Tools
Tools
Deshmukh, Nikhil Vishnupant (2023) Transfer Learning for Identification of disaster tweets using fine-tuning DistilBERT. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (1MB) | Preview |
Preview |
PDF (Configuration manual)
Download (911kB) | Preview |
Abstract
This study presents an in-depth exploration of tweet classification for disaster-related content detection, addressing the challenges posed by the proliferation of social media during crises. Three distinct models were developed and evaluated: CNN + LSTM with Glove Embedding, RoBERTa Transformer, and FineTune DistilBERT Transformer. These models were rigorously assessed for their ability to distinguish between disaster and non-disaster-related tweets, with a focus on accuracy, sensitivity, and specificity. The CNN + LSTM model exhibited promising precision but lacked recall for disastrous tweets. RoBERTa demonstrated enhanced performance owing to its extensive training data and methodology. However, the FineTune DistilBERT model emerged as the standout performer, showcasing a balanced sensitivity-specificity trade-off and achieving an impressive 89% accuracy. Leveraging its DistilBERT architecture, this model offers a compact yet powerful solution for accurate tweet classification. The findings underline the potential of transformer-based models in crisis informatics, specifically for identifying disaster-related content in social media streams. This study contributes to advancing rapid and accurate crisis response tools, empowering humanitarian organizations with improved insights and aiding decision-making in disaster scenarios. Further research avenues may explore ensemble methods and domain-specific fine-tuning to enhance model performance across diverse disaster contexts.
Actions (login required)
 |
View Item |