Who Is Being Mocked? Sarcasm Detection and Entity Linking as One Fuzzy Fixpoint

Entity linking systems decide which real-world entity a name refers to ("Jordan" the country, the basketball player, or the river) by trusting the surrounding words. In sarcastic text that trust is misplaced. "Great season, Eagles, truly a dynasty" uses triumphant words about a team the writer thinks is failing. A linker reads the praise as evidence, and a sarcasm detector misses the cue, because it treats "Eagles" as a bare string and does not know that this team just finished last. Each problem needs the other's answer. Sarcasm can only be recognised once we know which entity is meant and what is true of it. The right entity can only be chosen once we know whether the context should be read literally or flipped.

We propose to treat this circular dependency as a cycle to be solved, not as a pipeline ordering problem. Sarcasm degree, literal-reading degree, and per-candidate link degrees are written as a small system of mutually recursive fuzzy definitions, meaning statements whose truth values lie in $[0,1]$ instead of being true or false. We take the system's greatest fixed point: the largest assignment of degrees that is consistent with itself. Under min/max (Gödel) fuzzy semantics this fixed point exists and is reached exactly in a finite number of steps. The result is a joint, inspectable decision about who is being mocked and whether they are. We test it on a new entity-annotated layer over the iSarcasm tweets, where authors themselves supplied literal rephrasings.

Core Claim

The claim is that one specific class of tweets is mishandled by text-only sarcasm detectors, by literal-context entity linkers, and by one-directional pipelines between them. These are tweets whose sarcasm is detectable only through facts about a correctly disambiguated entity. A monotone fuzzy fixpoint over the joint system is the minimal machinery that handles this class correctly while staying exact and terminating.

The specific new element is the cycle. It is neither the fuzziness alone nor the knowledge base alone. If a one-pass or two-pass pipeline recovers the same gains, the claim is false, and the experiments are designed to show that.

This also bears on an unresolved disagreement in the literature. CSDGCN adds commonsense sentiment inferences from COMET to dependency graphs and reports gains on three Twitter benchmarks. Does Commonsense help in detecting Sarcasm? finds that COMET inferences added through a graph network do no better than DistilBERT, because generated knowledge on literal sentences makes the model see sarcasm that is not there. In our system, external knowledge enters only in proportion to how confidently an entity is linked. A literal tweet with no confidently linked entity therefore receives almost no incongruity evidence. We predict this lowers false positives on non-sarcastic tweets, and that prediction is tested directly.

The Mechanism

Take a tweet with an ambiguous mention whose candidate entities are $e_1,\dots,e_k$, retrieved from Wikipedia/Wikidata. For each candidate the system computes fixed input degrees:

  • $p(e)$ is a prior/context link score from a standard neural linker.
  • $c(e)$ is the congruity of the tweet's literal claim with the entity's known facts, scored by an entailment model over the entity's description and properties.
  • $\bar c(e)$ is the congruity of the inverted claim with those facts. The inverted claim is a sentiment-flipped paraphrase produced by a sarcasm interpreter in the spirit of Sarcasm SIGN, which translates sarcastic text into literal text.

The unknowns are the link degrees $\ell(e)$, a sarcasm degree $s$, and a literal-reading degree $t$. They are defined in terms of each other, with $\wedge$ as min and $\vee$ as max:

$$\ell(e) = p(e) \wedge \big[(t \wedge c(e)) \vee (s \wedge \bar c(e))\big], \qquad s = \bigvee_{e} \big(\ell(e) \wedge \bar c(e) \wedge (1-c(e))\big), \qquad t = \bigvee_{e} \big(\ell(e) \wedge c(e)\big).$$

The equations say three things in plain terms. A candidate is linked to the degree that it is plausible and fits whichever reading (literal or inverted) is currently supported. A tweet is sarcastic to the degree that some well-linked entity contradicts the literal claim and fits the inverted one. A tweet is literal to the degree that some well-linked entity fits the literal claim.

Literal and sarcastic readings are deliberately kept as two separate positive variables instead of setting $t = 1-s$. Negation of an unknown would make the system non-monotone, and its fixed point could then fail to exist or fail to be unique. Negation is applied only to the fixed inputs $c(e)$. Every operator on the unknowns is therefore min or max, so the system is monotone.

Monotonicity gives a greatest fixed point by the Knaster–Tarski theorem. Iterating downward from the all-ones assignment reaches it. Under min/max, every iterate takes values only from the finite set of input constants together with 0 and 1, and the iterates only decrease. The iteration therefore stops after at most (number of unknowns) × (number of distinct constants) steps. The final output is sarcastic if $s > t$, along with the entity of highest $\ell(e)$.

Three properties of this design are worth stating.

First, it follows Baader's observation that greatest-fixpoint semantics are the natural reading of self-referential definitions. It applies that observation to degrees, mirroring how cyclic $\mathcal{EL}$ definitions under greatest-fixpoint semantics make otherwise ill-defined constructions well defined and computable (Least Common Subsumers and Most Specific Concepts...).

Second, it is deliberately scoped to stay decidable. Syntactic Labelled Tableaux for Łukasiewicz Fuzzy ALC shows that general cyclic axioms under Łukasiewicz semantics are undecidable, and A Fuzzy Description Logic leaves cyclic fuzzy terminologies open. We avoid both by grounding the cycle in a finite per-tweet system and choosing Gödel connectives. We are not claiming to solve cyclic fuzzy terminologies in general; this is an instance-level fixpoint.

