Answers are grounded in this paper and cite their sources.
Sarcasm detection is a challenging problem in computational linguistics because the sentiment expressed on the surface of a sarcastic utterance frequently contradicts the sentiment its author intends. Existing detectors are also brittle: models that perform well on one platform typically degrade sharply when applied to text from a different domain or writing style. This paper proposes a context-aware hybrid model that couples a pre-trained BERT encoder with a bidirectional LSTM and an additive attention layer, and fuses the resulting contextual representation with two complementary signals — an unsupervised topic-context vector obtained from BERTopic and a set of nine handengineered linguistic cues. The model is trained only on a labelled Twitter corpus and evaluated in two regimes: an in-domain held-out Twitter test set and a zero-shot cross-domain test on a balanced, formally-written News-Headlines corpus that the model never sees during training. Against four classical TF-IDF baselines and pure transformer baselines evaluated under an identical protocol, the proposed hybrid attains the strongest in-domain results, reaching an accuracy of 0.8695 and an F1-score of 0.7945, and the zero-shot cross-domain evaluation quantifies the extent to which a tweet-trained detector generalizes to a different register. The results indicate that combining a contextual transformer encoder with unsupervised topic context and linguistic cues is an effective and transferable design for sarcasm detection.
No takes yet. Share an insight, caveat, or question.
Nikam et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: