Task-Oriented Dialog Systems for the Senegalese Wolof Language

Recent advances in deep learning have propelled NLP forward, especially with models like BERT, GPT, and large-scale pretraining. However, these models rely heavily on resource-rich languages, leaving low-resource languages behind. African languages, including Wolof, face severe data scarcity, lack of standardized orthography, and limited digital presence. Initiatives like Masakhane have sought to collect data and develop models, but progress remains slow. Cross-lingual transfer, where knowledge from high-resource languages is leveraged, has emerged as a promising approach. Task-oriented dialog systems (ToDS) are vital for automating services but demand extensive annotated datasets. This paper addresses these challenges by proposing a hybrid method combining machine translation, annotation projection, and language-agnostic embeddings to rapidly develop Wolof dialogue systems.

Developing effective dialogue systems for Wolof is hindered by scarce annotated data, inconsistent orthography, and limited NLP resources. Traditional data collection methods are costly and time-consuming, making it difficult to achieve high accuracy in intent recognition and slot filling. While cross-lingual transfer offers hope, its success depends on translation quality and annotation alignment. Existing approaches struggle with maintaining label consistency during translation, especially in low-resource settings. The core problem is how to generate high-quality, annotated datasets in Wolof with minimal native data, ensuring models can accurately interpret user intents and facilitate natural interactions in real-world applications.

The paper introduces a novel annotation projection pipeline that combines in-house French-Wolof machine translation with identifier-based label preservation, ensuring accurate transfer of annotations. It leverages LaBSE embeddings for language-independent intent classification, enabling the system to generalize across languages without retraining. This hybrid approach reduces data collection costs, accelerates system deployment, and maintains high performance. Additionally, the integration of Rasa’s modular architecture with custom translation and projection algorithms creates a scalable framework adaptable to other low-resource languages, addressing a significant gap in multilingual NLP.

• �� Build a high-quality French-Wolof machine translation system, ensuring accurate translation.
• �� Extract intent and slot annotations from Amazon Massive dataset in French.
• �� Use a novel annotation projection algorithm:
• Replace labels with unique identifiers ($0, $1, etc.) in source sentences.
• Translate sentences using the custom MT system, keeping identifiers intact.
• Back-translate and replace identifiers with corresponding Wolof translations.
• �� Train Rasa’s DIET model on the synthetic Wolof dataset for intent detection and slot filling.
• �� Use LaBSE embeddings for language-agnostic intent classification, enabling cross-lingual support.
• �� Evaluate model performance on a test set, comparing with French baseline, analyzing translation impact and label alignment.

The experiments involve training the intent classifier on both original French data and synthetic Wolof data generated via translation and annotation projection. The dataset from Amazon Massive covers 18 domains, with 1 million utterances in total, including 16,500 in French. The model is trained with an 80/20 split, and performance is measured using macro F1, accuracy, and confidence scores. Multiple translation strategies are tested, including marker-based and identifier-based projection, with the latter showing superior stability. Results demonstrate that the synthetic Wolof dataset yields nearly identical intent classification performance to the French dataset, validating the approach’s robustness. Ablation studies analyze the effect of translation quality and annotation accuracy on overall system performance.

The Wolof intent classifier achieved a macro F1 of 0.995, with an accuracy of 94%, closely matching the French baseline (F1 0.999). Confidence scores indicate high reliability, though some confusion arises in similar intents like calendar_query and recommendation_events, due to translation nuances. The annotation projection method maintained label integrity with the custom identifier approach, significantly outperforming marker-based methods. These results confirm that cross-lingual transfer, combined with high-quality machine translation, can produce effective low-resource dialogue systems, reducing data annotation efforts by over 80%. The system’s scalability suggests potential extension to other African languages with minimal adaptation.

This methodology enables rapid deployment of task-oriented dialogue systems in low-resource languages, supporting applications like customer service, information retrieval, and voice assistants in African contexts. Minimal native data requirements lower entry barriers for developers and organizations, facilitating multilingual AI solutions. The approach can be integrated into existing platforms like Rasa, providing a scalable, cost-effective way to serve diverse linguistic communities. Long-term, it paves the way for inclusive AI, where underrepresented languages gain access to advanced NLP tools, promoting digital equity and local language preservation.

The system’s performance heavily depends on the quality of the machine translation system; poor translations lead to misaligned labels and reduced accuracy. Variations in Wolof orthography and informal writing styles pose challenges for model robustness. Currently, the focus is on intent classification; extending to slot filling and multi-turn dialogues requires further research. Computational costs for training and translation remain significant, especially for larger datasets. Future improvements should address translation errors, dialectal variations, and multi-task capabilities to enhance practical deployment in diverse real-world scenarios.