0
$\begingroup$

I'm working on a Twitter classification task and while analyzing the errors I found quite a few strange predictions. I'm searching for a tool (preferably open-source) similar to https://towardsdatascience.com/how-does-bert-reason-54feb363211 that is able to compute the highest positive/negative attribution given to the words (the reason why I'm not able to use the approach presented in the aforementioned article is the price). In this way, I hope that I'll be able to better understand (and possibly correct) these misclassifications. I tried looking at the attention heads but I don't feel that I'm able to fully understand and draw conclusions based on this information.

Any help would be greatly appreciated!

Improve this question
$\endgroup$

3 Answers 3

Reset to default
1
$\begingroup$

You can use SHAP for the model interpretation.

Link: https://github.com/slundberg/shap#natural-language-example-transformers

Improve this answer
$\endgroup$
0
$\begingroup$

You can try to implement Integrated Gradients yourself (https://www.tensorflow.org/tutorials/interpretability/integrated_gradients), see also the paper for reference

Another alternative is using the attention mechanism (Here's a little example i found), or as the previous answer pointed out, you can use SHAP.

Improve this answer
$\endgroup$
-1
$\begingroup$

BERT is free. Any Sentiment Analysis model will achieve the task.

Improve this answer
$\endgroup$
1
  • $\begingroup$ I know that BERT is free. I was actually asking about a tool similar to the one presented in the linked article (developed at Fiddler). $\endgroup$
    – moz_szt
    Sep 8, 2020 at 22:39

Your Answer

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

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