matched and requested:
alpha = upd * (match + 1. + slot_match[slot])
elif matched and not requested:
alpha = upd * (match + slot_match[slot])
elif not matched and requested:
alpha = upd * (match + 1.)
else:
alpha = upd * match
self.state['inform_slots'][slot][y] += alpha
# inform_slots记录的其实是每个slot的每个value的已知程度，用于手工计算p，同理dont_care用于手工计算q
self.state['slot_tracker'].add(slot)
def _build_vocab_from_db(self, corpus):
'''
根据database构造全局vocabulary存到grams中，N-Gram实际上包括了全部[1,N]-Grams
:param corpus: database文本文件path
:return: None
'''
# 中文中就不需要N-Gram了！
with open(corpus, 'r') as f:
for line in f:
elements = line.strip().split('\t')[1:]
for ele in elements:
# if '·' in ele:
# tokens = ele.split('·')
# else:
tokens = to_tokens(ele)
for ngram in tokens:
if ngram.strip() != "" and ngram not in self.grams:
# print(ngram.encode("utf8") if ngram is not None and type(ngram) == unicode else ngram)
self.grams[ngram] = self.n
self.n += 1
def _prepare_for_search(self):
self.slot_tokens = {}
for slot in self.slots:
self.slot_tokens[slot] = to_tokens(slot)
def _prepare_for_search(self):
self.slot_tokens = {}
for slot in self.slots:
self.slot_tokens[slot] = to_tokens(slot)
