Machines’ Interpretation of Sarcasm A Comparative : Study of Different NLP Models

Sammanfattning

Human language is complex and full of nuances, including irony, sarcasm, and jokes. The ability of humans to understand and differentiate between these nuances depends on our social capability, our capacity to identify various patterns and contradictions, as well as non-verbal expressions, among others. As technology has become an integrated part of our daily lives, much has been digitized—including customer service functions on websites that are often managed by chatbots, artificial systems like Chat-GPT, as well as the analysis of social media content for marketing purposes and product development. These applications are just a few examples of the broad scope of use, which requires computers to understand human language. This involves understanding sarcasm, which is inherently subtle and influenced by a range of cultural contexts, making it challenging to interpret. Within this area, Natural Language Processing (NLP), a branch of Artificial Intelligence (AI), plays a crucial role. Despite extensive research in NLP and sarcasm detection, there is a notable lack of comparative analysis among different NLP techniques and their abilities to interpret and correctly classify sarcastic content in text-based communication. This gap in research constitutes the primary issue addressed in this thesis. This thesis focuses on sarcasm and natural language processing by comparing selected NLP models—BERT, GPT-2, and T5-base—and evaluating their ability to detect sarcasm. The purpose is to expand knowledge in this area and contribute new perspectives on the strengths and weaknesses of the selected models. The goal is to contribute to the academic world as well as support companies that use or develop this technology. The study uses a qualitative method supported by quantitative data. An extensive literature review was conducted to deepen the understanding of the area. Datasets were used for analysis, and the results form the basis for evaluating the selected models’ ability to identify sarcasm. The results indicate that among the selected models, both BERT and GPT-2 are effective for sarcasm classification and detection, while T5-base has shown significantly poorer performance, making it less suitable for such a task. This answers the thesis question.