Frustration as Evidence: Using Multi-Faction Signed Graph Structure to Detect Sarcastic Replies

When people argue online, a reply's literal tone often misrepresents the relationship behind it. "Thanks for the incredible wisdom" reads as friendly, yet it is usually an attack. Researchers who build networks of agreement and disagreement from discussion text record such replies as positive links. As a result, the factions they later recover are drawn from corrupted edges. This proposal reverses the usual direction of information flow. We ask whether the shape of the disagreement network itself can tell us which positive-looking replies are sarcastic. Suppose two users sit in hostile camps and every other signal between their camps is negative. Then a warm reply from one to the other is suspicious, and the amount of structural inconsistency it would introduce can be measured.

The measuring tool is the signed Laplacian. It is a matrix built from a network whose links carry plus or minus signs. Its smallest eigenvalue is zero exactly when the network splits cleanly into two camps that are friendly inside and hostile across, and it grows as the network becomes "frustrated", meaning inconsistent with such a split. That two-camp picture is the catch. Real debates often have three or more mutually hostile factions plus many bystanders. Under the classic two-camp theory, three mutually hostile groups are themselves frustrating, so a frustration-based prior would flag sincere hostility as an anomaly. It might also push the model to "explain away" genuine attacks. Our central claim is that a network-derived sarcasm signal is useful only if it is computed under the right notion of balance: multi-faction, tolerant of bystanders, and scored with the edge being judged left out. We further claim that the standard two-camp frustration signal will help in two-sided debates and actively hurt in multi-sided ones.

Where this sits

Signed graph methods supply the machinery, but all of them treat edge signs as given. SPONGE clusters signed networks into $k \ge 3$ groups by solving a generalized eigenproblem that balances attraction and repulsion, with recovery guarantees under a signed block model. Robust Deep Signed Graph Clustering via Weak Balance Theory (DSGC) adopts weak balance, in which three mutually hostile groups are allowed. It also corrects noisy signs using random-walk consistency, but its noise model is topological and it is evaluated on non-text correlation networks. Discovering Polarized Communities in Signed Networks lets nodes be neutral bystanders but allows only two active camps. Finding large balanced subgraphs and Balance Maximization via Edge Deletions use perturbation of the smallest signed-Laplacian eigenvalue to locate the nodes or edges most responsible for frustration. We reuse that perturbation idea, but apply it to score sign flips rather than deletions. These papers supply our method. None of them asks whether a frustrating edge is a linguistic error.

On the text side, Hassan et al. build signed networks from forum text with dependency-path attitude extraction and partition them greedily into more than two subgroups. Their extractor is literal, so sarcastic praise becomes a positive edge. This is the pipeline we repair, and it motivates the work. Sarcasm detectors condition on the sentence (ADGCN, whose "graph" is words within one sentence), on the parent comment (Gregory et al.), or on the author's timeline (Oprea and Magdy). None of them uses the network of who opposes whom. The gap that motivates this work is concrete. Gregory et al. report that 67.7% of political tweets in their sample were sarcastic, so text-inferred signed graphs in political domains are heavily corrupted. The closest idea to ours, DSGC's topological sign refinement, corrects signs without reading text. We instead combine text evidence with structural evidence, and we make the balance notion the experimental variable.

Core Claim

Let each reply edge $(i,j)$ carry a text-model probability $q_{ij}$ that the reply is hostile. Let $V$ be a $k$-faction embedding of users computed from the signed graph with edge $(i,j)$ removed. Removing the edge prevents it from vouching for itself. We define a structural flip score $s_{ij}$ as the first-order change in a multi-faction frustration objective when the edge is set negative rather than positive. For the two-camp baseline, this is the usual eigenvalue perturbation. If $v$ is the eigenvector for the smallest eigenvalue $\lambda_1$ of the signed Laplacian and the edge has weight $w$, flipping it from positive to negative changes $\lambda_1$ by approximately

$$\Delta\lambda_1 \approx 4w\, v_i v_j,$$

which is negative (frustration drops) when $i$ and $j$ sit on opposite sides. The multi-faction version replaces $v$ with SPONGE's embedding $V$. It sets the score to zero for users whose embedding norm falls below a bystander threshold, following the neutral-state idea of 2PC. It also counts negative edges between different factions as non-frustrating, following weak balance. The prediction combines the two sources of evidence as

$$\text{logit}\,P(\text{sarcastic}_{ij}) = f_\theta(\text{text}_{ij}, \text{parent}_{ij}) - \beta\, s_{ij},$$

where $f_\theta$ is a text sarcasm classifier and $\beta$ is fit on a development split. A few rounds of alternation then follow: flip the high-confidence sarcastic edges, recompute factions, and rescore.

The claim is falsifiable in two directions. The structural term must add to the text model mainly on low-activity users, where author history is thin. The two-camp version of $s_{ij}$ must degrade relative to text-only in threads with three or more factions.

Research Program

The main build is a reply-level signed graph corpus with sarcasm ground truth. We will use the Self-Annotated Reddit Corpus (SARC). It provides the /s self-labels, with a reported 1.0% false-positive rate, along with full comment trees, user identities, and subreddits, which is exactly what is needed. For each political subreddit and time window, we will form a user graph whose edge signs come from a literal stance/sentiment model applied to replies. The /s labels then mark which edges are sarcastic. Subreddits will be stratified by estimated faction count, using SPONGE's eigengap and DSGC's reported stability over $k=4$ to $8$ as a sanity check. This stratification provides a natural test of the binary-versus-multi-faction prediction.

