We describe the Seals-NLP system for SemEval-2026 Task 9 (POLAR) Subtask 1, binary polarization detection. Our study compares (i) fully fine-tuned encoder-only transformers, (ii) QLoRA-based fine-tuned open-weight LLMs, and (iii) zero-shot prompted LLMs. ModernBERT-large emerges as the most cost-effective option, matching or surpassing larger fine-tuned and zero-shot LLMs in macro-F1 while requiring substantially less memory and lower latency. An error analysis by failure mode and polarization subtype reveals systematic over-triggering on political cue words and under-detection of sarcastic vilification and multifaceted attacks in the POLAR dataset across all models.

The intended sarcasm cannot be understood until the listener observes that the text’s literal meaning violates truthfulness. Consequently, words and meanings play an essential role in specifying sarcasm. Enriched feature extraction techniques were proposed to capture both words and meanings in the contexts. Due to the overlapping features in sarcastic and non-sarcastic texts, a CNN model extracts local features from the combined class-dependent statistical embedding of sarcastic texts with contextualized embedding. Another component BiLSTM extracts long dependencies from combined non-sarcastic statistical and contextualized embeddings. This work combines a classifier that uses the combined high-level features of CNN and BiLSTM for sarcasm detection to produce the final predictions. The experimental analysis presented in this paper shows the effectiveness of the proposed method.