Skip to content

C³T: Counterfactual Causal Reasoning for Sentiment Shifts in Social-Media Conversation Trees

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Conversation Trees, Causal Sentiment Reasoning, Counterfactual Inference, Ancestor Attribution, Sentiment Shifts

TL;DR

This paper addresses the lack of causal explainability behind emotional dynamics in social-media conversation trees by introducing the CASIRE benchmark and C³T (Counterfactual Causal Conversation Transformer)—a model integrating tree-temporal message passing, sparse ancestor attribution, and dual-stream forced counterfactual intervention—achieving superior event-OOD robustness and principled estimation of discourse-move treatment effects on downstream negativity.

Background & Motivation

Conversations surrounding breaking news, crises, and online rumors inherently unfold as branching reply trees where participants react to emerging claims, debunking attempts, and hostile remarks. Within these cascades, user sentiment is neither static nor homogeneous; rather, it shifts dynamically across parent-child exchanges and propagates down disparate conversation paths. Prior research has overwhelmingly focused on veracity prediction and stance classification, largely overlooking the affective trajectory of how sentiment evolves over time and which specific conversational inputs act as true causal drivers of affective shifts.

Attributing sentiment transitions across branching reply trees is fundamentally a causal problem, yet observational social-media data are fraught with severe confounding, endogenous participation, and event-specific lexical shortcuts. Existing graph neural networks and temporal graph architectures easily overfit to surface entities and keywords (such as event hashtags or place names), leading to catastrophic performance degradation in out-of-event scenarios. Conversely, while open-weight large language models benefit from vast world knowledge via prompt-based inference, they struggle to faithfully track deep branching dependencies, frequently defaulting to the immediate parent post and failing to provide reliable, calibrated ancestor attribution.

This work addresses this challenge by conceptualizing conversational discourse moves (e.g., corrections, evidence links, toxicity) as candidate interventions within a potential-outcomes framework. Core idea: Introduce the CASIRE benchmark layer with calibrated multi-label discourse interventions and expert causal-source labels, and propose C³T, an end-to-end framework combining tree-temporal encoding, sparse ancestor attribution, and dual-stream forced counterfactual representation learning to robustly predict sentiment shifts while isolating the downstream causal treatment effects of conversational moves.

Method

Overall Architecture

C³T takes a rooted, timestamped conversation tree \(G=(V, E)\) as input, where each post node possesses text, relative posting time, structural depth, and multi-label discourse intervention tags. The framework operates across three core stages: first, an intervention tagging module extracts and calibrates an 8-dimensional multi-label treatment vector \(T_i\); second, a tree-temporal encoder interleaves direct parent message passing with scaled dot-product attention over bounded ancestor histories to construct dynamic causal representations; finally, dedicated heads output node sentiment distributions, parent-to-child shift categories, and sparse ancestor attribution weights, while a dual-stream counterfactual module isolates the treatment embedding to evaluate potential outcomes and compute Average Treatment Effects (ATE) across descendant subtrees.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["Conversation Tree Input<br/>Post texts, timestamps, and branching reply structure"] --> B["Intervention Tagging and Calibration<br/>8-class multi-label discourse tagging with abstention"]
    B --> C["Tree-Temporal Graph Encoder<br/>Parent message passing + ancestor temporal attention"]
    C --> D["Dual-Head Prediction and Sparse Attribution<br/>Node sentiment/shift heads + Sparsemax causal ancestor head"]
    D --> E["Dual-Stream Counterfactual Intervention<br/>Disentangled treatment embedding on/off for downstream ATE"]
    E --> F["Potential Outcome Outputs<br/>Event-OOD robust sentiment predictions & causal effect analysis"]

Key Designs

1. Intervention Tagging and Calibration: Disentangling Compound Communicative Acts

Individual social-media posts frequently perform concurrent conversational functions, such as pairing a factual correction with an external authoritative URL or embedding sarcasm within an inquisitive challenge. To capture these nuances without artificial mutual exclusion, the framework establishes a taxonomy of \(M=8\) conversational interventions (Claim/Assertion, Denial/Correction, Evidence/Link, Authority Citation, Toxicity/Attack, Sarcasm/Irony, Question/Challenge, Derail/Off-topic) encoded as a multi-hot vector \(T_i \in \{0, 1\}^M\). To mitigate poor probability calibration from weak labelers, predicted scores are adjusted via temperature scaling on held-out validation sets with category-specific decision thresholds (e.g., \(\tau=0.65\) for denial, \(\tau=0.70\) for evidence, \(\tau=0.80\) for toxicity), supplemented by a selective prediction abstention rule that discards instances with maximum confidence below 0.4 from counterfactual loss computation.

