Elamur Research⌕
☰
home / papers / 2025 / Act2P: LLM-Driven Online Dialogue Act Classification for Power Analysis

Act2P: LLM-Driven Online Dialogue Act Classification for Power Analysis

Paper · 2025 · Wenbo Zhang and Yuhan Wang

Summary

Act2P tackles two linked problems. First, it labels dialogue acts in a conversation as it unfolds, without any annotated training data. Second, it uses those labels to track how power shifts among speakers. The framework runs Deepseek-v3 zero-shot and improves it in three steps: engineered prompts (category descriptions, added context), a coarse-to-fine hierarchical classification, and an online feedback step in which the latest prediction can revise earlier labels. A graph-based algorithm, PowerRank, gives each dialogue act a power weight based on its pragmatic function. It then follows interactions between speakers to compute each person's changing power score. Each refinement raised accuracy at all three label granularities. Supervised baselines remain more accurate (about 92% against 84.5% on the coarsest labels), but Act2P needs almost no manual labels. Visualizing a multi-party academic meeting shows the professor generally keeps the leading power position while other participants fluctuate. The work matters because it links speech-act analysis to social structure and does so in a streaming, annotation-free setting.

Key points

  • Accuracy with Deepseek-v3 on Basic_label / General_label / Full_label:
    • zero-shot: 70.30 / 60.65 / 29.12
    • with prompt optimization: 80.34 / 70.23 / 36.43
    • with hierarchical classification: 82.83 / 73.56 / 44.00
    • with online feedback: 84.53 / 75.97 / 45.53
  • The supervised systems used for comparison, from Miah et al. (2023), Chapuis et al. (2020) and He et al. (2021), achieve 91.8–92.4 accuracy, but their manually labelled data reduction (MLDR) is 0%. Act2P reports 99.9% MLDR.
  • The online feedback rule is C_t = f_pred(X_t, C_{t−1}): the current classification depends on both the new input and the previous labels.
  • Power quantification is evaluated with Rank Accuracy (RA) and Dominant Speaker Accuracy (DSA). General_label is recommended: it scores slightly below Full_label on RA but better on DSA, and it is more stable.
  • Stated limitations: speech features such as tone and emotion are not used, and there is no prior baseline to compare the power analysis against.

Interesting ideas

  • I1. Retroactive relabeling. Later turns are allowed to revise labels already given to earlier utterances, because the meaning of a turn often becomes clear only after the reply. This is useful for any incremental interpretation system.
  • I2. Coarse-then-fine labeling. The model first picks a broad class and only then chooses among its subclasses. This narrows the choice and reduces confusion between similar fine labels.
  • I3. Power as a graph over typed acts. Each act type carries a pragmatic weight (for example, issuing a command versus accepting one), and the weights flow along speaker interactions to produce a per-speaker score that changes over time. This turns a stream of local labels into an interpretable social state.
  • I4. Choosing granularity by downstream usefulness. The paper selects the label set by how well it serves the final task, not by raw classification accuracy. The middle granularity won because it was more stable.
  • I5. Reporting annotation cost next to accuracy. Showing the percentage of manual labels saved alongside accuracy makes the trade-off against supervised baselines explicit.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (frames): The speech-act layer in prag could keep earlier act labels provisional and re-score them when a following utterance (an answer, a refusal) arrives, logging each revision so the change stays explainable.
  • I2 (learn): LLM teacher tiers could first agree on a coarse category, such as an act family or a verb-class group, and only then on the fine label, so silver data can be accepted at the coarse level even when fine labels disagree.
  • I3 (world): Dialogue scenes in the text world could hold a per-character "dominance" state updated by explicit, weighted speech-act edges, so questions like "who was in charge?" are answered from a traceable chain.
  • I4 (learn): When designing evaluations, Mova could measure which granularity of frame or speech-act labels best supports the answer stage, rather than maximizing tagger accuracy on the finest inventory.
  • I5 (learn): Evaluation reports could pair every score with the share of gold data used, making silver-data and self-play gains comparable to fully supervised setups.

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

🎓Beijing University of TechnologyBeijing · China13 works · 1,910 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 worksSociology and Political Science · 577 works

🧭 Topics 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksDialogue systems · 624 worksSpeech act theory · 229 works

Cite

Cite as Wenbo Zhang and Yuhan Wang. 2025. Act2P: LLM-Driven Online Dialogue Act Classification for Power Analysis. Findings of the Association for Computational Linguistics: ACL 2025.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.findings-acl.1052

@inproceedings{zhang2025act2p,
  title = {{Act2P: LLM-Driven Online Dialogue Act Classification for Power Analysis}},
  author = {Wenbo Zhang and Yuhan Wang},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2025},
  year = {2025},
  doi = {10.18653/v1/2025.findings-acl.1052},
  url = {https://doi.org/10.18653/v1/2025.findings-acl.1052}
}

