C$^{3}$T: Counterfactual Causal Reasoning for Sentiment Shifts in Social-Media Conversation Trees

S M Rafiuddin; Atriya Sen

Source-linked AI summary

C$^{3}$T: Counterfactual Causal Reasoning for Sentiment Shifts in Social-Media Conversation Trees

S M Rafiuddin, Atriya Sen

arXiv:2609.02131v1Sep 2, 2026cs.CLcs.AI

TL;DR

The paper addresses why sentiment changes within branching rumor conversations and develops causal sentiment reasoning for identifying conversational drivers. It introduces CaSiRe and C3T, combining intervention-aware counterfactual modeling with thread structure and temporal context. C3T improves event-level out-of-distribution robustness and attribution, while its effects remain model-based estimates from observational data.

Problem
Existing work under-models how sentiment changes over time in rumor conversation trees and which discourse moves plausibly drive those changes, while observational data contain confounding.
Method
The paper introduces CaSiRe and C3T, using time-stamped thread structure, conversational intervention tags, joint sentiment and shift prediction, ancestor attribution, and forced-intervention counterfactual queries.
Results
C3T performs best overall across node sentiment, edge-level shift, and causal ancestor attribution, with the largest gains in attribution and event-OOD robustness.
Takeaways & Limitations
Under event-level OOD evaluation, integrating thread structure, temporal context, and intervention-aware modeling improves attribution and generalization for sentiment reasoning in rumor trees.
Takeaways & Limitations
Estimated intervention effects are conditional, model-based counterfactuals rather than randomized causal effects because unmeasured confounding may remain.

Abstract

from arXiv · show

Sentiment in social-media threads does not only vary across posts; it shifts as users react to claims, corrections, evidence, and hostility within a branching reply tree. We study why sentiment changes in rumor-centric conversation trees by treating discourse moves (e.g., denial/correction, evidence/link, toxicity/attack) as candidate interventions and asking (i) what sentiment a reply expresses, (ii) whether the sentiment shifts relative to its parent, and (iii) which prior message most plausibly drove the reply's sentiment. To support this setting, we introduce CaSiRe, a causal sentiment reasoning layer over public rumor conversation datasets that adds post-level sentiment labels, induced parent-child shift labels, calibrated multi-label intervention tags, and explicitly annotated causal-source labels. We then propose C$^{3}$T (Counterfactual Causal Conversation Transformer), a thread-structured temporal model that jointly predicts node sentiment and shifts, learns sparse ancestor attribution, and supports counterfactual queries by forcing conversational intervention embeddings on or off to estimate potential outcomes. Under an event-level split, C$^{3}$T improves out-of-event robustness and attribution over text-only, graph-based, and temporal baselines, and yields interpretable model-based effects: denials/corrections and evidence reduce downstream negativity, while toxicity increases it. We also benchmark open-weight LLM prompting baselines and find that added conversational context helps, but attribution remains less reliable, motivating structure-aware counterfactual modeling for social-media analysis.

1 Introduction

The paper frames sentiment shifts in branching social-media conversations as a causal reasoning problem and proposes structure- and time-aware modeling to identify plausible conversational drivers.

Sentiment shifts in reply threads may reflect new claims, corrections, or hostile exchanges, not merely the sentiment expressed in individual posts.
Rumor conversation trees capture branching interactions, but prior work has emphasized stance and veracity rather than affective trajectories and their conversational drivers.
Observational participation and latent confounding make discourse-move-specific attribution of sentiment change fundamentally causal.
The proposed framework treats discourse moves as candidate interventions and estimates their downstream impact on expressed sentiment.
The study evaluates structure- and time-aware sentiment reasoning on public rumor benchmarks against neural and open-weight LLM prompting baselines.
The contributions include counterfactual intervention estimation and ablations of counterfactual loss, intervention tags, event-level invariance, and ancestor-window attribution.

2 Tasks and Causal Setup

The paper represents rumor discussions as time-stamped conversation trees, predicts sentiment and parent-child shifts, and defines intervention-based downstream negativity estimands under observational assumptions.

