AI Models Detect Suicidal Ideation in Amharic Social Media Posts

Researchers have built and compared machine learning, deep learning, and ensemble learning models for detecting suicidal ideation in Amharic social media posts, publishing the first publicly available dataset for this low-resource language. Suicide contributes to more than 800,000 deaths worldwide each year, and people who post about distress, self-harm, and death online offer a potentially crucial avenue for timely psychological intervention and surveillance. Most computational studies of suicidal ideation detection, however, have focused on English and other widely spoken languages, leaving linguistically diverse populations, including Amharic speakers, largely understudied.

The team, comprising Andualem Woldegiorgis, Mohammed Abebe, Durga Prasad Sharma, and Worku Jimma, collected 5,428 postings from platforms including Facebook and Telegram. Each post was manually annotated and processed to create the dataset, which the authors describe as the first publicly available resource for Amharic suicidal ideation detection. The work addresses what they identify as a pronounced global insufficiency in detecting and mitigating suicidal ideation across linguistically diverse environments, specifically in low-resource languages.

For feature engineering, the researchers tested three word embedding techniques: Word2Vec, GloVe, and FastText. They then trained and compared five deep learning models and four machine learning models on the annotated corpus. The deep architectures included bidirectional sequence models, which can capture contextual information flowing in both directions through Amharic text, an important property for interpreting posts in which meaning depends on surrounding words.

The experiments showed that ensemble learning models achieved slightly higher performance than the individual deep learning models. Among the individual deep learning models, CNN-BiLSTM reached an accuracy of 81.45 percent, while BiLSTM achieved 81.4 percent, indicating that bidirectional sequence learning effectively captures contextual information in Amharic texts. Beyond the empirical results, the study contributes a robust detection framework and a public Amharic dataset, establishing what the authors describe as a foundational methodological resource for mental health research in low-resource language settings.

The authors outline directions for future work: developing scalable datasets, building context-aware deep learning models that handle multimodal data, applying multilingual transformer models, engineering fusion-based features, and addressing bias in low-resource computational linguistics.

Subject of Research: Machine learning detection of suicidal ideation in Amharic social media posts

Article Title: A comparative analysis of machine learning, deep learning, and ensemble learning models for suicidal ideation detection on social media

Article References: Woldegiorgis, A., Abebe, M., Sharma, D. P., & Jimma, W. (2026). A comparative analysis of machine learning, deep learning, and ensemble learning models for suicidal ideation detection on social media. PLOS Mental Health, 3(10), e0000721. https://doi.org/10.1371/journal.pmen.0000721

Image Credits: AI Generated

DOI: 10.1371/journal.pmen.0000721

Keywords: suicidal ideation, Amharic, social media, machine learning, deep learning, ensemble learning, BiLSTM, word embeddings, low-resource languages, natural language processing, dataset annotation, mental health