Emojis as Affective Signals for Sarcasm Detection: An Empirical Analysis Informing EmoCentricSarcBERT and Its Application in the RADMAD Framework for Toxicity Mitigation in Online Discourse

Abstract The absence of paralinguistic cues in online discourse creates a contextual void, complicating classification of communicative intent such as sarcasm. Given the prevalent use of sarcasm to veil online toxicity, robust detection mechanisms are critical for effective content moderation. This thesis presents an integrated pipeline for an emoji-centric approach to sarcasm detection, transitioning from foundational statistical validation to the deployment of a resource-constrained moderation framework.

Initial hypothesis testing quantitatively validates that over 20% of online discourse utilizes emojis, establishing their viability as computational signals in text analytics. To address the inherent scarcity of emoji-rich sarcastic data, a novel dataset, SarcOji, is curated alongside two test sets: SarcOjiTest1 from existing benchmarks and SarcOjiTest2 from X (Twitter). Leveraging this data, novel Sarcasm-aware emoji embeddings are engineered to capture context from surrounding text based on word-emoji co-occurrences.

Empirical evaluations isolate three primary drivers required for emoji-centric sarcasm classification: explicit modeling of user intent via their most frequent emoji (MaxEmoji), contextual grounding through the proposed embeddings, and the integration of attention mechanisms. These drivers are subsequently unified within the EmoCentricSarcBERT architecture, which successfully captures deep bidirectional relationships between text and emojis to achieve robust F1, MCC, and ROC-AUC scores of 65.54%, 0.33, and 71.3% for SarcOjiTest1, and 53.33%, 0.35, and 74.18% for SarcOjiTest2, respectively.

This empirical success motivates the real-world deployment of EmoCentricSarcBERT, where knowledge distillation and quantization systematically compress this model into highly efficient low-resource architectures. This optimization culminates in the Resource-Aware Decentralized Moderation and Deployment (RADMAD) framework for sarcastic toxicity. Designed for low-resource edge devices, RADMAD functions as a proactive, nudge-based moderation system, dynamically calculating sarcasm scores to mitigate this toxicity at the source.

This deployable pipeline for real-time content moderation on resource-constrained devices yields ten core contributions: three datasets, Sarcasm-aware embeddings, EmoCentricSarcBERT, four compressed models, and the RADMAD framework.