Enhancing Sarcasm Detection with Contextual Factors for Improved Model Robustness Based on BMLRF

Detecting sarcasm in text remains a critical yet challenging task in natural language processing (NLP). Despite significant advances through deep learning, particularly the use of transformer-based architecture like BERT, sarcasm detection models still face challenges in achieving high accuracy. A major limitation lies in their insufficient incorporation of contextual awareness, including conversational history, social inter- actions, emotional cues, and cultural factors. To address this, this paper proposes the BERT with Meta-Feature Logistic Regression Fusion (BMLRF) model, which in- tegrates sentence-level embeddings from a pre-trained transformer with meta-features capturing social, emotional, and cultural contexts. The model also leverages multi-turn dialogue history and social interaction embeddings to enhance contextual understanding. This fusion approach aims to improve sarcasm detection robustness across domains such as social media and political discourse, emphasizing the critical role of emotion, social dynamics, and cultural context in sarcastic communication. Keywords: Natural Language Processing (NLP), BMLRF model, BERT, Multi-turn dialogue, Meta-features.