Visualising and Analysing the Performance of ...

Visualising and Analysing the Performance of Bert on Stanford Sentiment Treebank

Author Name : Nishkarsh Mittal

ABSTRACT This paper investigates the performance of the Bidirectional Encoder Representations from Transformers (BERT) on the Stanford Sentiment Treebank (SST), a benchmark dataset for sentiment analysis. Utilizing BERT's advanced natural language processing capabilities, we conduct a series of experiments to evaluate its effectiveness in sentiment classification tasks. The study involves fine-tuning BERT on the SST, analyzing its performance metrics, and visualizing the results through various methods, including confusion matrices and attention heatmaps. Our findings indicate that BERT significantly outperforms traditional sentiment analysis models, achieving a high accuracy rate while effectively capturing nuanced sentiments. The visualizations provide insights into BERT's decision-making process, illustrating how attention mechanisms contribute to its predictions. This research contributes to the understanding of transformer-based models in sentiment analysis and offers practical implications for their deployment in real-world applications.