Python基于LSI和CNN实现深度智能阅读模型

阳光穿透心脏的1/2处 2024-03-26 23:49 66阅读 0赞

目录  
一、 引言 3  
二、 模型框架 3  
三、 方案介绍 4  
3.1 数据分析与预处理 4  
3.1.1 数据分析 4  
3.1.2 数据预处理 6  
3.2 关键词匹配 9  
3.2.1 词频-逆向文件频率模型 9  
3.2.2 潜在语义索引模型 10  
3.3 精准匹配 11  
3.3.1 卷积神经网络 11  
3.3.2 模型设计 12  
四、 实验结果 14  
4.1 实验环境 14  
4.2 评价指标 14  
4.3 实验结果 15  
4.3.1 数据预处理阶段实验结果 16  
4.3.2 关键词匹配阶段实验结果 16  
4.3.3 精准匹配阶段实验结果 18  
五、 总结与展望 23  
5.1 总结 23  
5.2 展望 23  
5.2.1 改进损失函数——Triplet Loss 23  
5.2.2 构建基于Web的问答系统 24  
5.2.3 研究基于众包的智能阅读数据处理服务 25  
参考文献 27  
附录 28  
（一）FastText模型 28  
（二）CNN with Word2Vec模型 29  
（三）Bi-LSTM模型 31  
（四）基于Attention 模型 32  
二、模型框架  
为了使智能阅读模型能正确理解用户的问题，并跳转到答案所在文本所在行，我们提出了一种基于潜在语义索引(LSI)及卷积神经网络(CNN)的智能阅读模型，该模型主要包括三个部分：数据分析与预处理、关键词匹配以及精准匹配。模型架构图如图1所示。  
![在这里插入图片描述][f9f628d1fcaf4dca84f7d1b310cc379a.png]

图1 智能阅读模型框架  
第一步：数据分析与预处理。我们对问题给出的数据集进行统计分析，提出该数据集进行处理时的关键挑战，并给出相应的预处理步骤；  
第二步：关键词匹配。首先对用户提出的问题进1行分词，并将需要在其中寻找答案的文本构建成问答数据库。进而使用词频-逆向文件词频(TF-IDF)计算出问题以及段落的词频矩阵，再利用基于奇异值分解(SVD)的LSI方法将其转化为奇异矩阵，计算相似度，将相似度较大的若干个可能答案段落作为问题的粗匹配结果；  
第三步：精准匹配。我们在经典的TextCNN模型上进行优化，提出一个新的CNN模型在粗匹配结果上进行二次优化达到精确匹配的目的。在这个过程中首先进行词嵌入，然后分别对问题和粗匹配结果中的回答进行两次卷积核大小为3、4、5的卷积操作，经过最大池化层后，将池化的向量连接起来。并通过使用ReLU 激活函数，防止反向传播过程中的梯度问题（梯度消失和梯度爆炸）以及使用 Batch Normalization批规范化，加速收敛，最终输出排序后的较佳结果。

from keras.models import Model
    from keras.layers import Input, Dense, Activation, Dropout, Flatten, concatenate
    from keras.layers.normalization import BatchNormalization
    from keras.layers.convolutional import Conv1D
    from keras.layers.pooling import MaxPooling1D, GlobalAveragePooling1D
    from keras.layers.embeddings import Embedding
    from keras.layers.recurrent import LSTM
    from keras.layers.wrappers import Bidirectional, TimeDistributed
    
    ####################
    MAX_SEQUENCE_LENGTH = 200  # 问题/答案 上限200个词
    EMBEDDING_DIM = 100  # 100d 词向量
    
    QA_EMBED_SIZE = 64
    DROPOUT_RATE = 0.5
    ####################
    
    
    import pickle
    
    token_path = 'data/tokenizer.pkl'
    tokenizer = pickle.load(open(token_path, 'rb'))
    word_index = tokenizer.word_index
    
    
    def embedding_raw():
        return Embedding(len(word_index) + 1,
                         EMBEDDING_DIM,
                         input_length=MAX_SEQUENCE_LENGTH)
    
    
    def embedding():
        import numpy as np
        embedding_matrix = np.load('data/embedding_matrix.npy')
        return Embedding(len(word_index) + 1,
                         EMBEDDING_DIM,
                         weights=[embedding_matrix],
                         input_length=MAX_SEQUENCE_LENGTH,
                         trainable=False)
    
    
    def FastText():
        input_q = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        q = embedding_raw()(input_q)
        q = GlobalAveragePooling1D()(q)
        q = Dropout(DROPOUT_RATE)(q)
    
