metadata
language:
- en
- bem
- ny
tags:
- multi-task
- sentiment-analysis
- topic-classification
- language-identification
- multilingual
- transformer
- zambia
- lusaka
license: apache-2.0
library_name: transformers
pipeline_tag: text-classification
base_model:
- Kelvinmbewe/mbert_Lusaka_Language_Analysis
- Kelvinmbewe/mbert_LusakaLang_Sentiment_Analysis
- Kelvinmbewe/mbert_LusakaLang_Topic
model-index:
- name: LusakaLang-MultiTask
results:
- task:
type: text-classification
name: Language Identification
dataset:
name: LusakaLang Language Data
type: lusakalang
split: test
metrics:
- type: accuracy
value: 0.97
name: accuracy
- type: f1
value: 0.96
name: f1_macro
- type: accuracy
value: 0.9322
name: accuracy
- type: f1
value: 0.9216
name: f1_macro
- type: f1
value: 0.8649
name: f1_negative
- type: f1
value: 0.95
name: f1_neutral
- type: f1
value: 0.95
name: f1_positive
- type: accuracy
value: 0.91
name: accuracy
- type: f1
value: 0.9
name: f1_macro
LusakaLang MultiTask Model
This model is a unified transformer architecture built on top of bert-base-multilingual-cased, designed to perform three tasks simultaneously:
- Language Identification
- Sentiment Analysis
- Topic Classification
The system integrates three fine‑tuned LusakaLang checkpoints:
- mbert_Lusaka_Language_Analysis
- mbert_LusakaLang_Sentiment_Analysis
- mbert_LusakaLang_Topic
All tasks share a single mBERT encoder, supported by three independent classifier heads. This architecture enhances computational efficiency, reduces memory overhead and promotes consistent, harmonized predictions across all tasks.
Why This Model Matters
Zambian communication is inherently multilingual, fluid, and deeply shaped by context. A single message may blend English, Bemba, Nyanja, local slang, and frequent code‑switching, often expressed through culturally grounded idioms and subtle emotional cues. This model is designed specifically for that environment, where meaning depends not only on the words used but on how languages interact within a single utterance.
It excels at identifying the dominant language or detecting when multiple languages are being used together, interpreting sentiment even when it is conveyed indirectly or through culturally specific phrasing, and classifying text into practical topics such as driver behaviour, payment issues, app performance, customer support, and ride availability. By capturing these nuances, the model provides a more accurate and context‑aware understanding of real Zambian communication.
How to Use This Model
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn.functional as F
class LusakaLangMultiTask:
"""
LusakaLang MultiTask Model:
- Language Identification
- Sentiment Analysis
- Topic Classification
"""
def __init__(self, path="Kelvinmbewe/LusakaLang-MultiTask",
lang_temp=1.0, sent_temp=1.0, topic_temp=1.0):
# Load tokenizer and models
self.tokenizer = AutoTokenizer.from_pretrained(path)
self.lang_model = AutoModelForSequenceClassification.from_pretrained(
"Kelvinmbewe/mbert_Lusaka_Language_Analysis"
)
self.sent_model = AutoModelForSequenceClassification.from_pretrained(
"Kelvinmbewe/mbert_LusakaLang_Sentiment_Analysis"
)
self.topic_model = AutoModelForSequenceClassification.from_pretrained(
"Kelvinmbewe/mbert_LusakaLang_Topic"
)
# ID2Label mappings
self.lang_id2label = self.lang_model.config.id2label
self.sent_id2label = self.sent_model.config.id2label
self.topic_id2label = self.topic_model.config.id2label
# Temperature scaling
self.lang_temp = lang_temp
self.sent_temp = sent_temp
self.topic_temp = topic_temp
# Device setup
self.device = "cuda" if torch.cuda.is_available() else "cpu"
self.lang_model.to(self.device)
self.sent_model.to(self.device)
self.topic_model.to(self.device)
def predict_batch(self, texts, conf_threshold=0.5, batch_size=16):
"""
Predict language, sentiment, and topic for a list of texts.
Returns a list of dicts: [{"text": ..., "language":..., "sentiment":..., "topic":...}, ...]
