Document Type
Dissertation
Rights
This item is available under a Creative Commons License for non-commercial use only
Disciplines
1.2 COMPUTER AND INFORMATION SCIENCE, Computer Sciences, Information Science, Bioinformatics
Abstract
It is a common practice in today’s world for the public to use different micro-blogging and social networking platforms, predominantly Twitter, to share opinions, ideas, news, and information about many things in life. Twitter is also becoming a popular channel for information sharing during pandemic outbreaks and disaster events. The world has been suffering from economic crises ever since COVID-19 cases started to increase rapidly since January 2020. The virus has killed more than 800 thousand people ever since the discovery as per the statistics from Worldometer [1] which is the authorized tracking website. So many researchers around the globe are researching into this new virus from different perspectives. One such area is analysing micro-blogging sites like twitter to understand public sentiments. Traditional sentiment analysis methods require complex feature engineering. Many embedding representations have come these days but, their context-independent nature limits their representative power in rich context, due to which performance gets degraded in NLP tasks. Transfer learning has gained the popularity and pretrained language models like BERT(bi-directional Encoder Representations from Transformers) and XLNet which is a Generalised autoregressive model have started overtaking traditional machine learning and deep learning models like Random Forests, Naïve Bayes, Convolutional Neural Networks etc. Despite the great performance results by pretrained language models, it has been observed that finetuning a large pretrained model on downstream task with less training instances is prone to degrade the performance of the model. This research is based on a regularization technique called Mixout proposed by Lee (Lee, 2020). Mixout stochastically mixes the parameters of vanilla network and dropout network. This work is to understand the performance variations of finetuning BERT and XLNet base models on COVID-19 tweets by using Mixout regularization for sentiment classification.
DOI
https://doi.org/10.21427/fhx8-vk25
Recommended Citation
Dussa, A. (2020) Finetuning Pre-trained language models for sentiment classification of COVID19 tweets,Dissertation, Technological University Dublin. doi:10.21427/fhx8-vk25
Publication Details
A dissertation submitted in partial fulfilment of the requirements of Technological University Dublin for the degree of M.Sc. in Computer Science (Data Analytics)