peithonking/fire2022

Me and Diptarko participating in FIRE 2022.

FIRE 2022

The problem presented to us was sentiment classification in nature

  Given the massive success of Transformers in the field of NLP, we decided to use Transformer-based architectures for our model. We used a variant of the Bert Model. To build our pipeline, we used the Hugging face interface since it already has state-of-the-art architectures predefined. We used Pytorch (Vanilla) for training a Pre-trained Bert model which was specifically trained on a sentiment analysis task of tweets. Since our data corpus was very small we did not try to train the architecture from scratch. We also used some state-of-the-art data augmentation techniques to increase our data corpus and make our training more stable. We used AdamW optimiser along with a cosine decay schedule for our training.