Kelvinmbewe/mbert_LusakaLang_MultiTask

LusakaLang Multi‑Task Model (Language + Sentiment + Topic)

Model Description

LusakaLang‑MultiTask is a unified transformer model built on top of bert-base-multilingual-cased, designed to perform three tasks simultaneously:

1. Language Identification
2. Sentiment Analysis
3. Topic Classification

The model integrates three fine‑tuned LusakaLang checkpoints:

• Kelvinmbewe/mbert_Lusaka_Language_Analysis
• Kelvinmbewe/mbert_LusakaLang_Sentiment_Analysis
• Kelvinmbewe/mbert_LusakaLang_Topic

All tasks share a single mBERT encoder, with three independent classifier heads.
This architecture improves efficiency, reduces memory footprint, and enables consistent predictions across tasks.

---

Why This Model Matters

Zambian communication is multilingual, fluid, and highly context‑dependent.
A single message may include:

• English
• Bemba
• Nyanja
• Slang
• Code‑switching
• Cultural idioms
• Indirect emotional cues

This model is designed specifically for that environment.

It excels at:

• Identifying the dominant language or code‑switching
• Detecting sentiment polarity in culturally nuanced text
• Classifying topics such as:
  • driver behaviour
  • payment issues
  • app performance
  • customer support
  • ride availability

---

Training Architecture

The model uses:

• Shared Encoder: mBERT
• Head 1: Language classifier
• Head 2: Sentiment classifier
• Head 3: Topic classifier

This multi‑task setup improves generalization and reduces inference cost.

---

Performance Summary

Language Identification

Metric	Score
Accuracy	0.97
Macro‑F1	0.96

Sentiment Analysis (Epoch 30 — Final Checkpoint)

Metric	Score
Accuracy	0.9322
Macro‑F1	0.9216
Negative F1	0.8649
Neutral F1	0.95
Positive F1	0.95

Topic Classification

Metric	Score
Accuracy	0.91
Macro‑F1	0.90

---

How to Use This Model

Load the Multi‑Task Model

from transformers import AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("Kelvinmbewe/LusakaLang-MultiTask")
model = torch.load("Kelvinmbewe/LusakaLang-MultiTask/model.pt")
model.eval()

predict_language([
    "Ndeumfwa bwino lelo",
    "Galimoto inachedwa koma driver anali bwino",
    "The service was terrible today"
])

predict_sentiment([
    "Driver was rude and unprofessional",
    "Ndimvela bwino lelo",
    "The ride was okay, nothing special"
])

predict_topic([
    "Payment failed but money was deducted",
    "Support siyankhapo, waited long",
    "Driver was over speeding"
])

@model{LusakaLangMultiTask,
  author = {Kelvin Mbewe},
  title = {LusakaLang Multi-Task Model},
  year = 2025,
  url = {https://huggingface.co/Kelvinmbewe/LusakaLang-MultiTask}
}

                         ┌──────────────────────────────────────┐
                         │         Input Text (Any Language)     │
                         └──────────────────────────────────────┘
                                         │
                                         ▼
                         ┌──────────────────────────────────────┐
                         │        Tokenizer (mBERT-based)        │
                         └──────────────────────────────────────┘
                                         │
                                         ▼
                         ┌──────────────────────────────────────┐
                         │      Shared mBERT Encoder Layer       │
                         │  (bert-base-multilingual-cased)       │
                         └──────────────────────────────────────┘
                                         │
                                         ▼
                         ┌──────────────────────────────────────┐
                         │        [CLS] Pooled Representation    │
                         └──────────────────────────────────────┘
                                         │
         ┌───────────────────────────────┼───────────────────────────────┐
         │                               │                               │
         ▼                               ▼                               ▼
┌──────────────────────┐     ┌────────────────────────┐     ┌────────────────────────┐
│  Language Head        │     │  Sentiment Head        │     │   Topic Head            │
│ (Kelvinmbewe/         │     │ (Kelvinmbewe/          │     │ (Kelvinmbewe/           │
│  mbert_Lusaka_        │     │  mbert_LusakaLang_     │     │  mbert_LusakaLang_      │
│  Language_Analysis)   │     │  Sentiment_Analysis)   │     │  Topic)                 │
└──────────────────────┘     └────────────────────────┘     └────────────────────────┘
         │                               │                               │
         ▼                               ▼                               ▼
┌──────────────────────┐     ┌────────────────────────┐     ┌────────────────────────┐
│  Language Label       │     │  Sentiment Label       │     │  Topic Label            │
│ (e.g., Bemba, Nyanja, │     │ (Negative/Neutral/     │     │ (Driver, Payment,       │
│  English, Code‑Switch)│     │  Positive)             │     │  Support, etc.)         │
└──────────────────────┘     └────────────────────────┘     └────────────────────────┘