        input_a = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        a = embedding_raw()(input_a)
        a = GlobalAveragePooling1D()(a)
        a = Dropout(DROPOUT_RATE)(a)
    
        merged = concatenate([q, a])
        merged = Dense(64)(merged)
        merged = BatchNormalization()(merged)
        merged = Activation('relu')(merged)
        merged = Dropout(DROPOUT_RATE)(merged)
        merged = Dense(1, activation="sigmoid")(merged)
    
        model = Model([input_q, input_a], [merged])
        return model
    
    
    def CNN1():
        ##### Q
        input_q = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        q = embedding_raw()(input_q)
        # cnn1模块，kernel_size = 3
        conv1_1 = Conv1D(256, 3, padding='same')(q)
        bn1_1 = BatchNormalization()(conv1_1)
        relu1_1 = Activation('relu')(bn1_1)
        conv1_2 = Conv1D(128, 3, padding='same')(relu1_1)
        bn1_2 = BatchNormalization()(conv1_2)
        relu1_2 = Activation('relu')(bn1_2)
        cnn1 = MaxPooling1D(pool_size=2)(relu1_2)
        # cnn2模块，kernel_size = 4
        conv2_1 = Conv1D(256, 4, padding='same')(q)
        bn2_1 = BatchNormalization()(conv2_1)
        relu2_1 = Activation('relu')(bn2_1)
        conv2_2 = Conv1D(128, 4, padding='same')(relu2_1)
        bn2_2 = BatchNormalization()(conv2_2)
        relu2_2 = Activation('relu')(bn2_2)
        cnn2 = MaxPooling1D(pool_size=2)(relu2_2)
        # cnn3模块，kernel_size = 5
        conv3_1 = Conv1D(256, 5, padding='same')(q)
        bn3_1 = BatchNormalization()(conv3_1)
        relu3_1 = Activation('relu')(bn3_1)
        conv3_2 = Conv1D(128, 5, padding='same')(relu3_1)
        bn3_2 = BatchNormalization()(conv3_2)
        relu3_2 = Activation('relu')(bn3_2)
        cnn3 = MaxPooling1D(pool_size=2)(relu3_2)
        # 拼接三个模块
        cnn_q = concatenate([cnn1, cnn2, cnn3], axis=-1)
        cnn_q = Flatten()(cnn_q)
        cnn_q = Dropout(DROPOUT_RATE)(cnn_q)
    
        ##### A
        input_a = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        a = embedding_raw()(input_a)
        # cnn1模块，kernel_size = 3
        conv1_1 = Conv1D(256, 3, padding='same')(a)
        bn1_1 = BatchNormalization()(conv1_1)
        relu1_1 = Activation('relu')(bn1_1)
        conv1_2 = Conv1D(128, 3, padding='same')(relu1_1)
        bn1_2 = BatchNormalization()(conv1_2)
        relu1_2 = Activation('relu')(bn1_2)
        cnn1 = MaxPooling1D(pool_size=2)(relu1_2)
        # cnn2模块，kernel_size = 4
        conv2_1 = Conv1D(256, 4, padding='same')(a)
        bn2_1 = BatchNormalization()(conv2_1)
        relu2_1 = Activation('relu')(bn2_1)
        conv2_2 = Conv1D(128, 4, padding='same')(relu2_1)
        bn2_2 = BatchNormalization()(conv2_2)
        relu2_2 = Activation('relu')(bn2_2)
        cnn2 = MaxPooling1D(pool_size=2)(relu2_2)
        # cnn3模块，kernel_size = 5
        conv3_1 = Conv1D(256, 5, padding='same')(a)
        bn3_1 = BatchNormalization()(conv3_1)
        relu3_1 = Activation('relu')(bn3_1)
        conv3_2 = Conv1D(128, 5, padding='same')(relu3_1)
        bn3_2 = BatchNormalization()(conv3_2)
        relu3_2 = Activation('relu')(bn3_2)
        cnn3 = MaxPooling1D(pool_size=2)(relu3_2)
        # 拼接三个模块
        cnn_a = concatenate([cnn1, cnn2, cnn3], axis=-1)
        cnn_a = Flatten()(cnn_a)
        cnn_a = Dropout(DROPOUT_RATE)(cnn_a)
    
