@@ -7,28 +7,78 @@ def __init__(self):
77 self .TAGS = np .arange (4 )
88 self .TAG_MAPS = np .array ([[0 , 1 ], [2 , 3 ], [2 , 3 ], [0 , 1 ]], dtype = np .int32 )
99 self .tags_count = len (self .TAG_MAPS )
10+ self .dictionary = {}
11+ self .skip_window_left = 0
12+ self .skip_window_right = 1
1013
11- def viterbi (self , emission , A , init_A ):
14+ def viterbi (self , emission , A , init_A , return_score = False ):
1215 """
13- 维特比算法的实现,
14- :param emission: 发射概率矩阵,对应于本模型中的分数矩阵
15- :param A: 转移概率矩阵
16- :return:
16+ 维特比算法的实现,所有输入和返回参数均为numpy数组对象
17+ :param emission: 发射概率矩阵,对应于本模型中的分数矩阵,4*length
18+ :param A: 转移概率矩阵,4*4
19+ :param init_A: 初始转移概率矩阵,4
20+ :param return_score: 是否返回最优路径的分值,默认为False
21+ :return: 最优路径,若return_score为True,返回最优路径及其对应分值
1722 """
1823
19- path = np .array ([[0 ], [1 ]], dtype = np .int32 )
20- print (emission .shape )
21- path_score = np .array ([[init_A [0 ] + emission [0 , 0 ]], [init_A [1 ] + emission [0 , 1 ]]], dtype = np .float32 )
24+ length = emission .shape [1 ]
25+ path = np .ones ([4 , length ], dtype = np .int32 ) * - 1
26+ corr_path = np .zeros ([length ], dtype = np .int32 )
27+ path_score = np .ones ([4 , length ], dtype = np .float64 ) * (np .finfo ('f' ).min / 2 )
28+ path_score [:, 0 ] = init_A + emission [:, 0 ]
2229
23- for line_index in range (1 , emission .shape [0 ]):
24- last_index = path [:, - 1 ]
25- cur_index = self .TAG_MAPS [last_index ] # 当前所有路径的可选的标记矩阵,2x2
26- # 当前所有可能路径的分值
27- cur_res = A [last_index , cur_index ] + emission [line_index , cur_index ] + np .expand_dims (path_score [:, - 1 ], 1 )
28- cur_max_index = np .argmax (cur_res , 1 )
29- path = np .insert (path , [path .shape [1 ]], np .expand_dims (np .choose (cur_max_index , cur_index .T ), 1 ), 1 )
30- path_score = np .insert (path_score , [path_score .shape [1 ]], np .expand_dims (np .choose (cur_max_index , cur_res .T ), 1 ),
31- 1 )
30+ for pos in range (1 , length ):
31+ for t in range (4 ):
32+ for prev in range (4 ):
33+ temp = path_score [prev ][pos - 1 ] + A [prev ][t ] + emission [t ][pos ]
34+ if temp >= path_score [t ][pos ]:
35+ path [t ][pos ] = prev
36+ path_score [t ][pos ] = temp
3237
3338 max_index = np .argmax (path_score [:, - 1 ])
34- return path [max_index , :]
39+ corr_path [length - 1 ] = max_index
40+ for i in range (length - 1 , 0 , - 1 ):
41+ max_index = path [max_index ][i ]
42+ corr_path [i - 1 ] = max_index
43+ if return_score :
44+ return corr_path , path_score [max_index , :]
45+ else :
46+ return corr_path
47+ 48+ def sentence2index (self , sentence ):
49+ index = []
50+ for word in sentence :
51+ if word not in self .dictionary :
52+ index .append (0 )
53+ else :
54+ index .append (self .dictionary [word ])
55+ 56+ return index
57+ 58+ def index2seq (self , indices ):
59+ ext_indices = [1 ] * self .skip_window_left
60+ ext_indices .extend (indices + [2 ] * self .skip_window_right )
61+ seq = []
62+ for index in range (self .skip_window_left , len (ext_indices ) - self .skip_window_right ):
63+ seq .append (ext_indices [index - self .skip_window_left : index + self .skip_window_right + 1 ])
64+ 65+ return seq
66+ 67+ def tags2words (self , sentence , tags ):
68+ words = []
69+ word = ''
70+ for tag_index , tag in enumerate (tags ):
71+ if tag == 0 :
72+ words .append (sentence [tag_index ])
73+ elif tag == 1 :
74+ word = sentence [tag_index ]
75+ elif tag == 2 :
76+ word += sentence [tag_index ]
77+ else :
78+ words .append (word + sentence [tag_index ])
79+ word = ''
80+ # 处理最后一个标记为I的情况
81+ if word != '' :
82+ words .append (word )
83+ 84+ return words
0 commit comments