python - 在 Twitter 情感分析上用 Python 中的 LIME 文本解释器解释我的深度学习模型

Question

我有一个带有情绪标签的推文数据集。我已经对数据进行了预处理并完成了词性标注（全部通过 Python 中的 NLTK）。预处理后的数据如下所示：

预处理推文

预处理训练数据后，使用以下代码准备：

full_text = list(train['content'].values) + list(test['content'].values)
tokenizer = Tokenizer(num_words=20000,lower = True, filters = '')
tokenizer.fit_on_texts(full_text)
train_tokenized = tokenizer.texts_to_sequences(train['content'])
test_tokenized = tokenizer.texts_to_sequences(test['content'])

max_len = 50
X_train = pad_sequences(train_tokenized, maxlen = max_len)
X_test = pad_sequences(test_tokenized, maxlen = max_len)
embed_size = 300
max_features = 20000

def get_coefs(word,*arr):
    return word, np.asarray(arr, dtype='float32')
def get_embed_mat(embedding_path):
    
    embedding_index = dict(get_coefs(*o.strip().split(" ")) for o in open(embedding_path,encoding="utf8"))

    word_index = tokenizer.word_index
    nb_words = min(max_features, len(word_index))
    print(nb_words)
    embedding_matrix = np.zeros((nb_words + 1, embed_size))
    for word, i in word_index.items():
        if i >= max_features:
            continue
        embedding_vector = embedding_index.get(word)
        if embedding_vector is not None:
            embedding_matrix[i] = embedding_vector
        
    return embedding_matrix

深度学习模型是用词嵌入作为层构建的。模型构建代码如下：

def build_model1(lr = 0.0, lr_d = 0.0, units = 0, dr = 0.0):
    inp = Input(shape = (max_len,))
    x = Embedding(20001, embed_size, weights = [embedding_matrix], trainable = False)(inp)
    x1 = SpatialDropout1D(dr)(x)

    x_lstm = Bidirectional(LSTM(units, return_sequences = True))(x1)
    x1 = Conv1D(32, kernel_size=2, padding='valid', kernel_initializer='he_uniform')(x_lstm)
    avg_pool1_lstm1 = GlobalAveragePooling1D()(x1)
    max_pool1_lstm1 = GlobalMaxPooling1D()(x1)
    
    
    x_lstm = Bidirectional(LSTM(units, return_sequences = True))(x1)
    x1 = Conv1D(32, kernel_size=2, padding='valid', kernel_initializer='he_uniform')(x_lstm)
    avg_pool1_lstm = GlobalAveragePooling1D()(x1)
    max_pool1_lstm = GlobalMaxPooling1D()(x1)
    
    
    
    x = concatenate([avg_pool1_lstm1, max_pool1_lstm1,
                    avg_pool1_lstm, max_pool1_lstm])
    #x = BatchNormalization()(x)
    x = Dropout(0.1)(Dense(128,activation='relu') (x))
    x = BatchNormalization()(x)
    x = Dropout(0.1)(Dense(64,activation='relu') (x))
    x = Dense(8, activation = "sigmoid")(x)
    model = Model(inputs = inp, outputs = x)
    model.compile(loss = "binary_crossentropy", optimizer = Adam(lr = lr, decay = lr_d), metrics = ["accuracy"])
    history = model.fit(X_train, y_one_hot, batch_size = 128, epochs = 20, validation_split=0.1, 
                        verbose = 1, callbacks = [check_point, early_stop])
    model = load_model(file_path)
    return model

我想用 LIME 来解释这个模型的预测（如下图所示）。但它不起作用。

模型的石灰文字说明

标签： pythonnlpsentiment-analysislime