Pashto offensive language detection: a benchmark dataset and monolingual Pashto BERT

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Artificial Intelligence

Background:

• Offensive language on social media poses a threat to online communities.
• Detecting offensive content in low-resource languages like Pashto is an underexplored research area.
• Existing research primarily focuses on high-resource languages, leaving a gap for Pashto.

Purpose of the Study:

• To develop an AI model for automatic offensive text detection in the Pashto language.
• To create a benchmark dataset for Pashto offensive language identification.
• To evaluate and compare various deep learning and transfer learning models for this task.

Main Methods:

• Developed the Pashto Offensive Language Dataset (POLD) from Twitter data.
• Implemented and evaluated deep learning models (CNNs, RNNs) with static word embeddings (Word2Vec, fastText, GloVe).
• Investigated transfer learning using XLM-R and a custom-trained Pashto BERT model.

Main Results:

• The Pashto BERT model achieved superior performance compared to other evaluated models.
• The Pashto BERT model attained an F1-score of 94.34% and an accuracy of 94.77%.
• Deep learning and transfer learning approaches were benchmarked on the POLD dataset.

Conclusions:

• The developed Pashto BERT model is effective for offensive text detection in Pashto.
• This research contributes a valuable dataset and a high-performing model for a low-resource language.
• Automatic detection of offensive content in Pashto is feasible and crucial for online safety.