2. Tree-Temporal Graph Encoder: Synchronizing Structural Topology with Temporal Delays

Information propagation in reply threads exhibits irregular temporal gaps and deep hierarchical divergence, meaning flat sequential or static graph encoders fail to preserve the causal sequence of replies. C³T encodes post text via DeBERTa to yield \(e_i \in \mathbb{R}^d\), concatenating learned depth embeddings \(p_i = \text{Emb}_{\text{depth}}(d(i))\), continuous sinusoidal temporal encodings \(t_i = \phi(\Delta \tau_i)\) based on elapsed time from the parent post, and projected treatment embeddings \(u_i = \sum_m T_{i,m} r_m\). Across \(L\) layers, each node representation updates dynamically through a GRU gate combining the direct parent message \(m_i^{\text{par}} = W_p h_{\pi(i)}^{(\ell)}\) and an attention-aggregated ancestor context \(m_i^{\text{anc}}\) over the nearest \(D\) ancestors \(A_D(i)\):

\[h_i^{(\ell+1)} = \text{GRU}\left(h_i^{(\ell)}, \left[m_i^{\text{par}};\, m_i^{\text{anc}}\right]\right)\]

3. Dual-Head Prediction and Sparse Attribution: Explicit Shift Modeling and Targeted Cause Localization

To prevent models from relying on static sentiment priors that ignore contextual trajectory, C³T employs decoupled prediction heads: a node sentiment head \(\hat{y}_i = \text{softmax}(W_y z_i)\) and a transition shift head \(\hat{s}_{p\to i} = \text{softmax}(W_s [z_p; z_i; z_i - z_p])\) supervised against gold ordinal transitions (downward, unchanged, upward). For causal attribution, the model evaluates candidate ancestors in \(A_D(i)\) via bilinear matching scores \(e_{i,a} = (W_c z_i)^\top (W_a z_a)\). By substituting standard Softmax with Sparsemax normalization, the attribution head enforces sparse, peaked probability mass, eliminating diffuse attention distributions where "everything causes everything" and precisely pinpointing the primary causal driver.

4. Dual-Stream Counterfactual Intervention: Isolating Treatment Signals for Potential Outcome Estimation

Standard predictive models entangle intervention tokens with confounding text patterns, preventing genuine queries about what would happen if a specific discourse move were withheld. C³T resolves this via a two-stream representation \(z_i^{(m=v)} = b_i + W_u u_i^{(m=v)}\), where \(b_i\) represents the treatment-free contextual backbone and \(u_i^{(m=v)} = u_i + (v - T_{i,m})r_m\) forces intervention \(m\) on (\(v=1\)) or off (\(v=0\)). By propagating this modified source embedding down to descendant nodes \(j \in D_k(i)\) within a \(k\)-hop neighborhood (\(k=2\)), C³T measures the shift in predicted negativity probabilities, estimating the empirical Average Treatment Effect (ATE):

\[\Delta_{i,m}(k) = \frac{1}{|D_k(i)|} \sum_{j \in D_k(i)} \left(\hat{p}_{j,1}^{\text{neg}} - \hat{p}_{j,0}^{\text{neg}}\right), \quad \text{ATE}_m(k) = \mathbb{E}_i \left[\Delta_{i,m}(k)\right]\]

Loss & Training

The full model is trained end-to-end under a composite objective balancing supervised accuracy, causal sparsity, and cross-event representation invariance:

\[\mathcal{L} = \mathcal{L}_{\text{sent}} + \lambda_{\text{shift}}\mathcal{L}_{\text{shift}} + \lambda_{\text{att}}\mathcal{L}_{\text{att}} + \lambda_{\text{cf}}\mathcal{L}_{\text{cf}} + \lambda_{\text{spar}}\mathcal{L}_{\text{spar}} + \lambda_{\text{irm}}\mathcal{L}_{\text{irm}}\]

