📄 CSECU-DSG at SemEval-2026 Task 6: Imbalance-Aware Transformers for Unmasking Political Question Evasions

Summary

The paper describes the CSECU-DSG system for SemEval-2026 Task 6 (CLARITY), Subtask 1. The goal is to label how clearly a politician's answer responds to an interview question, using three classes: Clear Reply, Ambivalent Reply and Clear Non-Reply. The data is the QEvasion dataset, drawn from U.S. presidential news conferences. Its label distribution is skewed, with ambivalent answers far more common than the other two classes.

The authors treat the task as sentence-pair classification. Each question and answer are concatenated and fed into a fine-tuned DeBERTa-v3-base model, and the [CLS] representation produces the prediction. To counter the skew, they combine two techniques:

• oversampling minority classes by repetition;
• a class-weighted cross-entropy loss.

To keep computation cheap, inputs are truncated at the character level. The system reached a macro F1 of 0.67 and placed 31st, well behind the winning system at 0.89. The work is a modest but transparent baseline showing what a single transformer with imbalance-aware training achieves on detecting political evasion.

Key points

• Dataset: QEvasion covers press conferences of four presidents (G. W. Bush, Obama, Trump, Biden). It has about 3,700 annotated pairs: 3,448 for training and roughly 300 for testing. It is distributed in Parquet format via Hugging Face datasets.
• Input and training setup:
  • Questions are cut to 80 characters and answers to 120, with a maximum length of 192 tokens.
  • Training runs for 10 epochs with AdamW, learning rate 2e-5, weight decay 0.01 and 10% warmup.
  • The best checkpoint is chosen by validation macro F1.
• Imbalance handling:
  • Each class is oversampled to 1,500 examples (about 4,500 in total).
  • Class weights are computed as w_c = N/(k·n_c).
  • Oversampling was done before splitting, so the authors acknowledge possible leakage between splits. They therefore report results on a separate stratified 85/15 split of the original data.
• Results: The system scored 0.67 macro F1 and ranked 31st. The top four teams were TeleAI (0.89), AsymVerify (0.85), CSE-UOI (0.85) and Rasende Rakete (0.83).
• Not covered: The team took part only in Subtask 1 and did not attempt the fine-grained nine-class evasion subtask. The abstract mentions a regression formulation, but the paper does not describe one.