Sarcasm detectors trained on large, hashtag-labeled corpora look strong until they meet sarcasm as the writer actually meant it. On iSarcasm, a dataset where Twitter users labeled their own posts, models that scored F1 of 0.874 on hashtag benchmarks fell to at most 0.364 (iSarcasm). The likely reason is that these models learn surface cues tied to how a dataset was collected, not the mechanism of sarcasm itself. That mechanism is fairly simple to state. The literal words sound positive, the situation they describe is bad, and the reader has to notice the clash. This proposal builds a sarcasm detector around a single measured quantity: how strongly the literal tone conflicts with the described situation. A small, monotone calibration curve then maps that quantity to a probability.
The method comes from an unlikely source, constitutive modeling of fiber-reinforced concrete. That field fits compact laws relating stress to strain from a few hundred laboratory tests. It works with small data by compressing interacting factors into one physically meaningful index. In the UHPC case, the fiber reinforcement index is a product of fiber volume and fiber aspect ratio. A shape-constrained response curve with an explicit yield point is then fitted to that index. We borrow this recipe rather than the physics. The test is whether a detector built this way survives the shift from hashtag-labeled to author-labeled sarcasm better than fine-tuned transformers do. A second test is whether its fitted "yield threshold" gives a measurable account of why authors and outside readers disagree about what is sarcastic.
Core Claim
A sarcasm detector whose decision rests on one interpretable incongruity index, with a monotone calibration curve on top, will lose substantially less accuracy under label-regime shift (hashtag-labeled training, author-labeled testing) than a fine-tuned transformer. It should do this while staying competitive on in-distribution data. Fitting the same curve separately to author labels and to third-party labels on the same tweets will also give two yield thresholds, the index values at which predicted sarcasm crosses 50%. Their difference measures the intent–perception gap, which until now has been reported only as aggregate disagreement rates.
The index is
$$I(x) = L_+(x)\, S_-(x)\, \bigl(1 + \lambda\, m(x)\bigr), \qquad p(\text{sarcastic}\mid x) = g\bigl(I(x)\bigr),$$
where $L_+(x)$ is the strength of positive literal sentiment in tweet $x$, $S_-(x)$ is the negativity of the situation the tweet describes, and $m(x)$ is the intensity of rhetorical markers such as hyperbole, intensifiers, and punctuation. The weight $\lambda \ge 0$ is fitted. The function $g$ is constrained to be non-decreasing, and the yield point $I^*$ is the index value where $g(I^*) = 0.5$. The product form is deliberate. Sarcasm of the dominant iSarcasm type needs both a positive surface and a negative reality, so a large value of either term alone should not trigger it. This mirrors how the fiber reinforcement index multiplies dosage by aspect ratio because neither alone bridges cracks.
Research Program
The main technical work is measuring $S_-$, the sentiment of the situation rather than of the words. iSarcasm provides the supervision for this. For every sarcastic tweet, the author wrote a literal, non-sarcastic rephrasing (iSarcasm). The sentiment of that rephrasing is a direct measurement of intended polarity, much as a strain gauge reading measures a specimen. We train a small "situation sentiment" regressor to predict, from the tweet alone, the sentiment of its author's rephrasing. Non-sarcastic tweets serve as their own rephrasings. As a second instrument, an instruction-tuned LLM is prompted to describe the situation neutrally, without evaluative words, and the description is scored by an off-the-shelf sentiment model. $L_+$ comes from a standard sentiment classifier applied to the raw tweet, in the spirit of the sentiment-word clustering in Sarcasm SIGN. SIGN used sentiment to rewrite sarcasm into literal text; we use it to measure the gap that the rewrite would close. $m$ is a short lexicon-and-punctuation score, motivated by the Sarcasm-R1 ablation in which removing the linguistic dimension cost 7.6 macro-F1 points.
The calibration curve $g$ is fitted with isotonic regression or a monotone spline on the training split. Following the small-data practice of Explainable Prediction of UHPC Tensile Strength, we then use SHAP to report what share of the model's behavior each index component carries. That paper reached R² of 0.955 with support vector regression on only 178 tests, and attributed 37.5% of variance to the engineered fiber index. This is the methodological precedent: a well-chosen compound index lets a tiny model generalize. That paper also showed that linear Bayesian ridge regression on raw inputs failed. This constrains our design, because an additive linear model over raw sentiment features is not expected to capture the interaction. On Ptacek, where author history exists, we add an optional disposition term as a prior shift on $g$. Exploring Author Context showed that author history alone reaches F1 of 0.922 there, so this term cannot be ignored on that corpus, but it should be reported separately.
The multi-regime constitutive curve for PP-SACC, with elastic, first-crack, hardening, and pull-out softening stages (Xie et al.), suggests letting $g$ have a flat region below a threshold and a saturating region above it. We test this shape against plain logistic calibration rather than assuming it. Mechanics supplies the hypothesis about the shape; the data decide.
Baselines
The naive reference point is a majority-class or prior-ratio classifier. Since iSarcasm is roughly 1:5 sarcastic, any reported accuracy must be read against this prior. We therefore treat sarcastic-class F1 and macro-F1 as primary.
The first strong baseline is a fine-tuned encoder classifier: RoBERTa-large, plus the T5 encoder with Bi-LSTM and attention architecture of MarSan, which reported English sarcastic F1 of 58.89% on SemEval-2022 iSarcasmEval subtask A. It is run on the official train/test split with the same development-set tuning budget as our model, measured in GPU-hours and number of hyperparameter trials. In the transfer setting it is trained on Ptacek and tested on iSarcasm. It is the most relevant competitor because the transfer claim is specifically that such models collapse.