Here, \(\mathcal{L}_{\text{sent}}\) and \(\mathcal{L}_{\text{shift}}\) represent standard cross-entropy losses; \(\mathcal{L}_{\text{att}}\) optimizes ancestor attribution against gold annotations (validated at Fleiss' \(\kappa = 0.75\)), regularized by an entropy penalty \(\mathcal{L}_{\text{spar}} = \sum_i H(\alpha_i)\) to ensure sparse attribution. The counterfactual margin loss \(\mathcal{L}_{\text{cf}} = \sum_{i,m} \left[T_{i,m}\max(0, \gamma - \delta_{i,m}) + (1 - T_{i,m})\beta \delta_{i,m}\right]\) enforces non-trivial sensitivity \(\delta_{i,m}\) for active treatments while penalizing unwarranted sensitivity for absent ones. Finally, an Invariant Risk Minimization (IRM) gradient penalty \(\mathcal{L}_{\text{irm}} = \sum_e \|\nabla_{w|w=1.0} \mathcal{L}_{\text{pred}}^{(e)}(w \cdot \Phi_\theta)\|_2^2\) across event environments \(E_e\) prevents the network from latching onto spurious event-specific lexical shortcuts.

Key Experimental Results

Main Results

Evaluations are conducted on the PHEME/RumourEval conversation datasets. The primary setting is the Event-OOD split, partitioning threads strictly by root event to evaluate unseen crisis dynamics: training on charliehebdo, ferguson, and germanwings-crash (980 threads), tuning on ottawashooting (175 threads), and evaluating on sydneysiege (205 threads).

Model In-Domain Sent. F1 In-Domain Shift F1 In-Domain Attr. Top-1 / MRR Event-OOD Sent. F1 Event-OOD Shift F1 Event-OOD Attr. Top-1 / MRR
Text-only Transformer 63.4 ± 0.7 48.2 ± 0.9 – / – 49.5 ± 1.2 39.1 ± 1.5 – / –
Bi-GCN (adapted) 65.1 ± 0.5 55.7 ± 0.6 23.4 / 35.8 51.2 ± 1.0 44.3 ± 1.1 19.8 / 31.2
GACL (adapted) 66.8 ± 0.6 57.2 ± 0.5 26.1 / 38.5 55.9 ± 0.9 48.6 ± 0.8 22.4 / 34.1
Temporal GNN (TGN/TGAT) 66.2 ± 0.8 56.5 ± 0.7 25.8 / 37.9 54.1 ± 1.1 47.2 ± 1.3 21.5 / 33.7
C³T (Ours) 68.5 ± 0.4 61.3 ± 0.5 41.7 / 59.4 60.4 ± 0.6 54.8 ± 0.7 36.2 / 52.1

Ablation Study and Treatment Effects

The ablation study assesses the contributions of counterfactual regularization, explicit intervention tags, event-invariance penalties, and ancestor-window size under the Event-OOD regime. In parallel, C³T estimates the 2-hop Average Treatment Effect (\(\text{ATE}_m(k)\) with \(k=2\)) of all eight discourse interventions on downstream negativity, with 95% bootstrap confidence intervals computed over 1,000 thread-level resamples.

Model Variant / Ablation Sent. F1 Shift F1 Attr. Top-1 / MRR Discourse Intervention Type (m) 2-Hop Downstream Negativity ATE [95% CI] Effect Dynamic
C³T (Full Model) 60.4 ± 0.6 54.8 ± 0.7 36.2 / 52.1 Denial / Correction -0.142 [-0.178, -0.106] De-escalation
A1: w/o Counterfactual Loss (\(\lambda_{\text{cf}}=0\)) 58.1 ± 0.8 52.9 ± 0.9 32.4 / 48.5 Evidence / Link Provision -0.118 [-0.153, -0.083] De-escalation
A2: w/o Intervention Tags (\(u_i=0\)) 55.7 ± 1.1 49.6 ± 1.2 25.1 / 38.9 Authority Citation -0.096 [-0.135, -0.058] De-escalation
A3: w/o Event Invariance (\(\lambda_{\text{irm}}=0\)) 56.9 ± 0.9 51.3 ± 1.0 34.1 / 50.2 Question / Challenge -0.015 [-0.062, +0.032] Context-dependent
A4: Parent-only Attribution (\(A_D(i)=\{\pi(i)\}\)) 59.2 ± 0.5 53.7 ± 0.6 21.5 / 21.5 Claim / Assertion +0.032 [-0.012, +0.076] Context-dependent
Ref: Llama 3 8B Few-Shot (AncSum) 59.2 ± 1.2 – 25.1 / 38.4 Derail / Off-topic +0.048 [+0.005, +0.091] Slight Escalation
Ref: Qwen 2.5 7B Few-Shot (AncSum) 58.9 ± 1.3 – 25.8 / 39.1 Sarcasm / Irony +0.064 [+0.015, +0.113] Escalation
Ref: Mistral 7B Few-Shot (AncSum) 56.7 ± 1.5 – 23.4 / 36.8 Toxicity / Attack +0.235 [+0.192, +0.278] Strong Escalation

