When someone reposts a news story or quotes another user's post, most network analyses count that as agreement. A large share of reposts are the opposite: people quote their opponents in order to mock them, debunk them, or pile on. In English, fine-tuned text classifiers already fail on exactly these cases. The error analysis in A Few Topical Tweets are Enough for Effective User Stance Detection names "quoting the other side" and sarcasm as two of its four main error classes. In low-resource and code-mixed languages the situation is worse. There is no labeled data for hostile sharing at all, sarcasm detectors barely work, and machine translation tends to turn ironic praise into literal praise. The result is that stance maps, echo-chamber measurements, and misinformation-spread estimates built on "share = endorse" are silently wrong in exactly the communities that have the fewest tools.
We propose to treat hostile sharing as a two-text relational judgment: does the sharer's added comment endorse or oppose the content it quotes? We attack it with two ingredients that address the two separate bottlenecks. The first bottleneck is missing labels. We obtain silver training labels in the target language from the social network itself. For heavy users, the stance a user shows across their posting history reveals whether they agree with the author they quote, even when the comment's wording says nothing explicit. The second bottleneck is reading. A large language model writes a short, fixed-slot rationale in English that separates what the quoted content claims, what the comment literally says, and whether the comment's meaning is inverted. A small fine-tuned encoder then reads the post pair together with that rationale and makes the final call. Chain-of-thought prompting, meaning asking a model to write out intermediate reasoning before answering, is used here in a deliberately constrained form. The prior evidence is that free-form reasoning over-reads hostility in informal speech.
Core Claim
Hostile-share errors in low-resource languages are concentrated in pairs where the comment's literal tone disagrees with the sharer's actual relation to the quoted content, which is the sarcasm case. The claim has two parts. First, a slot-structured rationale that forces the model to state literal tone and inversion separately fixes most of these errors when a discriminative encoder consumes it. Free-form reasoning does not; it raises false "hostile" predictions on informal, heated endorsements. Second, silver labels derived from users' posting histories let the whole system be trained in a target language with no hand-annotated hostile-sharing data. The new element is the combination of relational (pair-level) supervision drawn from the network with an explicit literal-versus-intended decomposition. We test it where the literature has none: code-mixed Bangla-English and dialectal Arabic.
Relation to Prior Work
Three papers supply the method and one supplies the main constraint. Chain-of-Thought Embeddings for Stance Detection on Social Media concatenates a tweet with a one-shot ChatGPT reasoning chain and fine-tunes Twitter-RoBERTa on the result. It reaches 76.3 F1 on Tweet-Stance, 6.1 points above ChatGPT's own chain-of-thought labels, because the encoder ignores the model's sometimes self-contradictory final label. We reuse this "rationale as input feature" design. It is English-only, single-text, and uses free-form reasoning, whereas we work on pairs, in low-resource languages, and with a fixed slot structure.
Intermediate-Task Transfer Learning: Leveraging Sarcasm Detection for Stance Detection shows that pre-training on sarcasm before stance training rectifies 85% of sarcastic "Against" posts that standard classifiers call "InFavor." It also shows that the sarcasm source must match the target platform's style. This motivates our focus on inversion. It also becomes a baseline, since it is the obvious way to inject sarcasm knowledge without any reasoning chain.
A Reinforcement Learning Framework for Cross-Lingual Stance Detection Using Chain-of-Thought Alignment uses high-resource-language reasoning traces as a scaffold for target-language stance, followed by policy-gradient training. It reports gains of up to 12.8% average F1 over plain chain-of-thought prompting. We borrow its idea of an English reasoning scaffold. We differ in three ways: we evaluate on non-Latin and code-mixed text rather than Czech, Catalan, and French; we model sharer-to-quoted relations rather than isolated posts; and we skip reinforcement learning in favor of distillation into an encoder.
The main constraint comes from Enhanced Detection of Conversational Mental Manipulation Through Advanced Prompting Techniques. There, zero-shot chain-of-thought gave GPT-4o its lowest macro F1 of any prompting technique, because it read manipulation into confrontational style and fragmented syntax. Only few-shot reasoning with 42 human-written exemplars stabilized it. The same style-versus-intent confusion is exactly what would mislabel a heated, slangy Banglish endorsement as mockery, so our design must be tested against it rather than assumed to escape it.
The low-resource evidence comes from two sources. MixSarc finds sarcasm F1 of only 0.3938 for Banglish-BERT on code-mixed Bangla-English. It also estimates that 42.13% of negative-sentiment items in BnSentMix are sarcastic. ArSarcasm Shared Task: An Ensemble BERT Model reaches 59.89 sarcastic-class F1 on ArSarcasm-v2. To Translate or Not to Translate shows that translate-train and translate-test beat zero-shot multilingual transfer across 40 low-resource languages. That study covered NLI, sentiment, and NER, not non-literal language, so translation is a serious competitor we must beat on our task rather than dismiss.