The strongest existing method is Sarcasm-R1. It is a Qwen2.5-7B model trained with group-relative policy optimization against three reward models (linguistic, contextual, emotional), and reports 85.4% accuracy and 72.3% macro-F1 on iSarcasm, against GPT-4o at 76.0% and 59.4%. Its evaluation is on the authors' Sarcasm-Reason construction, so we will re-run it on the official iSarcasmEval test split if weights are available. If they are not, we quote the published number with that caveat and run GPT-4o-style zero-shot and chain-of-thought prompting ourselves on the identical split. Beating Sarcasm-R1 in-distribution is not required for the claim. The bar is a smaller degradation in the hashtag-to-intended transfer setting, plus in-distribution macro-F1 within a few points of the best fine-tuned encoder.
Three ablations isolate the key ingredient. The first replaces the product index with an additive linear combination of the same three components. The second replaces the monotone curve with unconstrained gradient boosting over the raw components $L_+$, $S_-$, and $m$. This is the reviewer's tabular alternative and the direct analogue of Zhang et al.. The third feeds $L_+$ and $S_-$ as extra inputs to the fine-tuned RoBERTa. If any of these matches the full model under transfer, the compression into one index is not what helps.
Why not just…?
Why not just fine-tune a larger transformer or prompt a frontier LLM? Under in-distribution evaluation this may well win, and Sarcasm-R1's 72.3 macro-F1 is a high bar. The evidence that it falls short concerns robustness rather than peak score. Hashtag-trained detectors dropped from 0.874 to at most 0.364 on author-labeled data (iSarcasm), and Sarcasm-R1's authors report that unconstrained chain-of-thought led GPT-4o to read surface tokens literally. Our transfer experiment settles this directly. If fine-tuned RoBERTa degrades as little as the index model, the proposal's main claim is false.
Why not just add the incongruity features to the transformer? This is the most serious threat. It keeps the measurement idea while discarding the low-dimensional bottleneck. If the transformer-plus-features ablation transfers as well as the index model, the honest conclusion is that the measured situation sentiment is the useful contribution and the calibrated curve is not. We would report it that way. The bottleneck version retains one advantage that is not about F1: an explicit yield threshold that can be compared across label sources, which the hybrid model does not provide.
Why not just run logistic regression or gradient boosting on the components? Logistic regression on a scalar index is already a monotone threshold model. If it matches the shaped curve, we adopt it and drop the multi-regime shape, keeping the product index. Gradient boosting can learn the interaction without our engineered product. On the material side, though, engineered indices still carried the small-data result even for tree ensembles, and CatBoost needed 1208.4 seconds of tuning (Zhang et al.). The ablation reports transfer gaps and fitted interaction plots for both.
Proposed Experiments
The in-distribution experiment trains all systems on iSarcasmEval English. It reports sarcastic-class F1, macro-F1, and expected calibration error on the official test split, alongside MarSan's 58.89 F1 and Sarcasm-R1's 72.3 macro-F1. The decisive experiment trains on Ptacek (hashtag-labeled, about 27,000 tweets) and tests on iSarcasm, and the reverse where data size allows. The quantity of interest is the drop in sarcastic F1 relative to in-distribution performance. Success means the index model's drop is at most half that of fine-tuned RoBERTa, with non-overlapping bootstrap confidence intervals.
The intent–perception experiment uses the iSarcasm test tweets that carry both author labels and third-party labels. Third-party annotators reached only F of 0.616 against author labels, missing 30% and over-flagging 45% (iSarcasm). We fit $g$ separately to each label source and estimate $I^*_{\text{author}}$ and $I^*_{\text{audience}}$ with bootstrap intervals. We also inspect whether audience false positives cluster at high marker intensity $m$ with low $S_-$, meaning readers react to tone rather than situation. The same comparison on Riloff (perceived, 701 tweets; inclusive author model F1 0.739 against 0.711 baseline in Exploring Author Context) checks whether the perceived-label threshold replicates across corpora.
A secondary, zero-shot check targets subtask B, where MarSan scored 7.43 macro-F1 and 0.00 on understatement and overstatement. Without training a category classifier, we test whether the ratio of literal intensity to situation intensity separates overstatement (literal far exceeds situation) from understatement (literal well below situation) better than chance. That would show the index components carry category information that data-starved classifiers cannot learn from one or ten examples.
The proposal fails cleanly if three things hold together: transfer gaps are indistinguishable from the fine-tuned encoder's, the transformer-plus-features ablation matches the index model, and the two yield thresholds have overlapping intervals. Each outcome is informative on its own terms.
References
- iSarcasm: A Dataset of Intended Sarcasm
- Sarcasm SIGN: Interpreting Sarcasm with Sentiment Based Monolingual Machine Translation
- Sarcasm-R1: Enhancing Sarcasm Detection through Focused Reasoning
- Explainable Prediction of UHPC Tensile Strength Using Machine Learning with Engineered Features and Multi-Algorithm Comparative Evaluation
- Exploring Author Context for Detecting Intended vs Perceived Sarcasm
- An experimental study and axial tensile constitutive model of the toughness of PP-SACC for rapid repairs
- MarSan at SemEval-2022 Task 6: iSarcasm Detection via T5 and Sequence Learners