Amharic Deep Learning Chatbot Model Shows High Accuracy for HIV Care Information

Researchers developing conversational artificial intelligence for healthcare have built and tested a deep learning text-based chatbot tailored specifically to the Amharic language to deliver accurate HIV and AIDS awareness and care advice. The project directly confronts a persistent disparity in digital health tools, where standard global patient-assistance chatbots primarily operate in English and fail to accommodate the linguistic structures, character variations, and specific living conditions of diverse populations in countries like Ethiopia. To provide a functional alternative, the investigative team curated a specialized health corpus, applied localized text-normalization protocols, and compared multiple recurrent neural network architectures. Across their experimental trials, a bidirectional gated recurrent unit architecture achieved a peak test classification accuracy of 95.01 percent and an objective loss of 0.372, outperforming unidirectional baselines and establishing an effective computational framework for automated Amharic medical guidance.

The public health context motivating this technological development remains severe. Human Immunodeficiency Virus targets human CD4 T cells to weaken the immune system, and despite therapeutic advances since its identification in the early 1980s, the resulting Acquired Immunodeficiency Syndrome presents an ongoing burden across resource-limited settings. In Ethiopia, estimates cited from the Ethiopian Public Health Institute for 2022 and 2023 indicated that approximately 610,350 individuals live with HIV, with an estimated 8,257 new infections and 11,322 AIDS-related deaths occurring each year. While national initiatives have expanded prevention and clinical care programs, broader progress is constrained by widespread gaps in basic understanding, persistent cultural stigmas, and socioeconomic hurdles. Automated conversational systems offer a mechanism to disseminate timely and accessible guidance, yet the absence of annotated Amharic healthcare datasets has historically stalled the deployment of reliable natural language processing models within Ethiopian clinical communication.

To overcome the acute scarcity of localized training resources, the researchers assembled a corpus comprising 10,291 text entries dedicated to core topics including HIV and AIDS causes, symptoms, prevention strategies, and medical treatments. The primary material was compiled from established international and clinical health repositories, specifically the World Health Organization and World Health Line, alongside relevant medical texts, frequently asked questions, and structured interviews conducted with healthcare professionals and patients living with HIV. Because online health documentation was largely available in English, the team initially translated the entire corpus into Amharic using Google Translate before organizing the content into a semi-structured JSON database. This database was structured into three operational categories: 218 discrete tags defining intended user classes, 9,426 patterns capturing varied user queries or frequently asked questions, and 647 distinct candidate responses designed to address those categorized intents.

Processing Amharic text poses distinct computational challenges rooted in the script’s orthography and morphological richness. The Amharic writing system incorporates multiple Fidel characters that share identical phonological pronunciations despite having distinct graphical forms. For example, characters such as ሀ, ኀ, ሃ, ሓ, and ኃ produce identical sounds, as do combinations like ሠ and ሰ, ሐ, ዐ, and አ, or ፀ and ጸ. Without standardization, these character variations unnecessarily inflate vocabulary size and diminish algorithmic learning efficiency. The researchers solved this problem during the text-cleaning and normalization stage by substituting redundant phonetic letters with single representative characters, such as replacing ሠ with ሰ, ኀ with ሐ, and ፀ with ጸ across all corresponding morphological variants. They systematically stripped out punctuation marks including Amharic word separators, full stops, commas, and colons, alongside English-style question marks, special symbols, Arabic numerals, and native Ethiopic numbering characters before executing space-delimited word tokenization.

For the critical step of feature extraction, the researchers deployed FastText embeddings to convert the cleaned word tokens into continuous 200-dimensional numerical vectors. Traditional word embedding techniques and bag-of-words paradigms often falter when applied to morphologically complex languages because rare grammatical inflections generate severe out-of-vocabulary complications. FastText overcomes this limitation by representing individual words as compositions of subword character n-grams, calculating the ultimate vector as the cumulative sum of those constituent segments. This mechanism enabled the retrieval-based chatbot pipeline to construct robust mathematical representations even for unseen word inflections and specialized medical terminology. In the subsequent modeling phase, user queries mapped into these vector representations were matched against intent classes governed by classification decision thresholds, allowing the system to identify the appropriate tag and retrieve a linked response.