Third, it differs from the closest fuzzy sarcasm model, CFN. CFN uses a learned sigmoid membership function to weight dialogue context inside a complex-valued network. It uses no formal connectives, no entity grounding, and no recursion.

The greatest fixed point rather than the least is a substantive choice. Starting from all zeros, the system has no base support and stays at zero unless extra base facts are added. Starting from the top, the iteration keeps every jointly consistent reading and removes only what lacks support. Comparing the two, with an lfp variant that has explicit seeds, is one of our ablations.

Research Program

There are three pieces of work.

The first is a dataset. iSarcasm (iSarcasm: A Dataset of Intended Sarcasm) contains 4,484 tweets in which sarcasm is labelled by the authors themselves, and each sarcastic tweet comes with the author's own literal rephrase. We will annotate every named-entity mention with a Wikipedia link. The rephrase helps annotators resolve who is meant, and it later serves as an oracle inverted claim. Annotators will also mark whether the sarcasm cue depends on facts about the entity. This tag defines the "entity-dependent" slice on which the core claim lives. We will add the same entity layer to a subset of the Riloff and Ghosh tweets, and evaluate pure linking on the 502-tweet benchmark of Entity Linking for Tweets.

The second is the input scorers. For $p(e)$ we use a BERT linker trained with hard negatives, as in Deep Entity Linking via Eliminating Semantic Ambiguity With BERT, where training negatives are candidates with near-identical names but different referents. For $c$ and $\bar c$ we use an off-the-shelf entailment model applied to entity facts. For the inverted claim we use a sentiment-flipping paraphraser.

The third is the solver and its analysis. This covers convergence statistics, the trace of how $s$, $t$, and $\ell$ evolve during iteration, and a per-example explanation naming the entity fact that drove the decision.

Baselines

The naive reference points are the most-popular-candidate linker (the popularity bias documented on WhoQA in Who's Who) and a text-only fine-tuned encoder for sarcasm. Both are trained on our splits.

For linking, the strongest existing method is the hard-negative BERT linker of Yin et al. It reports 95.04% micro-accuracy on CoNLL 2003 and 90.34% on TAC 2010. On tweets, the collective linker of Liu et al. reports F1 0.711. We found no reported result on Lacuna for any linker on sarcastic tweets, which is the gap our dataset fills. Beating the linker means a significant accuracy gain on the sarcastic slice of iSarcasm with no loss on the literal slice or on the Liu benchmark.

For sarcasm, the strongest reference on iSarcasm is MarSan's T5 system (MarSan at SemEval-2022 Task 6), with F1 58.89% on English Subtask A. The iSarcasm paper reports a maximum F-score of 0.364 for detectors trained on hashtag-labelled data. CSDGCN reports 87.16% accuracy on Riloff. All of these are retrained on identical splits, with the same backbone size and an equal hyperparameter search budget. Beating them means higher sarcastic-class F1 on iSarcasm overall and a large gain on the entity-dependent slice.

The ablations isolate the cycle and its components:

  • A one-pass system computes $s$ from priors alone, then links.
  • A two-pass pipeline detects sarcasm, then links with the inverted context.
  • A learned joint multitask model shares one encoder for both tasks.
  • A product t-norm version of the same equations, iterated to tolerance, tests whether Gödel exactness matters.
  • An lfp-with-seeds variant tests the fixpoint choice.
  • A variant without entity gating feeds knowledge evidence regardless of $\ell$, which mimics the ungated setting that produced noise.

Why not just…?

Why not just run a two-pass pipeline, sarcasm first and then linking? This is the most serious threat. Our fixpoint is in effect that pipeline iterated until it stabilises. The difference should appear only where the first sarcasm guess needs entity facts, which is the annotated entity-dependent slice. If the two-pass system matches us there, the cycle adds nothing and we will report that. We also report how often the iteration actually needs more than two rounds.

Why not just prompt a large language model to do both? Who's Who shows that unflagged entity ambiguity drops Llama 3 8B from 98.2% to 72.4% accuracy, with strong bias toward popular entities. Sarcasm is exactly the setting in which the context is also misleading. We will nonetheless run zero- and few-shot LLM baselines on our dataset under the same evaluation. Whether they collapse on the sarcastic slice is an empirical question this study answers.

Why not just append the entity's description to the sarcasm classifier's input? This is cheap and could work. It needs the entity to be chosen first, and the choice is made from misleading context. We include an oracle version that appends the gold entity and a predicted version that appends the top prior candidate. The gap between the two measures how much of the problem is the linking error. If the predicted version matches us, the fixpoint is unnecessary.

Why not just learn a joint probabilistic model end to end? A multitask encoder is in the baseline set. It may match accuracy. What it cannot give is a guarantee of termination or an explicit explanation of which entity fact flipped the decision. If it wins on accuracy, we would claim only interpretability, and we would say so.

Decisive Checks

The central test is accuracy on the entity-dependent slice of iSarcasm. Sarcasm F1 and linking accuracy are both reported with bootstrap confidence intervals against the two-pass pipeline and the entity-appended classifier. The core claim survives only if the fixpoint beats both there. A second check uses the author rephrase as the inverted claim. If the gains disappear when the predicted paraphrase replaces the oracle, the bottleneck is interpretation and not the cycle. A third check is the false positive rate on non-sarcastic tweets, gated against ungated, which tests whether gating knowledge by link confidence removes the noise reported by Roy Chowdhury and Chaturvedi. The final check compares Gödel and product connectives, reporting iteration counts and accuracy, to confirm that exact finite convergence comes without an accuracy cost.