DarijaBanking: A New Resource for Overcoming Language Barriers in Banking Intent Detection for Moroccan Arabic Speakers

TL;DR

This work introduces DarijaBanking, a multilingual banking intent dataset tailored to Moroccan Darija and MSA, derived from English banking corpora and expanded through phase-wise cleaning, expansion with IDOOS/OODOOS, automated translation, and manual verification. It conducts a comprehensive benchmarking across BERT-like finetuning, retrieval-based embeddings, LLM prompting, and an NMT pipeline, identifying domain-specific models such as Arabertv02 as the most effective for this setting. The study reveals that cross-lingual transfer alone often underperforms for dialectal Arabic without targeted data, and that retrieval-based approaches offer practical, cost-efficient alternatives when data is limited. The results underscore the importance of domain-specific resources and classifiers for accurate banking intent detection in Moroccan Darija, with BERTouch providing an accessible open-source baseline and the dataset offering a foundation for future work in dialect-aware financial NLP.

Abstract

Navigating the complexities of language diversity is a central challenge in developing robust natural language processing systems, especially in specialized domains like banking. The Moroccan Dialect (Darija) serves as the common language that blends cultural complexities, historical impacts, and regional differences. The complexities of Darija present a special set of challenges for language models, as it differs from Modern Standard Arabic with strong influence from French, Spanish, and Tamazight, it requires a specific approach for effective communication. To tackle these challenges, this paper introduces \textbf{DarijaBanking}, a novel Darija dataset aimed at enhancing intent classification in the banking domain, addressing the critical need for automatic banking systems (e.g., chatbots) that communicate in the native language of Moroccan clients. DarijaBanking comprises over 1,800 parallel high-quality queries in Darija, Modern Standard Arabic (MSA), English, and French, organized into 24 intent classes. We experimented with various intent classification methods, including full fine-tuning of monolingual and multilingual models, zero-shot learning, retrieval-based approaches, and Large Language Model prompting. One of the main contributions of this work is BERTouch, our BERT-based language model for intent classification in Darija. BERTouch achieved F1-scores of 0.98 for Darija and 0.96 for MSA on DarijaBanking, outperforming the state-of-the-art alternatives including GPT-4 showcasing its effectiveness in the targeted application.