2.1 Conversation tree and notation: Each discussion is modeled as a rooted, time-stamped tree whose nodes are posts or comments and whose directed edges encode replies to parent nodes.
2.2 Prediction targets: The framework uses node sentiment labels and induced edge-level shifts that encode down, unchanged, or up movement from parent to child.
2.3 Interventions: Each post receives a multi-label intervention vector covering claims, corrections, evidence, authority citations, toxicity, sarcasm, questions, and derailment.
2.4 Causal estimands: The causal setup uses potential outcomes to estimate how toggling one intervention at a source node changes downstream descendant sentiment.
2.4 Causal estimands: Primary causal-effect reporting is restricted to downstream negativity, while heterogeneous effects by event, platform, veracity, or conversational stage remain exploratory unless reported with confidence intervals.
2.5 Identification assumptions: The observational analysis assumes temporal precedence, consistency, positivity, limited interference, and scope-limited conditional ignorability.

3 Method: C3T

C3T models sentiment and sentiment shifts in time-stamped conversation trees while attributing replies to prior ancestors and estimating intervention-specific counterfactual outcomes. Its architecture combines calibrated intervention tags, structure- and time-aware encoding, multi-task prediction, sparse attribution, and forced treatment representations.

Model overview: C3T predicts node sentiment, edge-level sentiment shifts, ancestor sources, and counterfactual potential outcomes under conversational interventions.The model is designed to jointly represent what sentiment is expressed, how it changes from the parent, which prior message plausibly drove it, and how outcomes change under forced interventions.
Intervention tagging: Each post receives a multi-label intervention vector covering claims, corrections, evidence, authority citations, toxicity, sarcasm, questions, and derailment.Intervention labels may be produced by LLM-assisted tagging with calibrated, type-specific thresholds and abstention, or by a learned classifier trained on limited manual supervision.
Conversation encoder: The encoder combines text, depth, reply timing, optional author/platform features, and intervention embeddings, then updates nodes using parent message passing and ancestor attention.Sibling aggregation is optional, and the architecture avoids using gold or predicted parent sentiment as an input feature.
Prediction heads: Prediction heads estimate sentiment distributions and parent–child shifts from node representations and paired parent–child features.The shift head uses the parent representation, child representation, and their difference to classify down, same, or up sentiment movement.
Counterfactual representation learning: For intervention type m, C3T forces its assignment off or on while holding other intervention components fixed, producing potential-outcome predictions for downstream effect estimation.A two-stream decomposition separates structure-and-text representations from treatment-sensitive embeddings, reducing counterfactual computation.
Training and effects: The objective combines sentiment, shift, attribution, counterfactual, sparsity, and event-invariance losses, with sparse ancestor distributions trained against causal-source labels.The model estimates descendant negativity effects over k-hop descendants and reports average treatment effects, while ancestor attention costs O(nDd) per layer.

4 Experimental Setup

The experiments compare C3T with text-only, graph-based, temporal, and prompt-only open-weight LLM baselines under a fixed event-level evaluation setup. Evaluation covers sentiment, shift, causal-source attribution, and downstream causal effects.

Baselines: C3T is compared with text-only Transformers, rumor-propagation graph models, temporal models, and prompt-only open-weight LLM systems.Graph and temporal attribution variants use the same ancestor-scoring head and causal-source labels, but not C3T’s treatment embeddings, counterfactual loss, or event-invariance penalty.
LLM prompting: Open-weight LLMs are evaluated with fixed templates, demonstrations, option order, deterministic decoding, and three context settings: node-only, node+parent, and node+ancestor summary.The evaluated models include Llama 3, Gemma 2, Qwen2.5, and Mistral.
Metrics: Macro-F1 measures node sentiment and reply-edge shift prediction, while attribution uses Top-1 accuracy and MRR over candidate ancestors.Causal effects are reported as ATEm(k) for each intervention type with 95% bootstrap confidence intervals from thread-level resampling.
Training protocol: The main C3T model uses DeBERTa-base, 2 or 3 tree-temporal layers, ancestor windows of 8, 16, or 32, dropout 0.1, AdamW, three random seeds, and early stopping on development sentiment macro-F1.All non-LLM baselines share preprocessing, event-level splits, and applicable evaluation metrics; graph models tune learning rate and dropout.

