A Comparative Evaluation of Deep Learning Architectures for Amharic Riddle Meaning Detection
TL;DR
The finding establishes the viability of automated semantic detection in Ethiopian folk literature and should expand to other Ethiopian languages to enhance real-time, multilingual cultural preservation.
Abstract
Language is a rule-governed system of symbols that enables communication, expresses thought, and preserves culture across diverse contexts. Amharic, a morphologically rich Semitic language and the official working language of Ethiopia, serves as a vital medium for cultural heritage, particularly through the tradition of riddles (እንቆቅልሽ). Riddles are sophisticated cultural tools characterized by metaphorical ambiguity and deep semantic complexity. Despite their educational value, automated interpretation remains a significant challenge due to Amharic’s morphological complexity and the scarcity of specialized Natural Language Processing (NLP) frameworks. This study aims to develop a model for Amharic riddle meaning detection. Using an experimental research design with a dataset of 8,457 samples, we implemented a robust pre-processing pipeline involving normalization, tokenization, encoding, and sequence padding. We evaluated four deep learning architectures: Convolutional Neural Network (CNN), Deep Neural Network (DNN), Bidirectional Long Short-Term Memory (Bi-LSTM), and Bidirectional Gated Recurrent Unit (Bi-GRU). Experimental results demonstrate that the Bi-LSTM model achieved superior performance, reaching a peak accuracy of 98.64%. The finding establishes the viability of automated semantic detection in Ethiopian folk literature. Future research should expand to other Ethiopian languages to enhance real-time, multilingual cultural preservation.