On the network side, Tackling Fake News Detection by Continually Improving Social Context Representations explicitly lists stance-aware edges as missing from its graph model. Analysing the Extent of Misinformation in Cancer Related Tweets shows that corrective and misinformation posts share keywords, which confuses systems that ignore stance. These motivate the downstream use of the classifier.
Research Program
The first piece of work is the silver-label generator. Following the timeline-expansion approach of A Few Topical Tweets are Enough for Effective User Stance Detection, we estimate, for each dense user and each contested topic, a stance from their repost and posting history. This is done by unsupervised clustering and label propagation. That paper reports 95.6% accuracy for its unsupervised expansion on English users, and 98% for distant labels on active users. A quote-share becomes silver "endorse" when the sharer and the quoted author sit on the same side of a topic, and silver "oppose" when they sit on opposite sides. A separate silver inversion flag is set when the comment's literal sentiment, as read by an off-the-shelf multilingual sentiment model, disagrees with the silver relation. These labels depend only on network behavior, not on the comment's wording. That is the property that lets them teach a text model to see past literal tone. Their weakness is that they come only from dense users and treat same-side sharing as endorsement. We therefore measure their precision against a gold sample before trusting them.
The second piece is the rationale. For every pair, a prompted LLM fills four short English slots:
- what the quoted content asserts and about whom;
- the literal tone of the comment;
- whether the comment's meaning is inverted, with the cue that signals it (an emoji, an exaggeration, or a code-switch into English for a sarcastic phrase);
- the resulting relation.
The prompt includes a handful of native-speaker-written exemplars per language. This follows the finding of Yang et al. that exemplars, not free reasoning, stabilize subjective judgments. English is the working language of the rationale because the target text is code-mixed or dialectal. It also lets the reasoning draw on the high-resource side, in the spirit of RCCA's scaffold.
The third piece is the classifier. A target-language encoder (Banglish-BERT for Bangla-English, AraBERTv02 for Arabic) reads the comment, the quoted text, and the rationale, and is fine-tuned on the silver labels. At test time it needs neither the network nor a dense user history, so it applies to the sparse users who dominate real data.
The fourth piece is a small gold benchmark, since none exists. Native bilingual annotators will label roughly 2,000 quote-share pairs per language for relation (endorse, oppose, neutral) and inversion. Where authors can be contacted, as in iSarcasm, we also collect author-intended labels for a subset. That paper showed third-party annotators reach only 0.616 F-score against authors' own sarcasm labels, so a gold set built solely by outsiders would be a biased yardstick. The benchmark is split by whether the sharer is a dense or sparse user and by whether literal tone and relation disagree. This second split defines the "inverted subset" on which the central claim is tested.
Baselines
The naive reference point is the endorsement-by-sharing assumption itself: every quote-share is labeled "endorse." This is what propagation analyses such as Sharing political news implicitly do. Next to it sits a literal-reading baseline that maps the comment's sentiment directly to endorse or oppose. By construction it should fail on the inverted subset, and it tells us how large that subset's cost is.
The strongest existing method is the TR-Tweet+COT recipe of Gatto et al. applied to our pairs: the same encoders, the same silver training data, but free-form one-shot chain-of-thought rationales instead of slots. It reports 76.3 F1 on Tweet-Stance. We will first reproduce it on Tweet-Stance to confirm our pipeline before moving to new languages. No reported result found on Lacuna for hostile-sharing classification in any low-resource or code-mixed language, so the target-language comparison is new. Beating this baseline means a statistically significant macro-F1 gain on the gold test set, with most of the gain on the inverted subset and no loss of precision on informal endorsements.
A sarcasm-transfer baseline follows Nkhata and Gauch. The same encoder is first fine-tuned on in-language sarcasm data (MixSarc for Bangla-English, ArSarcasm-v2 for Arabic), then on the silver relation labels, with no rationale. This isolates whether explicit reasoning adds anything beyond sarcasm knowledge baked into the weights. For reference, sarcasm-detection quality on these intermediate tasks is 0.3938 F1 on MixSarc for Banglish-BERT and 59.89 sarcastic-class F1 on ArSarcasm-v2 for the ensemble system. Both numbers indicate how weak that injected knowledge may be.
A translation baseline follows Ebing and Glavaš. Pairs are translated to English with NLLB-200 3.3B and classified by an English TR-Tweet+COT model; a translate-train variant is also run.
Direct LLM baselines use the same model that writes our rationales. They cover zero-shot labeling, zero-shot free chain-of-thought, and few-shot chain-of-thought with the same exemplars we use.
We will not run RCCA's full reinforcement-learning stage because of its cost. We will run its supervised stage, fine-tuning the LLM on English-scaffolded target-language rationales, as a partial reproduction and label it as such. The file gives only relative gains for RCCA, with no absolute F1 per dataset.
