I have built a multinomial naive bayes model and a CNN model for binary sentiment analysis. Trained and tested on the same dataset, yet the MNB model has 0.88 accuracy and the CNN model has 0.89. From previous literature on the subject I expected a more substantial difference. What could be the reason?
I noticed the CNN model might be overfitting, since training accuracy kept going up and val accuracy plateued quickly, so I tried putting my dropout layers to 0.50 and adding regularization to my convoluted layers. The model's performance marginally went up with 0.50 dropout but quickly dropped when I added regularization too.
