Keras: Multiclass CNN Text Classifier predicts the same class for all input data - keras

Question

I am trying to predict one out of 8 classes for some short text data with word embeddings and CNN. The occuring problem is that the CNN Classifier predicts everything in one class (the biggest one of the training data, distribution ~40%). The text data should be preprocesed quite well, because the classification works fine with other classifiers like SVM or NB.
'''
#first, build index mapping words in the embeddings set to their embedding vector
import os
embeddings_index = {}
f = open(os.path.join('', 'embedding_word2vec.txt'), encoding='utf-8')
for line in f:
values = line.split()
words = values[0]
coefs = np.asarray(values[1:])
embeddings_index[word] = coefs
f.close()
#vectorize text samples in 2D Integer Tensor
from tensorflow.python.keras.preprocessing.text import Tokenizer
tokenizer_obj = Tokenizer()
tokenizer_obj.fit_on_texts(stemmed_text)
sequences = tokenizer_obj.texts_to_sequences(stemmed_text)
#pad sequences
from tensorflow.python.keras.preprocessing.sequence import pad_sequences
from keras.utils import to_categorical
max_length = max([len(s.split(',')) for s in total_texts])
print('MaxLength :', max_length)
word_index = tokenizer_obj.word_index
print('Found %s unique Tokens.' % len(word_index))
text_pad = pad_sequences(sequences)
labels = to_categorical(np.asarray(labels))
print('Shape of label tensor:', labels.shape)
print('Shape of Text Tensor: ', text_pad.shape)
embedding_dim = 100
num_words = len(word_index) + 1
#prepare embedding matrix
embedding_matrix = np.zeros((vocab_size, embedding_dim))
for index, word in enumerate(vocab):
embedding_vector = w2v.wv[word]
if embedding_vector is not None:
embedding_matrix[index] = embedding_vector
#train test split
validation_split = 0.2
indices = np.arange(text_pad.shape[0])
np.random.shuffle(indices)
text_pad = text_pad[indices]
labels = labels[indices]
num_validation_samples = int(validation_split * text_pad.shape[0])
x_train = text_pad[:-num_validation_samples]
y_train = labels[:-num_validation_samples]
x_test = text_pad[-num_validation_samples:]
y_test = labels[-num_validation_samples:]
from keras.models import Sequential
from keras.layers.core import Dense, Flatten, Dropout
from keras.layers.convolutional import Conv1D, MaxPooling1D
from keras.layers.embeddings import Embedding
model = Sequential()
model.add(Embedding(vocab_size, embedding_dim, input_length=max_length,
#embeddings_initializer = Constant(embedding_matrix),
weights=[embedding_matrix],
trainable=False))
model.add(Dropout(0.2))
model.add(Conv1D(filters=64, kernel_size=5, activation='relu'))
model.add(MaxPooling1D(pool_size=4))
#model.add(Conv1D(filters=64, kernel_size=5, activation='relu'))
#model.add(MaxPooling1D(pool_size=4))
model.add(Flatten())
model.add(Dense(8, activation='sigmoid'))
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
history = model.fit(x_train, y_train,
epochs=25,
validation_data=(x_test, y_test))
print(model.summary())
'''
I am grateful for every hint.
Thank you.