Distinguishing legitimate critical speech from coordinated conspiracy narratives is a pressing challenge for the integrity of the online information ecosystem, yet prevailing approaches rely on labeled entity roles and rigid predefined schemas that fail to capture the diffuse and context-dependent nature of participation in such narratives. We propose a modular framework that extracts an entity-interaction graph directly from narratives and reasons over it for classification, without depending on annotated node-level labels. Our pipeline enumerates candidate entity spans from a domain-adapted BERT encoder and learns a bipartite graph structure, which is then classified with a Heterogeneous Graph Transformer. To make predictions interpretable, we cast the graph as a hypergraph and apply a counterfactual distillation procedure that estimates the Individual Treatment Effect of each node, pruning extraneous entities to recover a minimal predictive subgraph. On the standard PAN Oppositional Narratives benchmark, our model achieves a macro F1-score of 0.93 and an MCC of 0.84, reaching state-of-the-art performance with roughly one-third of the parameters of the next best system, while pruning the graphs by 75.2% on average without altering predictions. Our results demonstrate that graph-based representations coupled with counterfactual inference offer both strong predictive accuracy and transparency into the entities driving conspiratorial discourse.
Figures & tables
Figure 1: Overview of the proposed methodology. Stage 1 defines the input, which in this case consists of oppositional narratives. Stage 2 generates tokens and contextual embeddings using BERT. Stage 3 extracts the entities V from the tokens and embeddings by applying Algorithm 1 (Appendix A ). Stage 4 produces enriched low-dimensional representations of the entities. Stage 5 constructs the incidence matrix A . Stage 6 performs graph classification. The final stage presents the counterfactual analysis. The key point is that every classification decision is routed through an explicit entity-interaction graph, which makes the counterfactual distillation of the last stage possible. Entity labels are abbreviated for readability; actual nodes are token spans (cf. Figure 2 ).
Component
Parameters
Percentage (%)
BERT (frozen)
109,483,778
99.841
Enricher
2,240
0.002
HyperLink classifier
759
0.001
Hypergraph classifier
126,745
0.116
Entity projector
24,592
0.022
Span classifier
11,525
0.011
Table 1: Decomposition of the total number of model parameters by component. Over 99.8% of the parameters live in the frozen text encoder: the entire graph construction and reasoning machinery adds fewer than 200K trainable parameters.
Team
MCC ↑
F1 -Macro ↑
F1 -Cons. ↑
F1 -Crit. ↑
Our model
0.840
0.930
0.895
0.945
IUCL
0.838
0.919
0.894
0.944
AI_Fusion
0.830
0.914
0.886
0.942
SINAI
0.829
0.914
0.888
0.941
Fine-tuned BERT
0.796
0.897
0.863
0.931
BERT+LOCO
0.655
0.844
0.787
0.901
Table 2: Classification results of our model compared to the PAN benchmark [ 21 ] . Our model outperforms IUCL [ 30 ] , AI_Fusion [ 1 ] and SINAI [ 39 ] on every metric while using roughly one-third of the parameters of the runner-up. Fine-tuned BERT is adapted to the specific context of the target class, whereas BERT+LOCO is a fine-tuned version trained on a larger dataset, enabling broader domain adaptation across several types of conspiracy-related texts.
Figure 2: Per-entity ITE contributions for an oppositional narrative text correctly classified as conspiracy, with a predicted probability of 0.836. A small number of entities account for most of the positive contribution which allows the graph to be pruned aggressively without flipping the prediction.
Ablation Approach
F1 -Score
Relative Decrease
Original model
0.9300
–
No enrichment module
0.8600
7.53%
No entity classifier
0.5256
43.48%
No Gumbel-Sigmoid module
0.3293
64.59%
Table 3: Ablation study results and performance decrease. Learning the graph structure with Gumbel-Sigmoid is the single most important component: forcing a fully connected graph costs almost two-thirds of the model’s performance.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Description
Span and entity representation
max_span_width
5
Maximum number of tokens allowed in a candidate span.
num_span_type
4
Number of span/entity categories considered by the model.
entity_dim
16
Dimensionality to represent each entity.
dropout_rate
0.15
Dropout probability applied for regularization.
