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

Introduction

Nowadays, sexually explicit comments are widespread on digital platforms, including social media, forums, and messaging applications. The ease of access to online spaces has spread this content quickly, turning it into a global issue that crosses borders and cultures. Such behavior has sparked important conversations about regulation, responsibility, and the need for better digital literacy (Radanliev et al., 2024).

Sexually explicit comments are often characterized by improper, unrefined, or sexually suggestive language (Betts et al., 2018). Reducing the spread of sexually explicit content on social media is necessary to guarantee the safety and dignity of all users. However, the lack of effective models for detecting sexually explicit content in various languages impacts individuals, communities, and digital platforms (Mukred et al., 2024). If not properly filtered, this content has the potential to disrupt the traditional values and norms of a community, which are vital for peaceful coexistence.

While detection using Natural Language Processing (NLP) and machine learning has significantly advanced for well-resourced languages like English, Chinese, and Arabic, challenges persist for under-resourced languages (Supriyono & Suyono, 2024). Advanced deep learning models such as BERT, Generative Pre-trained Transformer (GPT), mbrets, and Cross-lingual Language Model (XLM-R) are readily available for this task in well-resourced languages (Zhang & Shafiq, 2024). However, building these models from scratch is challenging for under-resourced languages, which lack sufficient data and preprocessing tools. Therefore, for tasks such as explicit content detection, question answering, sentiment analysis, and hate speech detection in under-resourced languages like Amharic, transfer learning and fine-tuning emerge as the most viable strategies (Zhao et al., 2024). These methods leverage knowledge from models pre-trained on well-resourced languages, allowing for efficient adaptation to new tasks with significantly less data.

Amharic, a Semitic language mainly spoken in Ethiopia, serves as the country’s official working language (Asker et al., 2009). The Amharic language is spoken by more than 32 million people as their mother tongue and by over 25 million as a second language in the country (Belay et al., 2020). Most Amharic speakers reside in Ethiopia, and significant Amharic-speaking communities also exist in Eritrea, the USA, Israel, Sweden, Somalia, and Djibouti due to diaspora populations (Meshesha & Jawahar, 2007). Amharic is written using a script adapted from the ancient Ethiopian language known as Ge’ez, which is still used in the Ethiopian Orthodox Church (Girma, 2019). This script uses 34 consonant characters with the seven vowel variants of each (called basic characters). For example, the character for ‘h’ (ሀ) is systematically changed to represent hu (ሁ), hi (ሂ), ha (ሃ), he (ሄ), hə (ህ), and ho (ሆ).

Beyond its unique script, Amharic also exhibits complex morphological processes involving derivation and inflection, which are particularly evident in word classes like verbs, nouns, and adjectives (Yohannes et al., 2025; Abate & Assabie, 2014). Amharic verbs are normally formed in two stages from verbal roots: first, formation of the verbal stem, and second, addition of inflectional markers (Gewe & Gasser, 2012; Yeshambel, Mothe & Assabie, 2022). To illustrate, the Amharic verb “መጠነ” (met’ene), meaning “he measures/adjusts,” beautifully demonstrates its origin in the verbal stem “መጠን” (met’en), which denotes “amount/size.” This stem originates from the verbal root “ምጥን” (mtn), which means “measured/adjusted.” Through a combination of derivation and inflection, a verbal stem can be marked for various grammatical categories, including person, case, gender, number, tense, and negation. This process results in a highly complex morphological structure (Booij, 2006).

In addition to its morphological complexity, Amharic exhibits significant dialectal variation throughout Ethiopia, reflecting the country’s diverse ethnic and cultural landscape (Mengistu & Melesew, 2017). The most common dialects include Addis Ababa, Gojjam, Gonder, Wollo, and Shewa (Gebre, Firisa & Dash, 2024). The Addis Ababa dialect is the standard for education and media because it combines features of many Ethiopian dialects (Salawu & Aseres, 2015). Each of the other dialects, such as Gojjam, Gonder, Wollo, and Shewa, shows unique phonetic, lexical, meaning, and grammatical variations. For instance, the Amharic word “ክፈተኝ” (kifetenyi) has different meanings across dialects. In the Gondar dialect, it means “open to me,” whereas in the Gojjam dialect, it means “open me.” This illustrates how dialectal variations in Amharic create significant semantic and grammatical differences, accordingly increasing the complexity of text processing in the language.

