A Comparative Study of DeepSeek R1 (Reasoning) and GPT-4.0 (Non-Reasoning) Agentic AI Models for Irony and Sarcasm Detection in Urdu and English

Abstract

Artificial Intelligence (AI) systems often face difficulties in accurately identifying irony and sarcasm because these forms of figurative language convey meanings that differ from their literal expressions. Successful detection of such linguistic phenomena requires AI models to possess contextual understanding, cultural awareness, and the ability to interpret subtle linguistic cues. English is considered a high-resource language with extensive training data available for AI systems, whereas Urdu is a comparatively low-resource language with limited datasets. As a result, AI models generally perform better in English than in Urdu. Detecting sarcasm and irony across both languages remains challenging due to the linguistic and cultural complexities associated with Urdu. Furthermore, limited research has been conducted on comparing the performance of reasoning and non-reasoning AI models in this domain.

This study presents a comparative analysis of ChatGPT-4.0, a non-reasoning agentic AI model, and DeepSeek R1, a reasoning-based AI model, to evaluate their effectiveness in detecting irony and sarcasm in English and Urdu texts. A mixed-methods research design was adopted, combining quantitative performance evaluation with qualitative error analysis. The study utilized a total of 10,000 text samples, comprising 5,000 English and 5,000 Urdu instances. The English dataset included sarcastic and non-sarcastic content collected from Twitter and Reddit, while the Urdu dataset was based on the Urdu Sarcastic Tweets (UST) corpus and additional Urdu-language data. By comparing the performance of the two AI models across both languages, the study aims to provide insights into the strengths and limitations of reasoning and non-reasoning approaches in multilingual sarcasm and irony detection. The findings contribute to the growing field of Natural Language Processing (NLP) and highlight the challenges and opportunities associated with developing AI systems for low-resource languages such as Urdu.