ALERT: A benchmark Bengali dataset for identifying and categorizing religiously aggressive texts.

Area of Science:

• Computational Linguistics
• Social Media Analysis
• Natural Language Processing

Background:

• Religiously aggressive content on social media threatens social harmony.
• Existing tools for detecting such content are limited for regional languages like Bengali.
• There's a need for resources to develop effective detection and prevention strategies.

Purpose of the Study:

• To introduce ALERT, a novel Bengali dataset for religiously aggressive text classification.
• To address the scarcity of resources for analyzing extremist content in Bengali.
• To facilitate the development of tools for identifying and mitigating harmful online content.

Main Methods:

• Creation of the ALERT dataset with 4027 annotated instances across four categories: hate speech, vandalism, atrocity, and no aggression.
• Data sourced from diverse online platforms (Facebook, YouTube, blogs, etc.).
• Annotation by diverse annotators, with conflict resolution by a domain expert; preprocessing included data cleaning and integrity checks.

Main Results:

• The ALERT dataset demonstrates strong inter-annotator agreement (Cohen's kappa score of 72%).
• Experiments with machine learning, deep learning, and transformer models show promising results for text classification.
• The dataset provides valuable insights into the characteristics and classification of religiously aggressive text.

Conclusions:

• ALERT serves as a benchmark dataset for religiously aggressive text classification in Bengali.
• The dataset's public accessibility promotes research, innovation, and collaboration in Bengali NLP.
• This resource is expected to advance the development of tools for combating online religious extremism.