So I have a task to classify the relation between 2 texts (4 classes possible) and one of the requirements is to preprocess them with TfidfVectorizer or CountVectorizer.

Since every sample has 2 strings, I figured it would be useful to make a CNN with a 1D conv layer so that it captures the relation between feature1 from the first string (vectorizer output for word 1 from the vocabulary or max_features if used) and feature1 from the second string (so the shape would be n_samples x n_features x 2).

Can anyone guide me on how I should create the layers after the first CONV1D input layer?

Also, I should add that the dataset is imbalanced (2k 2k 25k 25k aprox samples per classes) so I will use class_weights and also a dropout layer and earlystopping would also help on regularization

Thanks in advance for any tips!



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.