Elamur Research⌕
☰
home / papers / 2026 / RES2 at IMPOLS 2026: Implicit Content Classification via Span-Awareness, Hierarchical Multi-Task Learning and Synthetic Data

RES2 at IMPOLS 2026: Implicit Content Classification via Span-Awareness, Hierarchical Multi-Task Learning and Synthetic Data

Paper · 2026 · Salvatore Spezia, Salvatore Ferrara, Roberto Dioguardi et al.

Summary

The RES2 team's system for IMPOLS at EVALITA 2026 detects and classifies implicit content in Italian political speeches. The IMPOLS data comes from the multimodal IMPAQTS corpus. The challenge has three subtasks: detecting whether implicit content is present, telling implicatures from presuppositions, and assigning implicature subtypes. The team fine-tuned the encoder-only UmBERTo model and used text only. The annotated trigger span is wrapped in special marker tokens. The model's [CLS] embedding, which captures overall context, is joined with a mean-pooled embedding of the span, and an MLP classifier makes the prediction from both. Each finer-grained subtask has much less data, so the team took two countermeasures. First, they generated synthetic training data with Gemini 3.0, rewriting the context while leaving the trigger untouched. Second, for subtask 3 they trained a multi-task model whose three heads mirror the task hierarchy. The system ranked first in subtasks 2 and 3 and second in subtask 1. This suggests that, in specialized pragmatic tasks, a fine-tuned encoder can beat general-purpose generative LLMs. Audio features were tested and then dropped because they added little.

Key points

  • Data sizes: the balanced training sets shrink from 1,824 samples (subtask 1) to 912 (subtask 2) and 456 (subtask 3). Two Gemini paraphrases per sample tripled the training data for subtasks 2 and 3, with the largest gain on subtask 3.
  • Model setup: the backbone is UmBERTo-commoncrawl-cased-v1, a RoBERTa-based model with a SentencePiece tokenizer and 768-dimensional embeddings. Trigger spans are marked with <TGT_START>/<TGT_END>. Training used stratified 5-fold cross-validation.
  • Multi-task loss: subtasks 1 and 2 serve as auxiliary tasks for subtask 3. Their losses are weighted λ1=1.5, λ2=0.5 and λ3=5.0.
  • Audio experiments: the team tried Wav2Vec 2.0 (large-xlsr-53-italian), WhisperX alignment and cross-attention fusion, including a re-implementation of MTAMW.
    • An audio-only model reached just 0.5917 macro-F1 on subtask 1.
    • The multimodal model reached 0.9636, but removing audio at inference barely changed the result.
    • The authors attribute this to speaker bias and unreliable alignment.
  • Limitation: the system depends on gold trigger spans supplied by the organizers. Automatic trigger detection is left for future work.

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

 Fields, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

🧭 Topics 1Research subjects and languages the work is about.Implicature · 138 works

Cite

Cite as Salvatore Spezia, Salvatore Ferrara, Roberto Dioguardi, Ernesto Davì, Irene Siragusa and Roberto Pirrone. 2026. RES2 at IMPOLS 2026: Implicit Content Classification via Span-Awareness, Hierarchical Multi-Task Learning and Synthetic Data. Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{spezia2026res2,
  title = {{RES2 at IMPOLS 2026: Implicit Content Classification via Span-Awareness, Hierarchical Multi-Task Learning and Synthetic Data}},
  author = {Salvatore Spezia, Salvatore Ferrara, Roberto Dioguardi, Ernesto Davì, Irene Siragusa and Roberto Pirrone},
  booktitle = {Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026)},
  year = {2026},
  url = {https://aclanthology.org/2026.evalita-1.70/}
}

