Projects                                     Home

Unsupervised Domain Adaption for Sentiment Analysis using BERT  

Collaborators | Abdul Waheed, Bhavitvya Malik, Yifan Zhang
Using domain adversarial training we reduce the distance(MMD) between the source domain representations and the target domain representations for the task of sentiment analysis using the Amazon reviews dataset(Books being the source domain and DVDs being the target domain) leading to a clear increase in the accuracy on the target domain. This procedure is performed on all the layers of BERT and the perfomance is analysed. We also won the first place in the 17th SoC Term Project Showcase, NUS. [Poster]

Using Transfer Learning for Drug Detection from Tweets  

Using the SMM4H 2020 Shared Task dataset, I carried out an analysis of the performance of different pretrained sentence encoders on automatic classification of tweets mentioning a drug, that differ in size, pretraining objective, and pretraining data. This was submitted as my thesis in my final year during my undergrad. [pdf].