Published April 30, 2026 | Version v1
Dataset Open

Touché26-Fallacy-Detection

Authors/Creators

  • 1. ROR icon Bauhaus-Universität Weimar

Description

Dataset for the Fallacy Detection task at Touché 2026, based on the informal fallacies dataset by Sahai et al. (2021).

Data Fields

Each JSON entry contains the following fields:

Source Text

Field Description
id Unique identifier for the entry.
text_raw The original Reddit comment.
text_raw_parent The parent comment of text_raw.
text_raw_title Title of the thread text_raw originates from.

Processed Arguments

Field Description
text_base Self-contained rewrite of text_raw that incorporates relevant context from text_raw_parent and text_raw_title.
argument_base {claim, supports} — the argument structure extracted from text_base.
text_enhanced Rewrite of text_base that explicitly surfaces fallacy and scheme information.
argument_enhanced {claim, supports} — the argument structure extracted from text_enhanced.

Labels

Field Description
fallacy_exists Binary label (0 / 1) indicating whether the entry contains a fallacy.
fallacy_type The specific type of fallacy (e.g., authority, slippery_slope, hasty_generalization).
resembles_fallacy The fallacy type this argument most closely resembles. Equals fallacy_type for fallacious entries. For non-fallacious entries, identifies which fallacy the reasoning pattern could plausibly be confused with.
classification {argument_goal, argument_basis} — argumentation scheme dimensions based on Macagno (2022).

Base vs. Enhanced Fields

Reddit arguments are often unclear and rely on implicit assumptions. The dataset therefore provides two levels of processing:

  • Base fields (text_base, argument_base) produce a self-contained version of the argument by integrating context from the parent comment and thread title.
  • Enhanced fields (text_enhanced, argument_enhanced) are additionally rewritten with knowledge of the corresponding fallacy and scheme labels, making the reasoning patterns more explicit.

Resembles Fallacy

Each non-fallacious entry was deliberately selected because its reasoning pattern resembles a specific fallacy type. For example, an entry with resembles_fallacy: authority contains a legitimate appeal to authority — structurally similar to a fallacious one, but without committing the fallacy.

Files

Files (3.0 MB)

Name Size Download all
md5:1b5d4c8a811f9ea8e20eb89cebc4ef07
561.9 kB Download
md5:5ebcc132a2a739e4fe34e5902016126b
2.4 MB Download