Skip to content

The Multi-task Bangla Hate Speech Identification shared task is designed to address the complex and nuanced problem of detecting and understanding hate speech in Bangla across multiple related subtasks such as type of hate, severity, and target group.

Notifications You must be signed in to change notification settings

AridHasan/blp25_task1

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Task Website

https://multihate.github.io/

Objective

The Bangla Multi-task Hate Speech Identification shared task is designed to address the complex and nuanced problem of detecting and understanding hate speech in Bangla across multiple related subtasks such as type of hate, severity, and target group. In contrast to single-task approaches, this shared task embraces a multi-task learning framework, where models are trained to jointly learn several hate speech detection objectives. This approach is more reflective of real-world scenarios, where identifying hate speech requires understanding not just its presence, but also its type, target, and severity. Please see the Task Description below.

Table of contents:

Important Dates

  • 10 July 2025: Registration on codalab and beginning of the development cycle
  • 08 September 2025: Beginning of the evaluation cycle (test sets release and run submission)
  • 15 September 2025: End of the evaluation cycle
  • 17 September 2025: Publish rank list and share paper submission details
  • 29 September 2025: Paper Submission Deadline (Shared Task System Papers Due)
  • 03 November 2025: Notification of acceptance
  • 11 November 2025: Camera-ready due
  • 23-14 December 2025: Workshop co-located with IJCNLP-AACL 2025 (Mumbai, India)

Proceedings

Instructions to Prepare Your Shared Task Paper

The title of paper should be in the following format: < Team Name > at BLP-2025 Task 1: < Descriptive title of your paper >

For example, team AlphaX would have their title as follows: AlphaX at BLP-2025 Task 1: Transformer Models for Hate Speech Detection

  • The shared task papers may consist of up to four (4) pages of content.

Templates: The Shared tasks papers must follow the ACL 2025 two-column format, using the supplied official templates. The templates can be downloaded in style files and formatting. Please do not modify these style files, nor should you use templates designed for other conferences. Submissions that do not conform to the required styles, including paper size, margin width, and font size restrictions, will be rejected without review. Verification to guarantee conformance to publication standards, we will be using the ACL pubcheck tool. The PDFs of camera-ready papers must be run through this tool prior to their final submission, and we recommend its use also at submission time.

Submissions are open to only for the teams who submitted their systems during the evaluation phase and listed in the leaderboard. The working notes are to be submitted in both anonymously and non-anonymously.

Recent Updates

Recent Updates

  • [11/07/2025] Release example scripts using DistilBERT model for subtask 1A and subtask 1B
  • [10/07/2025] Development phase starts
  • [10/07/2025] Training and dev data released

Contents of the Directory

  • Main folder: data
    This directory contains data files for the task.

  • Main folder: baselines
    Contains scripts provided for baseline models of the task.

  • Main folder: example_scripts
    Contains an example script provided to run DistilBERT model for subtask 1A and subtask 1B.

  • Main folder: format_checker
    Contains scripts provided to check the format of the submission file.

  • Main folder: scorer
    Contains scripts provided to score the output of the model when provided with the label (i.e., dev).

  • README.md
    This file!

Task Description

This shared task is designed to identify the type of hate, its severity, and the targeted group from social media content. The goal is to develop robust systems that advance research in this area. In this shared task, we will have three subtasks:

  • Subtask 1A: Given a Bangla text collected from YouTube comments, categorize whether it contains Abusive, Sexism, Religious Hate, Political Hate, Profane, or None.
  • Subtask 1B: Given a Bangla text collected from YouTube comments, categorize whether the hate towards Individuals, Organizations, Communities, or Society.
  • Subtask 1C: This subtask is a multi-task setup. Given a Bangla text collected from YouTube comments, categorize it into type of hate, severity, and targeted group.

Dataset

For a brief overview of the dataset, kindly refer to the README.md file located in the data directory.

Input data format

Subtask 1A

Each file uses the tsv format. A row within the tsv adheres to the following structure:

id	text	label

Where:

  • id: an index or id of the text
  • text: text
  • label: Abusive, Sexism, Religious Hate, Political Hate, Profane, or None.
Example
490273	আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই	Political Hate

Subtask 1B

Each file uses the tsv format. A row within the tsv adheres to the following structure:

id	text	label

Where:

  • id: an index or id of the text
  • text: text
  • label: Individuals, Organizations, Communities, or Society.
Example
490273	আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই	Organization

Subtask 1C

Each file uses the tsv format. A row within the tsv adheres to the following structure:

id	text	hate_type   hate_severity   to_whom

Where:

  • id: an index or id of the text
  • text: text
  • hate_type: Abusive, Sexism, Religious Hate, Political Hate, Profane, or None.
  • hate_severity: Little to None, Mild, or Severe.
  • to_whom: Individuals, Organizations, Communities, or Society.
Example
490273	আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই	"Political Hate"  "Little to None"  Organization

Example Fine-tuning Script

We are pleased to release a set of example scripts to support participants in the Hate Speech Detection Shared Task. These scripts are designed to help you get started with data loading, preprocessing, and baseline model development for the three subtasks: subtask 1A, subtask 1B, and subtask 1C. We encourage you to use and adapt these examples to build and improve your own systems. The scripts are available in the shared task repository: example_scripts

Scorer and Official Evaluation Metrics

Scorers

The scorer for the task is located in the scorer module of the project. The scorer will report official evaluation metrics and other metrics of a prediction file. The scorer invokes the format checker for the task to verify the output is properly shaped. It also handles checking if the provided predictions file contains all tweets from the gold one.