5 Results

C3T performs best overall across sentiment, shift, and causal-ancestor attribution, with especially strong event-OOD robustness and attribution. Open-weight LLMs improve with conversational context but remain less stable at attribution.

C3T performs best overall across node sentiment, edge-level shift, and causal ancestor attribution in both in-domain and event-OOD evaluation.The largest gains occur in attribution and event-OOD robustness.
C3T reduces the event-OOD gap, suggesting representations less tied to event-specific surface cues and more tied to conversational function.The comparison emphasizes intervention-aware, event-invariant modeling.
Causal ancestor attribution shows the largest gains, indicating more reliable identification of earlier messages that shape affective reactions than reliance on the immediate parent.
Prompted LLM baselines benefit from Node+Parent and Node+AncSum context but remain less stable for attribution.The fixed-protocol evaluation compares zero-shot and few-shot prompting across node-only, parent, and ancestor-summary contexts.
Structure- and intervention-aware counterfactual learning improves sentiment-dynamics modeling beyond post-level classification.

6 Ablations and Robustness

The ablation study isolates counterfactual regularization, intervention tags, event-level invariance, and ancestor-window modeling on the held-out event-OOD test set. Table 4 additionally summarizes two-hop intervention effects on downstream negativity.

Ablation protocol: All core ablations use the same event-level split, training budget, and hyperparameter search ranges as the full model.Results are reported on the held-out event-OOD test set.
Ablation protocol: The core ablations test counterfactual loss, intervention tags, event-level invariance, and ancestor-window modeling as distinct design choices.
Robustness and effects: Table 4 estimates 2-hop average treatment effects for all eight conversational interventions on downstream negativity using thread-level bootstrap confidence intervals.Rows are ordered from de-escalation to escalation.
Robustness and effects: Question and Claim cross zero in Table 4, indicating context-dependent effects.

7 Causal Effects and Social Insights

The counterfactual analysis estimates how conversational interventions shape downstream negativity, while case studies illustrate ancestor attribution and intervention removal. Denials/corrections and evidence tend to de-escalate, whereas toxicity/attack escalates negativity, with some effects depending on context.

Causal estimands: All reported intervention effects are observational model-based causal estimates conditional on measured context.
Causal estimands: The k-hop ATE toggles one intervention while holding remaining intervention components fixed and measures expected downstream negativity within k hops.
Social insights: Denials/corrections and evidence provision tend to reduce downstream negativity when introduced early in a thread.The interpretation is consistent with clarifying uncertainty reducing antagonistic replies.
Social insights: Toxicity/attack has a positive ATE on negativity, acting as an escalation trigger whose effect propagates beyond the immediate exchange.
Social insights: Questions/challenges and sarcasm show context-dependent effects across events.Questions can reduce or increase negativity depending on their form and event context.
Qualitative case studies: Removing denial/correction increased predicted negativity by Ri,denial = +0.18, equivalently ∆i,denial = −0.18 under treatment-on minus treatment-off.C3T selected the early correction as the dominant ancestor for a later neutral reply.
Qualitative case studies: Removing the attack label reduced predicted negativity by Ri,tox = −0.22, equivalently ∆i,tox = +0.22 under treatment-on minus treatment-off.The toxic ancestor received most attribution mass for multiple negative descendants.

8 Additional Analysis

Additional analyses delimit C3T’s evidence: its reported effects are short-range, conditional, model-based sensitivities, while ablations support several design choices without establishing effect-magnitude invariance.

