The dataset viewer is not available for this split.
Error code: UnexpectedError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Dataset Card for Chinese Debate Documents
ASR-transcribed corpus of competitive Mandarin Chinese university debates, speaker-segmented and timestamped, with topic / round / team metadata parsed from the source filenames.
Loading the Dataset
from datasets import load_dataset
ds = load_dataset("DBWBD/Chinese_Debate_Documents", split="train")
print(ds[0]["topic"], "—", ds[0]["team_a"], "vs", ds[0]["team_b"])
for seg in ds[0]["segments"][:3]:
print(f" 说话人{seg['speaker_id']} {seg['timestamp']} {seg['text'][:60]}")
Streaming mode works the same way (streaming=True) since the corpus ships as
a single Parquet shard.
Dataset Summary
Chinese_Debate_Documents is a corpus of speaker-segmented transcripts of
competitive university debates conducted in Mandarin Chinese. The corpus draws
primarily from the 第十一届新国辩 (11th New International Debate Open) and
related Sinophone collegiate tournaments, covering topics ranging from
contemporary social commentary ("恋爱脑" 错 / 没错; "原生家庭批判" 是否让年轻人更
幸福) to public-policy debates (死刑废除; 当今中国是否应延迟退休; 第一学历歧视)
and philosophical 1v1s (人终将孤独 / 不孤独; "真爱"是 / 不是谎言).
The raw transcripts were produced by automatic speech recognition (ASR) from
the original tournament audio recordings, then segmented by the ASR system into
turns labelled 说话人N HH:MM (speaker-N timestamp). This release reformats the
two source archives (legacy .doc and modern .docx) into a single
HuggingFace-loadable Parquet shard with typed fields for the transcript, the
per-speaker segments, and any structured metadata recoverable from the original
filenames (topic, tournament round, university name of each team).
Supported Tasks
- Stance / position classification — given a debate topic and a transcript segment, predict which side (正方 / 反方) the speaker argues for.
- Argument mining / component classification — segment the transcript into argumentative units (claim, evidence, rebuttal) at the speaker-turn level.
- Long-context summarization — produce judge-style summaries of an entire
debate session (a single
transcriptfield can run >20 000 Chinese characters). - Conversational / dialogue modeling — multi-speaker structured turns make this useful for training Chinese turn-taking and rebuttal models.
- ASR post-correction / disfluency removal — the transcripts are raw ASR output and contain typical recognition errors; useful as a training set for Chinese ASR cleanup models.
This dataset does not include the source audio. Tasks that require the audio signal (acoustic ASR training, prosody analysis) are out of scope.
Languages
The corpus is in Mandarin Chinese (zh), primarily in standard Putonghua
with regional accents reflecting the speakers' universities (mainland Chinese
institutions plus several international/Asia-Pacific universities including
Malaya, NUS, Cardiff, Monash, LSE, the University of Hong Kong, Macau
University). A small number of English code-switches occur (e.g. "GPT", "AI",
proper nouns) but the transcripts are functionally monolingual.
Dataset Structure
Data Instances
A single record represents one debate session. Example (truncated):
{
"id": "GPT 式大语言模型发展应该_不该以\"对人类友善\"为前提丨初赛F组第一场 马来亚大学vs上海大学_音频.wav06-转写结果",
"source_collection": "docs",
"source_file": "GPT 式大语言模型发展应该_不该以\"对人类友善\"为前提丨初赛F组第一场 马来亚大学vs上海大学_音频.wav06-转写结果.doc",
"source_format": "doc",
"topic": "GPT 式大语言模型发展应该_不该以\"对人类友善\"为前提",
"round": "初赛F组第一场",
"team_a": "马来亚大学",
"team_b": "上海大学",
"num_segments": 412,
"num_chars": 47312,
"transcript": "0.wav\n说话人1 00:00\n大家好...\n说话人2 00:48\n请越南马来亚...\n",
"segments": [
{"speaker_id": 1, "timestamp": "00:00", "text": "大家好,欢迎收看..."},
{"speaker_id": 2, "timestamp": "00:48", "text": "请越南马来亚..."}
]
}
Data Fields
| Field | Type | Description |
|---|---|---|
id |
string |
Stable identifier (<source_collection>/<stem> for ordered batches, plain stem for docs/). |
source_collection |
string |
Either docs (the named-title batch) or one of the 机器-订单转写结果-<timestamp> order batches. |
source_file |
string |
Original Word filename, preserving Chinese characters and the full-width 丨 separator. |
source_format |
string |
doc (legacy OLE Word97-2003) or docx (Office Open XML). |
topic |
string (nullable) |
Debate motion, parsed from the filename. null when the source file uses an anonymous numeric name (e.g. 0.wav29-转写结果). |
round |
string (nullable) |
Tournament round (e.g. 初赛F组第一场, 小组赛, 决赛). |
team_a |
string (nullable) |
First team / university (e.g. 马来亚大学). |
team_b |
string (nullable) |
Opposing team / university (e.g. 上海大学). |
num_segments |
int32 |
Count of speaker-segmented turns in segments. |
num_chars |
int32 |
Length of transcript in (UTF-8 code-point) characters. |
transcript |
string |
Full debate transcript as flat text, with original ASR speaker headers (说话人N HH:MM) on their own lines. |
segments |
list<struct> |
Structured turn list. Each entry has speaker_id (int32), timestamp (MM:SS or HH:MM string), and text (the turn body, stripped of header). |
Data Splits
This release ships a single train split. No held-out evaluation split is
provided; downstream users should construct their own train/dev/test based on
the task at hand (we recommend splitting by topic to test out-of-topic
generalization, or by team_a/team_b to test out-of-school generalization).
