ZambiaABSA

ZambiaABSA is a multilingual Aspect-Based Sentiment Analysis (ABSA) model for Zambian ride-hailing social-media reviews. It uses an encoder that was first adapted to Zambian social-media language through ZambiaSocialBERT, before being fine-tuned for aspect-conditioned sentiment classification.

Task

The model performs aspect-conditioned sentiment classification. Given a review and a target service aspect, it predicts the sentiment expressed toward that specific aspect, rather than an overall sentiment for the whole review.

Sentiment classes: negative (0), neutral (1), positive (2).

Service aspects: driver_behavior, pricing, app_performance, payment, ride_quality, customer_support, service_quality, booking, safety, waiting_time.

Base model

kelvinmbewe/ZambiaSocialBERT — a multilingual BERT encoder adapted, using QLoRA, on a Zambian language-identification task across English, Bemba_Cibemba, Nyanja_Cinyanja, Lusaka_Slang, and noisy social-media content. This prior adaptation exposes the encoder to Zambian multilingual and code-switched text before ABSA fine-tuning.

Training data

The model was fine-tuned on the same synthetic multilingual ride-hailing ABSA dataset used for baselineABSA, containing English, Bemba_Cibemba, Nyanja_Cinyanja, and Lusaka_Slang reviews with aspect-level sentiment annotations.

Training procedure

The ABSA fine-tuning stage used the same configuration as baselineABSA: 7 epochs, learning rate 2e-5, train and evaluation batch size 16, weight decay 0.01, AdamW optimizer, evaluation and checkpoint saving per epoch, best checkpoint selected on macro F1-score. The only difference between ZambiaABSA and baselineABSA is the prior ZambiaSocialBERT adaptation of the encoder; the ABSA fine-tuning itself was identical.

Intended use

ZambiaABSA was developed as part of a master's dissertation on multilingual Aspect-Based Sentiment Analysis for low-resource Zambian ride-hailing social-media discourse. It is compared against baselineABSA to evaluate whether prior adaptation improves downstream sentiment interpretation.

Limitations

The model was trained on synthetic data; performance on naturally occurring reviews may differ. Improvements over the baseline were selective rather than uniform, and neutral sentiment and aspect entanglement remained difficult, as documented in the associated dissertation.