Graph-Augmented Transformers for Hausa Sentiment Analysis: When Does Graph Structure Help?

• Hafsa K. AHMAD

  Author

Hausa is still under-represented within natural language processing (NLP) research, despite having one of the largest speaker populations on the African continent. This study asks whether graph-augmented transformer architectures deliver measurable gains for sentiment analysis under those conditions. A hybrid model was developed and evaluated that combines a pre-trained transformer encoder with a Graph Attention Network (GAT) built over a document-word corpus graph, whose node features are obtained from Term Frequency–Inverse Document Frequency (TF-IDF) representations and fused via a learned scalar attention gate. Through systematic experiments across three transformer backbones, namely AfriBERTa (Hausa-inclusive pretraining), XLM-RoBERTa (general multilingual), and mDeBERTa-v3 (general multilingual), across four labeled data regimes (10%, 25%, 50%, 100%), the results show that the value of graph augmentation depends critically on the transformer’s language-specific pretraining. When paired with AfriBERTa, the hybrid offers no significant benefit (ΔF1 = −0.003), as language-specific pretraining already captures sufficient lexical and structural information. When paired with XLM-RoBERTa, the hybrid achieves a consistent gain of +0.017 Macro F1, with the advantage growing with data availability. The findings further demonstrate that TF-IDF node feature initialization is essential for enabling genuine transformer-GAT fusion, and that random initialization causes the attention gate to collapse to transformer-only behavior. These results suggest that for Hausa and other comparable low-resource languages, a well-adapted language-specific model is often enough, and that graph augmentation becomes truly useful as a complementary technique when only general multilingual transformers are available. Taken together, these findings offer concrete guidance for researchers developing NLP systems for Hausa and similar low-resource languages.

• 8_Ahmad (2026)

Copyright (c) 2026 Hafsa K. AHMAD (Author)

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License.