Key Findings

  • Divergent Causal Impacts of Discourse Moves: Factual corrections (\(\text{ATE}=-0.142\)), verifiable evidence links (\(-0.118\)), and authoritative citations (\(-0.096\)) consistently exert a calming, de-escalating influence by reducing predicted downstream negativity across subsequent turns. Conversely, toxic attacks (\(+0.235\)) act as potent triggers for escalation, sharply propagating negative affect well beyond the immediate parent-child exchange.
  • Critical Necessity of Deep Ancestor Attribution: Restricting attribution solely to the immediate parent (Ablation A4) causes Top-1 attribution accuracy to plummet from 36.2% to 21.5%. This empirical gap confirms that in non-linear social conversations, a substantial portion of affective shifts are provoked by root claims or earlier provocations rather than the post being directly replied to.
  • Counterfactual Regularization Enhances OOD Robustness: Removing intervention tags (A2) drops sentiment F1 by 4.7 points and attribution Top-1 by 11.1 points; omitting counterfactual margin regularization (A1) reduces attribution Top-1 by 3.8 points. Explicit intervention modeling shields representations from event-specific spurious correlations, directly powering generalizability.

Highlights & Insights

  • Lightweight Additive Formulation for Counterfactual Inference: Rather than retraining separate graph modules or executing computationally prohibitive graph mutations, C³T cleanly decomposes node representations into a treatment-agnostic structural backbone and an additive treatment embedding, rendering multi-hop potential-outcome queries highly efficient.
  • Actionable Diagnostic Utility for Online Governance: Beyond passive sentiment classification, C³T enables counterfactual "what-if" auditing (e.g., verifying that removing an abusive comment would have reduced subsequent negativity probability by 0.22 in qualitative case analyses), providing an objective tool for assessing the impact of content moderation.
  • Bridging Structural Deep Learning and LLM Capabilities: While open-weight LLMs achieve respectable classification when fed flat ancestor summaries, their attribution accuracy lags by over 10 points compared to C³T, illustrating that explicit graph-inductive biases and dedicated causal heads remain irreplaceable for structured conversation modeling.

Limitations & Future Work

  • Residual Unobserved Confounding: While C³T conditions on observed textual content, tree topology, temporal delays, and discourse tags, latent confounders such as users' external socio-political leanings, unrecorded offline exposure, and algorithmic feed prioritization remain unmeasured.
  • Representation-Level vs. Naturalistic Intervention: The counterfactual mechanism operates by toggling treatment embeddings on or off while keeping non-intervention context fixed; it does not simulate full text regeneration or how other participants would alter their participation decisions in response.
  • Local Interference Bounds: Descendant effects are bounded within a \(k=2\) hop radius, leaving broader cross-thread spillover dynamics, quote-tweet amplification, and cross-platform propagation unmodeled.
  • vs. Graph-Based Rumor Detection (e.g., Bi-GCN, GACL): Prior rumor architectures focus almost exclusively on macroscopic event veracity or coarse stance detection, viewing user sentiment merely as an implicit feature. C³T pivots to the micro-level affective dynamics of conversation trees, explicitly predicting parent-to-child emotional transitions and identifying their causal sources.
  • vs. Prompt-Based Open-Weight LLMs (e.g., Llama 3, Qwen 2.5): Prompting LLMs with flattened conversational histories frequently incurs recency bias and hallucinated attributions on deep branching threads. C³T integrates topological inductive biases with Sparsemax ancestor attention, establishing significantly superior attribution accuracy and out-of-event stability.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Formulates causal sentiment reasoning and counterfactual discourse estimation on social-media conversation trees for the first time.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous Event-OOD evaluation protocol, 8-class intervention ATE estimation, multi-seed variance reporting, and extensive ablations against both GNNs and LLMs.
  • Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally rigorous mathematical and causal grounding, clear problem formulation, and polished visual and empirical presentations.
  • Value: ⭐⭐⭐⭐⭐ Offers crucial methodological and empirical insights for computational social science, online toxicity mitigation, and automated dialogue moderation.