        ###### Q-A
        merged = concatenate([cnn_q, cnn_a])
        merged = Dense(512)(merged)
        merged = BatchNormalization()(merged)
        merged = Activation('relu')(merged)
        merged = Dropout(DROPOUT_RATE)(merged)
        merged = Dense(1, activation="sigmoid")(merged)
    
        model = Model([input_q, input_a], [merged])
        return model
    
    
    def CNN2():
        ##### Q
        input_q = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        q = embedding()(input_q)
        # cnn1模块，kernel_size = 3
        conv1_1 = Conv1D(256, 3, padding='same')(q)
        bn1_1 = BatchNormalization()(conv1_1)
        relu1_1 = Activation('relu')(bn1_1)
        conv1_2 = Conv1D(128, 3, padding='same')(relu1_1)
        bn1_2 = BatchNormalization()(conv1_2)
        relu1_2 = Activation('relu')(bn1_2)
        cnn1 = MaxPooling1D(pool_size=2)(relu1_2)
        # cnn2模块，kernel_size = 4
        conv2_1 = Conv1D(256, 4, padding='same')(q)
        bn2_1 = BatchNormalization()(conv2_1)
        relu2_1 = Activation('relu')(bn2_1)
        conv2_2 = Conv1D(128, 4, padding='same')(relu2_1)
        bn2_2 = BatchNormalization()(conv2_2)
        relu2_2 = Activation('relu')(bn2_2)
        cnn2 = MaxPooling1D(pool_size=2)(relu2_2)
        # cnn3模块，kernel_size = 5
        conv3_1 = Conv1D(256, 5, padding='same')(q)
        bn3_1 = BatchNormalization()(conv3_1)
        relu3_1 = Activation('relu')(bn3_1)
        conv3_2 = Conv1D(128, 5, padding='same')(relu3_1)
        bn3_2 = BatchNormalization()(conv3_2)
        relu3_2 = Activation('relu')(bn3_2)
        cnn3 = MaxPooling1D(pool_size=2)(relu3_2)
        # 拼接三个模块
        cnn_q = concatenate([cnn1, cnn2, cnn3], axis=-1)
        cnn_q = Flatten()(cnn_q)
        cnn_q = Dropout(DROPOUT_RATE)(cnn_q)
    
        ##### A
        input_a = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        a = embedding()(input_a)
        # cnn1模块，kernel_size = 3
        conv1_1 = Conv1D(256, 3, padding='same')(a)
        bn1_1 = BatchNormalization()(conv1_1)
        relu1_1 = Activation('relu')(bn1_1)
        conv1_2 = Conv1D(128, 3, padding='same')(relu1_1)
        bn1_2 = BatchNormalization()(conv1_2)
        relu1_2 = Activation('relu')(bn1_2)
        cnn1 = MaxPooling1D(pool_size=2)(relu1_2)
        # cnn2模块，kernel_size = 4
        conv2_1 = Conv1D(256, 4, padding='same')(a)
        bn2_1 = BatchNormalization()(conv2_1)
        relu2_1 = Activation('relu')(bn2_1)
        conv2_2 = Conv1D(128, 4, padding='same')(relu2_1)
        bn2_2 = BatchNormalization()(conv2_2)
        relu2_2 = Activation('relu')(bn2_2)
        cnn2 = MaxPooling1D(pool_size=2)(relu2_2)
        # cnn3模块，kernel_size = 5
        conv3_1 = Conv1D(256, 5, padding='same')(a)
        bn3_1 = BatchNormalization()(conv3_1)
        relu3_1 = Activation('relu')(bn3_1)
        conv3_2 = Conv1D(128, 5, padding='same')(relu3_1)
        bn3_2 = BatchNormalization()(conv3_2)
        relu3_2 = Activation('relu')(bn3_2)
        cnn3 = MaxPooling1D(pool_size=2)(relu3_2)
        # 拼接三个模块
        cnn_a = concatenate([cnn1, cnn2, cnn3], axis=-1)
        cnn_a = Flatten()(cnn_a)
        cnn_a = Dropout(DROPOUT_RATE)(cnn_a)
    
        ###### Q-A
        merged = concatenate([cnn_q, cnn_a])
        merged = Dense(512)(merged)
        merged = BatchNormalization()(merged)
        merged = Activation('relu')(merged)
        merged = Dropout(DROPOUT_RATE)(merged)
        merged = Dense(1, activation="sigmoid")(merged)
    
        model = Model([input_q, input_a], [merged])
        return model
    
    
    def BiLSTM():
        input_q = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        q = embedding()(input_q)
        q = Bidirectional(LSTM(QA_EMBED_SIZE, return_sequences=True, dropout=DROPOUT_RATE, recurrent_dropout=DROPOUT_RATE),
                          merge_mode="sum")(q)
        q = TimeDistributed(Dense(QA_EMBED_SIZE))(q)
        q = Flatten()(q)
        q = Dropout(DROPOUT_RATE)(q)
    
