Abstarct: This thesis focuses on the problem of detecting inappropriate discourse in social networks. Considering the importance of online communications and the need to maintain a healthy and ethical environment, the main objective of this research is to develop an accurate and efficient method for identifying and classifying inappropriate discourse in social media platforms. With the increasing use of social networks and the prevalence of inappropriate content, numerous studies have been conducted to identify various types of inappropriate discourse in different languages. However, limited research has been done specifically for the Persian language, which is considered a low-resource language. The aim of this thesis is to address the automatic detection of inappropriate discourse in the Persian language. To achieve this, we collected social media texts, labeled them, and prepared a dataset for training and evaluation. In addition to our own dataset, we also utilized existing datasets in the Persian language. Considering the low-resource nature of the Persian language, we employed transfer learning models, particularly cross-lingual transfer, and utilized multilingual data resources. Leveraging a large dataset of inappropriate discourse in the English language, we propose three transfer learning-baxsed architectures. Two architectures utilize cross-lingual translation methods, while the third architecture utilizes multilingual language models such as M-BERT and XLM. The experimental results demonstrate that our proposed architectures outperform other machine learning and deep learning approaches, such as recurrent neural networks, convolutional neural networks, and single-language models like BERT-Persian, achieving higher accuracy (approximately ۵% to ۱۰% improvement) in detecting inappropriate discourse.