python实现命名实体识别指标（实体级别）

pre = "0 0 B_SONG I_SONG I_SONG 0 B_SONG I_SONG I_SONG 0 0 B_SINGER I_SINGER I_SINGER 0 O O O B_ALBUM I_ALBUM I_ALBUM O O B_TAG I_TAG I_TAG O"
true = "0 0 B_SONG I_SONG I_SONG 0 0 0 0 0 0 B_SINGER I_SINGER I_SINGER 0 O O O B_ALBUM I_ALBUM I_ALBUM O O B_TAG I_TAG I_TAG O"

tags = [("B_SONG","I_SONG"),("B_SINGER","I_SINGER"),("B_ALBUM","I_ALBUM"),("B_TAG","I_TAG")]


def find_tag(labels,B_label="B_SONG",I_label="I_SONG"):
    result = []
    if isinstance(labels,str): # 如果labels是字符串
        labels = labels.strip().split() # 将labels进行拆分
        labels = ["O" if label =="0" else label for label in labels] # 如果标签是O就就是O，否则就是label
        # print(labels)
    for num in range(len(labels)): # 遍历Labels
        if labels[num] == B_label: 
            song_pos0 = num # 记录B_SONG的位置
        if labels[num] == I_label and labels[num-1] == B_label: # 如果当前lable是I_SONG且前一个是B_SONG
            lenth = 2 # 当前长度为2 
            for num2 in range(num,len(labels)): # 从该位置开始继续遍历
                if labels[num2] == I_label and labels[num2-1] == I_label: # 如果当前位置和前一个位置是I_SONG
                    lenth += 1 # 长度+1
                if labels[num2] == "O": # 如果当前标签是O
                    result.append((song_pos0,lenth)) #z则取得B的位置和长度
                    break # 退出第二个循环
    return result


def find_all_tag(labels):

    result = {}
    for tag in tags:
        res = find_tag(labels,B_label=tag[0],I_label=tag[1])
        result[tag[0].split("_")[1]] = res # 将result赋值给就标签
    return result
res = find_all_tag(pre)

结果：

{'ALBUM': [(18, 3)], 'SINGER': [(11, 3)], 'SONG': [(2, 3), (6, 3)], 'TAG': [(23, 3)]}

接下来计算精确率precision、召回率（查全率）recall、F1：

def precision(pre_labels,true_labels):
    '''
    :param pre_tags: list
    :param true_tags: list
    :return:
    '''
    pre = []
    if isinstance(pre_labels,str):
        pre_labels = pre_labels.strip().split() # 字符串转换为列表
        pre_labels = ["O" if label =="0" else label for label in pre_labels]
    if isinstance(true_labels,str):
        true_labels = true_labels.strip().split()
        true_labels = ["O" if label =="0" else label for label in true_labels]

    pre_result = find_all_tag(pre_labels) # pre_result是一个字典，键是标签，值是一个元组，第一位是B的位置，第二位是长度
    for name in pre_result: # 取得键，也就是标签
        for x in pre_result[name]: # 取得值：也就是元组，注意元组可能有多个
            if x: # 如果x存在
                if pre_labels[x[0]:x[0]+x[1]] == true_labels[x[0]:x[0]+x[1]]: # 判断对应位置的每个标签是否一致
                    pre.append(1) # 一致则结果添加1
                else:
                    pre.append(0) # 不一致则结果添加0
    return sum(pre)/len(pre) #为1的个数/总个数




def recall(pre_labels,true_labels):
    '''
    :param pre_tags: list
    :param true_tags: list
    :return:
    '''
    recall = []
    if isinstance(pre_labels,str):
        pre_labels = pre_labels.strip().split()
        pre_labels = ["O" if label =="0" else label for label in pre_labels]
    if isinstance(true_labels,str):
        true_labels = true_labels.strip().split()
        true_labels = ["O" if label =="0" else label for label in true_labels]

    true_result = find_all_tag(true_labels)
    for name in true_result: # 取得键，也就是标签，这里注意和计算precision的区别，遍历的是真实标签列表
        for x in true_result[name]: # 以下的基本差不多
            if x:
                if pre_labels[x[0]:x[0]+x[1]] == true_labels[x[0]:x[0]+x[1]]:
                    recall.append(1)
                else:
                    recall.append(0)
    return sum(recall)/len(recall)


def f1_score(precision,recall):

    return (2*precision*recall)/(precision+recall) # 有了precision和recall，计算F1就简单了

if __name__ == '__main__':
    precision = precision(pre,true)
    recall = recall(pre,true)
    f1 = f1_score(precision,recall)
    print(precision)
    print(recall)
    print(f1)

结果：

0.8

1.0

0.888888888888889

参考：http://www.manongjc.com/detail/15-ochyrivhdccrvka.html