---

Twitter Sentiment Analysis

Binary classification experiments for the Twitter dataset

Notebook viewer

‼️ Because of memory restrictions, GitHub and Browsers can't open always big jupyter notebooks. For this reason I have every notebook linked with the ✔️ jupyter nbviewer ✔️ in the following table. If you have any problems opening the notebooks, follow the links.

Notebook	Link to jupyter nbviewer	Link to Colab
BiRNN_LSTM_GRU-BestModel.ipynb
BiRNN_LSTM_GRU-Experiments.ipynb
FeedForwardNN_GloVe.ipynb
FeedForwardNN_TfiDf.ipynb
LogisticRegression.ipynb

---

Logistic regression

Developed a sentiment classifier using logistic regression for the Twitter sentiment classification dataset available in this link. I used the toolkit Scikit-Learn again.

Vectorization: Tf-Idf

Tf-Idf vectorization of the tweets. No pretrained vectors

Evaluation

Model metrics for evaluation: F1 score, Recall and Precision

Visualization: Confusion matrices

---

Feed-Forward Neural Net

Developed two sentiment classifier using feed-forward neural (pyTorch) networks for the Twitter sentiment analysis dataset.

Experimented with:

• the number of hidden layers, and the number of their units
• the activation functions (only the ones presented in the lectures)
• the loss function
• the optimizer, etc

Vectorization-1: Tf-Idf

Tf-Idf vectorization of the tweets. No pretrained vectors

Vectorization-2: Pre-trained word embendding vectors - GloVe

Vectorization made with GloVe (Stanford pre-trained embenddings)

Evaluation

Model metrics for evaluation: F1 score, Recall and Precision

Visualization: ROC curves, Loss vs Epochs, Accuracy vs Epochs and Confusion matrix

---

Bidirectional stacked RNN with LSTM/GRU cells

Experimented with:

• the number of stacked RNNs,
• the number of hidden layers,
• type of cells,
• skip connections,
• gradient clipping and
• dropout probability

Used the Adam optimizer and the binary cross-entropy loss function and transformed the predicted logits to probabilities using a sigmoid function.

Vectorization: GloVe

Pre-trained word embeddings (GloVe) as the initial embeddings to input on models.

Evaluation

Model metrics for evaluation: F1 score, Recall and Precision

Visualization: ROC curves, Loss vs Epochs, Accuracy vs Epochs and Confusion matrix

---

© Konstantinos Nikoletos | 2020 - 2021