Hellisotherpeople/OpenDebateEvidence-Anonymized

Dataset Card for OpenDebateEvidence (Anonymized)

A collection of evidence used in collegiate and high school debate competitions, with all debater-identifying columns removed.

This is an anonymized redistribution of Yusuf5/OpenCaselist. The argumentative content is byte-for-byte unchanged. 26 of the original 45 columns have been dropped. See Anonymization for exactly what was removed and why.

Dataset Details

Dataset Description

This dataset is a follow up to DebateSum, increasing its scope and amount of metadata collected.

It expands the dataset to include evidence used during debate tournaments, rather than just evidence produced during preseason debate "camps." The total amount of evidence is approximately 20x larger than DebateSum. It includes evidence from the Policy, Lincoln Douglas, and Public Forum styles of debate.

Evidence was deduplicated to detect the use of the same piece of evidence in different debates.

• Curated by: Yusuf Shabazz (original), anonymized redistribution by Allen G. Roush
• Language(s) (NLP): English
• License: MIT

Dataset Sources

openCaselist

• Repository: Data Processing Code and GraphQL Search API
• Paper: OpenDebateEvidence

Anonymization

The upstream dataset truncated debater first and last names to their first two characters, but several other columns re-exposed the full identities. Most significantly, filePath and opensourcePath contained the uploaded document filenames, which embed full surnames and school names:

./documents/ndtceda15/Army/JoSo/Army-Johnson-Sowatzke-Aff-Navy-Round1.docx

Team codes such as JoSo are themselves built from debater surnames, and roundId / fileId join back to the public caselist wiki where the names remain. Because many of these debaters were minors at the time, the following 26 columns were dropped entirely:

Category	Columns dropped
Direct names	debater1First, debater1Last, debater2First, debater2Last
Team codes derived from surnames	teamId, teamName, teamDisplayName
School and geography	schoolId, schoolName, schoolDisplayName, state, chapterId
Paths embedding names and schools	filePath, opensourcePath
Named third parties	opponent, judge
Free text that may name people	notes, report
Join keys back to the source wiki	fileId, roundId
Round context that joins to the public Tabroom record	tournament, round
Caselist identifiers that narrow a row to a league and season	caselistId, caselistName, caselistUpdatedAt
Team composition	teamSize

The last three categories were removed in a second pass. Tabroom.com publishes tournament entries and results by name, so a tournament plus a round number is a lookup key against a public record even with every other identifier stripped. caselistDisplayName is retained because it names a league and season (NDT/CEDA 2015-16) and identifies nobody.

Residual risk

If you find a record that somehow slipped through and identifies someone and you would like it removed, open a discussion on this repo and we will remove it ASAP.

Uses

Direct Use

The most direct use of the dataset is for research in argument mining. The dataset contains rich metadata about the argumentative content of the text it contains. It can be used to train models for identifying arguments, classifying arguments, detecting relations between arguments, generating arguments, and other related tasks. It also contains metadata useful for abstractive and extractive summarization, natural language inference, and is useful as a pretraining dataset for many tasks.

Out-of-Scope Use

The abstractive summaries generated by debaters make heavy use of jargon and are often in context of a broader argument, so they may not be useful directly.

Attempting to re-identify debaters from the retained round metadata is explicitly out of scope.

Dataset Structure

Each year is a separate config:

from datasets import load_dataset

ds = load_dataset("Hellisotherpeople/OpenDebateEvidence-Anonymized", "2019")

Configs are 2014 through 2022, plus openev for the preseason Open Evidence project.

Summaries

Debate evidence is typically annotated with three different summaries.

• tag is typically a one sentence abstract summary written by the debater. It often makes heavy use of jargon.
• summary is the underlined text in the original document. Typically a phrase level summary of the evidence.
• spoken is the highlighted text in the original document. This is the text the debater plans to read out loud during the debate.