        input_a = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        a = embedding()(input_a)
        a = Bidirectional(LSTM(QA_EMBED_SIZE, return_sequences=True, dropout=DROPOUT_RATE, recurrent_dropout=DROPOUT_RATE),
                          merge_mode="sum")(a)
        a = TimeDistributed(Dense(QA_EMBED_SIZE))(a)
        a = Flatten()(a)
        a = Dropout(DROPOUT_RATE)(a)
    
        merged = concatenate([q, a])
        merged = Dense(512)(merged)
        merged = BatchNormalization()(merged)
        merged = Activation('relu')(merged)
        merged = Dropout(DROPOUT_RATE)(merged)
        merged = Dense(1, activation="sigmoid")(merged)
    
        model = Model([input_q, input_a], [merged])
        return model
    
    
    def Attention():
        # https://github.com/farizrahman4u/seq2seq or https://github.com/datalogue/keras-attention
    
        import seq2seq
        from seq2seq.models import AttentionSeq2Seq
        
        input_q = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        q = embedding()(input_q)
        q = AttentionSeq2Seq(input_dim=EMBEDDING_DIM, input_length=MAX_SEQUENCE_LENGTH,
                             output_length=MAX_SEQUENCE_LENGTH, output_dim=QA_EMBED_SIZE,
                             depth=1)(q)
        q = Flatten()(q)
        q = Dropout(DROPOUT_RATE)(q)
    
        input_a = Input(shape=(MAX_SEQUENCE_LENGTH,), dtype='float64')
        a = embedding()(input_a)
        a = AttentionSeq2Seq(input_dim=EMBEDDING_DIM, input_length=MAX_SEQUENCE_LENGTH,
                             output_length=MAX_SEQUENCE_LENGTH, output_dim=QA_EMBED_SIZE,
                             depth=1)(a)
        a = Flatten()(a)
        a = Dropout(DROPOUT_RATE)(a)
    
        merged = concatenate([q, a])
        merged = Dense(512)(merged)
        merged = BatchNormalization()(merged)
        merged = Activation('relu')(merged)
        merged = Dropout(DROPOUT_RATE)(merged)
        merged = Dense(1, activation="sigmoid")(merged)
    
        model = Model([input_q, input_a], [merged])
        return model

![在这里插入图片描述][c1d5d9b1181940148aac5565dc61e067.png_pic_center]  
![在这里插入图片描述][8bcda9c779d94ad6bc95a409876e8249.png_pic_center]  
![在这里插入图片描述][552c36551626433e95fed770dcdb427b.png_pic_center]  
![在这里插入图片描述][2bd51a264f774f1bacbb86e39504966e.png_pic_center]  
![在这里插入图片描述][0f8d8d821a0b4bf59023b3e83da6ceb1.png_pic_center]  
![在这里插入图片描述][69ba0d348e67401c882c6be7c8ebbfb7.png_pic_center]  
![在这里插入图片描述][7bcd12f2267442279b542ea36934cf9b.png_pic_center]  
![在这里插入图片描述][493e6661c3354f76930233980e9cb5dd.png_pic_center]  
![在这里插入图片描述][0ded5d1f18904ac5b6ac3a4f5ebd1012.png_pic_center]  
![在这里插入图片描述][282c7cf7e9014d34a3aacb1d46c1032f.png_pic_center]  
![在这里插入图片描述][f5f03ef2dd13486bafb8a2439c42a756.png_pic_center]  
![在这里插入图片描述][ddb6aeecd5d14bacb14a465384db8277.png_pic_center]  
![在这里插入图片描述][04f0036e59d84d77b26eaf83308b991e.png_pic_center]  
![在这里插入图片描述][6d3a26d4cef44446af18e0c435e5d767.png_pic_center]  
![在这里插入图片描述][2a2a06b78cde4a71a0c8af4d3301cacf.png_pic_center]  
![在这里插入图片描述][fd82aa21c9c34727b108ed8c3f08e22e.png_pic_center]  
![在这里插入图片描述][42e0392194e04f6d85045aead3a97f96.png_pic_center]  
![在这里插入图片描述][192128a01cf94b5e8002e4cd4f947872.png_pic_center]  
![在这里插入图片描述][1a38587fdd6d42eb8238592b7a92d868.png_pic_center]  
![在这里插入图片描述][69810859b63a452eb7fe802963192504.png_pic_center]  
![在这里插入图片描述][21ef91c5a8ff4b8c815188781db506d4.png_pic_center]  
![在这里插入图片描述][4a56ae5ade8943d89e672f6fa566cf06.png_pic_center]  
![在这里插入图片描述][4dfc3bf9c258400d91de100d3e21dcc9.png_pic_center]  
![在这里插入图片描述][d5de9eeb1c3c42c298a67af6a4c79584.png_pic_center]  
![在这里插入图片描述][c8451f1828e34773a4f89c24945fe8c1.png_pic_center]  
![在这里插入图片描述][ef3e36415cc3469fb44969d7c499f28f.png_pic_center]  
![在这里插入图片描述][e7a19e69cce14e46b2cd74b99494e6bd.png_pic_center]  
![在这里插入图片描述][824745a748674fd8b8259a94b75cfb0a.png_pic_center]  
![在这里插入图片描述][b8ccaf54740f425fbca9de8b6c562f79.png_pic_center]  
![在这里插入图片描述][e7758cfe1dd7430abdc3baeac3de5dfb.png_pic_center]

