This study aims to determine whether the divergence between context-free and context-sensitive emoji sentiment scores can serve as a complementary signal for sarcasm detection in Arabic. This phenomenon has received little systematic attention so far. To address this aim, we conducted a multi-scale empirical investigation. At the lexicon level, we compared 1,031 overlapping emoji from two Arabic sentiment resources: the context-sensitive Arab-ESL, derived from 144,196 tweets, and the context-free CF-Arab-ESL, annotated by 53 native speakers across seven Arab regions. At the corpus level, we examined the divergence pattern in a stratified random sample of 199,999 tweets drawn from the 1,414,523-tweet TEAD collection. At the task level, we developed a divergence-augmented transformer architecture by concatenating eight cross-lexicon features with the [CLS] representation produced by three Arabic BERT encoders: MARBERT, AraBERTv2, and CAMeLBERT. We evaluated it on the ArSarcasMoji dataset of 24,630 samples across five random seeds. The results indicate that only 56.8% of overlapping emoji maintain consistent sentiment labels across the two annotation paradigms (W = 91,029, p = 2.44× 10-74, Cohen's d = 0.656). The corpus-level analysis confirms that context-free scores are systematically more positive than their contextual counterparts (d = 1.124, p < 10-300). At the classification level, the divergence-augmented MARBERT achieves the highest mean F1-sarcasm of 0.5234 ± 0.015 across five random seeds, compared with 0.5149 ± 0.017 for the text-only baseline, and the augmentation produces a directionally consistent gain across all three encoders. Ablation analysis shows that continuous divergence scores contribute most of this improvement, whereas binary irony-risk indicators tend to introduce noise at low prevalence. The findings establish cross-lexicon emoji sentiment divergence as a measurable and pervasive feature of Arabic digital discourse, and indicate a modest but directionally consistent benefit when the divergence signal is supplied as a complementary feature for downstream sarcasm detection. A controlled emoji-aware comparison further shows that retaining emoji as input tokens allows encoders whose vocabularies cover emoji to recover the dataset construction rule almost perfectly (F1-sarcasm up to 0.998), which establishes the emoji-token-free protocol adopted here, in which no raw emoji tokens reach the text encoder while emoji-derived lexicon features are retained, as the methodologically sound setting for this corpus.
