Deep learning models for sentiment analysis classification of greek general election tweets into 3 categories:
- Positive
- Neutral
- Negative
Four different classifiers have been developed on the same datasets using real twitter data:
In short, experimented with:
- Data preprocessing
- Different vectorizers (count, BoW, TF-IDF)
- Removal of features with very low/high frequency by adding max-df / min-df parameters
- Hyperparameter optimization using the Optuna framework
Place 36/114 at the course's kaggle competition.
In short, experimented with:
- Data preprocessing
- Different number of layers, neurons per layer & learning rates
- Activation functions (Linear & non-linear)
- Dropout layers
- Different optimizers (Stochastic Gradient Descent, Adam)
- Early stopping
- Hyperparameter optimization using the Optuna framework
Place 1/95 at the course's kaggle competition.
In short, experimented with:
- Data preprocessing
- Bidirectional stacked RNN's with LSTM/GRU cells
- Gradient Clipping
- Early stopping
- Skip connections
- Attention to the best model
- Hyperparameter optimization using the Optuna framework
Place 1/84 at the course's kaggle competition.
In short, experimented with:
- Data preprocessing
- BertModel
- DistilBertModel
- Adding dropout layers
- Hyperparameter optimization using the Optuna framework
Place 6/70 at the course's kaggle competition.
The models were created for the Artificial Intelligence II course under prof. Manolis Koubarakis.