References

34 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. L. Gregori, W. Paci, V. Saccone, IMPOLS at EVALITA 2026: Overview of the IMplicit contents in POLitical Speech Task, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026), CEUR.org, Bari, Italy, 2026. find
  2. F. Cutugno, A. Miaschi, A. P. Aprosio, G. Rambelli, L. Siciliani, M. A. Stranisci, Evalita 2026: Overview of the 9th evaluation campaign of natural language processing and speech tools for italian, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for I… find
  3. G. Comandini, G. Ferrari, A. Panunzi, Impaqts: A multimodal corpus of implicit manipulatory strategies in political speech, in: Proceedings of the Ninth Italian Conference on Computational Linguistics (CLiC-it 2023), CEUR-WS.org, Venice, Italy, 2023. URL: https://ceur-ws.org/Vol-3596/. ↗
  4. W. Paci, A. Panunzi, S. Pezzelle, They want to pretend not to understand: The limits of current llms in interpreting implicit content of political discourse, in: Findings of the Association for Computational Linguistics: ACL 2025, Association for Computational Linguistics, Vienna, Austria, 2025, pp.… find
  5. L. Parisi, S. Francia, P. Magnani, Umberto: an italian language model trained with whole word masking, https://github.com/musixmatchresearch/umberto, 2020. ↗
  6. Google DeepMind, Gemini 3.0: A new era of intelligence, 2025. URL: https://deepmind.google/ gemini, technical Report. Retrieved from https://deepmind.google/gemini. ↗
  7. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, I. Polosukhin, Attention is All you Need, in: Advances in Neural Information Processing Systems, 2017.
  8. M. Joshi, D. Chen, Y. Liu, D. S. Weld, L. Zettlemoyer, O. Levy, SpanBERT: Improving pre-training by representing and predicting spans, Transactions of the Association for Computational Linguistics 8 (2020) 64–77. URL: https://aclanthology.org/2020.tacl-1.5/. doi: 10.1162/tacl_a_00300 . ↗
  9. L. Baldini Soares, N. FitzGerald, J. Ling, T. Kwiatkowski, Matching the blanks: Distributional similarity for relation learning, in: A. Korhonen, D. Traum, L. Màrquez (Eds.), Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Lingui…
  10. D. Hendrycks, K. Gimpel, Gaussian error linear units (gelus), 2023. URL: https://arxiv.org/abs/1606. 08415. arXiv:1606.08415 . ↗
  11. D. I. Beaver, B. Geurts, Presupposition, Stanford Encyclopedia of Philosophy, 2011. URL: https://plato.stanford.edu/entries/presupposition/, first published 2011; substantive revision 2021. ↗
  12. C. N. Silla, A. A. Freitas, A survey of hierarchical classification across different application domains, Data Mining and Knowledge Discovery 22 (2011) 31–72. find
  13. R. Caruana, Multitask learning, Machine Learning 28 (1997) 41–75. find
  14. L. Liebel, M. Körner, Auxiliary tasks in multi-task learning, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018, pp. 1–10. find
  15. Y. Liu, M. Ott, N. Goyal, J. Du, M. Joshi, et al., RoBERTa: A Robustly Optimized BERT Pretraining Approach, arXiv preprint arXiv:1907.11692 (2019).
  16. T. Kudo, J. Richardson, Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing, in: Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, Association for Computational Linguistics, Brussel… — [doi:10.18653/v1/d18-](https://doi.org/10.18653/v1/d18-)
  17. T. Akiba, S. Sano, T. Yanase, T. Ohta, M. Koyama, Optuna: A next-generation hyperparameter optimization framework, in: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, ACM, Anchorage, AK, USA, 2019, pp. 2623–2631. doi: 10. 1145/3292500.3330701 .
  18. A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, et al., Pytorch: An imperative style, high-performance deep learning library, in: Advances in Neural Information Processing Systems, volume 32, Curran Associates, Inc., 2019, pp. 8024–8035. URL: https://proceedings.neurips.cc/ paper/90… ↗
  19. J. Bai, S. Bai, Y. Chu, Z. Cui, K. Dang, et al., Qwen Technical Report, arXiv preprint arXiv:2309.16609 (2023).
