Sarcasm usually works by praising something everyone knows is bad: "I love being stuck in traffic." A model can only catch this if it knows that traffic is unpleasant, and that knowledge is not in the sentence. Adding outside commonsense knowledge to sarcasm detectors has given mixed results. The most careful negative study found that the knowledge was mostly ignored, and on sincere sentences it sometimes misled the model. Our diagnosis is that these systems ask the knowledge source the wrong question. They generate commonsense about the whole sentence, or about every phrase in it, when the only question that matters is a narrow one: what is the speaker evaluating, and how do people normally feel about that thing?
In morphologically rich languages, the answer to "what is being evaluated" is often hidden inside words. In Arabic, a single written token can carry a verb, its subject, and its object as attached pronouns. Word order is loose, and gender and number agreement signal which adjective describes which noun. We propose to use explicit morphology to locate the evaluation and its target, meaning attached pronouns and agreement between a head word and its dependent. We then query commonsense knowledge only about that target, and feed the detector the clash between the expressed attitude and the expected one. The test bed is author-labelled ("intended") sarcasm in Arabic, where the best shared-task systems are weak, and English serves as a control.
Core Claim
Commonsense knowledge helps sarcasm detection when it is anchored to the specific evaluation–target pair in the utterance, and hurts or does nothing when it is not. In Arabic, morphological analysis finds that pair more reliably than the space-delimited word graphs used by current systems. We therefore predict two things. First, anchored knowledge will beat unanchored knowledge in both languages. Second, the advantage of morphological anchors over plain word-level anchors will be clearly larger in Arabic than in English, and largest on tweets where the evaluative word and its target share a single orthographic token.
Where This Sits
Three lines of prior work motivate the gap.
The first is the negative result. Does Commonsense help in detecting Sarcasm? added generated intents and effects from COMET, a model trained to produce everyday "if X, then people feel Y" inferences, as flat nodes in a graph network. The result matched DistilBERT: 96.13% versus 96.18% accuracy on News Headlines. Hiding the commonsense changed confidence by only 1.38%, compared with 27.99% for hiding the sentence. We read this as evidence that unanchored knowledge is noise, and we treat it as the hypothesis to beat rather than a dead end.
The second line is the positive results, which used structure to place knowledge. Commonsense Knowledge Enhanced Sentiment Dependency Graph for Sarcasm Detection substitutes COMET sentiment for particular nodes in an English dependency tree and reports 87.16% accuracy on Riloff. It explicitly fails on situational sarcasm. Detecting Emotional Incongruity of Sarcasm by Commonsense Reasoning (EICR) retrieves web passages, has a large language model write background knowledge, prunes a dependency graph, and reports 83.3% accuracy / 70.4% macro-F1 on English iSarcasm, beating GPT-4o prompting. The affective-graph models Affective Dependency Graph for Sarcasm Detection and Augmenting Affective Dependency Graph via Iterative Incongruity Graph Learning for Sarcasm Detection link emotionally distant words along spaCy parses. The latter treats arcs as untyped and undirected, and notes that typed relations are needed to stop contrasting adjectives on different heads from triggering false alarms.
All four systems build graphs over English whitespace tokens. None uses morphology. They also merge two things we separate: deciding where to ask a question, and how to fuse the answer. The pipeline in EICR is heavy. Our anchoring step replaces most of that pipeline with a single targeted query per tweet.
The third line supplies the method. Getting More from Morphology in Multilingual Dependency Parsing showed that arc-level agreement indicators, such as matching gender and number between head and dependent, gave up to 5.3% absolute unlabelled-attachment gains across 20 treebanks. Multilingual Dependency Analysis with a Two-Stage Discriminative Parser used morphological features to rule out ungrammatical arcs. Improved Arabic Base Phrase Chunking with a new enriched POS tag set found that packing definiteness, gender and number into a compact tag set beat feeding them as loose features. The same paper found that adding more dimensions such as case and mood hurt, so we restrict ourselves to the agreement features that bind evaluations to targets.
Tsarfaty's work, Integrated Morphological and Syntactic Disambiguation for Modern Hebrew and A Unified Morpho-Syntactic Scheme of Stanford Dependencies, constrains the design in two ways. First, attached pronouns and standalone pronouns should be treated as the same relation. Second, pipelines that segment first lose accuracy on raw text: 89.8% on gold segments versus 81.3% on raw text. This is why we measure anchor accuracy directly rather than assume it.
