看 understanding LSTM: http://colah.github.io/posts/2015-08-Understanding-LSTMs/
LSTM通过gate控制信息的舍弃和更新
看论文《Effective LSTMs for Target-Dependent Sentiment Classification》
TD-LSTM 用关键词前和关键词后的语句，分布作为left_LSTM和right_LSTM的输入，最后用softmax分类, left=left+target, right=right+target
TC-LSTM: 和TD-LSTM类似，不同之处在 left=left+targe_average_vector, right=right+targe_average_vector
