Multi-Task Learning for Sarcasm Detection and Sentiment Analsysis Using BERT

This paper proposes a Multi –Task Learning Model approach to detect Sarcasm and to analyze Sentiment within the given text. Sarcasm detection is important for enhancing natural language understanding particularly in customer reviews and social media where sarcastic sentences are mostly used. This system uses BERT (Bi-directional Encoder Representations from Transformers) to processes both tasks concurrently. It uses separate layers to handle each task individually. The text processing pipeline includes techniques such as text normalization, tokenization, and lemmatization and then followed by encoding using BERT’s tokenizer before using it for model training. BERT can understand the complex patterns of language which helps in easily recognize the small clues present in sarcastic statements. The architecture of the model incorporates specific layers for sarcasm and sentiment classification which optimizes prediction with loss functions for each task. If there is imbalance in class distributions class weighting is applied and model evaluation is performed using K-Fold cross validation to ensure robustness. To measure the model effectiveness performance metrics such as ROC-AUC curves, confusion matrix and detailed classification reports are used. Therefore, including MTL and BERT improves the performance and effectiveness in sarcasm detection and sentiment analysis.