| Split | # Examples | # Bytes (parquet, snappy) | # Bytes (uncompressed) |
|---|---|---|---|
| train | 2 028 | 131 544 115 (≈ 125 MB) | 238 676 609 (≈ 228 MB) |
The corpus totals approximately 62 million Chinese characters of transcript text across 2 028 debate sessions, with a mean of ~30 700 characters and ~183 speaker turns per session.
Dataset Creation
Curation Rationale
Chinese-language NLP resources for long-form argumentative discourse are scarce. Most Chinese conversational corpora are short customer-service or chat-style turns; most Chinese summarization corpora are news articles. A debate corpus offers something distinct: multi-speaker, multi-turn, topically diverse, argumentatively structured text where each utterance is positionally committed to one of two declared sides. This makes the corpus particularly useful for studying stance detection, rebuttal generation, and judge-style long-context summarization in Chinese.
Source Data
Initial Data Collection and Normalization
The raw audio was sourced from publicly available recordings of the 第十一届新国辩 (11th New International Debate Open) and related Sinophone university debate competitions. The audio was transcribed by a commercial automatic speech recognition service (the original filenames carry the "机器-订单转写" prefix — "machine order transcription result") and exported as Microsoft Word documents.
The dataset arrives in two physical generations:
docs/— 107 files in.doc(Word 97-2003 OLE) format. Mostly tournament-named titles preserving topic/round/team metadata, mixed with a smaller set of anonymous numeric IDs (0.wav29-转写结果.doc).机器-订单转写结果-<timestamp>/— 43 batch directories, each containing ~45.docxfiles in modern Office Open XML format. Filenames follow the<id>-<round> <team_a>VS<team_b> <topic>-<resolution>P …pattern.
This release does not alter the raw .doc / .docx files. The Parquet shard
under data/ is a derived artifact, produced by scripts/extract_transcripts.py
using Microsoft Word automation (COM) to extract the text losslessly, plus a
small parser that splits the text on the ASR delimiters (\r between segments,
\v between segment header and body) and recovers structured metadata from
the filenames.
Who are the source language producers?
University student debaters from mainland Chinese institutions and several international universities (Malaya, NUS, Cardiff, Monash, LSE, Hong Kong, Macau, etc.) competing in Mandarin Chinese. A small number of recordings feature professional debaters from the Chinese debate-show circuit (黄执中, 马薇薇, 詹青云, 庞颖, 梁秋阳, 肖磊) — these appear in the named "表演赛" and "哲理辩论" sessions.
Annotations
This dataset is not human-annotated. The "annotations" present are:
- Speaker turn boundaries and speaker IDs — generated by the ASR diarizer.
IDs are consistent within a single debate session but not stable across
sessions (
说话人1in one transcript is a different person than说话人1in another). - Timestamps — generated by the ASR system, accurate to the second.
- Topic / round / team metadata — parsed deterministically from the source
filenames by a regex in
scripts/extract_transcripts.py. When the parser cannot match a known pattern, the metadata fields are set tonull.
Annotation process
Automatic. The transcript bodies and segment headers come from a commercial Chinese ASR system; the metadata fields come from filename parsing. No human verification has been performed on the transcripts.