Amharic’s complex morphology and dialects, along with the lack of essential NLP resources like processing tools, language models, and training datasets, pose challenges for NLP development (Endalie, 2025). As a result, no NLP models have been developed or integrated with social media platforms to filter out Amharic sexually explicit content from posts and comments. This increases the risk of sexual harassment for Amharic speakers on social media (Ababu, Woldeyohannis & Getaneh, 2025). However, cultural and religious norms in Ethiopia often restrict open discussions about sexually explicit contents (Baraki & Thupayagale-Tshweneagae, 2023). The presence of such content online in Amharic creates conflict between traditional Ethiopian values and global digital culture. Therefore, developing a sexually explicit content detection system is essential to address this problem.

However, creating such a detection model is also challenging due to the cultural variation found in Ethiopia (Endalie, Haile & Taye, 2023). Cultural variation means that a comment considered sexually explicit by one group may not hold the same connotation for another. For example, the phrase “እናትህን ልብዳ” (inatihini libida), which translates to “fuck your mother,” is a culturally sensitive expression with varying interpretations across regions. In Eastern Ethiopia, such as Dire Dawa, the phrase may be considered a normal part of casual discourse. However, in Northern Ethiopia, including Gojjam and Gonder, the same phrase can incite significant conflict or physical harm. Such cultural variations complicate the development of universal detection systems, necessitating a specialized and robust approach.

Despite these challenges, this research aims to develop a fine-tuned BERT + XAI model for detecting sexually explicit content in Amharic, building on a BERT model initially trained for hate speech detection. The absence of a comprehensive Amharic dataset for BERT training presented clear limitations, and hate speech and sexually explicit content often exhibit shared linguistic patterns, prompting us to adopt this approach. We assembled a diverse dataset of both sexually explicit and non-sexual comments from various social media platforms, creating a valuable resource for future research. To enhance transparency and build confidence, we integrate XAI into our detection model, which helps to illustrate the decision-making process’s reliability and trustworthiness. Finally, this research will play an important role in future studies and provide a deeper understanding of Amharic language processing.

The key contributions of this study are as follows: (1) We created a freely available dataset of Amharic sexually explicit and non-explicit comments gathered from social media. (2) We built a deep learning model for Amharic sexually explicit content detection by adapting a BERT model originally trained on hate speech and using XAI to understand its classification decisions.

In the remainder of this article, we will describe the following: First, the second section will review current models used to detect sexually explicit comments. Next, the third section will explain the methodology we employed in our study. Following that, the fourth section will present our experimental results and discuss their implications. Finally, the fifth section will conclude with some suggestions for future research.

Related works

The detection of sexually explicit content on social media platforms has become a major area of interest, with many studies focusing on natural language processing and machine learning techniques (Jahan & Oussalah, 2023). However, research on underrepresented languages, such as Amharic, remains limited. The following articles outline the methodologies used to detect inappropriate content across various languages. Each study explores different approaches, including machine learning algorithms and NLP techniques.

As discussed in Ababu, Woldeyohannis & Getaneh (2025), the authors proposed four deep-learning classifiers combined with three feature extraction techniques to detect hate speech in Amharic and Afaan Oromo. Their Bidirectional Long Short-Term Memory (Bi-LSTM) model, using FastText embeddings, achieved 78.05% accuracy in bilingual hate speech detection and handled out-of-vocabulary words effectively. Additionally, the study incorporated supplementary linguistic features to improve detection performance and support future research on under-resourced Ethiopian languages.

The study conducted by Hamzah & Dhannoon (2021) utilized deep learning techniques, specifically Bi-LSTM, to detect sexual harassment and cyber predators in English. In their approach, word representations were carefully analyzed and mapped to real-number vectors. Their proposed model for detecting online sexual predators achieved an F_0.5 score of 0.927 and an accuracy of 97.27%. Additionally, the model demonstrated strong performance in identifying sexual harassment-related comments, attaining an F_0.5 score of 0.925 and a remarkable accuracy of 99.12%.

