Multimodal Argumentative Fallacy Detection and Classification on Political Debates Shared Task.

Co-located with The 12th Workshop on Argument Mining in Vienna, Austria.

Overview

This shared task focuses on detecting and classifying fallacies in political debates by integrating text and audio data. Participants will tackle two sub-tasks:

• Argumentative Fallacy Detection
• Argumentative Fallacy Classification

We offer three input settings:

• Text-only: Analyze textual arguments.
• Audio-only: Explore paralinguistic features.
• Text + Audio: Combine both for a multimodal perspective.

Join us to advance multimodal argument mining and uncover new insights into human reasoning! 💬

Tasks

Task A

• Input: a sentence, in the form of text or audio or both, extracted from a political debate.
• Task: to determine whether the input contains an argumentative fallacy.

Task B

• Input: a sentence, in the form of text or audio or both, extracted from a political debate, containing a fallacy.
• Task: to determine the type of fallacy contained in the input, according to the classification introduced by Goffredo et al. (2022). We only refer to macro categories.

For each sub-task, participants can leverage the debate context of a given input: all its previous sentences and corresponding aligned audio samples. For instance, consider the text-only input mode. Given a sentence from a political debate at index i, participants can use sentences with indexes from 0 to i - 1, where 0 denotes the first sentence in the debate.

Data

We use MM-USED-fallacy and release a version of the dataset specifically designed for argumentative fallacy detection. This dataset includes 1,278 sentences from Haddadan et al.’s (2019) dataset on US presidential elections. Each sentence is labeled with one of six argumentative fallacy categories, as introduced by Goffredo et al. (2022).

Inspired by observations from Goffredo et al. (2022) on the benefits of leveraging multiple argument mining tasks for fallacy detection and classification, we also provide additional datasets to encourage multi-task learning. A summary is provided in the table below:

All datasets will be available through MAMKit.

Since many multimodal datasets cannot release audio samples due to copyright restrictions, MAMKit provides an interface to dynamically build datasets and promote reproducible research.

Datasets are formatted as torch.Dataset objects, containing input values (text, audio, or both) and corresponding task-specific labels. More details about data formats and dataset building are available in MAMKit’s documentation. ## Retrieving the Data through MAMKit

To retrieve the datasets through MAMKit, you can use the following code interface:

from mamkit.data.datasets import MMUSEDFallacy, MMUSED, UKDebates, MArg, InputMode

import logging
from pathlib import Path

def loading_data_example():
    base_data_path = Path(__file__).parent.parent.resolve().joinpath('data')

    # MM-USED-fallacy dataset
    mm_used_fallacy_loader = MMUSEDFallacy(
        task_name='afc', # Choose between 'afc' or 'afd'               
        input_mode=InputMode.TEXT_AUDIO, # Choose between TEXT_ONLY, AUDIO_ONLY, or TEXT_AUDIO
        base_data_path=base_data_path
    )

    # MM-USED dataset
    mm_used_loader = MMUSED(
        task_name='asd',#Choose between 'asd' or 'acc'  
        input_mode=InputMode.TEXT_AUDIO, # Choose between TEXT_ONLY, AUDIO_ONLY, or TEXT_AUDIO
        base_data_path=base_data_path
    )

    # UKDebates dataset
    uk_debates_loader = UKDebates(
        task_name='asd', 
        input_mode=InputMode.TEXT_AUDIO, # Choose between TEXT_ONLY, AUDIO_ONLY, or TEXT_AUDIO
        base_data_path=base_data_path
    )

    # M-Arg dataset
    m_arg_loader = MArg(
        task_name='arc',
        input_mode=InputMode.TEXT_AUDIO, # Choose between TEXT_ONLY, AUDIO_ONLY, or TEXT_AUDIO
        base_data_path=base_data_path
    )

Each loader is initialized with the appropriate task name (afc for argumentative fallacy classification, asd for argumentative sentence detection, and ‘arc’ for argumentative relation classification), input mode (InputMode.TEXT_ONLY, InputMode.AUDIO_ONLY, or InputMode.TEXT_AUDIO), and the base data path.

Ensure that you have MAMKit installed and properly configured in your environment to use these loaders.

For more details, refer to the MAMKit GitHub repository and website .

Test Set Access 🔍

The test set for mm-argfallacy-2025 is now available! To use it, please:

1. Create a fresh environment
2. Clone the repository and install the requirements:

git clone git@github.com:nlp-unibo/mamkit.git
cd mamkit
pip install -r requirements.txt
pip install --editable .