Who are the annotators?
Software (commercial ASR + this repository's filename parser). No human annotators.
Personal and Sensitive Information
Debaters are public participants in publicly-broadcast tournaments, and their
school affiliation is part of the public record (it is the school's team
they represent). The dataset does not intentionally identify individual
debaters by name in metadata fields. However, transcripts may contain
self-introductions where debaters state their own name on-mic, and the
ASR-recovered text reproduces this verbatim. Users training models on the
transcript or segments fields should be aware of this. If you intend to
publish derived datasets, consider running a Chinese-NER pass over the
transcripts and redacting PER entities outside of the public-figure
allowlist (the small set of named television debaters listed above).
Considerations for Using the Data
Social Impact of Dataset
Competitive debate is an adversarial discourse format: each speaker is assigned a side and is required to argue it persuasively regardless of personal belief. Models trained on this corpus may learn rhetorical patterns that are effective at making any position sound reasonable. This is useful for research on argumentation, stance modeling, and rebuttal generation, but the same capability has obvious dual-use implications for persuasion and disinformation. We recommend that downstream applications include explicit disclosure when their outputs are debate-style argumentation rather than sincere belief.
Discussion of Biases
- Topic selection bias. The motions debated reflect the curatorial choices of the source tournaments — heavily weighted toward Chinese social, cultural, and policy issues of the 2020s (年轻人压力, 原生家庭, 恋爱脑, GPT 与 AI 伦理, 直播带货, 退休年龄). Models trained on this corpus will reflect that topical distribution.
- School-prestige bias. Top-tier mainland universities (清华, 北大, 中山, 人大, 国防科技) and elite international programs (LSE, NUS, Malaya) are overrepresented, reflecting which schools qualify for these tournaments.
- ASR error bias. The transcripts contain systematic ASR errors: homophone substitutions (e.g. 因 / 应), missing punctuation, occasional speaker-attribution swaps. These errors are not uniformly distributed — they are more frequent for non-standard accents and code-switched English terms.
- Language register. Competitive debate Mandarin is a deliberately formal, rhetorically dense register. Models trained on this corpus will not reflect colloquial spoken Chinese.
Other Known Limitations
- No audio. Only transcripts are released. Tasks requiring acoustic features are out of scope.
- No ground-truth ASR. The transcripts have not been corrected against the source audio; they are raw ASR output.
- Speaker IDs are not stable across files. Do not assume
说话人1is the same person across two different debate sessions. - Metadata coverage is partial. Filename patterns vary across the two
source generations, so the deterministic filename parser only recovers
structured fields for a minority of records:
team_a/team_b/roundare populated for ~30% of records (611 / 2 028)topicis populated for ~19% of records (384 / 2 028)- the remaining ~70% have transcripts and
source_fileonly — users wanting metadata on those should write their own filename heuristics (the original filename is preserved verbatim insource_file) The anonymous numeric-id files (e.g.0.wav29-转写结果.doc) are by construction metadata-free; the larger gap comes from order-batch filenames where the topic andteam_aare glued together without a whitespace separator, which a regex parser cannot reliably disambiguate without Chinese-NER knowledge of school names.
- Filename text in metadata fields may include full-width punctuation
(
,,、,",丨) and resolution suffixes leaked from the original upload pipeline. Normalize before exact-match search.
Additional Information
Dataset Curators
Curated and packaged by the dataset owner (DBWBD). The original audio belongs to the respective tournament organizers; transcripts were generated by a third-party commercial ASR provider.
Licensing Information
The packaging, parsing scripts, and dataset card are released under the MIT
License (see LICENSE). The raw transcript text reflects the underlying
audio recordings, which are publicly available; users redistributing derived
works should respect the source tournaments' broadcast terms.
Citation Information
@misc{chinese_debate_documents_2025,
title = {Chinese Debate Documents: ASR-transcribed competitive debate corpus},
author = {{DBWBD}},
year = {2025},
url = {https://huggingface.co/datasets/DBWBD/Chinese_Debate_Documents},
note = {Mandarin Chinese collegiate debate transcripts, speaker-segmented}
}
A CITATION.cff is also provided at the repository root for citation tooling.
Contributions
Issues and improvements are welcome via the Community tab on the dataset repository. Especially welcome:
- Manual ASR corrections for individual debates
- A hand-labeled stance / argument-component subset
- A speaker-name redaction pass with a Chinese NER model
- Downloads last month
- 5