Text Classification Competition - Twitter Scarcasm Detection

This task is to detect sarcasm from contextual tweets and beat the baseline performance of F1 = 0.723.

Voiced Presentation (Demo) Link

https://mediaspace.illinois.edu/media/t/1_685r9kih

Setup for prediction generation on test dataset

• Please open the Twitter_Scarcasm_Detection_Source_Code.ipynb file from Google Colab. (Link directly to Colab)
• Go to Runtime -> Change runtime type, and make sure it has GPU selected as Hardware accelerator and High-RAM as Runtime shape.
• Go to Runtime -> Run all. It takes approximately 5 minutes to complete.
• Before you download the answer.txt, you can also look at the validation F1 score, which is usually ~0.83. You can use the Table of contents toolbar on the left to navigate to section 7. Evaluation.
• Use the Files toolbar on the left, go to outputs -> Twitter_Sarcasm_Detection, and you should be able to see answer.txt.

Final Result on test dataset

F1 = 0.7626858

More Details

Please review the project documentation file, which includes all models I have tried and three different methods I used for Context text string. It also covers model performance comparison and different method comparison specification for this task.