The use of pre-trained embeddings or pre-trained architectures provides a solid basis for achieving high results on the identification of hatefull tweets. The RoBERTa-based model achieves better results than the CNN model. The increase in performance is due to the pre-trained embeddings provided by a model trained on tweets. Table 1 shows the achieved results.
See report.pdf for more information
- install pytorch: https://pytorch.org/get-started/locally/
- Install the following packages:
torchtext torchmetrics mlxtend matplotlib install nltk emoji transformers regex requests hydra-core omegaconf statsmodels seaborn scipy - install spaCy: https://spacy.io/usage
- download dataset at https://bit.ly/3HUgNPo and extract into the folder
raw_data - run
run.pyfrom the main project folder