United States

This paper presents the English-language submissions of The Counterfactuals team for the three subtasks of Task 9 at SemEval 2026. The task aims to detect multicultural online polarization, how it is expressed, and in what contexts. The task provides a high-quality annotation dataset of posts that follows a three-level schema: polarized or not (subtask 1), polarization type classification (subtask 2), and manifestation identification (subtask 3). I construct a pointwise mutual information-based lexicon that identifies highly-correlated words with the polarized class as labeled in subtask 1. Using this lexicon, I implement a large language model data augmentation technique. I then use the preprocessed datasets to finetune a BERT model (BERTweet) for each subtask. My highest performing models placed 48th out of 60, 35th out of 36, and 17th out of 24 on subtasks 1, 2, and 3 respectively. All code is available on GitHub.

ACL 2026

The Counterfactuals at SemEval-2026 Task 9: Can Counterfactually-Inspired Preprocessing help Detect Polarization?

llm-assisted synthetic data

counterfactuality

polarization

It is my great pleasure and honor to welcome you to the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, California, USA, July 2–7, 2026. ACL 2026 continues our fields tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.

It is exciting to see the many developments in the field reflected at our conference, such as the advancement of large language models (born out of our community) into many application directions. This is not without challenges, partly due to unsolved problems, and partly due to misuse of the technology. All this is reflected well at ACL 2026. The conference also reflects that the field is not only about advancing technology, but also the science of fundamentally understanding language and how it is use by many different speaker communities. 
[Full Message from General Chair Philipp Koehn](https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing) 
<html><button style="display: inline-flex; align-items: center; justify-content: center; white-space: nowrap; border-radius: 9999px; font-weight: bold; background: #7c3aed; color: white; font-family: 'Space Grotesk', sans-serif; height: 40px; font-size: 16px; padding: 0 20px; border: none; cursor: pointer" onclick="window.open('https://drive.google.com/file/d/1QiJFOmMjvDgsDo3i1XzlQnImxe3RU_Cy/view?usp=sharing')">Conference Handbook</button></html>

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://2026.aclweb.org/registration/) first.

Registration Required

The 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026) will take place in San Diego, California, from July 2 - 7, 2026.

workshop paper

For the Chinese track of SemEval-2026 Task 9 (Detecting Online Polarization), we address two key challenges: polarized content frequently uses implicit language (e.g., homophones and coded terms) to evade moderation, and class distributions exhibit severe long-tail imbalance. We propose a metric learning approach that frames polarization detection as semantic similarity matching, which captures implicit language patterns better than linear decision boundaries. We fine-tune an ERNIE-3.0 encoder with SoftTriple loss and apply kNN retrieval for binary detection (Subtask 1). For multi-label categorization (Subtasks 2 and 3), we transfer learned representations from the detection model and fine-tune with Asymmetric Loss. A priority-based stratified cross-validation strategy ensures minority classes appear across all training folds despite extreme label skew. Evaluated on the official 1,927-sample test set using an end-to-end pipeline, our system achieved Macro-F1 scores of 0.9190 (Rank 6) on Polarization Detection, 0.8244 (Rank 5) on Type Classification, and 0.6670 (Rank 4) on Manifestation Identification.

AlphaLyrae at SemEval-2026 Task 9: Metric Learning and Asymmetric Loss for Chinese Polarization Analysis

This paper presents the Duluth approach to
SemEval-2026 Task 6 on CLARITY: Unmask-
ing Political Question Evasions. We address
Task 1 (clarity-level classification) and Task 2
(evasion-level classification), both of which in-
volve classifying question–answer pairs from
U.S. presidential interviews using a two-level
taxonomy of response clarity. Our system is
based on DeBERTa-V3-base, extended with
focal loss, layer-wise learning rate decay, and
boolean discourse features. To address class
imbalance in the training data, we augment
minority classes using synthetic examples gen-
erated by Gemini 3 and Claude Sonnet 4.5. Our
best configuration achieved a Macro F1 of 0.76
on the Task 1 evaluation set, placing 8th out
of 40 teams. The top-ranked system (TeleAI)
achieved 0.89, while the mean score across par-
ticipants was 0.70. Error analysis reveals that
the dominant source of misclassification is con-
fusion between Ambivalent and Clear Reply
responses, a pattern that mirrors disagreements
among human annotators. Our findings demon-
strate that LLM-based data augmentation can
meaningfully improve minority-class recall on
nuanced political discourse tasks.

Duluth at SemEval-2026 Task 6: DeBERTa with LLM-Augmented Data for Unmasking Political Question Evasions

We present a lightweight, feature-based regression system for predicting \textbf{valence} (pleasantness) and \textbf{arousal} (activation) from longitudinal language data. The language data ranges from longer free-form ecological essays to short affect-word, organized by user and time, reflecting natural variation in affective expression and experience. Our approach combines three complementary signals: (i) sentence-level semantic embeddings, (ii) psycholinguistic category features capturing affect- and function-related word usage, (iii) similarity measures between the language data with archetypal sentences, and (iv) trainable user-embeddings to account for between-user differences. The resulting feature vector is passed to a multi-layer perceptron trained to jointly predict valence and arousal. Our design provides a strong and interpretable baseline by making it possible to isolate the contribution of semantic, psycholinguistic, similarity, and user-specific signals. We further analyze our model's predictions to identify which feature groups are most informative and where errors are concentrated across users and input types.

