Zongrui Yang, Haoyuan Li, Zhongsheng Wang, Zhirui Zeng, Pengqian Han, Yi Zhou, Yuting Wang, Jiamou Liu · Under review as a conference paper at ICLR 2027 · preprint arXiv:2609.21637
Evaluating debate requires interpreting contributions in context and relating them to local assessments and overall outcomes. We introduce the Chinese Competitive Debate Dataset (CCDD), linking competition transcripts to original adjudication under a predefined rubric. CCDD contains 148 matches, 2,690 stage–side scoring instances, and 20,542 turns, with reviewed transcripts, aligned individual judge scores, match decisions, best-debater ballots, and commentary. Linked appeal records preserve reassessments. Three benchmark tasks — winner prediction, stage-score prediction, and best-debater prediction — demonstrate the use of this supervision. Zero-shot results reveal challenges in reproducing stage assessments, while judge-agreement analyses and structural references contextualize performance. CCDD supports research on contextual debate evaluation and judge disagreement.
Authentic Chinese competitive debate linked to individual stage scores, match decisions, best-debater votes and commentary collected under a predefined rubric.
Transcripts and assessments organised with explicit stage and side associations; individual judgments and linked appeal adjudications preserved; provenance of original and generated materials documented.
Three prediction tasks with initial zero-shot results, judge-agreement analyses and structural references for reproducible evaluation and interpretation.
| Resource and scope | Textual organisation and focus | Annotations or assessment targets | Annotation provenance |
|---|---|---|---|
| ORCHID (2023) · 1,218 debates | Debate transcripts and stance-specific summaries | Target-independent stance; argumentative dialogue summarisation | Curated competition transcripts with task-specific annotations and summaries |
| CEDAR (2026) · 600 debates | Complete opposing-side transcripts with content, role and timing information | Claims, stances, evidence, rhetorical figures, modal words, debate results | Collected competition records annotated for computational argumentation |
| Conch (2026) · 3 debates across languages | Sessions, turns, argument blocks and linked clash structures | Clash points, disagreements, viewpoints and strategies | LLM-assisted structural analysis with expert annotation and validation; designed for visual analysis |
| Chen et al. (2026) · 94 debates | Contributions in open discussion and initial presentations | Support, response and language ratings; debate outcomes | LLM-generated quality ratings; competition outcomes; separate human pairwise validation |
| DEFINED (2026) · statement-level | Competition statements and augmented materials | Eight-dimensional creativity-related assessment and overall scoring | Expert-scored authentic statements with mixed-granularity supervision and augmentation |
| CCDD · 148 retained matches | Matches, scheduled stages and turns, linked to evaluated sides and roles | Stage scores, match decisions, best-debater ballots and judge commentary | Individual competition-time judgments under a predefined rubric; linked appeal records; separately identified generated supplements |
Counts denote debates; Conch includes one Chinese and two English debates. DEFINED is described by its statement-level assessment scope.
CCDD covers the collected Chinese-language competitions and their formats; repeated participants, teams, judges and motions introduce potential dependencies, and filtering for usable scoring records may affect corpus composition. Manual review supports record integrity but does not estimate residual transcription or segmentation error. The experiments measure text-based zero-shot agreement with recorded adjudication; judges observed audiovisual performance and assessed the completed match, whereas stage-scoring models receive text and preceding context, which limits direct comparison.
Generative AI tools were used to repair ASR transcripts under specified constraints, to process data and to summarise judges' rationales as supplementary annotations; processed transcripts and generated supplements underwent the review procedures described in the paper. AI tools also assisted with language polishing. The authors remain responsible for all claims, numerical results, citations and materials.