The research in Akhter et al. (2021) focused on automatically detecting abusive language in Urdu and Roman Urdu comments, tackling the challenges of their complex morphology and the mixed-script format of Roman Urdu. The study employed a range of machine learning and deep learning models. The traditional algorithms include Naive Bayes, Support Vector Machine (SVM), and logistic regression. The deep neural networks are like Convolutional Neural Networks (CNN), long short-term memories (LSTMs), and Bi-LSTMs. The datasets employed varied in size from two thousand to ten thousand comments. Notably, the CNN model achieved the highest accuracy, outperforming other methodologies and underscoring the effectiveness of deep learning techniques in this domain. Additionally, their findings indicated that single-layer deep learning architectures provided superior results compared to their multi-layered counterparts.

The authors of Saha et al. (2024) proposed a BERT-CNN model to detect hate speech and harassment in the Bengali language. This model integrates BERT with CNN to enhance performance. It was trained and evaluated using data collected from Bengali social media profiles. The BERT-CNN model outperformed baseline models, achieving notable results in detecting hate speech and fostering a more respectful online environment for Bengali language speakers.

The study by Islam et al. (2023) proposed machine learning and deep learning algorithms to detect and classify sexual harassment in Bangla text. The study achieved accuracy rates ranging from 65.2% to 76.5% using TF-IDF with various algorithms, including Naive Bayes, Decision Tree, Random Forest, AdaBoost, Stochastic Gradient Descent (SGD), Logistic Regression, K-Nearest Neighbors (KNN), and SVM. Deep learning algorithms like CNN, LSTM, and hybrid CNN-LSTM achieved an accuracy of 89%, making them comparatively better than machine learning techniques.

The work of Yenala et al. (2018) proposed a deep learning-based technique to detect and filter inappropriate language in online discussion forums and user conversations in English. The authors used a convolutional bidirectional LSTM architecture to find query suggestions and filter conversations that were not suitable. They did this by using LSTM and Bi-LSTM sequential models. The model outperforms hand-crafted feature-based baselines, enhancing user experiences and preventing abusive language in online discussions.

The study conducted by El-Alami, El Alaoui & En Nahnahi (2022) utilized a Multilingual Offensive Language Detection (MOLD) model that leverages transfer learning and fine-tuning techniques. This model employs BERT to effectively capture semantic and contextual information in offensive texts. It comprises three main components: preprocessing, text representation, and classification, enabling the detection of both offensive and non-offensive content. Experiments conducted on a bilingual dataset demonstrated that the translation-based method achieved an accuracy of over 93%, while the Arabic BERT model reached an accuracy of 91%.

Although various researchers have attempted to detect inappropriate content on social media platforms using machine learning and deep learning across different languages, it has been established that deep learning models are generally more effective in identifying such content. Integrating deep learning with XAI enhances model interpretability by revealing the features influencing each prediction. Moreover, research on inappropriate content detection in the Amharic language is limited, and sexually explicit content represents a significant concern in Ethiopia. Therefore, we propose a model for detecting sexually explicit content in Amharic using a transfer learning strategy, integrated with XAI to ensure that the model’s decisions are easily understandable and interpretable.

Materials and Methods

This section describes the materials and methods employed to create a model for detecting sexually explicit content in Amharic. The detection model utilized Explainable AI to illustrate its decision-making process for users. Figure 1 presents a flowchart that visually illustrates the systematic process of developing and evaluating the proposed model. The process begins with the collection of raw data, followed by several preprocessing steps. These steps encompass data cleaning, normalization, feature extraction, model training, and evaluation. This methodology progresses from initial data input to the final assessment of detection accuracy. Lastly, we incorporate the explainable AI component to improve our understanding of the model’s decision-making process.

The following section provides a concise overview of the modules included in the proposed Amharic sexually explicit content detection model. Each module plays an important role in how the model works, helping it to accurately identify sexually inappropriate content in Amharic text.

Results and discussion

Results

Utilizing this data, we trained the fine-tuned model using the hyperparameter values detailed in Table 2. To determine the optimal number of training epochs, we conducted a comparative analysis of the model’s performance across different epoch settings. This analysis revealed that the model achieved optimal performance with four epochs. We present the training results over these four epochs in Table 3. The table presents the training and validation losses, as well as the precision, recall, F1-score, and accuracy. Throughout the training process, the model’s performance improved significantly.