The experimental phase evaluated four recurrent neural network architectures: Long Short-Term Memory, Bidirectional Long Short-Term Memory, Gated Recurrent Units, and Bidirectional Gated Recurrent Units. To establish a rigorous baseline, all models shared fixed structural hyperparameters during comparative trials, including an embedding dimension of 200, a hidden dimension of 128 neurons, a training duration of 50 epochs, a batch size of 32, a dropout rate of 0.3 to mitigate overfitting, and a learning rate of 0.0001. Categorical intent classification was optimized using Sparse categorical cross-entropy loss alongside the Softmax activation function. Furthermore, the experiments contrasted the Adaptive Moment Estimation optimizer against Stochastic Gradient Descent, proving that Adam’s adaptive learning rate and momentum mechanisms converged significantly faster and navigated the complex multi-class intent space with greater precision.

When evaluated on a standard 80:20 train-test data split, the comparative metrics revealed substantial performance differences across the four deep learning frameworks. The basic unidirectional LSTM model delivered a training accuracy of 93 percent and a testing accuracy of 87 percent, with precision and recall scores of 86 percent and an F1-score of 85 percent. Incorporating bidirectionality in the BiLSTM architecture raised performance, securing a training accuracy of 97 percent, a testing accuracy of 92 percent, precision of 91 percent, recall of 91 percent, and an F1-score of 90 percent. The unidirectional GRU model demonstrated efficient training dynamics and rapid convergence, achieving 98 percent training accuracy and 94 percent testing accuracy, accompanied by 93 percent precision, 92 percent recall, and an F1-score of 92 percent. Across all experiments, the bidirectional networks exhibited consistent advantages over unidirectional configurations by retaining contextual linguistic information from both past and future sequence tokens.

The BiGRU architecture emerged as the top-performing model, demonstrating superior capability in capturing bidirectional contextual dependencies across user queries. When trained with the Adam optimizer and Softmax activation under an 80:20 data split, the BiGRU model achieved a training accuracy of 98 percent and a peak test accuracy of 95.01 percent, paired with a categorical cross-entropy loss of 0.372. Its broader evaluation metrics remained balanced at 93 percent precision, 93 percent recall, and an F1-score of 92 percent. Comparative experiments on activation functions within the recurrent architectures further demonstrated that Softmax consistently surpassed the Rectified Linear Unit function by providing coherent probability distributions across the 218 intent classes; for instance, in GRU tests, Softmax achieved 95.01 percent accuracy compared to 92.29 percent for ReLU.

The research team also investigated how different data partitioning ratios influenced model behavior, comparing splits of 70:30, 80:20, and 90:10 across all four network architectures. In accordance with machine learning empirical standards, allocating larger proportions of data to the training phase generally improved classification metrics. Across both the 70:30 and 90:10 partitions, the relative performance hierarchy remained stable: the BiGRU model consistently maintained the leading position across accuracy, precision, recall, and F1-score, followed in order by GRU, BiLSTM, and LSTM. Nevertheless, the 80:20 split configuration yielded the most stable generalization balance when combined with early stopping routines, which terminated iterative updates whenever validation performance ceased to gain, successfully halting overtraining before parameter degradation occurred.

To evaluate whether algorithmic benchmark success translated into operational usability, the researchers conducted a human evaluation using an interactive graphical user interface. Real users submitted queries directly to the chatbot prototype, allowing evaluators to measure answer correctness as the percentage of questions accurately addressed relative to the total inquiries presented. In this interactive trial, the chatbot attained an overall human evaluation correctness score of 87.6 percent, accurately interpreting intents when matching patterns existed in the underlying repository and generating fallback exception messages when queries fell outside its learned scope. The assessment documented that the model missed or failed to provide desired responses on 10 out of 72 presented queries, demonstrating effective domain competence while outlining clear avenues for dataset expansion, speech interface integration, and further Amharic natural language refinements in future work.

Subject of Research: Deep learning conversational AI model for HIV/AIDS awareness and care in the Amharic language

Article Title: Developing amharic text-based chatbot model for HIV/AIDS awareness and care using deep learning approaches

Article References: Sintayehu, A., & Emiru, E. D. (2025). Developing amharic text-based chatbot model for HIV/AIDS awareness and care using deep learning approaches. BMC Artificial Intelligence, 1(1), Article 2. https://doi.org/10.1186/s44398-025-00002-9

Image Credits: AI Generated

DOI: 10.1186/s44398-025-00002-9

Keywords: Artificial Intelligence, Deep Learning, Chatbots, Natural Language Processing, BiGRU, HIV/AIDS Awareness, Amharic Language, Healthcare AI, Developing, amharic, text-based, chatbot