Hyperlink module
Appendix
Table 4: Hyperparameters of the experiment and their roles in the architecture. Note that the hyperparameter num_hyperlink refers to the set size k , and all hyperlink references correspond to the set E .
Causal graphs in text are typically populated by observable, predefined events. In contrast, we study implicit causal graph construction from text by treating each described cause-effect pair as the begin- and endpoint of an underlying latent causal graph and using large language models (LLMs) to infer intermediate causal events. We compare end-to-end graph construction with methods that frame the task as causal chain discovery. In the latter, graphs are built either by aggregating inferred chains or by progressively expanding partial chains through an iterative search process. We further explore Wisdom of the Crowd extensions that access causal knowledge from multiple LLMs in post-hoc aggregation and collaborative inference settings. We analyze trade-offs among these approaches and evaluate the validity of inferred causal relations using a manually curated database of 1,560 scientifically validated causal pairs. This database-based evaluation is proposed as reliable, resource-efficient, and transferable to settings where ground-truth graphs are unavailable.
Liesbeth Allein, Marie-Francine Moens
Department of Computer Science, KU Leuven, Belgium · Department of Electronics and Information Systems, Ghent University, Belgium
A fundamental obstacle to causal discovery from text is the lack of causally annotated text data for use as ground truth, due to high annotation costs. This motivates an important task of generating text with causal graph annotations. Early template-based generation methods sacrifice text naturalness in exchange for high causal graph annotation accuracy. Recent Large Language Model (LLM)-dependent methods directly generate natural text from target graphs through LLMs, but do not guarantee causal graph annotation accuracy. Therefore, we propose iTAG, which performs real-world concept assignment to nodes before converting causal graphs into text in existing LLM-dependent methods. iTAG frames this process as an inverse problem with the causal graph as the target, iteratively examining and refining concept selection through Chain-of-Thought (CoT) reasoning so that the induced relations between concepts are as consistent as possible with the target causal relationships described by the causal graph. iTAG demonstrates both extremely high annotation accuracy and naturalness across extensive tests, and the results of testing text-based causal discovery algorithms with the generated data show high statistical correlation with real-world data. This suggests that iTAG-generated data can serve as a practical surrogate for scalable benchmarking of text-based causal discovery algorithms.
Wenshuo Wang, Boyu Cao, Nan Zhuang +1
School of Future Technology, South China University of Technology, Guangzhou, China · China Mobile GBA Innovation Institute, Guangzhou, China · Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
Touché 2026 extends causality extraction to counter-causal claims: news sentences whose surface form appears causal but whose meaning denies the causation, as in "It is falsely believed that X caused Y." A system that relies on surface cues such as "caused" or "led to" will accept such a sentence as causal and give it the wrong polarity. On the Countercausal News Corpus (CCNC), the task has three subtasks: deciding whether a sentence is causal (detection), locating its cause and effect spans (extraction), and labeling its polarity as procausal, counter-causal, or uncausal. We build one model per subtask. Detection is a fine-tuned classifier with a single cross-task rule that uses the extracted spans to remove false positives. For extraction, we ensemble three RoBERTa-large BILOU+CRF taggers by averaging their token-level scores before decoding, rather than voting on the spans each tagger produces. For polarity, where labeled counter-causal examples are scarcest, we add training sentences generated by a large language model prompted with nine patterns of counter-causal expression adapted from Hagen et al., keeping only those that pass automatic structural checks. On the held-out CCNC test set, the system reaches F1 0.869 on detection and macro-F1 0.817 on polarity, and in the organizers' final causal-only evaluation of extraction it scores granularity-adjusted F1 0.728, the highest extraction score among all submissions including the organizers' baseline. The development split is used only for component selection and the ablations reported in the paper.
Roham Zendehdel Nobari, Shayan Sooratgar
Department of Computational Linguistics, University of Zurich, Switzerland