Deep Learning Based Amharic Chatbot for FAQs in Universities

1. Problem Statement

Universities in Ethiopia face significant inefficiencies in handling Frequently Asked Questions (FAQs) from students. The current reliance on physical visits to administrative offices or repetitive email exchanges creates a bottleneck for both students and staff. While chatbots exist for major global languages, there is a critical lack of solutions for Amharic, a low-resource language characterized by:

• Morphological richness: Complex word structures and variations.
• Lexical gaps: Limited digital resources and standardized datasets.
• Script variations: Challenges in distinguishing Amharic from other languages using similar scripts (e.g., Ge'ez script variations).
• Dialectal nuances: Regional accents and linguistic varieties that standard translation tools often fail to capture.

The absence of an intelligent, Amharic-specific FAQ system forces students to waste time searching for information, while administrators are burdened with repetitive tasks.

2. Methodology

The study employed Design Science Research (DSR) to develop a functional artifact. The process followed a six-step model: Problem Identification, Objectives, Design/Development, Demonstration, Evaluation, and Communication.

A. Data Collection and Preprocessing

• Dataset: A purposive sampling method was used to collect data from 80 engineering students (across Mekelle and Aksum Universities).
• Structure: The data was organized into a JSON format containing 60 distinct topics (tags). Each entry included patterns (sample queries), responses, and context filters.
• Preprocessing: The Amharic text underwent standard NLP preprocessing:
  • Tokenization: Breaking text into words.
  • Normalization: Standardizing character variations.
  • Stop Word Removal: Eliminating non-essential words.
  • Stemming: Reducing words to their root forms to handle morphological variations.

B. Chatbot Architecture

The system consists of three main components:

1. User Interface: Integrated with Facebook Messenger via a Flask webhook, allowing natural language interaction.
2. Backend Server: Deployed on Heroku for 24/7 accessibility.
3. Core Model: An intent classification engine designed to map user queries to specific FAQ categories.

C. Model Development and Comparison

The researchers compared three machine learning algorithms for intent classification:

1. Support Vector Machine (SVM)
2. Multinomial Naïve Bayes (MNB)
3. Deep Neural Network (DNN)

The DNN Architecture (Selected Model):

• Framework: Built using TensorFlow and Keras.
• Structure: A Sequential model with:
  • Input layer (matching feature vector size).
  • Two hidden dense layers with 128 and 64 neurons respectively.
  • ReLU activation for hidden layers and Softmax for the output layer.
  • Dropout layers (0.2 probability) to prevent overfitting.
• Training: Optimized using the Adam optimizer with a learning rate of 0.0001. The model was trained for 150 epochs with a batch size of 8.
• Inference: User input is tokenized and converted to a vector using NLTK and the Bag of Words method. The model calculates probability scores for each intent class, selecting the highest probability to trigger a response.

3. Key Contributions

• First Amharic University FAQ Chatbot: This is the first documented system specifically designed to handle university FAQs in the Amharic language, addressing a significant gap in low-resource NLP.
• Deep Learning Framework: Demonstrated the superiority of Deep Neural Networks over traditional machine learning (SVM, MNB) for handling the morphological complexity of Amharic.
• Scalable Architecture: Provided a reusable JSON-based framework that allows for easy integration of new datasets and intents without retraining the entire system from scratch.
• Real-World Deployment: Successfully deployed a functional prototype on Facebook Messenger, proving the feasibility of AI assistants in Ethiopian higher education contexts.

4. Experimental Results

The study evaluated the models using Accuracy, Precision, Recall, and F1-Score on a test set (20% of the 850 collected queries).

Model	Accuracy	Precision	Recall	F1-Score
Deep Neural Network (DNN)	91.55%	85.98%	87.13%	85.23%
Support Vector Machine (SVM)	87.96%	-	-	-
Multinomial Naïve Bayes (MNB)	60.19%	-	-	-

• Optimization: The DNN with ReLU activation significantly outperformed Sigmoid, Tanh, and Linear activation functions.
• User Satisfaction: A survey of 15 participants (students and lecturers) yielded an 86.2% overall satisfaction rate. Users rated the system highly on "Ease of Use" (98.6%) and "Amharic Alphabet Variation Support" (96.0%).

5. Significance and Future Work

Significance:
The research validates that deep learning techniques can effectively overcome the linguistic challenges of Amharic, such as morphological variation and lexical gaps. It provides a foundational framework for developing conversational agents in under-resourced languages, moving beyond keyword matching to context-aware intent recognition.

Future Directions:

• Sentiment Analysis: Integrating emotion detection to provide empathetic responses.
• Expanded Language Support: Adapting the model for other Ethiopian languages.
• Complex Query Handling: Moving beyond simple FAQs to answer comparative, procedural, and complex questions.
• Voice Integration: Incorporating speech-to-text and text-to-speech for voice-based interaction.
• Resource Enhancement: Utilizing Amharic WordNet and expanding the FAQ dataset to further reduce lexical gaps and improve accuracy.

In conclusion, this paper presents a robust, high-accuracy solution for automating student support in Ethiopian universities, serving as a critical stepping stone toward comprehensive, multi-domain Amharic conversational AI.