Comparative Analysis of Large Language Model as Feature Extraction Methods in Sarcasm Detection Using Classification Algorithms

Sarcasm Detection is an important task in Natural Language Processing (NLP) because sarcasm expression can distort sentiment analysis and mislead autamated decision making system. The urgency of this research lies in the limitations of traditional methods such as Bag of Words (BoW) and TF-IDF, which fail to capture deep contextual understanding, while word embedding techniques like Word2Vec and Glove have improvement, difficulties remained in comprehending full sentence meanings. Large Language Model (LLM) such as BERT and RoBERTa have Transformed NLP by capturing contextual word representation, making them more effective for sarcasm detection. This study presents a comparative analysis of various feature extraction method like Word2Vec, Glove, BERT, and RoBERTa combined with classification algorithms such as Support Vector Machine (SVM), XGBoost, and Random Forest. This study uses the Knowledge Discovery in Database (KDD) Framework which includes data selection, preprocessing, transformation, modeling, and evaluation. The dataset consists of news headlines labeled as sarcastic or non sarcastic, Principal Component Analysis used for dimensionality reduction by eliminating redundant features. The result show that the RoBERTa-SVM combination achieves the highest accuracy of 88.00 %, indivating the superiority of transformer based models over traditional embedding techniques. This study confludes that integrating contextual embeddings and feature selection improves sarcasm detection perfomance while maintaining computational efficiency. However, the model still faces challenges in identifying implicit sarcasm due the absence of explicit linguistic cuesm which represent a limitation of this research.