Regression problem in NLP with 100M labeled examples; data randomly shuffled and split 90/10 train/test. Train RMSE is twice as high as test RMSE. How should you improve model performance?

The problem states that the training RMSE is twice as high as the test RMSE. This indicates underfitting, meaning the model is too simple to capture the underlying patterns in the training data. Increasing model complexity, such as adding layers or expanding vocabulary/n-gram size, allows the model to learn more intricate relationships and reduce training error. Increasing the test sample share would not address underfitting. Collecting more data is generally good, but if the current model is underfitting, it won't effectively utilize the additional data. Regularization techniques are used to combat overfitting (when training error is low but test error is high), which is the opposite of the scenario described.