Multitask learning with bidirectional encoder representations from transformers for sentiment analysis and sarcasm detection / Tan Yik Yang

• Main Author: Tan , Yik Yang
• Format: Thesis
• Published: 2024
• Subjects: QA76 Computer software; TK Electrical engineering. Electronics Nuclear engineering
• Online Access: http://studentsrepo.um.edu.my/15493/1/Tan_Yik_Yang.pdf; http://studentsrepo.um.edu.my/15493/2/Tan_Yik_Yang.pdf; http://studentsrepo.um.edu.my/15493/

In recent years, sentiment analysis has garnered significant interest in social media analytics, aiming to categorize people's thoughts, emotions, and feelings into positive, negative, or neutral categories. However, the increasing volume, complexity, and authenticity of social media data have introduced challenges such as misunderstanding, uncertainty, and inaccuracy. Particularly notable is the difficulty of identifying sarcasm in textual data, where negative intentions are expressed through positive sentences, presents a significant obstacle to sentiment analysis on social media platforms. This thesis proposes a novel multi-task learning framework that leverages Bidirectional Encoder Representations from Transformers (BERT), a state-of-the-art language model, to establish a correlation between sentiment analysis and sarcasm detection. The primary objective is to enhance the overall performance of sentiment analysis by identifying instances of sarcasm. The model's efficacy is demonstrated through comprehensive experiments, showing a notable improvement in F1-scores ranging from 2.5 to 6.5 percent upon incorporating sarcasm detection. The proposed approach not only enhances the sentiment classifier's performance but also significantly reduces training time and computational resources, offering substantial practical advantages. The findings underscore the importance of recognizing sarcasm in sentiment analysis and highlight how improved sentiment analysis aids in understanding sarcastic expressions in social media data. In the past, most sentiment analysis work treated the task as a standalone process. However, this thesis provides valuable insights into the influence of sarcasm on sentiment analysis, showing that accuracy can be improved in sentiment analysis by detecting sarcasm.