Comparison is wrong between tfidf on words and CNN char. You should use char ngrams along with LR and this will beat all your classifiers with high probability.

Comparison is wrong between tfidf on words and CNN char. You should use char ngrams along with LR and this will beat all your classifiers with high probability. This is because your CNN char does not have enough data to draw all the useful chat ngrams. Doing it as preprocessing and passing it to LR is in practice always better on small datasets. You can go one step forward and add layers and test an MLP on your char ngrams.