An Amharic sexually explicit content detection model using fine-tuned bidirectional encoder representation from transformers and explainable artificial intelligence

Nowadays, people are posting a significant amount of sexually explicit content on digital platforms and social media, particularly in under-resourced languages like Amharic. This rise in explicit content presents a significant challenge to the cultural and religious values of Ethiopia, which have traditionally emphasized humbleness and respect. To mitigate this growing issue, we present an Amharic sexually explicit content detection model. Our approach utilizes a fine-tuned Bidirectional Encoder Representation from Transformers (BERT) model that was initially trained to detect hate speech in Amharic. This is due to the lack of a large-scale Amharic dataset for training BERT from scratch and the shared linguistic characteristics between hate speech and sexually explicit content. We integrate this model with Explainable Artificial Intelligence (XAI) techniques to enhance the transparency and interpretability of its predictions. The model was trained on a diverse, manually annotated dataset of 34,710 comments and posts (comprising 17,199 explicit and 17,511 non-explicit examples) collected from social media platforms such as TikTok, YouTube, Facebook, and X (formerly known as Twitter). Data instances were annotated based on the presence of explicit terms, sexually offensive words, sexual slang, and content highly susceptible to sexual interpretation. To understand the model’s decision-making process, we incorporated Local Interpretable Model-agnostic Explanations (LIME) as our XAI framework. The experimental results demonstrate strong performance, yielding accuracy, precision, recall, and F1-score values of 94%, 95%, 94%, and 94%, respectively. The XAI techniques employed reveal how the model classifies comments by computing probabilities for each category and highlighting significant words associated with them. Furthermore, we compare the performance of our proposed model with several state-of-the-art text classification techniques, including base BERT, support vector machines (SVM), multilayer perceptrons (MLP), and convolutional neural networks (CNN). The results indicate that our model consistently outperforms these existing methods.