python實現命名實體識別指標（實體級別）

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() # 將labels進行拆分
        labels = ["O" if label =="0" else label for label in labels] # 如果標簽是O就就是O，否則就是label
        # print(labels)
    for num in range(len(labels)): # 遍歷Labels
        if labels[num] == B_label: 
            song_pos0 = num # 記錄B_SONG的位置
        if labels[num] == I_label and labels[num-1] == B_label: # 如果當前lable是I_SONG且前一個是B_SONG
            lenth = 2 # 當前長度為2 
            for num2 in range(num,len(labels)): # 從該位置開始繼續遍歷
                if labels[num2] == I_label and labels[num2-1] == I_label: # 如果當前位置和前一個位置是I_SONG
                    lenth += 1 # 長度+1
                if labels[num2] == "O": # 如果當前標簽是O
                    result.append((song_pos0,lenth)) #z則取得B的位置和長度
                    break # 退出第二個循環
    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 # 將result賦值給就標簽
    return result
res = find_all_tag(pre)

結果：

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

接下來計算精確率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() # 字符串轉換為列表
        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是一個字典，鍵是標簽，值是一個元組，第一位是B的位置，第二位是長度
    for name in pre_result: # 取得鍵，也就是標簽
        for x in pre_result[name]: # 取得值：也就是元組，注意元組可能有多個
            if x: # 如果x存在
                if pre_labels[x[0]:x[0]+x[1]] == true_labels[x[0]:x[0]+x[1]]: # 判斷對應位置的每個標簽是否一致
                    pre.append(1) # 一致則結果添加1
                else:
                    pre.append(0) # 不一致則結果添加0
    return sum(pre)/len(pre) #為1的個數/總個數




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: # 取得鍵，也就是標簽，這里注意和計算precision的區別，遍歷的是真實標簽列表
        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，計算F1就簡單了

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

結果：

0.8

1.0

0.888888888888889

 

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