McMaster NLP at SemEval-2026 Task 2: A Lightweight Multi-Feature System for Predicting Emotional Valence and Arousal over Time

We describe our system for SemEval-2026 Task 6 (CLARITY: Unmasking Political Question Evasions), which classifies English political interview responses by coarse-grained clarity (3-way) and fine-grained evasion strategy (9-way). Since responses frequently exceed the 512-token limit of standard Transformer encoders, we apply an overlapping sliding-window chunking strategy with element-wise Max-Pooling aggregation over chunk representations. A shared RoBERTa-large encoder supplies two task-specific heads trained jointly via a multi-task objective, with inference-time ensembling over 7-fold stratified cross-validation. Our system achieves a Macro-F1 of 0.80 on Subtask 1 and 0.51 on Subtask 2, ranking 11th in both subtasks.

SG-UniBuc-NLP at SemEval-2026 Task 6: Multi-Head RoBERTa with Chunking for Long-Context Evasion Detection

This paper presents approach to narrative similarity prediction for SemEval-2026 Task 4 Track A. We introduce an LLM-based system that operationalizes the three core dimensions—Abstract Theme, Course of Action, and Outcomes—via schema-constrained prompting to enforce structured outputs and alignment with the annotation protocol. The system proceeds in three stages: structured aspect decomposition and scoring, weak-signal gating for low-confidence cases, and a targeted LLM-based tiebreak. The final model achieved near-human performance and ranked second on the Track A leaderboard.

FactUEP at SemEval-2026 Task 4: Structured Narrative Similarity Scoring with Aspect Decomposition and Weak-Signal Gating

Machine-generated code detection is examined under out-of-distribution conditions where robust generalization is required. A hybrid feature representation is used in which code snippets are encoded through character-level TF–IDF patterns together with explicit structural indicators capturing properties such as verbosity and formatting behavior. Variability across generators is handled through clustering-based expert specialization, and predictions are produced using an ensemble of logistic regression and Naïve Bayes models with calibrated thresholds. Experimental results show that the proposed approach performs competitively despite relying on simple linear classifiers. The findings suggest that persistent structural patterns in code provide reliable cross-domain signals for identifying machine-generated programs.

Stylometry at SemEval-2026 Task 13: Clustered Stylometric Modeling for Machine-Generated Code Detection

We present our systems for SemEval-2026 Task 13, built on the Droid resource suite and benchmark setting. For Subtask A (binary classification of human-written vs. machine-generated code), lexical baselines such as TF–IDF and character n-grams transferred poorly from the LeetCode training distribution to the production-code evaluation split. After correcting pipeline errors that obscured true performance and selecting stable AST features under domain shift, our final system uses 5 uncorrelated features and achieves 0.57 macro F1 on the public test set.
For Subtask C (4-way authorship classification of human, AI, hybrid, and adversarial) lexical baselines performed poorly under a significant vocabulary shift. Deep semantic models proved more promising, and a per-class weighted ensemble which included these models achieved 0.57 macro F1 on the public test set

HyperparameterOmens at SemEval-2026 Task 13: Various approaches to detecting machine- generated code

We present a hybrid system for SemEval-2026 Task 5 on graded word-sense plausibility in narrative contexts. Our approach combines prompt-based large language model (LLM) scoring with three complementary features: semantic embedding similarity, story-conditioned definition generation, and a synthetic gaze signal based on predicted fixation time. We combine these signals using an ordinary least squares regressor. On the official test set, our best system achieves 90.10 Acc±SD and 79.19 Spearman correlation. The system surpasses the reported human reference score on Acc±SD, highlighting the value of combining LLM-based judgments with targeted linguistic and cognitive-inspired features.

CiNet-Handai-Kyodai at SemEval-2026 Task 5: Combining LLM Prompting, Semantic Similarity, and Synthetic Gaze for Graded Sense Plausibility

We present a winning three-stage system for SemEval 2026 Task 12: Abductive Event Reasoning that combines graph-based retrieval, LLM-driven abductive reasoning with prompt design informed by reflective prompt evolution, and post-hoc consistency enforcement; our system ranks first on the evaluation-phase leaderboard with an accuracy score of 0.95. Cross-model error analysis across 14 models (7 families) reveals three shared inductive biases: causal chain incompleteness, proximate cause preference, and salience bias, whose cross-family convergence (51% cause-count reduction) indicates systematic rather than model-specific failure modes in multi-label causal reasoning.

AILS-NTUA at SemEval-2026 Task 12: Graph-Based Retrieval and Reflective Prompting for Abductive Event Reasoning

We describe our system for SemEval-2026 Task 13 on detecting machine-generated code across eight programming languages and three subtasks: binary human-vs-AI detection, 11-way source identification, and 4-way generator classification. Our approach uses a task-specific combination of Qwen2.5-Coder-1.5B with LoRA fine-tuning, abstract syntax tree (AST) features, CodeBERT with head-tail chunking, and TF-IDF features. Experiments reveal three main findings. For Task A, neural detectors degrade markedly on the official test split, while AST-based structural features remain more stable, suggesting substantial distribution shift. For Task B, inverse-frequency class weighting is essential under extreme label imbalance and greatly improves macro-F1. For Task C, combining neural and statistical models performs better than relying on a single model alone, indicating complementary strengths across representations. Our final system achieves 0.638 macro-F1 on Task A, 0.449 macro-F1 on Task B, and 0.714 macro-F1 on Task C, offering practical insights into robustness, imbalance handling, and model complementarity for AI-generated code detection.

Downloads

Next from ACL 2026

AlphaLyrae at SemEval-2026 Task 9: Metric Learning and Asymmetric Loss for Chinese Polarization Analysis

Similar lecture

SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2026

AlphaLyrae at SemEval-2026 Task 9: Metric Learning and Asymmetric Loss for Chinese Polarization Analysis

Similar lecture

SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization

Downloads