Kelvinmbewe/ZambianABSA

Lusaka Multilingual Aspect-Based Sentiment Analysis (ABSA) Model

Model Identifier

Kelvinmbewe/mbert_Lusaka_Language_Analysis

---

1. Model Summary

We present a multilingual Aspect-Based Sentiment Analysis (ABSA) model fine-tuned from mBERT for ride-hailing service reviews in a Lusaka urban context. The model performs token-level BIO tagging to extract aspect spans and classify their sentiment polarity.

The system is designed for multilingual and code-switched inputs involving:

• English
• Bemba
• Nyanja
• English–Bemba code-switch
• English–Nyanja code-switch

This work contributes toward low-resource African language sentiment modeling under multilingual transformer architectures.

---

2. Task Definition

The model performs joint:

1. Aspect extraction (span detection)
2. Sentiment classification (Positive, Negative, Neutral)

Formally:

Given a sentence ( S = {w_1, w_2, ..., w_n} ),

The model predicts:

• A BIO tag for each token
• Sentiment polarity for extracted aspect spans

Label schema:

O
B-POS / I-POS
B-NEG / I-NEG
B-NEU / I-NEU

---

3. Model Architecture

• Base model: bert-base-multilingual-cased
• Architecture: Transformer encoder (12-layer)
• Fine-tuning objective: Token classification
• Maximum sequence length: 64 tokens
• Optimizer: AdamW
• Learning rate: 1e-5 (recommended configuration)
• Scheduler: Linear decay
• Mixed precision (FP16) training enabled

---

4. Training Data

The training corpus consists of approximately 1,000+ annotated ride-hailing reviews and complaints reflecting Lusaka urban language use.

Dataset Properties

• Balanced sentiment distribution (POS / NEG / NEU)
• Monolingual and code-switched samples
• BIO-annotated aspect spans
• Domain: ride-hailing service (driver, payment, app, ride, safety, etc.)

The dataset contains synthetic and semi-synthetic examples constructed to simulate realistic Lusaka multilingual patterns.

---

5. Evaluation

Evaluation is conducted at the token level using macro-averaged metrics.

Metrics reported:

• Accuracy
• Precision (macro)
• Recall (macro)
• F1-score (macro)

Typical performance range:

Metric	Approximate Range
Accuracy	0.70–0.80
Macro F1	0.65–0.75

These scores reflect token-level evaluation and may not directly correspond to span-level exact-match performance.

---

6. Intended Use

This model is intended for:

• Research on low-resource ABSA
• Multilingual sentiment modeling
• Code-switch robustness experiments
• Zambian NLP system development
• Ride-hailing feedback analytics

It is suitable for exploratory and research-driven deployments.

---

7. Limitations

1. Evaluation is token-level; span-level metrics are not provided.

2. Synthetic data may introduce distribution bias.

3. Limited dialectal variation beyond Lusaka urban patterns.

4. Performance may degrade on:

   • Long-form paragraphs
   • Heavy slang
   • Out-of-domain text

5. Code-switch handling depends on subword tokenization quality.

---

8. Ethical Considerations

• The model is trained on simulated ride-service feedback and does not include personal user data.
• Predictions may contain bias introduced through synthetic pattern repetition.
• The system should not be used for automated decision-making affecting individuals.

---

9. Reproducibility

training_args = TrainingArguments(
    output_dir=OUTPUT_DIR,            # Directory where model checkpoints and outputs will be saved
    
    eval_strategy="epoch",            # Run evaluation at the end of each training epoch
    save_strategy="epoch",            # Save a checkpoint at the end of each epoch
    save_total_limit=1,               # Keep only the most recent/best checkpoint to save disk space
    
    learning_rate=1e-5,               # Initial learning rate for the optimizer
    num_train_epochs=10,              # Total number of full passes through the training dataset
    
    warmup_ratio=0.1,                 # Warm up learning rate for the first 10% of total training steps
    lr_scheduler_type="linear",       # Linearly decay learning rate after warmup
    
    per_device_train_batch_size=8,    # Batch size per GPU/CPU during training
    per_device_eval_batch_size=8,     # Batch size per GPU/CPU during evaluation
    
    weight_decay=0.01,                # Apply L2 regularization to help prevent overfitting
    
    logging_strategy="epoch",         # Log training metrics at the end of each epoch
    
    load_best_model_at_end=True,      # After training, automatically load the best-performing checkpoint
    metric_for_best_model="f1_macro", # Metric used to determine which checkpoint is "best"
    greater_is_better=True,           # Higher F1 score indicates better performance
    
    fp16=True                         # Enable mixed precision (16-bit) training for faster performance and lower memory usage
)

10. Example

Input:

Payment yachoka twice but driver ali musuma sana.

Output:

[
  {"aspect": "Payment yachoka twice", "sentiment": "neg"},
  {"aspect": "driver ali umusuma sana", "sentiment": "pos"}
]

---

11. Research Contribution

This model demonstrates:

• Feasibility of ABSA under multilingual Zambian mixed language
• mBERT adaptability to low-resource sentiment tasks
• Token-level ABSA in mixed-language urban settings
• Synthetic dataset viability for bootstrapping low-resource domains

---