You can install all prerequisites through,

pip install -r requirements.txt

Launch the scorer for the task as follows:

python scorer/task.py --gold-file-path=<path_gold_file> --pred-file-path=<predictions_file>
Example
python scorer/task.py --pred_files_path task_dev_output.txt --gold_file_path data/dev.tsv

Official Evaluation Metrics

The official evaluation metric for the subtask 1A and 1B is micro-F1 and weighted micro-F1 for subtask 1C. However, the scorer also reports accuracy, precision and recall.

Baselines

The baselines module currently contains a majority, random and a simple n-gram baseline.

Subtask 1A

Baseline Results for the task on Test set (Evaluation Phase)

Model micro-F1
Random Baseline 0.1638
Majority Baseline 0.5638
n-gram Baseline 0.6020

Baseline Results for the task on Dev-Test set

Model micro-F1
Random Baseline 0.1465
Majority Baseline 0.5760
n-gram Baseline 0.6075

Subtask 1B

Baseline Results for the task on Test set (Evaluation Phase)

Model micro-F1
Random Baseline 0.2043
Majority Baseline 0.5974
n-gram Baseline 0.6209

Baseline Results for the task on Dev-Test set

Model micro-F1
Random Baseline 0.2118
Majority Baseline 0.6083
n-gram Baseline 0.6279

Subtask 1C

Baseline Results for the task on Test set (Evaluation Phase)

Model weighted micro-F1
Random Baseline 0.2304
Majority Baseline 0.6072
n-gram Baseline 0.6305

Baseline Results for the task on Dev-Test set

Model weighted micro-F1
Random Baseline 0.2300
Majority Baseline 0.6222
n-gram Baseline 0.6401

Format checker

The format checkers for the task are located in the format_checker module of the project. The format checker verifies that your generated results file complies with the expected format.

Before running the format checker please install all prerequisites,

pip install -r requirements.txt

To launch it, please run the following command:

python format_checker/task.py -p results_files
Example
python format_checker/task.py -p ./subtask_1A.tsv

results_files: can be one path or space-separated list of paths

Submission

Guidelines

Evaluation consists of two phases:

  1. Development phase: This phase involves working on the dev-test set.
  2. Evaluation phase: This phase involves working on the test set, which will be released during the evaluation cycle.

For each phase, please adhere to the following guidelines:

  • We request each team to establish and manage a single account for all submissions. Hence, all runs should be submitted through the same account. Any submissions made from multiple accounts by the same team may lead to your system being not ranked from the final ranking in the overview paper.
  • The most recently uploaded file on the leaderboard will serve as your final submission.
  • Adhere strictly to the naming convention for the output file, which must be labeled as 'task.tsv'. Deviation from this standard could trigger an error on the leaderboard.
  • Submission protocol requires you to compress the '.tsv' file into a '.zip' file (for instance, zip task.zip task.tsv) and submit it through the Codalab page.
  • With each submission, ensure to include your team name along with a brief explanation of your methodology.
  • Each team is allowed a maximum of 100 submissions per day for the given task. Please adhere to this limit.

Submission Format

Subtask 1A and 1B

Submission file format is tsv (tab seperated values). A row within the tsv adheres to the following structure:

id	label   model

Where:

  • id: a id of the text
  • label: [Abusive, Sexism, Religious Hate, Political Hate, Profane, or None] or [Individuals, Organizations, Communities, or Society.]
  • model: model name

Subtask 1C

Submission file format is tsv (tab seperated values). A row within the tsv adheres to the following structure:

id	hate_type   hate_severity   to_whom   model

Where:

  • id: a id of the text
  • hate_type: Abusive, Sexism, Religious Hate, Political Hate, Profane, or None.
  • hate_severity: Little to None, Mild, or Severe.
  • to_whom: Individuals, Organizations, Communities, or Society.
  • model: model name

Submission Site

Subtask 1A

https://www.codabench.org/competitions/9559/

Subtask 1B

https://www.codabench.org/competitions/9560/

Subtask 1C

https://www.codabench.org/competitions/9561/

Citation

There are various papers associated with the task. Details for the papers specific to the task as well as an overall overview will be posted here as they come out. Bib entries for each paper are included here.

