This repository contains the official dataset and experimental codebase for the study "A Statistically Validated Bidirectional Neural Machine Translation Framework for the Low-Resource Tigrigna–Kunama Language Pair."
The dataset includes:
An expert-verified parallel corpus of 4,712 sentence pairs.
Original, authentic Ge'ez script morphology preserved without homophone normalization.
The codebase includes:
A unified joint-bidirectional Bi-LSTM architecture with Luong global attention.
Preprocessing pipelines, BPE subword modeling (4,000 units), and automated training/evaluation scripts.
Statistical validation protocols (paired bootstrap resampling, 5-fold cross-validation).