[f9f628d1fcaf4dca84f7d1b310cc379a.png]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/cf6ff576fe0e4d5ab5b2ec55a45bc3bd.png
[c1d5d9b1181940148aac5565dc61e067.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/3fc877c81df146119b5ce096c3139346.png
[8bcda9c779d94ad6bc95a409876e8249.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/a502ad4153a54d4a964ed7bef2500f97.png
[552c36551626433e95fed770dcdb427b.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/33a8ab89af7b4e738c612a17e922c124.png
[2bd51a264f774f1bacbb86e39504966e.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/08858b3e6bde40bb956cddc56a7eda58.png
[0f8d8d821a0b4bf59023b3e83da6ceb1.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/01d6bc658ec3423c891f8379d859317d.png
[69ba0d348e67401c882c6be7c8ebbfb7.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/cb805260b13f4c7bae26d0a63134c7a0.png
[7bcd12f2267442279b542ea36934cf9b.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/76a6db3b69a34ebcb01d0bff9180e8bd.png
[493e6661c3354f76930233980e9cb5dd.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/a8ca6d5d79c347739c0c74bd4903fc58.png
[0ded5d1f18904ac5b6ac3a4f5ebd1012.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/a268112f9881416eb910bbf8c143ee98.png
[282c7cf7e9014d34a3aacb1d46c1032f.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/44b2814f394444af9a046247ce3d919d.png
[f5f03ef2dd13486bafb8a2439c42a756.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/20c4faa9056748e6a263f3b25ac838ed.png
[ddb6aeecd5d14bacb14a465384db8277.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/66f68ea4dd50424cbd78b2e530a5edce.png
[04f0036e59d84d77b26eaf83308b991e.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/ce51572d5b49445aaba903fe4a206c0d.png
[6d3a26d4cef44446af18e0c435e5d767.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/831f039af68a45e0bfe81b55cc351bc7.png
[2a2a06b78cde4a71a0c8af4d3301cacf.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/70536546929144aa88b3db9ab2b4cc80.png
[fd82aa21c9c34727b108ed8c3f08e22e.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/77027d04a9cf44a097f67a4d08084598.png
[42e0392194e04f6d85045aead3a97f96.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/4291590d3d0547be92285d13b392859d.png
[192128a01cf94b5e8002e4cd4f947872.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/3de6652297b94f088a381ded2dc4fae3.png
[1a38587fdd6d42eb8238592b7a92d868.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/c7c05976a066465484afa98a85fbcc31.png
[69810859b63a452eb7fe802963192504.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/32b6d7cd7f4643cd833b843bf299e205.png
[21ef91c5a8ff4b8c815188781db506d4.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/4aec6f3146084a41869e0f2ff58ac91a.png
[4a56ae5ade8943d89e672f6fa566cf06.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/0f4be6604faf45d6aef5508123c2a9d7.png
[4dfc3bf9c258400d91de100d3e21dcc9.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/b25b2c3bf2d240b0889e0dda40a047d9.png
[d5de9eeb1c3c42c298a67af6a4c79584.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/ca8a41b743c647df9f159364a4699d32.png
[c8451f1828e34773a4f89c24945fe8c1.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/d9426f592aa7473c985b09e05b8f1098.png
[ef3e36415cc3469fb44969d7c499f28f.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/03277db334a94038aa7e389c899376e9.png
[e7a19e69cce14e46b2cd74b99494e6bd.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/2f91664447ca438b9be8ab67f05f371e.png
[824745a748674fd8b8259a94b75cfb0a.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/c3003b6aa5f841bc97bf176c3cdd0770.png
[b8ccaf54740f425fbca9de8b6c562f79.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/18e7f9beacb4473eb650f1035f72bfc2.png
[e7758cfe1dd7430abdc3baeac3de5dfb.png_pic_center]: https://image.dandelioncloud.cn/pgy_files/images/2024/03/11/1983c5427c044fe588ff81a952395093.png