3. Access MAMKit in your Python code:

import mamkit

Then, retrieve the data using the following code:

For Fallacy Classification (afc):

from mamkit.data.datasets import MMUSEDFallacy, InputMode
from pathlib import Path

def loading_data_example():
    base_data_path = Path(__file__).parent.parent.resolve().joinpath('data')
    loader = MMUSEDFallacy(
        task_name='afc',
        input_mode=InputMode.TEXT_ONLY,  # or TEXT_AUDIO or AUDIO_ONLY
        base_data_path=base_data_path
    )
    split_info = loader.get_splits('mm-argfallacy-2025')

For Fallacy Detection (afd):

from mamkit.data.datasets import MMUSEDFallacy, InputMode
from pathlib import Path

def loading_data_example():
    base_data_path = Path(__file__).parent.parent.resolve().joinpath('data')
    loader = MMUSEDFallacy(
        task_name='afd',
        input_mode=InputMode.TEXT_ONLY,  # or TEXT_AUDIO or AUDIO_ONLY
        base_data_path=base_data_path
    )
    split_info = loader.get_splits('mm-argfallacy-2025')

References

• MM-USED-fallacy: Mancini et al. (2024). The version provided through MAMKit includes updated samples, with refinements in the alignment process. This results in a different number of samples compared to the original dataset.
• MM-USED: Mancini et al. (2022). The version provided through MAMKit includes updated samples, with refinements in the alignment process. This results in a different number of samples compared to the original dataset.
• UK-Debates: Lippi and Torroni (2016).
• M-Arg: Mestre et al. (2021).

Note: By “updated version,” we mean that the datasets have undergone a refinement in the alignment process, which has resulted in adjustments to the number of samples included compared to the original versions published in the referenced papers.

Evaluation

For argumentative fallacy detection, we will compute the binary F1-score on predicted sentence-level labels.
For argumentative fallacy classification, we will compute the macro F1-score on predicted sentence-level labels.
Metrics will be computed on the hidden test set to determine the best system for each sub-task and input mode.

Evaluation will be performed via the CodaLab platform.
On CodaLab, participants will find the leaderboard, along with the results of the provided baselines.
Submission guidelines can be found under the Evaluation section of the CodaLab competition page.

🚨 Important: In the evaluation website, you will also find a link to a mandatory participation survey.
Filling out this survey is required in order to participate in the task.
We also provide the survey link here for convenience: https://tinyurl.com/limesurvey-argfallacy

Baseline Results on Test Set

Argumentative Fallacy Classification (AFC) – Macro F1-score

Argumentative Fallacy Detection (AFD) – Binary F1-score

Submission

All evaluated submissions are required to commit to submitting a system description paper. You can choose between two options:

• Non-Archival Paper:
  A 2-page paper describing your system, with unlimited pages for appendices and bibliography. These papers will not be published in the workshop proceedings, but your system will be mentioned in the Overview Paper of the shared task, upon acceptance.

• Archival Paper:
  A 4-page paper describing your system, also with unlimited pages for appendices and bibliography. These papers will be published in the official ACL workshop proceedings and must be presented at the workshop (poster or oral session).
  ⚠️ In accordance with ACL policy, at least one team member must register for the workshop in order to present an archival paper if aaccepted to be published at the ACL proceedings.

All papers must use the official ACL style templates, available in both LaTeX and Word. We strongly recommend using the official Overleaf template for convenience.

We have sent an email to each team with all the details regarding the system description paper submission for MM-ArgFallacy2025. Please check your inbox (and spam folder just in case).

• 🗓️ Submissions open: May 1st, 2025 (the day after the end of the evaluation period)
• 🗓️ Submissions close: May 15th, 2025
• 📢 Notification of acceptance: May 20th, 2025
• 📝 Camera-ready deadline: May 25th, 2025

Important notes:

• All accepted archival papers will be presented during the workshop’s poster session and require at least one registered author.
• Non-archival papers do not require registration and are not presented at the workshop, but their systems will be acknowledged in the Overview Paper.

We look forward to receiving your submissions!

🏆 Leaderboard – Shared Task Results

AFC Task – Argumentative Fallacy Classification

📝 Text-only

🔊 Audio-only

🔁 Text-Audio

AFD Task – Argumentative Fallacy Detection

📝 Text-only

🔊 Audio-only

🔁 Text-Audio

Key Dates (Anywhere on Earth)

• Release of Training Data: February 25th
• Release of Test Set: March 24th → April 7th
• Evaluation Start: April 14th → April 21st
• Evaluation End: April 25th → April 30th
• Paper Submissions Open: May 1st
• Paper Submission Close: May 15th
• Notification of acceptance: May 20th
• Camera-ready Due: May 25th
• Workshop: July 31st

Task Organizers

Eleonora Mancini

Language Technologies Lab, University of Bologna, Italy

Federico Ruggeri

Language Technologies Lab, University of Bologna, Italy

Paolo Torroni

Language Technologies Lab, University of Bologna, Italy

Serena Villata

Inria-I3S WIMMICS Laboratoire I3S, CNRS, Sophia Antipolis, France

Contacts

Join the MM-ArgFallacy2025 Slack Channel!

Credits

This shared task is partially supported by the project European Commission's NextGeneration EU programme, PNRR -- M4C2 -- Investimento 1.3, Partenariato Esteso, PE00000013 - FAIR - Future Artificial Intelligence Research'' -- Spoke 8 Pervasive AI’’.