The second strand is the downstream question. Does sarcasm-corrected edge construction improve subgroup recovery? We will test this on the Hassan et al. forum threads, which have gold subgroup partitions. We will also test on the Reddit Israel–Palestine stance data, which has three classes including a neutral one, if its comments can be linked back to reply structure. The file does not say whether they can. If they cannot, we will annotate a small comparable set ourselves.

The analysis strand characterizes when structure helps. It will stratify gains by author activity, by whether the parent comment is available, and by the reply's position relative to faction boundaries.

Baselines

The naive reference points are two. The first predicts from the literal sign alone, calling nothing sarcastic. The second is a sarcasm classifier using the target comment only. On SARC's unbalanced label regime, the SARC paper reports that baseline F1 falls to around 25–27%, and a bag-of-bigrams baseline reaches 75–76% pairwise accuracy against 82–92% for humans. These numbers anchor the task's difficulty.

The strongest text competitor is a transformer that reads both the reply and its parent comment, in the style of Gregory et al.. Conversational incongruity is the best-established sarcasm cue in the file. The same model is our $f_\theta$, so the comparison isolates the structural term exactly. It will be trained with the same data, splits, and tuning budget. We found no reported result on Lacuna for a parent-context transformer on SARC in the unbalanced regime, so we will measure it ourselves. For intended sarcasm, the file's best reference is the iSarcasmEval English top score of 0.605 F-sarcastic, but that benchmark has no reply graph and serves only as context.

Author history, following Oprea and Magdy, is run as a third baseline and also combined with $f_\theta$. The paper reports 0.934 F1 on hashtag-labeled Ptacek tweets and 0.739 on Riloff; the file has no SARC number for it. An LLM prompted with the thread, following the reflective re-read prompting of the conflict stance paper, is a further baseline, with prompts tuned on the same development split. For subgroup recovery, the reference is Hassan et al.'s reported Rand index of 0.60 on Wikipedia discussions, against 0.49 for their spectral clustering baseline. We will rerun their literal pipeline and SPONGE on uncorrected edges under identical splits.

The ablations isolate the key ingredient. The first swaps the multi-faction flip score for the two-camp $\Delta\lambda_1$. The second removes bystander zeroing. The third removes leave-edge-out, to measure leakage. The fourth replaces $s_{ij}$ with DSGC's text-blind random-walk sign refinement. The fifth replaces it with a plain "are $i$ and $j$ in different clusters" indicator taken from the same SPONGE partition. Beating the baselines means a statistically reliable gain in sarcastic-class F1 over the parent-context model at SARC's natural ratio, concentrated on low-history users. It also means a higher Rand index than the uncorrected pipeline on forum subgroups.

Why not just…?

Why not just use a strong LLM or context-aware classifier and flip the signs it flags? This is the main threat, and it is also our $f_\theta$. The evidence that text alone plateaus is indirect. The best prompted model on three-way conflict stance reaches 74.08% accuracy and is weakest on neutral comments (conflict stance paper). Hashtag-trained detectors fall to 0.364 F on author-intended sarcasm (iSarcasm). Hassan et al. found that opposing factions share vocabulary, so structure carries information that content does not. The same-model, with-and-without-$s_{ij}$ comparison settles the question. If $\beta$ is fit near zero, the proposal fails.

Why not just condition on author history? History helps when it exists, but many forum participants post rarely. Its reported gain on third-party-labeled sarcasm is small, 0.739 versus 0.711 F1 (Oprea and Magdy). We test whether the structural score is complementary by stratifying results by user activity and by combining both signals.

Why not just use the faction labels from any clustering, with no eigenvalue perturbation? This is a real possibility. If the plain different-cluster indicator matches the perturbation score, the honest conclusion is that a multi-faction network prior matters and the spectral perturbation does not. We would report exactly that. The perturbation score should win where cluster assignments are uncertain, near faction boundaries and for weakly attached users, because it is continuous and weighted by embedding confidence. The boundary-stratified analysis checks this.

Why not just use DSGC's topological sign correction? It flags inconsistent edges without reading them. So it cannot distinguish a sarcastic compliment from a sincere cross-faction compliment, and the latter is exactly what bridge users produce. Running it as an ablation measures how much of the gain needs text.

Decisive Checks

Three results would decide the proposal. First, on SARC political subreddits at the natural label ratio, adding the multi-faction $s_{ij}$ must raise sarcastic-class F1 over the parent-context model, with the gain concentrated on low-history users. Second, the two-camp $\Delta\lambda_1$ prior must help in threads estimated to be two-sided and hurt in threads with three or more factions, while the multi-faction score helps in both. This crossover is the direct test of the binary polarization assumption. Third, sarcasm-corrected edges must raise subgroup Rand index on the Hassan forum threads above the uncorrected literal pipeline. If the leave-edge-out ablation shows large gains that vanish once leakage is removed, or if the plain cluster indicator equals the spectral score everywhere, we will report the narrower finding rather than the full claim.