1 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 1 / 1
Yayın Sarcasm detection in text using deep neural networks(Işık Üniversitesi, 2024-02-25) Gümüşçekiçci, Gizem; Dehkharghani, Rahim; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Yüksek Lisans Programı; Işık University, School of Graduate Studies, Computer Science Engineering Master ProgramSarcasm is a form of irony which is generally used in expressing negative opinions. Sarcasm poses a linguistic challenge due to its figurative nature where intended meaning contradicts with literal interpretation. Sarcasm is widely used in our Daily lives and also upon many social platforms. Detecting sarcasm in written text is a challenging process that has captured the interest of many researchers. Hence, sarcasm has become a crucial task in the Natural Language Processing (NLP) field. This thesis study explores the concept of sarcasm, and its importance on existing sarcasm research. The automatic process of sarcasm detection involves dataset selection, preprocessing steps, and selecting proper approaches, including rule-based methods, Machine Learning (ML), Deep Learning (DL) and Transformer architectures. This study surveys previous research on sarcasm detection, specifically examining the dataset, methodology and performance. This thesis study attempts to automatically detect sarcasm by utilizing various ML, DL and transformer and hybrid neural network architectures on news headlines datasets. To overcome the dataset and performance limitations on existing approaches, we propose various methodologies to detect sarcastic text mostly focusing on DL, hybrid neural networks and transformer architectures. We combine appropriate architectures with several hand-crafted features and utilizing different word embedding models. To further extend the performance of our proposed methods and also enhance the existing news headlines dataset, we proposed several modifications. We contribute to the existing dataset by applying augmentation to increase the dataset size to help enhance the performance of the proposed models with overcoming dataset limitations. Our methodologies correctly identify sarcasm with 97.68% F1 score.