@article{hasan2025llm,
      title={LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target}, 
      author={Hasan, Md Arid and Alam, Firoj and Hossain, Md Fahad and Naseem, Usman and Ahmed, Syed Ishtiaque},
      year={2025},
      journal={arXiv preprint arXiv:2510.01995},
      url={https://arxiv.org/abs/2510.01995},
}

@inproceedings{blp2025-overview-task1,
    title = "Overview of BLP 2025 Task 1: Bangla Hate Speech Identification",
    author = "Hasan, Md Arid and Alam, Firoj and Hossain, Md Fahad and Naseem, Usman and Ahmed, Syed Ishtiaque",
    booktitle = "Proceedings of the Second International Workshop on Bangla Language Processing (BLP-2025)",
    editor = {Alam, Firoj
          and Kar, Sudipta
          and Chowdhury, Shammur Absar
          and Hassan, Naeemul
          and Prince, Enamul Hoque
          and Tasnim, Mohiuddin
          and Rony, Md Rashad Al Hasan,
          and Rahman, Md Tahmid Rahman
    },
    month = dec,
    year = "2025",
    address = "India",
    publisher = "Association for Computational Linguistics",
}

Leaderboard

Subtask 1A

Rank username F1-Micro
1 shifat_islam 0.7362
2 SyntaxMind 0.7345
3 zannatul_007 0.734
4 mahim_ju 0.7331
5 reyazul 0.7328
6 mohaiminulhoque 0.7323
7 nahidhasan 0.7305
8 adib709 0.7282
9 sahasourav17 0.7275
10 ashraf_989 0.7273
11 CUET-NLP_Zenith 0.7263
12 nsu_milab 0.725
13 abid_al_hossain 0.7238
14 Penta Global Ltd 0.7178
15 mohaymen 0.7133
16 ttprama 0.7111
17 minjacodes9 0.7075
18 samin007 0.707
19 pritampal98 0.7057
20 bahash_ai 0.7028
21 programophile 0.7013
22 fatin_anif 0.6954
23 heytamjid 0.6941
24 adriti12 0.6921
25 im_tushu_221 0.6901
26 sadman03samir 0.6871
27 cuet_sntx_srfrs 0.6867
28 abir_bot69 0.684
29 antara_n_15 0.6815
30 UB 0.6761
31 quasar 0.6733
32 shahriar_9472 0.6689
33 intfloat 0.6634
34 naim-parvez 0.6587
35 Organizers 0.5638
36 teddymas 0.4589
37 mizba 0.1077

Subtask 1B

Rank username F1-Micro
1 mahim_ju 0.7356
2 shifat_islam 0.7335
3 mohaiminulhoque 0.7328
4 reyazul 0.7317
5 SyntaxMind 0.7317
6 zannatul_007 0.7315
7 abid_al_hossain 0.7286
8 nahidhasan 0.7279
9 adib709 0.7275
10 sahasourav17 0.7269
11 Penta Global Ltd 0.7256
12 mohaymen 0.7254
13 CUET-NLP_Zenith 0.7213
14 adriti12 0.7125
15 ashraf_989 0.7114
16 ttprama 0.7095
17 nsu_milab 0.6981
18 heytamjid 0.6979
19 pritampal98 0.6974
20 bahash_ai 0.6954
21 cuet_sntx_srfrs 0.6817
22 sadman03samir 0.676
23 Organizers 0.5974
24 lamiaa 0.2848

Subtask 1C

Rank username F1-Micro
1 mahim_ju 0.7392
2 CUET-NLP_Zenith 0.7378
3 shifat_islam 0.7361
4 reyazul 0.7332
5 adib709 0.7312
6 mohaiminulhoque 0.731
7 sahasourav17 0.7262
8 abid_al_hossain 0.725
9 nur_163 0.7241
10 nahidhasan 0.724
11 ttprama 0.7233
12 zannatul_007 0.7181
13 Penta Global Ltd 0.7159
14 pritampal98 0.7153
15 abir_bot69 0.7129
16 sadman03samir 0.7129
17 bahash_ai 0.6969
18 cuet_sntx_srfrs 0.6842
19 aacontest 0.673
20 Organizers 0.6072
21 adriti12 0.3898

Communication

Please join us in Slack channel for discussion and doubts:

Organizers

  • Md Arid Hasan, PhD Student, The University of Toronto
  • Firoj Alam, Senior Scientist, Qatar Computing Research Institute
  • Md Fahad Hossain, Lecturer, Daffodil International University
  • Usman Naseem, Assistant Professor, Macquarie University
  • Syed Ishtiaque Ahmed, Associate Professor, The University of Toronto

About

The Multi-task Bangla Hate Speech Identification shared task is designed to address the complex and nuanced problem of detecting and understanding hate speech in Bangla across multiple related subtasks such as type of hate, severity, and target group.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published