Furthermore, we evaluated the performance of the fine-tuned BERT model using the 20% testing set. The testing set is completely unseen during training. This evaluation assessed the model’s ability to generalize to new data and accurately classify Amharic explicit and non-explicit text. Table 4 shows the model’s performance in classifying Amharic sexually explicit content, including precision, recall, F1-score, and accuracy for each category. These metrics offer a detailed view of the model’s ability to correctly identify sexually explicit content while minimizing false positives and false negatives.

Our fine-tuned BERT model demonstrated balanced performance in classifying Amharic sexually explicit comments, achieving an overall accuracy of 94%. The model’s precision for category 0 is 0.95, indicating a very low rate of false positives, which means it is highly reliable when predicting non-explicit comments. For category 1, the precision is 0.94, showing strong reliability in identifying explicit content. The recall for category 0 is 0.94, reflecting its strong ability to correctly identify non-explicit content and minimize false negatives. In contrast, the recall for category 1 is 0.95, demonstrating an equally strong capacity to identify explicit content accurately.

The F1-scores of 0.94% for both classes highlight the model’s balanced performance, indicating that it does not favor one type of error over the other. The closely aligned macro-average and weighted average scores, both at 0.94%, confirm that this robust performance is consistent across the entire dataset. These results establish that our model is a highly effective and reliable tool for this binary classification task.

This high performance is a direct result of a robust training process designed to prevent overfitting. We used the training and validation loss curves to demonstrate how the model learns from the training dataset and generalizes to unseen validation data. A training accuracy of 0.98 and a testing accuracy of 0.94 initially suggested overfitting. Figure 2 illustrates the validation loss curve, which was generated to diagnose this issue and show the effect of the weight decay regularization we employed. The loss curve illustrates a healthy training pattern: as the training loss decreases, the validation loss also declines and, importantly, converges closely with the training loss after 3.0 epochs. The final minimal gap between the two loss curves, nearly 0.05, is a strong indication that weight decay was successful in preventing severe overfitting.

The small but persistent accuracy gap is a reflection of the slight divergence in the loss curve between 1.5 and 3.0 epochs. This divergence is a generalization gap that regularization aims to minimize. Although weight decay did not eliminate this gap, which would be impossible without compromising the model’s capacity, it successfully reduced it to a minimal level, as confirmed by the loss curves.

We further analyzed the model’s performance using a confusion matrix. This matrix offers a detailed view of the model’s classification accuracy, including the number of instances in each category and the model’s prediction accuracy for each class (Vujovic, 2021). Table 5 presents the confusion matrix for our explicit comment classification model, comparing actual classifications with the model’s predictions and categorizing them into four quadrants (True Positive, True Negative, False Positive, and False Negative).

The above table presents the confusion matrix summarizing the model’s performance in distinguishing between explicit and non-explicit Amharic comments. Out of a total of 3,518 non-sexually explicit comments, 3,308 were correctly identified. However, 210 non-sexually explicit comments were incorrectly classified as explicit, and 179 explicit comments were misclassified as non-explicit. Additionally, 3,320 explicit comments were accurately identified.

To strengthen the validity of our findings, we analyzed misclassifications from our model. We examined using the comment, “አብረን ብናድርና ተሞካክረን አሸናፊው ቢለይ ምን ትያለሽ?” (“What do you think if we spend the night together and compete to see who wins?”), was an explicit comment that our model misclassified as non-explicit. The model likely failed to detect the sexual nature of this comment because it depends on keyword detection. In Amharic, this kind of proposition uses culturally nuanced metaphors and euphemisms such as አሸናፊው (“compete”) and ቢለይ ምን ትያለሽ (“decide who wins”) rather than direct sexual terminology.

To determine whether such errors were isolated or part of a broader pattern, we analyzed the probability distribution of borderline classifications. As shown in Fig. 3, the red dashed line at 0.500 is our decision threshold, separating explicit from non-explicit comments. The histogram highlights comments that the model struggles to classify with high confidence. For instance, one comment was classified as non-explicit with a probability score of about 0.485, and two were classified as explicit with a score of about 0.508. The proximity of these scores to 0.5 underscores the model’s hesitation. It found these comments ambiguous and lacking the strong indicators needed for a confident prediction.