The key ablation removes the rationale entirely: the encoder fine-tuned on silver labels with only the post pair as input. A second ablation keeps the rationale but deletes the inversion slot.
Fairness rules apply to all trained systems. They see identical silver data and identical gold dev sets. Each gets the same hyperparameter search budget, 20 configurations with three seeds each. All LLM-based systems use the same model and the same exemplar pool.
Why not just…?
Why not just fine-tune a multilingual encoder on the silver labels and skip the reasoning chain? This is the most serious threat. The silver labels already encode the true relation independent of wording, so an encoder might learn inversion cues directly. The rationale-free ablation settles it. If that ablation matches the full system on the inverted subset, we will report that the network-derived supervision, not chain-of-thought, is the contribution, and drop the rationale. We expect it to fall short for two reasons. The inverted subset is small, echoing the roughly 1:5 imbalance in iSarcasm and MixSarc. And recognizing what a quoted news item claims requires world knowledge a 100M-parameter code-mixed encoder lacks.
Why not just translate everything into English and use a good English model? Translation won consistently in Ebing and Glavaš, but on tasks without non-literal meaning. MixSarc reports that translation engines struggle with transliterated Banglish. The concrete failure we predict is that literal translation of ironic praise produces English that an English classifier reads as praise. We will measure this directly as the rate at which the literal-tone slot flips between original and translated input on gold-inverted pairs. If translate-test matches us on the inverted subset, the rationale's language handling is unnecessary and we will say so.
Why not just prompt a frontier LLM, with or without chain-of-thought? MixSarc reports zero-shot frontier models at 0.3905 sarcasm F1 at best, with near-zero recall for others. Gatto et al. document reasoning chains whose final labels contradict their own text. Yang et al. document free chain-of-thought hallucinating hostility from informal style. Few-shot chain-of-thought is still a fair competitor and is run with our exact exemplars. If it matches the distilled encoder, the remaining argument for our system is cost and latency, which we will report but not oversell.
Why not just use sarcasm-transfer pre-training? It works in English, but its benefit depended on stylistic match between source and target. In-language sarcasm detectors here are weak (0.3938 F1 on MixSarc), so the knowledge they transfer may be mostly noise. The sarcasm-transfer baseline tests this directly.
Proposed Experiments and Decisive Checks
The English sanity run reproduces TR-Tweet+COT on Tweet-Stance and SemEval 2016 Task 6A, where the sarcasm-transfer paper reports 0.775 F1. This confirms our implementation before any claim about new languages.
The main experiment trains all systems on silver labels for Bangla-English and Arabic quote-shares and evaluates macro F1 on the gold benchmark, reported separately for dense versus sparse sharers and for inverted versus literal pairs.
Four checks decide whether the core claim holds. First, on the inverted subset, the slot-rationale system must beat both the rationale-free ablation and free-form TR-Tweet+COT. On the literal subset it should roughly tie them; a uniform gain would mean the rationale helps for some reason other than resolving inversion. Second, the false-hostile rate on gold informal endorsements (heated, slangy, or profane but agreeing) must be lower for the slot system than for free zero-shot chain-of-thought. This directly tests the over-thinking failure from Yang et al. in a new language setting. Third, a counterfactual faithfulness test negates the inversion slot at test time. If the encoder's prediction does not flip on a substantial fraction of inverted pairs, the encoder is not using the reasoning, and gains must come from elsewhere. Fourth, we measure silver-label precision against gold, overall and on inverted pairs. If silver precision on inverted pairs is near chance, the data-scarcity half of the claim fails regardless of downstream scores.
A final downstream check applies the best classifier to re-weight edges in a quote-share graph for one contested topic per language. We then ask how much estimated cross-group endorsement shrinks when hostile shares are removed. This makes concrete, for low-resource communities, the stance-aware graph edges that the fake-news GNN work named as missing.
References
- A Few Topical Tweets are Enough for Effective User Stance Detection
- Chain-of-Thought Embeddings for Stance Detection on Social Media
- Intermediate-Task Transfer Learning: Leveraging Sarcasm Detection for Stance Detection
- A Reinforcement Learning Framework for Cross-Lingual Stance Detection Using Chain-of-Thought Alignment
- Enhanced Detection of Conversational Mental Manipulation Through Advanced Prompting Techniques
- MixSarc: A Bangla-English Code-Mixed Corpus for Implicit Meaning Identification
- ArSarcasm Shared Task: An Ensemble BERT Model for SarcasmDetection in Arabic Tweets
- To Translate or Not to Translate: A Systematic Investigation of Translation-Based Cross-Lingual Transfer to Low-Resource Languages
- Tackling Fake News Detection by Continually Improving Social Context Representations using Graph Neural Networks
- Analysing the Extent of Misinformation in Cancer Related Tweets
- iSarcasm: A Dataset of Intended Sarcasm
- Sharing political news: the balancing act of intimacy and socialization in selective exposure