I'm training a bidirectional lstm with attention on a dataset with text data and six target classes.

The F1 score on the test set by class is about 0.7 for four of the classes, and about 0.35 for the other two. Not, only that, the model is confusing those two classes and I can't find the pattern in the data by which it's making the prediction for them.

I'm using fasttext pretrained word vectors.

I couldn't find a proper solution for such a problem.

So, my question is: Is there a research on this topic that you could guide me to, or a solution that I could use?

  • 2
    $\begingroup$ Probably these two classes are much less frequent than the other 4, what is their frequency? if you want to be able to find the pattern used by the model, you should probably use a traditional approach rather than DL, like decision trees. $\endgroup$ – Erwan Dec 9 '19 at 0:14
  • $\begingroup$ Thank you for your answer. I forgot to mention that the dataset is balanced. I'll try out your suggestion. $\endgroup$ – donots Dec 9 '19 at 8:50

One option is to print the confusion matrix to see the rate of misclassification by class.

Another option is to read through the misclassified documents to see if a pattern is apparent.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.