📄 Deep Learning Based Amharic Chatbot for FAQs in Universities

Summary

The paper tackles a practical gap: Ethiopian university students repeatedly ask staff the same questions, and no chatbot existed for answering such FAQs in Amharic, a morphologically rich, low-resource language. The authors framed the work as Design Science Research, following the six-step process of Peffers et al. They gathered real student questions through a questionnaire at Mekelle University and Aksum University, translated and corrected them into Amharic with Google Translator and native experts, and organized them into 60 intent topics in a JSON file. Input text is tokenized, normalized to handle Amharic Fidel character variants, cleaned of stop words, and stemmed. Three intent classifiers were compared: SVM, Multinomial Naïve Bayes, and a deep neural network built with TensorFlow, Keras and NLTK. The neural model performed best, at 91.55% accuracy. The bot was connected to Facebook Messenger and hosted on Heroku for round-the-clock use. The work matters as an early end-to-end, openly released Amharic FAQ assistant and a template for other under-resourced languages.

Key points

• The dataset came from questionnaires filled in by 80 engineering students at two universities (Mekelle and Aksum), chosen by purposive sampling, and was reduced to 60 topics.
• Each intent in the JSON file stores a tag, sample patterns, candidate responses (one is picked at random), a context set and a context filter.
• The best model, a deep neural network trained with the Adam optimizer and a SoftMax output, reached 91.55% accuracy and 0.3548 validation loss, beating the SVM and Multinomial Naïve Bayes baselines.
• Preprocessing targets Amharic-specific issues: Fidel (script) variation, morphological variation and lexical gaps.
• The bot was deployed via Facebook Messenger on Heroku. The dataset, model and code are released on GitHub, and integrating an Amharic WordNet is proposed as future work.

Interesting ideas

• I1. Intents carry a "context set" that changes the dialogue state and a "context filter" that limits which intents can match in that state, giving a cheap, inspectable way to model conversation flow. — For Mova (frames): A comparable explicit state flag could restrict which speech-act or frame readings are allowed given the preceding turn, keeping the disambiguation choice traceable.
• I2. Script normalization collapses interchangeable Fidel characters before classification, so surface spelling variants do not split the vocabulary. — For Mova (tokens): A transparent normalization table for English spelling variants (British vs. American, hyphenation) placed before lemmatization would help lemmas and lookups line up, with each mapping logged for explanation.
• I3. The authors gathered evaluation material by asking real users what they actually ask, rather than inventing questions. — For Mova (answer): Test sets for the QA readers could be seeded from questions collected from real learners, giving a more realistic difficulty profile than synthetic items.
• I4. Classic linear classifiers (SVM, Naïve Bayes) were kept as baselines next to the neural model on the same data. — For Mova (learn): Every new learned component could be reported against a simple, cheap baseline on identical splits, which shows how much the added complexity really buys.
• I5. The authors see a WordNet as the remedy for lexical gaps, where users phrase a question with synonyms absent from the training patterns. — For Mova (world): WordNet synonym and hypernym links in the compiled ontology could map unseen wording onto known concepts, so matching relies on meaning rather than surface overlap.