Towards an NLP Model for the Algerian Dialect: Building a Corpus for Machine Translation Processing

Abstract

Neural Machine Translation (NMT) for Arabic dialects presents a significant challenge within Natural Language Processing due to a critical scarcity of digital resources. This issue is particularly acute for the Maghrebi continuum, where Algerian dialects like Oranian (ORN) are technologically underserved. This dissertation addresses this absolute void by tackling the dual problem of resource absence and the lack of MT benchmarks for the ORN-English language pair. The study details the creation of the novel Oranian-English Parallel Corpus (OEPC), an 8.5K sentence-pair dataset compiled through meticulous processes. OEPC was subsequently used to design and train a Sequence-to-Sequence (Seq2Seq) NMT system. The model’s architecture is based on a two-layer stacked Long Short-Term Memory (LSTM) network, critically augmented with an additive attention mechanism. The final model achieves a robust BLEU score of 30.9 on a held-out test set, significantly outperforming an attention-less baseline. Notably, this result is highly competitive, surpassing even larger, semi-supervised systems for other Arabic dialects, which highlights the impact of a high-quality, linguistically consistent corpus. Qualitative error analysis confirms the model’s competency with core grammatical structures while identifying classic low-resource challenges related to out-of-vocabulary words, idiomatic expressions, and semantic ambiguity. The findings validate a "depth-over-breadth" corpus strategy as an effective approach for initiating research on zero-resource dialects. This work provides a methodological blueprint for future work pointing towards Transformer architectures and semi-supervised learning.