Overview
The Bangla Multi-task Hatespeech 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.
Task Details
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.
Official Evaluation Metrics
We choose the following evaluation metrics considering imbalance across classes:
- Subtask 1A: Micro-F1 score
- Subtask 1B: Micro-F1 score
- Subtask 1C: Averaged Micro-F1 score
| Rank | username | F1-Macro |
|---|---|---|
| 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 |
| Rank | username | F1-Macro |
|---|---|---|
| 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 |
| Rank | username | F1-Macro |
|---|---|---|
| 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 |
Participation
Please follow the steps to participate:
- To participate in this competition, you must have an account in Codabench
- Register to the competition to participate.
Data Repository: https://github.com/AridHasan/blp25_task1
For a brief overview of the dataset, kindly refer to the README.md file located in the data directory.
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.
490273 আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই Political Hate
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.
490273 আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই Organization
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.
490273 আওয়ামী লীগের সন্ত্রাসী কবে দরবেন এই সাহস আপনাদের নাই "Political Hate" "Little to None" Organization
Timeline
- 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)
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>
python scorer/task.py --pred_files_path task_dev_output.txt --gold_file_path data/dev.tsv
The baselines module currently contains a majority, random and a simple n-gram baseline.
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 |
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 |
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 |
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
python format_checker/task.py -p ./subtask_1A.tsv
results_files: can be one path or space-separated list of paths
Evaluation consists of two phases:
- Development phase: This phase involves working on the dev-test set.
- 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 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
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
Organizers
Resources
For updates and resources, visit the GitHub repository.
To communicate join our Slack Channel
Data Uses
Participants must agree to use the dataset for research purposes only and cite the shared task paper and dataset source in any publication or derivative work.
Citation
@inprocedding{hasan2025multihate,
title="BanglaMultiHate",
}