Machine learning Text emotion classifications using BERT, RoBERTa, DistilBERT, XLNet

Text classifications

BERT, double BERT, triple BERT
*Get measures for (BERT)
Type 'This is doable' at the start of your proposal
*Make Bi BERT (double BERT)-> take the output 150 features from BERT as input to biBERT layer
*Get measures for (biBERT): 150, 100,50 features as output from BERT and input to biBERT
*Get measures for(TriBERT): TriBERT (triple BERT) take the output 100 or 50 output feature of biBERT as input to triBERT

*Repeat the pervious by changing Bert to be ( RoBERTa, DistilBERT, XLNet)
Check the optimization of each method of them as a standalone method before making bi and tri

-Code should be general to any dataset with any number of classes
-classifiy emotions to 7 classes (anger, disgust, sadness, fear, joy, shame, and guilt.)

- Data Preprocessing
- Data Augmentation
- Skew data optimization
-
- Plots and measures:
- Give full measures values with accuracies like (Accuracy, fscore, recall, precision, Heatmap, PR, Roc curve …) for all method and for each class.
- Comparisons for all methods in one graph for each class separately and for all classes together in one plot and average
- 10 fold cross validation
- Get high accuracies more than 97%
- Do on Kaggle
- Write full documentation of methods used contains detailed description and figures and diagram of the architectures used.



- Use datasets:
- ISEAR data contains 7666 sentences, 7 types of emotions (joy, anger, guilt, sadness, disgust, fear, and shame).
- COVID-19 Survey Data 2408 records (Anger, Anxiety, Fear, Relaxation, Sadness)
- GoEmotions Data 553 records (Anger, Disgust, Fear, Joy, Sadness, and Surprise)
- WASSA-21 Data train1585 records and tested 245 records (Anger, Disgust, Fear, Joy, Sadness, and Surprise)
- Use preprocessing of data before using it : we turned text in all datasets to lower case and removed stop words using NLTK
- use of data augmentation in order to improve performance.
-


Datasets:
1)ISEAR
Text Classification: Emotions Description
https://www.kaggle.com/datasets/radedaevi/emotions-description
7666 sentences
7 classes (joy, fear, anger, sadness, disgust, shame, and guilt)


2)COVID-19 Survey Data 2408 records (Anger, Anxiety, Fear, Relaxation, Sadness)
https://github.com/ben-aaron188/covid19worry

3)GoEmotions Data 553 records (Anger, Disgust, Fear, Joy, Sadness, and Surprise)
https://github.com/google-research/google-research/tree/master/goemotions

4)WASSA-21 Data train1585 records and tested 245 records (Anger, Disgust, Fear, Joy, Sadness, and Surprise)
https://competitions.codalab.org/competitions/28713