Note that the tag will be biased to support some broader argument, and the summary and spoken summaries will be biased to support the tag. Some arguments called "analytics" do not contain any body text, and only have the debater written tag.

Headings

Debate documents are usually organized with a hierarchy of headings.

• pocket Top level heading. Usually the speech name.
• hat Medium level heading. Usually the broad type of argument.
• block Low level heading. Usually the specific type of argument.

Headings for speeches later in the debate will often contain the abbreviations "A2" or "AT" (Answer to) when responding to an argument in a previous speech.

Deduplication

Debaters will often reuse the same evidence across different debates and repurpose evidence used by other debaters for their own arguments. A deduplication procedure detects this repeated use.

• bucketId Duplicate evidence will share the same bucketId.
• duplicateCount is the number of other pieces of evidence in the same bucket. This is useful as a rough measure of argument quality, since good arguments tend to be used more.

Dataset Fields

Column Name	Description
id	Unique identifier for the evidence
tag	Abstract summary of the evidence/argument made by the debater with evidence
cite	String indicating the short citation of the source used for the evidence
fullcite	Full citation of the source used for the evidence
summary	Longer extractive summary of the evidence
spoken	Shorter extractive summary of the evidence
fulltext	The full text of the evidence
textLength	The length of the text in the evidence in characters
markup	The full text of the evidence with HTML markup for parsing/visualization purposes
pocket	String indicating the virtual "pocket"
hat	String indicating the virtual "hat"
block	String indicating the virtual "block"
bucketId	Unique identifier for the deduplication bucket of the evidence
duplicateCount	The number of duplicates of the evidence
side	The debate side on which the evidence was used (Affirmative or Negative)
caselistDisplayName	The league and season the evidence was uploaded under, e.g. NDT/CEDA 2015-16
year	The year in which the debate round took place
event	The event in which the debate round took place
level	The level of the debate (college, high school)

Dataset Creation

Curation Rationale

Existing argument mining datasets are limited in scope or contain limited metadata. This dataset collects debate evidence produced by high school and collegiate competitive debaters to create a large dataset annotated for argumentative content.

Source Data

Debaters will typically source evidence from news articles, journal articles, books, and other sources. Many debaters upload annotated excerpts from these sources to the openCaselist website where this data is collected from.

Data Collection and Processing

This dataset was collected from Word Document files uploaded to openCaselist. The .docx is unzipped to access the internal XML file, then the individual pieces of evidence in the file are extracted and organized using the formatting and structure of the document. The deduplication procedure works by identifying other evidence that contains identical sentences.

The anonymization pass streamed the upstream CSVs, dropped the 20 columns listed above, and re-encoded the result as zstd-compressed Parquet. No row filtering and no text modification was performed.

Who are the source data producers?

The source data is produced by high school and collegiate competitive debaters.

Personal and Sensitive Information

See Anonymization. Debater names, schools, states, team codes, opponents, judges, and file paths have been removed. Free-text evidence fields have not been scrubbed and may contain card-cutter initials.

Bias, Risks, and Limitations

The dataset inherits the biases in the types of arguments used in a time limited debate format, and the biases of academic literature as a whole. Arguments tend to be mildly hyperbolic and one sided, and the arguments are on average "left" leaning and may overrepresent certain positions. As each debate was about a certain broad topic, those topics will be represented heavily in the dataset. A list of previous topics is available from the National Speech and Debate Association. There may be occasional errors in parsing that lead to multiple pieces of evidence being combined into one record.

Citation

BibTeX:

@inproceedings{
    roush2024opendebateevidence,
    title={OpenDebateEvidence: A Massive-Scale Argument Mining and Summarization Dataset},
    author={Allen G Roush and Yusuf Shabazz and Arvind Balaji and Peter Zhang and Stefano Mezza and Markus Zhang and Sanjay Basu and Sriram Vishwanath and Ravid Shwartz-Ziv},
    booktitle={The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
    year={2024},
    url={https://openreview.net/forum?id=43s8hgGTOX}
}