Data and structural scope: The event-OOD evaluation uses disjoint rumor threads across training, development, and testing events, with attribution limited to the nearest 16 ancestors and two-hop descendants.Replies without candidate ancestors are excluded, and sources without two-hop descendants receive zero by convention.
Counterfactual estimand and assumptions: C3T’s counterfactual branch toggles one intervention embedding while holding text, reply structure, temporal context, and other interventions fixed.This estimates conditional, model-based sensitivity rather than simulating rewritten or deleted posts and their changed participation or structure.
Objective and ablations: Removing counterfactual loss reduces event-OOD sentiment macro-F1 from 60.4 to 58.1, shift macro-F1 from 54.8 to 52.9, and attribution Top-1/MRR from 36.2/52.1 to 32.4/48.5.Removing intervention tags yields 55.7, 49.6, and 25.1/38.9, while parent-only attribution yields 21.5/21.5.
Intervention provenance and label uncertainty: Intervention tags use calibrated, type-specific thresholds and abstention, excluding 4.2% of test nodes from the counterfactual loss.Remaining errors in sarcasm, indirect hostility, coded language, and ambiguous challenges may propagate into effect estimates; Fleiss’ κ = 0.75 concerns causal-source labels, not LLM tags.
Interpreting reported effects: Forcing denial/correction, evidence/link, or authority-citation embeddings on is associated with lower predicted two-hop negativity, whereas toxicity/attack is associated with higher negativity.The reported effects are −0.142, −0.118, −0.096, and +0.235 respectively, with bootstrap intervals not crossing zero; question/challenge and claim/assertion have no stable aggregate direction.
Robustness boundary: The reported effects do not establish stable magnitudes or rankings under permuted tags, placebo checks, alternative hyperparameters, or removal of counterfactual loss.The conclusions remain limited to the reported signs, intervals, predictive ablations, and selected examples.

9 Conclusion

The paper introduces CaSiRe and C3T to model sentiment, shifts, ancestor attribution, and intervention counterfactuals in rumor trees. Its conclusions remain conditional on observational data, bounded local interference, noisy labels, representation-level counterfactuals, and rumor-centric domains.

Conclusion: CaSiRe and C3T address sentiment, parent-child shifts, causal-ancestor attribution, and counterfactual intervention reasoning in rumor-centered conversation trees.C3T integrates thread structure, temporal context, and intervention-aware modeling under event-level out-of-distribution evaluation.
Limitations: The estimated intervention effects are conditional, model-based counterfactual estimates rather than randomized causal effects because unmeasured confounding may remain.Potential confounders include prior beliefs, offline exposure, community affiliation, and platform-specific norms.
Limitations: Bounded descendant modeling only approximates interference, excluding cross-branch spillovers, quote-post behavior, external updates, and recommendation-system effects.Exposure logs, cross-thread links, and temporal event timelines are suggested directions for modeling broader interaction effects.
Limitations: LLM-assisted intervention labels remain noisy for implicit, context-dependent, culturally specific moves such as sarcasm, indirect hostility, coded language, and ambiguous questions.Calibration, thresholding, abstention, and human validation are used to reduce some risks, but label noise can propagate into attribution and effect estimates.
Limitations: Sparse ancestor attribution identifies the most plausible modeled source within a bounded context, although important causes may lie outside the window or observed thread.Focused explanations may also underrepresent conversations with multiple interacting causes.
Limitations: Generalization beyond public rumor and crisis-style conversations is not guaranteed, motivating extensions across platforms, languages, and conversation genres.The authors specifically identify casual conversations, long-form forums, multilingual communities, political deliberation, and moderation-heavy platforms as potentially different settings.
Limitations: Forcing intervention embeddings on or off while fixing text and structure yields representation-level approximations, not fully realistic alternative conversation histories.Real interventions could change wording, participation, subsequent replies, and thread structure.
Broader impact: The framework should be used for diagnostic and aggregate analysis rather than automated moderation or behavioral manipulation.Attribution outputs could otherwise be misused to target or blame individual users, while modeling tools could optimize manipulative messaging or suppress disagreement.

