Iranian Journal of Electrical and Electronic Engineering

This systematic review documents the state-of-the-art, methods, corpora, and challenges of detecting hate speech from Persian text. Persian is a low-resource language with unique linguistic, cultural, and resource-based challenges distinguishing it from other languages ​​such as English, Arabic, and European languages ​​in developing effective automatic hate speech detection systems. Persian's complex morphology, script, and cultural nuances, in contrast to most high-resource languages, require tailored solutions. The survey brings together the contributions from classical machine learning, deep learning, transformer models, and hybrid models, encouraging their adaptation to Persian characteristics and noisy social media datasets. It brings to the fore the absence of labeled data, class imbalance, domain shift, annotator subjectivity, and resource availability variability compared to other languages. The survey also touches upon the background of datasets, annotation strategies, and evaluation measures and contains information about the performance of various solutions. It finishes with presenting strategic recommendations to further enhance Persian hate speech detection, including developing benchmarked datasets, expanding resource-constrained corpora, and designing culturally appropriate models for low-resource settings. This review aims to guide future studies towards more improved and dependable NLP applications for Persian and other low-resource languages ​​based on the importance of language-specific settings and cultural factors.

Keywords: Persian language, Hate speech detection, Low-resource languages, Social media analysis, Offensive speech detection.

Full-Text [PDF 1684 kb]   (1 Downloads)