I am working on a text classification model predicting classes for text. We have languages from many parts of the world and some of our classes are dominated by specific languages. The model we are using is:


Even though the model is multilingual it shows bias pushing certain languages towards specific classes. If I translate text from English to Thai I will receive different predictions. Given the dataset imbalance in classes/languages, this is understandable but I'd like to improve it.

I'm wondering if someone has a good solution for decreasing this bias? I'm thinking of simply translating training data between the languages to reduce it



Your Answer

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