Elamur Research⌕
☰
home / papers / 2026 / SilkPeak at SemEval-2026 Task 6: When Politicians Dodge –- Unmasking Evasion in Political Interviews through Joint Multi-Task Transformer Learning

SilkPeak at SemEval-2026 Task 6: When Politicians Dodge –- Unmasking Evasion in Political Interviews through Joint Multi-Task Transformer Learning

Paper · 2026 · Amruth Tetakali and Lavanya Tetakali

Summary

The paper describes the SilkPeak entry to SemEval-2026 Task 6 (CLARITY), which targets the automatic detection of evasive answers in political interviews. The task has two levels. Subtask 1 classifies an answer as a Clear Reply, Ambivalent, or Clear Non-Reply. Subtask 2 assigns one of nine evasion strategies, such as Dodging, Deflection or Claims Ignorance. The authors observe that the coarse clarity labels are essentially a merged form of the fine-grained strategy labels, so they train both subtasks jointly. A single DeBERTa-v3-Large encoder reads the concatenated question and answer. Two linear heads sit on the [CLS] vector, and the two cross-entropy losses are summed so that both shape the shared encoder. On the official Task 1 evaluation this reaches 0.76 macro F1, compared with 0.58 for a single-task DeBERTa-Base baseline. It also outperforms every alternative the authors tried, including a LoRA-tuned LLaMA-3-8B. The work suggests that joint supervision and full use of the data matter more than model size for small pragmatic classification datasets.

Key points

  • Hard parameter sharing (one encoder, a 3-way head and a 9-way head, unweighted loss sum) gives 0.76 macro F1 on Task 1 in evaluation, against 0.63 on dev. Removing joint training lowers this to 0.58.
  • Alternative designs all scored lower on the eval set:
    • Hierarchical bi-encoder with cross-attention: 0.53 (0.51 in ablation).
    • Weighted pooling of all 24 layers plus multi-sample dropout: 0.59 in ablation.
    • 5-fold ensemble: 0.54 in ablation.
  • The generative LLaMA-3-8B with LoRA reached only 0.45 after 60 steps and 0.49 after 450 steps. Its errors frequently confused Deflection with General and Implicit with Partial/Half-Answer.
  • On the 308-item public test split, Ambivalent scores best (F1 0.76) and Clear Reply worst (F1 0.52). 35 of the 79 Clear Reply items are labelled Ambivalent, often because a long hedging preamble precedes the actual answer.
  • The training data contains roughly 3,000 English examples (QEvasion dataset). The authors see data volume as the main bottleneck and propose augmentation and weighted task losses as next steps. Code is available on GitHub (amruth6002/clarity).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Indian Institute of Technology RoorkeeRoorkee · India3 works · 6 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksGeneral Social Sciences · 227 works

Cite

Cite as Amruth Tetakali and Lavanya Tetakali. 2026. SilkPeak at SemEval-2026 Task 6: When Politicians Dodge –- Unmasking Evasion in Political Interviews through Joint Multi-Task Transformer Learning. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.253

@inproceedings{tetakali2026silkpeak,
  title = {{SilkPeak at SemEval-2026 Task 6: When Politicians Dodge –- Unmasking Evasion in Political Interviews through Joint Multi-Task Transformer Learning}},
  author = {Amruth Tetakali and Lavanya Tetakali},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.253},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.253}
}

References

13 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Peter Bull. 2003. The Microanalysis of Political Communication: Claptrap and Ambiguity. Routledge, London, UK. find
  2. Rich Caruana. 1997. Multitask learning. Machine Learning, 28:41–75. find
  3. Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pre-training text encoders as discriminators rather than generators. In International Conference on Learning Representations. find
  4. Steven Clayman and John Heritage. 2002. The News Interview: Journalists and Public Figures on the Air. Cambridge University Press, Cambridge, UK. find
  5. Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. DeBERTa: Decoding-enhanced bert with disentangled attention. In International Conference on Learning Representations.
  6. Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  7. Kian Kenyon-Dean and 1 others. 2018. Sentiment, evasion, and discourse structure in political speech. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics. find
  8. Tanya Ko and 1 others. 2020. Answering questions about charts and generating data questions using transfer learning. In Proceedings of KDD. find
  9. Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. 2019. MT-DNN: Multi-task deep neural networks for natural language understanding. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4487–4496. Association for Computational Linguistics.
  10. Xiaofei Sun, Xiaoya Ji, Baohang Ma, and 1 others. 2023. Text classification via large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023. Association for Computational Linguistics.
  11. Konstantinos Thomas, Giorgos Filandrianos, Maria Lymperaiou, Chrysoula Zerva, and Giorgos Stamou. 2024. “I never said that”: A dataset, taxonomy and baselines on response clarity classification. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 5204–5233, Miami, Florida… find
  12. Konstantinos Thomas, Giorgos Filandrianos, Maria Lymperaiou, Chrysoula Zerva, and Giorgos Stamou. 2026. Semeval-2026 task 6: CLARITY – unmasking political question evasions. Preprint, arXiv:2603.14027.
  13. Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tetakali-2026-silkpeak-semeval-2026-task-6/