"""
results = []
for i in range(0, len(texts), batch_size):
batch_texts = texts[i:i + batch_size]
inputs = self.tokenizer(
batch_texts,
return_tensors="pt",
truncation=True,
padding=True
).to(self.device)
with torch.no_grad():
# Language
lang_logits = self.lang_model(**inputs).logits / self.lang_temp
lang_probs = F.softmax(lang_logits, dim=-1)
lang_conf, lang_idx = torch.max(lang_probs, dim=-1)
# Sentiment
sent_logits = self.sent_model(**inputs).logits / self.sent_temp
sent_probs = F.softmax(sent_logits, dim=-1)
sent_conf, sent_idx = torch.max(sent_probs, dim=-1)
# Topic
topic_logits = self.topic_model(**inputs).logits / self.topic_temp
topic_probs = F.softmax(topic_logits, dim=-1)
topic_conf, topic_idx = torch.max(topic_probs, dim=-1)
for j, text in enumerate(batch_texts):
results.append({
"text": text,
"language": self.lang_id2label[lang_idx[j].item()]
if lang_conf[j].item() >= conf_threshold else "unknown",
"language_conf": round(lang_conf[j].item(), 3),
"sentiment": self.sent_id2label[sent_idx[j].item()],
"sentiment_conf": round(sent_conf[j].item(), 3),
"topic": self.topic_id2label[topic_idx[j].item()]
if topic_conf[j].item() >= conf_threshold else "unknown",
"topic_conf": round(topic_conf[j].item(), 3)
})
return results
# ================= Example Usage =================
llm = LusakaLangMultiTask(lang_temp=1.2, sent_temp=0.93, topic_temp=1.5)
samples = [
"Driver was rude, shouting all the way",
"Payment failed, money deducted but no ride",
"Support did not reply to my complaint",
"Umudriver alisala sana, alelanda ifintu ifipusa",
]
predictions = llm.predict_batch(samples, conf_threshold=0.5)
for p in predictions:
print(f"TEXT: {p['text']}")
print(f" Language : {p['language']} (conf={p['language_conf']})")
print(f" Sentiment: {p['sentiment']} (conf={p['sentiment_conf']})")
print(f" Topic : {p['topic']} (conf={p['topic_conf']})\n")
Sample Output
TEXT: Driver was rude, shouting all the way
Language : English (conf=0.999)
Sentiment: Negative (conf=0.968)
Topic : Driver Behaviour (conf=0.807)
TEXT: Payment failed, money deducted but no ride
Language : English (conf=0.996)
Sentiment: Neutral (conf=0.708)
Topic : Payment Issue (conf=0.634)
TEXT: Support did not reply to my complaint
Language : English (conf=1.0)
Sentiment: Negative (conf=0.960)
Topic : Customer Support (conf=0.984)
TEXT: Umudriver alisala sana, alelanda ifintu ifipusa
Language : Bemba (conf=0.958)
Sentiment: Negative (conf=0.874)
Topic : Driver Behaviour (conf=0.812)
Sentiment by Topic
negative neutral positive
driver_behaviour 82% 14% 4%
payment_issues 76% 20% 4%
app_issues 68% 25% 7%
support_issues 74% 21% 5%
others 29% 56% 15%
Language vs Topic
driver payment app support others
english 38% 24% 15% 10% 13%
bemba 41% 18% 12% 14% 15%
nyanja 36% 19% 17% 13% 15%
mixed 22% 11% 18% 16% 33%
Sample Executive Summary
Out of 10,000 customer complaints:
61% are negative
34% relate to driver behaviour
21% involve payment issues
17% of texts were classified as mixed language
Topic model shows lower confidence than sentiment model
"Others" category remains relatively high (18%)
=========================== Training Architecture ===========================
📥 Input → 🧠 Core Engine → 📈 Output
------------------------------------------------------------------------------------
Text (Any Language) → Tokenizer 🔤 → Language 🌍
→ Shared mBERT Encoder 🧠 → Bemba / Nyanja /
→ CLS Vector 🎯 → English / Mixed
------------------------------------------------------------------------------------
User Feedback 💬 → Tokenizer 🔤 → Sentiment ❤️
→ Shared Encoder 🧠 → Negative / Neutral /
→ CLS Vector 🎯 → Positive
------------------------------------------------------------------------------------
Ride Context 🚗 → Tokenizer 🔤 → Topic 🗂️
→ Shared Encoder 🧠 → Driver / Payment /
→ CLS Vector 🎯 → Support / App / Availability
------------------------------------------------------------------------------------