def __init__(self):
# 加载语言模型
self.lm = LanguageModel()
# 待求解的拼音输入
self.pinyins = []
# 有向图
self.graph = None
# viterbi递归的缓存
self.viterbi_cache = {}
def __init__(self):
# 加载语言模型
self.lm = LanguageModel()
# 待求解的拼音输入
self.pinyins = []
# 有向图
self.graph = None
# viterbi递归的缓存
self.viterbi_cache = {}
class InputMethod(object):
def __init__(self):
# 加载语言模型
self.lm = LanguageModel()
# 待求解的拼音输入
self.pinyins = []
# 有向图
self.graph = None
# viterbi递归的缓存
self.viterbi_cache = {}
def get_key(self, *words):
return '_'.join([ str(w) for w in words])
def translate(self, pinyins):
'''
@param pinyins: 拼音列表
@return: 汉字串
'''
self.graph = Graph(pinyins, self)
self.viterbi_cache = {}
self.pinyins = pinyins
# 从第一个字开始使用viterbi算法求解最大路径
words = self.graph.sequence[0].keys()
max_node = None
max_score = 0.0
for k in words:
node = self.graph.sequence[0][k]
score = self.viterbi(0, k)
if score > max_score:
max_score = score
max_node = node
# 输出中文路径
result = []
while True:
result.append(max_node.word)
if not max_node.next_node:
break
max_node = max_node.next_node
print (' '.join(pinyins))
return (''.join(result)).decode('utf8')
def viterbi(self, t, k):
'''第 t 个位置出现 k 词的概率
@param t: pinyin数组下标
@param k: 词
@return: 最大分值
'''
if self.get_key(t,k) in self.viterbi_cache:
return self.viterbi_cache[self.get_key(t,k)]
node = self.graph.sequence[t][k]
# 当前词长度
length_self = len(k.decode('utf8'))
# 开始时加载句首词词频作为初始概率
if t == 0:
init_prop = self.lm.get_init_score(k)
else:
init_prop = 1
# 到达结尾
if t == len(self.pinyins)-length_self:
pinyin = '|'.join(self.pinyins[t:t+length_self])
emission_prop = 1/self.lm.emission[pinyin][k]
node.max_score = emission_prop
self.viterbi_cache[self.get_key(t,k)] = node.max_score
return node.max_score
# 获得下一个状态所有可能的词
next_words = self.graph.sequence[t+length_self].keys()
for word in next_words:
# 下一个词长度
length_next = len(word.decode('utf8'))
state_transfer = self.lm.get_trans_pro(word, k)
pinyin = '|'.join(self.pinyins[t+length_self : t+length_self+length_next])
emission_prop = 1/self.lm.emission[pinyin][word]
# 递归调用，直到最后一个拼音结束
score = self.viterbi(t+length_self, word) * state_transfer * emission_prop * init_prop
if score > node.max_score:
node.max_score = score
node.next_node = self.graph.sequence[t+length_self][word]
self.viterbi_cache[self.get_key(t,k)] = node.max_score
return node.max_score
class InputMethod(object):
def __init__(self):
# 加载语言模型
self.lm = LanguageModel()
# 待求解的拼音输入
self.pinyins = []
# 有向图
self.graph = None
# viterbi递归的缓存
self.viterbi_cache = {}
def get_key(self, *words):
return '_'.join([str(w) for w in words])
def translate(self, pinyins):
'''
@param pinyins: 拼音列表
@return: 汉字串
'''
self.graph = Graph(pinyins, self)
self.viterbi_cache = {}
self.pinyins = pinyins
# 从第一个字开始使用viterbi算法求解最大路径
words = self.graph.sequence[0].keys()
max_node = None
max_score = 0.0
for k in words:
node = self.graph.sequence[0][k]
score = self.viterbi(0, k)
if score > max_score:
max_score = score
max_node = node
# 输出中文路径
result = []
while True:
result.append(max_node.word)
if not max_node.next_node:
break
max_node = max_node.next_node
print(' '.join(pinyins))
return (''.join(result)).decode('utf8')
def viterbi(self, t, k):
'''第 t 个位置出现 k 词的概率
@param t: pinyin数组下标
@param k: 词
@return: 最大分值
'''
if self.get_key(t, k) in self.viterbi_cache:
return self.viterbi_cache[self.get_key(t, k)]
node = self.graph.sequence[t][k]
# 当前词长度
length_self = len(k.decode('utf8'))
# 开始时加载句首词词频作为初始概率
if t == 0:
init_prop = self.lm.get_init_score(k)
else:
init_prop = 1
# 到达结尾
if t == len(self.pinyins) - length_self:
pinyin = '|'.join(self.pinyins[t:t + length_self])
emission_prop = 1 / self.lm.emission[pinyin][k]
node.max_score = emission_prop
self.viterbi_cache[self.get_key(t, k)] = node.max_score
return node.max_score
# 获得下一个状态所有可能的词
next_words = self.graph.sequence[t + length_self].keys()
for word in next_words:
# 下一个词长度
length_next = len(word.decode('utf8'))
state_transfer = self.lm.get_trans_pro(word, k)
pinyin = '|'.join(self.pinyins[t + length_self:t + length_self +
length_next])
emission_prop = 1 / self.lm.emission[pinyin][word]
# 递归调用，直到最后一个拼音结束
score = self.viterbi(
t + length_self,
word) * state_transfer * emission_prop * init_prop
if score > node.max_score:
node.max_score = score
node.next_node = self.graph.sequence[t + length_self][word]
self.viterbi_cache[self.get_key(t, k)] = node.max_score
return node.max_score