def caculate_wer(train_data,sentence):
logging.debug('=' * 10 + 'WER' + '=' * 10)
min_score_list = []
for label,group in train_data.groupby(by=['LABEL']):
logging.debug( '-'*20)
logging.debug(u'正在计算类别（%s）,个数:%d'%(label,len(group)))
min_score = 1000
min_reference = None
for reference in group['SEGMENT_EVERYWORD']:
wer_score = wer(reference.split('|'),sentence.split('|'))
if wer_score < min_score:
min_score = wer_score
min_reference = reference
# logging.debug('%f,'%(wer_score)+reference)
logging.debug(u'最小错误率：%f,句子是：%s'%(min_score,min_reference))
min_score_list.append((label,min_score,min_reference))
min_score_list = sorted(min_score_list,key=lambda x:x[1])
most_similary_label = min_score_list[0][0]
most_similary_score = min_score_list[0][1]
most_similary_sentence = min_score_list[0][2]
logging.debug( '-'*20)
logging.debug(u'最近似句子：%s（%s），分数（%f）'%(most_similary_sentence,
most_similary_label,most_similary_score))
return most_similary_label,min_score_list
def caculate_wer(train_data,sentence,k=3):
'''
计算sentence跟训练库中所有句子的相似性，然后使用前k个最相似的实例进行投票
:param train_data:
:param sentence:
:param k:
:param k:
:return:
'''
logging.debug('=' * 10 + 'WER' + '=' * 10)
# 创建一个列表存放每个实例的分数
get_wer_score = lambda x: wer(x.split('|'), sentence.split('|'))
sentence_wer_score_list = train_data[config['train_segment_column']].apply(get_wer_score).as_matrix()
# print sentence_wer_score_list
# 排序
sorted_index = np.argsort(sentence_wer_score_list)
sorted_index = sorted_index[0:k]
# print sorted_index
most_similary_score = sentence_wer_score_list[sorted_index]
most_similary_sentence = train_data['SEGMENT'].iloc[sorted_index].as_matrix()
most_similary_label = train_data['LABEL'].iloc[sorted_index].as_matrix()
logging.debug(u'前k个最相似的句子为：%s'%(' , '.join(most_similary_sentence)))
logging.debug('分数分别为：%s'%most_similary_score)
logging.debug(u'类别分别为：%s'%(' , '.join(most_similary_label)))
return most_similary_score.tolist(),most_similary_sentence.tolist(),most_similary_label.tolist()
