Keras - Seq2Seqで文章校正に挑戦したが同じ文字しか出力されない

Question

前提・実現したいこと

pythonのkerasを使って文章校正のためのseq2seqを組もうとしています。
実装中に以下の問題が発生しました。

発生している問題・エラーメッセージ

以下のコードを実行しても予測結果は全て同じ文字しか出力されません。
storyとnew_storyには校正前文章と校正後文章が500（文書数）×250（１文章内の単語数）で格納されています。

該当のソースコード

import pickle
import numpy as np

from keras.models import Sequential
from keras.layers.core import Dense, Activation, RepeatVector
from keras.layers.recurrent import LSTM
from keras.layers.wrappers import TimeDistributed
from keras.optimizers import Adam
from keras.callbacks import EarlyStopping

with open('word_indices.pickle', 'rb') as f :
    word_indices = pickle.load(f)

with open('indices_word.pickle', 'rb') as f :
    indices_word = pickle.load(f)

with open('story.pickle', 'rb') as ff :
    story = pickle.load(ff)

with open('new_story.pickle', 'rb') as ff :
    new_story = pickle.load(ff)

X = []
Y = []
T = []
maxlen = 250

print("#1")

enc = []
dec = []
N = 499
for i in range(N):
    if len(story[i]) >= 250:
        story[i] = story[i]
        story[i] = story[i][:250]
        enc.append(story[i])
        new_story[i] = new_story[i]
        new_story[i] = new_story[i][:250]
        dec.append(new_story[i])

print("#2")

X = np.zeros((len(enc), maxlen, len(word_indices)))
Y = np.zeros((len(dec), maxlen, len(word_indices)))

print("#3")

for i in range(len(enc)):
    for t, char in enumerate(enc[i]):
        X[i, t, word_indices[char]] = 1
    for t, char in enumerate(dec[i]):
        Y[i, t, word_indices[char]] = 1 

del(story)
del(new_story)

X_train = X[0:int(len(enc)*0.8)][:][:]
X_test = X[int(len(enc)*0.8):int(len(enc)*0.9)][:][:]
X_validation = X[int(len(enc)*0.9):][:][:]
Y_train = X[0:int(len(enc)*0.8)][:][:]
Y_test = X[int(len(enc)*0.8):int(len(enc)*0.9)][:][:]
Y_validation = X[int(len(enc)*0.9):][:][:]

del(enc)
del(dec)
N_validation = len(X_validation)
print(X_train.shape)
print(X_test.shape)
print(Y_train.shape)
print(Y_test.shape)

n_in = len(word_indices)
n_hidden = 128
n_out = len(word_indices)

def weight_variable(shape, name=None):
    return np.random.normal(scale=.01, size=shape)

print("#4")

early_stopping = EarlyStopping(monitor='val_loss', patience=10, verbose=1)

model = Sequential()
model.add(LSTM(n_hidden, input_shape=(maxlen, n_in)))
model.add(RepeatVector(maxlen))
model.add(LSTM(n_hidden, return_sequences=True))
model.add(TimeDistributed(Dense(n_out, kernel_initializer=weight_variable)))
model.add(Activation('softmax'))
model.compile(loss='categorical_crossentropy',
              optimizer=Adam(lr=0.01, beta_1=0.9, beta_2=0.999),
              metrics=['accuracy'])

print("#5")

epochs = 20
batch_size = 10

for epoch in range(epochs):
    print("#")
    model.fit(X_train, Y_train, batch_size=batch_size, epochs=1,
              validation_data=(X_validation, Y_validation),
              callbacks=[early_stopping])
    for i in range(2):
        index = np.random.randint(0, N_validation)
        input_text = X_validation[np.array([index])][:][:]
        output_text = Y_validation[np.array([index])][:][:]
        prediction = model.predict_classes(input_text, verbose=0)

        input_text = input_text.argmax(axis=-1)
        output_text = output_text.argmax(axis=-1)

        q = ''.join(indices_word[i] for i in input_text[0])
        a = ''.join(indices_word[i] for i in output_text[0])
        p = ''.join(indices_word[i] for i in prediction[0])
        print('Q:  ', q)
        print('A:  ', p)

###出力

Using TensorFlow backend.
#1
#2
#3
#4
WARNING:tensorflow:From /Users/PPP/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py:263: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version.
Instructions for updating:
Colocations handled automatically by placer.
WARNING:tensorflow:From /Users/PPP/.pyenv/versions/3.6.8/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py:1190: calling reduce_sum_v1 (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.
Instructions for updating:
keep_dims is deprecated, use keepdims instead
WARNING:tensorflow:From /Users/PPP/.pyenv/versions/3.6.8/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py:1154: calling reduce_max_v1 (from tensorflow.python.ops.math_ops) with keep_dims is deprecated and will be removed in a future version.
Instructions for updating:
keep_dims is deprecated, use keepdims instead
#5
#
WARNING:tensorflow:From /Users/PPP/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.cast instead.
WARNING:tensorflow:From /Users/PPP/anaconda3/envs/tensorflow/lib/python3.6/site-packages/tensorflow/python/ops/math_grad.py:102: div (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Deprecated in favor of operator or tf.math.divide.
Train on 147 samples, validate on 19 samples
Epoch 1/1
2019-05-31 12:37:58.383911: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
2019-05-31 12:37:58.384368: I tensorflow/core/common_runtime/process_util.cc:71] Creating new thread pool with default inter op setting: 4. Tune using inter_op_parallelism_threads for best performance.

147/147 [==============================] - 332s - loss: 8.1267 - acc: 0.0393 - val_loss: 7.7482 - val_acc: 2.1053e-04
Q: 文章（著作権の問題から掲載できません）
A: …………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………………

試したこと

X_trainに正しく数字が格納されているか確かめましたがそこは大丈夫でした。

補足情報（FW/ツールのバージョンなど）

keras 2.0.5