On the application side, Arabic sarcasm systems such as ArSarcasm Shared Task: An Ensemble BERT Model for SarcasmDetection in Arabic Tweets use only subword transformers and decoded emojis, with no syntax and no outside knowledge. SemEval-2022 Task 6: iSarcasmEval, Intended Sarcasm Detection in English and Arabic reports a top Arabic F-sarcastic of only 0.563 and names missing world knowledge as a main limitation. iSarcasm: A Dataset of Intended Sarcasm shows why the intended-sarcasm setting is the honest one: models scoring 0.874 on hashtag-labelled data fall to 0.364 there.
Research Program
The system has three parts.
The anchor extractor. Each tweet is segmented and morphologically analysed into stems, attached pronouns, and gender, number, person and definiteness. The file indexes no dialectal Arabic analyser, so choosing one is part of the work. We will report its error rate on our data rather than assume formal-text accuracy.
We then build a small typed graph over morphemes rather than words. Its arcs are scored with Hohensee-style agreement indicators and typed relations: subject, object, modifier, and attached-pronoun argument. Candidate evaluative items are morphemes with strong polarity in an Arabic sentiment resource or a learned polarity head. The anchor is the evaluative item plus the argument it agrees with or governs, expanded to its phrase. This argument is the "evaluated situation." For example, in a dialectal tweet meaning roughly "how wonderful, they cut-it-on-us again," the evaluation is "wonderful," and the target is the verb with its attached object and beneficiary pronouns, which a whitespace graph would treat as one opaque token. The extractor may abstain when no agreement-consistent pair exists.
The targeted knowledge query. For the evaluated situation only, a fixed language model, optionally with retrieval, answers one constrained question: how do people typically feel when this happens to them? It returns a short answer and an expected polarity. Sincere tweets with contrasting words, such as "bad memory but great battery," yield two anchors whose expected polarities agree with their expressed ones, so no clash is produced. This is the mechanism by which anchoring should remove the false alarms that flat injection created.
The detector. A fine-tuned MARBERT (English: the same backbone family used by the strongest English entrants) receives the tweet. The knowledge answer is injected through a gate that may attend only to anchor tokens. A scalar clash feature is computed as
$$c = -\,p_{\text{expr}}(e)\,p_{\text{exp}}(s),$$
where $p_{\text{expr}}(e)\in[-1,1]$ is the polarity the speaker expresses toward evaluation $e$ and $p_{\text{exp}}(s)\in[-1,1]$ is the knowledge-derived expected polarity of situation $s$. The feature is positive when praise meets a normally bad situation. When the extractor abstains, the gate is closed and the model reduces to the plain encoder. This guarantees that anchoring cannot do worse than the baseline on tweets without a clear target.
The diagnostic set. We will hand-annotate gold evaluation–target pairs for about 500 Arabic and 500 English iSarcasmEval tweets. This set lets us separate anchor-finding errors from knowledge errors and from fusion errors.
Baselines
The naive reference point is the trivial predictor and the organisers' baselines reported in SemEval-2022 Task 6: iSarcasmEval, Intended Sarcasm Detection in English and Arabic. For English Subtask A these reach F-sarcastic 0.348 with BERT and 0.275 with SVM. Any knowledge method that fails to clear a plainly fine-tuned encoder by a wide margin over these is uninteresting.
The strongest existing competitor on the Arabic question is a fine-tuned Arabic transformer. The best reported Arabic Subtask A result is F-sarcastic 0.563 from the top team (CS-UM6P) in the same task paper. On ArSarcasm-v2, the AraBERTv02 + Sentence-BERT ensemble of ArSarcasm Shared Task: An Ensemble BERT Model for SarcasmDetection in Arabic Tweets reports F1-sarcastic 59.89 and accuracy 78.30%. We will rerun a MARBERT fine-tune ourselves on identical splits and seeds, with the same hyperparameter search budget as our model. The reproduced number, not only the published one, is the bar.
For English, the strongest knowledge-based competitor is EICR at 70.4% macro-F1 on iSarcasm, as reported in Detecting Emotional Incongruity of Sarcasm by Commonsense Reasoning. That figure is on the original iSarcasm data, not the iSarcasmEval test set. It is therefore a reference rather than a like-for-like target, and we will rerun a simplified EICR under our splits. For the Arabic split, no reported result was found on Lacuna for any commonsense-augmented model.
The ablations that isolate the key ingredient all use the same encoder, the same knowledge model, and the same number of knowledge calls per tweet. We will compare four sources of knowledge placement:
- knowledge generated for the whole tweet and concatenated, which replicates the flat setting of Does Commonsense help in detecting Sarcasm?;
- anchors from a word-level dependency graph without morphology, mirroring the existing English graph models;
- anchors chosen at random among content words;
- gold anchors from the diagnostic set, as an upper bound.