References

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

  1. Pavel Berkhin. 2005. A survey on pagerank computing. Internet mathematics, 2(1):73–120. ↗
  2. Reihane Boghrati and Morteza Dehghani. 2018. Follow my language! effect of power relations on syntactic alignment. In Proceedings of the Annual Meeting of the Cognitive Science Society, volume 40. find
  3. Emile Chapuis, Pierre Colombo, Matteo Manica, Matthieu Labeau, and Chloé Clavel. 2020. Hierarchical pre-training for sequence labelling in spoken dialog. volume EMNLP 2020. find
  4. Zheqian Chen, Rongqin Yang, Zhou Zhao, Deng Cai, and Xiaofei He. 2018. Dialogue act recognition via crf-attentive structured network. In The 41st international acm sigir conference on research & development in information retrieval, pages 225–234. ↗
  5. Minje Choi, Luca Maria Aiello, Krisztián Zsolt Varga, and Daniele Quercia. 2020. Ten social dimensions of conversations and relationships. In Proceedings of The Web Conference 2020, pages 1514–1525. find
  6. Pierre Colombo, Emile Chapuis, Matteo Manica, Emmanuel Vignon, Giovanna Varni, and Chloe Clavel. 2020. Guiding attention in sequence-to-sequence models for dialogue act prediction. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 7594–7601. ↗
  7. Cristian Danescu-Niculescu-Mizil, Lillian Lee, Bo Pang, and Jon Kleinberg. 2012. Echoes of power: Language effects and power differences in social interaction. In Proceedings of the 21st international conference on World Wide Web, pages 699–708. ↗
  8. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. CoRR, abs/2407.21783. ↗
  9. Changzeng Fu, Yikai Su, Kaifeng Su, Yinghao Liu, Jiaqi Shi, Bowen Wu, Chaoran Liu, Carlos Toshinori Ishi, and Hiroshi Ishiguro. 2025. HAM-GNN: A hierarchical attention-based multi-dimensional edge graph neural network for dialogue act classification. Expert Syst. Appl., 261:125459. ↗
  10. Deepanway Ghosal, Navonil Majumder, Soujanya Poria, Niyati Chhaya, and Alexander F. Gelbukh. 2019. Di-aloguegcn: A graph convolutional neural network for emotion recognition in conversation. pages 154–164. Association for Computational Linguistics. find
  11. Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. ↗
  12. Zihao He, Leili Tavabi, Kristina Lerman, and Mohammad Soleymani. 2021. Speaker turn modeling for dialogue act classification. In Findings of the Association for Computational Linguistics: EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pages 2150–2157. Association for Computational Linguistics.
  13. Geert Hofstede. 1984. Culture’s consequences: International differences in work-related values, volume 5. sage. find
  14. Harshit Kumar, Arvind Agarwal, Riddhiman Dasgupta, and Sachindra Joshi. 2018. Dialogue act sequence labeling using hierarchical encoder with crf. In Proceedings of the aaai conference on artificial intelligence, volume 32. ↗
  15. Michelle Lam, Catherina Xu, and Vinodkumar Prabhakaran. 2018. Power networks: A novel neural architecture to predict power relations. In Proceedings of the Second Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 97–102. ↗
  16. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. 2024. Deepseek-v3 technical report. CoRR, abs/2412.19437. ↗
  17. Md Messal Monem Miah, Adarsh Pyarelal, and Ruihong Huang. 2023. Hierarchical fusion for online multimodal dialog act classification. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 7532–7545.
  18. OpenAI. 2023. GPT-4 technical report. abs/2303.08774. ↗
  19. CoRR, Vipul Raheja and Joel R. Tetreault. 2019. Dialogue act classification with context-aware self-attention. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pages 3727– 3733. Association for Computational Linguistics. ↗
  20. Purva Raut, Rohit Chawhan, Tejas Joshi, and Pratik Kasle. 2020. Classification of power relations based on email exchange. In 2020 IEEE International Conference on Computing, Power and Communication Technologies (GUCON), pages 486–489. IEEE. find
  21. Laria Reynolds and Kyle McDonell. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended abstracts of the 2021 CHI conference on human factors in computing systems, pages 1–7.
  22. John R. Searle. 1969. Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press. ↗
  23. Yang Sun, Nan Yu, and Guohong Fu. 2021. Integrating rich utterance features for emotion recognition in multi-party conversations. In Neural Information Processing: 28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8–12, 2021, Proceedings, Part IV 28, pages 51–62. Springer. find
  24. Dong Wang, Ziran Li, Haitao Zheng, and Ying Shen. 2020. Integrating user history into heterogeneous graph for dialogue act recognition. In Proceedings of the 28th International Conference on Computational Linguistics, pages 4211–4221.
  25. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. 2022. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS ’22, Red Hook, NY, USA. Curran Associates Inc.
  26. Tao Wen, Yu-wang Chen, Tahir Abbas Syed, and Dar-minder Ghataoura. 2025. Examining communication network behaviors, structure and dynamics in an organizational hierarchy: A social network analysis approach. Information Processing & Management, 62(1):103927. ↗
  27. Philine Witzig, Rares Constantin, Nikola Kovacevic, and Rafael Wampfler. 2024. Multimodal dialog act classification for digital character conversations. In ACM Conversational User Interfaces 2024, CUI 2024, Luxembourg, July 8-10, 2024, page 12. ACM. find
  28. Jerrold H Zar. 2005. Spearman rank correlation. Encyclopedia of biostatistics, 7. A ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/zhang-2025-act2p-llm-driven-online-dialogue/