For example, the comment “ለንስሀ እራቁትን መቅርብ ጥሩ ነው” (“It is important to bring yourself naked for repentance”) is a borderline case. Although its meaning is non-explicit, a metaphor for confessing one’s sins without hiding anything, the model likely assigned it a probability score near the threshold (e.g., 0.485), classifying it as non-explicit but with low confidence. The model identified the literal keyword “naked” as a potential indicator of explicit content, which conflicted with the non-explicit context and produced an uncertain prediction. In contrast, the absolute classifier labeled it as explicit (category 1).

Explainable artificial intelligence

We utilized XAI in addition to the model developed for identifying sexually explicit content. This approach helps clarify how the model differentiates between sexually explicit and non-sexual text, making the process more transparent for users. By clarifying how decisions are reached, users can better understand the factors that influence these classifications, which in turn fosters greater trust in the model’s outcomes.

As an illustration of the explainability framework, two examples were selected: one comment classified as sexually explicit by the model and one classified as non-sexual. The comment “እምስዎ በጣም ጣፋጭ ይመስላል” (Your pussy looks very delicious) received a classification of sexually explicit. Conversely, the comment “ታደለ ጥበቡን ያለማድነቅ እንዴት አይቻልም” (How can one not admire Tadele’s wisdom?) was classified as non-sexual. The factors underpinning these classifications are detailed in Figs. 4 and 5, which provide insight into the model’s feature weighting and contextual understanding.

Figure 4 illustrates the trained model’s classification of the example “እምስዎ በጣም ጣፋጭ ይመስላል” (Your pussy looks very delicious). The model analyzes the sentence for features indicative of sexually explicit content. The presence of “እምስዎ” (your pussy) suggests the message is directed at a specific individual and refers to a private body part. “ይመስላል” (looks) implies an assessment of the body part’s appearance. The intensifier “በጣም” (very) emphasizes the description. Finally, “ጣፋጭ” (delicious) carries a connotation of desirability, contributing to the sexually suggestive nature of the sentence. Based on these linguistic features, the model predicts a label of 1, which aligns with the actual label.

Figure 5 illustrates the trained model’s interpretation of the sentence “ታደለ ጥበቡን ያለማድነቅ እንዴት አይቻልም” (“How can one not admire Tadele’s wisdom?”). The model correctly identifies “ታደለ” (Tadele) as a proper noun. ‘ጥበቡን’ (wisdom) functions as a noun referring to Tadele’s intellect. ‘ያለማድነቅ’ (admire) serves as a verb expressing positive sentiment. ‘እንዴት አይቻልም’ (how can one not) operates as a rhetorical question to enhance the preceding sentiment. As a result of this linguistic analysis, the model predicts a non-sexual classification, assigning a label of 0, which aligns with the actual label.

Conclusion

In this study, we introduced a BERT + XAI model designed to detect sexually explicit comments in the Amharic language. The proposed model comprises several modules, including document collection, preprocessing, feature extraction, data splitting, model training, and the application of XAI. We employed a fine-tuned BERT model specifically designed to identify hate speech in the Amharic language. Our findings demonstrate that this BERT-based model is highly effective at distinguishing between sexually explicit and non-sexually explicit content, showcasing its ability to handle the unique linguistic features of Amharic. Additionally, we incorporated explainability techniques that enhance our model’s decision-making process, making it easier for users to understand how comments are categorized by clearly highlighting the clues used. This model achieved significant improvements in performance metrics compared to traditional models like base BERT, SVM, CNN, and MLP. This study contributes to the field of natural language processing, with a particular focus on advancing research for the Amharic language. It also highlights the pressing need for advanced AI methods to effectively manage and mitigate sensitive issues such as online sexually explicit or inappropriate content. Looking ahead, we plan to expand our dataset to include code-switching and further refine our model to enhance its detection capabilities, ultimately aiming to create a safer online environment for Amharic speakers. Overall, our findings highlight the promise of XAI in tackling classification challenges across various linguistic contexts.

Supplemental Information