Beating the field means two things on Arabic Subtask A. First, exceeding our reproduced MARBERT and the 0.563 reported best with a gap larger than seed variance. Second, morphological anchors beating word-level anchors by a margin that is significant under paired bootstrap, while the same gap in English is smaller.
Why not just…?
Why not just fine-tune a larger Arabic encoder? Subword tokenisers do split off many attached pronouns, so a bigger MARBERT may already find the target implicitly. What it cannot do is know that the target is unpleasant when nothing in the tweet says so. Deep contextualized word representations for detecting sarcasm and irony argues that contextual encoders capture how something is said rather than what is true. The Arabic ceiling of 0.563 among transformer-based entrants is consistent with that. Our gate design makes the encoder-only model a strict special case. If the gain disappears as encoder size grows, the scaling curve we run on two backbone sizes will show it.
Why not just concatenate generated commonsense to the input? This is the setting that already failed in Does Commonsense help in detecting Sarcasm?, where the model learned to ignore the extra text. That study used COMET with a flat graph and an older encoder, and a modern LLM-generated paragraph might fare better. So we run concatenation as ablation 1 with a matched call budget, and repeat the occlusion analysis from that paper on both variants. Our claim is falsified if concatenation matches anchoring.
Why not just have a large language model find the anchor, or answer directly? This is the real threat. An LLM can plausibly identify "what is being praised" without any morphological analyser, and do so more robustly on dialect than a rule-and-agreement extractor trained on formal treebanks. We therefore add an LLM-extracted-anchor condition and score all anchor sources against gold anchors on the diagnostic set. If the LLM anchors match morphological anchors in accuracy and downstream F-sarcastic, then the contribution reduces to "anchor the query." We would report morphology's role as cost and interpretability only, not accuracy. Direct end-to-end LLM prompting is also run. EICR's report that a trained system beat GPT-4o prompting suggests it will not dominate, but we will not assume this for Arabic.
Proposed Experiments
The main experiments use iSarcasmEval Subtask A in Arabic and English with author labels, plus ArSarcasm-v2 for a larger Arabic training signal. Twitter-Riloff and Ghosh are reported only for continuity with the graph-model literature, given their hashtag-label bias. Metrics are F-sarcastic, the official measure, plus macro-F1. Results are averaged over five seeds with paired bootstrap significance.
The first decisive check is the interaction test. We predict a larger morphological-minus-word-level gap in Arabic than in English. We further predict it will be concentrated in the subset where evaluation and target share one orthographic token, a subset we can identify automatically from the analyser. Absence of this interaction would mean morphology is not the operative ingredient.
The second check uses Subtask C, where each sarcastic tweet is paired with the author's sincere rephrase. The rephrase keeps the situation and changes the evaluation. Anchored knowledge should therefore return the same expected polarity for both members, and the decision should flip only through the expressed polarity. We will measure how often this holds and compare pairwise accuracy against the reported tops of 0.930 (Arabic) and 0.870 (English).
The third check is a false-alarm audit on sincere tweets containing contrasting sentiment words. It tests whether anchoring removes the hallucinated-sarcasm errors that flat injection caused.
The final check uses the gold-anchor upper bound to split the remaining error between anchor-finding and knowledge quality. This tells us whether the field's next investment should go to morphological analysis of dialect or to better world knowledge.
References
- Does Commonsense help in detecting Sarcasm?
- Commonsense Knowledge Enhanced Sentiment Dependency Graph for Sarcasm Detection
- Detecting Emotional Incongruity of Sarcasm by Commonsense Reasoning
- Affective Dependency Graph for Sarcasm Detection
- Augmenting Affective Dependency Graph via Iterative Incongruity Graph Learning for Sarcasm Detection
- Getting More from Morphology in Multilingual Dependency Parsing
- Multilingual Dependency Analysis with a Two-Stage Discriminative Parser
- Improved Arabic Base Phrase Chunking with a new enriched POS tag set
- Integrated Morphological and Syntactic Disambiguation for Modern Hebrew
- A Unified Morpho-Syntactic Scheme of Stanford Dependencies
- ArSarcasm Shared Task: An Ensemble BERT Model for SarcasmDetection in Arabic Tweets
- SemEval-2022 Task 6: iSarcasmEval, Intended Sarcasm Detection in English and Arabic
- iSarcasm: A Dataset of Intended Sarcasm
- Deep contextualized word representations for detecting sarcasm and irony