Essential characteristics of convolution operation +textcnn text classification

1. CNN The essence of
   Three points ： Sparse interaction 、 Parameters of the Shared 、 Translation invariance , For details, see link
2. textCNN Length of convolution kernel
   Commonly used convolution kernels are 3 Kind of ： 3、 ... and 、 Four 、 5、 ... and . Width is the width of our vector .（ Specific details , I want to see others ,300）
3. cnn And global vector Glove Combined thinking of
   Cnn The core of a sentence is to grasp the local information in the sentence , Is it possible to consider the global vector GLOVE Exist side by side and play a part together .
4. CNN Why is it better than RNN Easier parallel computing
   Reference 1 link
   Reference 2 link
5. Specific process ：
   （1） Extended dimension and convolution ： This is included at the end 1 dimension , namely shape=(?, 135, 128, 1),filter_shape: [3, 128, 1, 128], The result is ： shape=(?, 133, 1, 128)
   Be careful ： The general addition is in the 2 dimension , namely shape=(?, 1,135, 128)
   （2） Relu Function activation , The result is ：shape=(?, 133, 1, 128);
   （3） Maximum pooling .
   Be careful ： Convolution Padding Operation adopts valid The way , Its output dimension is ：sequence_length - filter_size + 1. thus , Make a sentence into a vector .
   （4） The last three output vectors concat become shape[None,1,1,384];flat Post output is [None,384].
   （5） Dropout after , Projection output .
   （6） Loss function ：softmax_cross_entropy（_with_logits）.
6. textcnn Reference articles for
   link 1
7. Some common problems of training parameter initialization ：
   link 1