  20. S. Schweter, Italian bert and electra models, 2020. URL: https://github.com/dbmdz/berts. doi: 10. 5281/zenodo.4271493 . ↗
  21. P. He, J. Gao, W. Chen, Debertav3: Improving deberta using electra-style pre-training with gradient-disentangled embedding sharing, in: Eleventh International Conference on Learning Representations (ICLR), 2023. URL: https://openreview.net/forum?id=sE7-XhLxHA. ↗
  22. A. Conneau, K. Khandelwal, N. Goyal, V. Chaudhary, G. Wenzek, F. Guzmán, et al., Unsupervised cross-lingual representation learning at scale, in: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 8440–8451. find
  23. D. Atzeni, M. Guredda, A. Moretti, Gilberto: a pre-trained language model for italian, 2020. URL: https://github.com/idb-ita/GilBERTo. ↗
  24. P. Boersma, D. Weenink, Praat: doing phonetics by computer [computer program], 2024. URL: http://www.praat.org/, retrieved from http://www.praat.org/. ↗
  25. B. McFee, C. Raffel, D. Liang, D. P. W. Ellis, M. McVicar, E. Battenberg, O. Nieto, librosa: Audio and music signal analysis in python, in: K. Huff, J. Bergstra (Eds.), Proceedings of the 14th Python in Science Conference, 2015, pp. 18–25. doi: 10.25080/Majora- 7b98e3ed- 003 . — [doi:10.25080/majora-](https://doi.org/10.25080/majora-) find
  26. F. Eyben, K. R. Scherer, B. W. Schuller, J. Sundberg, E. André, et al., The geneva minimalistic acoustic parameter set (gemaps) for voice research and affective computing, IEEE Transactions on Affective Computing 7 (2016) 190–202. doi: 10.1109/TAFFC.2015.2457417 . — [doi:10.1109/taffc.2015.2457417](https://doi.org/10.1109/taffc.2015.2457417) find
  27. F. Eyben, M. Wöllmer, B. Schuller, opensmile: The munich versatile and fast open-source audio feature extractor, in: Proceedings of the 18th ACM International Conference on Multimedia, ACM, Firenze, Italy, 2010, pp. 1459–1462. doi: 10.1145/1873951.1874246 . — [doi:10.1145/1873951.1874246](https://doi.org/10.1145/1873951.1874246) find
  28. A. Baevski, Y. Zhou, A. Mohamed, M. Auli, wav2vec 2.0: A framework for self-supervised learning of speech representations, in: H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, H. Lin (Eds.), Advances in Neural Information Processing Systems (NeurIPS), volume 33, 2020, pp. 12449–12460. URL: https://…
  29. S. Chen, C. Wang, Z. Chen, Y. Wu, S. Liu, et al., Wavlm: Large-scale self-supervised pre-training for full stack speech processing, IEEE Journal of Selected Topics in Signal Processing 16 (2022) 1505–1518. doi: 10.1109/JSTSP.2022.3188113 . find
  30. M. Bain, J. Huh, T. Han, A. Zisserman, Whisperx: Time-accurate speech transcription of long-form audio, in: INTERSPEECH 2023, ISCA, 2023, pp. 148–152. URL: https://doi.org/10.21437/Interspeech. 2023-78. doi: 10.21437/Interspeech.2023- 78 . ↗
  31. Y.-H. H. Tsai, S. Bai, P. P. Liang, J. Z. Kolter, L.-P. Morency, R. Salakhutdinov, Multimodal transformer for unaligned multimodal language sequences, in: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Association for Computational Linguistics, Florence, Ita…
  32. K. Yang, H. Xu, K. Gao, Cm-bert: Cross-modal bert for text-audio sentiment analysis, in: Proceedings of the 28th ACM International Conference on Multimedia, ACM, Seattle, WA, USA, 2020, pp. 521–528. doi: 10.1145/3394171.3413690 . — [doi:10.1145/3394171.3413690](https://doi.org/10.1145/3394171.3413690) find
  33. N.-C. Ristea, A. Anghel, R. T. Ionescu, Cascaded cross-modal transformer for audio-textual classification, arXiv preprint arXiv:2401.07575 (2024). URL: https://arxiv.org/abs/2401.07575. — [arXiv:2401.07575](https://arxiv.org/abs/2401.07575) ↗
  34. Y. Wang, J. He, D. Wang, Q. Wang, B. Wan, X. Luo, Multimodal transformer with adaptive modality weighting for multimodal sentiment analysis, Neurocomputing 572 (2024) 127181. doi: 10.1016/j.neucom.2023.127181 . — [doi:10.1016/j.neucom.2023.127181](https://doi.org/10.1016/j.neucom.2023.127181) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/spezia-2026-res2-impols-2026-implicit-content/