Bidirectional CRNN and BERT Embeddings for Sentiment Analysis on Twitter Data

This project aims to perform sentiment analysis on a dataset consisting of tweets. We start by considering two baseline results obtained through simple machine learning techniques with different embeddings and compare these to a more complex model centered around a state-of-the-art architecture for this kind of task, using BERT embeddings. In the comparative study, we evaluate the classification accuracy of each of the models. The full report can be found at the following link.

Setting up

To run the code you first need to install the required packages. A requirements.txt file is provided.

To run the exeriments, first save the dataset in a folder called twitter-datasets.

Running different experiments

We provide the different logs of the experiments in the folder called lightning_logs.

To execute the experiments on clean data, first execute the following command and change the training file in the train_model.py:

python src/clean_data.py

Training

To run the training, execute inside the root folder the following command:

python src/train_model.py --lr=<learning_rate> --n_epochs=<n_epochs> --batch_size=<batch_size> --n_workers=<n_workers> --label_smoothing=<label_smoothing> --sched_step_size=<sched_step_size> --sched_gamma=<sched_gamma>

Predicting

To predict the sentiment of a test file using a model, execute inside the root folder the following command:

python src/predict_model.py --n_epochs=<n_epochs> --batch_size=<batch_size> --n_workers=<n_workers> --model=<model>

Our model

Our model checkpoint is the final result of the following command:

python src/train_model.py --n_epochs=2

To obtain the predictions with our model, we provide the checkpoint file in the following link. This checkpoint should be saved inside a folder in lightning_logs/final/checkpoints. To reproduce the predictions that we provide in the folder predictions/final, just execute the commad:

python src/predict_model.py