465对语料，每句话15个词语，60维向量表示。
训练20000轮次。
model.summary()
Layer (type) Output Shape Param #
=================================================================
lstm_1 (LSTM) (None, 15, 60) 29040

---

lstm_2 (LSTM) (None, 60) 29040

---

dropout_1 (Dropout) (None, 60) 0

---

repeat_vector_1 (RepeatVecto (None, 15, 60) 0

---

lstm_3 (LSTM) (None, 15, 60) 29040

---

lstm_4 (LSTM) (None, 15, 60) 29040

---

dropout_2 (Dropout) (None, 15, 60) 0

---

time_distributed_1 (TimeDist (None, 15, 60) 3660

Total params: 119,820
Trainable params: 119,820
Non-trainable params: 0

某些问题能给出比较精确的回答，另外一些问题就瞎扯了。

分析：
数据集质量不高，网络也不深。
更好的数据集它是能学到一些措辞规律的。