https://neuroscirn.org/ojs/index.php/nrnotes A high quality, free open access and peer-reviewed journal from scientists to scientists. Neurotak Publishing (BC0008786-U) en-US Neuroscience Research Notes 2576-828X <p>The observations and associated materials published or posted by NeurosciRN are licensed by the authors for use and distribution in accord with the <a href="https://creativecommons.org/licenses/by-nc/4.0/" target="_blank" rel="external noopener">Creative Commons Attribution license CC BY-NC 4.0 international</a>, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.</p> https://neuroscirn.org/ojs/index.php/nrnotes/article/view/580 <p>Subtle sarcasm and contextual sarcasm are particularly important topics of research due to their nature as linguistic phenomena. Indeed, sarcasm detection is a complex NLP task because sarcasm conveys meaning beyond the literal one and depends heavily on context. Thus, it requires considering discourse-level features and long-term relations between statements. Machine learning and deep learning models often perform poorly in recognising sarcasm because they cannot effectively incorporate contextual cues and polarity shifts. To address this problem, this paper introduces an improved sarcasm detection framework based on the Reformer transformer model to improve efficiency in modelling long sequences and extracting semantic features. We evaluated the proposed model on benchmark datasets (news, social media texts, conversations), using accuracy, precision, recall, F1-score, and AUC-ROC as performance indicators. The results showed that the recommended model outperformed several other available frameworks (BiLSTM, BERT, RoBERTa, DistilBERT).</p> Ankur Srivastava Sifatullah Siddiqi Copyright (c) 2026 Ankur Srivastava, Sifatullah Siddiqi https://creativecommons.org/licenses/by-nc/4.0 2026-11-04 2026-11-04 9 4 580.1 580.12 10.31117/neuroscirn.v9i4.580