BMT-Net: Broad Multitask Transformer Network for Sentiment Analysis.

Area of Science:

• Natural Language Processing
• Machine Learning
• Artificial Intelligence

Background:

• Sentiment analysis faces challenges with the vast scale of online opinionated text.
• Pretrained language models offer improved contextual representations over traditional methods.
• Current approaches often treat feature-based and fine-tuning methods for pretrained models separately.

Purpose of the Study:

• To propose a Broad Multitask Transformer Network (BMT-Net) to address limitations in sentiment analysis.
• To leverage both feature-based and fine-tuning methods for enhanced contextual representations.
• To create universal representations across different sentiment analysis tasks.

Main Methods:

• Developed BMT-Net, a multitask transformer network integrating feature-based and fine-tuning approaches.
• Utilized a broad learning system for deep and broad feature searching.
• Employed multitask transformers to ensure learned representations are universal across tasks.

Main Results:

• Achieved an F1-score of 0.778 on the Twitter dataset (SemEval Sentiment Analysis in Twitter).
• Reached an accuracy of 94.0% on the SST-2 dataset (Stanford Sentiment Treebank).
• Demonstrated superior performance compared to existing state-of-the-art methods.

Conclusions:

• BMT-Net effectively enhances sentiment analysis by learning robust contextual representations.
• The proposed method highlights the importance of deep and broad contextual feature searching.
• The universal representations generated are beneficial for various sentiment analysis tasks.