TY - GEN
T1 - Prompt strategies for sarcastic meme detection
T2 - PhD Symposium, Posters, Demos, and A Web for more inclusive, sustainable and prosperous societies, WEB-for-GOOD 2024 and 1st International Workshop on AI and Web Data Analytics, AIWDA 2024 form the 25th International Conference on Web Information Systems Engineering, WISE 2024
AU - Abdullakutty, Faseela
AU - Al-Maadeed, Somaya
AU - Naseem, Usman
PY - 2025
Y1 - 2025
N2 - Memes, often characterized by subtle humour and irony, have become a prominent digital communication medium. Detecting sarcasm in memes presents a significant challenge due to its context-dependent nature, negatively impacting user experiences on social media platforms. To improve the ability of social media systems to recognize and manage sarcastic content, this study investigates the effectiveness of Large Language Models (LLMs) for sarcasm detection in memes. Specifically, we evaluate three prompting techniques: Standard Prompt, Chain of Thought (CoT), and Concise Chain of Thought (CCoT) to determine their impact on the classification of sarcastic memes. Using the GOAT dataset as a benchmark, the study employs four pre-trained LLMs: Flan-T5-XXL, Llama-2, Mistral 7B, and GPT-2. The research identifies the most effective prompting strategies for sarcasm detection through a comparative analysis. The results demonstrate that CoT and CCoT significantly enhance performance over the Standard Prompt, with CCoT achieving the highest accuracy, particularly with advanced models like Mistral 7B. However, the choice of prompting technique depends on both the model and task requirements, emphasizing the need for tailored approaches in sarcastic meme analysis.
AB - Memes, often characterized by subtle humour and irony, have become a prominent digital communication medium. Detecting sarcasm in memes presents a significant challenge due to its context-dependent nature, negatively impacting user experiences on social media platforms. To improve the ability of social media systems to recognize and manage sarcastic content, this study investigates the effectiveness of Large Language Models (LLMs) for sarcasm detection in memes. Specifically, we evaluate three prompting techniques: Standard Prompt, Chain of Thought (CoT), and Concise Chain of Thought (CCoT) to determine their impact on the classification of sarcastic memes. Using the GOAT dataset as a benchmark, the study employs four pre-trained LLMs: Flan-T5-XXL, Llama-2, Mistral 7B, and GPT-2. The research identifies the most effective prompting strategies for sarcasm detection through a comparative analysis. The results demonstrate that CoT and CCoT significantly enhance performance over the Standard Prompt, with CCoT achieving the highest accuracy, particularly with advanced models like Mistral 7B. However, the choice of prompting technique depends on both the model and task requirements, emphasizing the need for tailored approaches in sarcastic meme analysis.
KW - Meme detection
KW - Prompting
KW - LLMs
UR - https://www.scopus.com/pages/publications/105000406595
U2 - 10.1007/978-981-96-1483-7_25
DO - 10.1007/978-981-96-1483-7_25
M3 - Conference proceeding contribution
AN - SCOPUS:105000406595
SN - 9789819614820
T3 - Lecture Notes in Computer Science
SP - 285
EP - 298
BT - Web Information Systems Engineering – WISE 2024 PhD Symposium, Demos and Workshops
A2 - Barhamgi, Mahmoud
A2 - Wang, Hua
A2 - Wang, Xin
A2 - Aïmeur, Esma
A2 - Mrissa, Michael
A2 - Chikhaoui, Belkacem
A2 - Boukadi, Khouloud
A2 - Grati, Rima
A2 - Maamar, Zakaria
PB - Springer, Springer Nature
CY - Singapore
Y2 - 2 December 2024 through 5 December 2024
ER -