How does Keras 1d convolution layer work with word embeddings - text classification problem? (Filters, kernel size, and all hyperparameter)

I am currently developing a text classification tool using Keras. It works (it works fine and I got up to 98.7 validation accuracy) but I can't wrap my head around about how exactly 1D-convolution layer works with text data.

What hyper-parameters should I use?

I have the following sentences (input data):

• Maximum words in the sentence: 951 (if it's less - the paddings are added)
• Vocabulary size: ~32000
• Amount of sentences (for training): 9800
• embedding_vecor_length: 32 (how many relations each word has in word embeddings)
• batch_size: 37 (it doesn't matter for this question)
• Number of labels (classes): 4

It's a very simple model (I have made more complicated structures but, strangely it works better - even without using LSTM):

model = Sequential()
model.add(Embedding(top_words, embedding_vecor_length, input_length=max_review_length))
model.add(Conv1D(filters=32, kernel_size=2, padding='same', activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(Flatten())
model.add(Dense(labels_count, activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model.summary())

My main question is: What hyper-parameters should I use for Conv1D layer?

model.add(Conv1D(filters=32, kernel_size=2, padding='same', activation='relu'))

If I have following input data:

• Max word count: 951
• Word-embeddings dimension: 32

Does it mean that filters=32 will only scan first 32 words completely discarding the rest (with kernel_size=2)? And I should set filters to 951 (max amount of words in the sentence)?

Examples on images:

So for instance this is an input data: http://joxi.ru/krDGDBBiEByPJA

It's the first step of a convoulution layer (stride 2): http://joxi.ru/Y2LB099C9dWkOr

It's the second step (stride 2): http://joxi.ru/brRG699iJ3Ra1m

And if filters = 32, layer repeats it 32 times? Am I correct? So I won't get to say 156-th word in the sentence, and thus this information will be lost?

Solution 1:

I would try to explain how 1D-Convolution is applied on a sequence data. I just use the example of a sentence consisting of words but obviously it is not specific to text data and it is the same with other sequence data and timeseries.

Suppose we have a sentence consisting of m words where each word has been represented using word embeddings:

Now we would like to apply a 1D convolution layer consisting of n different filters with kernel size of k on this data. To do so, sliding windows of length k are extracted from the data and then each filter is applied on each of those extracted windows. Here is an illustration of what happens (here I have assumed k=3 and removed the bias parameter of each filter for simplicity):

As you can see in the figure above, the response of each filter is equivalent to the result of its convolution (i.e. element-wise multiplication and then summing all the results) with the extracted window of length k (i.e. i-th to (i+k-1)-th words in the given sentence). Further, note that each filter has the same number of channels as the number of features (i.e. word-embeddings dimension) of the training sample (hence performing convolution, i.e. element-wise multiplication, is possible). Essentially, each filter is detecting the presence of a particular feature of pattern in a local window of training data (e.g. whether a couple of specific words exist in this window or not). After all the filters have been applied on all the windows of length k we would have an output of like this which is the result of convolution:

As you can see, there are m-k+1 windows in the figure since we have assumed that the padding='valid' and stride=1 (default behavior of Conv1D layer in Keras). The stride argument determines how much the window should slide (i.e. shift) to extract the next window (e.g. in our example above, a stride of 2 would extract windows of words: (1,2,3), (3,4,5), (5,6,7), ... instead). The padding argument determines whether the window should entirely consists of the words in training sample or there should be paddings at the beginning and at the end; this way, the convolution response may have the same length (i.e. m and not m-k+1) as the training sample (e.g. in our example above, padding='same' would extract windows of words: (PAD,1,2), (1,2,3), (2,3,4), ..., (m-2,m-1,m), (m-1,m, PAD)).

You can verify some of the things I mentioned using Keras:

from keras import models
from keras import layers

n = 32  # number of filters
m = 20  # number of words in a sentence
k = 3   # kernel size of filters
emb_dim = 100  # embedding dimension

model = models.Sequential()
model.add(layers.Conv1D(n, k, input_shape=(m, emb_dim)))

model.summary()

Model summary:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv1d_2 (Conv1D)            (None, 18, 32)            9632      
=================================================================
Total params: 9,632
Trainable params: 9,632
Non-trainable params: 0
_________________________________________________________________

As you can see the output of convolution layer has a shape of (m-k+1,n) = (18, 32) and the number of parameters (i.e. filters weights) in the convolution layer is equal to: num_filters * (kernel_size * n_features) + one_bias_per_filter = n * (k * emb_dim) + n = 32 * (3 * 100) + 32 = 9632.