A.5 LLMs for social-media analysis and weak supervision

The paper uses LLMs both as calibrated annotation aids and as deterministic prompt-only baselines for social-media sentiment and attribution. It evaluates whether conversational context enables reliable reasoning over branching threads.

LLMs as annotation aids and baselines: LLMs supply calibrated, thresholded, abstention-capable candidate intervention tags and serve as prompt-only sentiment and attribution baselines.The baselines use fixed templates, decoding settings, and context budgets.
Research question: The results test whether prompt-only LLMs can reason over long, branching threads or whether structure-aware models remain necessary.The paper argues that counterfactual sentiment reasoning requires explicit structure, time, intervention signals, and causal attribution.
Evaluation protocol: The implementation and evaluation cover dataset partitioning, CASIRE annotation, intervention tagging, causal-effect estimation, and LLM prompting baselines.All splits are defined at thread or event level so no conversation tree is shared across partitions.

B.1 Dataset Splits and Partitioning Protocols

The evaluation uses complementary thread-, event-, and platform-based splits, with Event-OOD as the primary protocol. Full reply-tree structures are preserved whenever reconstructable, and sentiment shifts are induced from parent–child sentiment labels.

Event-OOD split: Event-OOD is the primary evaluation setting, holding out all threads associated with one root event from training and development.Sydneysiege is test-only, ottawashooting is development, and the remaining events are used for training.
In-domain split: The in-domain protocol uses a stratified 80/10/10 thread-level split within the training-event pool, with no thread crossing partitions.Stratification approximately preserves imbalance in available node-level sentiment labels.
Cross-platform split: The cross-platform protocol trains on Twitter threads and tests on Reddit threads from the same events for robustness analysis, not final model selection.Training events are charliehebdo, ferguson, ottawashooting, and germanwings-crash.
Tree preservation: Full reply-tree structures are preserved, while threads with unreconstructable trees due to missing parent references or malformed timestamps are removed.This maintains structural information within valid conversation trees.
Label construction: Edge-level shift labels are deterministically induced from ordinal parent–child sentiment labels rather than separately annotated.The ordering is negative < neutral < positive, yielding upward, downward, or unchanged shifts.

B.2 C3T Hyperparameters and Compute

C3T is tuned on Event-OOD development sentiment macro-F1, with controlled hyperparameter exploration, fixed compute and prompting protocols, and structured intervention and attribution components. Its evaluation also uses calibrated intervention tags and conditional counterfactual effects.

Hyperparameter selection: The final C3T configuration is selected using Event-OOD development sentiment macro-F1 after exploring encoder layers L ∈ {2, 3} and ancestor windows D ∈ {8, 16, 32}.Main reported predictive scores are averaged over three random seeds.
Compute: C3T training uses one NVIDIA A100 GPU with 80GB memory, averaging approximately 3.5 hours per run.Baselines share data splits, preprocessing, maximum token budget, and applicable evaluation metrics.
Intervention tagging: The intervention taxonomy contains eight multi-label discourse moves, including assertion, correction, evidence, authority citation, toxicity, sarcasm, questioning, and derailment.Posts may receive multiple intervention labels because they can perform several functions simultaneously.
Intervention calibration: LLM-assisted intervention scores are temperature-calibrated and converted to labels using type-specific development-set thresholds.The thresholds are 0.65 for denial, 0.70 for evidence, 0.80 for toxicity, and 0.50 for remaining types.
Causal-effect estimation: Counterfactual intervention effects toggle one intervention type while holding observed node, structure, and other intervention features fixed.The resulting estimates are conditional, model-based counterfactual effects rather than randomized causal effects.
Attribution: Causal-source attribution selects the most plausible ancestor from a bounded candidate set using a normalized score distribution, with sparse normalization encouraging peaked assignments.Instances with unclear sources or empty ancestor sets are excluded from attribution training and evaluation.
LLM baselines: Open-weight LLM baselines use deterministic decoding, a 4,096-token context cap, and fixed context variants from node-only input through bounded ancestor summaries.Few-shot prompts reuse k = 8 fixed demonstrations across models and evaluation examples.